SAFETY, RELIABILITY AND RISK ANALYSIS: THEORY, METHODS

AND APPLICATIONS

PROCEEDINGS OF THE EUROPEAN SAFETY AND RELIABILITY CONFERENCE, ESREL 2008,

AND 17TH SRA-EUROPE, VALENCIA, SPAIN, SEPTEMBER, 2225, 2008

Safety, Reliability and Risk Analysis:

Theory, Methods and Applications

Editors

Sebastin Martorell
Department of Chemical and Nuclear Engineering,
Universidad Politcnica de Valencia, Spain

C. Guedes Soares
Instituto Superior Tcnico, Technical University of Lisbon, Lisbon, Portugal

Julie Barnett
Department of Psychology, University of Surrey, UK

VOLUME 1

Cover picture designed by Centro de Formacin Permanente - Universidad Politcnica de Valencia

CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business
2009 Taylor & Francis Group, London, UK
Typeset by Vikatan Publishing Solutions (P) Ltd., Chennai, India
Printed and bound in Great Britain by Antony Rowe (A CPI-group Company), Chippenham, Wiltshire.
All rights reserved. No part of this publication or the information contained herein may be reproduced, stored
in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by photocopying,
recording or otherwise, without written prior permission from the publisher.
Although all care is taken to ensure integrity and the quality of this publication and the information herein, no
responsibility is assumed by the publishers nor the author for any damage to the property or persons as a result
of operation or use of this publication and/or the information contained herein.
Published by: CRC Press/Balkema
P.O. Box 447, 2300 AK Leiden, The Netherlands
e-mail: [email protected]
www.crcpress.com www.taylorandfrancis.co.uk www.balkema.nl
ISBN: 978-0-415-48513-5 (set of 4 volumes + CD-ROM)
ISBN: 978-0-415-48514-2 (vol 1)
ISBN: 978-0-415-48515-9 (vol 2)
ISBN: 978-0-415-48516-6 (vol 3)
ISBN: 978-0-415-48792-4 (vol 4)
ISBN: 978-0-203-88297-9 (e-book)

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Table of contents

Preface

XXIV

Organization

XXXI

Acknowledgment

XXXV

Introduction

XXXVII

VOLUME 1

Thematic areas
Accident and incident investigation
A code for the simulation of human failure events in nuclear power plants: SIMPROC
J. Gil, J. Espern, L. Gamo, I. Fernndez, P. Gonzlez, J. Moreno, A. Expsito,
C. Queral, G. Rodrguez & J. Hortal

A preliminary analysis of the Tlahuac incident by applying the MORT technique

J.R. Santos-Reyes, S. Olmos-Pea & L.M. Hernndez-Simn

11

Comparing a multi-linear (STEP) and systemic (FRAM) method for accident analysis
I.A. Herrera & R. Woltjer

19

Development of a database for reporting and analysis of near misses in the Italian
chemical industry
R.V. Gagliardi & G. Astarita

27

Development of incident report analysis system based on m-SHEL ontology

Y. Asada, T. Kanno & K. Furuta

33

Forklifts overturn incidents and prevention in Taiwan

K.Y. Chen, S.-H. Wu & C.-M. Shu

39

Formal modelling of incidents and accidents as a means for enriching training material
for satellite control operations
S. Basnyat, P. Palanque, R. Bernhaupt & E. Poupart

45

Hazard factors analysis in regional traffic records

M. Mlynczak & J. Sipa

57

Organizational analysis of availability: What are the lessons for a high risk industrial company?
M. Voirin, S. Pierlot & Y. Dien

63

Thermal explosion analysis of methyl ethyl ketone peroxide by non-isothermal

and isothermal calorimetry application
S.H. Wu, J.M. Tseng & C.M. Shu

71

Crisis and emergency management

A mathematical model for risk analysis of disaster chains
A. Xuewei Ji, B. Wenguo Weng & Pan Li

79

Effective learning from emergency responses

K. Eriksson & J. Borell

83

On the constructive role of multi-criteria analysis in complex decision-making:

An application in radiological emergency management
C. Turcanu, B. Carl, J. Paridaens & F. Hardeman

89

Decision support systems and software tools for safety and reliability
Complex, expert based multi-role assessment system for small and medium enterprises
S.G. Kovacs & M. Costescu

99

DETECT: A novel framework for the detection of attacks to critical infrastructures

F. Flammini, A. Gaglione, N. Mazzocca & C. Pragliola

105

Methodology and software platform for multi-layer causal modeling

K.M. Groth, C. Wang, D. Zhu & A. Mosleh

113

SCAIS (Simulation Code System for Integrated Safety Assessment): Current

status and applications
J.M. Izquierdo, J. Hortal, M. Snchez, E. Melndez, R. Herrero, J. Gil, L. Gamo,
I. Fernndez, J. Espern, P. Gonzlez, C. Queral, A. Expsito & G. Rodrguez
Using GIS and multivariate analyses to visualize risk levels and spatial patterns
of severe accidents in the energy sector
P. Burgherr
Weak signals of potential accidents at Seveso establishments
P.A. Bragatto, P. Agnello, S. Ansaldi & P. Pittiglio

121

129
137

Dynamic reliability
A dynamic fault classification scheme
B. Fechner

147

Importance factors in dynamic reliability

R. Eymard, S. Mercier & M. Roussignol

155

TSD, a SCAIS suitable variant of the SDTPD

J.M. Izquierdo & I. Caamn

163

Fault identification and diagnostics

Application of a vapour compression chiller lumped model for fault detection
J. Navarro-Esbr, A. Real, D. Ginestar & S. Martorell

175

Automatic source code analysis of failure modes causing error propagation

S. Sarshar & R. Winther

183

Development of a prognostic tool to perform reliability analysis

M. El-Koujok, R. Gouriveau & N. Zerhouni

191

Fault detection and diagnosis in monitoring a hot dip galvanizing line using
multivariate statistical process control
J.C. Garca-Daz

201

Fault identification, diagnosis and compensation of flatness errors in hard turning

of tool steels
F. Veiga, J. Fernndez, E. Viles, M. Arizmendi, A. Gil & M.L. Penalva

205

VI

From diagnosis to prognosis: A maintenance experience for an electric locomotive

O. Borgia, F. De Carlo & M. Tucci

211

Human factors
A study on the validity of R-TACOM measure by comparing operator response
time data
J. Park & W. Jung
An evaluation of the Enhanced Bayesian THERP method using simulator data
K. Bladh, J.-E. Holmberg & P. Pyy

221
227

Comparing CESA-Q human reliability analysis with evidence from simulator:

A first attempt
L. Podofillini & B. Reer

233

Exploratory and confirmatory analysis of the relationship between social norms

and safety behavior
C. Fugas, S.A. da Silva & J.L. Meli

243

Functional safety and layer of protection analysis with regard to human factors
K.T. Kosmowski

249

How employees use of information technology systems shape reliable operations

of large scale technological systems
T.K. Andersen, P. Nsje, H. Torvatn & K. Skarholt

259

Incorporating simulator evidence into HRA: Insights from the data analysis of the
international HRA empirical study
S. Massaiu, P.. Braarud & M. Hildebrandt

267

Insights from the HRA international empirical study: How to link data
and HRA with MERMOS
H. Pesme, P. Le Bot & P. Meyer

275

Operators response time estimation for a critical task using the fuzzy logic theory
M. Konstandinidou, Z. Nivolianitou, G. Simos, C. Kiranoudis & N. Markatos

281

The concept of organizational supportiveness

J. Nicholls, J. Harvey & G. Erdos

291

The influence of personal variables on changes in driver behaviour

S. Heslop, J. Harvey, N. Thorpe & C. Mulley

299

The key role of expert judgment in CO2 underground storage projects

C. Vivalda & L. Jammes

305

Integrated risk management and riskinformed decision-making

All-hazards risk frameworkAn architecture model
S. Verga

315

Comparisons and discussion of different integrated risk approaches

R. Steen & T. Aven

323

Management of risk caused by domino effect resulting from design

system dysfunctions
S. Sperandio, V. Robin & Ph. Girard

331

On some aspects related to the use of integrated risk analyses for the decision
making process, including its use in the non-nuclear applications
D. Serbanescu, A.L. Vetere Arellano & A. Colli

341

On the usage of weather derivatives in AustriaAn empirical study

M. Bank & R. Wiesner

VII

351

Precaution in practice? The case of nanomaterial industry

H. Kastenholz, A. Helland & M. Siegrist

361

Risk based maintenance prioritisation

G. Birkeland, S. Eisinger & T. Aven

365

Shifts in environmental health risk governance: An analytical framework

H.A.C. Runhaar, J.P. van der Sluijs & P.P.J. Driessen

369

What does safety margin really mean?

J. Hortal, R. Mendizbal & F. Pelayo

379

Legislative dimensions of risk management

Accidents, risk analysis and safety managementDifferent perspective at a
Swedish safety authority
O. Harrami, M. Strmgren, U. Postgrd & R. All
Evaluation of risk and safety issues at the Swedish Rescue Services Agency
O. Harrami, U. Postgrd & M. Strmgren
Regulation of information security and the impact on top management commitment
A comparative study of the electric power supply sector and the finance sector
J.M. Hagen & E. Albrechtsen
The unintended consequences of risk regulation
B.H. MacGillivray, R.E. Alcock & J.S. Busby

391
399

407
415

Maintenance modelling and optimisation

A hybrid age-based maintenance policy for heterogeneous items
P.A. Scarf, C.A.V. Cavalcante, R.W. Dwight & P. Gordon

423

A stochastic process model for computing the cost of a condition-based maintenance plan
J.A.M. van der Weide, M.D. Pandey & J.M. van Noortwijk

431

A study about influence of uncertain distribution inputs in maintenance optimization

R. Mullor, S. Martorell, A. Snchez & N. Martinez-Alzamora

441

Aging processes as a primary aspect of predicting reliability and life of aeronautical hardware

J. Zurek,
M. Zieja, G. Kowalczyk & T. Niezgoda

449

An alternative imperfect preventive maintenance model

J. Clavareau & P.E. Labeau

455

An imperfect preventive maintenance model with dependent failure modes

I.T. Castro

463

Condition-based maintenance approaches for deteriorating system influenced

by environmental conditions
E. Deloux, B. Castanier & C. Brenguer

469

Condition-based maintenance by particle filtering

F. Cadini, E. Zio & D. Avram

477

Corrective maintenance for aging air conditioning systems

I. Frenkel, L. Khvatskin & A. Lisnianski

483

Exact reliability quantification of highly reliable systems with maintenance

R. Bri

489

Genetic algorithm optimization of preventive maintenance scheduling for repairable

systems modeled by generalized renewal process
P.A.A. Garcia, M.C. SantAna, V.C. Damaso & P.F. Frutuoso e Melo

VIII

497

Maintenance modelling integrating human and material resources

S. Martorell, M. Villamizar, A. Snchez & G. Clemente

505

Modelling competing risks and opportunistic maintenance with expert judgement

T. Bedford & B.M. Alkali

515

Modelling different types of failure and residual life estimation for condition-based maintenance
M.J. Carr & W. Wang

523

Multi-component systems modeling for quantifying complex maintenance strategies

V. Zille, C. Brenguer, A. Grall, A. Despujols & J. Lonchampt

531

Multiobjective optimization of redundancy allocation in systems with imperfect repairs via

ant colony and discrete event simulation
I.D. Lins & E. Lpez Droguett

541

Non-homogeneous Markov reward model for aging multi-state system under corrective
maintenance
A. Lisnianski & I. Frenkel

551

On the modeling of ageing using Weibull models: Case studies

P. Praks, H. Fernandez Bacarizo & P.-E. Labeau

559

On-line condition-based maintenance for systems with several modes of degradation

A. Ponchet, M. Fouladirad & A. Grall

567

Opportunity-based age replacement for a system under two types of failures

F.G. Bada & M.D. Berrade

575

Optimal inspection intervals for maintainable equipment

O. Hryniewicz

581

Optimal periodic inspection of series systems with revealed and unrevealed failures
M. Carvalho, E. Nunes & J. Telhada

587

Optimal periodic inspection/replacement policy for deteriorating systems with explanatory

variables
X. Zhao, M. Fouladirad, C. Brenguer & L. Bordes

593

Optimal replacement policy for components with general failure rates submitted to obsolescence
S. Mercier

603

Optimization of the maintenance function at a company

S. Adjabi, K. Adel-Aissanou & M. Azi

611

Planning and scheduling maintenance resources in a complex system

M. Newby & C. Barker

619

Preventive maintenance planning using prior expert knowledge and multicriteria method
PROMETHEE III
F.A. Figueiredo, C.A.V. Cavalcante & A.T. de Almeida
Profitability assessment of outsourcing maintenance from the producer (big rotary machine study)
P. Fuchs & J. Zajicek

627
635

Simulated annealing method for the selective maintenance optimization of multi-mission

series-parallel systems
A. Khatab, D. Ait-Kadi & A. Artiba

641

Study on the availability of a k-out-of-N System given limited spares under (m, NG )
maintenance policy
T. Zhang, H.T. Lei & B. Guo

649

System value trajectories, maintenance, and its present value

K.B. Marais & J.H. Saleh

IX

659

The maintenance management framework: A practical view to maintenance management

A. Crespo Mrquez, P. Moreu de Len, J.F. Gmez Fernndez, C. Parra Mrquez & V. Gonzlez
Workplace occupation and equipment availability and utilization, in the context of maintenance
float systems
I.S. Lopes, A.F. Leito & G.A.B. Pereira

669

675

Monte Carlo methods in system safety and reliability

Availability and reliability assessment of industrial complex systems: A practical view
applied on a bioethanol plant simulation
V. Gonzlez, C. Parra, J.F. Gmez, A. Crespo & P. Moreu de Len
Handling dependencies between variables with imprecise probabilistic models
S. Destercke & E. Chojnacki
Monte Carlo simulation for investigating the influence of maintenance strategies on the production
availability of offshore installations
K.P. Chang, D. Chang, T.J. Rhee & E. Zio

687
697

703

Reliability analysis of discrete multi-state systems by means of subset simulation

E. Zio & N. Pedroni

709

The application of Bayesian interpolation in Monte Carlo simulations

M. Rajabalinejad, P.H.A.J.M. van Gelder & N. van Erp

717

Occupational safety
Application of virtual reality technologies to improve occupational & industrial safety
in industrial processes
J. Rubio, B. Rubio, C. Vaquero, N. Galarza, A. Pelaz, J.L. Ipia, D. Sagasti & L. Jord
Applying the resilience concept in practice: A case study from the oil and gas industry
L. Hansson, I. Andrade Herrera, T. Kongsvik & G. Solberg

727
733

Development of an assessment tool to facilitate OHS management based upon the safe
place, safe person, safe systems framework
A.-M. Makin & C. Winder

739

Exploring knowledge translation in occupational health using the mental models approach:
A case study of machine shops
A.-M. Nicol & A.C. Hurrell

749

Mathematical modelling of risk factors concerning work-related traffic accidents

C. Santamara, G. Rubio, B. Garca & E. Navarro

757

New performance indicators for the health and safety domain: A benchmarking use perspective
H.V. Neto, P.M. Arezes & S.D. Sousa

761

Occupational risk management for fall from height

O.N. Aneziris, M. Konstandinidou, I.A. Papazoglou, M. Mud, M. Damen, J. Kuiper, H. Baksteen,
L.J. Bellamy, J.G. Post & J. Oh

767

Occupational risk management for vapour/gas explosions

I.A. Papazoglou, O.N. Aneziris, M. Konstandinidou, M. Mud, M. Damen, J. Kuiper, A. Bloemhoff,
H. Baksteen, L.J. Bellamy, J.G. Post & J. Oh

777

Occupational risk of an aluminium industry

O.N. Aneziris, I.A. Papazoglou & O. Doudakmani

787

Risk regulation bureaucracies in EU accession states: Drinking water safety in Estonia

K. Kangur

797

Organization learning
Can organisational learning improve safety and resilience during changes?
S.O. Johnsen & S. Hbrekke

805

Consequence analysis as organizational development

B. Moltu, A. Jarl Ringstad & G. Guttormsen

813

Integrated operations and leadershipHow virtual cooperation influences leadership practice

K. Skarholt, P. Nsje, V. Heps & A.S. Bye

821

Outsourcing maintenance in services providers

J.F. Gmez, C. Parra, V. Gonzlez, A. Crespo & P. Moreu de Len

829

Revising rules and reviving knowledge in the Norwegian railway system

H.C. Blakstad, R. Rosness & J. Hovden

839

Risk Management in systems: Learning to recognize and respond to weak signals

E. Guillaume

847

Author index

853

VOLUME 2
Reliability and safety data collection and analysis
A new step-stress Accelerated Life Testing approach: Step-Down-Stress
C. Zhang, Y. Wang, X. Chen & Y. Jiang

863

Application of a generalized lognormal distribution to engineering data fitting

J. Martn & C.J. Prez

869

Collection and analysis of reliability data over the whole product lifetime of vehicles
T. Leopold & B. Bertsche

875

Comparison of phase-type distributions with mixed and additive Weibull models

M.C. Segovia & C. Guedes Soares

881

Evaluation methodology of industry equipment functional reliability

J. Kamenick

891

Evaluation of device reliability based on accelerated tests

E. Nogueira Daz, M. Vzquez Lpez & D. Rodrguez Cano

899

Evaluation, analysis and synthesis of multiple source information: An application to nuclear

computer codes
S. Destercke & E. Chojnacki

905

Improving reliability using new processes and methods

S.J. Park, S.D. Park & K.T. Jo

913

Life test applied to Brazilian friction-resistant low alloy-high strength steel rails
D.I. De Souza, A. Naked Haddad & D. Rocha Fonseca

919

Non-homogeneous Poisson Process (NHPP), stochastic model applied to evaluate the economic
impact of the failure in the Life Cycle Cost Analysis (LCCA)
C. Parra Mrquez, A. Crespo Mrquez, P. Moreu de Len, J. Gmez Fernndez & V. Gonzlez Daz

929

Risk trends, indicators and learning rates: A new case study of North sea oil and gas
R.B. Duffey & A.B. Skjerve

941

Robust estimation for an imperfect test and repair model using Gaussian mixtures
S.P. Wilson & S. Goyal

949

XI

Risk and evidence based policy making

Environmental reliability as a requirement for defining environmental impact limits
in critical areas
E. Calixto & E. Lbre La Rovere

957

Hazardous aid? The crowding-out effect of international charity

P.A. Raschky & M. Schwindt

965

Individual risk-taking and external effectsAn empirical examination

S. Borsky & P.A. Raschky

973

Licensing a Biofuel plan transforming animal fats

J.-F. David

981

Modelling incident escalation in explosives storage

G. Hardman, T. Bedford, J. Quigley & L. Walls

987

The measurement and management of Deca-BDEWhy the continued certainty of uncertainty?

R.E. Alcock, B.H. McGillivray & J.S. Busby

993

Risk and hazard analysis

A contribution to accelerated testing implementation
S. Hossein Mohammadian M., D. At-Kadi, A. Coulibaly & B. Mutel

1001

A decomposition method to analyze complex fault trees

S. Contini

1009

A quantitative methodology for risk assessment of explosive atmospheres according to the

ATEX directive
R. Lisi, M.F. Milazzo & G. Maschio
A risk theory based on hyperbolic parallel curves and risk assessment in time
N. Popoviciu & F. Baicu
Accident occurrence evaluation of phased-mission systems composed of components
with multiple failure modes
T. Kohda

1019
1027

1035

Added value in fault tree analyses

T. Norberg, L. Rosn & A. Lindhe

1041

Alarm prioritization at plant design stageA simplified approach

P. Barbarini, G. Franzoni & E. Kulot

1049

Analysis of possibilities of timing dependencies modelingExample of logistic support system

J. Magott, T. Nowakowski, P. Skrobanek & S. Werbinska

1055

Applications of supply process reliability model

A. Jodejko & T. Nowakowski

1065

Applying optimization criteria to risk analysis

H. Medina, J. Arnaldos & J. Casal

1073

Chemical risk assessment for inspection teams during CTBT on-site inspections of sites
potentially contaminated with industrial chemicals
G. Malich & C. Winder

1081

Comparison of different methodologies to estimate the evacuation radius in the case

of a toxic release
M.I. Montoya & E. Planas

1089

Conceptualizing and managing risk networks. New insights for risk management
R.W. Schrder

XII

1097

Developments in fault tree techniques and importance measures

J.K. Vaurio

1103

Dutch registration of risk situations

J.P. vant Sant, H.J. Manuel & A. van den Berg

1113

Experimental study of jet fires

M. Gmez-Mares, A. Palacios, A. Peiretti, M. Muoz & J. Casal

1119

Failure mode and effect analysis algorithm for tunneling projects

K. Rezaie, V. Ebrahimipour & S. Shokravi

1125

Fuzzy FMEA: A study case on a discontinuous distillation plant

S.S. Rivera & J.E. Nez Mc Leod

1129

Risk analysis in extreme environmental conditions for Aconcagua Mountain station

J.E. Nez Mc Leod & S.S. Rivera

1135

Geographic information system for evaluation of technical condition and residual life of pipelines
P. Yukhymets, R. Spitsa & S. Kobelsky

1141

Inherent safety indices for the design of layout plans

A. Tugnoli, V. Cozzani, F.I. Khan & P.R. Amyotte

1147

Minmax defense strategy for multi-state systems

G. Levitin & K. Hausken

1157

Multicriteria risk assessment for risk ranking of natural gas pipelines

A.J. de M. Brito, C.A.V. Cavalcante, R.J.P. Ferreira & A.T. de Almeida

1165

New insight into PFDavg and PFH

F. Innal, Y. Dutuit, A. Rauzy & J.-P. Signoret

1173

On causes and dependencies of errors in human and organizational barriers against major
accidents
J.E. Vinnem

1181

Quantitative risk analysis method for warehouses with packaged hazardous materials
D. Riedstra, G.M.H. Laheij & A.A.C. van Vliet

1191

Ranking the attractiveness of industrial plants to external acts of interference

M. Sabatini, S. Zanelli, S. Ganapini, S. Bonvicini & V. Cozzani

1199

Review and discussion of uncertainty taxonomies used in risk analysis

T.E. Nkland & T. Aven

1207

Risk analysis in the frame of the ATEX Directive and the preparation of an Explosion Protection
Document
A. Pey, G. Suter, M. Glor, P. Lerena & J. Campos

1217

Risk reduction by use of a buffer zone

S.I. Wijnant-Timmerman & T. Wiersma

1223

Safety in engineering practice

Z. Smalko & J. Szpytko

1231

Why ISO 13702 and NFPA 15 standards may lead to unsafe design
S. Medonos & R. Raman

1239

Risk control in complex environments

Is there an optimal type for high reliability organization? A study of the UK offshore industry
J.S. Busby, A. Collins & R. Miles

1251

The optimization of system safety: Rationality, insurance, and optimal protection

R.B. Jongejan & J.K. Vrijling

1259

XIII

Thermal characteristic analysis of Y type zeolite by differential scanning calorimetry

S.H. Wu, W.P. Weng, C.C. Hsieh & C.M. Shu
Using network methodology to define emergency response team location: The Brazilian
refinery case study
E. Calixto, E. Lbre La Rovere & J. Eustquio Beraldo

1267

1273

Risk perception and communication

(Mis-)conceptions of safety principles
J.-T. Gayen & H. Schbe

1283

Climate change in the British press: The role of the visual

N.W. Smith & H. Joffe

1293

Do the people exposed to a technological risk always want more information about it?
Some observations on cases of rejection
J. Espluga, J. Farr, J. Gonzalo, T. Horlick-Jones, A. Prades, C. Oltra & J. Navajas

1301

Media coverage, imaginary of risks and technological organizations

F. Fodor & G. Deleuze

1309

Media disaster coverage over time: Methodological issues and results

M. Kuttschreuter & J.M. Gutteling

1317

Risk amplification and zoonosis

D.G. Duckett & J.S. Busby

1325

Risk communication and addressing uncertainties in risk assessmentsPresentation of a framework

J. Stian strem, H. Thevik, R. Flage & T. Aven

1335

Risk communication for industrial plants and radioactive waste repositories

F. Gugliermetti & G. Guidi

1341

Risk management measurement methodology: Practical procedures and approaches for risk
assessment and prediction
R.B. Duffey & J.W. Saull

1351

Risk perception and cultural theory: Criticism and methodological orientation

C. Kermisch & P.-E. Labeau

1357

Standing in the shoes of hazard managers: An experiment on avalanche risk perception

C.M. Rheinberger

1365

The social perception of nuclear fusion: Investigating lay understanding and reasoning about
the technology
A. Prades, C. Oltra, J. Navajas, T. Horlick-Jones & J. Espluga

1371

Safety culture
Us and Them: The impact of group identity on safety critical behaviour
R.J. Bye, S. Antonsen & K.M. Vikland

1377

Does change challenge safety? Complexity in the civil aviation transport system
S. Hyland & K. Aase

1385

Electromagnetic fields in the industrial enviroment

J. Fernndez, A. Quijano, M.L. Soriano & V. Fuster

1395

Electrostatic charges in industrial environments

P. LLovera, A. Quijano, A. Soria & V. Fuster

1401

Empowering operations and maintenance: Safe operations with the one directed team
organizational model at the Kristin asset
P. Nsje, K. Skarholt, V. Heps & A.S. Bye

XIV

1407

Leadership and safety climate in the construction industry

J.L. Meli, M. Becerril, S.A. Silva & K. Mearns

1415

Local management and its impact on safety culture and safety within Norwegian shipping
H.A Oltedal & O.A. Engen

1423

Quantitative analysis of the anatomy and effectiveness of occupational safety culture

P. Trucco, M. De Ambroggi & O. Grande

1431

Safety management and safety culture assessment in Germany

H.P. Berg

1439

The potential for error in communications between engineering designers

J. Harvey, R. Jamieson & K. Pearce

1447

Safety management systems

Designing the safety policy of IT system on the example of a chosen company
I.J. Jzwiak, T. Nowakowski & K. Jzwiak
Determining and verifying the safety integrity level of the control and protection systems
under uncertainty
T. Barnert, K.T. Kosmowski & M. Sliwinski

1455

1463

Drawing up and running a Security Plan in an SME type companyAn easy task?
M. Gerbec

1473

Efficient safety management for subcontractor at construction sites

K.S. Son & J.J. Park

1481

Production assurance and reliability managementA new international standard

H. Kortner, K.-E. Haugen & L. Sunde

1489

Risk management model for industrial plants maintenance

N. Napolitano, M. De Minicis, G. Di Gravio & M. Tronci

1495

Some safety aspects on multi-agent and CBTC implementation for subway control systems
F.M. Rachel & P.S. Cugnasca

1503

Software reliability
Assessment of software reliability and the efficiency of corrective actions during the software
development process
R. Savic

1513

ERTMS, deals on wheels? An inquiry into a major railway project

J.A. Stoop, J.H. Baggen, J.M. Vleugel & J.L.M. Vrancken

1519

Guaranteed resource availability in a website

V.P. Koutras & A.N. Platis

1525

Reliability oriented electronic design automation tool

J. Marcos, D. Bveda, A. Fernndez & E. Soto

1533

Reliable software for partitionable networked environmentsAn experience report

S. Beyer, J.C. Garca Ortiz, F.D. Muoz-Esco, P. Galdmez, L. Froihofer,
K.M. Goeschka & J. Osrael

1539

SysML aided functional safety assessment

M. Larisch, A. Hnle, U. Siebold & I. Hring

1547

UML safety requirement specification and verification

A. Hnle & I. Hring

1555

XV

Stakeholder and public involvement in risk governance

Assessment and monitoring of reliability and robustness of offshore wind energy converters
S. Thns, M.H. Faber, W. Rcker & R. Rohrmann

1567

Building resilience to natural hazards. Practices and policies on governance and mitigation
in the central region of Portugal
J.M. Mendes & A.T. Tavares

1577

Governance of flood risks in The Netherlands: Interdisciplinary research into the role and
meaning of risk perception
M.S. de Wit, H. van der Most, J.M. Gutteling & M. Bockarjova

1585

Public intervention for better governanceDoes it matter? A study of the Leros Strength case
P.H. Linde & J.E. Karlsen

1595

Reasoning about safety management policy in everyday terms

T. Horlick-Jones

1601

Using stakeholders expertise in EMF and soil contamination to improve the management
of public policies dealing with modern risk: When uncertainty is on the agenda
C. Fallon, G. Joris & C. Zwetkoff

1609

Structural reliability and design codes

Adaptive discretization of 1D homogeneous random fields
D.L. Allaix, V.I. Carbone & G. Mancini

1621

Comparison of methods for estimation of concrete strength

M. Holicky, K. Jung & M. Sykora

1629

Design of structures for accidental design situations

J. Markov & K. Jung

1635

Developing fragility function for a timber structure subjected to fire

E.R. Vaidogas, Virm. Juocevicius & Virg. Juocevicius

1641

Estimations in the random fatigue-limit model

C.-H. Chiu & W.-T. Huang

1651

Limitations of the Weibull distribution related to predicting the probability of failure

initiated by flaws
M.T. Todinov

1655

Simulation techniques of non-gaussian random loadings in structural reliability analysis

Y. Jiang, C. Zhang, X. Chen & J. Tao

1663

Special features of the collection and analysis of snow loads

Z. Sadovsk, P. Fako, K. Mikulov, J. Pecho & M. Vojtek

1671

Structural safety under extreme construction loads

V. Juocevicius & A. Kudzys

1677

The modeling of time-dependent reliability of deteriorating structures

A. Kudzys & O. Lukoeviciene

1685

Author index

1695

VOLUME 3
System reliability analysis
A copula-based approach for dependability analyses of fault-tolerant systems with
interdependent basic events
M. Walter, S. Esch & P. Limbourg

XVI

1705

A depth first search algorithm for optimal arrangements in a circular

consecutive-k-out-of-n:F system
K. Shingyochi & H. Yamamoto

1715

A joint reliability-redundancy optimization approach for multi-state series-parallel systems

Z. Tian, G. Levitin & M.J. Zuo

1723

A new approach to assess the reliability of a multi-state system with dependent components
M. Samrout & E. Chatelet

1731

A reliability analysis and decision making process for autonomous systems

R. Remenyte-Prescott, J.D. Andrews, P.W.H. Chung & C.G. Downes

1739

Advanced discrete event simulation methods with application to importance measure

estimation
A.B. Huseby, K.A. Eide, S.L. Isaksen, B. Natvig & J. Gsemyr

1747

Algorithmic and computational analysis of a multi-component complex system

J.E. Ruiz-Castro, R. Prez-Ocn & G. Fernndez-Villodre

1755

An efficient reliability computation of generalized multi-state k-out-of-n systems

S.V. Amari

1763

Application of the fault tree analysis for assessment of the power system reliability

A. Volkanovski, M. Cepin
& B. Mavko

1771

BDMP (Boolean logic driven Markov processes) as an alternative to event trees

M. Bouissou

1779

Bivariate distribution based passive system performance assessment

L. Burgazzi

1787

Calculating steady state reliability indices of multi-state systems using dual number algebra
E. Korczak

1795

Concordance analysis of importance measure

C.M. Rocco S.

1803

Contribution to availability assessment of systems with one shot items

D. Valis & M. Koucky

1807

Contribution to modeling of complex weapon systems reliability

D. Valis, Z. Vintr & M. Koucky

1813

Delayed system reliability and uncertainty analysis

R. Alzbutas, V. Janilionis & J. Rimas

1819

Efficient generation and representation of failure lists out of an information flux model
for modeling safety critical systems
M. Pock, H. Belhadaoui, O. Malass & M. Walter

1829

Evaluating algorithms for the system state distribution of multi-state k-out-of-n:F system
T. Akiba, H. Yamamoto, T. Yamaguchi, K. Shingyochi & Y. Tsujimura

1839

First-passage time analysis for Markovian deteriorating model

G. Dohnal

1847

Model of logistic support system with time dependency

S. Werbinska

1851

Modeling failure cascades in network systems due to distributed random disturbances

E. Zio & G. Sansavini

1861

Modeling of the changes of graphite bore in RBMK-1500 type nuclear reactor

I. utautaite-eputiene, J. Augutis & E. Upuras

1867

XVII

Modelling multi-platform phased mission system reliability

D.R. Prescott, J.D. Andrews & C.G. Downes
Modelling test strategies effects on the probability of failure on demand for safety
instrumented systems
A.C. Torres-Echeverria, S. Martorell & H.A. Thompson

1873

1881

New insight into measures of component importance in production systems

S.L. Isaksen

1891

New virtual age models for bathtub shaped failure intensities

Y. Dijoux & E. Ide

1901

On some approaches to defining virtual age of non-repairable objects

M.S. Finkelstein

1909

On the application and extension of system signatures in engineering reliability

J. Navarro, F.J. Samaniego, N. Balakrishnan & D. Bhattacharya

1915

PFD of higher-order configurations of SIS with partial stroke testing capability

L.F.S. Oliveira

1919

Power quality as accompanying factor in reliability research of electric engines

I.J. Jzwiak, K. Kujawski & T. Nowakowski

1929

RAMS and performance analysis

X. Quayzin, E. Arbaretier, Z. Brik & A. Rauzy

1937

Reliability evaluation of complex system based on equivalent fault tree

Z. Yufang, Y. Hong & L. Jun

1943

Reliability evaluation of III-V Concentrator solar cells

N. Nez, J.R. Gonzlez, M. Vzquez, C. Algora & I. Rey-Stolle

1949

Reliability of a degrading system under inspections

D. Montoro-Cazorla, R. Prez-Ocn & M.C. Segovia

1955

Reliability prediction using petri nets for on-demand safety systems with fault detection
A.V. Kleyner & V. Volovoi

1961

Reliability, availability and cost analysis of large multi-state systems with ageing components
K. Kolowrocki

1969

Reliability, availability and risk evaluation of technical systems in variable operation conditions
K. Kolowrocki & J. Soszynska

1985

Representation and estimation of multi-state system reliability by decision diagrams

E. Zaitseva & S. Puuronen

1995

Safety instrumented system reliability evaluation with influencing factors

F. Brissaud, D. Charpentier, M. Fouladirad, A. Barros & C. Brenguer

2003

Smooth estimation of the availability function of a repairable system

M.L. Gmiz & Y. Romn

2013

System design optimisation involving phased missions

D. Astapenko & L.M. Bartlett

2021

The Natvig measures of component importance in repairable systems applied to an offshore

oil and gas production system
B. Natvig, K.A. Eide, J. Gsemyr, A.B. Huseby & S.L. Isaksen

2029

The operation quality assessment as an initial part of reliability improvement and low cost
automation of the system
L. Muslewski, M. Woropay & G. Hoppe

2037

XVIII

Three-state modelling of dependent component failures with domino effects

U.K. Rakowsky

2045

Variable ordering techniques for the application of Binary Decision Diagrams on PSA
linked Fault Tree models
C. Ibez-Llano, A. Rauzy, E. Melndez & F. Nieto

2051

Weaknesses of classic availability calculations for interlinked production systems

and their overcoming
D. Achermann

2061

Uncertainty and sensitivity analysis

A critique of Info-Gaps robustness model
M. Sniedovich

2071

Alternative representations of uncertainty in system reliability and risk analysisReview

and discussion
R. Flage, T. Aven & E. Zio

2081

Dependence modelling with copula in probabilistic studies, a practical approach

based on numerical experiments
A. Dutfoy & R. Lebrun

2093

Event tree uncertainty analysis by Monte Carlo and possibility theory

P. Baraldi & E. Zio

2101

Global sensitivity analysis based on entropy

B. Auder & B. Iooss

2107

Impact of uncertainty affecting reliability models on warranty contracts

G. Fleurquin, P. Dehombreux & P. Dersin

2117

Influence of epistemic uncertainties on the probabilistic assessment of an emergency operating

procedure in a nuclear power plant
M. Kloos & J. Peschke

2125

Numerical study of algorithms for metamodel construction and validation

B. Iooss, L. Boussouf, A. Marrel & V. Feuillard

2135

On the variance upper bound theorem and its applications

M.T. Todinov

2143

Reliability assessment under Uncertainty Using Dempster-Shafer and Vague Set Theories
S. Pashazadeh & N. Grachorloo

2151

Types and sources of uncertainties in environmental accidental risk assessment: A case study
for a chemical factory in the Alpine region of Slovenia
M. Gerbec & B. Kontic
Uncertainty estimation for monotone and binary systems
A.P. Ulmeanu & N. Limnios

2157
2167

Industrial and service sectors

Aeronautics and aerospace
Condition based operational risk assessment for improved aircraft operability
A. Arnaiz, M. Buderath & S. Ferreiro

2175

Is optimized design of satellites possible?

J. Faure, R. Laulheret & A. Cabarbaye

2185

XIX

Model of air traffic in terminal area for ATFM safety analysis

J. Skorupski & A.W. Stelmach

2191

Predicting airport runway conditions based on weather data

A.B. Huseby & M. Rabbe

2199

Safety considerations in complex airborne systems

M.J.R. Lemes & J.B. Camargo Jr

2207

The Preliminary Risk Analysis approach: Merging space and aeronautics methods
J. Faure, R. Laulheret & A. Cabarbaye

2217

Using a Causal model for Air Transport Safety (CATS) for the evaluation of alternatives
B.J.M. Ale, L.J. Bellamy, R.P. van der Boom, J. Cooper, R.M. Cooke, D. Kurowicka, P.H. Lin,
O. Morales, A.L.C. Roelen & J. Spouge

2223

Automotive engineering
An approach to describe interactions in and between mechatronic systems
J. Gng & B. Bertsche

2233

Influence of the mileage distribution on reliability prognosis models

A. Braasch, D. Althaus & A. Meyna

2239

Reliability prediction for automotive components using Real-Parameter Genetic Algorithm

J. Hauschild, A. Kazeminia & A. Braasch

2245

Stochastic modeling and prediction of catalytic converters degradation

S. Barone, M. Giorgio, M. Guida & G. Pulcini

2251

Towards a better interaction between design and dependability analysis: FMEA derived from
UML/SysML models
P. David, V. Idasiak & F. Kratz

2259

Biotechnology and food industry

Application of tertiary mathematical models for evaluating the presence of staphylococcal
enterotoxin in lactic acid cheese
I. Steinka & A. Blokus-Roszkowska

2269

Assessment of the risk to company revenue due to deviations in honey quality

E. Domnech, I. Escriche & S. Martorell

2275

Attitudes of Japanese and Hawaiian toward labeling genetically modified fruits

S. Shehata

2285

Ensuring honey quality by means of effective pasteurization

E. Domnech, I. Escriche & S. Martorell

2289

Exposure assessment model to combine thermal inactivation (log reduction) and thermal injury
(heat-treated spore lag time) effects on non-proteolytic Clostridium botulinum
J.-M. Membr, E. Wemmenhove & P. McClure

2295

Public information requirements on health risk of mercury in fish (1): Perception

and knowledge of the public about food safety and the risk of mercury
M. Kosugi & H. Kubota

2305

Public information requirements on health risks of mercury in fish (2): A comparison of mental
models of experts and public in Japan
H. Kubota & M. Kosugi

2311

Review of diffusion models for the social amplification of risk of food-borne zoonoses
J.P. Mehers, H.E. Clough & R.M. Christley

XX

2317

Risk perception and communication of food safety and food technologies in Flanders,
The Netherlands, and the United Kingdom
U. Maris
Synthesis of reliable digital microfluidic biochips using Monte Carlo simulation
E. Maftei, P. Pop & F. Popentiu Vladicescu

2325
2333

Chemical process industry

Accidental scenarios in the loss of control of chemical processes: Screening the impact
profile of secondary substances
M. Cordella, A. Tugnoli, P. Morra, V. Cozzani & F. Barontini
Adapting the EU Seveso II Directive for GHS: Initial UK study on acute toxicity to people
M.T. Trainor, A.J. Wilday, M. Moonis, A.L. Rowbotham, S.J. Fraser, J.L. Saw & D.A. Bosworth

2345
2353

An advanced model for spreading and evaporation of accidentally released hazardous

liquids on land
I.J.M. Trijssenaar-Buhre, R.P. Sterkenburg & S.I. Wijnant-Timmerman

2363

Influence of safety systems on land use planning around seveso sites; example of measures
chosen for a fertiliser company located close to a village
C. Fivez, C. Delvosalle, N. Cornil, L. Servranckx, F. Tambour, B. Yannart & F. Benjelloun

2369

Performance evaluation of manufacturing systems based on dependability management

indicators-case study: Chemical industry
K. Rezaie, M. Dehghanbaghi & V. Ebrahimipour

2379

Protection of chemical industrial installations from intentional adversary acts: Comparison

of the new security challenges with the existing safety practices in Europe
M.D. Christou

2389

Quantitative assessment of domino effect in an extended industrial area

G. Antonioni, G. Spadoni, V. Cozzani, C. Dondi & D. Egidi

2397

Reaction hazard of cumene hydroperoxide with sodium hydroxide by isothermal calorimetry

Y.-P. Chou, S.-H. Wu, C.-M. Shu & H.-Y. Hou

2405

Reliability study of shutdown process through the analysis of decision making in chemical plants.
Case of study: South America, Spain and Portugal
L. Amendola, M.A. Artacho & T. Depool

2409

Study of the application of risk management practices in shutdown chemical process

L. Amendola, M.A. Artacho & T. Depool

2415

Thirty years after the first HAZOP guideline publication. Considerations

J. Dunj, J.A. Vlchez & J. Arnaldos

2421

Civil engineering
Decision tools for risk management support in construction industry
S. Mehicic Eberhardt, S. Moeller, M. Missler-Behr & W. Kalusche

2431

Definition of safety and the existence of optimal safety

D. Proske

2441

Failure risk analysis in Water Supply Networks

A. Carrin, A. Debn, E. Cabrera, M.L. Gamiz & H. Solano

2447

Hurricane vulnerability of multi-story residential buildings in Florida

G.L. Pita, J.-P. Pinelli, C.S. Subramanian, K. Gurley & S. Hamid

2453

Risk management system in water-pipe network functioning

B. Tchrzewska-Cieslak

2463

XXI

Use of extreme value theory in engineering design

E. Castillo, C. Castillo & R. Mnguez

2473

Critical infrastructures
A model for vulnerability analysis of interdependent infrastructure networks
J. Johansson & H. Jnsson
Exploiting stochastic indicators of interdependent infrastructures: The service availability of
interconnected networks
G. Bonanni, E. Ciancamerla, M. Minichino, R. Clemente, A. Iacomini, A. Scarlatti,
E. Zendri & R. Terruggia
Proactive risk assessment of critical infrastructures
T. Uusitalo, R. Koivisto & W. Schmitz

2491

2501

2511

Seismic assessment of utility systems: Application to water, electric power and transportation
networks
C. Nuti, A. Rasulo & I. Vanzi

2519

Author index

2531

VOLUME 4
Electrical and electronic engineering
Balancing safety and availability for an electronic protection system
S. Wagner, I. Eusgeld, W. Krger & G. Guaglio
Evaluation of important reliability parameters using VHDL-RTL modelling and information
flow approach
M. Jallouli, C. Diou, F. Monteiro, A. Dandache, H. Belhadaoui, O. Malass, G. Buchheit,
J.F. Aubry & H. Medromi

2541

2549

Energy production and distribution

Application of Bayesian networks for risk assessment in electricity distribution system
maintenance management
D.E. Nordgrd & K. Sand

2561

Incorporation of ageing effects into reliability model for power transmission network
V. Matuzas & J. Augutis

2569

Mathematical simulation of energy supply disturbances

J. Augutis, R. Kriktolaitis, V. Matuziene & S. Peciulyte

2575

Risk analysis of the electric power transmission grid

L.M. Pedersen & H.H. Thorstad

2581

Security of gas supply to a gas plant from cave storage using discrete-event simulation
J.D. Amaral Netto, L.F.S. Oliveira & D. Faertes

2587

SES RISK a new framework to support decisions on energy supply

D. Serbanescu & A.L. Vetere Arellano

2593

Specification of reliability benchmarks for offshore wind farms

D. McMillan & G.W. Ault

2601

Health and medicine

Bayesian statistical meta-analysis of epidemiological data for QRA
I. Albert, E. Espi, A. Gallay, H. De Valk, E. Grenier & J.-B. Denis

XXII

2609

Cyanotoxins and health risk assessment

J. Kellner, F. Boek, J. Navrtil & J. Dvork

2613

The estimation of health effect risks based on different sampling intervals of meteorological data
J. Jeong & S. Hoon Han

2619

Information technology and telecommunications

A bi-objective model for routing and wavelength assignment in resilient WDM networks
T. Gomes, J. Craveirinha, C. Simes & J. Clmaco

2627

Formal reasoning regarding error propagation in multi-process software architectures

F. Stre & R. Winther

2635

Implementation of risk and reliability analysis techniques in ICT

R. Mock, E. Kollmann & E. Bnzli

2641

Information security measures influencing user performance

E. Albrechtsen & J.M. Hagen

2649

Reliable network server assignment using an ant colony approach

S. Kulturel-Konak & A. Konak

2657

Risk and safety as system-theoretic conceptsA formal view on system-theory

by means of petri-nets
J. Rudolf Mller & E. Schnieder

2665

Insurance and finance

Behaviouristic approaches to insurance decisions in the context of natural hazards
M. Bank & M. Gruber
Gaming tool as a method of natural disaster risk education: Educating the relationship
between risk and insurance
T. Unagami, T. Motoyoshi & J. Takai
Reliability-based risk-metric computation for energy trading
R. Mnguez, A.J. Conejo, R. Garca-Bertrand & E. Castillo

2675

2685
2689

Manufacturing
A decision model for preventing knock-on risk inside industrial plant
M. Grazia Gnoni, G. Lettera & P. Angelo Bragatto

2701

Condition based maintenance optimization under cost and profit criteria for manufacturing
equipment
A. Snchez, A. Goti & V. Rodrguez

2707

PRA-type study adapted to the multi-crystalline silicon photovoltaic cells manufacture

process
A. Colli, D. Serbanescu & B.J.M. Ale

2715

Mechanical engineering
Developing a new methodology for OHS assessment in small and medium enterprises
C. Pantanali, A. Meneghetti, C. Bianco & M. Lirussi

2727

Optimal Pre-control as a tool to monitor the reliability of a manufacturing system

S. San Matas & V. Giner-Bosch

2735

The respirable crystalline silica in the ceramic industriesSampling, exposure

and toxicology
E. Monfort, M.J. Ibez & A. Escrig

XXIII

2743

Natural hazards
A framework for the assessment of the industrial risk caused by floods
M. Campedel, G. Antonioni, V. Cozzani & G. Di Baldassarre

2749

A simple method of risk potential analysis for post-earthquake fires

J.L. Su, C.C. Wu, K.S. Fan & J.R. Chen

2757

Applying the SDMS model to manage natural disasters in Mexico

J.R. Santos-Reyes & A.N. Beard

2765

Decision making tools for natural hazard risk managementExamples from Switzerland
M. Brndl, B. Krummenacher & H.M. Merz

2773

How to motivate people to assume responsibility and act upon their own protection from flood
risk in The Netherlands if they think they are perfectly safe?
M. Bockarjova, A. van der Veen & P.A.T.M. Geurts

2781

Integral risk management of natural hazardsA system analysis of operational application

to rapid mass movements
N. Bischof, H. Romang & M. Brndl

2789

Risk based approach for a long-term solution of coastal flood defencesA Vietnam case
C. Mai Van, P.H.A.J.M. van Gelder & J.K. Vrijling

2797

River system behaviour effects on flood risk

T. Schweckendiek, A.C.W.M. Vrouwenvelder, M.C.L.M. van Mierlo, E.O.F. Calle & W.M.G. Courage

2807

Valuation of flood risk in The Netherlands: Some preliminary results

M. Bockarjova, P. Rietveld & E.T. Verhoef

2817

Nuclear engineering
An approach to integrate thermal-hydraulic and probabilistic analyses in addressing
safety margins estimation accounting for uncertainties
S. Martorell, Y. Nebot, J.F. Villanueva, S. Carlos, V. Serradell, F. Pelayo & R. Mendizbal

2827

Availability of alternative sources for heat removal in case of failure of the RHRS during
midloop conditions addressed in LPSA
J.F. Villanueva, S. Carlos, S. Martorell, V. Serradell, F. Pelayo & R. Mendizbal

2837

Complexity measures of emergency operating procedures: A comparison study with data

from a simulated computerized procedure experiment
L.Q. Yu, Z.Z. Li, X.L. Dong & S. Xu

2845

Distinction impossible!: Comparing risks between Radioactive Wastes Facilities and Nuclear
Power Stations
S. Kim & S. Cho

2851

Heat-up calculation to screen out the room cooling failure function from a PSA model
M. Hwang, C. Yoon & J.-E. Yang
Investigating the material limits on social construction: Practical reasoning about nuclear
fusion and other technologies
T. Horlick-Jones, A. Prades, C. Oltra, J. Navajas & J. Espluga

2861

2867

Neural networks and order statistics for quantifying nuclear power plants safety margins
E. Zio, F. Di Maio, S. Martorell & Y. Nebot

2873

Probabilistic safety assessment for other modes than power operation

M. Cepin
& R. Prosen

2883

Probabilistic safety margins: Definition and calculation

R. Mendizbal

2891

XXIV

Reliability assessment of the thermal hydraulic phenomena related to a CAREM-like

passive RHR System
G. Lorenzo, P. Zanocco, M. Gimnez, M. Marqus, B. Iooss, R. Bolado Lavn, F. Pierro,
G. Galassi, F. DAuria & L. Burgazzi
Some insights from the observation of nuclear power plant operators management of simulated
abnormal situations
M.C. Kim & J. Park
Vital area identification using fire PRA and RI-ISI results in UCN 4 nuclear power plant
K.Y. Kim, Y. Choi & W.S. Jung

2899

2909
2913

Offshore oil and gas

A new approach for follow-up of safety instrumented systems in the oil and gas industry
S. Hauge & M.A. Lundteigen
Consequence based methodology to determine acceptable leakage rate through closed safety
critical valves
W. Red, K. Haver, H.S. Wiencke & T.E. Nkland

2921

2929

FAMUS: Applying a new tool for integrating flow assurance and RAM analysis
. Grande, S. Eisinger & S.L. Isaksen

2937

Fuzzy reliability analysis of corroded oil and gas pipes

M. Singh & T. Markeset

2945

Life cycle cost analysis in design of oil and gas production facilities to be used in harsh,
remote and sensitive environments
D. Kayrbekova & T. Markeset
Line pack management for improved regularity in pipeline gas transportation networks
L. Frimannslund & D. Haugland

2955
2963

Optimization of proof test policies for safety instrumented systems using multi-objective
genetic algorithms
A.C. Torres-Echeverria, S. Martorell & H.A. Thompson

2971

Paperwork, management, and safety: Towards a bureaucratization of working life

and a lack of hands-on supervision
G.M. Lamvik, P.C. Nsje, K. Skarholt & H. Torvatn

2981

Preliminary probabilistic study for risk management associated to casing long-term integrity
in the context of CO2 geological sequestrationRecommendations for cement plug geometry
Y. Le Guen, O. Poupard, J.-B. Giraud & M. Loizzo

2987

Risk images in integrated operations

C.K. Tveiten & P.M. Schiefloe

2997

Policy decisions
Dealing with nanotechnology: Do the boundaries matter?
S. Brunet, P. Delvenne, C. Fallon & P. Gillon

3007

Factors influencing the public acceptability of the LILW repository

N. eleznik, M. Polic & D. Kos

3015

Risk futures in Europe: Perspectives for future research and governance. Insights from a EU
funded project
S. Menoni
Risk management strategies under climatic uncertainties
U.S. Brandt

XXV

3023
3031

Safety representative and managers: Partners in health and safety?

T. Kvernberg Andersen, H. Torvatn & U. Forseth

3039

Stop in the name of safetyThe right of the safety representative to halt dangerous work
U. Forseth, H. Torvatn & T. Kvernberg Andersen

3047

The VDI guideline on requirements for the qualification of reliability engineersCurriculum

and certification process
U.K. Rakowsky

3055

Public planning
Analysing analysesAn approach to combining several risk and vulnerability analyses
J. Borell & K. Eriksson
Land use planning methodology used in Walloon region (Belgium) for tank farms of gasoline
and diesel oil
F. Tambour, N. Cornil, C. Delvosalle, C. Fivez, L. Servranckx, B. Yannart & F. Benjelloun

3061

3067

Security and protection

Protection from half-criminal windows breakers to mass murderers with nuclear weapons:
Changes in the Norwegian authorities discourses on the terrorism threat
S.H. Jore & O. Nj

3077

A preliminary analysis of volcanic Na-Tech risks in the Vesuvius area

E. Salzano & A. Basco

3085

Are safety and security in industrial systems antagonistic or complementary issues?

G. Deleuze, E. Chatelet, P. Laclemence, J. Piwowar & B. Affeltranger

3093

Assesment of energy supply security indicators for Lithuania

J. Augutis, R. Kriktolaitis, V. Matuziene & S. Peciulyte

3101

Enforcing application securityFixing vulnerabilities with aspect oriented programming

J. Wang & J. Bigham

3109

Governmental risk communication: Communication guidelines in the context of terrorism

as a new risk
I. Stevens & G. Verleye

3117

On combination of Safety Integrity Levels (SILs) according to IEC61508 merging rules

Y. Langeron, A. Barros, A. Grall & C. Brenguer

3125

On the methods to model and analyze attack scenarios with Fault Trees
G. Renda, S. Contini & G.G.M. Cojazzi

3135

Risk management for terrorist actions using geoevents

G. Maschio, M.F. Milazzo, G. Ancione & R. Lisi

3143

Surface transportation (road and train)

A modelling approach to assess the effectiveness of BLEVE prevention measures on LPG tanks
G. Landucci, M. Molag, J. Reinders & V. Cozzani

3153

Availability assessment of ALSTOMs safety-relevant trainborne odometry sub-system

B.B. Stamenkovic & P. Dersin

3163

Dynamic maintenance policies for civil infrastructure to minimize cost and manage safety risk
T.G. Yeung & B. Castanier

3171

FAI: Model of business intelligence for projects in metrorailway system

A. Oliveira & J.R. Almeida Jr.

3177

XXVI

Impact of preventive grinding on maintenance costs and determination of an optimal grinding cycle
C. Meier-Hirmer & Ph. Pouligny

3183

Logistics of dangerous goods: A GLOBAL risk assessment approach

C. Mazri, C. Deust, B. Nedelec, C. Bouissou, J.C. Lecoze & B. Debray

3191

Optimal design of control systems using a dependability criteria and temporal sequences
evaluationApplication to a railroad transportation system
J. Clarhaut, S. Hayat, B. Conrard & V. Cocquempot

3199

RAM assurance programme carried out by the Swiss Federal Railways SA-NBS project
B.B. Stamenkovic

3209

RAMS specification for an urban transit Maglev system

A. Raffetti, B. Faragona, E. Carfagna & F. Vaccaro

3217

Safety analysis methodology application into two industrial cases: A new mechatronical system
and during the life cycle of a CAFs high speed train
O. Revilla, A. Arnaiz, L. Susperregui & U. Zubeldia

3223

The ageing of signalling equipment and the impact on maintenance strategies

M. Antoni, N. Zilber, F. Lejette & C. Meier-Hirmer

3231

The development of semi-Markov transportation model

Z. Mateusz & B. Tymoteusz

3237

Valuation of operational architecture dependability using Safe-SADT formalism: Application

to a railway braking system
D. Renaux, L. Cauffriez, M. Bayart & V. Benard

3245

Waterborne transportation
A simulation based risk analysis study of maritime traffic in the Strait of Istanbul
B. zbas, I. Or, T. Altiok & O.S. Ulusu

3257

Analysis of maritime accident data with BBN models

P. Anto, C. Guedes Soares, O. Grande & P. Trucco

3265

Collision risk analyses of waterborne transportation

E. Vanem, R. Skjong & U. Langbecker

3275

Complex model of navigational accident probability assessment based on real time

simulation and manoeuvring cycle concept
L. Gucma

3285

Design of the ship power plant with regard to the operator safety
A. Podsiadlo & W. Tarelko

3289

Human fatigue model at maritime transport

L. Smolarek & J. Soliwoda

3295

Modeling of hazards, consequences and risk for safety assessment of ships in damaged
conditions in operation
M. Gerigk

3303

Numerical and experimental study of a reliability measure for dynamic control of floating vessels
B.J. Leira, P.I.B. Berntsen & O.M. Aamo

3311

Reliability of overtaking maneuvers between ships in restricted area

P. Lizakowski

3319

Risk analysis of ports and harborsApplication of reliability engineering techniques

B.B. Dutta & A.R. Kar

3323

XXVII

Subjective propulsion risk of a seagoing ship estimation

A. Brandowski, W. Frackowiak, H. Nguyen & A. Podsiadlo

3331

The analysis of SAR action effectiveness parameters with respect to drifting search area model
Z. Smalko & Z. Burciu

3337

The risk analysis of harbour operations

T. Abramowicz-Gerigk

3343

Author index

3351

XXVIII

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Preface

This Conference stems from a European initiative merging the ESRA (European Safety and Reliability
Association) and SRA-Europe (Society for Risk AnalysisEurope) annual conferences into the major safety,
reliability and risk analysis conference in Europe during 2008. This is the second joint ESREL (European Safety
and Reliability) and SRA-Europe Conference after the 2000 event held in Edinburg, Scotland.
ESREL is an annual conference series promoted by the European Safety and Reliability Association. The
conference dates back to 1989, but was not referred to as an ESREL conference before 1992. The Conference
has become well established in the international community, attracting a good mix of academics and industry
participants that present and discuss subjects of interest and application across various industries in the fields of
Safety and Reliability.
The Society for Risk AnalysisEurope (SRA-E) was founded in 1987, as a section of SRA international
founded in 1981, to develop a special focus on risk related issues in Europe. SRA-E aims to bring together
individuals and organisations with an academic interest in risk assessment, risk management and risk communication in Europe and emphasises the European dimension in the promotion of interdisciplinary approaches of
risk analysis in science. The annual conferences take place in various countries in Europe in order to enhance the
access to SRA-E for both members and other interested parties. Recent conferences have been held in Stockholm,
Paris, Rotterdam, Lisbon, Berlin, Como, Ljubljana and the Hague.
These conferences come for the first time to Spain and the venue is Valencia, situated in the East coast close
to the Mediterranean Sea, which represents a meeting point of many cultures. The host of the conference is the
Universidad Politcnica de Valencia.
This year the theme of the Conference is "Safety, Reliability and Risk Analysis. Theory, Methods and
Applications". The Conference covers a number of topics within safety, reliability and risk, and provides a
forum for presentation and discussion of scientific papers covering theory, methods and applications to a wide
range of sectors and problem areas. Special focus has been placed on strengthening the bonds between the safety,
reliability and risk analysis communities with an aim at learning from the past building the future.
The Conferences have been growing with time and this year the program of the Joint Conference includes 416
papers from prestigious authors coming from all over the world. Originally, about 890 abstracts were submitted.
After the review by the Technical Programme Committee of the full papers, 416 have been selected and included
in these Proceedings. The effort of authors and the peers guarantee the quality of the work. The initiative and
planning carried out by Technical Area Coordinators have resulted in a number of interesting sessions covering
a broad spectre of topics.
Sebastin Martorell
C. Guedes Soares
Julie Barnett
Editors

XXIX

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Organization

Conference Chairman
Dr. Sebastin Martorell Alsina

Universidad Politcnica de Valencia, Spain

Conference Co-Chairman
Dr. Bls Galvn Gonzlez

University of Las Palmas de Gran Canaria, Spain

Conference Technical Chairs

Prof. Carlos Guedes Soares
Dr. Julie Barnett

Technical University of LisbonIST, Portugal

University of Surrey, United Kingdom

Board of Institution Representatives

Prof. Gumersindo Verd
Vice-Rector for International Actions
Universidad Politcnica de Valencia, Spain
Dr. Ioanis Papazoglou
ESRA Chairman
Dr. Roberto Bubbico
SRA-Europe Chairman
Technical Area Coordinators
Aven, TerjeNorway
Bedford, TimUnited Kingdom
Berenguer, ChristopheFrance
Bubbico, RobertoItaly
Cepin, MarcoSlovenia
Christou, MichalisItaly
Colombo, SimoneItaly
Dien, YvesFrance
Domnech, EvaSpain
Eisinger, SiegfriedNorway
Enander, AnnSweden
Felici, MassimoUnited Kingdom
Finkelstein, MaximSouth Africa
Goossens, LouisThe Netherlands
Hessami, AliUnited Kingdom
Johnson, ChrisUnited Kingdom
Kirchsteiger, ChristianLuxembourg
Technical Programme Committee
Ale B, The Netherlands
Alemano A, Luxembourg
Amari S, United States
Andersen H, Denmark
Aneziris O, Greece
Antao P, Portugal
Arnaiz A, Spain

Leira, BertNorway
Levitin, GregoryIsrael
Merad, MyriamFrance
Palanque, PhilippeFrance
Papazoglou, IoannisGreece
Preyssl, ChristianThe Netherlands
Rackwitz, RuedigerGermany
Rosqvist, TonyFinland
Salvi, OlivierGermany
Skjong, RolfNorway
Spadoni, GigliolaItaly
Tarantola, StefanoItaly
Thalmann, AndreaGermany
Thunem, Atoosa P-JNorway
Van Gelder, PieterThe Netherlands
Vrouwenvelder, TonThe Netherlands
Wolfgang, KrgerSwitzerland

Badia G, Spain
Barros A, France
Bartlett L, United Kingdom
Basnyat S, France
Birkeland G, Norway
Bladh K, Sweden
Boehm G, Norway

XXXI

Bris R, Czech Republic

Brndl M, Switzerland
Burgherr P, Switzerland
Bye R, Norway
Carlos S, Spain
Castanier B, France
Castillo E, Spain
Cojazzi G, Italy
Contini S, Italy
Cozzani V, Italy
Cha J, Korea
Chozos N, United Kingdom
De Wit S, The Netherlands
Droguett E, Brazil
Drottz-Sjoberg B, Norway
Dutuit Y, France
Escriche I, Spain
Faber M, Switzerland
Fouladirad M, France
Garbatov Y, Portugal
Ginestar D, Spain
Grall A, France
Gucma L, Poland
Hardman G, United Kingdom
Harvey J, United Kingdom
Hokstad P, Norway
Holicky M, Czech Republic
Holloway M, United States
Iooss B, France
Iung B, France
Jonkman B, The Netherlands
Kafka P, Germany
Kahle W, Germany
Kleyner A, United States
Kolowrocki K, Poland
Konak A, United States
Korczak E, Poland
Kortner H, Norway
Kosmowski K, Poland
Kozine I, Denmark
Kulturel-Konak S, United States
Kurowicka D, The Netherlands
Labeau P, Belgium

Webpage Administration
Alexandre Janeiro

Le Bot P, France
Limbourg P, Germany
Lisnianski A, Israel
Lucas D, United Kingdom
Luxhoj J, United States
Ma T, United Kingdom
Makin A, Australia
Massaiu S, Norway
Mercier S, France
Navarre D, France
Navarro J, Spain
Nelson W, United States
Newby M, United Kingdom
Nikulin M, France
Nivolianitou Z, Greece
Prez-Ocn R, Spain
Pesme H, France
Piero B, Italy
Pierson J, France
Podofillini L, Italy
Proske D, Austria
Re A, Italy
Revie M, United Kingdom
Rocco C, Venezuela
Rouhiainen V, Finland
Roussignol M, France
Sadovsky Z, Slovakia
Salzano E, Italy
Sanchez A, Spain
Sanchez-Arcilla A, Spain
Scarf P, United Kingdom
Siegrist M, Switzerland
Srensen J, Denmark
Storer T, United Kingdom
Sudret B, France
Teixeira A, Portugal
Tian Z, Canada
Tint P, Estonia
Trbojevic V, United Kingdom
Valis D, Czech Republic
Vaurio J, Finland
Yeh W, Taiwan
Zaitseva E, Slovakia
Zio E, Italy

Instituto Superior Tcnico, Portugal

Local Organizing Committee

Sofa Carlos Alberola
Eva Ma Domnech Antich
Antonio Jos Fernandez
Bls Galvn Gonzlez
Aitor Goti Elordi
Sebastin Martorell Alsina
Rubn Mullor Ibaez

Universidad Politcnica de Valencia

Universidad Politcnica de Valencia
Iberinco, Chairman Reliability Committee AEC
Universidad de Las Palmas de Gran Canaria
Universidad de Mondragn
Universidad Politcnica de Valencia
Universidad de Alicante

XXXII

Rafael Prez Ocn

Ana Isabel Snchez Galdn
Vicente Serradell Garca
Gabriel Winter Althaus

Universidad de Granada
Universidad Politcnica de Valencia
Universidad Politcnica de Valencia
Universidad de Las Palmas de Gran Canaria

Conference Secretariat and Technical Support at Universidad Politcnica de Valencia

Gemma Cabrelles Lpez
Teresa Casquero Garca
Luisa Cerezuela Bravo
Fanny Collado Lpez
Mara Luca Ferreres Alba
Angeles Garzn Salas
Mara De Rus Fuentes Manzanero
Beatriz Gmez Martnez
Jos Luis Pitarch Catal
Ester Srougi Ramn
Isabel Martn Lluch
Alfredo Moreno Manteca
Maryory Villamizar Leon
Jos Felipe Villanueva Lpez
Sponsored by
Ajuntament de Valencia
Asociacin Espaola para la Calidad (Comit de Fiabilidad)
CEANI
Generalitat Valenciana
Iberdrola
Ministerio de Educacin y Ciencia
PMM Institute for Learning
Tekniker
Universidad de Las Palmas de Gran Canaria
Universidad Politcnica de Valencia

XXXIII

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Acknowledgements

The conference is organized jointly by Universidad Politcnica de Valencia, ESRA (European Safety and
Reliability Association) and SRA-Europe (Society for Risk AnalysisEurope), under the high patronage of
the Ministerio de Educacin y Ciencia, Generalitat Valenciana and Ajuntament de Valencia.
Thanks also to the support of our sponsors Iberdrola, PMM Institute for Learning, Tekniker, Asociacin
Espaola para la Calidad (Comit de Fiabilidad), CEANI and Universidad de Las Palmas de Gran Canaria. The
support of all is greatly appreciated.
The work and effort of the peers involved in the Technical Program Committee in helping the authors to
improve their papers are greatly appreciated. Special thanks go to the Technical Area Coordinators and organisers
of the Special Sessions of the Conference, for their initiative and planning which have resulted in a number of
interesting sessions. Thanks to authors as well as reviewers for their contributions in the review process. The
review process has been conducted electronically through the Conference web page. The support to the web
page was provided by the Instituto Superior Tcnico.
We would like to acknowledge specially the local organising committee and the conference secretariat and technical support at the Universidad Politcnica de Valencia for their careful planning of the practical arrangements.
Their many hours of work are greatly appreciated.
These conference proceedings have been partially financed by the Ministerio de Educacin y Ciencia
de Espaa (DPI2007-29009-E), the Generalitat Valenciana (AORG/2007/091 and AORG/2008/135) and the
Universidad Politcnica de Valencia (PAID-03-07-2499).

XXXV

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Introduction

The Conference covers a number of topics within safety, reliability and risk, and provides a forum for presentation
and discussion of scientific papers covering theory, methods and applications to a wide range of sectors and
problem areas.
Thematic Areas

Accident and Incident Investigation

Crisis and Emergency Management
Decision Support Systems and Software Tools for Safety and Reliability
Dynamic Reliability
Fault Identification and Diagnostics
Human Factors
Integrated Risk Management and Risk-Informed Decision-making
Legislative dimensions of risk management
Maintenance Modelling and Optimisation
Monte Carlo Methods in System Safety and Reliability
Occupational Safety
Organizational Learning
Reliability and Safety Data Collection and Analysis
Risk and Evidence Based Policy Making
Risk and Hazard Analysis
Risk Control in Complex Environments
Risk Perception and Communication
Safety Culture
Safety Management Systems
Software Reliability
Stakeholder and public involvement in risk governance
Structural Reliability and Design Codes
System Reliability Analysis
Uncertainty and Sensitivity Analysis

Industrial and Service Sectors

Aeronautics and Aerospace

Automotive Engineering
Biotechnology and Food Industry
Chemical Process Industry
Civil Engineering
Critical Infrastructures
Electrical and Electronic Engineering
Energy Production and Distribution
Health and Medicine
Information Technology and Telecommunications
Insurance and Finance
Manufacturing
Mechanical Engineering
Natural Hazards

XXXVII

Nuclear Engineering
Offshore Oil and Gas
Policy Decisions
Public Planning
Security and Protection
Surface Transportation (road and train)
Waterborne Transportation

XXXVIII

Thematic areas
Accident and incident investigation

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A code for the simulation of human failure events in nuclear

power plants: SIMPROC
J. Gil, J. Espern, L. Gamo, I. Fernndez, P. Gonzlez & J. Moreno
Indizen Technologies S.L., Madrid, Spain

A. Expsito, C. Queral & G. Rodrguez

Universidad Politnica de Madrid (UPM), Madrid, Spain

J. Hortal
Spanish Nuclear Safety Council (CSN), Madrid, Spain

ABSTRACT: Over the past years, many Nuclear Power Plant (NPP) organizations have performed Probabilistic
Safety Assessments (PSAs) to identify and understand key plant vulnerabilities. As part of enhancing the PSA
quality, the Human Reliability Analysis (HRA) is key to a realistic evaluation of safety and of the potential
weaknesses of a facility. Moreover, it has to be noted that HRA continues to be a large source of uncertainly in
the PSAs. We developed SIMulator of PROCedures (SIMPROC) as a tool to simulate events related with human
actions and to help the analyst to quantify the importance of human actions in the final plant state. Among others,
the main goal of SIMPROC is to check if Emergency Operating Procedures (EOPs) lead to safe shutdown plant
state. First pilot cases simulated have been MBLOCA scenarios simulated by MAAP4 severe accident code
coupled with SIMPROC.

Also, this software tool is part of SCAIS software

package, (Izquierdo 2003), (Izquierdo et al. 2000) and
(Izquierdo et al. 2008), which is a simulation system
able to generate dynamic event trees stemming from an
initiating event, based on a technique able to efficiently
simulate all branches and taking into account different
factors which may affect the dynamic plant behavior
in each sequence. The development of SIMPROC is
described in detail in this paper.

INTRODUCTION

In addition to the traditional methods for verifying procedures, integrated simulations of operator and plant
response may be useful to:
1. Verify that the plant operating procedures can be
understood and performed by the operators.
2. Verify that the response based on these procedures
leads to the intended results.
3. Identify potential situations where judgment of
operators concerning the appropriate response is
inconsistent with the procedures.
4. Study the consequences of errors of commission
and the possibilities for recovering from such
errors, (CSNI 1998) and (CSNI-PWG1 and CSNIPWG5 1997).
5. Study time availability factors related with procedures execution.

WHY SIMPROC?

The ascribing of a large number of accidents to human

error meant that there was a need to consider it in
risk assessment. In the context of nuclear safety studies (Rasmussen 1975) concluded that human error
contributed to 65% of the considered accidents. This
contribution is as great or even greater than the contribution associated to system failures. The NRC estimates human error is directly or indirectly involved in
approximately 65% of the abnormal incidents (Trager
1985). According to the data of the Incident Reporting
System of the International Atomic Energy Agency
(IAEA), 48% of the registered events are related
to failures in human performances. Out of those,

Indizen Technologies, in cooperation with the

Department of Energy Systems of Technical University of Madrid and the Spanish Nuclear Safety Council
(CSN), has developed during the last two years a tool
known as SIMPROC. This tool, coupled with a plant
simulator, is able to incorporate the effect of operator actions in plant accident sequences simulations.

63% were into power operation and the remaining 37%

into shutdown operation. Additionally, analyzing the
events reported using the International Nuclear Event
Scale (INES) during the last decade, most of the major
incidents of Level 2 or higher could be attributed to
causes related to human performance. Moreover, a
study based on a broad set of PSA data states that
between 15 and 80% of Core Damage Frequency is
related with execution failure of some operator action
(NEA 2004). Reason (Reason 1990), in a study of a
dozen meaningful accidents in the 15 years prior to
their publication, including Three Mile Island (TMI),
Chernobyl, Bhopal and the fire of London underground, concludes that at least 80% of the system
failures were caused by humans, specially by inadequate management of maintenance supervision. In
addition, it determines that other aspects have relevance, mainly technical inaccuracy or incomplete
training in operation procedures. This last aspect is of
great importance in multiple sectors whose optimization of operation in abnormal or emergency situations
is based on strict operation procedures. In this case,
it stands out that four of the investigations carried out
by the NRC and nearly 20 of the additional investigations from Three Miles Island (TMI) concluded that
intolerable violations of procedures took place. The
TMI accident illustrated clearly how the interaction
of technical aspects with human and organizational
factors can help the progression of events. After the
incident, big efforts on investigation and development have focused on the study of human factors in
accidents management. The management of accidents
includes the actions that the operation group must perform during beyond design basis accidents, with the
objective of maintaining the basic functions of reactor safety. Two phases in emergencies management
are distinguished: The preventive phase, in which performances of the operator are centered in avoiding
damage to the core and maintaining the integrity of the
installation, and the phase of mitigation, in which once
core damage occurs, operator actions are oriented to
reduce the amount of radioactive material that is going
to be released. Management of accidents is carried out
by following EOPs and Severe Accident Management
Guides (SAMGs), and its improvement was one of
the activities carried out after TMI accident. The accident sequence provides the basis for determining the
frequencies and uncertainties of consequences. The
essential outcome of a PSA is a quantitative expression of the overall risks in probabilistic terms. The
initial approach to import human factor concerns into
engineering practices was to use existing PSA methods and extend them to include human actions. We will
use SIMPROC to extend this functionality to include
the human factors into the plant evolution simulation.
This is done in a dynamic way instead of the static
point of view carried out by PSA studies.

BABIECA-SIMPROC ARCHITECTURE

The final objective of the BABIECA-SIMPROC system is to simulate accidental transients in NPPs considering human actions. For this purpose is necessary
to develop an integrated tool that simulates the dynamics of the system. To achieve this we will use the
BABIECA Simulation Engine to calculate the time
evolution state of the plant. Finally we will model
the influence of human operator actions by means of
SIMPROC. We have modeled the operators influence
over the plant state as a separate module to emphasize the significance of operator actions in the final
state of the plant. It is possible to plug or unplug
SIMPROC to consider the operator influence over the
simulation state of the plant in order to compare end
states in both cases. The final goal of the BABIECASIMPROC overall system, integrated in SCAIS, is to
simulate Dynamic Event Trees (DET) to describe the
time evolution scheme of accidental sequences generated from a trigger event. During this calculation it
must be taken into account potential degradations of
the systems associating them with probabilistic calculations in each sequence. Additionally EOPs execution
influence is defined for each plant and each sequence.
In order to achieve this objective the integrated scheme
must fit the following features:
1. The calculation framework must be able to integrate
other Simulation Codes (MAAP, TRACE, . . . ). In
this case, BABIECA-SIMPROC acts as a wrapper
to external codes. This will allow to work with different codes in the same time line sequence. In case
the simulation reaches core damage conditions, it is
possible to unplug the best estimate code and plug
a severe accident code to accurately describe the
dynamic state of the plant.
2. Be able to automatically generate the DET associated to an event initiator, simulating the dynamic
plant evolution.
3. Obtain the probability associated to every possible
evolution sequence of the plant.
All the system is being developing in C++ code
in order to meet the requirements of speed and
performance needed in this kind of simulations. Parallelization was implemented by means of a PVM architecture. The communication with the PostGresSQL
database is carried out by the libpq++ library. All the
input desk needed to initialize the system is done using
standard XML.
The main components of the Global System Architecture can be summarized as follows:
1. DENDROS event scheduler. It is in charge of
opening branches of the simulation tree depending
on the plant simulation state. DENDROS allows
the modularization and parallelization of the tree

2.

3.
4.

5.

generation. Calculation of the probability for each

branch is based on the true condition of certain
logical functions. The scheduler arranges for the
opening of the branch whenever certain conditions
are met, and stops the simulation of any particular branch that has reached an absorbing state. The
time when the opening of the new branch occurs can
be deterministically fixed by the dynamic conditions (setpoint crossing) or randomly delayed with
respect to the time when the branching conditions
are reached. The latter option especially applies
to operator actions and may include the capability to use several values of the delay time within
the same dynamic event tree. The scheduler must
know the probability of each branch, calculated
in a separate process called Probability Wrapper,
in order to decide which branch is suitable for
further development. The applications of a tree
structured computation extend beyond the scope
of the DETs. In fact, the branch opening and cutoff
can obey any set of criteria not necessarily given by
a probability calculation as, for instance, sensitivity studies or automatic initialization for Accident
Management Strategy analyses. More details on
how dynamic event trees are generated and handled
and their advantages for safety analysis applications are given in another paper presented in this
conference (Izquierdo et al. 2008).
BABIECA plant simulator. It is adapted to execute the sequence simulation launched by Dendros Event Scheduler. As mentioned previously,
BABIECA can wrap other nuclear simulation codes
(i.e., MAAP). This simulation code is able to
extend the simulation capacities of BABIECA to
the context of severe accidents.
SIMPROC procedures simulator. This simulation
module allows us to interact with the Simulation
Engine BABIECA to implement the EOPs.
Probability engine. It calculates the probabilities
and delays associated with the set points of the
DET. The initial implementation will be based in
PSA calculations, but we have developed a probability wrapper able to use calculations from BDDs
structures in the future.
Global database. It will be used to save data from
the different simulation modules, providing restart
capability to the whole system and allowing an
easier handling of the simulation results.

Figure 1.

BABIECA-SIMPROC architecture.

modules allow us to represent relevant plant systems

in great detail. In this publication we will focus our
attention on the MAAP4 Wrapper, which allows us to
connect BABIECA with this severe accident code. The
SIMPROC interaction over the system is illustrated in
2. The process can be summarized in the following
steps:
1. BABIECA starts the simulation and SIMPROC
is created when a EOP execution is demanded
according to a set of conditions over plant variables
previously defined.
2. SIMPROC is initialized with the plant state variables at that time instant. As a previous step the
computerized XML version of the set of EOPs must
be introduced in the SIMPROC database.
3. The BABIECA calculation loop starts and the outcome of EOPs executions are modeled as boundary conditions over the system. Each topology
block can modify its state according to the specific actions of SIMPROC EOPs execution. Once
boundary conditions are defined for the current
step, the solution for the next step is calculated
for each topology block. The calculation sequence
includes continuous variables, discrete variables
and events recollection for the current time step.
Finally, all variables information are saved in the
database and, depending on the system configuration, a simulation restart point is set.
4. The procedures simulator SIMPROC does not have
its own time step but adapts its execution to the
BABIECA pace. Additionally, it is possible to set
a default communication time between BABIECA
and SIMPROC. This time represents the average
time the operator needs to recognize the state of
the plant, which is higher than the time step of the
simulation.

If we focus our attention on the SIMPROC integration of the system, the BABIECA-SIMPROC
architecture can be illustrated accordingly (Fig. 1).
BABIECA acts as a master code to encapsulate
different simulation codes in order to build a robust
system with a broad range of application and great
flexibility. BABIECA Driver has its own topology,
named BABIECA Internal Modules in Fig. 1. These

Figure 3.
Figure 2.

BABIECA-SIMPROC calculation flow.

feature allows for a better modelling of the distribution of the operation duties among the members
of the operation team.

5. Once BABIECA reaches the defined simulation

end time, it sends a message to SIMPROC to quit
procedure execution (if it has not already ended)
and saves the overall plant state in the database.

SIMPROC-BABIECA-MAAP4 connection.

As shown in Fig. 3, BABIECA implements a block

topology to represent control systems that model operator actions over the plant. SIMPROC has access to
those blocks and can change its variables according
to the directives associated with the active EOPs. The
block outputs will be the time dependent boundary
conditions for MAAP4 calculation during the current time step. It can also be seen in Fig. 3 that
BABIECA has another type of blocks: SndCode and
RcvCode. These blocks were designed to communicate BABIECA with external codes through a PVM
interface. SndCode gathers all the information generated by control systems simulation and sends it
to MAAP4 wrapper, which conforms all the data
to be MAAP4 readable. Then, MAAP4 will calculate the next simulation time step. When this calculation ends, MAAP4 sends the calculated outputs
again to MAAP4 wrapper and then the outputs reach
the RcvCode block of the topology. At this point,
any BABIECA component, especially control systems, has the MAAP4 calculated variables available.
A new BABIECA-SIMPROC synchronization point
occurs and the BABIECA variable values will be the
boundary conditions for SIMPROC execution.

SIMPROC-BABIECA-MAAP4 CONNECTION
SCHEME

Providing system connectivity between BABIECA,

SIMPROC and MAAP4 is one of the primary targets.
This allows to simulate severe accident sequences in
NPPs integrating operator actions simulated by SIMPROC into the plant model. In order to simulate the
operator actions over the plant with MAAP, the user
has to define these actions in the XML input file
used as a plant model, considering logical conditions
(SCRAM, SI, . . .) to trigger EOPs execution. MAAP4
can simulate operator functions, although in a very
rudimentary way. The key advantages of using the
integrated simulator SIMPROC-BABIECA-MAAP4
as opposed of using MAAP4 alone are:
1. The actions take a certain time while being executed, whereas in MAAP4 they are instantaneous.
This delay can be modelled by taking into account
different parameters. For example, the number of
simultaneous actions the operator is executing must
influence his ability to start new required actions.
2. Actions must follow a sequence established by the
active EOPs. In MAAP, operator actions have no
order. They are executed as logical signals are
triggered during plant evolution.
3. Although SIMPROC does not contain an operator
model, it allows for the use of pre-defined operator skills, such as reactor operator or turbine
operator, each one with specific attributes. This

APPLICATION EXAMPLE

The example used to validate the new simulation package simulates the operator actions related with the level
control of steam generators during MBLOCA transient in a PWR Westinghouse design. These operator
actions are included in several EOPs, like ES-1.2 procedure, Post LOCA cooldown and depressurization,
which is associated with primary cooling depressurization. The first step in this validation is to run

the mentioned transient using MAAP4 alone. After

that, the same simulation was run using BABIECASIMPROC system. Finally, the results from both
simulations are compared.
5.1

SG water level control with MAAP

Level control of steam generator is accomplished by

either of two modes of operation, automatic and manual. In the automatic mode, the main feed water control
system senses the steam flow rate, Ws , and the downcomer water level, z, and adjusts the feed water control
valves to bring the water level to a program (desired)
water level, z0 . If the water level is very low, z < 0.9z0 ,
the control valve is assumed to be fully open, that is,
W = Wmax

Figure 4. Representation of the key parameters to implement the narrow-band control over SGWL.

The first step is to define the topology file for

BABIECA to work. In this file we need to set the block
structure of the systems we want to simulate and we
need to specify which blocks are going to be used by
SIMPROC to model Operator actions over the plant
simulation. The EOP has a unique code to identify it
in the database, and some tags to include a description
of the actions we are going to execute. Moreover it has
a load-delay parameter designed to take into account
the time the operator needs since the EOP demand
trigger starts until the operator is ready to execute the
proper actions.
This simple EOP has only one step designed to
control the Steam Generator Water Level. The main
parameters of this step are:

(1)

where W is the feed water flow rate.

If the water level is above 0.9z0 , the model applied
a limited proportional control in two steps:
1. Proportional control. The resulting feed water flow
rate, W, returns the water level to z0 at a rate
proportional to the mismatch,
(W Ws ) = (z z0 ).

(2)

The coefficient of proportionality, in eq. 2

is chosen so that the steam generator inventory
becomes correct after a time interval , which is
set to 100 s at present.
2. A limited flow rate. The feed water flow rate is
limited to values between 0 (valve closed) and Wmax
(valve fully opened).

Skill. Sets the operator profile we want to execute

the action. In this example there is a REACTOR
profile.
Texec. Defines the time the operator is going to be
busy each time he executes an action.
Taskload. Defines a percentage to take account of
the attention that the operator needs to execute an
action properly. The sum of all the taskload parameters of the actions the operator is executing must be
less than 100%. In future works, Texec and Taskload
parameters will be obtained from the Probability
Engine in executing time according probability distributions of human actuation times for the different
actions required by EOPs. These time distributions
could be obtained from experimental studies.
Time window. Sets the temporal interval during
which the MONITOR is active.
Targets. Sentences that define the logical behavior
of the MONITOR. We tell the MONITOR what it
has to do, when and where.

The other control mode is manual. In this mode,

the control tries to hold the water level within an envelope defined by a user-supplied deadband zDEAD . The
feedwater flow rate is set as follows:

Wmax , if z < z0 zDEAD

0, if z > z0 + zDEAD 2

Wmin , if z0 + zDEAD
<z
2
W =
(3)
+
z
<
z

0
DEAD

DEAD

W , if z0 z 2 < z

< z0 + DEAD
2
where Wmin is the flow rate used on the decreasing part
of the cycle.
Operation in manual mode results in a sawtoothlike level trajectory which oscillates about the desired
level z0 .
The parameters used to implement the narrowband control over the steam generator water level are
illustrated in Fig. 4.
To simulate the BABIECA-SIMPROC version of
the same transient we must create the XML files
needed to define the input desk of the overall system.

In more complex applications, a great effort must

be made in computerizing the specific EOPs of each
nuclear plant under study, (Expsito and Queral
2003a) and (Expsito and Queral 2003b). This topic
is beyond the scope of this paper.

Finally, it is necessary to define the XML simulation files for BABIECA and SIMPROC. The main
difference with the previous XML files is that they do
not need to be parsed and introduced in the database
prior to the simulation execution. They are parsed
and stored in memory during runtime execution of the
simulation.
The BABIECA simulation file parameters are:
Simulation code. Must be unique in the database.
Start input. Informs about the XML BABIECA
Topology file linked with the simulation.
Simulation type. It is the type of simulation: restart,
transient or steady.
Total time. Final time of the simulation.
Delta. Time step of the master simulation.
Save output frequency. Frequency to save the outputs in the database.
Initial time. Initial time for the simulation.
Initial topology mode. Topology block can be in
multiple operation modes. During a simulation execution some triggers can lead to mode changes that
modify the calculation loop of a block.
Save restart frequency. Frequency we want to save
restart points to back up simulation evolution.
SIMPROC active. Flag that allow us to switch on
SIMPROC influence over the simulation.

Figure 5. Steam Generator Water Level comparison (black

line: MAAP4 output; dashed red line: BABIECA-SIMPROC
output).

The main parameters described in the XML SIMPROC simulation file are:
Initial and end time. These parameters can be different to the ones used for the simulation file and
define a time interval for SIMPROC to work.
Operator parameters. These are id, skill and slowness. The first two identify the operator and his type
and the latter takes account of his speed to execute
the required actions. It is known that this parameter
depends on multiple factors like operator experience
and training.
Initial variables. These are the variables that are
monitored continuously to identify the main parameters to evaluate the plant state. Each variable has a
procedure code to be used in the EOP description,
a BABIECA code to identify the variable inside the
topology and a set of logical states.
Variables. These variables are not monitored in a
continuous way but have the same structure as Initial
Variables. They are only updated under SIMPROC
request.

Figure 6. Mass Flow Rate to the Cold Leg (red line: MAAP4
output; dashed black line: BABIECA-SIMPROC output).

the simulation. Then water level starts to raise going

through the defined dead band until the level reaches
12.5 m. At this point we set FWFR to its minimum
value (7.36 kgs ). When SGWL is higher than 15 m, the
flow rate is set to zero.
The concordance between MAAP4 and BABIECASIMPROC results is good in general. The differences
can be explained due to different time steps of both
codes. MAAP4 chooses its time step according to convergence criteria, while BABIECA-SIMPROC has a
fixed time step set in the XML BABIECA Simulation File. Additionally, BABIECA-SIMPROC takes
into consideration the time needed by the operator to
execute each action whereas MAAP4 implementation
of the operator automatically execute the requested
actions.

Once we have defined the XML input files, we have

met all the required conditions to run the BABIECASIMPROC simulation.
Compared simulation results of MAAP4 simulation
and BABIECA-SIMPROC simulation can be seen in
Fig. 5 and Fig. 6. As shown in the figures, feed water
flow rate is set to 9,9 kgs when SGWL is lower than
7.5 m. This situation occurs in the very first part of

CSN Spanish Nuclear Safety Council

EOP Emergency Operation Procedure
SAMG Severe Accident Management Guide
LOCA Loss of Coolant Accident
DET Dynamic Event Tree
PVM Parallel Virtual Machine
BDD Binary Decision Diagram
XML Extensible Markup Language
SGWL Steam Generator Water Level
FWFR Feed Water Flow Rate

CONCLUSIONS

The software tool BABIECA-SIMPROC is being

developed. This software package incorporates operator actions in accidental sequences simulations in
NPP. This simulation tool is not intended to evaluate the probability of human errors, but to incorporate
in the plant dynamics the effects of those actions performed by the operators while following the operating
procedures. Nonetheless, human errors probabilities
calculated by external HRA models can be taken into
account in the generation of dynamic event trees under
the control of DENDROS. We have tested this application with a pilot case related with the steam generator
water level control during a MBLOCA transient in a
PWR Westinghouse NPP. The results have been satisfactory although further testing is needed. At this stage
we are in this process of validation to simulate a complete set of EOPs that are used in a PWR Westinghouse
NPP. Moreover, we are extending the capabilities of
the system to incorporate TRACE as an external code
with its corresponding BABIECA wrapper. When this
part of the work is completed, a wider simulation will
be available. This will allow to analyze the impact of
EOPs execution by operators in the final state of the
plant as well as the evaluation of the allowable response
times for the manual actions.

REFERENCES
CSNI (Ed.) (1998). Proceedings from Specialists Meeting
Organized: Human performance in operational events,
CSNI.
CSNI-PWG1, and CSNI-PWG5 (1997). Research strategies
for human performance. Technical Report 24, CSNI.
Expsito, A. and C. Queral (2003a). Generic questions about
the computerization of the Almaraz NPP EOPs. Technical
report, DSE-13/2003, UPM.
Expsito, A. and C. Queral (2003b). PWR EOPs computerization. Technical report, DSE-14/2003, UPM.
Izquierdo, J.M. (2003). An integrated PSA approach to independent regulatory evaluations of nuclear safety assessment of Spanish nuclear power stations. In EUROSAFE
Forum 2003.
Izquierdo, J.M., J. Hortal, M. Sanchez-perea, E. Melndez,
R. Herrero, J. Gil, L. Gamo, I. Fernndez, J. Espern,
P. Gonzlez, C. Queral, A. Expsito, and G. Rodrguez
(2008). SCAIS (Simulation Code System for Integrated
Safety Assesment): Current status and applications. Proceedings of ESREL 08.
Izquierdo, J.M., C. Queral, R. Herrero, J. Hortal, M. Sanchezperea, E. Melandez, and R. Muoz (2000). Role of fast
Running TH Codes and Their Coupling with PSA Tools,
in Advanced Thermal-hydraulic and Neutronic Codes:
Current and Future Applications. In NEA/CSNI/R(2001)2,
Volume 2.
NEA (2004). Nuclear regulatory challenges related to human
performance. Isbn 92-64-02089-6, NEA.
Rasmussen, N.C. (1975). Reactor safety study, an assessment
of accident risks in u. s. nuclear power plants. In NUREG
NUREG-75/014, WASH-1400.
Reason, J. (1990). Human Error. Cambridge University
Press.
Trager, E.A. (1985). Case study report on loss of safety system function events. Technical Report AEOD/C504, ffice
for Analysis and Evaluation of Operational Data. Nuclear
Regulatory Commission (NRC).

ACKNOWLEDGMENTS
SIMPROC project is partially funded by the Spanish
Ministry of Industry (PROFIT Program) and SCAIS
project by the Spanish Ministry of Education and Science (CYCIT Program). Their support is gratefully
acknowledged.
We want to show our appreciation to the people
who in one way or another, have contributed to the
accomplishment project.
NOMENCLATURE
ISA Integrated Safety Analysis
SCAIS Simulation Codes System for Integrated Safety
Assessment
PSA Probabilistic Safety Analysis
HRA Human Reliability Analysis

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A preliminary analysis of the Tlahuac incident by applying

the MORT technique
J.R. Santos-Reyes, S. Olmos-Pea & L.M. Hernndez-Simn
Safety, Risk & Reliability Group, SEPI-ESIME, IPN, Mexico

ABSTRACT: Crime may be regarded as a major source of social concern in the modern world. Very often
increases in crime rates will be treated as headline news, and many people see the law and order issue as
one of the most pressing in modern society. An example of such issues has been highlighted by the Tlhuac
incident which occurred in Mexico City on 23 November 2004. The fatal incident occurred when an angry
crowd burnt alive two police officers and seriously injured another after mistaking them for child kidnappers.
The third policeman who was finally rescued by colleagues (three and half hours after the attack began) suffered
serious injuries. The paper presents some preliminary results of the analysis of the above incident by applying the
MORT (Management Over-sight Risk Three) technique. The MORT technique may be regarded as a structured
checklist in the form of a complex fault tree model that is intended to ensure that all aspects of an organizations
management are looked into when assessing the possible causes of an incident. Some other accident analysis
approaches may be adopted in the future for further analysis. It is hoped that by conducting such analysis lessons
can be learnt so that incidents such as the case of Tlhuac can be prevented in the future.

other hand, rape is generally defined, in the literature,

as a sexual act imposed to the victim by means of
violence or threats of violence (Griffiths 2004, Puglia
et al. 2005). It is recognised that the most common
motives for rape include power, opportunity, pervasive
anger, sexual gratification, and vindictiveness. Also,
it is emphasised that is important to understand the
behavioural characteristics of rapists, and the attitudes
towards rape victims.

INTRODUCTION

Crime and disorder may comprise a vast set of events

involving behaviour formally deemed against the law
and usually committed with evil intent (Ekblom
2005). These events range from murder to fraud, theft,
vandalism, dealing in drugs, kidnappings and terrorist
atrocities that threaten the public safety. Public safety
may be defined as a state of existence in which people, individually and collectively, are sufficiently free
from a range or real and perceived risks centering on
crime and disorder, are sufficiently able to cope with
those they nevertheless experience, and where unable
to cope unaided, are sufficiently-well protected from
the consequences of these risks (Ekblom 2005).

1.1

1.1.2 Murdering
Very often murdering is a product of conflict between
acquaintances or family members or a by-product
of other type of crime such as burglary or robbery (McCabe & Wauchope 2005, Elklit 2002). It
is generally recognised that the role of cultural attitudes, reflected in and perpetuated by the mass media
accounts, has a very significant influence in the public
perception of crime.

The nature of crime

1.1.1 Sexual and rape

It is well documented that sexual assault and abuse
have a profound effect on the victims emotional functions, such as grief, fear, self-blame, emotional liability, including cognitive reactions such as flashbacks,
intrusive thoughts, blocking of significant details of
the assault and difficulties with concentration (Alaggia
et al. 2006). The offenders on the other hand display
problems with empathy, perception about others, and
management of negative emotions, interpersonal relationships, and world views (Alalehto 2002). On the

1.1.3 Organized crime

The phenomenon of organised crime, such as smuggling or trafficking, is a kind of enterprise driven
as a business: i.e., the pursuit of profits (Tucker
2004, Klein 2005). Smuggling or trafficking is usually understood as the organised trade of weapons or
drugs, women and children, including refugees, but
unlike the legal trade, smuggling or trafficking provides goods or services, based on market demand

11

much more sophisticated. This role includes such as

developments as DNA testing and evidence, the use
of less-than-lethal weapons, increasingly sophisticated
forms of identification, and crimes such as identity
theft and a variety of frauds and scams. The use of
scientific evidence and expert witnesses is also a significant issue in the prosecution and adjudication of
offenders. Third, he puts emphasis on public security
and terrorism. This has shaped, and will continue to
shape, criminal justice in a variety of ways.
Ratcliffe (2005) emphasises on the practice of intelligence driven policing as a paradigm in modern law
enforcement in various countries (e.g., UK, USA,
Australia & New Zealand). The author proposes a
conceptual model of intelligence driven policing; the
model essentially entails three stages: (1) law enforcement interpret the criminal environment, (2) influence
decision-makers, and (3) decision-makers impact on
the criminal environment.

and the possibilities of realising it, that are illegal

or socially unacceptable. Lampe and Johansen (2004)
attempt to develop a thorough understanding of organised crime, by clarifying and specifying the concept
of thrust in organise crime networks. They propose
four typologies that categorise trust as: (a) individualised trust, (b) trust based on reputation, (c) trust
based on generalisations, and (d) abstract trust. They
show that the existence of these types of trust through
the examination of illegal markets.
1.1.4 Battering
Chiffriller et al. (2006) discuss the phenomenon
called battering, as well as batterer personality and
behavioural characteristics. Based on cluster analysis,
they developed five distinct profiles of men who batter
women. And based on the behavioural and personality
characteristics that define each cluster, they establish
five categories: (1) pathological batterers, (2) sexually violent batterers, (3) generally violent batterers,
(4) psychologically violent batterers, and (5) familyonly batterers. The authors discuss the implications for
intervention of these categories.

1.2

Crime in Mexico City

In 1989 the International Crime Victim Survey

(ICVS) was born and since then it has contributed to
the international knowledge of crime trends in several
countries. It is believed that since its conception standardized victimization surveys have been conducted in
more than 70 countries worldwide. The ICVS has been
surrounded by a growing interest by both the crime
research community and the policy makers. A part for
providing an internationally standardized indicators
for the perception and fear of crime across different
socio-economic contexts; it also has contributed to
an alternative source of data on crime. Similarly, in
Mexico, four National Crime Victim Surveys (known
as ENSI-1, 2, 3 & 4 respectively) have been conducted
since 2001. The surveys are intended to help to provide a better knowledge of the levels of crime which
affect the safety of the Mexican citizens (ICESI 2008).
Some key findings of the fourth (i.e., ENSI-4) are
summarized in Tables 1 & 2.

1.1.5 Bullying
Burgess et al. (2006) argue that bullying has become a
major public health issue, because of its connection to
violent and aggressive behaviours that result in serious injury to self and to others. The authors define
bullying as a relationship problem in which power
and aggression are inflicted on a vulnerable person
to cause distress. They further emphasise that teasing
and bullying can turn deadly.
1.1.6 Law enforcement
Finckenauer (2004) emphasizes that a major push of
the expansion of higher education in crime and justice
studies came particularly from the desire to professionalise the policewith the aim of improving police
performance. He suggests new and expanded subjectmatter coverage. First, criminal-justice educators must
recognise that the face of crime has changedit has
become increasingly international in nature. Examples
include cyber-crime, drug trafficking, human trafficking, other forms of trafficking and smuggling,
and money laundering. Although global in nature,
these sorts of crimes have significant state and local
impact. The author argues that those impacts need to
be recognised and understood by twenty-first-century
criminal-justice professors and students. Increasingly,
crime and criminals do not respect national borders.
As a result, law enforcement and criminal justice cannot be bond and limited by national borders. Second,
he emphasises the need to recognise that the role
of science in law enforcement and the administration of justice has become increasingly pervasive and

Table 1.

Victimization (ICESI 2008).

Types of crime

Percentage (%)

Robbery
56.3
Other types of robbery
25.8
Assault
7.2
Theft of items from cars (e.g., accessories) 2.9
Burglary
2.4
Theft of cars
1.5
Other types of crime
0.4
Kidnappings
2.1
Sexual offences
0.8
Other assaults/threat to citizens
0.6

12

Table 2.

18:25 hrs. One of the PFP-police officers tried to

communicate with his superiors at the Headquarters
but failed in the attempt. It is believe that the reason
was because there was a meeting going on at the time
and his superior was unable to assist the call. At about
the same time, the crowd cheered, chanted obscenities
as they attacked the officers.
18:30 hrs. A first live TV coverage is being
broadcasted.
19:30 hrs. Again, one of the PFP-police officers
attempted for the second time to communicate with
their superiors and it was broadcasted live, and he
said They are not allowing us to get out, come and
rescue us.
21:20 hrs the regional police director from the
regional police headquarters (it will be called here
as RPHQ) was acknowledged that two police officers
have been burnt alive, he ordered the back-up units to
intervene and rescue the officers.
22:30 hrs after three and a half hours the bodies of
the two PFP-police officers were recovered.

Publics perceptions of crime (ICESI 2008).

Activities that have been given

up by the public in Mexico City

Percentage (%)

The Public have been given up:

going out at night
going to the football stadium
going to dinner out
going to the cinema/Theater
carrying cash with them
taking public transport
wearing jewellery
taking taxis
carrying credit cards with them
visiting friends or relatives
other

49.1
17.1
19.8
21.3
45.4
28.1
56.0
37.0
38.0
30.5
1.6

Overall, the publics perception of crime shows that 9

out of 10 citizens feel unsafe in Mexico City. More than
half of the population believes that crime has affected
their quality of life; for example, Table 2 shows that
one in two citizens gave up wearing jewellery, going
out at night and taking cash with them.

THE TLAHUAC INCIDENT

3.1

THE MANAGEMENT OVERSIGHT RISK

TREE (MORT)
Overview of the MORT method

The Management Oversight and Risk Tree (MORT)

is an analytical procedure for determining causes and
contributing factors (NRI-1 2002).
In MORT, accidents are defined as unplanned
events that produce harm or damage, that is, losses
(NRI-1 2002). Losses occur when a harmful agent
comes into contact with a person or asset. This contact can occur either because of a failure of prevention
or, as an unfortunate but acceptable outcome of a
risk that has been properly assessed and acted-on (a
so-called assumed risk). MORT analysis always
evaluates the failure route before considering the
assumed risk hypothesis. In MORT analysis, most
of the effort is directed at identifying problems in the
control of a work/process and deficiencies in the protective barriers associated with it. These problems are
then analysed for their origins in planning, design,
policy, etc. In order to use MORT key episodes in
the sequence of events should be identified first; each
episode can be characterised as: {a} a vulnerable target
exposed to; {b} an agent of harm in the; {c} absence
of adequate barriers.
MORT analysis can be applied to any one or more
of the episodes identified; it is a choice for you to
make in the light of the circumstances particular to
your investigation. To identify these key episodes, you
will need to undertake a barrier analysis (or Energy
Trace and Barrier Analysis to give it its full title).
Barrier analysis allows MORT analysis to be focussed;

On 23 November 2004 an angry crowd burnt a live

two police officers and seriously injured another after
mistaken them for child kidnappers. The third police
officer was finally rescued by colleagues three and
half hours after the attack began. (The three police
officers were members of the branch called Federal Preventive Police, known as PFP. They were
under the management of the Federal Police Headquarters; here it will be called FPHQ). The incident
occurred at San Juan Ixtayoapan, a neighborhood of
approximately 35,000 people on Mexico Citys southern outskirts; however, the incident is better known
as the Tlhuac case. It is believed that the police
officers were taken photographs of pupils at a primary school, where two children had recently gone
missing. TV reporters reached the scene before police
reinforcements, and live cameras caught a mob beating the police officers. However, the head of the FPHQ
said that back-up units were unable to get through for
more than three and a half hours because of heavy traffic. The main events are thought to be the following
(FIA 2004, 2005):
November 23rd 2004
17:55 hrs. Residents caught the PFP-police officers
in plain clothes taking photographs of pupils leaving
the school in the San Juan Ixtlayopan neighbourhood.
18:10 hrs. The crowd was told that the three PFPpolice officers were kidnappers.

13

it is very difficult to use MORT, even in a superficial

way, without it.

Losses

Oversights
and omissions

Assumed
risks

Specific
Control factors
LTA
S

SA1

SA2

Potentially
Harmful
condition
SB1

Information
systems
SD1

R2

The Barrier analysis is intended to produce a clear

set of episodes for MORT analysis. It is an essential
preparation for MORT analysis. The barrier analysis
embraces three key concepts, namely: {a} energy;
{b} target; and {c} barrier.
Energy refers to the harmful agent that threatens
or actually damages a Target that is exposed to it.
Targets can be people, things or processesanything, in fact, that should be protected or would be
better undisturbed by the Energy. In MORT, an incident can result either from exposure to an energy flow
without injuries or damage, or the damage of a target with no intrinsic value. Barrier part of the title
refers to the means by which Targets are kept safe
from Energies.

Rn

Mitigation
LTA

Incident

R1

Management
system factors
LTA

3.2 Barrier analysis

Implementation of
policy LTA

Policy
LTA
MA1

MA2

Inspection
LTA
SD2

SB3

Operational
readiness
LTA
SD3

MA3
Energy flows
leading
accident/incident

Controls &
Barriers
LTA

Vulnerable
people/objects
SB2

Risk
Assessment
System LTA

SB4

Maintenance
LTA
SD4

Super
vision
LTA
SD5

Supervision support LTA

Table 3.

Barrier analysis for the Tlahuac incident.

SD6

Energy flow Target

Figure 1.

Basic MORT structure.

Barrier

3-PFP
Tlahuac PFP-police officers in plain(Police
Suburb
clothes taking pictures of pupils
Officers)
leaving school.
OperaPFP-police officers did not have
tions
an official ID or equivalent to
demonstrate that they were
effectively members of the PFP.
No communication between the
PFP-police officers and the local
& regional authorities of the
purpose of their operations in
the neighborhood.

Figure 2. SB1 branchPotentially harmful energy flow or

condition. (Red: problems that contributed to the outcome;
Green: is judged to have been satisfactory)

14

Angry
crowd

3-PFP

PFP-police officers in plainclothes taking pictures of pupils

leaving school.
PFP-police officers did not have
an official ID or equivalent to
demonstrate that they were
effectively PFP-police officers.
Failure to investigate previous
kidnappings by the Police or
them?

Crowd
attack

3-PFP

PFP-police officers in plainclothes taking pictures of pupils

leaving school.
PFP-police officers did not have
an official ID or equivalent to
demonstrate that they were
effectively PFP-police officers.
No back-up units.

3.3

MORT structure

On the other hand, the Specific and Management

branches are regarded as the two main branches in
MORT (see Fig. 2). Specific control factors are broken
down in to two main classes: {a} those related to the

Figure 1 shows the basic MORT structure. The top

event in MORT is labelled Losses, beneath which
are its two alternative causes; i.e., {1} Oversights &
Omissions, {2} Assumed risks. In MORT all
the contributing factors in the accident sequence are
treated as oversights and omissions unless they are
transferred to the Assumed risk branch. Input to
the Oversights and Omissions event is through and
AND logic gate. This means that problems manifest
in the specific control of work activities, necessarily
involve issues in the management process that govern
them.

Figure 5. SC2 branchBarriers LTA. (Red: problems

that contributed to the outcome; Blue: need more information. Green: is judged to have been satisfactory).

Figure 3. SB2 branchVulnerable people. (Red: problems that contributed to the outcome; Blue: need more
information. Green: is judged to have been satisfactory).

Figure 6. SD1 branchtechnical information LTA.

(Red: problems that contributed to the outcome; Green: is
judged to have been satisfactory).

Figure 4. SB3 branchBarriers & Controls. (Red: problems that contributed to the outcome).

15

Figure 7. SD1 branchCommunication LTA. (Red:

problems that contributed to the outcome; Green: is judged
to have been satisfactory).
Figure 9. SD6 branchSupport of supervision LTA.
(Red: problems that contributed to the outcome; Green: is
judged to have been satisfactory).

Figure 8. SD1 branchData collection LTA. (Red: problems that contributed to the outcome; Blue: need more
information. Green: is judged to have been satisfactory).

Figure 10. MB1 branchHazard analysis process LTA.

(Red: problems that contributed to the outcome; Blue:
need more information. Green: is judged to have been
satisfactory).

incident or accident (SA1), and {b} those related to

restoring control following an accident (SA2). Both
of them are under an OR logic gate because either can
be a cause of losses.

energy flow (crowd attack) was considered for the

analysis.
The results of the MORT analysis are presented in
two ways, i.e., the MORT chart and brief notes on
the MORT chart. Both need to be read in conjunction with the NRI MORT Users manual (NRI-12002).

ANALYSIS & RESULTS

A barrier analysis for the Tlahuac incident has been

conducted and Table 3 shows the three possible phases
or episodes. At this stage of the investigation the third

16

blue, to indicate where there is a need to find more

information to properly assess it. Figures 211 show
several branches of the MORT chart for the Tlahuac
incident.
Table 4 summarizes some of the findings in brief
of the third phase identified by the Barrier analysis.

A preliminary analysis of the Tlahuac incident has

been conducted by applying the MORT technique.
Some preliminary conclusions have been highlighted.
However, it should be pointed out that some of the
branches of the Tree were not relevant or difficult
to interpret in the present case; for example: SD2Operational readiness LTA, SD3-Inspection LTA,
SD4-Maintenance LTA, etc. It may be argued that
these branches are designed to identify weaknesses in
a socio-technical system; however it may be the case
that they are not suitable for the case of a human activity system; such as crime. A constraint of the present
study was the difficulty of gathering reliable information of the chain of command within the police force
at the time of the incident.
More work is needed to draw some final conclusions of this fatal incident. The future work includes:
{1} the circles shown in blue in the chart MORT need
further investigation; {2} the other two energy flows
identified by the Barrier analysis should be investigated; i.e., Angry crowd & PFP-police officer
operations. Other energy flows may be identified and investigated. Finally, other accident analysis
approaches should be adopted for further analysis;
these include for example, PRISMA (Van der Schaaf
1996), STAMP (Levenson et al. 2003), SSMS model
(Santos-Reyes & Beard 2008), and others that may be
relevant. It is hoped that by conducting such analysis
lessons can be learnt so that incidents such as the case
of Tlhuac can be prevented in the future.

Figure 11. MB1 branch Management system factors.

(Red: problems that contributed to the outcome; Blue:
need more information. Green: is judged to have been
satisfactory).

Table 4.

DISCUSSION

Summary of main failures that were highlighted.

Poor attitude from management (i.e., federal Police &

Mexico Citys Police organizations)
Management laid responsibility to the traffic jams preventing them sending reinforcements to rescue the POs.
A hazard assessment was not conducted. Even if it was
conducted it did not highlight this scenario.
No one in management responsible for the POs safety
Management presumed/assumed that the three POs were
enough for this operation and did not provide reinforcements
Both the Pos and the management did not learn from
previous incidents
Lack of training. The Pos did not know what to do in such
circumstances
Lack of support from the management
No internal checks/monitoring
No independent checks/monitoring
Lack of coordination between the Federal & Local police
organizations
Lack of an emergency plan of such scenarios. It took four
hours to rescue the bodies of two POs and rescue the third
alive

ACKNOWLEDGEMENTS
This project was funded by CONACYT & SIP-IPN
under the following grants: CONACYT: No-52914 &
SIP-IPN: No-20082804.

REFERENCES
Alaggia, R., & Regehr, C. 2006. Perspectives of justice
for victims of sexual violence. Victims and Offenders 1:
3346.
Alalehto, T. 2002. Eastern prostitution from Russia to
Sweden and Finland. Journal of Scandinavian Studies in
Criminology and Crime Prevention 3: 96111.

In practice, to make the process of making brief notes

on the MORT chart easier to review, it is customary
to colour-code the chart as follows: {a} red, where
a problem is found; {b} green, where a relevant
issue is judged to have been satisfactory, and; {c}

17

Burgess, A.W., Garbarino, C., & Carlson, M.I. 2006. Pathological teasing and bulling turned deadly: shooters and
suicide. Victims and Offenders 1: 114.
Chiffriller, S.H., Hennessy, J.J., & Zappone, M. 2006. Understanding a new typology of batterers: implications for
treatment. Victims and Offenders 1: 7997.
Davis, P.K., & Jenkins, B.M. 2004. A systems approach to
deterring and influencing terrorists. Conflict Management
and Peace Science 21: 315.
Ekblom, P. 2005. How to police the future: scanning for
scientific and technological innovations which generate
potential threats & opportunities in crime, policing &
crime reduction (Chapter 2). M.J. Smith & N. Tilley (eds),
Crime Sciencenew approaches to prevent & detecting
crime: 2755. Willan Publishing.
Elklit, A. 2002. Attitudes toward rape victimsan empirical
study of the attitudes of Danish website visitors. Journal of Scandinavian Studies in Criminology and Crime
Prevention 3: 7383.
FIA, 2004. Linchan a agentes de la PFP en Tlhuac. Fuerza
Informativa Azteca (FIA), http://www.tvazteca.com/
noticias (24/11/2004).
FIA, 2005. Linchamiento en Tlhuac pareca celebracin.
Fuerza Informativa Azteca (FIA), http://www.tvazteca.
com/noticias (10/01/2005).
Finckenauer, J.O. 2005. The quest for quality in criminal
justice education. Justice Quarterly 22 (4): 413426.
Griffiths, H. 2004. Smoking guns: European cigarette
smuggling in the 1990s. Global Crime 6 (2): 185200.
Instituto Ciudadano de Estudios Sobre la Inseguridad
(ICESI), online www.icesi.org.mx.
Klein, J. 2005. Teaching her a lesson: media misses boys
rage relating to girls in school shootings. Crime Media
Culture 1 (1): 9097.
Lampe, K.V., & Johansen, P.O. 2004. Organized crime and
trust: on the conceptualization and empirical relevance of
trust in the context of criminal networks. Global Crime
6 (2): 159184.

Levenson, N.G., Daouk, M., Dulac, N., & Marais, K. 2003.

Applying STAMP in accident analysis. Workshop on the
investigation and reporting of accidents.
McCabe, M.P., & Wauchope, M. 2005. Behavioural characteristics of rapists. Journal of Sexual Aggression 11 (3):
235247.
NRI-1. 2002. MORT Users Manual. For use with the Management Oversight and Risk Tree analytical logic diagram.
Generic edition. Noordwijk Risk Initiative Foundation.
ISBN 90-77284-01-X.
Puglia, M., Stough, C., Carter, J.D., & Joseph, M. 2005.
The emotional intelligence of adult sex offenders: ability
based EI assessment. Journal of Sexual Aggression 11 (3):
249258.
Ratcliffe, J. 2005. The effectiveness of police intelligence management: A New Zealand case study. Police
Practice & Research 6 (5): 435451.
Santos-Reyes, J., & Beard, A.N. 2008. A systemic approach
to managing safety. Journal of Loss Prevention in the
Process Industries 21 (1): 1528.
Schmid, A.P. 2005. Root Causes of Terrorism: Some conceptual Notes, a Set of Indicators, and a Model, Democracy
and Security 1: 127136.
The international Crime Victim Survey (ICVS). Online
http://www.unicri.it/wwd/analysis/icvs/index.php.
Tucker, J. 2004. How not to explain murder: a sociological
critique of bowling for columbine. Global Crime 6 (2):
241249.
Van der Schaaf, T.W. 1996. A risk management tool based
on incident analysis. International Workshop on Process
Safety Management and Inherently Safer process, Proc.
Inter. Conf. 811 October, Orlando, Florida, USA:
242251.

18

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Comparing a multi-linear (STEP) and systemic (FRAM) method

for accident analysis
I.A. Herrera
Department of Production and Quality Engineering, Norwegian University of Science and Technology,
Trondheim, Norway

R. Woltjer
Department of Computer and Information Science, Cognitive Systems Engineering Lab,
Linkping University, Linkping, Sweden

ABSTRACT: Accident models and analysis methods affect what accident investigators look for, which contributing factors are found, and which recommendations are issued. This paper contrasts the Sequentially Timed
Events Plotting (STEP) method and the Functional Resonance Analysis Method (FRAM) for accident analysis and modelling. The main issues addressed in this paper are comparing the established multi-linear method
(STEP) with the systemic method (FRAM) and evaluating which new insights the latter systemic method provides for accident analysis in comparison to the former established multi-linear method. Since STEP and FRAM
are based on a different understandings of the nature of accidents, the comparison of the methods focuses on
what we can learn from both methods, how, when, and why to apply them. The main finding is that STEP helps
to illustrate what happened, whereas FRAM illustrates the dynamic interactions within socio-technical systems
and lets the analyst understand the how and why by describing non-linear dependencies, performance conditions,
variability, and their resonance across functions.

Researchers have argued that linear approaches fail

to represent the complex dynamics and interdependencies commonly observed in socio-technical systems
(Amalberti, 2001; Dekker, 2004; Hollnagel, 2004;
Leveson, 2001; Rochlin, 1999; Woods & Cook, 2002).
Recently, systemic models and methods have been
proposed that consider the system as a whole and
emphasize the interaction of the functional elements.
FRAM (Hollnagel, 2004) is such a systemic
model. FRAM is based on four principles (Hollnagel,
Pruchnicki, Woltjer & Etcher, 2008). First, the principle that both successes and failures result from the
adaptations that organizations, groups and individuals perform in order to cope with complexity. Success
depends on their ability to anticipate, recognise, and
manage risk. Failure is due to the absence of that ability (temporarily or permanently), rather than to the
inability of a system component (human or technical)
to function normally. Second, complex socio-technical
systems are by necessity underspecified and only
partly predictable. Procedures and tools are adapted
to the situation, to meet multiple, possibly conflicting
goals, and hence, performance variability is both normal and necessary. The variability of one function is
seldom large enough to result in an accident. However,

INTRODUCTION

Analysing and attempting to understand accidents is an

essential part of the safety management and accident
prevention process. Many methods may be used for
this, however they often implicitly reflects a specific
view on accidents. Analysis methodsand thus their
underlying accident modelsaffect what investigators
look for, which contributing factors are found, and
which recommendations are made. Two such methods
with underlying models are the Sequentially Timed
Events Plotting method (STEP; Hendrick & Benner,
1987) and the Functional Resonance Accident Model
with the associated Functional Resonance Analysis
Method (FRAM; Hollnagel, 2004, 2008a).
Multi-linear event sequence models and methods
(such as STEP) have been used in accident analysis
to overcome the limitations of simple linear causeeffect approaches to accident analysis. In STEP, an
accident is a special class of process where a perturbation transforms a dynamically stable activity into
unintended interacting changes of states with a harmful outcome. In this multi-linear approach, an accident
is viewed as several sequences of events and the system
is decomposed by its structure.

19

by the co-pilot as Pilot-Flying (PF) and the captain

as Pilot Non-Flying (PNF). Shortly after clearance
to 4000 ft, the crew was informed that runway 19R
was closed because of sweeping and that the landing
should take place on runway 19L. The aircraft was
guided by air traffic control to land on 19L. Changing
of the runway from 19R to 19L resulted in change in
the go-around-altitude from 4000 ft at 19R to 3000 ft
at 19L. The crew performed a quick briefing for a new
final approach.
During the last part of the flight, while the aircraft was established on the localizer (LLZ) and glide
slope (GS) for runway 19L, the glide slope signal
failed. It took some time to understand this for the
pilots, who had not yet switched to tower (TWR) frequency from APP frequency after acknowledging the
new frequency. Immediately after the glide path signal disappeared the aircraft increased its descent rate to
2200 ft/min while being flown manually towards LLZminima. The aircraft followed a significantly lower
approach than intended and was at its lowest only 460 ft
over ground level at DME 4,8. The altitude at this distance from the runway should have been 1100 ft higher.
The crew initiated go-around (GA) because the aircraft
was still in dense clouds and it drifted a little from the
LLZ at OSL. However, the crew did not notice the
below-normal altitude during approach. Later a new
normal landing was carried out.
The executive summary of the Norwegian Accident
Investigation Board (AIBN) explains that the investigation was focused on the glide slope transmission,
its technical status and information significance for
the cockpit instrument systems combined with cockpit
human factors. The AIBN understanding of the situation attributes the main cause of the incident to the
pilots incorrect mental picture of aircraft movements
and position. The report concludes that the in-cockpit
glide slope capture representation was inadequate. In
addition, the report points to a deficiency in the procedure for transfer of responsibility between approach
and tower air traffic control. (AIBN, 2004)
Five recommendations resulted from the AIBN
investigation. The first recommendation is that the
responsibility between controls centres should be
transferred 8 NM before landing or at acceptance
of radar hand-over. The second recommendation is
related to the certification of avionics displays, advising the verification of the information provided to
pilots, with special attention to glide slope and autopilot status information. Third, training should take
into account glide slope failures after glide slope capture under ILS approach. Fourth, Oslo airport should
consider the possibility of providing radar information
to the tower controller to be able to identify approach
paths deviations. The last recommendation is for the
airline to consider situational awareness aspects in the
crew resource management (CRM) training.

the third principle states that the variability of multiple

functions may combine in unexpected ways, leading to disproportionately large consequences. Normal
performance and failure are therefore emergent phenomena that cannot be explained by solely looking at
the performance of system components. Fourth, the
variability of a number of functions may resonate,
causing the variability of some functions to exceed
normal limits, the consequence of which may be an
accident. FRAM as a model emphasizes the dynamics and non-linearity of this functional resonance, but
also its non-randomness. FRAM as a method therefore aims to support the analysis and prediction of
functional resonance in order to understand and avoid
accidents.

RESEARCH QUESTIONS AND APPROACH

The main question addressed in this paper is which

new insights this latter systemic method provides for
accident analysis in comparison to the former established multi-linear method. Since the accident analysis
methods compared in this paper are based on a different understanding of the nature of accidents, the
comparison of the methods focuses on what we can
learn from both methods, how, when, and why to apply
them, and which aspects of these methods may need
improvement.
The paper compares STEP and FRAM in relation
to a specific incident to illustrate the lessons learned
from each method. The starting point of the study
is the incident investigation report. A short description of STEP and FRAM is included. For a more
comprehensive description, the reader is referred to
Hendrick and Benner (1987; STEP) and Hollnagel
(2004; FRAM). Since different methods invite for
different questions to be asked, it was necessary to
interview air traffic controllers, pilots, and accident
investigators to acquire more information. The information in this paper was collected through interviews
and workshops involving a total of 50 people. The analysis with STEP and FRAM was an iterative process
between researchers and operative personnel.

SUMMARY OF THE INCIDENT

A Norwegian Air Shuttle Boeing 737-36N with callsign NAX541 was en-route from Stavanger Sola airport to Oslo Gardermoen airport (OSL). The aircraft
was close to Gardermoen and was controlled by Oslo
Approach (APP). The runway in use at Gardermoen
was 19R. The aircraft was cleared to reduce altitude to
4000 ft. The approach and the landing were carried out

20

14:42:36

14:42:55

14:42:57

14:44:02

TIME LINE

ACTORS

AIR TRAFFIC

OSLO APP CONTROL

AIRCRAFT AC-1

TWR INFORMS
G/S FAIL AC-2
14:42:57

GARDERMOEN TWR
CONTROL

RUNWAY EQUIP.
RWY-E

Figure 1.

APP REQUEST AC-1 TO CHANGE TO

TWR FRQ 14:42:36

CAPTAIN, COPILOT
PNF, PF

RWY-E ACTIVATES
ALARM G/S FAIL
14:42:55
PNF ACCEPTS
TRANSFER
14:42:38

AIRCRAFT
AC-1

PNF CHANGES
TO TWR FRQ
AC-1 NOSE MOVES
DOWN, DISCONNECT
A/P 14:43:27

PNF
MANUAL

GO
AROUND

ALTITUDE
460ft

AC-1 CHANGES FRQ

TO TWR 14:44:02

STEP applied to NAX541 incident (simplified example).

represented in STEP are related to normal work and

help to predict future risks. The safety problems are
identified by analysing the worksheet to find events
sets that constitute the safety problem. The identified safety problems are marked as triangles in the
worksheet. These problems are evaluated in terms of
severity. Then, they are assessed as candidates for recommendations. A STEP change analysis procedure
is proposed to evaluate recommendations. Five activities constitute this procedure. The identification of
countermeasures to safety problems, the ranking of the
safety effects, assessment of the trade-off involved the
selection of the best recommendations and a quality
check.

SEQUENTIAL TIMED EVENTS PLOTTING

STEP provides a comprehensive framework for accident investigation from the description of the accident
process, through the identification of safety problems,
to the development of safety recommendations. The
first key concept in STEP is the multi-linear event
sequence, aimed at overcoming the limitations of the
single linear description of events. This is implemented in a worksheet with a procedure to construct a
flowchart to store and illustrate the accident process.
The STEP worksheet is a simple matrix. The rows are
labelled with the names of the actors on the left side.
The columns are labelled with marks across a time line.
Secondly, the description of the accident is performed by universal events building blocks. An event
is defined as one actor performing one action. To
ensure that there is a clear description the events are
broken down until it is possible to visualize the process and be able to understand its proper control. In
addition, it is necessary to compare the actual accident
events with what was expected to happen.
A third concept is that the events flow logically in
a process. This concept is achieved by linking arrows
to show proceed/follow and logical relations between
events. The result of the third concept is a cascading
flow of events representing the accident process from
the beginning of the first unplanned change event to the
last connected harmful event on the STEP worksheet.
The organization of the events is developed and
visualized as a mental motion picture. The completeness of the sequence is validated with three tests.
The row test verifies that there is a complete picture of
each actors actions through the accident. The column
test verifies that the events in the individual actor rows
are placed correctly in relation to other actors actions.
The necessary and sufficient test verifies that the early
action was indeed sufficient to produce the later event,
otherwise more actions are necessary.
The STEP worksheet is used to have a link between
the recommended actions and the accident. The events

APPLICATION OF STEP TO NAX541

The incident is illustrated by a STEP diagram. Due

to page and paper limitations, Figure 1 illustrates a
small part of the STEP diagram that was created based
on the incident report. In Figure 1, the time line is
on along the X-axis and the actors are on the Y-axis.
An event is considered to mean an actor performing
one action. The events are described in event building
blocks, for example APP request to A/C to change
to TWR frequency. An arrow is used to link events.
Safety problems are illustrated on the top line by triangles in the incident process. Three such problems
were identified: 1) no communication between aircraft 1 (NAX541) and tower (triangle 1 in Figure 1);
2) changed roles between PF and PNF not coordinated;
and 3) pilots not aware of low altitude (2 and 3 not
shown in simplified figure).
6

FUNCTIONAL RESONANCE ANALYSIS

METHOD

FRAM promotes a systemic view for accident analysis. The purpose of the analysis is to understand

21

the characteristics of system functions. This method

takes into account the non-linear propagation of events
based on the concepts of normal performance variability and functional resonance. The analysis consists of
four steps (that may be iterated):
Step 1: Identifying essential system functions, and
characterizing each function by six basic parameters. The functions are described through six aspects,
in terms of their input (I, that which the function uses or transforms), output (O, that which the
function produces), preconditions (P, conditions that
must be fulfilled to perform a function), resources
(R, that which the function needs or consumes), time
(T, that which affects time availability), and control
(C, that which supervises or adjusts the function), and
may be described in a table and subsequently visualized in a hexagonal representation (FRAM module,
Figure 2). The main result from this step is a FRAM
model with all basic functions identified.
Step 2: Characterizing the (context dependent)
potential variability through common performance
conditions. Eleven common performance conditions
(CPCs) are identified in the FRAM method to be
used to elicit the potential variability: 1) availability
of personnel and equipment, 2) training, preparation,
competence, 3) communication quality, 4) humanmachine interaction, operational support, 5) availability of procedures, 6) work conditions, 7) goals, number
and conflicts, 8) available time, 9) circadian rhythm,
stress, 10) team collaboration, and 11) organizational
quality. These CPCs address the combined human,
technological, and organizational aspects of each function. After identifying the CPCs, the variability needs
to be determined in a qualitative way in terms of stability, predictability, sufficiency, and boundaries of
performance.
Step 3: Defining the functional resonance based on
possible dependencies/couplings among functions and
the potential for functional variability. The output of
the functional description of step 1 is a list of functions
each with their six aspects. Step 3 identifies instantiations, which are sets of couplings among functions for
specified time intervals. The instantiations illustrate
how different functions are active in a defined context.
Time T

Precondition
Figure 2.

C Control

Activity/
Function

Input I

The description of the aspects defines the potential

links among the functions. For example, the output of
one function may be an input to another function, or
produce a resource, fulfil a pre-condition, or enforce
a control or time constraint. Depending on the conditions at a given point in time, potential links may
become actual links; hence produce an instantiation
of the model for those conditions. The potential links
among functions may be combined with the results of
step 2, the characterization of variability. That is, the
links specify where the variability of one function may
have an impact, or may propagate. This analysis thus
determines how resonance can develop among functions in the system. For example, if the output of a
function is unpredictably variable, another function
that requires this output as a resource may be performed unpredictably as a consequence. Many such
occurrences and propagations of variability may have
the effect of resonance; the added variability under the
normal detection threshold becomes a signal, a high
risk or vulnerability.
Step 4: Identifying barriers for variability (damping
factors) and specifying required performance monitoring. Barriers are hindrances that may either prevent
an unwanted event to take place, or protect against
the consequences of an unwanted event. Barriers can
be described in terms of barrier systems (the organizational and/or physical structure of the barrier)
and barrier functions (the manner by which the barrier achieves its purpose). In FRAM, four categories
of barrier systems are identified: 1) physical barrier systems block the movement or transportation
of mass, energy, or information, 2) functional barrier systems set up pre-conditions that need to be
met before an action (by human and/or machine)
can be undertaken, 3) symbolic barrier systems are
indications of constraints on action that are physically present and 4) incorporeal barrier systems are
indications of constraints on action that are not physically present. Besides recommendations for barriers,
FRAM is aimed at specifying recommendations for
the monitoring of performance and variability, to be
able to detect undesired variability.

APPLICATION OF FRAM TO NAX541

Step 1 is related to the identification and characterization of functions: A total of 19 essential functions
were identified and grouped in accordance to the area
of operation. There are no specified rules for the level
of granularity, instead functions are included or split
up when the explanation of variability requires. In this
particular analysis some higher level functions, e.g.
Oslo APP control, and some lower level functions,
e.g. Change frq to TWR control.

O Output

R Resource

A FRAM module.

22

The operative areas and functions for this particular

incident are:

Table 2.

Function:
Manual approach

Crew operations: Change Runway (RWY) to 19L,

New final approach briefing, Auto-pilot approach
(APP), Change APP frequency (frq) to TWR
frq, Manual approach, GO-AROUND, Landing,
Approach, Receiving radio communication, Transmitting radio communication
Avionics Functions: Disconnect Autopilot (A/P),
Electronic Flight Instrument (EFIS), Ground Proximity Warning System (GPWS)
Air traffic control: Oslo APP control, RWY sweeping, Glideslope transmission, Gardermoen TWR
control
Aircraft in the vicinity: AC-2 communication, AC-3
communication

Availability
of resources
(personnel,
equipment)
Training,
preparation,
competence
Communication
quality
HMI operational
support
Avail. procedures
Work conditions

The NAX541 incident report contains information

that helps to define aspects of functional performance.
Essential functions are described with these aspects.
Table 1 shows an example of the aspects of the function
Manual Approach. Similar tables were developed
for 18 other functions.
In step 2 the potential for variability is described using
a list of common performance conditions (CPCs).
Table 2 presents an example of CPCs for the function
Manual Approach.
The description of variability is based on the information registered in the incident report combined with
a set of questions based on the CPCs. Since little of
this information regarding variability was available,
it was necessary to interview operational personnel
(air traffic controllers, pilots). An example is for CPC
HMI, operational support, a question was how aware
pilots are of these EFIS, GPWS discrepancies, a pilot
stated Boeing manuals explain which information is
displayed, it is normal to have contradictory information. In this case an understanding of the system as a
whole is required. Pilots needs to judge relevant information for each situation. An additional example of
questions for the function Runway change, was if
Table 1.

Input
Output
Preconditions
Resources
Time
Control

Performance
conditions

Rating
Adequate

PF little
experience on type

Temporarily
inadequate

Delay to contact
tower
Unclear alerts

Inefficient

Interruptions?

# Goals, conflicts

Overloaded

Available time

Task synchronisation

Circadian rhythm
Team collaboration
Org. quality

Switched roles

Inadequate
Adequate
Temporarily
inadequate?
More than
capacity
Temporarily
inadequate
Adjusted
Inefficient

it is normal and correct to request a runway change

with such a short notice? The interviews identified
that there are no formal operational limits for tower
air traffic controllers, but for pilots there are. Thus
an understanding of performance and variability was
obtained.
In step 3 links among functions are identified for
certain time intervals. States are identified to be valid
during specific time intervals, which define links
among the aspects of functions, hence instantiate
the model. An example instantiation is presented in
Figure 3, where some of the links during the time interval 14:42:3714:43:27 of the incident are described as
an instantiation of the FRAM that resulted from step
1. Many more such instantiations may be generated,
but here only one example can be shown.
To understand the events in relation to links and
functions in this instantiation, numbers 15 and letters
a-d have been used to illustrate two parallel processes.
Following the numbers first, the APP controller communicates to the pilot that they should contact TWR
at the TWR frequency (1). This is an output of Oslo
APP control, and an input to Receiving radio communication. This latter function thus has as output the
state that transfer is requested to the TWR frequency
(2), which matches the preconditions of Change APP
frq to TWR frq, and Transmitting radio communication. The fulfilment of this precondition triggers
the pilots to acknowledge the transfer to TWR to the
APP controller (3), an output of transmitting function,

A FRAM module function description.

Function:
Manual approach

Manual Flight Approach CPCs.

Aspect description
GPWS alarms,
pilot informed of G/S failure
Altitude in accordance with
approach path, Altitude lower/
higher than flight path
A/P disconnected
Pilot Flying, Pilot Non Flying
Efficiency Thoroughness TradeOff, time available varies
SOPs

23

A/C-1 pilot & A/C functions

Change
APP frq to
TWR frq

A/C-1 avionics ept

Oslo APP control

Auto-pilot
approach

Manual
approach

Gardermoen TWR control

Ground equipment

5,d)Pilotinformed
of G/S failure

2)Transfer requested
toTWRfrq
T

4) Frequency still
set to APP

C
Transmitting radio
comm

14:43:27

O
I

Receiving
rad io comm

T
1) APP -Pilot:
con tact TWR
on TWR frq

3) Pilot-APP :
to TWR frq
C

Oslo APP
con trol

6) Pilot-TWR :
Fligh t on TWRfrq
T

Gardermoen
TWR
con trol
P

Auto-p ilot

b)TWR-pilot:
informa/cof
G/Sfailure

R
I

a) G/S lost
14:42:55

Glide slope
transmission

P
X) Proactive TWR -APP comm:
check
frequency change

Figure 3.

R
P

c) A/P d isconn ected

a) no G/S sign al
14:42:55

A FRAM instantiation during the time interval 14:42:3714:43:27 with incident data.

(d). Concurrently, the loss of G/S no longer fulfils the

precondition of the auto-pilot function, with the resulting output of A/P being disconnected (c) about half a
minute after G/S loss. This in turn no longer fulfils
the precondition of an auto-pilot approach and instead
matches the precondition for a manual approach. All
of this in turn results in variability on the manual
approach, e.g. with decreased availability of time,
inadequate control because of PF-PNF collaboration
problems, and inadequate resources (e.g. displays
unclear indications of A/P and G/S) resulting in highly
variable performance (output) of the manual approach.
Step 4 addresses barriers to dampen unwanted variability and performance variability monitoring where
variability should not be dampened. AIBN recommendations could be modelled as barrier systems and
barrier functions, e.g. Responsibility between control centres should be transferred 8 NM before landing,
or at acceptance by radar hand over. (AIBN, p. 31,
our translation). In FRAM terminology this can be

input to Oslo APP control. The pilots however do

not switch immediately after the transfer is requested,
hence the output is that the frequency still is set to
APP, for a much longer time than would be intended
(indicated by the red O), and the pilots do not contact
TWR (6) until much later. This has consequences for
the precondition of receiving/transmitting (4), which
is being on the same frequency with the control centre
that has responsibility for the flight. With the delay in
frequency change, the link that the pilot is informed
of the G/S failure (5) is also delayed.
At about the same time, following the letters in
Figure 3, Glide slope transmission changes output to
that there is no G/S signal at 14:42:55 (a), because of a
failure of the G/S transmitting equipment (a resource,
R in red). This makes the TWR controller inform
pilots on the TWR frequency of the G/S failure (b),
excluding the incident aircraft crew because of the
unfulfilled precondition because of link (4), delaying the point that the pilot is informed of G/S failure

24

performance varied becomes apparent: For example,

the operational limits for runway change for different
operators were discussed; the question of why the frequency change was delayed gets answered based on
the normal variability in pilot-first-officer-interaction
patterns in cases of experience difference; the pilots
unawareness of the low altitude is understandable with
regard to variability related to e.g. team collaboration
and human-machine interface issues.
STEP provides a mental motion picture
(Hendrick & Benner, 1987, p. 75) illustrating
sequences of events and interactions between processes, indicating what happened when. FRAM
instead sketches a functional slide show with its
illustrations of functions, aspects, and emerging links
between them in instances, indicating the what and
when, and common performance conditions, variability, and functional resonance, indicating why.
FRAMs qualitative descriptions of variability provide more gradations in the description of functions
than the bimodal (success/failure) descriptions typical
for STEP.
In relation to the question of when each method
should be used, the type of incident and system to
be analysed needs to be taken into account. STEP
is suited to describe tractable systems, where it is
possible to completely describe the system, the principles of functioning are known and there is sufficient
knowledge of key parameters. FRAM is better suited
for describing tightly coupled, less tractable systems
(Hollnagel, 2008b), of which the system described in
this paper is an example. Because FRAM does not
focus only on weaknesses but also on normal performance variability, this provides a more thorough
understanding of the incident in relation to how work
is normally performed. Therefore the application of
FRAM may lead to a more accurate assessment of
the impact of recommendations and the identification of previously unexplored factors that may have
a safety impact in the future. While the chain of events
is suited for component failures or when one or more
components failed, they are less adequate to explain
systems accidents (Leveson, 2001). This can be seen
in the STEP-FRAM comparison here. The STEP diagram focuses on events and does not describe the
systems aspects: the understanding of underlying systemic factors affecting performance is left to experts
interpretation. FRAM enables analysts to model these
systemic factors explicitly.

described as an incorporeal prescribing barrier. This

barrier would have an effect on the variability of the
APP and TWR control functions through the aspect
of control and the links between input and output in
various instantiations describing communication and
transfer of responsibility. New suggestions for barriers
also result from the FRAM. For example, a proactive communication from TWR to APP when a flight
does not report on frequency would link their output and input (see link (X) in Figure 3), triggering
instantiations of links 16 so that control and contact
is re-established. This barrier may be implemented in
various systems and functions, such as through regulation, training, procedures, checklists and display
design, etc. The FRAM also points to the interconnectivity of air traffic control and pilot functions,
suggesting joint training of these operators with a wide
range of variability in the identified functions. As with
any method, FRAM enables the suggestion of barriers
(recommendations), which need to be evaluated by
domain experts in terms of feasibility, acceptability,
and cost effectiveness, among other factors.
The FRAM and the instantiations that were created
here also point to the future development of indicators
for matters such as overload and loss of control when
cockpit crew has significant experience differences.
8

COMPARISON

Accident models, implicitly underlying an analysis or

explicitly modelling an adverse event, influence the
elicitation, filtering, and aggregation of information.
Then, what can we learn from the applications of STEP
and FRAM to this incident?
STEP is relatively simple to understand and provides a clear picture of the course of the events.
However, STEP only asks the question of which events
happened in the specific sequence of events under
analysis. This means that events mapped in STEP are
separated from descriptions of the normal functioning of socio-technical systems and their contexts. For
example, the STEP diagram illustrates that the PNFs
switch to TWR frequency was delayed, but not why.
Instead, STEP only looks for failures and safety problems, and highlights sequence and interaction between
events. FRAM refrains from looking for human errors
and safety problems but tries to understand why the
incident happened. Since FRAM addresses both normal performance variability and the specifics of an
adverse event, FRAM broadens data collection of the
analysis compared to a STEP-driven analysis: Thus
the development of the incident is contextualized in a
normal socio-technical environment. Through asking
questions based on the common performance conditions and linking functions in instantiations, FRAM
identified additional factors and the context of why

CONCLUSIONS AND PRACTICAL

IMPLICATIONS

This paper presented two accident analysis methods:

The multi-sequential STEP and systemic FRAM. The
question of how to apply these methods was addressed

25

ACKNOWLEDGEMENTS

by discussing the steps of the methods, illustrated by

applying these methods to a missed approach incident.
This paper concluded that FRAM provides a different
explanation about how events are a result of variability of normal performance and functional resonance,
compared to STEP. The main finding is that STEP
helps to illustrate what happened, whereas FRAM covers what happened and also illustrates the dynamic
interactions within the socio-technical system and lets
the analyst understand the how and why by describing non-linear dependencies, performance conditions
and variability, and their resonance across functions.
Another important finding is that it was possible to
identify additional factors with FRAM. STEP interpretations and analysis depends on investigator experience, FRAM introduces questions for systemic factors
and enables the explicit identification of other relevant
aspects of the accident. The example also illustrates
how unwanted variability propagates such as the information about G/S failure and the undesired resonance
with the differences in pilots experience. However,
several incidents in different contexts would need to
be analysed to validate and generalize these findings.
Two practical implications are found. The first is
that FRAM provides new ways of understanding failures and successes, which encourages investigators to
look beyond the specifics of the failure under analysis
into the conditions of normal work. The second is that
it models and analyses an intractable socio-technical
system within a specific context. While FRAM as a
model has been accepted in the majority of discussions
with practitioners, and seems to fill a need for understanding intractable systems, FRAM as a method is
still young and needs further development. This paper
has contributed to the development of the method by
outlining a way to illustrate instantiations for a limited
time interval. An additional need is the identification
of normal and abnormal variability which this paper
has addressed briefly. Remaining challenges include
a more structured approach to generating recommendations in terms of barriers and indicators, as well
as evaluating how well FRAM is suited as a method
to collect and organize data during early stages of
accident investigation.

This work has benefited greatly from the help and support of several aviation experts and the participants in
the 2nd FRAM workshop. We are particularly grateful
to the investigators and managers of the Norwegian
Accident Investigation Board who commented on a
draft of the model. Thanks to Ranveig K. Tinmannsvik,
Erik Jersin, Erik Hollnagel, Jrn Vatn, Karl Rollenhagen, Kip Smith, Jan Hovden and the conference
reviewers for their comments on our work.

REFERENCES
AIBN. 2004. Rapport etter alvorlig luftfartshendelse ved
Oslo Lufthavn Gardermoen 9. Februar 2003 med Boeing 737-36N, NAX541, operert av Norwegian Air Shuttle.
Aircraft Investigation Board Norway, SL RAP.:20/2004.
Amalberti, R. 2001. The paradoxes of almost totally safe
transportation systems. Safety Science, 37, 109126.
Dekker, S.W.A. 2004. Ten questions about human error: A
new view of human factors and system safety. Mahwah,
NJ: Lawrence Erlbaum.
Hendrick, K., Benner, L. 1987. Investigating accidents with
STEP. Marcel Dekker Inc. New York.
Hollnagel, E. 2004. Barriers and accident prevention. Aldershot, UK: Ashgate.
Hollnagel, E. 2008a. From FRAM to FRAM. 2nd FRAM
Workshop, Sophia-Antipolis, France.
Hollnagel, E. 2008b. The changing nature of risks. Ecole des
Mines de Paris, Sophia Antipolis, France.
Hollnagel, E., Pruchnicki, S., Woltjer, R., & Etcher, S. 2008.
Analysis of Comair flight 5191 with the Functional Resonance Accident Model. Proc. of the 8th Int. Symp. of
the Australian Aviation Psychology Association, Sydney,
Australia.
Leveson, N. 2001. Evaluating accident models using recent
aerospace accidents. Technical Report, MIT Dept. of
Aeronautics and Astronautics.
Perrow, C. 1999. Normal accidents. Living with high risk
technologies. Princeton: Princeton University Press. (First
issued in 1984).
Rochlin, G.I. 1999. Safe operation as a social construct.
Ergonomics, 42, 15491560.
Woods, D.D., & Cook, R.I. 2002. Nine steps to move forward
from error. Cognition, Technology & Work, 4, 137144.

26

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Development of a database for reporting and analysis of near misses

in the Italian chemical industry
R.V. Gagliardi
ISPESL Department of Industrial Plants and Human Settlements, Rome, Italy

G. Astarita
Federchimica Italian Federation of Chemical Industries, Milan, Italy

ABSTRACT: Near misses are considered to be an important warning that an accident may occur and therefore
their reporting and analysis may have a significant impact on industrial safety performances, above all for those
industrial sectors involving major accident hazards. From this perspective, the use of a specific information
system, including a database ad hoc designed for near misses, constitutes an appropriate software platform
that can support company management in collecting, storing and analyzing data on near misses, and also
implementing solutions to prevent future accidents. This paper describes the design and the implementation of
such a system, developed in the context of a cooperation agreement with the Italian Chemical Industry Federation.
This paper also illustrates the main characteristics and utilities of the system, together with future improvements
that will be made.
1

large number of minor accidents and even more near

misses, representing the lower portion of the pyramid.
As a consequence, reducing the near misses should
mean reducing the incidence of major accidents. In
addition, due to the fact that near misses frequently
occur in large facilities, they represent a large pool of
data which is statistically more significant than major
accidents. This data could be classified and analyzed
in order to extract information which is significant in
improving safety.
The reporting of near misses, therefore, takes on
an important role in industrial safety, because it is
the preliminary step of subsequent investigative activities aimed at identifying the causes of near misses and
implementing solutions to prevent future near misses,
as well as more serious accidents. However, a real
improvement in industrial safety level as a whole can
only be reached if these solutions are disseminated
as widely as possible. They must also be in a form
that can be easily and quickly retrieved by all personnel involved at any stage of the industrial process.
From this perspective, the use of a specific and widely
accessible software system, which is designed for
facilitating near miss data collection, storage and analysis, constitutes an effective tool for updating process
safety requirements.
This paper describes the design and implementation of such a system, carried out in the context of
a cooperation agreement with the Italian Chemical

INTRODUCTION

The importance of post-accident investigation is

widely recognized in industrial sectors involving the
threat of major accident hazards, as defined in the
European Directive 96/82/CE Seveso II (Seveso II,
1997): in fact, by examining the dynamics and
potential causes of accidents, lessons can be learned
which could be used to identify technical and managerial improvements which need to be implemented
in an industrial context. This is in order to prevent the occurrence of accidents and/or mitigate their
consequences.
Such investigations are even more significant in the
case of near misses, that is, those hazardous situations (events or unsafe acts) that could lead to an
accident if the sequence of events is not interrupted
(Jones & al. 1999). Philley & al. 2003, also stated that
a near miss can be considered as an event in which
property loss, human loss, or operational difficulties could have plausibly resulted if circumstances had
been slightly different. Their significance, according
to a shared awareness in the scientific world, is due to
the fact that near misses are an important warning that
a more serious accident may occur. In fact, several
studies (CCPS Guidelines, 2003), using the pyramid
model, suggest a relationship between the number
of near misses and major accidents: for each major
accident located at the top of the pyramid there are a

27

the assumption that a similar approach can benefit

the whole industrial sector involving major accident
hazards, the development of an informative system
for the collection and analysis of near misses in the
chemical process industry, has been undertaken at a
national level. This is thanks to a strong cooperation
agreement between the Italian National Institute for
Prevention and Safety at Work (ISPESL), and Federchimica. ISPESL is a technical-scientific institution
within the Ministry of Health, which supports Italian Competent Authorities for the implementation of
the Seveso legislation in Italy; in recent decades it
has acquired wide expertise in post major accidents
investigations and reporting activities, as well as in
the inspection of establishments which are subject to
the Seveso legislation. Federchimica is the Italian federation of chemical industries and is comprised of
more than 1300 companies, including several establishments which are subject to the Seveso legislation.
In Italy, Federchimica leads the Responsible Care
voluntary worldwide programme for the improvement
of health, safety and environmental performances
in the chemical industry. The authors of this paper
hope that a joint effort in the field of near misses
reporting and investigation, incorporating different
expertises and perspectives from ISPESL and Federchimica, could result in a real step forward in industrial
safety enhancement.

Industry Federation. The various steps undertaken for

the development of such a system, involved in recording near misses in the Italian chemical industry, are
presented below.

2
2.1

FRAMEWORK FOR A NATIONWIDE

APPROACH
Legislative background

As a preliminary remark, from a legislative perspective, the recommendation that Member States report
near misses to the Commissions Major Accident
Reporting System (MARS) on a voluntary basis has
been introduced by the European Directive 96/82/EC
Seveso II. This is in addition to the mandatory requirements of major accident reporting. More
specifically, in annex VI of the Directive, in which
the criteria for the notification of an accident to the
Commission are specified, is included a recommendation that near misses of particular technical interest
for preventing major accidents and limiting their consequences should be notified to the Commission. This
recommendation is included in the Legislative Decree
n. 334/99 (Legislative Decree n. 334, 1999), the
national law implementing the Seveso II Directive in
Italy.
In addiction further clauses regarding near misses
are included in the above mentioned decree, referring to the provisions concerning the contents of the
Safety Management System in Seveso sites. In fact,
the Decree states that one of the issues to be addressed
by operators in the Safety Management System is
the monitoring of safety performances; this must be
reached by taking into consideration, among other
things, the analysis of near misses, functional anomalies, and corrective actions assumed as a consequence
of near misses.
2.2

SYSTEM DESIGN AND REQUIRED

ATTRIBUTES

In the context of the above mentioned cooperation

agreement, an informative system has been developed consisting of a web based software and a
database designed for near misses. The system is
specifically directed towards companies belonging to
Federchimica which have scientific knowledge and
operating experience in the prevention of accidents and
safety control; moreover ISPESL personnel, who are
involved in inspection activities on Seveso sites, are
authorized to enter the system in order to acquire all
elements useful for enhancing its performances in prevention and safety issues. Two main goals have been
taken into account with regard to the project design:
first, gather as much information as possible on near
misses events occurring in the national chemical process industry. This is in order to build a reliable and
exhaustive database on this issue. Second, provide
a tool which is effective in the examination of near
misses and extraction of lesson learned, and which
aims to improve industrial safety performances.
To meet these requirements a preliminary analysis
of specific attributes to be assigned to the system has
been carried out in order to define its essential characteristics. The first attribute which has been considered

State of the art

Although legislation clearly introduces the request for

the reporting of near misses, the present situation,
in Italy, is that this activity is carried out not on a
general basis but only by a number of companies,
which use their own databases to share learning internally. It should also be noted that similar initiatives
in this field have been undertaken by public authorities, in some cases, at a regional level. However
a wider, more systematic and nationwide approach
for the identification, reporting and analysis of near
misses is lacking. To fill this gap a common software
platform, not specifically for each individual situation,
facilitating the dissemination of information between
different establishments as well as different industrial
sectors, must be implemented. Therefore, based on

28

is accessibility: To allow a wider diffusion, the system has been developed in such a way that it is easy
to access when needed. A web-version has therefore
been built, which assures security, confidentiality and
integrity criteria of the data handled. Secondly, in order
to guarantee that the system can be used by both expert
and non-expert users, the importance of the system
being user-friendly has been an important consideration in its design. The user must have easy access
to the data in the system and be able to draw data from
it by means of pull-down menus, which allow several
options to be chosen; moreover, help icons clarify
the meaning of each field of the database to facilitate data entry. Third, the system must be designed in
such a way as to allow the user to extract information
from the database; to this end it has been provided
with a search engine. By database queries and the
subsequent elaboration of results, the main causes of
the near misses and the most important safety measures adopted to avoid a repetition of the anomaly can
be singled out. In this way the system receives data
on near misses from the user and, in return, provides
information to the user regarding the adoption of corrective actions aimed at preventing similar situations
and/or mitigating their consequences. Lastly, the system must guarantee confidentiality: this is absolutely
necessary, otherwise companies would not be willing to provide sensitive information regarding their
activities. In order to ensure confidentiality, the data
are archived in the database in an anonymous form.
More precisely, when the user consults the database,
the only accessible information on the geographical
location of the event regards three macro-areas, Northern, Central and Southern Italy respectively; this is in
order to avoid that the geographical data inserted in the
database (municipality and province) in the reporting
phase could lead to the identification of the establishment in which near misses have occurred. Another
factor which assures confidentiality is the use of usernames and passwords to enter the system. These are
provided by Federchimica, after credentials have been
vetted, to any company of the Federation, which has
scientific knowledge and operating experience in prevention and safety control. In this way, any company,
fulfilling the above conditions, is permitted to enter the
system, in order to load near misses data and consult
the database, and have the guarantee of data security
and confidentiality.
4

near miss

reporting phase

consultation phase
lesson
learned

Figure 1.

Concept of the software system.

of the near misses, and the second is the consultation

phase, for the extraction of the lessons learned from the
database, both described in the following paragraphs.
4.1 Near misses reporting phase
This part contains all data regarding the near miss and
is divided into two sections: the first contains all the
information about the specific event (time, place, substances involved etc.); the second provides the causes,
the post accident measures put into place, and the
lessons learned. It is worth noting that, in order to
encourage use of the database, in the selection of fields
to be filled in, a compromise was reached between the
need to collect as much information as possible, so
as to enrich the database and the need to simplify the
completion of the form by the user.
These forms are filled in on a voluntary basis, in
fact. The contents of the two sections are described
in paragraph 4.1.1 and 4.1.2 respectively. A printout illustrating the fields contained in section 1 is
presented in Fig. 2, in which the fields and the
explanations are shown in Italian.
4.1.1 Section 1
Geographical location: Municipality and province
where the near miss occurred.
Date: The date when the near miss took place.
Time: The time when the near miss took place.
Seveso classification: Upper or lower tier establishments.
Location: The location where the near miss occurred,
either inside the production units, or in the logistic
units, internal or external.

DATABASE STRUCTURE

On the basis of the above defined attributes, which

are assigned to the system, the software platform supporting the near misses database has been created and
performs two main functions, as illustrated in Fig. 1:
The first is the reporting phase, for the description

29

Figure 2.

Section 1 of the near misses database.

Causes: The analysis of causes, including equipment

failure, procedural deficiencies or human factors.
Each near miss may be triggered by more than
one cause. The operational status of the establishment (normal operation, shut down, restart,
maintenance), when the near miss took place must
be specified.
Damages: The cost of damages provoked by the near
miss.
Corrective actions: Corrective actions that have been
put in place after the near miss in order to avoid a
repetition of the event; these actions can be immediate or delayed, and if delayed, can be further divided
into technical, procedural or training.
Lessons learned: Descriptions of what we have learned
from the near miss and what improvements have
been introduced as a consequence.
Annex: It is possible to complete the record of any near
miss, adding files in pdf, gif o jpeg format.

Area: Area of establishment in which the near miss

took place, including production, storage, services,
utilities or transport units.
Unit: Depending on the area selected, unit directly
involved in the near miss.
Accident description: The description of the accident
as it was given by the compiler.
Substances: The substance or the substances involved
in the near miss.
4.1.2 Section 2
Type of event: The types of events which occurred after
a near miss; this is a multi-check field allowing the
user to select more options simultaneously.
Potential danger: Dangers that could have stemmed
from a near miss and that could have resulted in
more serious consequences if the circumstances had
been slightly different.
Safety measures in place: The kind of safety measures
which are in place, if any, specifying if they have
been activated and, if this is the case, if they are
adequate.

Once this descriptive part is completed, a classification code will be attributed to any near miss, in order
that identification is univocal.

30

Figure 3.

4.2

Near misses search criteria.

thus supporting effective training activities. In this

initial phase of the project the first priority has been the
collection of as much near misses data as possible. In
order to reach this goal, the use of the database has been
encouraged among the partners of Federchimica, who
are, in fact the main contributors of the data entered.
To publicize the use of the software system both the
institutions responsible for the project, ISPESL and
Federchimica, have organized a number of workshops
in different Italian cities to illustrate the performances
and potentialities of the database. The feedback from
the industries with respect to the initiative promoted
by ISPESL and Federchimica can be considered, on
the whole, positive, at least for those bigger companies which already dedicate economic and human
resources to the industrial safety; further efforts are
required to motivate also the smaller enterprises to use
the database.
The system is currently in a running in phase,
which will provide an in-depth evaluation of its performance and identify potential improvements. This
phase involves the analysis of the (approximately) 70
near misses which are now stored in the database.
A preliminary analysis of the first data collected shows
that, for any near misses reported, all fields contained
in the database have been filled; this fact represents a
good starting point for an assessment on the quality of
the technical and managerial information gathered, as
well as on the reported lessons learned, which will be
carried out in the next future. Further distribution to the
competent authorities and to other industrial sectors
will be considered in a second stage.

Near misses consultation phase

As specified at the beginning of the previous paragraph, the second part of the software platform
concerns the extraction of lessons learned by the consultation of the database contents. In fact, one of the
main objectives of the database is the transfer of technical and managerial corrective actions throughout the
chemical process industry. The database has therefore
been supplied with a search engine which aids navigation through near miss reports, allowing several
options. First, the user is able to visualize all near
misses collected in the database in a summary table,
illustrating the most important fields, that are, respectively, event code, submitter data, event description,
damages, immediate or delayed corrective actions,
lessons learned, annex; second, the details of a specific
event can be selected by clicking on the corresponding code. It is also possible to visualize any additional
documents, within the record of the specific event, by
clicking on Annex. Third, to extract a specific near
miss from the database, on the basis of one or more
search criteria, a query system has been included in
the software utilities. This is done by typing a keyword, relevant, for example, to the location, unit,
year of occurrence, potential danger, etc, as shown
in Fig. 3.

RESULTS AND DISCUSSION

Thanks to these functions, the analysis of the database

contents can provide company management with
the information required to identify weaknesses in the
safety of the industrial facilities, study corrective
actions performed to prevent the occurrence of accidents, prioritize preventive and/or mitigating measures
needed, and better understand the danger of specific
situations. Another important function is the transfer
of knowledge and expertise to a younger workforce,

CONCLUDING REMARKS

A software system for the collection, storage and analysis of near misses in the Italian chemical industry has
been developed for multiple purposes. It allows information regarding the different factors involved in a

31

training activities implemented in the chemical

industry; they should be for all personnel involved
in the use of the software system. These training
activities, to which both ISPESL and Federchimica
may vigorously contribute, are expected to lead to an
enhancement in the quality of the retrieval of the underlying causes of near misses. They should also lead to
an increase in the reliability of data stored as well as
of lessons learned extracted from the database.

near miss event to be shared and information regarding

lessons learned among all interested stakeholders to
be disseminated in a confidential manner. The system which has been developed attempts to fill a gap
which exists in the wide range of databases devoted to
process industry accidents. The system also can fulfil
the legal requirements for the implementation of the
Safety Management System; in fact, it offers industry management the possibility of verifying its safety
performances by analyzing near miss results. Its userfriendly character can stimulate the reporting of near
misses as well as the sharing of lessons learned, supporting industry management in the enhancement of
safety performances. It is the intention of the authors to
continue the work undertaken, creating a wider range
of functions and keeping the database up to date. From
a preliminary analysis carried out on the data collected,
we realized that, in the near future, a number of aspects
regarding the management of near misses will need
to be investigated in more depth. For example, the
screening criteria for the selection of near misses must
be defined. These criteria are necessary for identifying
near misses of particular technical interest for preventing major accidents and limiting their consequences,
as required by the Seveso legislation. Another element
which requires in-depth analysis is the root causes of
near misses, and the corrective measures needed. All
the above mentioned issues can benefit from specific

REFERENCES
Center for Chemical Process Safety, 2003. Investigating
Chemical Process Incidents. New York, American Institute of Chemical Engineers.
European, Council 1997. Council Directive 96/82/EC on
the major accident hazard of certain industrial activities (Seveso II). Official Journal of the European
Communities. Luxembourg.
Jones, S., Kirchsteiger, C. & Bjierke, W. 1999. The importance of near miss reporting to further improve safety
performance. Journal of Loss prevention in the Process
Industries, 12, 5967.
Legislative Decree 17 August 1999, n. 224 on the control of
major accident hazards involving dangerous substances.
Gazzetta Ufficiale n. 228, 28 September 1999 Italy.
Philley, J., Pearson, K. & Sepeda, A. 2003. Updated CCPS
Investigation Guidelines book, Journal of Hazardous
Materials, 104, 137147.

32

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Development of incident report analysis system

based on m-SHEL ontology
Yoshikazu Asada, Taro Kanno & Kazuo Furuta
Department of Quantum Engineering and Systems Science, The University of Tokyo, Japan

ABSTRACT: As increase of complex systems, simple mistakes or failures may cause serious accidents. One
of measures against this situation is to understand the mechanism of accidents and to use the knowledge for
accident prevention. However, analyzing incident reports is not kept up with the pace of their accumulation
at present, and a database of incident reports is utilized insufficiently. In this research, an analysis system of
incident reports is to be developed based on the m-SHEL ontology. This system is able to process incident reports
and to obtain knowledge relevant for accident prevention efficiently.

INTRODUCTION

Recently, the number of complex systems, such as

nuclear systems, air traffic control systems or medical
treatment systems, is increasing with progress in science and technology. This change makes our life easy,
but it makes the causes of system failures differently.
With advanced technology, simple mistakes or incidents may cause serious accidents. In other words, it
is important to understand the mechanism of incidents
for avoiding major accidents.
Small accidents or mistakes that occur behind a
huge accident are called incidents. Heinrich found
the relationship between huge accidents and incidents
(Heinrich, 1980). His study showed that there are 29
small accidents with light injury and 300 near misses
behind one big accident with seriously-injured people
(Figure 1). Bird also found the importance of nearmisses or incidents for accident prevention (Bird &
George, 1969).
Of course this ratio will change among different situations, but the important point is the fact that there
are a lot of non-safety conditions behind a serious
problem.
For this reason, to reduce the number of incidents also contributes to decrease the number of major
accidents. Nowadays many companies are operating
an incident reporting system. NUClear Information
Archives (NUCIA), a good example of such a system,
is a web site operated by the Japan Nuclear Technology
Institute [3]. In this site, lots of information such as

Figure 1.

Heinrichs law.

inspection results or incident reports in nuclear power

plants is available to the public.
This is an example of good practice for safety
management. However, though incident reports are
collected, they are not always analyzed rightly.
There are several reasons, but main reasons are as
follows:
1. Insufficient expertise for incident analysis
2. Inconsistent quality of incident reports
Even though a lot of incident reports are compiled, in most cases they are only stored in the
database. It is important to solve this problem by

33

analyzing incident reports using advanced information

processing technologies.
In this study, a conceptual design of an analysis
system of incident reports is proposed, and its verification is conducted experimentally. We adopted nuclear
industry as a particular application domain because of
availability of NUCIA.
2

PROBLEMS OF INCIDENT REPORTS

Figure 2.

A number of incident reporting systems are now operated in various areas, but there are some problems. The
problems are twofold.
The first problem is how to make the report.
There are two types of reporting schemes: multiplechoice and freestyle description. Multiple-choice is
easy to report, because the reporter only have to
choose one or a few items from alternatives already
given. It saves the time of reporting. For this reason, many of existing reporting systems use this
type. But, some important information may be lost
with this reporting scheme. The purpose of incident
reporting is to analyze the cause of an incident and
to share the obtained knowledge. If the reports are
ambiguous, it is hard to know the real causes of an
incident.
Second problem is about the error of the contents.
An estimation of the cause or investigation of preventative measures is the very important part of the
report. However, there are sometimes mistake of analysis. Of course, these parts are complicated, but exact
description must be needed.
It is not easy to get rid of these problems, and in
this study it is assumed that no incident reports are free
from these problems. The influence of these problems
will be discussed later.

3
3.1

Diagram of m-SHEL model.

interested person, and the bottom one is other people in

relation with the central person. This model is shown
in Figure 2.
In Figure 2, there are some gaps between two elements. This means that mismatches between different
elements may cause an accident. For example, misunderstanding of the operation manual may happen
in the gap between S and L. To fill the gaps, the
interface that connects different elements plays an
important role.

3.2 COCOM
There are many incidents caused by human behavior,
such as carelessness or rule violation. For this reason,
the liveware in the m-SHEL model must be focused on.
In this study, the COCOM model is applied to incident
analysis.
The COCOM model was proposed by E. Hollnagel
(Hollnagel, 1993). Hollnagel categorized the state
of human consciousness into the following four
classes:

ANALYSIS OF HUMAN FACTORS

m-SHEL model

For analysis of accidents, it is necessary to break

the whole view of an incident down into elementary
factors, such as troubles in devices, lack of mutual
understanding among staffs, or misunderstanding of
an operator. Some frameworks for accident analysis
have been proposed such as m-SHEL or 4M-4E. They
are originally proposed for risk management, and they
can support understanding incidents.
The m-SHEL model was proposed by Kawano,
et al. (Kawano, 2002). This model classifies accident factors into five elements of a system. They
are m (management), S (software), H (hardware),
E (environment), and L (liveware). In the m-SHEL
model, there are two livewares. The central one is the

Strategic control (high attention)

Tactical control
Opportunistic control
Scrambled control (low attention)

According to this model, human consciousness

changes by the attention level to the task. When workers have high attention to the task, they know well
how to do the task, and there are no environmental
causes of an accident. They are doing the task in a
strategic way.
Once some problem happens, however, the situation will change. Workers may lose their countenance,
and sometimes they become scrambled. Under such
situation, it is difficult for them to be able to do the
task in a strategic or tactical manner.
By focusing on human consciousness based on
COCOM, it is able to make a short list of incident
causes. It is useful both for causal analysis and for risk

34

prediction of a similar incident by checking specific

factors.
3.3

2. Hierarchical structure by is-a link between two

concepts, like BMW is a car or dog is a
mammal.
3. Some relationships between different concepts
other than is-a link, like car has an engine.
4. Axiom that defines or constrains concepts or
links.

Nuclear Information Archives (NUCIA)

Here, let us look into an example of an incident report

database.
In NUCIA, lots of information such as inspection
results or incident reports in nuclear power plants are
available to the public. Disclosure of the information
is conducted under the following ideas.

In this study, ontology is defined for analysis of

incidents occurred in nuclear systems as follows:
1. A set of concepts necessary for analysis of incidents
in nuclear power plants.
2. Hierarchical structure. In nuclear engineering, the
structure of a nuclear power plant and the organizational framework are described.
3. Relationships. Causality information is defined in
this part.
4. Definition of axioms. Synonyms are also defined.

Various opinions not only from electric power suppliers but also from industry-government-academia
communities are useful to solve problems in nuclear
safety.
Increasing transparency to the society contributes
to obtain trust from the public.

An ontology-based system is different from knowledge base systems, expert systems or artificial intelligence. These traditional ways have some problems.
They are:

Collected information, however, is not utilized

enough at present. There are several reasons as
follows.
The amount of data is enough, but no effective
methods of analysis are available.
While some reports are detailed, others are not.
There is little uniformity in description method.

Little knowledge is usable in common between

different fields.
New information cannot be added to the existent
knowledge base easily.

Under such circumstances, even though there is a

lot of useful information in the database, analysis and
application are not kept up with accumulation of data.
3.4

Compared with knowledge based system, ontology

is meta-level knowledge. In this research, the incident ontology contains not only detailed knowledge
or case examples (which are traditional knowledgebase), but also association of incidents or causality
information (which are meta-level information). For
this reason, analysis by ontology based system is more
multidirectional.

Ontology

Ontology here means relations between concepts like

synonyms or sub concepts. Mizoguchi defines ontology as a theory of vocabulary/concepts used for building artificial systems (Mizoguchi, Kozaki, Sano &
Kitamura). Figure 3 shows an example of ontology.
There are four elements in ontology. They are:

3.5 m-SHEL ontology

There are some previous works for computer based
analysis. Chris was proposed case based retrieval
with Semantic Network (Johnson 2003). In his
research, network were made by basic verbal phrases,
such as is_a, makes or resolved_by. David did
document synthesis with techniques like XML or
Semantic Web systems (Cavalcanti & Robertson
2003).
In this study, an m-SHEL ontology is developed
using XML (eXtensive Markup Language) (Tim et al.
1998). The m-SHEL ontology is based on incident
reports of NUCIA. In this ontology, each element has
at least one m-SHEL attribution, such as Software or
Environment. All of construction processes, like tagging of attribution or making a hierarchical structure
were done manually. There is an example of XML
description of a concept contained in the m-SHEL
ontology.

1. A conceptual set which consists of elementally

concepts of the intended field.

Figure 3.

An example of ontology.

35

<concept>
<label>
(Information about concept.
This label is for readability.)
</label>
<ont-id>
143
</ont-id>
<mSHEL>
H
</mSHEL>
<description>
(Information about what is this
concept.
This is used for searching)
</description>
<cause-id>
217
(id of the concept which cause No. 143)
</cause-id>
<effect-id>
314
(id of the concept which is caused by No. 143)
</effect-id>
</concept>
Figure 4.

The m-SHEL ontology has not only information of

a hierarchical structure of elements, but also of causal
association. Causality is defined by a link orthogonal
to the concept hierarchy.
Causality is described by XML as follows:

A part of m-SHEL ontology.

There is a weighing factor in the causality. It is

defined by how many cases are there which have the
same cause in the reports. This weighing factor is used
to decide which causality leads to the most or the least
likely outcome.
Figure 4 shows a part of the m-SHEL ontology.

<causality>
<label>
(Information about causality.
This label is for readability.)
</label>
<causality-id>
C-13
</causality-id>
<cause-id>
147
(id of the element which cause C-13)
</cause-id>
<effect-id>
258
(id of the element which is caused by C-13)
</effect-id>
<weight>
0.3 (weighing factor)
</weight>
</causality>

ANALYSIS SYSTEM

After having made the m-SHEL ontology, an analysis system was developed. The flow of this system is
shown below:
1. Input an incident data in a text form. This step can
be done both manually and by importing from XML
style files.
2. In Japanese language, since no spaces are placed
between words, morphological analysis is done to
the input text. It is explained in Chapter 4.1.
3. Keywords from the text are selected and mapped
onto the ontology. Keywords are the words included
in the m-SHEL ontology.
4. Required information of the incident is structurized both in the concept hierarchy and causality
relations.

In this example, causality labeled 13 is caused by

the concept No.147, and it causes the concept No. 258.
These concept numbers are the same as the id of each
element of ontology.

Figure 5 shows an example of mapping result. For

visualization, only specific elements related to the
incident are shown here.

36

Figure 5.

An example of mapping result.

Visual output function is not implemented in the

current system. Only output in an XML format is available now. Figure 5 was generated by another diagram
drawing software.
4.1

Morphological analysis

Japanese language is different from English. In

Japanese, no spaces are inserted between two words. If
English sentence is written like Japanese, it is written
as follows:

Figure 6.

Fraction of compound nouns detected.

This is a sample text. (a)

In the m-SHEL ontology, it is necessary to separate
each word before applying text mining methods. For
this step, Mecab (Yet Another Part-of-Speech and
Morphological Analyzer) is used [6]. Mecab is able
to convert Japanese style sentences like (a) into to
English style sentences like (b).

From these points, a following rule is added: continuous Kanji nouns are combined to form a single
technical word. This is very simple change, but around
60% of keywords can be found by this rule. Figure 4
shows the percentage of compound nouns detected for
ten arbitrary sample cases of incident reports.

This is a sample text. (b)

Mecab has an original dictionary for morphological
analysis. The default dictionary does not have many
technical terms. If no improvements have been made,
analysis by Mecab does not work well for incident
reports, which include a lot of technical words or
uncommon words.
To solve this problem, some change has been made
to the dictionary and the algorithm of morphological
analysis.
Firstly, some terms used in nuclear engineering
or human factor analysis were added. This way of
improvement is most simple and reliable, but there still
remain some problems. Even though the fields of text
which is dealt with is specialized (in this case, nuclear
engineering), there are so many necessary words. It is
hard to add all of these terms.
Moreover, this improvement is required every time
as the field of application is to be changed, i.e.,
important technical words are different in different
application fields.
Secondly, the algorithm to extract technical words
has been adjusted. In Japanese, many technical terms
are compound nouns, which are formed with more
than two nouns. From a grammatical viewpoint, two
nouns are rarely appeared continuously. Moreover,
Japanese Kanji is often used in those words.

4.2 Deletion of unnecessary parts

Another change to the algorithm is for deletion of
unnecessarypartsofreports. Incidentreportssometimes
contain information additional to causal analysis.
Exampleofsuchadditionalinformationisonhorizontal
development, which is to develop lessons learned from
an incident horizontally throughout the company or the
industry. Additional information may sometimes work
as noises for causal analysis; it is eliminated therefore
fromreportsbeforeprocessing, ifthepurposeofanalysis
does not include horizontal development.
5

VERIFICATION EXPERIMENT

In this chapter, test analysis of incident reports has

been done using the proposed system. Sample reports
were from the NUCIA database, which is open to
the public. Verification experiment was done by the
following steps:
1. Before making the m-SHEL ontology, some incident reports are set aside. These reports are used
for verification.
2. Some reports were analyzed with the system automatically.
3. An expert analyzed the same reports manually to
check the adequacy of analysis by the system.

37

Table 1.

m
S
H
E
L
c

Another reason for the high ratio is a trend of reporting incidents in nuclear industry of Japan that incidents
are reported from a viewpoint of failure mechanism of
hardware rather than human factors.
This trend also resulted in the low presence of Environment or management items. Not only the result of
the system, but also that of the expert marked low
counts for these items. It means that the low scores are
attributable not to the system but to some problems in
the original reports.
On the other hand, the results for Liveware and
causal association are different from those for Environment and management. The results of automatic
analysis for Liveware and causal association also
marked low scores. However, the expert did not mark
so low as the system. It seems this outcome is caused
by some defects in the m-SHEL ontology.

Result of verification test.

Case
1

Case
2

Case
3

Case
4

Case
5

Case
6

Case
7

0/0
2/6
7/18
0/1
2/4
1/4

1/2
3/7
5/12
0/0
1/5
0/3

0/1
5/6
6/13
0/0
2/4
2/4

4/7
2/4
5/12
2/2
2/5
3/6

2/2
1/3
9/10
0/1
3/4
4/5

0/1
1/2
0/8
3/3
3/5
0/2

2/4
4/7
5/7
0/1
4/3
1/2

Number of causal association.

4. Two results were compared. At this step, we

focused how many elements and causality relations
are extracted automatically.
In this study, seven reports were used for verification.
The incidents analyzed are briefly described below:

automatic trip of the turbine generator due to

breakdown of the excitation equipment,
automatic trip of the reactor due to high neutron flux
of the Intermediate Range Monitor (IRM),
impaired power generation due to switching over of
the Reactor Feed Pump (RFP), and
manual trip of the reactor because of degradation of
vacuum in a condenser.

An analysis system of incident reports has been developed for nuclear power plants. The method of analysis
adopted is based on the m-SHEL ontology. The result
of automatic analysis failed to mark high scores in
assessment, but it is partly because of the contents
of the original incident report data. Though there is
a room for improvement of the m-SHEL ontology,
the system is useful to process incident reports and
to obtain knowledge useful for accident prevention.

Result of the test is shown in Table 1. In this table,

each number shows how many concepts contained in
the ontology were identified in the input report. The
numbers on the left represent the result obtained by the
system and those on the right the result by the expert.
For example, 3/7 means three concepts were identified
by the system, and seven by the expert.
It should be noted here that not the numbers themselves are important, but coincidence of the both
numbers is significant for judging validity of analysis by the system. In this research, since the algorithm
of the system is quite simply designed it is reasonable
to think that analysis by the expert is the reference. If
the system identified more concepts than the expert,
it is probable that the system picked up some noises
rather than the system could analyze more accurately
than the expert.
6

CONCLUSION

REFERENCES
H.W. Heinrich. 1980. Industrial accident prevention: A safety
management approach, McGraw-Hill.
Frank E. Bird Jr. & George L. Germain. 1969. Practical Loss
Control Leadership, Intl Loss Control Inst.
NUCIA: Nuclear Information Archives, http://www.nucia.jp
R. Kawano. 2002. Medical Human Factor Topics. http://
www.medicalsaga.ne.jp/tepsys/MHFT_topics0103.html
E. Hollnagel. 1993. Human reliability analysis: Context and
control. Academic Press.
Riichiro Mizoguchi, Kouji Kozaki, Toshinobu Sano & Yoshinobu Kitamura. 2000. Construction and Deployment of a
Plant Ontology. Proc. of the 12th International Conference Knowledge Engineering and Knowledge Management (EKAW2000). 113128.
C.W. Johnson. 2003. Failure in Safety-Critical Systems:
A Handbook of Accident and Incident Reporting,
University of Glasgow Press, http://www.dcs.gla.ac.uk/
johnson/book/
J. Cavalcanti & D. Robertson. 2003. Web Site Synthesis based
on Computational Logic. Knowledge and Information
Systems Journal, 5(3):263287.
MeCab: Yet Another Part-of-Speech and Morphological
Analyzer, http://mecab.sourceforge.net/
Tim Bray, Jean Paoli, C.M. Sperberg-McQueen (ed.), 1998.
Extensible Markup Language (XML) 1.0: W3C Recommendation 10-Feb-1998. W3C.

DISCUSSION

In Table 1, Hardware elements are relatively well

extracted. Plenty of information related to some parts
of a power plant, such as CR (control rod) or coolant
water, are included in incident reports. These words
are all categorized as Hardware items, so the number
of Hardware items is larger than others. This is the
reason why Hardware has a high ratio of appearance.
A similar tendency is shown with Software elements.

38

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Forklifts overturn incidents and prevention in Taiwan

K.Y. Chen & Sheng-Hung Wu
Graduate School of Engineering Science and Technology, National Yunlin University of Science and Technology,
Douliou, Yunlin, Taiwan, ROC

Chi-Min Shu
Process Safety and Disaster Prevention Laboratory, Department of Safety, Health, and Environmental
Engineering, National Yunlin University of Science and Technology, Douliou, Yunlin, Taiwan, ROC

ABSTRACT: Forklifts are so maneuverable that they can move almost everywhere. With stack board, forklifts also have the capability of loading, unloading, lifting and transporting materials. Forklifts are not only
widely used in various fields and regions, but are common in industry for materials handling. Because they are
used frequently, for any incidents such as incorrect forklift structures, inadequate maintenance, poor working
conditions, the wrong operations by forklift operators, and so on, may result in property damages and casualties. The forklifts, for example, may be operated (1) over speed, (2) in reverse or rotating, (3) overloaded, (4)
lifting a worker, (5) on an inclined road, or (6) with obscured vision and so on, which may result in overturning, crushing the operator, hitting pedestrian workers, causing loads to collapse, lifting a worker high to fall
and so on. The abovementioned, therefore, will result in adjacent labor accidents and the loss of property.
According to the significant professional disaster statistical data of the Council of Labor Affairs, Executive
Yuan, Taiwan, approximately 10 laborers perish in Taiwan due to forklift accidents annually. This obviously
shows that forklift risk is extremely high. If the operational site, the operator, the forklifts and the work environment are not able to meet the safety criterion, it can possibly cause a labor accident. As far as loss prevention
is concerned, care should be taken to guard handling, especially for forklift operation. This study provides
some methods for field applications, in order to prevent forklift overturn accidents and any related casualties
as well.
1

1.1

INTRODUCTION

Being struck by forklifts

The driver was obscured by goods piled too high, driving, reversing or turning the forklift too fast, forgetting
to use alarm device, turn signal, head lamp, rear lamp
or other signals while driving or reversing, not noticed
by pedestrian workers or cyclists, or disturbed by the
work environment such as corner, exit and entrance,
shortage of illumination, noise, rain, etc., causing
laborers to be struck by forklifts.

According to the significant occupational disaster statistical data of the Council of Labor Affairs (CLA),
Executive Yuan, Taiwan, forklifts cause approximately
10 fatalities annually from 1997 to 2007, as listed
in Fig. 1 (http://www.iosh.gov.tw/, 2008). The manufacturing industry, the transportation, warehousing
and communication industry, construction industry
are the three most common sectors sharing the highest occupational accidents of forklifts in Taiwan, as
listed in Fig. 2. In addition, the forklift accidents were
ascribed to being struck by forklifts, falling off a forklift, being crushed by a forklift and overturning, in
order. Similar to the normal cases in the USA, forklift
overturns are the leading cause of fatalities involving
forklift accidents. They represent about 25% of all
forkliftrelated deaths (NIOSH alert, 2001). Figure 3
reveals that the forklift occupational accidents could
be classified into five main occupational accident
types, explained as follows (Collins et al., 1999; Yang,
2006).

1.2

Falling from forklifts

The forklift driver lifted the worker high to work by

truck forks, the stack board or other materials but not
established a work table, used the safety belt, safe
guard equipment and so on, causing the worker fall
off the forklifts easily.
1.3

Collapsing and crushed

The material handling was not suitable so that the

loads were piled up too high to stabilize well. The mast

39

Year
2007

include collisions with workplace materials, resulting

in the collapse of the materials and wounding operators
nearby.

6
7

2006

10

2005

13

2004
2003

1.4

10

2002

4
16

2000
1999

10

1998

1997
0

Because of the high speed of the forklifts reverse or

rotation, or because of the ascending, the descending, uneven ground, wet slippery ground, soft ground,
overly high lifted truck forks, or overloads, the forklifts
overturned to crush the operator.

2001

12
2

10

12

14

1.5

16

Number of fatalities

Others

Construction

10

Transportation,
warehousing and
communication

27

Manufacturing
0

60
10

20

30

40

50

60

Figure 2. Fatalities from forklift accidents are distributed

in different industries, Taiwan, from 1997 to 2007.

Struck
37

2.1

Number of deaths

30

20
15

Collapsing
Falling
and
Crushed Overturn
17
16
15

OVERTURN CASE REPORT

Case one

On December 4, 2005 ca. 1 p.m., a 34-year-old laborer

who drove a forklift to discharge mineral water was
crushed to death by the forklifts. The laborer reversed
the forklift into a trailer on a board. The forklift slipped
because the wheel deviated from the board. It, then,
was turned over, crushing the driver to death, as shown
in Fig. 4.

35

25

Becoming stuck or pinned

Getting stuck or pinned between the truck forks and

mast or the tires while repairing or maintaining the
forklifts. In other instances, the operator forgot to turn
the forklift off in advance and adjust the goods on
the truck fork, or got off and stood in front of the
forklift and next to the forklifts dashboard to adjust
the goods. Therefore, when the operator returned to
the driver seat, he (or she) touched the mast operating
lever carelessly, causing the mast backward, and got
his or her head or chest stuck between the mast and
overhead guard.
This study focused on reporting forklifts overturn
accidents and preventing them from occurring. Two
forklifts overturn accidents were described as below
(http://www.iosh.gov.tw/, 2008; http://www.cla.gov.
tw/, 2008).

Figure 1. Fatalities of forklift accidents in Taiwan, from

1997 to 2007.

40

Collapsing and crushed

Stuck
and
Pinned
11

10
5

Hitting
4

2.2
Others
2

Case two

Figure 5 shows that on September 27, 2005 about

11:00 am, an 18-year-old laborer had engaged in
forklift operation, but accidentally hit against the connecting rod of the shelf while driving, and then the
forklift overturned, crushing him, and he passed away.
The laborer operated the forklift to lift the cargo up
to a shelf. After laying aside the cargo 7 meters
high, the forklift left the shelf. The forklift moved
without the truck fork being lowered. The forklifts
mast accidentally hit against the connecting rod and
inclined. The operator was shocked by the disaster

Figure 3. The number of deaths in forkliftaccidents is

divided into different types, Taiwan, from 1997 to 2007.

inclined while the forklift operator was driving so a

worker stood on the forklift to assist material handling by holding the goods with hands and so on,
which easily made the goods collapse and crushed the
assistant worker or pedestrians nearby; other disasters

40

and escaped quickly from the drivers seat to the warehouse entrance. However, the forklift was unable to
stop, decelerate its movement, but was still moving
forward to the warehouse entrance. After the forklift
mast hit against the connecting rod, its center of gravity was changed to the driver seat side. The forklift
reversed in the direction of the warehouse entrance.
The laborer was hit and killed. A similar accident of a
forklift overturning and tipping is pictured in Fig. 6.
3

Figure 4.

The CLA has promulgated the rule of labor safety and

health facilities (CLA, 2007), and the machine tool
protection standard (CLA, 2001), which became effective and last modified in 2007 and 2001 for forklift
equipment, measures and maintenance, respectively.
The CLA also has the Rules of Labor Safety Health
Organization Management and Automatic Inspection
(CLA, 2002) for forklifts inspection. The standard
regulates operator training and licensing as well as
periodic evaluations of operator performance. The
standard also addresses specific training requirements for forklift operations, loading, seat belts,
overhead protective structures, alarms, and maintenance. Refresher training is required if the operator is
(a) observed operating the truck in an unsafe manner,
(b) involved in an accident or mistake, or (c) assigned
a different type of forklift.

Forklift overturns from the ladder.

3.1 Forklift equipment, measure and maintenance

Figure 5. Forklift overturns due to hitting against the

connecting rod of the cargo.

Figure 6.

RELATED REGULATIONS

a. The employer should have safe protective equipment for the moving forklifts and its setup should
be carried out according to the provision of the
machine apparatus protection standard.
b. The employer should make sure that the forklifts
cannot carry the laborer by the pallet or skid of
goods on the forks of forklifts or other part of forklifts (right outside the driver seat), and the driver or
relevant personnel should be responsible for it. But
forklifts those have been stopped, or have equipment or measures to keep laborers from falling, are
not subject to the restriction.
c. The pallet or skid at fork of forklifts, used by the
employer, should be able to carry weight of the
goods.
d. The employer should assure that the forks etc. are
placed on the ground and the power of the forklift
is shut off when the forklift operator alights.
e. The employer should not use a forklift without placing the back racket. The mast inclines and goods
fall off the forklift but do not endanger a laborer,
which is not subjected to the restriction.
f. The employer should have necessary safety and
health equipment and measures when the employee

Forklift overturning and tipping due to overloads.

41

which are supplied by all equipment manufacturers

and described the safe operation and maintenance of
forklifts.

uses a forklift in the place with dangerous

goods.
g. The employer can not exceed the biggest load that
forklifts can bear for the operation, and it transports of the goods should keep a firmness status
and prevent to turn over.
3.2

CONCLUSION AND RECOMMENDATIONS

The statistics on fatalities indicate that the three most

common forkliftrelated fatalities involve overturns,
workers being struck or crushed by forklifts, and workers falling from forklifts. The case studies of overturns
indicate that the forklifts, the operating environment,
and the actions of the operator all contribute to fatal
incidents related to forklifts. Furthermore, these fatalities reveal that many workers and employers are not
using, or even may be unaware of, safety procedures,
not to mention the proper use of forklifts to reduce
the risk of injury and death. Reducing the risk of
forklift incidents requires a safe work environment, a
sound forklift, secure work practices, comprehensive
worker training, and systematic traffic management.
This study recommends that employers and workers comply with national regulations and consensus
standards, and take the following measures to prevent injury when operating or working near forklifts,
especially avoiding overturn accidents.

Forklift operation and inspection

Under national regulations, CLA requirements for

forklifts operation and inspection are as follows:
a. The employer should assign someone who has gone
through the special safety and health education and
personnels operation training to operate a forklift with over one metric ton load. The employer
ordering personnel to operate more than one metric ton load forklift should assure that she or he
had accepted 18 hours special training courses for
safety and health.
b. The employer should periodically check for the
whole of the machine once every year to the
forklifts. The employer should carefully and regularly inspected and maintained the brakes, steering
mechanisms, control mechanisms, the oil pressure
equipment, warning devices, lights, governors,
overload devices, guard, back racket and safety
devices, lift and tilt mechanisms, articulating axle
stops, and frame members should be in a safe condition, at least periodically checked once monthly.
c. Based on the rule of labor safe health organization management and automatic inspection (CLA,
2002), CLA requires that industrial forklifts must
be inspected before being placed in service. They
should not be placed in service if the examination
shows any condition adversely affecting the safety
of the vehicle. Such examination should be made
at least daily. When defects are found, they should
be immediately reported and corrected.
d. Under all travel conditions, the forklift should be
operated at a speed that will permit it to be brought
safely to a stop.
e. The operator should slow down and sound the horn
at cross aisles and other locations where vision is
obstructed.
f. Unauthorized personnel should not be permitted to
ride on forklifts. A safe place to ride should be
provided and authorized.
g. An operator should avoid turning, if possible, and
should be alert on grades, ramps, or inclines. Normally, the operator should travel straight up and
down.
h. The operator of a forklift should stay with the truck
if it tips over. The operator should hold on firmly
and lean away from the point of impact.

4.1 Worker trainingemployer event

a. Ensure that a worker does not operate a forklift
unless she or he has been trained and licensed.
b. Develop, implement, and enforce a comprehensive written safety program that includes worker
training, operator licensing, and a timetable for
reviewing and revising the program. A comprehensive training program is important for preventing
injury and death. Operator training should address
factors that affect the stability of a forkliftsuch
as the weight and symmetry of the load, the speed
at which the forklift is traveling, operating surface,
tire pressure, driving behavior, and the like.
c. Train operators to handle asymmetrical loads when
their work includes such an activity.
d. Inform laborers that when the forklift overturns
jumping out the cockpit could possibly lead to being
crushed by the mast, the roof board, the back racket,
the guard rail and the fuselage equipment. In an
overturn, the laborer should remain in the cockpit,
and incline his or her body in the reverse direction
to which the forklift will turn over.
4.2 Forklift inspection, maintenance
and liftingemployer event
a. Establish a vehicle inspection and maintenance
program.
b. Ensure that operators use only an approved lifting cage and adhere to general safety practices for

In addition to the above regulations, employers

and workers should follow operators manuals,

42

c. Do not jump from an overturning forklift. Stay

there, hold on firmly and lean in the opposite
direction of the overturn, if a lateral tip over occurs.
d. Use extreme caution on grades, ramps, or inclines.
In general, the operator should travel only straight
up and down.
e. Do not raise or lower the forks while the forklift is
moving.
f. Do not handle loads that are heavier than the weight
capacity of the forklift.
g. Operate the forklift at a speed that will permit it to
be stopped safely.
h. Look toward the path of travel and keep a clear view
of it.
i. Do not allow passengers to ride on a forklift unless
a seat is provided.
j. When dismounting from a forklift, always set the
parking brake, lower the forks, and turn the power
off to neutralize the controls.
k. Do not use a forklift to elevate a worker who is
standing on the forks.
l. Whenever a truck is used to elevate personnel,
secure the elevating platform to the lifting carriage
or forks of the forklift.

elevating personnel with a forklift. Also, secure the

platform to the lifting carriage or forks.
c. Provide means for personnel on the platform to shut
power off whenever the forklift is equipped with
vertical only or vertical and horizontal controls for
lifting personnel.
d. When work is being performed from an elevated
platform, a restraining means such as rails, chains,
and so on, should be in place, or a safety belt with
lanyard or deceleration device should be worn by
the person on the platform.
4.3

Workers near forkliftsemployer event

a. Separate forklift traffic from other workers where

possible.
b. Limit some aisles to workers either on foot or by
forklifts.
c. Restrict the use of forklifts near time clocks, break
rooms, cafeterias, and main exits, particularly when
the flow of workers on foot is at a peak (such as at
the end of a shift or during breaks).
d. Install physical barriers where practical to ensure
that workstations are isolated from aisles traveled
by forklifts.
e. Evaluate intersections and other blind corners to
determine whether overhead dome mirrors could
improve the visibility of forklift operators or workers on foot.
f. Make every effort to alert workers when a forklift
is nearby. Use horns, audible backup alarms, and
flashing lights to warn workers and other forklift
operators in the area. Flashing lights are especially
important in areas where the ambient noise level
is high.
4.4

ACKNOWLEDGMENTS
The authors are deeply grateful to the Institute of Occupational Safety and Health, Council of Labor Affairs,
Executive Yuan, Taiwan, for supplying related data.
REFERENCES
http://www.iosh.gov.tw/frame.htm, 2008; http://www.cla.gov.
tw, 2008
U.S. National Institute for Occupational Safety and Health,
2001 NIOSH alert: preventing injuries and deaths of
workers who operate or work near forklift, NIOSH
Publication Number 2001-109.
Collins, J.W., Landen, D.D., Kisner, S.M., Johnston, J.J.,
Chin, S.F., and Kennedy, R.D., 1999. Fatal occupational injuries associated with forklifts, United States,
19801994, American Journal of Industrial Medicine,
Vol. 36, 504512.
Yang, Z.Z., 2006. Forklifts accidents analysis and prevention,
Industry Safety Technology, 2630.
The Rule of Labor Safety and Health Facilities, 2007.
Chapter 5, Council of Labor Affairs, Executive Yuan,
Taipei, Taiwan, ROC.
The Machine Tool Protection Standard, 2001. Chapter 5,
Council of Labor Affairs, Executive Yuan, Taipei, Taiwan,
ROC.
The Rule of Labor Safety Health Organization Management
and Automatic Inspection, 2002. Chapters 4 and 5, Council of Labor Affairs, Executive Yuan, Taipei, Taiwan,
ROC.

Work environmentemployer event

a. Ensure that workplace safety inspections are routinely conducted by a person who can identify
hazards and conditions that are dangerous to workers. Hazards include obstructions in the aisle, blind
corners and intersections, and forklifts that come
too close to workers on foot. The person who conducts the inspections should have the authority to
implement prompt corrective measures.
b. Enforce safe driving practices, such as obeying
speed limits, stopping at stop signs, and slowing
down and blowing the horn at intersections.
c. Repair and maintain cracks, crumbling edges, and
other defects on loading docks, aisles, and other
operating surfaces.
4.5

Workerslabor event

a. Do not operate a forklift unless trained and licensed.

b. Use seat belts if they are available.

43

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Formal modelling of incidents and accidents as a means for enriching

training material for satellite control operations
S. Basnyat, P. Palanque & R. Bernhaupt
IHCS-IRIT, University Paul Sabatier, Toulouse, France

E. Poupart
Products and Ground Systems Directorate/Generic Ground Systems Office, CNES, Toulouse, France

ABSTRACT: This paper presents a model-based approach for improving the training of satellite control room
operators. By identifying hazardous system states and potential scenarios leading to those states, suggestions
are made highlighting the required focus of training material. Our approach is grounded on current knowledge
in the field of interactive systems modeling and barrier analysis. Its application is shown on a satellite control
room incident.
1

1990) argues that operator training must be supported

by the development of appropriate displays. While
this is true, an appropriate interface of a safety-critical
system may not deter an incident or accident from
occurring (as shown in this paper) if operator training
does not accentuate critical actions.
The most efficient way to effectively alter operator
behaviour is through training. However, the design of
the training material faces the same constraints as the
design of the system itself i.e. that usability, functionality coverage, task-based structuring, . . . are critical
to make training efficient. Indeed, training operators with the wrong information or designing training
material that only offers partial coverage of available
system functions will certainly lead to further failures.
We have defined a formal description technique
dedicated to the specification of safety-critical interactive systems. Initial attempts to deal with incidents and
accidents was by means of reducing development time
to a minimum making it possible to iterate quickly and
thus modify an exiting model of the system in a RAD
(Rapid Application Development) way (Navarre et al.
2001). Less formal approaches to informing operator
training have been presented in (Shang et al. 2006) in
which the authors use 3D technology and scenarios as
a means of improving process safety.
As shown in Table 1, certain categories of systems cannot accommodate iterative re-designs and
thus require the definition and the integration of barriers on top of the existent degraded system. In (Schupp
et al. 2006), (Basnyat and Palanque 2006), we presented how the formal description technique could
be applied to the description of system and human

INTRODUCTION

Preventing incidents and accidents from recurring is

a way of improving safety and reliability of safetycritical systems. When iterations of the development
process can be rapid (as for instance for most web
applications), the system can be easily modified and
redeployed integrating behavioural changes that would
prevent the same incident or accident from recurring.
When the development process is more resource consuming by, for instance, the addition of certification
phases and the need to abide by standards, the design
and implementation of barriers (Hollnagel 2004) is
considered. Previous research (Basnyat et al. 2007)
proposes the specification and integration of barriers to existing systems in order to prevent undesired
consequences. Such barriers are designed so that they
can be considered as patches over an already existing and deployed system. These two aforementioned
approaches are potentially complementary (typically,
one would be preferred to the other depending on the
severity of the failures or incidents that occurred),
putting the system at the centre of the preoccupations
of the developers.
Another less common possibility is to adopt an
operator-centred view and to act on their behaviour
leaving the system as it is. This approach is contradictory to typical Human-Computer Interaction philosophies (that promote user-centred design approaches
which aim to produce systems adapted to the tasks
and knowledge of the operators) but may be necessary
if the system is hard to modify (too expensive, too
time consuming, unreachable . . .). Norman (Norman

45

Table 1.

Categorization of types of users and systems.

Identification
of users

Types of users

Examples
of systems

Training availability

Response to failure

Barrier
implementation

Anonymous

General public

Walk-up-and-use

None

Software/hardware patch

Technical barrier

Many
identifiable

General public

Microsoft Office

Software/hardware + online
documentation patch

Technical barrier

Few
identifiable

Specialists

Software/hardware +
documentation patch

Technical barrier

Pilots

Software
development
tools
Aircraft cockpits

Online documentation
onlytraining offered
by third parties
OptionalIn house and
third parties

Satellite
operators

Satellite control
room

Software/hardware/documentation patch & training

Software/hardware patch
(unlikely) training

Socio-technical
barrier
Human barrier

Very few
identifiable

Trainedin house only

Trainedin house only

Modifications to an error-prone version of the training

material (both the actual document and its ICO model)
are performed, so that operator behaviour that led to
the incident/accident is substituted by safer behaviour.
As for the work on barriers, the formal description technique allows for verifying that the training
material covers all the system states including the ones
known to lead to the accident/incident. We will provide examples of models based on an example of an
operational procedure in a spacecraft control room.
The approach provides a systematic way to deal with
the necessary increase of reliability of the operations
of safety-critical interactive systems.
The approach uses the following foundations.
A system model is produced, using the ICO notation,
representing the behaviour and all possible states of the
currently unreliable system on which an incident was
identified. Using formal analysis techniques, such as
marking graphs for Petri nets, an extraction of all possible scenarios leading to the same hazardous state in
which the incident occurred is performed. This ensures
that not only the actual scenario is identified, but any
additional scenarios as well. A barrier analysis is then
performed, in order to identify which technical, human
and/or socio-technical barriers must be implemented
in order to avoid such a scenario from reoccurring.
This however, depends on whether or not the system
is reachable and modifiable. In the case of technical
barriers, the system model is modified and the same
hazardous scenarios are run on the improved system
model as a means of verification. The human barriers, based on identification of hazardous states, must
be realised via modifications to training material and
selective training.
The following section presents related work in the
field of model-based training. We then present the
case study and an incident in section 3, followed by
our training approach to safety in section 4. Section 5
briefly presents the ICO formalism and the Petshop
support tool used for the formal description of the
interactive system. We present in section 6 system

barriers and how the model-level integration allows

to assess, a priori, the adequacy and efficiency of the
barrier.
The current paper focuses on a new problem corresponding to the highlighted lower row of Table 1.
It corresponds to a research project in collaboration
with CNES (the French Space Agency) and deals with
command and control systems for spacecraft ground
segments. In such systems, modifications to the spacecraft are extremely rare and limited in scope which
significantly changes incident and accident management with respect to systems we have formally worked
on such as Air Traffic Control workstations (Palanque
et al. 1997) or aircraft cockpit applications (Barboni
et al. 2007), (Navarre et al. 2004).
For the space systems we are currently considering,
only the ground system is partly modifiable while the
embedded system remains mostly unreachable.
To address the issue of managing incidents and
accidents in the space domain we promote a formal
methods approach focussing on operators. This paper
presents an approach that shows how incidents and
accidents can be prevented from recurring by defining
adequate training material to alter operator behaviour.
The paper illustrates how to exploit a graphical formal
description technique called Interactive Cooperative
Objects (ICOs) (Navarre et al. 2003), based on highlevel Petri nets, for describing training material and
operators tasks. Using such a formal method ensures
knowledge of complete coverage of system states as a
means of informing training material. When we talk
about complete coverage of system states, or unambiguous descriptions, we refer to the software part of
the global system, which alone can be considered as an
improvement to what is currently achieved with formal methods for describing states and events which
are at the core of any interactive system. Verification
techniques can then be used to confront such
models with the system model of the subsection of
the ground segment and the model of a subsection of
the spacecraft in order to assess their compatibility.

46

models relating to the reported incident and discuss

how these models can be used to inform training
material in section 7. We then conclude in the final
section.

RELATED WORK ON MODEL-BASED

TRAINING

Though there are few industrial methodological

model-based training systems, this section is dedicated to theoretical approaches and two established
industrial frameworks. Related fields of research that
have dedicated annual international conferences are
Intelligent Tutoring Systems (ITS), Knowledge Based
Systems and Artificial Intelligence etc.
Such model-based approaches focus on a wide
range of models, including cognitive processes, the
physical device, user tasks, knowledge, domain etc.
It appears that the system models, a key factor in
determining hazardous scenarios and states, are often
omitted or under specified. This will be exemplified
in the following paragraphs.
The Volpe Center (National Transportation Systems Centre) has designed and implemented training
programs to satisfy many different types of needs,
including systems use training, workforce skills training, awareness training, change management training,
and mission critical training.
Although modeling operator knowledge does not
guarantee the safety of an application (Johnson 1997),
Chris Johnson argues that epistemic logics (a form
of textual representation of operator knowledge) can
be recruited to represent knowledge requirements for
safety-critical interfaces. Similarly to our approach,
the formalism highlights hazardous situations but does
not suggest how to represent this on the interface
to inform and support the operator. Interface design,
based on usability principles, design guidelines, Fitts
law (Fitts 1954) etc. is a discipline in its own right and
should be applied after system analysis.
2.1

Figure 1.

Example of a training scenario (from (Lin 2001)).

While Kontogiannis targets safety and productivity

in process control, Lin (Lin 2001) presents a Petrinet based approach, again focusing on task models,
for dealing with online training systems, showing
that modelled training tasks and training scenarios
analysis can facilitate creating a knowledge base of
online training systems. Training plans (TP) and Training scenarios (TS) are proposed. The TS-nets are
separated into four areas (see Figure 1).
Simulation area: for simulating the online training
environment
Interface area: for trainee-system interaction
Evaluation sub-net: for monitoring and recording
trainee performance
Instruction area (the focus of the research)
Elizale (Elizalde et al. 2006) presents an Intelligent Assistant for Operators Training (IAOT). The
approach, based on Petri nets takes a derived action
plan, indicating the correct sequence of actions necessary to reach an optimum operation and translates it
into a Petri net. The Petri net is used to follow the operators action and provide advice in the case of a deviation
from recommended actions. What is different in this
approach, is that the recommended actions are not
user-centred, as with task modeling approaches, but
system-centred as designed by a decision-theoretic
planner.

Petri-net based methods

In the introduction, Petri nets were advocated as a

means of formal graphical representation of a system behavior. Within the literature are several Petri
net-based methods for training purposes.
Kontogiannis (Kontogiannis 2005) uses coloured
Petri nets for integrating task and cognitive models
in order to analyse how operators process information, make decisions or cope with suspended tasks and
errors. The aim is to enhance safety at work. While the
approach is based on formal methods, there is no proposed means of verifying that the identified tasks and
workload are supported by the intended system as no
system model is presented.

2.2 Frameworks
Within the Model Based Industrial Training (MOBIT)
project, Keith Brown (Brown 1999) developed Modelbased adaptive training framework (MOBAT). The
framework consists of a set of specification and realisation methods, for model-based intelligent training
agents in changing industrial training situations. After
a specification of training problem requirements, a
detailed analysis splits into three separate areas for:
(a) a specification of the tasks a trainee is expected

47

to learn; (b) the specification of expertise and (c) a

specification of the target trainees.
The MOBAT framework includes multiple models, such as the task model, cognitive model, physical
model and domain model. The closest representation
of a system model in the framework (physical model)
(modeled using specific notations), does not provide
explicit representation of all the system states or of the
human-computer interaction. It does however provide
the facility for a user to request information relating
to past, present and future states (Khan et al. 1998).
A second model-based framework dedicated to
safety-critical domains, is NASAs Man Machine
Design and Analysis System (MIDAS). MIDAS is a
a 3D rapid prototyping human performance modeling and simulation environment that facilitates the
design, visualization, and computational evaluation
of complex man-machine system concepts in simulated operational environments (NASA 2005). The
framework has several embedded models including,
an anthropometric model of human figure, visual
perception, Updatable World Representation (UWR),
decisions, task loading, mission activities. . .
While the related work discussed in this section
focus on model-based training, they do not specifically
target post-incident improvements. For this reason, we
argue that an explicit system model is required, as a
means of identifying hazardous states and scenarios
leading to those states for informing training material
with intention to improve operator awareness.
3

Figure 2.
mands.

CASE STUDY: SENDING A HAZARDOUS

TELECOMMAND

Typical application interface for sending telecom-

Figure 3. CNES Octave application interface for sending

telecommands.

Satellites and spacecraft, are monitored and controlled

via ground segment applications in control centres
with which satellite operators implement operational
procedures. A procedure contains instructions such
as send telecommand, check telemeasure, wait etc.
Figure 2 and Figure 3 provide illustrations of typical
applications for managing and sending telecommands
(TC) to spacecraft. To give the reader an idea, the
figures have been separated into sections for explanatory purposes. In Figure 2, section 1 indicates the
current view type, i.e. the commander, the monitor, the
support engineer etc as well as the toolbar. Section 2
gives the configuration of the current telecommand
and the mode that the operator must be in, in order to
send that telecommand. Section 3 shows the history
of the current procedure and the current mode (operational or preparation). Section 4 details the active stack
(current procedure).
The second example (Figure 3) is of Octave, A
Portable, Distributed, Opened Platform for Interoperable Monitoring Services (Cortiade and Cros 2008) and
(Pipo and Cros 2006) developed by SSII-CS (French

Software Engineering Services Company) and is to

be used with future satellites such as JASON-2 and
MeghaTropique. It has the similar interface sections as
the previous example, 1) toolbar, 2) information and
triggering area (detailed information about the TC and
triggering button for TCs), 3) the current procedure
and 4) the history.
Certain procedures contain telecommand that may
have potential catastrophic consequences or critical
consequences e.g. loss of mission, negative incidence on the mission (halt maneuver e.g. and/or on
the vehicle (switch off of an equipment e.g.). These
telecommands need an individual confirmation by
the telecommand operator. This classification is performed at flight segment level and related commands
are adequately flagged (catastrophic) in the MDB
(mission database). These hazardous telecommands
have a specific presentation (for example a different
color) in the local stack display.

48

Figure 4.

Hazardous telecommand icon.

The sending of a hazardous TC requires 3 mouse

clicks (as opposed to 2 for non-hazardous telecommands). With the exception of the Red Button CAM
command, all operator initiated commands involving
functions that could lead to catastrophic consequences
shall be three step operations with feedback from the
function initiator after each step prior to the acceptance
of the command. That requirement is extracted from
a specification document at CNES not referenced for
confidentiality reasons.
Before presenting the incident, we describe here the
correct behaviour for sending a hazardous telecommand based on the example Monitoring and Control
application in Figure 2. The Commanding operator
highlights the TC to be sent and clicks on the send
button. The operator should note that the icon inline
with the TC has a yellow ! indicating that it is a TC
is classified as hazardous (see lower line of Figure 4).
After clicking send, a Telecommand Inspection
window appears detailing the TCs parameters. The
Commanding operator verifies that the parameters
appear to be correct and clicks on the send TC button.
Following this 2nd click, a red window called Confirmation appears, in the top left hand corner of the
application. At this point, the Commanding operator
must use a dedicated channel on the intercom system to
request confirmation from the Flight Director before
sending a hazardous TC from the Flight Director. The
communication is similar to FLIGHT from COMMANDER on OPS EXEC: The next TC is hazardous,
do you confirm its sending? and the reply would be
Commander from FLIGHT on OPS EXEX: GO for
the TC which performs x, y, z. . .. After receiving the
go-ahead, the Commander can send the TC.
3.1

Figure 5.

Example of an intercom system.

from the Flight Director. The Commander was thus

clicking numerous times to send the hazardous TCs.
The Commander was distracted by technical conversation regarding a forthcoming operation; his vigilance
(see research on situation awareness (Lee 2005 and
Eiff 1999)) was therefore reduced. The Commander was involved in this conversation because there
were only 2 team operators at that time as opposed
to 3 or 4 that would have been necessary for the
workload. Two additional operators were holding
technical discussions at less than a meter from the
Commander resulting in perturbation to the auditory
environment.
A parallel can be drawn between the reduced number of operators in the test phase of the case study,
and the situation in ACC Zurich during the Ueberlingen accident (Johnson 2006). Our approach does not
attempt to encompass such aspects as we mainly focus
on the computer part of such a complex socio technical
system and even more restrictively, its software part.
As opposed to current research in the field of computer
science we take into account user related aspects of the
computer system including input devices, interaction
and visualisation techniques.

TRAINING APPROACH TO SAFETY

The introduction and related work sections argued for

the use of formal methods and more particularly modelbased design for supporting training of operators as
a means of encompassing all possible system states
and identifying hazardous states. Furthermore, it has
been identified (Table 1) that in response to an incident, such a satellite control room application can
often only be made safer via training (a human barrier), as making modifications to the ground system
can be a lengthy procedure after satellite deployment. We propose, in the following two paragraphs,

The incident

The incident described in this section occurred during the test phase of a satellite development. The
Commanding operator sent a TC without receiving
confirmation from the Flight Director. There was no
impact on the satellite involved in the tests.
The circumstances surrounding the incident were
the following. Several minutes prior to the incident,
45 hazardous TCs were sent with a global go-ahead

49

two model-based approaches for improving safety

following the hazardous TC incident reported in
section 3.1.
4.1

Changing operator behaviour without system

modification

This approach assumes that modifications to the system are not possible. Therefore, changes to operator
behavior are necessary. Using a model representing the
behavior of the system, it is possible to clearly identify the hazardous state within which operators must
be wary and intentionally avoid the known problem.
The system model will also indicate the pre-conditions
(system and/or external events) that must be met in
order to reach a particular hazardous state providing
an indicator of which actions the operator must not
perform.
The advantages of this human-barrier approach lie
in the fact that no modifications to the system are
necessary (no technical barriers required) and is efficient if only several operators lie in the fact that
no modifications to the system are necessary (no
technical barriers required) and is efficient if only
several users are involved (as within a control centre). We would argue that this is not an ideal solution
to the problem, but that the company would have
to make do with it as a last resort if the system is
inaccessible.
4.2

Changing operator behaviour with system

modification

Figure 6. Approach to formal modelling of incidents and

accidents as a means for enriching training material.

The second proposed approach assumes that the

ground system is modifiable. In the same way as
before, the hazardous state is identified using the
system model and preconditions are identified after
running scenarios on the system model. A barrier analysis and implementation is then performed (Basnyat,
Palanque, Schupp, and Wright 2007). This implies
localised changes (not spread throughout entire system) to the system behavior. The existing training
material (if any) must then be updated to reflect the
change in behaviour due to the technical barrier implementation. Similarly, localized changes to the training material are made from an identified hazardous
system state.
While modifications to the ground segment may
have an impact on the onboard system, it can be considered more reliable as the system will block the
identified hazardous situation. The operators would
receive additional differential training (selective training since the behaviour of the system before and
after modifications is known) for the new behavior but even if they do not remember, or slip into
old routines, the improved system will block the
actions.

The point here is not to force the operator to avoid

an action, but to accurately make them learn the new
system behavior.

INTERACTIVE COOPERATIVE
OBJECTS & PETSHOP

The Interactive Cooperative Objects (ICOs) formalism is a formal description technique dedicated
to the specification of interactive systems (Bastide
et al. 2000). It uses concepts borrowed from the
object-oriented approach (dynamic instantiation, classification, encapsulation, inheritance, client/server
relationship) to describe the structural or static aspects
of systems, and uses high-level Petri nets (Genrich 1991) to describe heir dynamic or behavioural
aspects.
An ICO specification fully describes the potential
interactions that users may have with the application. The specification encompasses both the input
aspects of the interaction (i.e. how user actions impact

50

Figure 7.

An ICO model for procedure management.

models to enable designers to break-down complexity

into several communicating models. This feature of
the ICO notation has not been exploited in this paper
as explanation through one single (even fairly illegible model) is better choice for explanatory purposes,
especially as on paper a model cannot be manipulated.
The diagram has been segregated for explanatory purposes. Part A represents the behavior of an available
list of instructions contained within the procedure.
In this case they are CheckTM (check telemeasure),
Switch (an operator choice), Wait, Warning and TC
(telecommand). These instructions can be seen globally in part B read from left to right. Parts C and D
(the focus of the incident) describe the behaviour for
sending of a TC (normal and hazardous). For brevity,
these are the only sections of the model that will be
discussed in detail.
The model is such that the operator may select any
instruction before selecting the instruction he wishes
to implement. However, the instructions must be carried out in a given order defined within the procedure
and correctly restricted by the application. While the
operator is effectuating an instruction (usually on a
separate window superimposed on the list of instructions), he cannot click on or select another instruction
until his current instruction is terminated. This behavior of instruction order restriction is represented with
a place OrderX after the terminating click of each
instruction, a transition Next/Ok.
The ICO tool, Petshop, provides a novel feature
with respect to common Petri nets, called virtual
places, to increase the explainability of models
(by reducing the number of crossing arcs in the net).

on the inner state of the application, and which

actions are enabled at any given time) and its output
aspects (i.e. when and how the application displays
information relevant to the user).
An ICO specification is fully executable, which
gives the possibility to prototype and test an application before it is fully implemented (Navarre et al.
2000). The specification can also be validated using
analysis and proof tools developed within the Petri
nets community. In subsequent sections, we use the
following symbols (see Figure 7).
States of the system are represented by the distribution of tokens into places
Actions triggered in an autonomous way by the
system are called transitions
Actions triggered by users are represented by half
bordered transition

SYSTEM MODEL OF PROCEDURE

BEHAVIOUR

Using the ICO formalism and its CASE tool, Petshop (Navarre, Palanque, and Bastide 2003), a model
describing the behavior of a ground segment application for the monitoring and control of a non-realistic
procedure, including the sending of a hazardous
telecommand has been designed. Figure 7 illustrates
this ICO model. It can be seen, that as the complexity
of the system increases, so does the Petri net model.
This is the reason why the ICO notation that we use
involves various communication mechanisms between

51

A virtual place is a partial copy of a normal place. It

adopts the display properties (such as markings) but
not the arc connections. The arcs can be connected
to normal places or to virtual places. The display is
therefore modified allowing easier reorganisation of
models (see (Barboni et al. 2006)). In the ICO procedure model, there are 3 uses of the virtual place;
Instruction_List, a place containing tokens representing the available instructions, Block_Selection, a place
restricting the operator from selecting an instruction
while manipulating another and Past_Instructions, a
place containing the tokens representing instructions
that have terminated.
Referring back to the incident described, we are
particularly interested in the sending of a hazardous
telecommand. Once the operator has terminated all of
the instructions preceding the telecommand instruction, transition becomes fireable TC (see upper transition of Figure 8) and parts C and D of the model
become the focus of the simulation.

6.1

x.equals(tc) and x.equals(tchaza) in transitions SendTc2_1, SendTc2_2 respectively. If a normal

TC is being treated, only transitions SendTc2_1 and
Cancel2 will be fireable. A click on the SendTC button
(Transition SendTC_1) will terminate the instruction, sending a token in place Past_instructions. If
the operator clicks CANCEL, transition Cancel2,
the TC is returned to the instruction list ready for
reimplementation.
6.2 Sending a hazardous telecommand
If the TC (token) is of type hazardous, only transitions
SendTc2_2 and Cancel2 will be fireable. Respecting the requirements, a 3rd confirmation click is
required. After transition SendTc2_2 is fired, place
Display_Confirmation receives a token. The key to
the incident described in section 3.1 lies in this
state.
Before clicking for the 3rd time to send the hazardous TC, the operator should verbally request the
Go-Ahead from the Flight Director. The application
modelled allows the operator to click SEND (or
CANCEL) without the Go-Ahead. It is therefore up
to the operator to recall the fact that dialogue must be
initiated.
Part D in Figure 7 represents the potential scenarios
available, if transition DialogWithFD is fired, a token
is received in place RequestPending (representing the
operator initiating dialogue with the Flight Director).
This part of the complete model is shown in Figure 9.
For improved legibility, the virtual places representing Block_selection and Past_instructions_have been
removed from the diagram. For simulation purposes,
the transition DialogWithFD is represented using the
autonomous type of transition even though the event
is external to the system behaviour. The action does

Sending a telecommand

When transition Tc is fired, place ReadyToSendTc

receives a token. From here, the application allows
the operator can click SEND (the red button next to
number 2 in Figure 2). This is represented by transition
SendTc1. After this external event has been received,
place Display_TC_inspection receives a token. In
this state, the application displays a pop-up window
and the operator can either click SEND or CANCEL. In the model, transitions SendTc2_1, SendTc2_2
and Cancel2 become fireable. The two Send transitions identify the value of the token (x) currently in
place Display_TC_inspection, to determine whether
it is a normal TC or a hazardous TC. The code

Figure 8. Send TC section of complete model (part C of

Figure 7).

Figure 9.

52

Request go-ahead (part D of Figure 7).

improvements to the modelled application assuming

system modifications cannot and subsequently can
be made. Each suggestion includes a barrier classification according to Hollnagels Barrier systems and
barrier functions (Hollnagel 1999).

not involve the operator clicking a button on the interface. Rather, this action can be considered as a human
cognitive task and subsequently physical interaction with an independent communication system (see
Figure 5).
Tokens in places RequestPending and FlightDirector allow transition RequestGoAheadFD to be fireable. This transition contains the statement y=fd.
requestGoAhead(x).
It is important to note, that data from the token
in place FlightDirector would come from a separate model representing the behaviour of the Flight
Director (his interactions with control room members etc). Once place Result receives a token,
4 transitions to be fireable: Go, WaitTimeAndGo,
WaitDataAndGo and NoGo, representing the possible responses from the Flight Director. Again, these
transitions (apart from the one involving time) are
represented using autonomous transitions. While they
are external events; they are not events existing
in the modelled application (i.e. interface buttons).
These are verbal communications. In the case of each
result, the operator can still click on both SEND
and FAIL on the interface, transitions SendTc3_x and
Cancel3_ x in Figure 9. We provide here the explanation of one scenario. If the token (x) from the
FlightDirector place contains the data WaitTime,
then transition WaitTimeAndGo, (containing the statement y==WaitTime) is fireable taking the value
of y.
Within the scenario in which the operator must
wait for data (from an external source) before clicking
SEND, the model contains references to an external service (small arrows on places SIP_getData,
SOP_getData and SEP_getData, Service Input, Output and Exception Port respectively. When in state
WaitingForData a second model, representing the service getData would interact with the current procedure
model.
7

7.1 Without system modifications

If the application cannot be modified, training should
focus on operator behaviour between the 2nd and
3rd click for sending a hazardous TC. Section 6.2
identifies that the process of establishing dialogue
with the Flight Director as being potentially completely excluded from the operational procedure. This
is why part D of Figure 7 is intentionally segregated.
Training of operators concerning this behaviour
should explicitly highlight the fact that although they
can click the SEND button without actually receiving the Go-Ahead from the Flight Director, that it is
imperative to request the Go-Ahead. A go-around
and potential support technique would be to use a
paper checklist. This suggestion is an immaterial barrier system with barrier functions Monitoring and
Prescribing.

7.2 With system modifications

Assuming the application is modifiable, an improvement in system reliability could be achieved using
several technical (software) barriers.
Support or replace the current verbal Go-Ahead
with an exchange of data between the two parties.
Functional barrier system with hindering barrier
function
Modify the application so that the Go-Ahead data is
requested and sent via interaction with the interface,
the packet would contain information necessary for
the Flight Director to make a decision. Functional
barrier system with soft preventing barrier function.
Change the 3rd SEND button behaviour, so that it
is disabled pending receipt of the Go-Ahead packet
(including timer and additional data if necessary).
Functional barrier system with soft preventing barrier function.
Although the application analysed has a very visible red window providing the 3rd SEND button (as demanded in the system requirements),
it does not include any form of reminder to
the operator that dialogue with Flight Director must established. The 3rd pop-up window
should thus display text informing the operator of the data exchange process. Symbolic barrier system with regulating and indicating barrier
functions.

MODEL-BASED SUPPORT FOR TRAINING

The aim of the approach presented in this paper, is

to use a model-based representation of current system
behaviour (in this case after an incident) to support
training of operators in order to improve reliability in
a satellite control room.
The analysis and modelling of the monitoring
and control application including operator interactions, and reported incident enables us to identify a
hazardous state: After transition SendTc2_2 is fired
(2nd confirmation of a hazardous TC) and place
Display_Confirmation contains a token.
Table 1 highlights that the only likely response to
a failure in a satellite control room is to modify operator training. The following two subsections provide

53

CONCLUSIONS AND FURTHER WORK

Cortiade, E. and PA Cros. 2008. OCTAVE: a data modeldriven Monitoring and Control system in accordance
with emerging CCSDS standards such as XTCE and
SM&C architecture. SpaceOps 2008 1216 May 2008,
Heidelberg, Germany.
Eiff, G.M. 1999. Organizational safety culture. In R.S. Jensen
(Ed.). Tenth International Symposium on Aviation Psychology (pp. 778783). Columbus. OH: The Ohio State
University.
Elizalde, Francisco, Enrique Sucar, and Pablo deBuen. 2006.
An Intelligent Assistant for Training of Power Plant
Operators. pp. 205207 in Proceedings of the Sixth
IEEE International Conference on Advanced Learning
Technologies. IEEE Computer Society.
Fitts, P.M. 1954. The Information Capacity of the
Human Motor System in Controlling the Amplitude of
Movement. . ..
Genrich, H.J. 1991. Predicate/Transitions Nets, HighLevels Petri-Nets: Theory and Application. pp. 343 in.
Springer Verlag.
Hollnagel, E. 1999. Accidents and barriers. pp. 175180
in Proceedings CSAPC99, In J.M. Hoc, P. Millot,
E. Hollnagel & P.C. Cacciabue (Eds.). Villeneuve dAsq,
France: Presses Universitaires de Valenciennes.
Hollnagel, E. 2004. Barriers and Accident Prevention.
Ashgage.
Johnson, C.W. 1997. Beyond Belief: Representing
Knowledge Requirements For The Operation of SafetyCritical Interfaces. pp. 315322 in Proceedings of
the IFIP TC13 International Conference on HumanComputer Interaction. Chapman & Hall, Ltd http://portal.
acm.org/citation.cfm?id=647403.723503&coll=GUIDE
&dl=GUIDE&CFID=9330778&CFTOKEN=19787785
(Accessed February 26, 2008).
Johnson, C.W. 2006. Understanding the Interaction Between
Safety Management and the Can Do Attitude in
Air Traffic Management: Modelling the Causes of
the Ueberlingen Mid-Air Collision. Proceedings of
Human-Computer Interaction in Aerospace 2006, Seattle, USA, 2022 September 2006. EDITORS F. Reuzeau
and K. Corker.Cepadues Editions Toulouse, France.
pp. 105113. ISBN 285428-748-7.
Khan, Paul, Brown, and Leitch. 1998. Model-Based Explanations in Simulation-Based Training. Intelligent Tutoring Systems. http://dx.doi.org/10.1007/3-540-68716-5_7
(Accessed February 20, 2008).
Kontogiannis, Tom. 2005. Integration of task networks and
cognitive user models using coloured Petri nets and its
application to job design for safety and productivity.
Cogn. Technol. Work 7:241261.
Lee, Jang R, Fanjoy, Richard O, Dillman, Brian G. The
Effects of Safety Information on Aeronautical Decision
Making. Journal of Air Transportation.
Lin, Fuhua. 2001. Modeling online instruction knowledge
using Petri nets. pp. 212215 vol.1 in Communications,
Computers and signal Processing, 2001. PACRIM. 2001
IEEE Pacific Rim Conference on, vol. 1.
NASA. 2005. Man-machine Integration Design and
Analysis System (MIDAS). http://human-factors.arc.
nasa.gov/dev/www-midas/index.html (Accessed February 19, 2008).
Navarre, D, P Palanque, and R Bastide. 2004. A Formal
Description Technique for the Behavioural Description

We have presented a model-based approach for

identifying human-computer interaction hazards and
their related mitigating human (training) and technical (software) barriers, classified according to
Hollnagels barrier systems and barrier functions
(Hollnagel 1999).
The approach, applied to a satellite ground segment
control room incident, aids in identifying barriers such
as symbolic and immaterial barriers, necessary in the
case where the degraded system is not modifiable.
These barriers may not have been considered within
the development process, but have been identified
using the system model.
We have used the ICO formal description technique
(a Petri net based formalism dedicated to the modelling of interactive systems) to model the monitoring
and control of an operational procedure using a satellite ground segment application. The model provides
explicit representation of all possible system states,
taking into account human interactions, internal and
external events, in order to accurately identify hazardous states and scenarios leading to those states via
human-computer interactions.

ACKNOWLEDGEMENTS
This research was financed by the CNES R&T Tortuga
project, R-S08/BS-0003-029.

REFERENCES
Barboni, E, D Navarre, P Palanque, and S Basnyat. 2007.
A Formal Description Technique for the Behavioural
Description of Interactive Applications Compliant with
ARINC Specification 661. Hotel Costa da Caparica,
Lisbon, Portugal.
Barboni, E, D Navarre, P Palanque, and S Basnyat. 2006.
Addressing Issues Raised by the Exploitation of Formal Specification Techniques for Interactive Cockpit
Applications.
Basnyat, S, and P Palanque. 2006. A Barrier-based
Approach for the Design of Safety Critical Interactive
Application. vol. Guedes Soares & Zio (eds). Estoril,
Portugal: Taylor & Francis Group.
Basnyat, S, P Palanque, B Schupp, and P Wright. 2007. Formal socio-technical barrier modelling for safety-critical
interactive systems design. Special Edition of Elseviers
Safety Science. Special Issue safety in design 45:545565.
Bastide, R, O Sy, P Palanque, and D Navarre. 2000. Formal
specification of CORBA services: experience and lessons
learned. . . ACM Press.
Brown, Keith. 1999. MOBIT a model-based framework for
intelligent training. pp. 6/16/4 in Invited paper at the
IEEE Colloqium on AI.

54

of Interactive Applications Compliant with ARINC 661

Specification.
Navarre, D, P Palanque, and R Bastide. 2003. A ToolSupported Design Framework for Safety Critical Interactive Systems. Interacting with computers 15/3:309328.
Navarre, D, P Palanque, R Bastide, and O Sy. 2001. A
Model-Based Tool for Interactive Prototyping of Highly
Interactive Applications.
Navarre, D, P Palanque, R Bastide, and O Sy. 2000.
Structuring Interactive Systems Specifications for Executability and Prototypability. pp. 97109 in, vol. no,
1946. Lecture Notes in Computer Science, Springer.
Norman, D.A. 1990. The Problem with Automation:
Inappropriate Feedback and Interaction, not OverAutomation Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 327:585593.

Palanque, P, R Bastide, F Paterno, R Bastide, and F Paterno.

1997. Formal Specification as a Tool for Objective
Assessment of Safety-Critical Interactive Systems. . .
Sydney, Australia: Chapman et Hall.
Pipo, C, and C.A Cros. 2006. Octave: A Portable, Distributed, Opened Platform for Interoperable Monitoring
Services. Rome, Italy.
Schupp, B, S Basnyat, P Palanque, and P Wright. 2006.
A Barrier-Approach to Inform Model-Based Design of
Safety-Critical Interactive Systems.
Shang, P.W.H., L Chung, K Vezzadini, K Loupos, and
W Hoekstra. 2006. Integrating VR and knowledgebased technologies to facilitate the development of operator training systems and scenarios to improve process
safety. vol. Guedes Soares & Zio (eds). Estoril, Portugal:
Taylor & Francis Group.

55

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Hazard factors analysis in regional traffic records

M. Mlynczak & J. Sipa
Wroclaw University of Technology, Poland

ABSTRACT: Road safety depends mainly on traffic intensity and conditions of road infrastructure. More
precisely, one may found many more factors influencing safety in road transportation but process of their
perception, processing and drawing conclusions by the driver while driving is usually difficult and in many cases
fall behind the reality. Diagnostic Aided Driving System (DADS) is an idea of supporting system which provides
driver with selected, most useful information which reduces risk due to environment, weather conditions as well as
statistics of past hazardous events corresponding to given road segment. Paper presents part of the project aiming
to describe relations and dependencies among different traffic conditions influencing number of road accidents.

Undesired road events are described and stored in

reports created by rescue services consisting mainly
by fire brigade and police divisions. Precise accident
identification requires giving the following facts:

FOREWORD

Road traffic is defined as movement of road vehicles,

bicycles and pedestrians. The most important features
of the traffic is effective and safe transportation of
passengers and goods. Traffic takes place in complex
system consisting of road infrastructure, vehicle operators, vehicles and weather conditions. Each of these
factors is complex and undergoes systematic and random changes. Safety is provided in that system by the
means of safety in soft (law, regulations, and local
rules) and hard form (traffic signals, barriers, etc.). In
this light an accident is a fact of violation of certain
safety condition. In the idea of MORT (Wixon 1992) it
might be understood as releasing some energy stored
in transportation system.
Analysis of accident reports, databases and papers
creates an image of necessary and sufficient conditions
leading to undesired road event.

2
2.1

exact date and time of the event (day time, week day,
month, season),
place of the event (number of lanes and traffic directions, compact/dispersed development, straight/
bend road segment, dangerous turn, dangerous
slope, top of the road elevation, road crossing
region, of equal importance or subordinated, roundabout, pedestrian crossing, public transport stop,
tram crossing, railway sub grade, railway crossing
unguarded/guarded, bridge, viaduct, trestle bridge,
pavement, pedestrian way, bike lane, shoulder,
median strip, road switch-over, property exit),
severity of consequences (collision, crash, catastrophe),
collision mode (collision in motion [front, side,
rear], run over [a pedestrian, stopped car, tree,
lamp-post, gate arm, pot-hole/hump/hole, animal],
vehicle turn-over, accident with passenger, other),
cause/accident culprit (driver, pedestrian, passenger, other reasons, complicity of traffic actors),
accidents caused by pedestrian (standing, lying in
the road, walking wrong lane, road crossing by
red light, careless road entrance [before driving
car; from behind the car or obstacle], wrong road
crossing [stopping, going back, running], crossing
at forbidden place, passing along railway, jumping
in to vehicle in motion, children before 7 year old
[playing in the street, running into the street], other),
cause/vehicle maneuver (not adjusting speed to traffic conditions, not observing regulations not giving
free way, wrong [overtaking, passing by, passing,

ACCIDENT DATABASE REVIEW

Accident description

Polish road rescue system is based on fire service.

In emergency the suitable database is created. The
database described in this paper concerns the southwestern region of Poland, Lower Silesia. Period contained in work-sheets covers years 19992005. From
thousands of records information dealing with road
accidents was selected.
The most important element of an accident description is its identification in time and road space. An
accident description includes always the same type of
information.

57

driving over pedestrian crossing, turning, stopping,

parking, reverse a car], driving wrong lane, driving by red light, not observing other signs and
signals, not maintaining safe distance between vehicles, rapid breaking, driving without required lights,
tiredness, falling asleep, decreasing of psychomotor
efficiency, other),
driver gender and experience,
age (age interval: 06, 714, 1517, 1824, 2539,
4059, 60 and over),
culprit due to the influence of alcohol (because of
driver, pedestrian, passenger, other, complicity of
traffic actors),
weather conditions (sunshine, clouds, wind, rain/
snow, fog/smog),
type of vehicle (bicycle, moped/scooter, motorbike,
car, tuck, bus, farm tractor, horse-drawn vehicle,
tram, train, other),
vehicle fault (self damage, damage of the other
vehicle),
event consequences (environment [soil, water, air],
road, road infrastructure [road facilities, buildings],
cargo [own, foreign]),
necessary actions (first-aid service, road rescue,
chemical neutralization, fire extinguishing, rebuild).

2.2

Figure 2. Ratio of number of people injured or victims in

total number of accidents.

It is seen that over 87% of accidents happen between

6 a.m. and 10 p.m. with the highest frequency range
between 8 a.m. and 6 p.m. Within observed period, the
frequency and distribution of accidents in day hours is
stable and does not change much despite from rising
motorization index in Poland. The average rate of traffic flow, according to the entire analyzed road segment
at motorway A-4, is presented in Fig. 4. It is shown
that since year 2001, volume of traffic is growing year
by year mainly due to economic development.
Similar analysis of accidents distribution in time
was done in relation to months over the year (Fig. 5)
and week days (Fig. 6).

Analysis of accidents conditions

General analysis of the database is directed on main

factors attendant circumstances. Major accidents consequences are caused proportionally by cars (over 80%
of injured and victims) but almost 30% of losses result
from trucks and road machines (tractors, combine
harvesters end the like) accidents (Fig. 1).
The worst consequences of accident dealing with
human life are usually death or bodily harm. Over
the entire observation period an average ratio of victims per accident is 0,22 and lies between 0,35 and
0,07. Average number of injured is about 1,8 per each
accident (Fig. 2).
Analysis of accidents taking place within day time
is shown in Fig. 3.

MODELLING OF ROAD EVENT NUMBER

3.1 Regression models

Statistical number of accidents along certain road
segments is a random number depending on many factors described precisely above. Randomness means
in this case that the number of expected accidents
may vary according to random events and is correlated to some other factors. Regression models allow
taking into consideration several factors influencing
accident number. Variation of accident number is
described as pure random and systematic variation
depending on some accident factors (Baruya 1999,
Evans 2005).
Two main regression models exist which are generally based on Poisson distribution (Fricker & Whitford
2004). Poisson model is adequate to random events
(accident number in a given road segment) providing
that randomness is independent for all variables taken
into account (Evans 2005, Kutz 2004, Miaou & Lum
1993).
Poisson regression function of accent number is:

Figure 1. Accident consequences fraction caused by different type of vehicle.

= e(0 +1 x1 +2 x2 ++k xk )

58

(1)

day

0,07

07:00

Figure 3.

17:00

Frequency of accidents in day time.

0,25

accident
frequency

0,2
0,15
0,1
0,05

no

Sunday

Saturday

Friday

Thursday

Wednesday

Tuesday

Figure 6. Distribution of accidents in week days over the

period 19992005.

Figure 4. Traffic volume in analyzed motorway segment in

years 19992005.
injured
accidents

120

Monday

week day
0

appropriate. Otherwise the Pascal model (2) should

be considered, because it assumes randomness of
parameter itself (Michener & Tighe 1999).

death
100
80

= e(0 +1 x1 +2 x2 ++k xk + )

60
40

(2)

where: is the Gamma distributed error term.

20
0
1

10

11
12
month

3.2
Figure 5. Distribution of accidents and victims in months
over the period 19992005.

Application of regression models

to the database

Regression analysis of road events uses database

of Provincial Command of the State Fire Service.
From that database the motorway segment of the
total length l = 131 km was selected and from
the period of 19992005 about 1015 accidents were
extracted.
Motorway segment is divided in 11 subsegments
of given length li = x1 and daily traffic volume
tvi = x2 . To the analysis are also taken into account
the following casual variables:

where: i = coefficients of regression, xi = independent variables.

Poisson distribution assumes that variance is equal
to mean number of events occurrence so that mean
and variance may be taken as measure of statistical
confidence. If variance is calculated as in the following formula = (1 + ), and (overdispersion
parameter) is not significantly different from 0, then
the regression analysis based on the Poisson model is

hvi = x3 heavy vehicle rate in traffic flow [%],

59

Ta ble 1. Regression models for all road events.

Independent variable

Model
1
2
3
4
5
6
7
8
9
10
11
12

7,06
7,23
9,22
5,19
6,75
2,33
4,56
9,35
9,94
9,83
11,49
9,91

ln l
Length
of road
segment

ln tv
Daily
traffic
volume

Hv
Heavy
vehicle
rate

1,65
1,55
1,49
2,25
1,68
2,45
2,60
2,16
2,15
2,26
2,36
1,59

0,56
0,89
0,82
0,60
0,83
0,22*
0,43*
0,90
0,99
0,95
1,12
1,06

0,09

0,11
0,09
0,09
0,10
0,05
0,06
0,06
0,06
0,05

J
Number
of all
junctions

rj
Number
of road
junctions

0,02*

0,09

0,01*
0,05

0,05

rn
Number
of road
junctions
(national
roads)

rc
Number of
road junctions
(communal roads)

rl
Number of
road junctions (local roads)

rp
Number of
exits to
parking

rm
Number of
modernization

0,16
0,08*

0,10
0,04

0,13
0,07

0,10
0,06*

0,09*

0,06*

0,77
0,86
0,85
0,80
0,65

0,05*
0,02*

0,07
0,12

* Insignificant parameters are written in italics.

Table 2.

Significance analysis of selected regression models.

Multiple correlation coefficient

Table 3.
Model

Model deviance

parameter

Model

R2

R2p

R2w

R2wp

R2ft

R2ftp

MD

8
9
10
11
12

0,72
0,74
0,76
0,78
0,70

0,88
0,90
0,92
0,95
0,85

0,69
0,70
0,72
0,75
0,67

0,87
0,88
0,91
0,94
0,84

0,69
0,70
0,72
0,75
0,67

0,87
0,90
0,92
0,96
0,85

261,26
267,63
269,76
269,83
248,61

105
105
105
105
105

0,034
0,027
0,020
0,011
0,053

0,1936
0,2617
0,3763
0,6067
0,1054

Comparison of regression models for number of road events, accidents, injured and fatalities.
0

Number of road events

0,0001
Number of accidents
0,0004
Number of injured
0,0028
Number of fatalities
2203,9357

ln l

ln tv

2,26
0,953
2,769
0,681
3,759
0,18
5,885 1,972

hv

rj rn

0,056
0,083
0,065
0,055

0,209

0,071
0,128
0,104

Rc

rl

rp

rm

0,099
0,209
0,077

0,1

0,854
0,152
0,053 1,005
0,093 0,114 1

1,718

rmi = x10 time fraction of the year with modernization road works of subsection (value from
0 to 1).

ji = x4 number of all junctions (exits and

approaches of motorway),
rji = x5 number of all types road junctions,
rni = x6 number of road junctions with national
roads,
rci = x7 number of road junctions with communal
roads,
rli = x8 number of road junctions with local
roads,
rpi = x9 number of exits and approaches of
parking places,

Accident database and construction data obtained

from the company managing motorway allowed for
building general regression model in the following
form (3):


i = li 1 tvi 2 e (0 +j xji ) ;

60

j>2

(3)

avg. number of road events, accidents,

injured, fatalities

14

number of road events

number of accidents
number of injured
number of fatalities

12
10
8
6
4
2
0
0

3000

6000

9000

12000

15000

18000

21000

24000

27000

traffic volume [veh./24h]

Figure 7. Regression functions of number of road events, accidents, injured and fatalities in function of traffic volume (all
remaining independent variables are constant). Shaded area relates to the range of traffic volume analyzed in database.

pr = 1, rm = 1. Fig. 7 shows number of road events,

number of accident, number of injured and number of
fatalities in function of daily traffic flow.
These theoretical functions are valid in the interval of traffic flow 900024000 veh./24 hours (shaded
area) what was observed in the database in the period
of years 19992005. Regression functions may be
extended on wider range of traffic flow to predict
number of interesting events in other conditions. The
phenomenon of decreasing number of fatalities in traffic flow may be explained that within seven years of
observation average quality of cars has risen. Traffic
flow is here strongly correlated with time. Better
cars move faster and generate higher traffic flow and
on the other hand may rise traffic culture and provide
higher safety level through installed passive and active
safety means. Estimation of victim number is burden higher randomness and proper and precise models
have not been elaborated jet.

Twelve regression models were created on the

basis of the formula (3) and 10 independent variables
described above. In the Table 1 the successive models
are presented containing logical combination of variables x3 to x10 . Regression functions were obtained and
selected using the squared multiple correlation coefficient R, the weighted multiple correlation coefficient
Rw and the Freeman-Tukey correlation coefficient Rft .
Five models no. 812 (marked in shaded records)
are considered as models of good statistical significance. In that models parameters are estimated properly and parameter is insignificantly different from 0.
In the estimated models number of accidents follows
Poisson distribution. Models significance analysis is
shown in Table 2. Regression coefficients denoted
with index p deal with adjusted models that describe
portion of explained systematic randomness.
Probability values given in Table 2 for parameter
greater then 0,05 prove that is insignificant and certify correct assumption of Poisson rather than Pascal
model for number of accidents.
Described calculation method and regression model
gave the basis of evaluating theoretical number of road
events. There are four important events discussed in
road safety: road event, accident with human losses,
accident with injured and fatalities. Having known historical data, these events were analyzed and created
regression models. In Table 3 parameters of regression models are put together to describe numbers of
the mentioned events. It is seen that some independent
variables does not influence on calculated number,
especially on number of fatalities.
In order to check the behavior of the obtained model
it is has been proposed that all independent variables
are constant contrary to traffic volume. It will give
regression function of number of certain events in relation to traffic volume. It is assumed that: l = 10 km,
hv = 0, 25, j = 3, rj = 1, rn = 1, rc = 0, rl = 1,

CONCLUSIONS

Two main subjects are discussed in the paper: theoretical modeling of traffic events and elaborating
of regression models on the basis of real accident
database. Proposed model consists in developing of
prediction function for various road events. There
were four types of road events concerning: all road
events, accidents, injuries and fatalities. Verification
of the assessment that number of, so called, rare
events undergoes Poisson distribution was done comparing elaborated real data with Poisson model with
parameter calculated as regression function of ten
independent variables. Conformity of five proposed
models was checked calculating statistical significance of parameter . Regression models applied to
online collected data are seen to be a part of active

61

supporting system for drivers, warning them about

rising accident risk at given road segment in certain
environmental conditions. Further development of the
models would be possible to use in road design process
especially in risk analysis and assessment.

Fricker J.D. & Whitford R.K. 2004. Fundamentals of Transportation Engineering. A Multimodal System Approach.
Pearson. Prentice Hall.
Kutz M. (ed.). 2004. Handbook of Transportation Engineering. McGraw-Hill Co.
Miaou S.P. & Lum H. 1993. Modeling Vehicle Accidents
And Highway Geometric Design Relationships. Accident
Analysis & Prevention, no. 25.
Michener R. & Tighe C. 1999. A Poisson Regression Model
of Highway Fatalities. The American Economic Review.
vol. 82, no. 2, pp. 452456.
Wixson J.R. 1992. The Development Of An Es&H Compliance Action Plan Using Management Oversight, Risk Tree
Analysis And Function Analysis System Technique. SAVE
Annual Proceedings.

REFERENCES
Baruya A. 1998. Speed-accident Relationships on European
Roads. Transport Research Laboratory, U.K.
Evans A.W. 2003. Estimating Transport Fatality Risk From
Past Accident Data. Accident Analysis & Prevention,
no. 35.

62

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Organizational analysis of availability: What are the lessons

for a high risk industrial company?
M. Voirin, S. Pierlot & Y. Dien
EDF R&DIndustrial Risk Management Department, Clamart, France

ABSTRACT: EDF R&D has developed an organisational analysis method. This method which was designed
from in depth examination of numerous industrial accidents, incidents and crises and from main scholar findings in the domain of safety, is to be applied for industrial safety purpose. After some thoughts regarding
(dis)connections between Safety and Availability, this paper analyses to what extend this method could be
used for event availability oriented analysis.

INTRODUCTION

Table 1.

Examples of events analyzed by EDF R&D.

Event

EDF Research & Development has been leading since

several years a sentinel system dedicated to the
study of worldwide cross-sectorial industrial events.
One purpose of this sentinel system is to define an
investigation method allowing to go beyond direct and
immediate causes of events, i.e. to find organizational
in-depth causes.
Based upon (i) studies of more than twenty industrial accidents, incidents and crises carried out in
the frame of this sentinel system (see table 1),
(ii) knowledge of more than fifty industrial disasters,
(iii) analysis of investigation methods used for several events (by event we mean accident or incident
or crisis, e.g. Cullen, 2000; Cullen, 2001; Columbia
Accident Investigation Board, 2003; US Chemical
Safety and Hazard Investigation Board, 2007) and
findings of scholars in domains of industrial safety and
accidents (e.g. Turner, 1978; Perrow, 1984; Roberts,
1993; Sagan, 1993; Sagan, 1994; Reason, 1997;
Vaughan, 1996; Vaughan, 1997; Vaughan, 1999; . . .),
EDF-R&D has drawn up a method for Organisational
Analysis. This method is safety oriented and can
be used for event investigation and for diagnosis of
organisational safety of an industrial plant as well.
After a brief description of main features of the
method and some remarks about similarities and differences between safety and availability, we will
examine to which extend the method could also be
used in case of events dealing with availability (versus
safety).

Texas City Explosion, March 23, 2005

Columbia Accident, February 1, 2003
Airplane Crashes above the Constance Lake, July 1, 2002
Crash of the Avianca Companys Boeing near JFK Airport,
New York, January 25, 1990
Failures of NASAs Mars Climate Orbiter and Mars Polar
Lander engines, 1999
Explosion of the Longford Gas Plant, Australia,
September 25, 1998
Crisis at the Millstone NPP from 1995 to 1999
Corrosion of the Davis Besses Vessel Head, March 2002
Tokamura reprocessing plants accident,
September 30, 1999
Fire in the Mont Blanc tunnel, March 24, 1999
Trains collision in the Paddington railways station (near
London), October 5, 1999

MAIN FEATURES OF ORGANISATIONAL

ANALYSIS METHOD

This section gives an overview of the method. More

detailed information is available (Dien, 2006; Pierlot
et al., 2007).
The method is grounded on an empirical basis (see
figure 1). But it is also based on scholar findings, and
more specifically on Reason work.

63

2.1

Historical dimension

As Llory states (1998): accident does not start

with triggering of final accidental sequence; therefore, analysis require to go back in time, [. . .] in
order to put in prominent place deterioration phenomena. Analysis has to go upstream in the History
of the organisations involved for highlighting significant malfunctioning aspects: what was not appreciated
in real time has to make sense when risk was
confirmed (i.e. when event has happened). Vaughan
reminds (2005): The O-ring erosion that caused the
loss of Challenger and the foam debris problem that
took Columbia out of the sky both had a long history.
Early warning signs has to be looked for and detected
long before time event occurrence.
Numerous industrial events show that weakness
of operational feedback could be incriminated for
their occurrencei.e. that previous relevant event(s)
was/were not taken into account or poorly treated after
their occurrence. Analysts have to pay a specific
attention at incidents, faults, malfunctioning occurred
prior to the event.
Analysis of the historical dimension is parallel
to detailed examination of parameters, of variables of
context which allow understanding of events. It has
to avoid a retrospective error. Fine knowledge of
event scenarioi.e. sequences of actions and decisions which led to itallows to assess actual mid and
long term effects of each action and decision. Analysts
have to keep in mind that this evaluation is easier to
make after the event than in real time. In other words,
analysts have to avoid a blame approach.

Figure 1. General description of the method building (from

Dien & Llory, 2006).

Figure 2. The organizational analysis three main axis.

According to Reason (1997), causes of an event is

made of three levels:
The person (having carried out the unsafe acts, the
errors);
The workplace (local error-provoking conditions);
The organization (organizational factors inducing
the event).
Development of event is bottom-up, i.e. direction causality is from organizational factors to person.
In the event analysis, direction is opposite. Starting
point of analysis is direct and immediate causes of
bad outcome (event). Then, step by step, analysis considers, as far as possible, how and when defences
failed.
In addition to results obtained by scholars in the
field of organisational studies, real event organisational analyses carried out allow us to define the three
main dimensions of an organisational approach, helping to go from direct causes to root organisational
causes (see figure 2):

2.2 Organizational network

Within an organisation, entities communicate (Entity
means a part of organisation more or less important in
terms of size, staffing. It could be a small amount
of people or even an isolated person, for instance a
whistle blower): entities exchange data, they make
common decisionsor at least they discuss for making a decision, they collaborate, . . . So it is of the
first importance to draw organisational network
between entities concerned in the event. This network
is not the formal organisation chart of entities. It is a
tool for showing numerous and complex interactions
involved for occurrence of event. It is a guideline for
carrying out the analysis; it is built all along analysis
itself.
Organisational network is hardly defined once and
for all for a given organisation. It is draft according to
the analysis goals. Parts of organisation can be ignored
because they were not involved in the event.
Organisational network allows visualising complexity of functional relationships between entities,

Historical dimension;
Organizational network;
Vertical relationships in the organization (from
field operators to plant top management).
We have to note that, if these dimensions are introduced in a independent way, they are interacting and
an analysis has to deal with them in parallel (and in
interaction).

64

and sometimes, it highlights absence of relationships

which had to be present.
2.3

Table 2. The seven Pathogenic Organizational Factors

(POF).
Production Pressures (Rupture of the equilibrium between
availability and safety to the detriment of safety)
Difficulty in implementing feedback experience
Short comings of control bodies
Short comings in the safety organizational culture
Poor handling of the organizational complexity
No re-examining of design hypothesis
Failure in daily safety management

Vertical relationships in the organization

This dimension is a part of organisational network

on which a specific focus is needed: It has to be
specifically studied because it helps to figure out
interactions between different hierarchical layers and
between managers and experts.
It covers top-down and bottom-up communications.
It is essential to isolate it since it makes easier understanding of interactions between various management
levels, experts and field operators. We have to
remind an obviousness often forgotten during event
analysis: organisation is a hierarchical system.
The main interests of this dimension are: modes of
relationships, of communication, of information flows
and modes of co-operation between hierarchical layers. Real events show that deterioration of these modes
are cause of their occurrence.
At least, thanks to this dimension, causes of an event
cannot be focussed only on field operators.
Some other concepts have to be kept in mind while
an organizational analysis is carried out as for instance:
incubation period of an event (Turner, 1978), existence or not of weak signals prior to occurrence
of the event and connected to it (Vaughan, 1996) and
existence or not of whistle blower(s), active and latent
defaults (Reason, 1997). Regarding these defaults,
analyses of events have shown that some organizational flaws (latent defaults) were present and that they
are part of the root causes of the event.
2.4

complexity of the event-causing phenomena, a factor in a certain situation may be deemed to be a

consequence in another situation.
POFs are useful during event analysis since they
allow to define goals/phenomena to look for.
A least, purpose of an organisational analysis is to
understand how organisation is working: it leads to (try
to) understand weaknesses and vulnerabilities, but also
resilience aspects. Its goal is not necessarily to explain
event from an expert point of view resulting in a list
of (more or less numerous) direct causes leading to
consequences (with, at the end the fatal consequence).
This approach leads, in addition of collecting and analysis of objective data, to take account of employees
speeches. In order to collect true information analysts have to have an empathic attitude toward people
met during analysis (collecting true information
does not mean to take what it is said as holly words.
Analysts have to understand speeches according to the
context, the role of the speaker(s), . . . Information
collected has to be cross-examined as well).

The Pathogenic Organizational Factors

2.5

Analyses have also shown some repetition of organizational root causes of events. We have defined,
according to Reasons terminology (1997) these causes
as Pathogenic Organizational Factors (POFs).
POFs cover a certain number of processes and phenomena in the organization with a negative impact
on safety. They may be identified by their various
effects within the organisation, which can be assimilated into symptoms. Their effects may be objective
or even quantifiable, or they could be subjective and so
only felt and suggested by employees. POFs represent
the formalisation of repetitive, recurrent or exemplary
phenomena detected in the occurrence of multiple
events. These factors can be described as the hidden
factors which lead to an accident.
A list of seven POFs was defined (Pierlot, 2007).
Table 2 gives a list of this seven POFs. This list should
not be considered to be definitive since new case studies findings could add some findings even though
it is relatively stable. Classification of the factors is
somewhat arbitrary since, given the multiplicity and

The Organizational Analysis Methodology

As stated above, this methodology has been already

fully described (Pierlot et al., 2007; Voirin et al., 2007).
Just let us recall here the main points of the
methodology:
Organizational Analysis relies on background
knowledge made of study cases, findings of scholars, definition of Pathogenic Organizational Factors,
Organizational Analysis explores the three main
axis described above,
Analysts use both written evidences and interviews.
The interviews are performed under a comprehensive approach, in which, not only the causes
are searched, but also the reasons, the intentions,
the motivations and the reasoning of the people
involved,
Analysts have to bring a judgment on the studied
situation, so that the puzzle of the situation could
be rebuilt,

65

 Results of the Organizational Analysis are built with

the help of a thick description (Geertz, 1998) and
then summarized.
3

also a safety mission because they are part of the second barrier against radioactive material leakage, and
also because their cooling capacities are necessary to
deal with some accidental events.
But all of these systems could have both an impact
on safety and availability. It is clear for systems dealing both with Safety and Availability, but, in case of
failures, some of the systems dealing only with availability could have clearly an impact on safety, as some
of the systems dealing only with safety could have
clearly an impact on availability (Voirin et al., 2007).
For example, if there is a turbine trip while the plant is
operating at full power, the normal way to evacuate the
energy produced by the nuclear fuel is no longer available without the help of other systems, and that may
endanger safety even if the turbine is only dedicated
to availability.
Safety and availability have then complex relationships: This is not because short term or medium
term availability is ensured that neither long term
availability or safety are also insured.
And we shall never forget that, before the accident, the Chernobyl nuclear power plant as the Bhopal
chemical factory had records of very good availability
levels, even if, in the Bhopal case, these records were
obtained with safety devices out of order since many
months (Voirin et al., 2007).

SAFETY AND AVAILABILITY: HOW

CLOSE OR DIFFERENT ARE THESE TWO
CONCEPTS?

Although the method described above was built for

safety purposes, considering the potential impacts of
the Pathogenic Organizational Factors on the safety
of a given complex system, the first applications of
this method within EDF were dedicated to availability
events.
According to the CEI 60050-191 standard, the system availability is the capacity of a system to feed its
expected mission, within given operating conditions at
a given time. Availability must not be confused with
reliability which has the same definition, excepted
that the time dimension is not at a given time, but
during a time interval.
Therefore Availability addresses two different
issues: the system capacity to avoid an damageable
event, and, if such an event had to occur, the system
capacity to recover from the event so that its initial
performances are fully met again.
Safety can be described as the system capacity to
avoid catastrophic failures, potentially hazardous for
workers, for the environment or for the public (Llory
& Dien 20062007).
Safety and Availability seem then, at the first view,
to be two un-related characteristics. But this is only
an impression. In fact, Availability and Safety are not
independent.
This is due to the fact that the equipment being
parts of a complex high risk system can deal, either
only with Availability, either only with Safety, either
with Safety and Availability (Voirin et al., 2007).
For instance, in a nuclear power plant, there systems dealing only with the plant availability, such as
the generator group which allows to produce electricity thanks to the steam coming from the Steam
Generators.
There are also systems dealing only with safety.
These systems are not required to produce electricity
and are not operating: Their purpose is only to fulfil a
safety action if needed. For example, the Safety Injection System in a nuclear power plant has for mission
to insure the water quantity within the primary circuit,
so that decay heat could be still evacuated, even in case
of a water pipe break.
At last, they are systems dealing both with safety
and availability. For example, the Steam Generators
of a nuclear power plant have an availability mission,
because they produce the sufficient steam to make the
generator group working; but Steam Generators fulfil

CAN AN ORGANIZATIONAL ANALYSIS

BE CARRIED OUT FOR AVAILABILITY
ORIENTED EVENT?

Due to this strong relationship between availability and

safety, is it possible to perform an Organisational Analysis of an Availability event in the same way Safety
oriented Organisational Analysis are performed?
We already gave a positive answer to that question (Voirin et al., 2007), even if they are noticeable
differences.
The POFs were established on the study of many
Safety Events (Pierlot et al., 2007), they rely on a
strong empirical basis However, the transposition of
such Safety oriented POFs into Availability oriented
Pathogenic Organisational Factors was performed only
from the theoretical point of view, with the help of
our knowledge of some study cases (Voirin et al.,
2007).
The main point is that, during an Availability Organisational Analysis, particular attention must be paid
to the equilibrium between Availability and Safety,
knowing that one of the Safety oriented POFs (production pressures) leads to the rupture of such a balance
(Dien et al., 2006; Pierlot et al., 2007).
It is with such a background knowledge that we
performed in 2007 the Organisational Analysis of an
event that occurred in one power plant. This event

66

most of them were reluctant to give materials to the

analysts.
It can be pointed out that the few managers who
agree to be interviewed were mainly managing entities
which could be seen, at first, in that case, as victims. In that case, managers appeared to be opened
because they thought that the organisational analysis
was the opportunity for them to let the entire organisation know that they can not be blamed for what
happened. . .
These reaction of defence of the managers could
have been forecasted. It can be over passed, knowing
that the analysts have to adopt a empathic attitude during the interviews but also, that they have to bring a
judgment of the information given by the interviewees.
The second fact that could be bring out to explain the
difficulty in that case to address the vertical dimension
is the analysts position. In this analysis, the analysts
were closed to the analysis sponsor. But, as analysts
can be stopped in their investigation by implicit stop
rules (Hopkins, 2003), and as the organisation culture and/or analysts interests could have effects on the
analysis results, the closer of the organisation the analysts are, the more difficult will be for them to address
the vertical dimension.
However these difficulties to address the vertical
dimension are not blocking points: in this case study,
even if this dimension is not treated with the extent
it should have been, it is nevertheless explored sufficiently so that, in the end, the analysts were able to
give a thick description of the event, and then a synthesis, which have been accepted by the people who
ordered the study.

was originally chosen because of its impact on the

Availability records of the plant.
We have checked these assumptions by carrying out
an Organizational Analysis of a real case occurred in a
power plant. This is discuss in the following chapters.

5
5.1

MAIN FINDINGS REGARDING. . .

. . . Organizational analysis method

5.1.1 The three main dimensions

The event we studied was chosen for Availability reasons, but it appeared very early in the analysis that
it had also a Safety dimension. This event involved
different technical teams of the power plant where it
occurred, but also several engineering divisions.
The Organization Analysis methodology requires
to address three dimensions: a historical one, a crossfunctional one, and the vertical relationships in the
organisation.
The organizational network
We have interviewed people working in each of
the concerned entities and involved in the related
event. We also succeeded in recovering official and
un-official (e-mails, power point presentations) data
written by members of these entities.
The combination of interviews and the study of the
written documents allowed us to draw the organisational network (Pierlot et al., 2007), describing how
all these entities really worked together on the issues
related to this event: We were able then to address the
cross functional dimension.
The historical dimension
It was easily possible to address the historical dimension, and then, to obtain not only a chronology of
related facts on a time scale, but also to draw a three
dimension organisational network, made of the usual
cross functional plan and a time dimension, allowing then to put in obvious place the evolution of these
cross functional dimension over the years before the
event occurrence. In this sense, we could make ours
the title of one chapter of the CAIB report History as
A Cause (CAIB, 2003).

5.1.2 The thick description

The thick description (Geertz, 1998) appears to be a
added value in the analysis process as in the restitution
phase. It allows the analysts to clearly formulate their
assumptions, to make sure that all the pieces of the
puzzle fit the overall picture. In this way, this thick
description is elaborated by a succession of tries and
errors, until the general sense of the studied issue is
made clear to the analysts.
The first attempts to build this thick description
were not made at the end of the analysis but during
the analysis. As a result, we can say that the analysis is
over when the collection of new data doesnt change
anymore this thick description, when these new data
dont change the picture given in the description.
This thick description is a necessary step in the analysis process, but it isnt the final result of the analysis.
As Pierlot mentioned it (Pierlot, 2006), there is a need
to establish a synthesis of this thick description so that
general learning could be deduced from the analysis.
Going from the thick description to this synthesis
requires the analysts to bring a judgment so that the

The vertical dimension

The study of the third dimension, the vertical one,
was more difficult to address. We could give several
explanations of this fact.
The first one is due to the position of the managers.
As organisational analysis has for purpose to point out
the organisational dysfunctions of a given organisation, as one of the managers missions is to make sure
that the organisation works smoothly, he/she could
have had a protection reaction that may explain why

67

also an impact on Availability. In that way, we can

say then that we are dealing with POFs having a
potential impact on the overall results of a complex
high-risk system because these factors could lead to
a decrease of the results on the safety scale or on the
availability scale. This is more specifically the case
for the factors Poor handling of the organisational
complexity and failures of operational feedback
system.
However, the performed analysis was focused on
the event occurrence. The analysis did not have for
purpose to look at the way the organisation recovered
from the situation. If it had, we can make the realistic
assumption that the Safety oriented POFs might not
be useful to study this issue. It is then a possibility that
there are also some pathogenic organisational factors
that could describe the different dysfunctions ways for
an organisation to recover from an Availability (and
perhaps Safety) event.

main conclusions of the analysis could be understood

and accepted by the concerned organisation and its
managers.
In this sense, the writing of the synthesis is an
important part of the analysis that may use, in a certain way, the same narrative techniques that are used
by the storytelling approaches although the goals to
be fulfilled are dramatically different: to give a general
sense to the collected data in the case of the organisational analysis; to meet propaganda purposes in the
case of the pure storytelling (Salmon, 2007).
During the building of the thick description and the
writing of the synthesis, a specific focus has to be
paid to the status that has to be given to the collected
data. As stated above, the data come from two different paths: written documents and specific information
given by interviewees.
5.1.3 The collected data status
If the reality of written documents cannot be discussed,
there could be always doubts on the quality, the reality
of information collected during interviews. In order
to overcome this difficulty, a written report of each
interview was established and submitted to the approbation of the interviewees. Each specific information
was also cross-checked, either with the collected written documents, or with information collected during
other interviews.
But, there is still the case where a specific information cannot be confirmed or denied by written
documents or during other interviews. What has to be
done in that case? There is no generic answer. A protective choice could be to say that, in this case, the
related information is never considered by the analysts.
This could protect the analysts against the criticism to
use un-verified (and perhaps even false) data. However, such a solution could deprive the analysts of
interesting and valuable materials.
We believe then, that the use of single subjective
data relies only on the analysts judgment. If the data
fits the overall picture, if it fits the general idea that the
analysts have of the situation, it can be used. However,
if this data doesnt cope with the frame of the analysis
results, the best solution is to avoid using it.
5.2

CONCLUSIONS

The performed Organisational Analysis on the occurrence of an Availability event confirms that the method
designed for a Safety purpose can be used for an event
dealing with Availability.
The three main dimensions approach (historical,
cross-functional, and vertical), the data collection and
checking (written data as information given by interviewees), the use of the thick description allow to
perform the analysis of an Availability event and to
give a general sense to all the collected data.
The knowledge of the Organisational Analysis
of many different safety events, of the different
Safety oriented Pathogenic Organisational Factors is a
required background for the analysts so that they could
know what must be looked for and where it must be
done, or which interpretation could be done from the
collected data.
This case confirms also that the usual difficulties encountered during the Organisational Analysis
of a Safety event are also present for the Organisational Analysis of an Availability event : the vertical
dimension is more difficult to address, the question
of single data is a tough issue that could deserve
deepener thoughts.
The performed analysis proves also that most of
the Safety oriented Pathogenic Organisational Factors
could be also seen as Availability oriented Pathogenic
Organisation Factors. However, these factors are
focused only on the event occurrence; they do not
intend to deal with the organisation capacity to recover
from the availability event. We believe that this
particular point must also be studied more carefully.
The last point is that an Availability event
Organisational Analysis must be performed with a

. . . The FOPs

We remind that the performed analysis was launched

due to Availability reasons, but that it had also a Safety
dimension.
Let us recall also that most of the Safety POFs
(Pierlot et al., 2007) can be seen also, from a theoretical point of view, as Availability oriented Pathogenic
Organisational Factors (Voirin et al., 2007).
The performed analysis confirms this theoretical
approach: most of the Safety oriented POFs have

68

close look on safety issues, and more specifically,

on the balance between Safety and Availability. Here
again, we believe that the complex relationships
between Availability and Safety deserves a closer look
and that some research based on field studies should
be carried out.
Eventually this case reinforces what we foresaw,
i.e. if an organization is malfunctioning, impacts can
be either on Safety or on Availability (or both) and
so, a method designed and used for a Safety purpose
can also be used, to a large extent, for Availability
purpose.

Llory, M., Dien. Y. 20062007. Les systmes sociotechniques risques : une ncessaire distinction entre
fiabilit et scurit. Performances. Issue No. 30 (SeptOct
2006); 31 (NovDec 2006); 32 (Janvfev 2007)
Perrow, C. (ed.) 1984. Normal Accidents. Living with HighRisk Technology. New York: Basic Books.
Pierlot, S. 2006. Risques industriels et scurit: les
organisations en question. Proc. Premier Sminaire de
SaintAndr. 2627 Septembre 2006, 1935.
Pierlot, S., Dien, Y., Llory M. 2007. From organizational
factors to an organizational diagnosis of the safety. Proceedings, European Safety and Reliability conference, T.
Aven & J.E. Vinnem, Eds., Taylor and Francis Group,
London, UK, Vol. 2, 13291335.
Reason, J (Ed.). 1997. Managing the Risks of Organizational
Accidents. Aldershot: Ashgate Publishing Limited.
Roberts, K. (Ed.). 1993. New challenges to Understanding Organizations. New York: Macmillan Publishing
Company.
Sagan, S. (Ed.). 1993. The Limits of Safety: Organizations,
Accidents and Nuclear Weapons. Princeton: Princeton
University Press.
Sagan, S. 1994. Toward a Political Theory of Organizational Reliability. Journal of Contingencies and Crisis
Management, Vol. 2, No. 4: 228240.
Salmon, C. (Ed.). 2007. Storytelling, la machine fabriquer
des histoires et formater les esprits. Paris: ditions La
Dcouverte.
Turner, B. (Ed.). 1978. Man-Made Disasters. London:
Wykeham Publications.
U.S. Chemical Safety and Hazard Investigation Board. 2007.
Investigation Report, Refinery Explosion and Fire, BP
Texas City, Texas, March 23, 2005, Report No. 2005-04I-TX.
Vaughan, D. (Ed.). 1996. The Challenger Launch Decision. Risky Technology, Culture, and Deviance at NASA.
Chicago: The Chicago University Press.
Vaughan, D. 1997. The Trickle-Down Effect: Policy Decisions, Risky Work, and the Challenger Tragedy. California Management Review, Vol. 39, No. 2, 80102.
Vaughan, D. 1999. The Dark Side of Organizations: Mistake,
Misconduct, and Disaster. Annual Review of Sociology,
vol. 25, 271305.
Vaughan, D. 2005. System Effects: On Slippery Slopes,
Repeating Negative Patterns, and Learning from Mistake, In: Starbuck W., Farjoun M. (Ed.), Organization at
the Limit. Lessons from the Columbia Disaster. Oxford:
Blackwell Publishing Ltd.
Voirin, M., Pierlot, S. & Llory, M. 2007. Availability
organisational analysis: is it hazard for safety? Proc.
33rd ESREDA SeminarFuture challenges of accident
investigation, Ispra, 1314 November 2007.

REFERENCES
Columbia Accident Investigation Board 2003. Columbia
Accident Investigation Board. Report Volume 1.
Cullen, W. D. [Lord] 2000. The Ladbroke Grove Rail Inquiry,
Part 1 Report. Norwich: HSE Books, Her Majestys
Stationery Office.
Cullen, W. D. [Lord] 2001. The Ladbroke Grove Rail Inquiry,
Part 2 Report. Norwich: HSE Books, Her Majestys
Stationery Office.
Dien, Y. 2006. Les facteurs organisationnels des accidents industriels, In: Magne, L. et Vasseur, D. (Ed.),
Risques industrielsComplexit, incertitude et dcision:
une approche interdisciplinaire, 133174. Paris: ditions
TED & DOC, Lavoisier.
Dien, Y., Llory, M., Pierlot, S. 2006. Scurit et performance: antagonisme ou harmonie? Ce que nous apprennent les accidents industriels. Proc. Congrs 15Lille,
October 2006.
Dien, Y., Llory M. 2006. Mthode danalyse et de diagnostic
organisationnel de la sret. EDF R&D internal report.
Dien, Y., Llory, M. & Pierlot, S. 2007. Laccident la raffinerie BP de Texas City (23 Mars 2005)Analyse et
premire synthse. EDF R&D internal report.
Geertz C. 1998. La description paisse. In Revue Enqute.
La Description, Vol. 1, 73105. Marseille: Editions
Parenthses.
Hollnagel, E., Woods, D.D., et Leveson, N.G. (Ed.) 2006.
Resilience Engineering: Concepts and Precepts. Aldershot: Ashgate Publishing Limited.
Hopkins, A. 2003. Lessons from Longford. The Esso Gas
Plant Explosion. CCH Australia Limited, Sydney, 7th
Edition (1st edition 2000).
Llory, M. 1998. Ce que nous apprennent les accidents industriels. Revue Gnrale Nuclaire. Vol. 1, janvier-fvrier,
6368.

69

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Thermal explosion analysis of methyl ethyl ketone peroxide by

non-isothermal and isothermal calorimetry application
S.H. Wu
Graduate School of Engineering Science and Technology, National Yunlin University of Science and Technology,
Douliou, Yunlin, Taiwan, ROC

J.M. Tseng
Graduate School of Engineering Science and Technology, National Yunlin University of Science
and Technology, Douliou, Yunlin, Taiwan, ROC

C.M. Shu
Department of Safety, Health, and Environmental Engineering, National Yunlin University of Science
and Technology, Douliou, Yunlin, Taiwan, ROC

ABSTRACT: In the past, process accidents incurred by Organic Peroxides (OPs) that involved near miss,
over-pressure, runaway reaction, thermal explosion, and so on occurred because of poor training, human error,
incorrect kinetic assumptions, insufficient change management, inadequate chemical knowledge, and so on, in
the manufacturing process. Calorimetric applications were employed broadly to test small-scale organic peroxides
materials because of its thermal hazards, such as exothermic behavior and self-accelerating decomposition in the
laboratory. In essence, methyl ethyl ketone peroxide (MEKPO) has a highly reactive and unstable exothermal
feature. In recent years, it has many thermal explosions and runaway reaction accidents in the manufacturing
process. Differential Scanning Calorimetry (DSC), Vent Sizing Package 2 (VSP2), and thermal activity monitor
(TAM) were employed to analyze thermokinetic parameters and safety index and to facilitate various auto-alarm
equipment, such as over-pressure, over-temperature, hazardous materials leak, etc., during a whole spectrum of
operations. Results indicated that MEKPO decomposed at lower temperature (3040 C) and was exposed on
exponential development. Time to Maximum Rate (TMR), self-accelerating decomposition temperature (SADT),
maximum of temperature (Tmax ), exothermic onset temperature (T0 ), and heat of decomposition (Hd ) etc.,
were necessary and mandatory to discover early-stage runaway reactions effectively for industries.

peroxide (BPO), hydrogen peroxide (H2 O2 ), lauroyl

peroxide (LPO), tert-butyl perbenzoate (TBPBZ), tertbutyl peroxybenzoate (TBPB), etc. are usually applied
as initiators and cross-linking agents for polymerization reactions. One reason for accidents involves the
peroxy group (-O-O-) of organic peroxides, due to
its thermal instability and high sensitivity for thermal
sources. The calorimetric technique is applied to evaluate the fundamental exothermic behavior of smallscale reactive materials during the stage of research
and development (R&D). Many thermal explosions
and runaway reactions have been caused globally by
MEKPO resulting in a large number of injuries and
even death, as shown in Table 1 (Yeh et al., 2003;
Chang et al., 2006; Tseng et al., 2006; Tseng et al.,
2007; MHIDAS, 2006). Major Hazard Incident Data

INTRODUCTION

The behavior of thermal explosions or runaway reactions has been widely studied for many years. A reactor
with an exothermic reaction is susceptible to accumulating energy and temperature when the heat generation rate exceeds the heat removal rate by Semenov
theory (Semenov, 1984). Unsafe actions and behaviors by operators, such as poor training, human error,
incorrect kinetic assumptions, insufficient change
management, inadequate chemical knowledge etc.,
lead to runaway reactions, thermal explosions, and
release of toxic chemicals, as have sporadically
occurred in industrial processes (Smith, 1982). Methyl
ethyl ketone peroxide (MEKPO), cumene hydroperoxide (CHP), di-tert-butyl peroxide (DTBP), tert-butyl

71

Table 1. Thermal explosion accidents caused by MEKPO

globally.
Year

Nation

Frequency I

19531978
19802004
19842001
2000
19731986
1962

Japan
14
China
14
Taiwan
5
Korea
1
Australia 2
1
UK

Worst case

115 23 Tokyo
13 14 Honan
156 55 Taipei
11 3 Yosu
0 0 NA
0 0 NA

Figure 1 (a). The aftermath

of the Yung-Hsin explosion
which devastated the entire
plant, including all buildings
within 100 m.

I: Injuries; F: Fatalities.
Data from MHIDAS (MHIDAS, 2006).
NA: Not applicable.

Methyl
ethyl
ketone

Service (MHIDAS) was employed to investigate three

accidents for MEKPO in Australia and UK (MHIDAS,
2006).
A thermal explosion and runaway reaction of MEKPO
occurred at Taoyuan County (the so-called YungHsin explosion) that killed 10 people and injured 47
in Taiwan in 1996. Figures 1 (a) and 1 (b) show
the accident damage situation from the Institute of
Occupational Safety and Health in Taiwan.
Accident development was investigated by a report
from the High Court in Taiwan. Unsafe actions (wrong
dosing, dosing too rapidly, errors in operation, cooling
failure) caused an exothermic reaction and the first
thermal explosion of MEKPO.
Simultaneously, a great deal of explosion pressure
led to the top of the tank bursting and the hot concrete
being broken and shot to the 10-ton hydrogen peroxide (H2 O2 ) storage tank (d = 1, h = 3). Under this
circumstance, the 10-ton H2 O2 storage tank incurred
the second explosion and conflagration that caused
10 people to perish (including employees and fire
fighters) and 47 injuries. Many plants near Yung-Hsin
Co. were also affected by the conflagration caused
by the H2 O2 tank. H2 O2 , dimethyl phthalate (DMP),
and methyl ethyl ketone (MEK) were applied to manufacture the MEKPO product. Figure 2 depicts the
MEKPO manufacture process flowchart. To prevent
any further such casualties, the aim of this study was
to simulate an emergency response process. Differential scanning calorimetry (DSC), vent sizing package
2 (VSP2), and thermal activity monitor (TAM) were
employed to integrate thermal hazard development.
Due to MEKPO decomposing at low temperature
(3040 C) (Yuan et al., 2005; Fu et al., 2003) and
exploding with exponential development, developing
or creating an adequate emergency response procedure
is very important.
The safety parameters, such as temperature of no
return (TNR ), and time to maximum rate (TMR),
self-accelerating decomposition temperature (SADT),
maximum temperature (Tmax ), etc., are necessary and
useful for studying emergency response procedures

Figure 1 (b). Reactor

bursts incurred by
thermal runaway.

Dimethyl
phthalate
(DMP)

Hydrogen
peroxide
(H2O2)

Reactor

H3PO4

Crystallization
tank

Cooling
to lower
than
10C

Dehydration tank

Drying tank

MEKPO in storage
tank

Figure 2.
Taiwan.

Typical manufacturing process for MEKPO in

in terms of industrial applications. In view of loss

prevention, emergency response plan is mandatory
and necessary for corporations to cope with reactive
chemicals under upset scenarios.

EXPERIMENTAL SETUP

2.1 Fundamental thermal detection by DSC

DSC has been employed widely for evaluating thermal hazards (Ando et al., 1991; ASTM, 1976) in
various industries. It is easy to operate, gives quantitative results, and provides information on sensitivity
(exothermic onset temperature, T0 ) and severity (heat
of decomposition, Hd ) at the same time. DSC was

72

Table 2.

Thermokinetics and safety parameters of 31 mass% MEKPO and 20 mass% H2 O2 by DSC under at 4 C min1 .

Mass (mg) T1 ( C) Hd (J g1 ) T2 ( C) Tmax ( C) Hd (J g1 ) T3 ( C) Tmax ( C) Hd (J g1 ) Htotal (J g1 )

1
2

4.00
2.47

42

41

83
67

135
100

304
395

160

200

768

1,113
395

1 MEKPO; 2 H O .
2 2

applied to detect the fundamental exothermic behavior of 31 mass% MEKPO in DMP that was purchased
directly from the Fluka Co. Density was measured and
provided directly from the Fluka Co. ca. 1.025 g cm3 .
It was, in turn, stored in a refrigerator at 4 C (Liao
et al., 2006; Fessas et al., 2005; Miyakel et al., 2005;
Marti et al., 2004; Sivapirakasam et al., 2004; Hou
et al., 2001). DSC is regarded as a useful tool for the
evaluation of thermal hazards and for the investigation
of decomposition mechanisms of reactive chemicals if
the experiments are carried out carefully.

2.5
20 mass% H2O2
31 mass% MEKPO

Heat flow (W g-1)

2.0
1.5
1.0
0.5
0.0
0

2.2

Adiabatic tests by VSP2

50

100

150

200

250

300

Temperature (C)

VSP2, a PC-controlled adiabatic calorimeter, manufactured by Fauske & Associates, Inc. (Wang et al.,
2001), was applied to obtain thermokinetic and thermal hazard data, such as temperature and pressure
traces versus time. The low heat capacity of the
cell ensured that almost all the reaction heat that
was released remained within the tested sample.
Thermokinetic and pressure behavior in the same test
cell (112 mL) usually could be tested, without any
difficult extrapolation to the process scale due to a
low thermal inertia factor () of about 1.05 and 1.32
(Chervin & Bodman, 2003). The low  allows for
bench scale simulation of the worst credible case, such
as incorrect dosing, cooling failure, or external fire
conditions. In addition, to avoid bursting the test cell
and missing all the exothermic data, the VSP2 tests
were run with low concentration or smaller amount
of reactants. VSP2 was used to evaluate the essential
thermokinetics for 20 mass% MEKPO and 20 mass%
H2 O2 . The standard operating procedure was repeated
by automatic heat-wait-search (HWS) mode.

Figure 3. Heat flow vs. temperature of 31 mass% and 20

mass% H2 O2 under 4 C min1 heating rate by DSC.

RESULTS AND DISCUSSION

3.1 Thermal decomposition analysis of 31 mass%

MEKPO and 20 mass% H2 O2 for DSC
Table 2 summarizes the thermodynamic data by the
DSC STAR e program for the runaway assessment.
MEKPO could decompose slowly at 3032 C, as
disclosed by our previous study for monomer. We
surveyed MEKPO decomposing at 30 C, shown in
Figure 3. We used various scanning rates by DSC to
survey the initial decomposition circumstances. As a
result, a quick temperature rise may cause violent initial decomposition (the first peak) of MEKPO under
external fire condition. Table 2 shows the emergency
index, such as T0 , H, Tmax , of 31 mass% MEKPO
by DSC under heating rate () 4 C min1 . The initial decomposition peak usually releases little thermal
energy, so it is often disregarded. The T2 of mainly
decomposition was about 83 C. The total heat of reaction (Htotal ) was about 1,113 J g1 . DSC was applied
to evaluate the Ea and frequency factor (A). The Ea
under DSC dynamic test was about 168 kJ mol1 and
A was about 3.5 1019 (s1 ).
Figure 3 displays the exothermic reaction of 20
mass% H2 O2 under 4 C min1 of by the DSC.
Due to H2 O2 being a highly reactive chemical, operators must carefully control flow and temperature.
H2 O2 was controlled at 10 C when it precipitated the

2.3 Isothermal tests by TAM

Isothermal microcalorimetry (TAM) represents a
range of products for thermal measurements manufactured by Thermometric in Sweden. Reactions can be
investigated within 1290 C, the working temperature
range of this thermostat. The spear-head products are
highly sensitive microcalorimeters for stability testing of various types of materials (19). Measurements
were conducted isothermally in the temperature range
from 60 to 80 C (The Isothermal Calorimetric Manual,
1998).

73

Table 3.

1
2
1

Thermokinetic parameters of MEKPO and H2 O2 from adiabatic tests of VSP2.

T0 ( C)

Tmax ( C)

Pmax (psi)

(dT dt1 )max ( C min1 )

(dP dt1 )max (psi min1 )

Tad ( C)

Ea(kJ mol1 )

90
80

263
158

530
841

394.7
7.8

140.0
100.0

178
78

97.54
NA

MEKPO; 2 H2 O2 .

280

300

220

200

200

100

180

Tmax = 158C

dT dt-1 (C min -1)

Temperature (C)

240

160
140
120
100
80
60

20 mass% MEKPO
20 mass% H2O2

40
20
0

50

100

150

200

(dT dt-1)max = 395C min -1

400

Tmax = 263C

260

20 mass% MEKPO

-3.6

-3.2

-3.0

Pmax = 841 psi

160
140
120
100
80
60
40
20
0
-20

Pressu re (psig)

700
Pmax = 530 psi

600
500
20 mass% MEKPO
20 mass% H2O2

160 -3.6
140
120
100
80
60
40
20
0
-20
-3.6

200
100
0
-100
0

50

100

150

-2.2

-2.0

-1.8

-3.4

-3.2

-3.0

-2.8

-2.6

-2.4

-2.2

-2.0

-1.8

Figure 6. Dependence of rate of temperature rise on temperature from VSP2 experimental data for 20 mass% MEKPO
and H2 O2 .

800

300

-2.4

-1000/T (K-1)

Figure 4. Temperature vs. time for thermal decomposition

of 20 mass% MEKPO and H2 O2 by VSP2.

400

-2.6

20 mass H2O2

Time (min)

900

-2.8

(dT dt-1)max = 7C min -1

-3.6

250

-3.4

14
12
10
8
6
4
2
0

200

250

Time (min)

Figure 5. Pressure vs. time for thermal decomposition of

20 mass% MEKPO and H2 O2 by VSP2.

20 mass% MEKPO

-3.4

-3.2

-3.0

-2.8

-2.6

-2.4

-2.2

-2.0

-1.8

-2.8

-2.6

-2.4

-2.2

-2.0

-1.8

max

20 mass H2O2

-3.4

-3.2

-3.0

Figure 7. Dependence of rate of pressure rise on temperature from VSP2 experimental data for 20 mass% MEKPO
and H2 O2 .

reaction. H2 O2 exothermic decomposition hazards are

shown in Table 2.
mass% MEKPO by VSP2 was more dangerous in
Figure 4 because it began selfheating at 90 C and
Tmax reached 263 C. H2 O2 decomposed at 80 C and
increased to 158 C. In parallel, H2 O2 decomposed
suddenly and was transferred to vapor pressure. Pmax
was detected from Figure 5. Figures 6 and 7 show
the rate of temperature increase in temperature and
pressure from VSP2 experimental data for 20 mass%
MEKPO and H2 O2 . According to the maximum of

3.2 Thermal decomposition analysis for VSP2

Table 3 indicates T0 , Tmax , Pmax , (dT dt1 )max ,
(dP dt1 )max for 20 mass% MEKPO and H2 O2 by
VSP2. Figure 4 displays temperature versus time
by VSP2 with 20 mass% MEKPO and H2 O2 . The
pressure versus time by VSP2 under 20 mass%
MEKPO and H2 O2 is displayed in Figure 5. Twenty

74

were about 263 C and 34.7 bar, respectively. Under 20

mass% MEKPO by VSP2 test, the (dT dt1 )max and
(dP dt1 )max were about 394.7 C min1 and 140.0 bar
min1 , respectively. During storage and transportation, a low concentration (<40 mass%) and a small
amount of MEKPO should be controlled. H2 O2 was
controlled 10 C when it joined a MEKPO manufacturing reaction. This is very dangerous for the MEKPO
manufacturing process, so the reaction was a concern
and controlled at less than 20 C in the reactor.

Table 4. Scanning data of the thermal runaway decomposition of 31 mass% MEKPO by TAM.
Isothermal
temperature
( C)

Mass
(mg)

Reaction time
(hr)

TMR
(hr)

Hd
(Jg1 )

60
70
80

0.05008
0.05100
0.05224

800
300
250

200
80
40

784
915
1,015

0.00015

ACKNOWLEDGMENT

0.00010

Isothermal test for 60 C

Heat flux (Wg-1)

0.00005

The authors would like to thank Dr. Kuo-Ming Luo for

valuable suggestions on experiments and the measurements of a runaway reaction. In addition, the authors
are grateful for professional operating techniques and
information.

0.00000
0.00015 0

200

400

600

0.00010

800

1000

1200

Isothermal test for 70 C

0.00005
0.00000
0.00015

50

100

150

0.00010

200

250

300

350

REFERENCES

Isothermal test for 80 C

0.00005
0.00000
0

50

100

150

200

250

300

Ando, T., Fujimoto, Y., & Morisaki, S. 1991. J. of Hazard.

Mater., 28: 251280.
ASTM E537-76, 1976. Thermal Stability of Chemical by
Methods of Differential Thermal Analysis.
Chervin, S. & Bodman, G.T. 2003. Process Saf. Prog, 22:
241243.
Chang, R.H., Tseng, J.M., Jehng, J.M., Shu, C.M. & Hou,
H.Y. 2006. J. Therm. Anal. Calorim., 83, 1: 5762.
Fu, Z.M., Li, X.R., Koseki, H.K., & Mok, Y.S. 2003. J. Loss
Prev. Process Ind., 16: 389393. Semenov, N.N. 1984. Z.
Phys. Chem., 48: 571.
Fessas, D. Signorelli, M. & Schiraldi, A. 2005. J. Therm.
Anal. Cal., 82: 691.
Hou, H.Y., Shu, C.M. & Duh, Y.S. 2001. AIChE J., 47(8):
18936.
Liao, C.C., Wu, S.H., Su, T.S., Shyu, M.L. & Shu, C.M.
2006. J. Therm. Anal. Cal., 85: 65.
Marti, E., Kaisersberger, E. & Emmerich, W.D. 2004. J.
Therm. Anal. Cal., 77: 905.
Miyakel, A., Kimura, A., Ogawa, T., Satoh, Y. & Inano, M.
2005. J. Therm. Anal. Cal., 80: 515.
MHIDAS, 2006. Mayor Hazard Incident Data Service,
OHS_ROM, Reference Manual.
Sivapirakasam, S.P. Surianarayanan, M. Chandrasekaran, F.
& Swaminathan, G. 2004. J. Therm. Anal. Cal., 78: 799.
Smith, D.W. 1982. Runaway Reaction, and Thermal Explosion, Chemical Engineering, 13: 7984.
The Isothermal Calorimetric Manual for Thermometric AB,
1998, Jarfalla, Sweden.
Tseng, J.M., Chang, R.H., Horng, J.J., Chang, M.K., &
Shu, C.M. 2006. ibid, 85 (1) : 189194.
Tseng, J.M., Chang, Y.Y., Su, T.S., & Shu, C.M. 2007. J.
Hazard. Mater., 142: 765770.
Wang, W.Y., Shu, C.M., Duh, Y.S. & Kao, C.S. 2001. Ind.
Eng. Chem. Res., 40: 1125.
Yeh, P.Y., Duh, Y.S., & Shu, C.M. 2003. Ind. Eng. Chem.
Res., 43: 1 5.
Yuan, M.H., Shu, C.H., & Kossoy, A.A. 2005. Thermochimica Acta,430: 6771.

350

Time (hr)

Figure 8. Thermal curves for 31 mass% MEKPO by TAM

at 6080 C isothermal temperature.

temperature rise rate ((dT dt1 )max ) and temperature

rise rate maximum of pressure rise rate ((dP dt1 )max )
from Figures 6 and 7, MEKPO is more dangerous
than H2 O2 . According to Figure 2, the MEKPO and
H2 O2 were the Ops that may cause a violent reactivity
behavior in the batch reactor.
3.3 Thermal decomposition analysis of 31
mass% MEKPO for TAM
TAM was employed to determine the reaction behavior for isothermal circumstance. Table 4 displays
the thermal runaway decomposition of 31 mass%
MEKPO under various isothermal environment by
TAM. Results showed the TMR for isothermal temperature at 60, 70, and 80 C that were about 200,
80, and 40 hrs, respectively. From data investigations,
isothermal temperature was rise, and then emergency
response time was short.

CONCLUSIONS

According to the DSC experimental data, MEKPO

decomposes at 3040 C. Under external fire
circumstances, MEKPO can decompose quickly and
cause a runaway reaction and thermal explosion. On
the other hand, for 20 mass% MEKPO by VSP2,
the maximum temperature (Tmax ) and pressure (Pmax )

75

Crisis and emergency management

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A mathematical model for risk analysis of disaster chains

A. Xuewei Ji
Center for Public Safety Research, Tsinghua University, Beijing, China
School of Aerospace, Tsinghua University, Beijing, China

B. Wenguo Weng & Pan Li

Center for Public Safety Research, Tsinghua University, Beijing, China

ABSTRACT: A disaster often causes a series of derivative disasters. The spreading of disasters can often form
disaster chains. A mathematical model for risk analysis of disaster chains is proposed in causality network,
in which each node represent one disaster and the arc represent interdependence between them. The nodes of
the network are classified into two categories, active nodes and passive nodes. The term inoperability risk,
expressing the possible consequences of each passive node due to the influences of all other nodes, is cited
to assess the risk of disaster chains. An example which may occur in real life is solved to show how to apply
the mathematical model. The results show that the model can describe the interaction and interdependence of
mutually affecting emergencies, and it also can estimate the risk of disaster chains.

of chain of losses and failures to the simple couple

hazardous event damages. Glickman et al. proposed
hazard networks to describe the interdependencies
among the hazards and determine risk reduction strategies. Some studies about risk assessment of domino
effect based physical models and quantitative area risk
analysis have been made great progress as for instance
Khan et al. and Spadoni et al.
The purpose of the paper is to present a mathematical model in causality network allowing to study the
interaction and interdependence of mutually affecting
among disasters and to conduct risk analysis of disaster
chains.

INTRODUCTION

Nowadays, on account of increasing congestion in

industrial complexes and increasing density of human
population around such complexes, any disaster is not
isolated and static, and many disasters are interactional
and interdependent. A disaster often causes a series of
derivative disasters. In many cases, derivative disasters
can cause greater losses than the original disaster itself.
The spreading of disasters can often be described by
interconnected causality chain, which is called disaster chains here. Disaster chains reflect how one factor
or sector of a system affects others. These cascadelike chain-reactions, by which a localized event in time
and space causes a large-scale disaster, may affect the
whole system. The recent example is the snowstorm
in several provinces of southern China this year. The
rare snowstorm made highways, railways, and shipping lines blocked leading to the poor transportation
of coal, which brought the strain of electricity in many
regions. All of these formed typical disaster chains.
Therefore, a great effort is necessary to prevent the
chain effect of disasters and to improve the ability
of emergency management. Researchers have made
some contribution to understand the domino effect and
chain effect behind disasters. For example, Gitis et al.
studied catastrophe chains and hazard assessment,
and they suggested a mathematical model describing mutually affecting catastrophes. Menoni et al.
analyzed the direct and indirect effects of the Kobe
earthquake and suggested to substitute the concept

MATHEMATIC MODEL

2.1 Some concepts and assumptions

Disaster chains mean that one critical situation triggers
another one and so on, so that the situation worsens
even more. It is a complex system which consists of
interacting disasters. All the disasters forms a causality
network, in which each node represent one disaster
and the arc represent interdependence between them.
In this section, a mathematical model is set up based
on the causality network (see Figure 1) for risk analysis
of disaster chains.
In the causality network above, we classify the
nodes of the network into two categories, active nodes
and passive nodes. The term active nodes will be used

79

Figure 1.

For each active node j = 1, 2, . . ., m, a random

variable i is used to represent the state, which takes
values 1 or 0 provided the node occurs or not. For each
passive node k(k = m +1, m+ 2, . . . , n) inoperability
is cited to represent its state, which is assumed to be a
continuous variable (k ) evaluated between 0 and 1,
with 0 corresponding to a normal state and 1 corresponding to completely destroyed state. We shall
consider conditional distribution of i and postulate
the Markov property: this distribution depends only on
the influence of the direct connected nodes. We postulate the following expressions for these conditional
probabilities:


P(j = 1|i = xi , k = yk ) = j +
xi aij +
yk akj

Causality network of disaster chains.

to refer to the nodes whose consequences (loss or damage) are embodied through passive nodes, for example,
the loss of earthquake comes from the damage of some
anthropogenic structures caused by it. Likewise, passive nodes will be used to refer to the nodes whose
consequences are possible loss events and are embodied by themselves. Taking power system as an example, the disaster is the consequence of inability of part
of its functions (called inoperability here). Risk is the
integration of possibility and consequence. Therefore,
the inoperability risk proposed by Haimes et al. which
expresses the possible consequences of each passive
node due to the influences of other nodes (including
passive and active), is cited to assess the risk of disaster
chains.
The proposed model is based on an assumption that
the state of active node is binary. That means that
active node has only two states (occurrence of nonoccurrence). The magnitude of active node (disaster) is
not taken into consideration in the model. In contrast
to that, the state of passive nodes can be in continuous states from normal state to completely destroyed
state.

j +

m


xi aij +

i=1

Pj = j +

m


Pi aij +

i=1

(3)

n


Rk akj

(4)

k=m+1
m

i=1

j = 1, 2, . . . , m

yk akj 1

Equation (2) means that the conditional probability

of the event i = 1 with given i = xi , k = yk is a linear combination of xi plus constant. Only the nodes i
directly connected with the nodes j are involved in this
linear combination. This hypothesis forms the basis of
the mathematical theory of the voter model (Malyshev
et al.). Let us denote by Pi the unconditional probability P(i = 1). It is interpreted as the probability
of the occurrence of the active node j. If we take the
expectation value of (2), we get the equations:

Let us define certain active node with certain occurrence probability as the start point of disaster chains.
Let us index all nodes i = 1, 2, . . . , m, m + 1, m +
2, . . . , n. The nodes represent active nodes if i =
1, 2, . . . , n, and passive nodes if i = m + 1, m +
2, . . . , n. We shall denote by aij the probability for any
node i to induce directly the active node j, and aij > 0
in case of the existence of an arrow from node i to node
j, otherwise aij = 0. By definition, we put aii = 0. So
we get the matrix:
i = 1, 2, . . . , n,

n

k=m+1

The model

A = (aij ),

(2)

where i denotes the probability of spontaneous

occurrence of the active node j. In the above equation,
it is supposed that

Pj = min j +
2.2

i, j = 1, . . . , m, k = m + 1, . . . , n

Pi aij +

n



Rk akj , 1

(5)

k=m+1

where Rk denotes the inoperability of the passive

node k. Next, we use the quantity bjk to represent
the inoperability of each passive node k caused by
active node i. So we have the inoperability (Ck ) of passive node k due to all the active nodes that are directly
connected with the passive node:
Ck =

m


Pi bik

(6)

i=1

To assess the risk of disaster chains, we propose a quantity Mij called direct influence coefficient.

(1)

80


m
n

Pi aij +
Rk akj
Pj = j +

i=1
k=m+1

m
n

Pi bki +
Ri Mik
Rk =

i=m+1
i=1

s.t.


m
n

P = min +

P
a
+
R
a
,
1

j
j
i
ij
k
kj

k=m+1

m i=1

Rk = min
Pi bki +
Ri Mik , 1

i=1
i=m+1

for j = 1, 2, . . . , m, k = m + 1, m + 2 . . . , n

It expresses the influence of inoperability between passive i and j, which are directly connected. We get direct
influence coefficient matrix:
M = (Mij ),

i, j = m + 1, m + 2, . . . , n

(7)

The expression M k reflects all influences over k 1

nodes and k links, i.e. k = 1 corresponds to direct
influences, k = 2 to feedback loops with one intermediate node, etc. So, total influence coefficient matrix
(Tij ) can be expressed:
T = M + M2 + M3 + + Mk + =

(14)
Finally, in the mathematical model, we need to
determine three matrix, i.e. the A matrix, the B matrix
and the M matrix. For the matrix A, several physical
models and probability models have been developed
to give the probability that a disaster induces another
one (Salzano et al.). Extensive data collecting, data
mining and expert knowledge or experience may be
required to help determine the M matrix. We can use
historical data, empirical data and experience to give
the matrix B.

Mk

k=1

(8)
As this converges only for the existence of the
matrix (I M )1 , we will obtain the formula:
T = (I M )1 I

(9)

where I denotes the unity matrix. Besides, the total

inoperability risk of the passive node k due to the
influence of all the nodes (active and passive) can be
expressed as:
Rk =

n


Ci Tik

To show how to apply the mathematical model, we

solve the following example. We consider the disaster
chains consisting of eight nodes and earthquake is considered as spontaneous with a known probability 1 .
The causality network of the disaster chains is illustrated by Figure 2. We think that D1 D4 are active
nodes and D5 D8 are passive nodes.
In principle, the matrix A, B and M are obtained
from expert data and historical data. For simplicity,
all the elements of the three matrixes are regarded as
constant. We have:

(10)

i=m+1

Based on the above equations, we can get:

Rk = Ck +

n


Ri Mik

EXAMPLE

(11)

i=m+1

0
0

0
A=
0

0
0

That is also:
Rk =

m


Pi bki +

i=1

n


Ri Mik

(12)

i=m+1

It is important to notice that, due to the constraint

0 Rk 1, equation (12) sometimes does not have
a solution. If this is the case, we will have to solve an
alternative problem, i.e.
Rk = min

 m

i=1

Pi bki +

n


0.6
0.1
0
0
0
0
0
0

0.2
0.6

0
B = 0.9
0
0

0
0.5

0
0

0
0.3
M =
0
0
0
0

0
0
0
0

0.3
0

0
0

0
0
0
0

0
0.9
0.5
0

0.5
0
0
0.1

0
0
0.4
0.2

Now, suppose 1 = 0.3. Due to the earthquake,

fires in commercial area, hazardous materials release
and social crisis are induced, and commercial area,
public transportation system and power supply system
are impacted. We can get the results:


Ri Mik , 1

0.4
0
0.4
0
0
0
0
0.3

(13)

i=m+1

(P1 , P2 , P3 , P4 ) = (0.30, 0.33, 0.21, 0.06)

To sum up, the mathematical model for risk analysis

of disaster chains can be summarized as:

(R5 , R6 , R7 , R8 ) = (0.58, 0.23, 0.37, 0.42)

81

Thus, because of the possible influences of the

earthquake and a series of disasters (fires and hazardous materials release) induced by it, the commercial area and power plant loss 58% and 42% of each
functionality, the economic situation deteriorate 23%
of its normal state, and the transportation system loses
63% of its operability. We conclude that the interdependences make the influence of the earthquake be
amplified according to the results above. From the
viewpoint of disaster chains, the interdependences can
be taken into consideration.
We can also evaluate the impact of the earthquake
of varying occurrence probability on the risk distribution. Solving the problem can yield the following
results. When 1 is 0.52, the value of R5 reaches 1.
The value of R8 reaches 1, when 1 increases to 0.72.
This means the operability of commercial area and
economic situation is 0.55 and 0.87 at 1 = 0.72,
while the other two systems have been in completely
destroyed. When 1 increases to 0.97, all of the systems completely fail except economic situation. The
results are shown in Fig. 3.

CONCLUSIONS

The mathematical model proposed here can be used to

study the interaction and interdependence of mutually
affecting among disasters and to conduct risk analysis
of disaster chains. It can also be used to analyze which
disaster is more important compared to other disasters,
and to provide scientific basis for disaster chains emergency management. However, disaster chains are so
complex that developing a meaningful mathematical
model capable of study disaster chains is an arduous
task. The work here is only a preliminary attempt to
contribute to the gigantic efforts.
In order to construct a mathematical model to study
disaster chains in causality network, we think that the
following steps are needed:
Enumerate all possible disasters in the causality
network.
For each node of the network, define a list of
other nodes which can be directly induced by it
and directly induce it, and define the interdependent
relationships between them.
Define appropriate quantities to describe the risk of
disaster chains.
Then risk analysis process of disaster chains can be
described by the model presented above.
REFERENCES

Figure 2.

Causality network of our example.

1.0

inoperability risk

0.8

Gitis, V.G., Petrova, E.N. & Pirogov, S.A. 1994. Catastrophe

chains: hazard assessment, Natural hazards 10: 117121.
Glickmanand, T.S. & Khamooshi, H. 2005. Using hazard
networks to determine risk reduction strategies, Journal
of the operational research society 56: 12651272.
Haimes, Y.Y. & Pu Jiang. 2001. Leontief-based model of
risk in complex interconnected infrastructures, Journal of
infrastructure systems 7(1): 112.
Khan, F.I. & Abbasi, S.A. 2001. An assessment of the likelihood of occurrence, and the damage potential of domino
effect (chain of accidents) in a typical cluster of industries, Journal of loss prevention in the process industries
14: 283306.
Malyshev, V.A., Ignatyuk, I.A. & Molchanov, S.A. 1989.
Momentum closed processes with local interaction,
Selecta mathematica Sovietica 8(4): 351384.
Salzano, E., Lervolino, I. & Fabbrocino, G. 2003. Seismic
risk of atmospheric storage tanks in the framework of
quantitative risk analysis, Journal of loss prevention in
the process industries 16: 403409.
Scira Menoni. 2001. Chains of damages and failures in
a metropolitan environment: some observations on the
Kobe earthquake in 1995, Journal of hazardous materials
86: 101119.
Spadoni, G., Egidi, D. & Contini, S. 2000. Through ARIPARGIS the quantified area risk analysis supports land-use
planning activities, Journal of hazardous materials 71:
423437.

power plant
commercial area

public transportation

0.6

economic situation
0.4

0.2

0.0
0.0

0.2

0.4

0.6

0.8

1.0

Probability of earthquake

Figure 3. Inoperability risk as function of the probability of

earthquake.

82

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Effective learning from emergency responses

K. Eriksson
LUCRAM (Lund University Centre for Risk Analysis and Management),
Department of Fire Safety Engineering and Systems Safety, Lund University, Lund, Sweden

J. Borell
LUCRAM (Lund University Centre for Risk Analysis and Management),
Department of Design Sciences, Lund University, Lund, Sweden

ABSTRACT: During recent years a couple of emergencies have affected the city of Malmo and its inhabitants
and forced the city management to initiate emergency responses. These emergencies, as well as other incidents,
are situations with a great potential for learning emergency response effectiveness. There are several methods
available for evaluating responses to emergencies. However, such methods do not always use the full potential
for drawing lessons from the occurred emergency situations. Constructive use of principles or rules gained
during one experience (in this case an emergency response) in another situation is sometimes referred to as
positive transfer. The objective of this paper is to develop and demonstrate an approach for improving learning
from the evaluation of specific response experiences through strengthening transfer. The essential principle in
the suggested approach is to facilitate transfer through designing evaluation processes so that dimensions of
variation are revealed and given adequate attention.

During the last four years three larger emergencies

have affected the city of Malmo. In 2004 the Boxing Day Tsunami resulted in an evacuation of people
to Sweden and thus also to Malmo, which received a
substantial number of evacuees. In the middle of July
2006 an armed conflict between Lebanon and Israel
broke out. Also this incident resulted in an evacuation
of Swedish citizens, of whom many were connected
to Malmo. Both these incidents resulted in a need
for the city of Malmo to use their central emergency
response organisation to support the people evacuated
from another part of the world. In April 2007 a riot
broke out in a district of the city. Also during this
incident there was a need for the city to initiate their
emergency response organisation but this time only
in the affected district. After each of these incidents
Malmo has performed evaluations. A central question
is: How can organisations be trained to face future
unknown incidents based on known cases?
The objective of this paper is to develop and
demonstrate an approach for strengthening emergency response capability through improving learning
from the evaluation of specific emergency response
experiences. An approach has been developed based
on learning theories. The approach has been applied
to evaluations of emergency responses in Malmo.
The preliminary findings indicate that the developed

INTRODUCTION

During recent years there has been an increased

demand from the public that society should prevent
emergencies to occur or at least minimize their negative outcomes. The demand for a more and more robust
society also enhances the need for the society to learn
from past experience. Experience and evaluations of
instances of emergency response are two possible
inputs to an emergency management preparedness
process. Other important inputs are risk analyses,
exercises as well as experience from everyday work.
Ideally the result from an evaluation improves the
organisations ability to handle future incidents. Furthermore, experience of a real emergency situation
commonly creates awareness and willingness to prepare for future emergencies (Tierney et al. 2001; Boin
et al. 2005). It is of central interest to question whether
the full potential of an evaluation of an emergency
situation is used. Sometimes there is a tendency to
plan for the current situation (Carley & Harrald 1997).
For example Lagadec (2006 p. 489) mention that it is
essential to . . . not prepare to fight the last war but
for future emergencies. For an overview of possible
barriers against learning from experience concerning
emergency management, see for example Smith &
Elliott (2007).

83

3.2 Transfer

approach can improve experience-based learning in

organisations.

Constructive use of principles or rules that a person gained during one experience (in this case an
emergency response operation) in another situation
is sometimes referred to as positive transfer (Reber
1995). Transfer may be quite specific when two situations are similar (positive or negative transfer), but
also more general, e.g. learning how to learn. The
concept of transfer is also discussed within organisational theory. At an organisational level the concept
transfer involves transfer at an individual level but
also transfer between different individuals or organisations. Transfer at an organisational level can be defined
as . . . the process through which one unit (e.g. group,
department, or division) is affected by experience of
another (Argote & Ingram 2000 p. 151).

METHOD

From theories of learning a first hypothetical approach

for improving learning from evaluations of emergency
responses was created. To further examine and refine
the approach it was tested through application on
an evaluation of emergency response in the city of
Malmo. The evaluation concerned the management
of emergency response activities during the Lebanon
war in July 2006. In addition, evaluations of the
Boxing Day Tsunami in 2004 and the emergency management organisations handling of a riot that broke
out in a district of the Malmo during April 2007
are used for further demonstration of the approach.
The evaluations are based on analyses of interviews
and collected documents. The interviews were carried out with mainly municipal actors involved during
the incidents. In total 19 interviews have been completed. The documents analysed were for example
notes from the response organisations information
systems, notes and minutes from managerial meetings during the events, written preparedness plans
and newspaper articles. The use of the three evaluations can be seen as a first test of the approach
for improved learning from evolutions of emergency
situations.

3
3.1

3.3 Variation
One essential principle for facilitating the transfer
process, established in the literature on learning, is
to design the learning process so that the dimensions of possible variation become visible to the
learners (Pang 2003, Marton & Booth 1999). Successful transfer for strengthening future capability
demands that critical dimensions of possible variation specific for the domain of interest are considered
(Runesson 2006).
When studying an emergency scenario two different kinds of variation are possible; variation of the
parameter values and variation of the set of parameters that build up the scenario. The first kind of
variation is thus the variation of the values of the
specific parameters that build up the scenario. In practice, it is not possible to vary all possible parameter
values. A central challenge is how to know which
parameters are critical in the particular scenario and
thus worth closer examination by simulated variation of their values. The variation of parameter values
can be likened to the concept of single-loop learning
(Argyris & Schn 1996). When the value of a given
parameter in a scenario is altered, that is analogous
to when a difference between expected and obtained
outcome is detected and a change of behaviour is
made.
The second kind of variation is variation of the
set of parameters. This kind of variation may be discerned through e.g. discussing similarities as well as
dissimilarities of parameter sets between different scenarios. The variation of the set of parameters can be
likened to the concept of double-loop learning (Argyris
& Schn 1996), wherein the system itself is altered
due to an observed difference between expected
and obtained outcome. A central question is what
the possible sets of parameters in future emergency
scenarios are.

THEORY
Organisational learning

For maintaining a response capability in an organisation over time there is a need that not only separate
individuals but the entire organisation has the necessary knowledge. According to Senge (2006 p. 129)
. . . Organizations learn only through individuals who
learn. Individual learning does not guarantee organizational learning. But without it no organizational
learning occurs. Also Argyris & Schn (1996) point
out that organisational learning is when the individual members learn for the organisation. Argyris &
Schn (1996) also discuss two types of organisational
learning: single-loop learning and double-loop learning. Single-loop learning occurs when an organisation
modifies its performance due to a difference between
expected and obtained outcome, without questioning
and changing the underlying program (e.g. changes
in values, norms and objectives). If the underlying program that led to the behaviour in the first
place is questioned and the organisation modifies it,
double-loop learning has taken place.

84

3.4

4.2

Scenario

The second step is to vary the value of the parameters that build up the scenario. This may be carried out
through imagining variation of the included parameters (that are seen as relevant) within the scenario
description. Typical examples of parameters can be
the length of forewarning or the number of people
involved in an emergency response.
Variation of parameter values makes the parameters
themselves as well as the possible variation of their
values visible. This can function as a foundation for
positive transfer to future emergency situations with
similar sets of relevant parameters. This in turn may
strengthen the capability to handle future emergencies
of the same kind as the one evaluated, but with for
example greater impact.

A specific emergency situation can be described as

a scenario, here seen as a description of a series of
occurred or future events arranged along a timeline.
Scenarios describing past emergencies . . . answers
the question: What happened? (Alexander 2000
pp. 8990). An emergency scenario can further be
seen as consisting of various parameters. The concept parameter is here defined in very broad terms and
every aspect with a potential for variation in a scenario
is seen as a parameter. For example the duration of a
scenario or the quantity of recourses that is needed
can be viewed as parameters. Alexander (2000 p. 90)
further mentions that when imagining the future the
question to ask is What if . . . ?.
One kind of parameters suitable for imagined
variation that is discussed in the literature is the
different needs or problems that arise during an emergency situation. Dynes (1994) discusses two different types of needs or problems that requires to be
responded to during an emergency. These are the
response generated and the agent generated needs.
The response generated needs are quite general and
often occur during emergencies. They are results of
the particular organisational response to the emergency, e.g. the needs for communication and coordination. The agent generated needs (Dynes et al.
1981) are the needs and problems that the emergency in itself creates, for example search, rescue,
care of injured and dead as well as protection against
continuing threats. These needs tend to differ more
between emergencies than the response generated
needs do.

Variation of the values of the parameters

4.3

Variation of the set of parameters

The third step is the variation of the set of parameters.

By comparing the current case with other cases both
occurred (e.g. earlier emergencies) and imagined (e.g.
results from risk and vulnerability analyses) different
sets of parameters can be discussed.
Obviously it is not possible to decide the ultimate
set of parameters to prepare for with general validity.
There is always a need for adaption to the situation.
Yet it is often possible for an organisation to observe a
pattern of similarities in the parameters after a couple
of evaluations. Even if two emergency scenarios differ when it comes to physical characteristics there are
often similarities in the managing of the scenarios.
A possible way to vary the set of parameter is to
be inspired by Dynes (1994) different types of needs
or problems that arise during an emergency situation. Commonly similarities are found when studying
response generated needs and it is thus wise to prepare
to handle them. Even if agent generated needs tend to
differ more between emergencies, paying them some
attention too may be worthwhile.
Greater experience means more opportunities for
positive transfer. Furthermore, with increasing experience of thinking in terms of varying the set of
parameters and their values, it is probable that the
organisation and its employees also develop the ability
to more general transfer, through the abstract ability
to think of variation of parameters.

THE PROPOSED APPROACH

The main goal of this paper is to develop and

demonstrate an approach to strengthening emergency
response capability through improving learning from
the evaluation of response experiences.
4.1 Description of the emergency scenario
The first step in the approach is to construct a description of the emergency scenario, i.e. create and document a description of an occurred emergency situation.
The description of the occurred scenario is needed for
further discussions on the organisations ability to handle future emergencies. By describing the series of
events that build up the scenario, the most relevant
parameters can be identified. From this description
it is then possible to vary the possible parameters as
well as the set of parameters that build up the scenario,
and thus answer Alexanders (2000) question: what
if . . . ?.

4.4

Transferring information and knowledge

A step that is often given inadequate attention is the

transferring of information and knowledge obtained
during the managing and the evaluation of an emergency to the entire organisation. Thus there is a need
for creating organisational learning (Carley & Harrald
1997). This task is not an easy one, and requires serious

85

Malmo also had an administrative diary system that

was supposed to be used as a way to spread information within the organisation. During the managing of
the Lebanon war this system was not used as intended.
Among other things, this depended on that some people had problems using it. The critical question to ask
is thus: If the situation had been worse, how would they
then disseminate information within the organisation?
In a worse situation, with many more people involved,
it is probably not manageable to only use direct contact.
Would Malmo have managed such a situation?
By asking what if-questions concerning these two
critical aspects or parameters (as well as others), and in
that way varying the parameters, the organisation and
its employees might get an improved understanding of
their capability to manage future emergencies. When
Malmo used the proposed approach and asked those
critical questions several problems became visible and
discussed. Due to this, Malmo has later developed
new routines for staffing their emergency response
organisation.

resources. Therefore it is essential for organisations

to create a planned structure or process for this task.
This step also includes activities such as education and
exercises. In the end it is essential that the findings are
carried by the individuals as well as codified in suitable
artefacts of the organisation.
In addition, to create a better transfer and organisational learning it is throughout all steps of the approach
recommendable to work in groups. One reason for this
is that more people can be potential messengers to the
rest of the organisation.
5

DEMONSTRATING THE APPROACH

The main goal of this paper was to develop and

demonstrate an approach for strengthening emergency
response capability through improving learning from
the evaluation of specific response experiences. From
theories of learning a hypothetical approach has been
constructed. Below the constructed approach will be
demonstrated through application on the evaluation of
the city of Malmos response to the Lebanon war. In
addition, the evaluations of Malmos managing of the
tsunami and the riot will be used in the demonstration.
5.1

5.3

Description of the emergency scenario

The test of the approach started from a construction

of the emergency scenario. During this work critical
parameters for the response of the managing of the
Lebanon war were observed.
5.2

Variation of the set of parameters

The second step was to vary the set of parameters

by e.g. comparing the evaluated scenario to other
scenarios.
During the interviews many people compared
Malmos managing of the Lebanon war with the
managing of the Tsunami. During both emergencies
the Swedish government evacuated Swedish people
to Sweden from other countries. Individuals arriving to Malmo caused Malmo to initiate emergency
responses. Malmos work during both the situations
aimed at helping the evacuated. During both situations
it is possible to identify more or less the same response
generated needs. Both situations required e.g. communication within the organisation and the creation of a
staff group to coordinate the municipal response.
Also the agent generated needs were similar in
the two situations. Some of the affected individuals
arriving to Sweden needed support and help from the
authorities. But the type of support needed varied. For
example, after the tsunami there were great needs for
psychosocial support, while the people returning from
Lebanon instead needed housing and food.
After the riot in 2007 a comparison with the managing of the Malmo consequences of the Lebanon war
was carried out. This comparison showed much more
dissimilarities than the comparison mentioned above,
especially when discussing in terms of agent generated needs. For example, during the riot efforts were
concentrated on informing the inhabitants on what the
city was doing and on providing constructive activities
for youths. The response generated needs, e.g. a need
to initiate some form of emergency response organisation and communication with media, were the same

Variation of the values of the parameters

During the evaluation of the Lebanon war two parameters were identified as especially critical. These were
the staffing of the central staff group and the spreading
of information within the operative organisation.
During the emergency situation the strained staffing
situation was a problem for the people working in the
central staff group. There was no plan for long-term
staffing. A problem was that the crisis happened during the period of summer when most of the people
that usually work in the organisation were on vacation. In addition, there seems to have been a hesitation
to bring in more than a minimum of staff. This resulted
in that some individuals on duty were overloaded with
tasks. After a week these individuals were exhausted.
This was an obvious threat to the organisations ability
to continue operations. Critical questions to ask are:
What if the situation was even worse, would Malmo
have managed it? What if it would have been even
more difficult to staff the response organisation?
The spreading of information within the operative central organisation was primarily done by direct
contact between people either by telephone or mail.

86

unproductive. But if we can do those things in a reasonably disciplined way, we can be smarter and more
imaginative (Clarke 2005 p. 84).
The application of the approach during the evaluation of the managing of the Lebanon war consequences appears to have strengthened the learning in
the organisation. This statement is partly based on
opinions expressed by the individuals in the organisation involved in the discussions during and after the
evaluation. During the reception of the written report
as well as during seminars and presentations of the subject we found that the organisation understood and had
use of the way of thinking generated by using the proposed approach. Subsequently the organisation used
the findings from the use of this way of thinking in
their revision of their emergency management plan.
The new way of thinking seems to have resulted
in providing the organisation a more effective way
of identifying critical aspects. Consequently, it comes
down to being sensitive to the critical dimensions of
variation of these parameters. There is still a need to
further study how an organisation knows which the
critical dimensions are. It is also needed to further
evaluate and refine the approach in other organisations
and on other forms of emergencies.

as during Malmos managing of the Lebanon related

events.
Really interesting question to ask are: What will
happen in the future? What if Malmo will be hit by a
storm or a terrorist attack? How would that influence
the city? Who will need help? Will there be any need
for social support or just a need for technical help?
Will people die or get hurt? Obviously, this is a never
ending discussion. The main idea is not to find all
possible future scenarios, but to create a capability for
this way of thinking. There is a need to always expect
the unexpected.
5.4

Transferring information and knowledge

To support transfer of the result throughout the organisation the evaluation of the Lebanon war resulted in
seminars for different groups of people within the
organisation, e.g. the preparedness planners and the
persons responsible for information during an emergency. These seminars resulted in thorough discussions in the organisation on emergency management
capability. In addition, an evaluation report was made
and distributed throughout the organisation. The discussions during and after the evaluation of the Lebanon
war also led to changes in Malmos emergency management planning and plans. Some of these changes
can be considered examples of double-loop learning,
with altering of parameters, expected to improve future
emergency responses.

CONCLUSION

Seeing scenarios as sets of parameters, and elaborating on the variation of parameter values as well as
the set of parameters, seems to offer possibilities for
strengthening transfer. It may thus support emergency
response organisations in developing rich and manysided emergency management capabilities based on
evaluations of occurred emergency events.

DISCUSSION

This paper has focused on the construction of an

approach for strengthening emergency response capability through improving learning from the evaluation
of specific response experiences.
We have described how imaginary variation of sets
of parameters and parameter values can be used in
evaluation processes around scenarios. Discussing in
terms of such variation has shown to be useful. This
holds for written reports as well as presentations. Similar views are discussed in the literature. For example
Weick & Sutcliffe (2001) discuss how to manage the
unexpected. They describe certain attitudes within an
organisation, e.g. that the organisation is preoccupied
with failures and reluctant to simplify interpretations, that help the organisation to create mindfulness.
A mindful organisation continues to question and
reconsider conceptualisations and models and thus
increases the reliability of their operations. Likewise Clarke (2005) discusses the need for playing
with scenarios and imagines different possible futures.
It is sometimes said that playing with hypothetical scenarios and concentrating on consequences is

REFERENCES
Alexander, D. 2000. Scenario methodology for teaching principles of emergency management. Disaster Prevention
and Management: An International Journal 9(2): 8997.
Argote, L. & Ingram, P. 2000. Knowledge Transfer: A
Basis for Competitive Advantage in Firms. Organizational Behavior and Human Decision Processes 82(1):
150169.
Argyris, C. & Schn, D.A. 1996. Organizational Learning II:
Theory, Method, and Practice. Reading, Massachusetts:
Addison-Wesley Publishing Company.
Boin, A., t Hart, P., Stern, E. & Sundelius, B. 2005. The
Politics of Crisis Management: Public Leadership Under
Pressure. Cambridge: Cambridge University Press.
Carley, K.M. & Harrald, J.R. 1997. Organizational Learning Under Fire. The American Behavioral Scientist 40(3):
310332.
Clarke, L. 2005. Worst cases: terror and catastrophe in
the popular imagination. Chicago: University of Chicago
Press.

87

Dynes, R.R. 1994. Community Emergency Planning: False

Assumptions and Inappropriate Analogies. International
Journal of Mass Emergencies and Disasters 12(2):
141158.
Dynes, R.R., Quarantelli, E.L. & Kreps, G.A. 1981. A perspective on disaster planning, 3rd edition (DRC Research
Notes/Report Series No. 11). Delaware: University of
Delaware, Disaster Research Center.
Lagadec, P. 2006. Crisis Management in the Twenty-First
Century: Unthinkable Events in Inconceivable Contexts. In H. Rodriguez, E.L. Quarantelli & R. Dynes (eds),
Handbook of Disaster Research: 489507. New York:
Springer.
Marton, F. & Booth, S. 1999. Learning and Awareness.
Mahawa NJ: Erlbaum.
Pang, M.F. 2003. Two Faces of Variation: on continuity in
the phenomenographic movement. Scandinavian Journal
of Educational Research 47(2): 145156.

Reber, A.S. 1995. The Penguin dictionary of psychology.

London: Penguin Books.
Runesson, U. 2006. What is it Possible to Learn? On Variation as a Necessary Condition for Learning. Scandinavian
Journal of Educational Research 50(4): 397410.
Senge, P.M. 2006. The Fifth Discipline: The Art and Practice
of the Learning Organization. London: Random House
Business.
Smith, D. & Elliott, D. 2007. Exploring the Barriers to
Learning from Crisis: Organizational Learning and Crisis.
Management Learning 38(5): 519538.
Tierney, K.J., Lindell, M.K. & Perry, R.W. 2001. Facing the
unexpected: Disaster preparedness and response in the
United States. Washington, D.C.: Joseph Henry Press.
Weick, K.E. & Sutcliffe, K.M. 2001. Managing the Unexpected: Assuring High Performance in an Age of Complexity. San Francisco: Jossey-Bass.

88

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On the constructive role of multi-criteria analysis in complex

decision-making: An application in radiological emergency management
C. Turcanu, B. Carl, J. Paridaens & F. Hardeman
SCKCEN Belgian Nuclear Research Centre, Inst. Environment, Health and Safety, Mol, Belgium

ABSTRACT: In this paper we discuss about the use of multi-criteria analysis in complex societal problems and
we illustrate our findings from a case study concerning the management of radioactively contaminated milk. We
show that application of multi-criteria analysis as an iterative process can benefit not only the decision-making
process in the crisis management phase, but also the activities associated with planning and preparedness. New
areas of investigation (e.g. zoning of affected areas or public acceptance of countermeasures) are tackled in
order to gain more insight in the factors contributing to a successful implementation of protective actions and
the stakeholders values coming into play. We follow the structured approach of multi-criteria analysis and we
point out some practical implications for the decision-making process.

structuring the decision process and achieving a common understanding of the decision problem and the
values at stake.
Using MCDA in an iterative fashion (as recommended e.g. by Dodgson et al. 2005) brings about
also an important learning dimension, especially for
an application field such as nuclear or radiological
emergency management, where planning and preparedness are two essential factors contributing to
an effective response. An elaborated exercise policy will serve as a cycle addressing all the steps of
the emergency management process: accident scenario definition, emergency planning, training and
communication, exercising and response evaluation,
fine-tuning of plans, etc (see Fig. 1).

INTRODUCTION

Management of nuclear or radiological events leading to contamination in the environment is a complex

societal problem, as proven by a number of events
ranging from nuclear power plant accidents to loss of
radioactive sources. In addition to the radiological consequences and the technical feasibility, an effective
response strategy must also take into account public acceptability, communication needs, ethical and
environmental aspects, the spatial variation of contamination and the socio-demographic background of
the affected people. This has been highlighted in the
international projects in the 1990s (French et al. 1992,
Allen et al. 1996) and reinforced in recent European projectse.g. STRATEGY, for site restoration
(Howard et al. 2005) and FARMING, for extending
the stakeholders involvement in the management of
contaminated food production systems (Nisbet et al.
2005)as well as in reports by the International
Atomic Energy Agency (IAEA 2006, p. 86).
In this context, multi-criteria decision aid (MCDA)
appears as a promising paradigm since it can deal
with multiple, often conflicting, factors and value
systems. It also helps overcoming the shortcomings
of traditional Cost-Benefit Analysis, especially when
dealing with values that cannot be easily quantified
or even less, translated in monetary terms due to their
intangible nature.
A number of nuclear emergency management decision workshops (e.g. French 1996, Hmlinen et al.
2000a, Geldermann et al. 2005) substantiate the role
of MCDA as a useful tool in stimulating discussions,

Risk management
cycle

Risk Assessment &

accident scenarios
Post crisis interventions,
recovery & rehabilitation
Evaluations &
lessons learned

89

Emergency
planning and response
scenarios
Exercises

Training &
communication

Simulated
response

Response, interventions
and crisis management

Figure 1.

Exercise
management
cycle

Mitigation

Accident

The emergency management cycle.

of MAVT software in decision conferences can be

more successful provided there is a careful planning in
advance, particularly for use in emergency exercises.
In what concerns the aggregation function used, the
usual choice is an additive aggregation. The use of
a utility function of exponential form had been earlier proposed by Papamichail & French (2000), but
it proved non-operational due to missing information
on the variance of actions scores. It is interesting to
notice that although risk aversion could seem as a natural attitude, Hmlinen et al. (2000b) report a case
study where half of the participants were risk averse
and half were risk seeking, possibly due to a different
perception of the decision problem. Related to this,
the authors of this case study refer to the hypothesis
of Kahneman & Tversky (1979) that people are often
risk seeking when it comes to losses (e.g. lives lost) and
risk averse when it comes to gains (e.g. lives saved).
Another type of MCDA approach, addressing at the
same time the definition of countermeasure strategies
as potential actions, as well as the exploration of the
efficient ones is due to Perny & Vanderpooten (1998).
In their interactive MCDA model, the affected area is
assumed to be a priori divided in a number of zones,
whereas the treatment of food products in each zone
can be again divided among various individual countermeasures. The practical implementation makes use
of a multi-objective linear programming model having
as objective functions cost, averted collective dose and
public acceptability. This modelling, although questionable as far as public acceptabilityand qualitative
factors in generalis concerned, allows a great deal
of flexibility at the level of defining constraints and
exploration of efficient, feasible strategies.
The various degrees of acceptance of such decision
aid tools in the different countries suggest however that
the processes envisaged must be tuned to fully accommodate the stakeholders needs (Carter & French
2005). At the same time, the role of such analyses
in a real crisis is still a subject of debate (Hmlinen
et al. 2000a, Mustajoki et al. 2007).
Early stakeholder involvement in the design of
models developed for decision support is needed in
order to accommodate their use to the national context,
e.g. to ensure that the hypotheses assumed, the models and the type of reasoning used reflect the actual
needs of the decision-making process. From an operational viewpoint, the MCDA process can be used to
check if the existing models fit e.g. the databases and
decision practices in place and, at the same time, if
they are sufficiently flexible at all levels: from definition of potential actions and evaluation criteria, up to
modelling of comprehensive preferences.
In what concerns the general use of MCDA methods, some authors (Belton & Stewart 2002) suggest
that value functions methods are well suited within
workshop settings, facilitating the construction of

It is therefore of interest to extend the study of

potential applications of MCDA in this particular field
and to try to integrate it into the socio-political context specific to each country. Drawing on these ideas,
the research reported in this paper aimed at adding
the knowledge and experience from key stakeholders
in Belgium to the multi-criteria analysis tools developed for decision-support, and at exploring various
MCDA methods. The management of contaminated
milk after a radioactive release to the environment has
been chosen as case study.
In Section 2 we discuss about the use of multicriteria analysis in nuclear emergency management.
In Section 3 we outline the Belgian decision-making
context. This will set the framework for Section 4 in
which we elaborate on the results from a stakeholder
process carried out in Belgium in order to set up a
theoretical and operational framework for an MCDA
model for the management of contaminated milk. The
constructive role of multi-criteria analysis, the conclusions and lessons learnt are summarised in the final
section.
2

MCDA IN NUCLEAR EMERGENCY

MANAGEMENT

MCDA provides a better approach than CBA for

nuclear emergency management and site restoration,
first and foremost because the consequences of potential decisions are of a heterogeneous nature, which
impedes coding them on a unique, e.g. monetary,
scale. In an overwhelming majority (e.g. French 1996,
Zeevaert et al. 2001, Geldermann et al. 2005, Panov
et al. 2006, Gallego et al. 2000), the MCDA methods
used up to now in connection with the management
of radiological contaminations mainly draw from the
multi-attribute value/utility theory (MAU/VT).
Research in this application field for MCDA has
started in the early 90s in the aftermath of the
Chernobyl accident (French et al. 1992) and subsequently linked to the European decision-support
system RODOS (Ehrhardt & Weiss. 2000), as reported
e.g. in French (1996), Hmlinen et al. (1998) and
more recently in Geldermann et al. (2006).
The reported use of MAU/VT for comparing and
ranking countermeasure strategies suggests that it
facilitates the identification of the most important
attributes, contributing thus to a shared understanding of the problem between the different stakeholders
concerned by the decision-making process. Among
the drawbacks revealed was the informational burden of setting criteria weights, especially for sociopsychological attributes (e.g. Hmlinen et al. 1998),
or even evaluating the impact of potential actions for
such non-quantifiable factors. Some recent studies
(Mustajoki et al. 2007) suggest that the application

90

agricultural countermeasures in case of a radioactive

contamination in the food chain is to reduce the radiological risk for people consuming contaminated foodstuff. For such cases, maximum permitted radioactivity levels for food products, also called the European
Council Food Intervention LevelsCFIL(CEC,
1989), have been laid down and are adopted by the
Belgian legislation as well.
The experience from the European project FARMING (Vandecasteele et al. 2005) clearly showed however that the characteristics of the agriculture and the
political environment, as well as the past experiences
of food chain contamination crises are also important
factors to be taken into account. This explains why
some countermeasures aiming at reducing the contamination in food products were considered hardly
acceptable by the stakeholders, if at all, even when
supported by scientific and technical arguments.
Decisions taken have potentially far-reaching consequences, yet they often have to be made under time
pressure and conditions of uncertainty. Also in case of
food countermeasures, time may become an important
issue. For example in case of milk, the large amounts
produced daily vs. the limited storage facilities of
diaries make it necessary to rapidly take a decision
for management of contaminated milk (processing,
disposal, etc).

preferences by working groups who mainly represent

stakeholder interests. Outranking methods and
what/if experiments with value functions might instead
be better fitted for informing political decisionmakers regarding the consequences of a particular
course of action.
The exploration of an outranking methodology is
motivated by some particularities of our decision problem (Turcanu 2007). Firstly, the units of the evaluation
criteria (e.g. averted dose, cost, and public acceptance) are heterogeneous and coding them into one
common scale appears difficult and not entirely natural. Secondly, the compensation issues between gains
on some criteria and losses on other criteria are not
readily quantifiable. In general, Kottemann & Davis
(1991) suggest that the degree to which the preference
elicitation technique employed requires explicit tradeoff judgments influences the decisional conflict that
can negatively affect the overall perception of a multicriteria decision support system. Thirdly, the process
of weighting and judging seems in general more qualitative than quantitative. When it comes to factors such
as social or psychological impact, the process of giving
weights to these non-quantifiable factors may appear
questionable. Furthermore, outranking methods relax
some the key assumptions of the MAUT approach, for
example comparability between any two actions and
transitivity of preferences and indifferences.
In the following sections, the focus is laid on
the management of contaminated milk. The continuous production of milk requires indeed urgent
decisions due to the limited storage facilities. Moreover, dairy products are an important element of the
diet, especially for children, who constitute the most
radiosensitive population group. Finally, for certain
radionuclides with high radiotoxicity such as radiocaesium and radioiodine, maximum levels of activity
concentration in milk are reached within few days
after the ground deposition of the radioactive material
(Nisbet, 2002).
3

MCDA MODEL FRAMEWORK

In this section we discuss the multi-criteria decision

aid framework developed for the management of contaminated milk. We describe the stakeholder process
and the key elements of the MCDA model. Additional
details on the stakeholder process and its results can
be found in Turcanu et al. (2006); in the following we
point out the new areas of investigation opened by this
research and we underline the constructive dimension
brought by the use of MCDA.
4.1

NUCLEAR AND RADIOLOGICAL

EMERGENCY MANAGEMENT IN BELGIUM

The stakeholder process

In the stakeholders process (Turcanu et al. 2006),

we carried out 18 individual interviews with governmental and non-governmental key actors for the
management of contaminated milk: decision-makers,
experts and practitioners. This included various areas
of activity: decision making (Ministry of Interior);
radiation protection and emergency planning; radioecology; safety of the food chain; public health; dairy
industry; farmers association; communication; local
decision making; social science; and management of
radioactive waste.
We have decided to make use of individual interviews rather than focus group discussions or questionnaires in order to allow sufficient time for discussions,

In the context of the Belgian radiological emergency

plan (Royal Decree, 2003 and subsequent amendments), the decision maker, also called the Emergency Director is the Minister of Interior or his
representative. He is chairman of the decision cell,
i.e. the Federal Co-ordination Committee. The decision cell is advised by the radiological evaluation cell,
mainly in charge with the technical and feasibility
aspects, and by the socio-economic evaluation cell.
Various other organisations and stakeholders have
a direct or indirect impact on the efficiency of the
chosen strategies. The primary goal of food and

91

to cover in a consistent way a range of stakeholders as

complete as possible and to facilitate free expression
of opinions. The discussion protocol used followed the
main steps described below.
4.2

fast calculation of e.g. the amount of production in

the selected zone, the implementation costs for the
selected countermeasure, the collective doses and
maximal individual doses due to ingestion of contaminated foodstuff (the dose is an objective measure of
the detriment to health), helping the decision makers
or advisers in choosing the set of potential actions to
be further evaluated.

Potential actions

Several individual or combined countermeasures can

be employed for the management of contaminated
milk (Howard et al. 2005), for instance disposal of
contaminated milk; prevention or reduction of activity
in milk (clean feeding, feed additives) and/or storage
and processing to dairy products with low radioactivity
retention factors. The need to make available a more
flexible way to generate potential actions, requirement which came out from the stakeholder process,
led to the development of a prototype tool allowing
the integration of various types of data: data from the
food agency (e.g. location of dairy farms and local
production), modelling and measurement data (e.g.
ground depositions of radioactive material at the location of dairy farms), other data such as administrative
boundaries, etc (see Fig. 2).
Simulated deposition data, for example for iodine
and caesium, were generated in discrete points of a grid
using dispersion and deposition models, e.g. as existing in the RODOS decision-support system (Ehrhardt
and Weiss 2000) and introduced into a geographical information system (GIS). Here, the data were
converted into continuous deposition data through a
simple natural neighbour algorithm. Next, the deposition data were evaluated at the locations of dairy farms,
and combined with production data, both provided
by the federal food agency. The GIS allows overlaying these data with other spatial information such as
administrative boundaries, topological maps, selected
zones, population data, etc.
A data fusion tool was subsequently implemented
in spreadsheet form. This tool can be used for a

4.3 Evaluation criteria

The evaluation criteria were built through a process
combining a top-down and a bottom-up approach
(Turcanu et al. 2006).
The stakeholders interviewed were first asked to
identify all the relevant effects, attributes and consequences of potential actions. Subsequently, they
commented on a list of evaluation criteria derived from
the literature and amended it, if felt necessary. Based
on the resulting list, a number of evaluation criteria
was built taking account, as much as possible, of the
properties of exhaustiveness, cohesiveness and nonredundancy (see Roy 1996 for a description of these
concepts), in order to arrive at a consistent set of evaluation criteria. The example discussed at the end of
this section illustrates the list of criteria proposed.
Here we should mention that one of the evaluation criteria highlighted by all stakeholders as very
important is public acceptance. To have a better
assessment of it, we included in a public survey in
Belgium (Turcanu et al. 2007) a number of issues
relevant for the decision-making process: i) public
acceptance of various countermeasures; ii) consumers
behaviour. The results showed that clean feeding
of dairy cattle and disposal of contaminated milk
are the preferred options in case of contaminations
above legal norms. For contaminations below legal
norms, normal consumption of milk seemed better
accepted than disposal. Nonetheless, the expressed
consumers behaviour revealed a precautionary tendency: the presence of radioactivity at some step
in the food chain could lead to avoiding purchasing products from affected areas. Finally, public trust
building was revealed as a key element of a successful
countermeasure strategy.
The resulting distributions of acceptance degrees
(from strong disagreement to strong agreement)
on the sample of respondents can be compared in several ways (Turcanu et al. 2007). To make use of all
information available, an outranking relation S (with
the meaning at least as good as) can be defined on
the set of individual countermeasures, e.g. based on
stochastic dominance:

GIS data

Food Agency
data

Model and
measurement data

aSb

Figure 2. Integration of various types of data and selection

of areas and countermeasures.


ji

92

aj


ji

bj + ,

i = 1, . . . , 5

be better fitted for a given criterion gi ,the following

discrimination thresholds were chosen:

where aj , bj are the percentages of respondents using

the j-th qualitative label to evaluate countermeasures
a and b, respectively, and is a parameter linked
to the uncertainty in the evaluation of aj and bj . A
simpler approach, but with loss of information, is to
derive a countermeasures public acceptance score as
e.g. the percentage of respondents agreeing with a
countermeasure.
Despite the inherent uncertainty connected to
assessing the public acceptance of countermeasures
in peace time we consider that such a study is useful for emergency planning purposes, especially for
situations when there is a time constraint as it is the
case for the management of contaminated milk.

4.4

ref

qi (gi (a)) = max {qi , q0i gi (a)}

and

pi (gi (a)) = p0 qi (gi (a)),

ref

with p0 > 1, a fixed value and qi 0 and 0 q0i < 1

parameters depending on the criterion gi .
4.5 Comprehensive preferences
In order to derive comprehensive preferences, we discussed with the stakeholders interviewed about the
relative importance of evaluation criteria. This notion
can be interpreted differently (Roy & Mousseau 1996),
depending on the type of preference aggregation
method used.
We investigated the adequacy for our application of
four types of inter-criteria information: i) substitution
rates (tradeoffs) between criteria; ii) criteria weights as
intrinsic importance coefficients; iii) criteria ranking
with possible ties; iv) a partial ranking of subfamilies
of criteria.
For each of these we asked the stakeholders interviewed if such a way to express priorities is suitable and, most importantly, if they are willing and
accept to provide/receive such information. Our discussions revealed a higher acceptance of the qualitative
approaches, which indicates that outranking methods
might be better suited. The concept of weights as
intrinsic importance coefficients proved hard to understand, but encountered a smaller number of opponents
than weights associated with substitution rates. The
main argument against the latter can be seen in ethical motivations, e.g. the difficulty to argue for a value
trade-off between the doses received and the economic
cost. Methods of outranking type that can exploit a
qualitative expression of inter-criteria information are
for instance the MELCHIOR method (Leclerq 1984)
or ELECTRE IV (Roy 1996).
Let us consider in the following the case when the
inter-criteria information is incomplete -because the
decision-maker is not able or not willing to give this
information- is the following. Lets suppose that the
information about the relative importance of criteria is
available in the form of a function assumed irreflexive
and asymmetric:

Formal modelling of evaluation criteria

The preferences with respect to each criterion were

modelled with "double threshold" model (Vincke
1992). Each criterion was thus represented by a realvalued positive function associated with two types of
discrimination thresholds: an indifference threshold
q() and a preference threshold p(). For a criterion g
to be maximised and two potential actions a and b, we
can define the following relations:
a I b (a indifferent to b)
g(a) g(b) + q(g(b)) and
g(b) g(a) + q(g(a));
a P b (a strictly preferred to b)
g(a) > g(b) + p(g(b));
a Q b (a weakly preferred to b)
g(b) + q(g(b)) < g(a) and
g(a) g(b) + p(g(b)).
Under certain consistency conditions for p and q,
this criterion model corresponds to what is called a
pseudo-criterion (see Roy 1996).
The choice of the double threshold model is motivated by the fact that for certain criteria (e.g. economic
cost) it might not be possible to conclude a strict
preference between two actions scoring similar values, e.g. due to the uncertainties involved, while an
intermediary zone exists between indifference and
strict preference. The double threshold model is a
general one, easy to particularise for other types
of criteria. For example, by setting both thresholds
to zero, one obtains the traditional, no-threshold
model.
In order to account in an intuitive way for both the
situations when a fixed or a variable threshold could

: G G {0, 1}, such that

(gm , gp ) = 1 criterion gm is more important
than criterion gp ,
where G is the complete set of evaluation criteria.
The function , comparing the relative importance
of individual criteria, can be extended to subsets of

93

criteria in a manner inspired from the MELCHIOR

method: we test if favourable criteria are more important than the unfavourable criteria.
Formally stated, we define recursively a mapping
: (G) (G) {0, 1} as:

(F, ) = 1,

= F G,

(, F) = 0,

F G,

Table 2.

Criterion
C1 Residual collective
effective dose
(person . Sv)
C2 Maximal individual
5% (thyroid)
dose (mSv)
C3 Implementation
cost (kC
=)
C4 Waste (tonnes)
C5 Public
acceptance
C6 (Geographical)
feasibility
C7 Dairy industrys
acceptance
C8 Uncertainty of
outcome
C9 Farmers
acceptance
C10 Environmental
impact
C11 Reversibility

({gm }, {gp }) = 1 (gm , gp ) = 1,

({gm } F, H) = 1, with

{gm } F G

and

H G (F, H) = 1 or gp H: (gm , gp ) = 1
and

(F, H\{gp }) = 1.

We further define a binary relation R representing comprehensive preferences on the set of potential
actions A as follows:
a, b A, R(a, b) = (F, H), where
F = {gi G|aPi b}, H = {gi G|b(Pi Qi )a},
and (Pi , Qi , Ii ) is the preference structure associated
with criterion gi .
4.6

An illustrative example

In this section we discuss a hypothetical (limited scale)

131
I milk contamination. The potential actions (suitable decision alternatives) are described in Table 1.
Table 2 gives the complete list of evaluation criteria
and Table 3 summarises the impact of potential actions
with respect to these criteria.
When inter-criteria information is not available,
the comprehensive preferences resulting from the
aggregation method given above are illustrated in
Fig. 3. For instance, the arrow A4 A3 means that
action A4 is globally preferred to action A3 . We can
see that there are also actions which are incomparable,
Table 1.

Description

A1
A2

Do Nothing
Clean feed in area defined by sector
(100 , 119 , 25 km):
Clean feed in area where deposit activity
>4000 Bq/m2 :
Clean feed in area where deposited activity
>4000 Bq/m2 , extended to full
administrative zones
Storage for 32 days in area where deposited
activity >4000 Bq/m2
Storage for 32 days in area where deposited
activity >4000 Bq/m2 , extended to full
administrative zones

A3
A4
A5
A6

Variable
indifference
threshold
(q0i )

Minimal
indifference
threshold Optimis.
(qi ref )
direction

10%

10 person
Sv

0.5 mSv

min

10%

20 kC
=

min

10%
0

1t
0

min
max

max

max

min

max

min

max

min

p0 = 2 for all cases.

Table 3.

Impact of potential actions .

Criterion C1

Potential actions: Example.

Action

Evaluation criteria: Example.

C2

Action

Person
Sv
mSv

A1
A2
A3
A4
A5
A6

4
0.1
0.3
0.3
0.8
0.8

C3

C4

C5

C6

C7

C8

=
kC

0
3
3
3
2
2

1
1
1
2
1
2

0
2
2
2
1
1

3
1
1
1
2
2

100
0
0
3.6 240
16
4.6 17
16
4.6 27
16
20
1.3 0
20
2.5 0

On criteria C C
9
11 all actions score the same, therefore
they are not mentioned in the table.

for instance actions A4 and A6 ; the situation changes

however, when some information is given concerning
the relative importance of criteria.
Let us assume that the decision-maker states that
the maximal individual dose is more important than
any other criterion and that public acceptance is more
important than the cost of implementation and the
geographical feasibility. We then obtain the results presented in Fig. 4, highlighting both actions A4 and A2
as possibly interesting choices.

94

A2

A4

Table 4.

A6

Results from the multi-criteria analysis.

Multi-criteria analysis Stakeholder process results

Decision context
A3

A5

Potential actions
Evaluation criteria

A1

Figure 3. Comprehensive preferences without inter-criteria

information.

Preference
aggregation

A4

A3

A6

makers and decision advisers, as well as practitioners in the field (e.g. from dairy industry or farmers
union) contributed to a better understanding of many
aspects of the problem considered. For our case study,
this process triggered further research in two directions: flexible tools for generating potential actions
and social research in the field of public acceptance of
food chain countermeasures.
MCDA can be thus viewed as bridging between various sciencesdecision science, radiation protection,
radioecological modelling and social scienceand a
useful tool in all emergency management phases.
The research presented here represents one step in
an iterative cycle. Further feedback from exercises and
workshops will contribute to improving the proposed
methodology.

A2
A5

A1
Figure 4. Comprehensive preferences with inter-criteria
information.

Naturally, results such as those presented above

must be subjected to a detailed robustness analysis
(Dias 2006) as they depend on the specific values chosen for the parameters used in the model,
i.e. discrimination thresholds. For instance, if
q2 ref = 1 mSv, instead of 0.5 mSv as set initially (see
Table 2), while the rest of the parameters remain at
their initial value, both actions A4 and A3 would be
globally preferred to action A2 .
5

Learn about the different

viewpoints.
New modelling tools connecting
decision-support tools with
existing databases.
Set of evaluation criteria that can
serve as basis in future exercises.
Better assessment of public acceptance (public opinion survey).
Revealed tendency towards a qualitative inter-criteria information.
Outranking methods highlighted
as potentially useful for future
studies.

REFERENCES
Allen, P., Archangelskaya, G., Belayev, S., Demin, V.,
Drotz-Sjberg, B.-M., Hedemann-Jensen, P., Morrey, M.,
Prilipko, V., Ramsaev, P., Rumyantseva, G., Savkin, M.,
Sharp, C. & Skryabin, A. 1996. Optimisation of health
protection of the public following a major nuclear
accident: interaction between radiation protection and
social and psychological factors. Health Physics 71(5):
763765.
Belton, V. & Stewart, T.J. 2002. Multiple Criteria Decision
Analysis: An integrated approach. Kluwer: Dordrecht.
Carter, E. & French, S. 2005. Nuclear Emergency Management in Europe: A Review of Approaches to Decision
Making, Proc. 2nd Int. Conf. on Information Systems for
Crisis Response and Management, 1820 April, Brussels,
Belgium, ISBN 9076971099, pp. 247259.
Dias, L.C. 2006. A note on the role of robustness analysis in decision aiding processes. Working paper, Institute
of Systems Engineering and Computers INESC-Coimbra,
Portugal. www.inesc.pt.

CONCLUSIONS

In this paper we have discussed the application of

multi-criteria analysis for a case study dealing with the
management of contaminated milk. Our main findings
are summarised in Table 4.
One important conclusion is that consultation with
concerned stakeholders is a key factor that can lead
to a more pragmatic decision aid approach and presumably to an increased acceptance of the resulting
models. The stakeholder process, involving decision

95

Dodgson, J., Spackman, M., Pearman, A. & Phillips, L. 2000.

Multi-Criteria Analysis: A Manual. London: Department of the Environment, Transport and the Regions,
www.communities.gov.uk.
Ehrhardt, J. & Weiss, A. 2000. RODOS: Decision Support
for Off-Site Nuclear Emergency Management in Europe.
EUR19144EN. European Community: Luxembourg.
French, S. 1996. Multi-attribute decision support in the event
of a nuclear accident. J Multi-Crit Decis Anal 5: 3957.
French, S., Kelly, G.N. & Morrey, M. 1992. Decision conferencing and the International Chernobyl Project. J Radiol
Prot 12: 1728.
Gallego, E., Brittain, J., Hkanson, L., Heling, R.,
Hofman, D. & Monte, L. 2000. MOIRA: A Computerised
Decision Support System for the Restoration of Radionuclide Contaminated Freshwater Ecosystems. Proc. 10th
Int. IRPA Congress, May 1419, Hiroshima, Japan.
(http://www2000.irpa.net/irpa10/cdrom/00393.pdf).
Geldermann, J., Treitz, M., Bertsch, V. & Rentz, O. 2005.
Moderated Decision Support and Countermeasure Planning for Off-site Emergency Management. In: Loulou, R.,
Waaub, J.-P. & Zaccour, G. (eds) Energy and Environment: Modelling and Analysis. Kluwer: Dordrecht,
pp. 6381.
Geldermann, J., Bertsch, V., Treitz, M., French, S.,
Papamichail, K.N. & Hmlinen, R.P. 2006. Multicriteria decision-support and evaluation of strategies for
nuclear remediation management. OMEGAThe International Journal of Management Science (in press). Also
downloadable at www.sal.hut.fi/Publications/pdf-files/
MGEL05a.doc.
Hmlinen, R.P., Sinkko, K., Lindstedt, M.R.K., Amman,
M. & Salo, A. 1998. RODOS and decision conferencing
on early phase protective actions in Finland. STUK-A 159
Report, STUKRadiation and Nuclear Safety Authority,
Helsinki, Finland, ISBN 951-712-238-7.
Hmlinen, R.P., Lindstedt, M.R.K. & Sinkko, K. 2000a.
Multi-attribute risk analysis in nuclear emergency management. Risk Analysis 20(4): 455468.
Hmlinen, R.P., Sinkko, K., Lindstedt, M.R.K., Amman,
M. & Salo, A. 2000b. Decision analysis interviews on
protective actions in Finland supported by the RODOS
system. STUK-A 173 Report, STUKRadiation and
Nuclear Safety Authority, Helsinki, Finland. ISBN 951712-361-2.
Howard, B.J., Beresford, N.A., Nisbet, A., Cox, G., Oughton,
D.H., Hunt, J., Alvarez, B., Andersson, K.G., Liland, A. &
Voigt, G. 2005. The STRATEGY project: decision tools
to aid sustainable restoration and long-term management
of contaminated agricultural ecosystems. J Env Rad 83:
275295.
IAEA. 2006. Environmental consequences of the Chernobyl
accident and their remediation: twenty years of experience/report of the Chernobyl Forum Expert Group Environment. STI/PUB/1239, International Atomic Energy
Agency. Vienna: Austria.
Kahneman, D. & Tversky, A. 1979. A. Prospect Theory: An
Analysis of Decision under Risk. Econometrica 47(2):
263291.
Kottemann, J.E. & Davis, F.D. 1991. Decisional conflict and
user acceptance of multi-criteria decision-making aids.
Dec Sci 22(4): 918927.

Leclercq, J.P. 1984. Propositions dextension de la notion

de dominance en prsence de relations dordre sur les
pseudo-critres: MELCHIOR. Revue Belge de Recherche
Oprationnelle, de Statistique et dInformatique 24(1):
3246.
Mustajoki, J., Hmlinen, R.P. & Sinkko, K. 2007. Interactive computer support in decision conferencing: Two cases
on off-site nuclear emergency management. Decision
Support Systems 42: 22472260.
Nisbet, A.F., Mercer, J.A., Rantavaara, A., Hanninen, R.,
Vandecasteele, C., Carl, B., Hardeman, F., Ioannides,
K.G., Papachristodoulou, C., Tzialla, C., Ollagnon, H.,
Jullien, T. & Pupin, V. 2005. Achievements, difficulties
and future challenges for the FARMING network. J Env
Rad 83: 263274.
Panov, A.V., Fesenko, S.V. & Alexakhin, R.M. 2006. Methodology for assessing the effectiveness of countermeasures
in rural settlements in the long term after the Chernobyl accident on the multi-attribute analysis basis. Proc.
2nd Eur. IPRA Congress, Paris, 1519 May, France.
www.irpa2006europe.com.
Papamichail, K.N. & French, S. 2000. Decision support in
nuclear emergencies, J. Hazard. Mater. 71: 321342.
Perny, P. & Vanderpooten, D. 1998. An interactive multiobjective procedure for selecting medium-term countermeasures after nuclear accidents, J. Multi-Crit. Decis. Anal.
7: 4860.
Roy, B. 1996. Multicriteria Methodology for Decision Aiding. Kluwer: Dordrecht.
Roy, B. & Mousseau, V. 1996. A theoretical framework for
analysing the notion of relative importance of criteria.
J Multi-Crit Decis Anal 5: 145159.
Royal Decree, 2003. Plan dUrgence Nuclaire et Radiologique pour le Territoire Belge, Moniteur Belge,
20.11.2003.
Turcanu, C. 2007. Multi-criteria decision aiding model for
the evaluation of agricultural countermeasures after an
accidental release of radionuclides to the environment.
PhD Thesis. Universit Libre de Bruxelles: Belgium.
Turcanu, C., Carl, B. & Hardeman, F. 2006. Agricultural countermeasures in nuclear emergency management: a stakeholders survey for multi-criteria
model development. J Oper Res Soc. DOI 10.1057/
palgrave.jors.2602337
Turcanu, C., Carl, B., Hardeman, F., Bombaerts, G. & Van
Aeken, K. 2007. Food safety and acceptance of management options after radiological contaminations of the food
chain. Food Qual Pref 18(8): 10851095.
Vandecasteele, C., Hardeman, F., Pauwels, O., Bernaerts, M.,
Carl, B. & Sombr, L. 2005. Attitude of a group of
Belgian stakeholders towards proposed agricultural countermeasures after a radioactive contamination: synthesis
of the discussions within the Belgian EC-FARMING
group. J. Env. Rad. 83: 319332.
Vincke, P. 1992. Multicriteria decision aid. John Wiley &
Sons, Chichester.
Zeevaert, T., Bousher, A., Brendler, V., Hedemann Jensen, P. &
Nordlinder, S. 2001. Evaluation and ranking of restoration
strategies for radioactively contaminated sites. J. Env. Rad.
56: 3350.

96

Decision support systems and software tools for safety and reliability

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Complex, expert based multi-role assessment system for small

and medium enterprises
S.G. Kovacs & M. Costescu
INCDPM Alexandru Darabont, Bucharest, Romania

ABSTRACT: The paper presents a complex assessment system developed for Small and Medium Enterprises.
It assesses quality, safety and environment on three layers starting from basic to complete assessment (including
management and organizational culture). It presents the most interesting attributes of this system together with
some of the results obtained by testing this system on a statistical lot of 250 Romanian SME.

GENERAL ASPECTS

In Romania, Small and Medium Enterprises (SMEs)

were somehow, from objective and subjective reasons, the Cinderella of safety activities. Complex
safety assessment procedures, like Hazop are too
complicated for SMEs. On the other side, more simple assessment systems are considered as too biased.
SMEs need not just to assess safety but quality and
environment as well.
The paper presents a three-layered multi-role
assessment system, based on expert-decision assisting structures (Kovacs 2003b)system developed
especially for SME and known as MRAS.
The system is conceived on three levels of complexity, from basic to extra. In this way, all the requests
regarding safety assessment are satisfied. Moreover,
on the third level, a dual, safety-quality assessment is
performed, together with a safety culture analysis.
The system is integrated into the Integrated Safety
Management Unit (ISMU) which is the management

Figure 2.

center of all the safety solutions for a SME. In this

respect it optimises the informational flows assuring
an efficient usage of all these tools in order to prevent
risks or mitigate them.
ISMU schema is presented in Figure 1.
The general structure of the system is shown in
Figure 2.
2

Figure 1.

MRAS structure.

VULNERABILITY ANALYSIS

In our assessment system vulnerability analysis is the

first step to understand how exposed are SMEs to
operational (safety) threats that could lead to loss
and accidents (Kovacs 2004).Vulnerability analysis
could be defined as a process that defines, identifies and classifies the vulnerabilities (security holes)

ISMU structure.

99

in an enterprise infrastructure. In adition vulnerability analysis can forecast the effectiveness of proposed
prevention measures and evaluate their actual effectiveness after they are put into use. Our vulnerability
analysis is performed in the basic safety assessment
layer and consist mainly on the next steps:
1. Definition and analysis of the existing (and available) human and material resources;
2. Assignment of relative levels of importance to the
resources;
3. Identification of potential safety threats for each
resource;
4. Identification of the potential impact of the threat
on the specific resource (will be later used in
scenario analysis).
5. Development of a strategy for solving the threats
hierarchically, from the most important to the last
important.
6. Defining ways to minimise the consequences if a
threat is acting(TechDir 2006).
We have considered a mirrored vulnerability
(Kovacs 2006b)this analysis of the vulnerability
being oriented inwards and outwards of the assessed
workplace. The Figure 3 shows this aspect.
So, the main attributes for which vulnerability is
assessed are:

Table 1.
Mark

Semnification

0
1
2
3
4
5

Non-vulnerable
Minimal
Medium
Severe vulnerability-loss
Severe vulnerability-accidents
Extreme vulnerability

Past Incident Coefficient.

Ov = Av Pik (1)
where
Pik = 0if there were no previous incidents;
Pik = 1.5if there were near misses, loss incidents
and/or minor accidents;
Pik = 2if there were severe accidents before.

1. the human operator:

2. the operation
3. the process;
The mirrored vulnerability analysis follows not just
how much are these three elements vulnerable but also
how much they affect the vulnerability of the whole
SME.
Vulnerability is estimated in our system on a 0 to 5
scale like in Table 1.
We are also performing a cross-over analysis taking
into account past five year incident statistical data.
So we are computing an Operational Vulnerability
as means of Assessed Vulnerability multiplied by an

Figure 3.

Vulnerability scores.

PRE-AUDIT

The pre-audit phase of the basic safety assessment

plays many roles (Mahinda 2006). The most important
of these are connected with the need to have regularly a preliminary safety assessment (Kovacs 2001a)
which:
identifies the most serious hazards;
maps the weak spots of the SME, where these
hazards could act more often and more severly;
estimates the impact of the hazards action upon man
and property;
verifies basic conformity with safety law and basic
safety provisions;
assures consistency to the following assessments,
recordings and also the management of previous
assessments;
Pre-audit is centered around two attributes of the
workplace:
Pre-audit of the human operator(Kovacs 2001b)its
findings leads to two main measures: the improvement
of training or the change of workplace for the operator
which performance could endanger the safety at the
workplace.
Pre-audit of machines and the technological processthe findings of this part of safety pre-audit
could lead to machine/process improvement or to
machine/process revision or to the improvement
of machines maintenance or if necessary to the
elimination of the machineif the machine is beyond
repair.

Vulnerability analysis schema.

100

Table 2.

Pre- Minimum
audit corresponding level
score of safety

Level of
Pretransition to
audit move to next
ranges level

Nothing in place

01

Demonstrate a basic
12
knowledge and
willingness to
implement safety
policies/procedures/
guidelines
Responsibility and
23
accountability
identified for most
safety related tasks

4
5

OHS Policies/
Procedures and
Guidelines
implemented
Comprehensive level
of implementation
across audit site
Pre-audit results
used to review
and improve
safety system;
demonstrates
willingness for
continual
improvement

Table 3.

The pre-audit assessment system uses the rating

presented in the Table 2.
Pre-audit main guideline is represented by ISO
9001.
Pre-audit main instruments are checklists. A very
short example of such a checklist is presented in the
Table 3.

Pre-audit scores.

34

45

Identification of
basic safety
needs, some
activities
performed
Local safety
procedures
developed and
implemented

Responsibility
documented and
communicated for
the main safety
tasks
Significant level
of implementation
through audit site

SAFETY SCENARIO ANALYSIS

As SMEs are performing often routine activities it is

possible to develop and use as a safety improvement
tool safety scenarios in order to be aware of what could
happen if risks are acting (TUV 2008).
A schema of our scenario analysis module is given
in the Figure 4.
The previous safety assessment data is used in order
to build a scenario planfirstly. For example, in a
specific area of the SME were recorded frequent near
misses in a very short period of time. The previous assessment data is used in order to estimate the
trend of components and their attributes. One of such
components is represented by the Human Operator

Complete
implementation
across audit site
Industry best
practice

Pre-audit checklist example.

Please estimate the truth of the following affirmations taking

into account the actual situation from the workplace
Question
nr.
Question
1
2
3
4

We have a comprehensive Health

and Safety Policy manual plan in
place
All our employees are familiar with
local and national safety regulations
and we abide by them
We have developed a comprehensive
safety inspection process
We are performing regularly this
safety inspection process taking
into account all the necessary
detailsour inspection is real and
not formal

Yes No

Figure 4.

101

Safety scenario module.

and one of its attributes would be the preparedness

in emergency situations. In our example we could
consider the sudden apparition of a fire at one of
the control panels at the workplace. If the preparedness in emergency situations is set on in scenario, the
operator is able to put down the fire and no other
loss occurs. If the attribute is set off the operator is
frightened and runsthe fire propagates and there is
an accidental spillbecause some valves got out of
control.
Our scenario analysis instrument assures the rule of
the 3 Ps (Kovacs 2006b):
Plan:
Establish goals and objectives to follow;
Develop procedures to follow;

The Human Operator;

The Machine(s)including facilities;
The Working Environment;
together with a derived component which includes the
specific work task and also the interaction between
components;
All these components- and their interactionare
analysed in the best case-worst case framework considering also an intermediary state, the medium case.
The medium case is (if not mentioned otherwise) is
the actual situation in the SME. One point of attention
is the transition between the cases, in order to analyse
what happens if, for example, safety resources allocation is postponed in order to allocate resources in a
more hotter places.

Predict:
Predict specific risk action;
Predict component failure;
Predict soft spots where risks could materialize
more often;
Prevent:
Foresee the necessary prevention measures; some
of these measures could be taken immediately;
other are possible to be postponed for a more
favourable period (for example the acquisition of
an expensive prevention mean), other are simply improvements of the existing situationfor
example a training that should include all the
workers at the workplace not just the supervisors;
The actual development of the scenario is performed in a best casemedium caseworst case
framework. This framework could be seen in the
Figure 5.
Mainly there are considered three essential components:

Figure 5.

Worst case-best case scenario framework.

OBTAINED RESULTS

We have tested our system on a statistically significant

lot of 250 Romanian SME (Kovacs 2006a) from all the
economic activities, on a half year period, taking into
account the following assessments:
General management assessment;
Safety management assessment;
Incident rate after implementation (comparatively
with incidents on a 5 year period);
Assessment of the ground floor work teams;
The SMEs test shown that such a systemwhich
is not very complicatedin order to be able to be used
by SMEs aloneis a very efficient tool in order to:
assess more objectively safety together with quality
and environment;
perform an efficient assessment of the SME management regarding quality, safety and environment;
offers a realistic image to the SME management
in order to be convinced about the necessity of
improvement, not just for safety but also for environment, as quality improvement is a must for every
SME in order to be able to remain in the market.
tells the SME manager that quality must not be considered alone but only in connexion with safety and
environment;
offers a valuable assessment instrument for external assessment firms, inspection authorities and risk
assurance companies (Kovacs 2003a);
The Figure 6 shows the vulnerability analysis results
for the statistical lot.
It is possible to see that taking into account the
vulnerability alone the majority of the statistical lot
has a vulnerability over 3, this showing a severe
vulnerability at risk action.

102

REFERENCES

Vulnerability analysis results

5
12%

0
14%

4
16%

1
14%

3
24%

Figure 6.

2
20%

Vulnerability analysis results.

CONCLUSIONS

MRAS (Multi-Role Assessment System) was developed initially as a self-audit tool which could allow
SME to have an objective and realistic image regarding their efforts to assure the continuous improvement
of quality and maintain a decent standard of safety
and environment protection. However, on the development period we have seen a serious interest from
control organs (like the Work Inspection) so we have
developed the system so that it can be used for selfaudit or it can be used by an external auditor. The
control teams are interested to have a quick referential
in order to be able to check quickly and optimally the
safety state inside a SME. In this respect our system
was the most optimal.
Romanian SME are progressing towards the full
integration into the European market. In this respect
they need to abide to the EU provisions, especially regarding safety and environment. Considering
this, the system assures not just the full conformity
with European laws but also an interactive forecastplan-atact-improve instrument. The compliance audits
included in the system together with the management
and culture audits are opening Romanian SMEs to the
European Union world.

Kovacs S, Human operator assessmentbasis for a safe

workplace in the Process Industry, in Proceedings of the
Hazards XVI SymposiumAnalysing the past, planning
the future, Manchester 68 November 2001, Symposium
Series no. 148, ISBN 0-85295-441-7, pp. 819833.
Kovacs S, Safety as a need-Environment protection as a
deed, in Proceedings of the Hazards XVI Symposium
Analysing the past, planning the future, Manchester
68 November 2001, Symposium Series no. 148, ISBN
0-85295-441-7, pp. 867881.
Kovacs S, Major risk avoidance-fulfilling our responsibilitiescosts and benefits of Romanias safety integration into
the European Union, in Proceedings of the Hazards XVII
Symposiumfulfilling our responsibilities, Manchester
2527 March 2003 Symposium Series no. 149, ISBN
0-85295-459-X, pp. 619633.
Kovacs S, Developing best practice safety procedures through
IT systems, in Proceedings of the Hazards XVII
fulfilling our responsibilities symposium, Manchester
2527 March 2003 Symposium Series no. 149, ISBN
0-85295-459-X, pp. 793807.
Kovacs S, To do no harm- measuring safety performance
in a transition economy, in Proceedings of the Hazards
XVIII Symposium-Process safety-sharing best practice,
Manchester 2325 November 2004, Symposium Series
No. 150, ISBN 0-85295-460-3, pp. 151171.
Kovacs S, Process safety trends in a transition economya case study, In Proceedings of the Hazards XIX
Symposium-Process Safety and Environmental Protection, Manchester 2527 March 2006, Symposium Series
No. 151, ISBN 0-85295-492-8, pp. 152166.
Kovacs S, Meme based cognitive models regarding risk and
loss in small and medium enterprises, In Proceedings of
the Seventh International Conference on Cognitive Modelling, ICCM06, Trieste, Edizioni Goliardiche, ISBN
88-7873-031-9, pp. 377379.
Mahinda Seneviratne and Wai on Phoon, Exposure assessment in SME: a low cost approach to bring oHS Services
to SMEs, in Industrial Health 2006, 44, pp. 2730.
TechDir 2006, SMP 12 Safety Case and Safety Case report.
TUV Rheinland: SEM 28 Safety Assessment service.

103

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

DETECT: A novel framework for the detection of attacks to critical

infrastructures
F. Flammini
ANSALDO STS, Ansaldo Segnalamento Ferroviario S.p.A., Naples, Italy
Universit di Napoli Federico II, Dipartimento di Informatica e Sistemistica, Naples, Italy

A. Gaglione & N. Mazzocca

Universit di Napoli Federico II, Dipartimento di Informatica e Sistemistica, Naples, Italy

C. Pragliola
ANSALDO STS, Ansaldo Segnalamento Ferroviario S.p.A., Naples, Italy

ABSTRACT: Critical Infrastructure Protection (CIP) against potential threats has become a major issue in
modern society. CIP involves a set of multidisciplinary activities and requires the adoption of proper protection
mechanisms, usually supervised by centralized monitoring systems. This paper presents the motivation, the
working principles and the software architecture of DETECT (DEcision Triggering Event Composer & Tracker),
a new framework aimed at the automatic and early detection of threats against critical infrastructures. The
framework is based on the fact that non trivial attack scenarios are made up by a set of basic steps which have to
be executed in a predictable sequence (with possible variants). Such scenarios are identified during Vulnerability
Assessment which is a fundamental phase of the Risk Analysis for critical infrastructures. DETECT operates
by performing a model-based logical, spatial and temporal correlation of basic events detected by the sensorial
subsystem (possibly including intelligent video-surveillance, wireless sensor networks, etc.). In order to achieve
this aim, DETECT is based on a detection engine which is able to reason about heterogeneous data, implementing
a centralized application of data fusion. The framework can be interfaced with or integrated in existing
monitoring systems as a decision support tool or even to automatically trigger adequate countermeasures.
1

INTRODUCTION & BACKGROUND

Critical Infrastructure Protection (CIP) against terrorism and any form or criminality has become
a major issue in modern society. CIP involves a
set of multidisciplinary activities, including Risk
Assessment and Management, together with the
adoption of proper protection mechanisms, usually supervised by specifically designed Security
Management Systems (SMS)1 (see e.g. (LENEL
2008)).
Among the best ways to prevent attacks and disruptions is to stop any perpetrators before they strike.
This paper presents the motivation, the working principles and the software architecture of DETECT
(DEcision Triggering Event Composer & Tracker),

1 In

some cases, they are integrated in the

traditional SCADA (Supervisory Control & Data
Acquisition) systems.

a new framework aimed at the automatic detection of

threats against critical infrastructures, possibly before
they evolve to disastrous consequences. In fact, non
trivial attack scenarios are made up by a set of basic
steps which have to be executed in a predictable
sequence (with possible variants). Such scenarios must
be precisely identified during Vulnerability Assessment which is a fundamental aspect of Risk Analysis
for critical infrastructures (Lewis 2006). DETECT
operates by performing a model-based logical, spatial
and temporal correlation of basic events detected by
intelligent video-surveillance and/or sensor networks,
in order to sniff sequence of events which indicate
(as early as possible) the likelihood of threats. In order
to achieve this aim, DETECT is based on a detection
engine which is able to reason about heterogeneous
data, implementing a centralized application of data
fusion (a well-known concept in the research field of
cognitive/intelligent autonomous systems (Tzafestas
1999)). The framework can be interfaced with or integrated in existing SMS/SCADA systems in order to
automatically trigger adequate countermeasures.

105

With respect to traditional approaches of infrastructure surveillance, DETECT allows for:

A quick, focused and fully automatic response
to emergencies, possibly independent from human
supervision and intervention (though manual confirmation of detected alarms remains an option). In
fact, human management of critical situations, possibly involving many simultaneous events, is a very
delicate task, which can be error prone as well as
subject to forced inhibition.
An early warning of complex attack scenarios since
their first evolution steps using the knowledge base
provided by experts during the qualitative risk analysis process. This allows for preventive reactions
which are very unlikely to be performed by human
operators given the limitation both in their knowledge base and vigilance level. Therefore, a greater
situational awareness can be achieved.
An increase in the Probability Of Detection (POD)
while minimizing the False Alarm Rate (FAR), due
to the possibility of logic as well as temporal correlation of events. While some SMS/SCADA software
offer basic forms of logic correlation of alarms,
the temporal correlation is not implemented in any
nowadays systems, to the best of our knowledge
(though some vendors provide basic options of onsite configurable sequence correlation embedded
in their multi-technology sensors).
The output of DETECT consists of:
The identifier(s) of the detected/suspected scenario(s).
An alarm level, associated to scenario evolution
(only used in deterministic detection as a linear
progress indicator; otherwise, it can be set to 100%).
A likelihood of attack, expressed in terms of probability (only used as a threshold in heuristic detection;
otherwise, it can be set to 100%).
DETECT can be used as an on-line decision support system, by alerting in advance SMS operators
about the likelihood and nature of the threat, as well
as an autonomous reasoning engine, by automatically activating responsive actions, including audio
and visual alarms, emergency calls to first responders, air conditioning flow inversion, activation of
sprinkles, etc.
The main application domain of DETECT is homeland security, but its architecture is suited to other
application fields like environmental monitoring and
control, as well. The framework will be experimented in railway transportation systems, which have
been demonstrated by the recent terrorist strikes to
be among the most attractive and vulnerable targets.
Example attack scenarios include intrusion and drop
of explosive in subway tunnels, spread of chemical

or radiological material in underground stations, combined attacks with simultaneous multiple train halting
and railway bridge bombing, etc. DETECT has proven
to be particularly suited for the detection of such articulated scenarios using a modern SMS infrastructure
based on an extended network of cameras and sensing
devices. With regards to the underlying security infrastructure, a set of interesting technological and research
issues can also be addressed, ranging from object
tracking algorithms to wireless sensor network integration; however, these aspects (mainly application
specific) are not in the scope of this work.
DETECT is a collaborative project carried out by
the Business Innovation Unit of Ansaldo STS Italy
and the Department of Computer and System Science
of the University of Naples Federico II.
The paper is organized as follows: Section 2
presents a brief summary of related works; Section 3
introduces the reference software architecture of the
framework; Section 4 presents the language used to
describe the composite events; Section 5 describes the
implementation of the model-based detection engine;
Section 6 contains a simple case-study application;
Section 7 draws conclusions and provides some hints
about future developments.
2

RELATED WORKS

Composite event detection plays an important role

in the active database research community, which
has long been investigating the application of Event
Condition Action (ECA) paradigm in the context of
using triggers, generally associated with update, insert
or delete operations. In HiPAC (Dayal et al. 1988)
active database project an event algebra was firstly
defined.
Our approach for composite event detection follows
the semantics of the Snoop (Chakravarthy & Mishra
1994) event algebra. Snoop has been developed at the
University of Florida and its concepts have been implemented in a prototype called Sentinel (Chakravarthy
et al. 1994, Krishnaprasad 1994). Event trees are used
for each composite event and these are merged to form
an event graph for detecting a set of composite events.
An important aspect of this work lies in the notion
of parameter contexts, which augment the semantics
of composite events for computing their parameters
(parameters indicate component events). CEDMOS
(Cassandra et al. 1999) refers to the Snoop model
in order to encompass heterogeneity problems which
often appear under the heading of sensor fusion. In
(Alferes & Tagni 2006) the implementation of an event
detection engine that detects composite events specified by expressions of an illustrative sublanguage of
the Snoop event algebra is presented. The engine has
been implemented as a Web Service, so it can also be

106

used by other services and frameworks if the markup

for the communication of results is respected.
Different approaches for composite event detection are taken in Ode (Gerani et al. 1992a, b) and
Samos (Gatziu et al. 1994, Gatziu et al. 2003). Ode
uses an extended Finite Automata for composite event
detection while Samos defines a mechanism based on
Petri Nets for modeling and detection of composite
events for an Object Oriented Data-Base Management
System (OODBMS).
DETECT transfers to the physical security the
concept of Intrusion Detection System (IDS) which
is nowadays widespread in computer (or logical)
security, also borrowing the principles of Anomaly
Detection, which is applied when an attack pattern is
known a priori, and Misuse Detection, indicating the
possibility of detecting unknown attacks by observing
a significant statistical deviation from the normality
(Jones & Sielken 2000). The latter aspect is strictly
related to the field of Artificial Intelligence and related
classification methods.
Intelligent video-surveillance exploits Artificial
Vision algorithms in order to automatically track
object movements in the scene, detecting several type
of events, including virtual line crossing, unattended
objects, aggressions, etc. (Remagnino et al. 2007).
Sensing devices include microwave/infrared/ultrasound volumetric detectors/barriers, magnetic detectors, vibration detectors, explosive detectors, and
advanced Nuclear Bacteriologic Chemical Radiological (NBCR) sensors (Garcia 2001). They can be
connected using both wired and wireless networks,
including ad-hoc Wireless Sensor Networks (WSN)
(Lewis 2004, Roman et al. 2007).
3

Event History database, containing the list of basic

events detected by sensors or cameras, tagged with
a set of relevant attributes including detection time,
event type, sensor id, sensor type, sensor group,
object id, etc. (some of which can be optional, e.g.
object id is only needed when video-surveillance
supports inter-camera object tracking).
Attack Scenario Repository, providing a database
of known attack scenarios as predicted in Risk
Analysis sessions and expressed by means of an
Event Description Language (EDL) including logical as well as temporal operators (derived from
(Chakravarthy et al. 1994)).
Detection Engine, supporting both deterministic (e.g. Event Trees, Event Graphs) and heuristic (e.g. Artificial Neural Networks, Bayesian
Networks) models, sharing the primary requirement
of real-time solvability (which excludes e.g. Petri
Nets from the list of candidate formalisms).
Model Generator, which has the aim of building
the detection model(s) (structure and parameters)
starting from the Attack Scenario Repository by
parsing all the EDL files.
Model Manager, constituted by four sub-modules
(grey-shaded boxes in Figure 1):
Model Feeder (one for each model), which
instantiates the inputs of the detection engine
according to the nature of the models by cyclically
performing proper queries and data filtering on
the Event History (e.g. selecting sensor typologies and zones, excluding temporally distant
events, etc.).
Model Executor (one for each model), which
triggers the execution of the model, once it has
been instantiated, by activating the related (external) solver. An execution is usually needed at each
new event detection.
Model Updater (one for each model), which
is used for on-line modification of the model

THE SOFTWARE ARCHITECTURE

The framework is made up by the following main

modules (see Figure 1):

MODEL(S)
GENERATOR

EDL
REPOSITORY

EVENT
HISTORY
MODEL k
SOLVER

QUERIES

MODEL k
EXECUTOR

MODEL
UPDATER k

MODEL k
FEEDER
INPUTS

DETECTION
MODEL k

SMS / SCADA
ALARMS ->
<- CONFIG
COUNTERMEASURES

OUTPUT
MANAGER

Figure 1.

DETECTION
ENGINE

The software architecture of DETECT.

107

(e.g. update of a threshold parameter), without regenerating the whole model (whenever
supported by the modeling formalism).
Output Manager (single), which stores the output of the model(s) and/or passes it to the interface
modules.

Standard communication protocols (OPC3 ,

ODBC4 , Web-Services, etc.) needed to interoperate with open databases, SMS/SCADA, or any
other client/server security subsystems which are
compliant to such standards.

Model Solver, that is the existing or specifically

developed tool used to execute the model.

The last two points are necessary to provide

DETECT with an open, customizable and easily
upgradeable architecture. For instance, by adopting a standard communication protocol like OPC,
an existing SMS supporting this protocol could integrate DETECT as it was just a further sensing
device.
At the current development state of DETECT:

Model Generator and Model Manager are dependent on the formalisms used to express the models
constituting the Detection Engine. In particular, the
Model Generator and Model Feeder are synergic in
implementing the detection of the event specified in
EDL files: in fact, while the Detection Engine plays
undoubtedly a central role in the framework, many
important aspects are demanded to the way the query
on the database is performed (i.e. selection of proper
events). As an example, in case the Detection Engine
is based on Event Trees (a combinatorial formalism),
the Model Feeder should be able to pick the set of last
N consecutive events fulfilling some temporal properties (e.g. total time elapsed since the first event of the
sequence <T), as defined in the EDL file. In case of
Event Graphs (a state-based formalism), instead, the
model must be fed by a single event at a time.
Besides these main modules, there are others which
are also needed to complete the framework with useful,
though not always essential, features (some of which
can also be implemented by external tools or in the
SMS):

A GUI has been developed to edit scenarios and

generate EDL files starting from the Event Tree
graphical formalism.
A Detection Engine based on Event Graphs (Buss
1996) is already available and fully working, using
a specifically developed Model Solver.
A Model Generator has been developed in order to
generate Event Graphs starting from the EDL files
in the Scenario Repository.
A Web Services based interface has been developed
to interoperate with external SMS.
The issues related to the use of ANN (Jain et al.
1996) for heuristic detection have been addressed
and the related modules are under development and
experimentation.

4
Scenario GUI (Graphical User Interface) used to
draw attack scenarios using an intuitive formalism
and a user-friendly interface (e.g. specifically
tagged UML Sequence Diagrams stored in the
standard XMI2 format (Object Management Group
UML 2008)).
EDL File Generator, translating GUI output into
EDL files.
Event Log, in which storing information about
composite events, including detection time, scenario type, alarm level and likelihood of attack
(whenever applicable).
Countermeasure Repository, associating to each
detected event or event class a set of operations to
be automatically performed by the SMS.
Specific drivers and adapters needed to interface
external software modules, possibly including antiintrusion and video-surveillance subsystems.

THE EVENT DESCRIPTION LANGUAGE

The Detection Engine needs to recognize combination

of events, bound each other with appropriate operators
in order to form composite events of any complexity.
Generally speaking, an event is a happening that occurs
in the system, at some location and at some point in
time. In our context, events are related to sensor data
variables (i.e. variable x greater than a fixed threshold,
variable y in a fixed range, etc.). Events are classified
as primitive events and composite events.
A primitive event is a condition on a specific sensor which is associated some parameters (i.e. event
identifier, time of occurrence, etc.). Event parameters
can be used in the evaluation of conditions. Each entry
stored in the Event History is a quadruple:
< IDev, IDs, IDg, tp >, where:
IDev is the event identifier;
IDs is the sensor identifier;

3 OLE
2 XML

(eXtended
Interchange.

Markup

Language)

Metadata

(Object Linking & Embedding) for Process

Communication.
4 Open Data-Base Connectivity.

108

 IDg is the sensor group identifier (needed for

geographical correlation);
tp is the event occurrence time which should be a
sensor timestamp (when a global clock is available
for synchronization) or the Event History machine
clock.
Since the message transportation time is not instantaneous, the event occurrence time can be different
from the registration time. Several research works
have addressed the issue of clock synchronization in
distributed systems. Here we assume that a proper
solution (e.g. time shifting) has been adopted at a lower
level.
A composite event is a combination of primitive
events defined by means of proper operators. The
EDL of DETECT is derived from Snoop event algebra (Chakravarthy & Mishra 1994). Every composite
event instance is a triple:
<IDec, parcont, te>, where:
IDec is the composite event identifier;
parcont is the parameter context, stating which
occurrences of primitive events need to be considered during the composite event detection (as
described below);
te is the temporal value related to the occurrence of
the composite event (corresponding to the tp of the
last component event).
Formally an event E (either primitive or composite)
is a function from the time domain onto the boolean
values, True and False:
E: T {True, False}, given by:

True, if E occurs at a time t
E (t) =
False, otherwise
The basic assumption of considering a boolean
function is quite general, since different events can
be associated to a continuous sensor output according
to a set of specified thresholds. Furthermore, negate
conditions (!E) can be used when there is the need for
checking that an event is no longer occurring. This
allows considering both instantaneous (occurs =
has occurred) and continuous (occurs = is
occurring) events. However, in order to simplify
EDL syntax, negate conditions on events can be substituted by complementary events. An event Ec is
complementary to E when:
Ec !E
Each event is denoted by an event expression, whose
complexity grows with the number of involved events.
Given the expressions E1 , E2 , . . . , En , every application on them through any operator is still an expression.

In the following, we briefly describe the semantics of

these operators. For a formal specification of these
semantics, the reader can refer to (Chakravarthy et al.
1994).
OR. Disjunction of two events E1 and E2 , denoted
(E1 OR E2 ). It occurs when at least one of its
components occurs.
AND. Conjunction of two events E1 and E2 ,
denoted (E1 AND E2 ). It occurs when both E1 and
E2 occur (the temporal sequence is ignored).
ANY. A composite event, denoted ANY (m, E1 ,
E1 , . . . , En ), where m n. It occurs when m out of n
distinct events specified in the expression occur (the
temporal sequence is ignored).
SEQ. Sequence of two events E1 and E2 , denoted
(E1 SEQ E2 ). It occurs when E2 occurs provided that
E1 has already occurred. This means that the time
of occurrence of E1 has to be less than the time of
occurrence of E2 .
The sequence operator is used to define composite
events when the order of its component events is relevant. Another way to perform a time correlation on
events is by exploiting temporal constraints.
The logic correlation could loose meaningfulness
when the time interval between component events
exceeds a certain threshold. Temporal constraints can
be defined on primitive events with the aim of defining a validity interval for the composite event. Such
constraints can be added to any operator in the formal
expression used for event description.
For instance, let us assume that in the composite
event E = (E1 AND E2 ) the time interval between the
occurrence of primitive events E1 and E2 must be at
most T. The formal expression is modified by adding
the temporal constraint [T] as follows:
(E1 AND E2 )[T] = True

t1 t|(E1 (t) E2 (t1 ) E1 (t1 ) E2 (t)) |t t1 | T

THE SOFTWARE IMPLEMENTATION

This section describes some implementation details of

DETECT, referring to the current development state
of the core modules of the framework, including the
Detection Engine. The modules have been fully implemented using the Java programming language. JGraph
has been employed for the graphical construction of
the Event Trees used in the Scenario GUI. Algorithms
have been developed for detecting composite events in
all parameter contexts.
Attack scenarios are currently described by Event
Trees, where leaves represent primitive events while

109

 Chronicle: the (initiator, terminator) pair is unique.

The oldest initiator is paired with the oldest terminator.
Continuous: each initiator starts the detection of the
event.
Cumulative: all occurrences of primitive events are
accumulated until the composite event is detected.

Figure 2. Event tree for composite event ((E1 OR E2) AND

(E2 SEQ (E4 AND E6))).

internal nodes (including the root) represent EDL language operators. Figure 2 shows an example Event
Tree representing a composite event.
After the user has sketched the Event Tree, the Scenario GUI module parses the graph and provides the
EDL expression to be added to the EDL Repository.
The parsing process starts from the leaf nodes representing the primitive events and ends at the root
node. Starting from the content of the EDL Repository,
the Model Generator module builds and instantiates
as many Event Detector objects as many composite
events stored in the database. The detection algorithm
implemented by such objects is based on Event Graphs
and the objects include the functionalities of both the
Model Solver and the Detection Engine.
In the current prototype, after the insertion of attack
scenarios, the user can start the detection process on
the Event History using a stub front-end (simulating
the Model Executor and the Output Manager modules). A primitive event is accessed from the database
by a specific Model Feeder module, implemented by a
single Event Dispatcher object which sends primitive
event instances to all Event Detectors responsible for
the detection process.
The Event Dispatcher requires considering only
some event occurrences, depending on a specific
policy defined by the parameter context. The policy
is used to define which events represent the beginning (initiator) and the end (terminator) of the scenario. The parameter context states which component
event occurrences play an active part in the detection process. Four contexts for event detection can be
defined:
Recent: only the most recent occurrence of the
initiator is considered.

The effect of EDL operators is then conditioned

by the specific context, which is implemented in the
Event Dispatcher. Theoretically, in the construction of
the model a different node should be defined for each
context. Whilst a context could be associated to each
operator, currently a single context is associated to
each detection model. Furthermore, a different node
object for each context has been implemented.
In the current implementation, Event Graphs are
used to detect the scenarios defined by Event Trees,
which are only used as a descriptive formalism. In
fact, scenarios represented by more Event Trees can
be detected by a single Event Graph produced by the
Model Generator. When an Event Detector receives
a message indicating that an instance of a primitive
event Ei has occurred, it stores the information in the
node associated with Ei . The detection of composite events follows a bottom-up process that starts from
primitive event instances and flows up to the root node.
So the composite event is detected when the condition
related to the root node operator is verified. The propagation of the events is determined by the user specified
context. After the detection of a composite event, an
object of a special class (Event Detected) is instantiated with its relevant information (identifier, context,
component event occurrences, initiator, terminator).

AN EXAMPLE SCENARIO

In this section we provide an application of DETECT

to the case-study of a subway station. We consider
a composite event corresponding to a terrorist threat.
The classification of attack scenarios is performed by
security risk analysts in the vulnerability assessment
process.
The attack scenario consists of an intrusion and drop
of an explosive device in a subway tunnel. Let us suppose that the dynamic of the scenario follows the steps
reported below:
1. The attacker stays on the platform for the time
needed to prepare the attack, missing one or more
trains.
2. The attacker goes down the tracks by crossing the
limit of the platform and moves inside the tunnel
portal.
3. The attacker drops the bag containing the explosive
device inside the tunnel and leaves the station.

110

Obviously, it is possible to think of several variants

of this scenario. For instance, only one between step 1
and step 2 could happen. Please note that the detection
of step 1 (person not taking the train) would be very
difficult to detect by a human operator in a crowded
station due to the people going on and off the train.
Le us suppose that the station is equipped with
a security system including intelligent cameras (S1 ),
active infrared barriers (S2 ) and explosive sniffers (S3 )
for tunnel portal protection. The formal description of
the attack scenario consists of a sequence of events
which should be detected by the appropriate sensors
and combined in order to form the composite event.
The formal specification of primitive events constituting the scenario is provided in following:
a.
b.
c.
d.
e.

extended presence on the platform (E1 by S1 );

train passing (E2 by S1 );
platform line crossing (E3 by S1 );
tunnel intrusion (E4 by S2 );
explosive detection (E5 by S3 ).

For the sake of brevity, further steps are omitted.

The composite event drop of explosive in tunnel
can be specified in EDL as follows:
(E1 AND E2 ) OR E3 SEQ (E4 AND E5 )
Figure 3 provides a GUI screenshot showing the
Event Tree for the composite event specified above.
The user chooses the parameter context and builds the
tree (including primitive events, operators and interconnection edges) by the user-friendly interface. If a
node represents a primitive event, the user has to specify event (Ex ) and sensor (Sx ) identifiers. If a node
is an operator, the user can optionally specify other
parameters such as a temporal constraint, the partial
alarm level and the m parameter (ANY operator). Also,
the user can activate/deactivate the composite events
stored in the repository carrying out the detection
process.

A partial alarm can be associated to the scenario

evolution after step 1 (left AND in the EDL expression), in order to warn the operator of a suspect
abnormal behavior.
In order to activate the detection process, a simulated Event History has been created ad-hoc. An
on-line integration with a real working SMS will
be performed in the near future for experimentation
purposes.

CONCLUSIONS & FUTURE WORKS

In this paper we have introduced the working principles

and the software architecture of DETECT, an expert
system allowing for early warnings in security critical
domains.
DETECT can be used as a module of a more complex hierarchical system, possibly involving several
infrastructures. In fact, most critical infrastructures
are organized in a multi-level fashion: local sites,
grouped into regions and then monitored centrally by
a national control room, where all the (aggregated)
events coming from lower levels are routed. When
the entire system is available, each site at each level
can benefit from the knowledge of significant events
happening in other sites. When some communication
links are unavailable, it is still possible to activate
countermeasures basing on the local knowledge.
We are evaluating the possibility of using a single
automatically trained multi-layered ANN to complement deterministic detection by: 1) classification of
suspect scenarios, with a low FAR; 2) automatic detection of abnormal behaviors, by observing deviations
from normality; 3) on-line update of knowledge triggered by the user when a new anomaly has been
detected. The ANN model can be trained to understand normality by observing the normal use of the
infrastructure, possibly for long periods of time. The
Model Feeder for ANN operates in a way which is
similar to the Event Tree example provided above. A
ANN specific Model Updater allows for on-line learning facility. Future developments will be aimed at a
more cohesive integration between deterministic and
heuristic detection, by making the models interact one
with each other.

REFERENCES

Figure 3.

Insertion of the composite event using the GUI.

Alferes, J.J. & Tagni, G.E. 2006. Implementation of a Complex Event Engine for the Web. In Proceedings of IEEE
Services Computing Workshops (SCW 2006). September
1822. Chicago, Illinois, USA.
Buss, A.H. 1996. Modeling with Event Graphs. In Proc.
Winter Simulation Conference, pp. 153160.

111

Cassandra, A.R., Baker, D. & Rashid, M. 1999. CEDMOS:

Complex Event Detection and Monitoring System. MCC
Tecnical Report CEDMOS-002-99, MCC, Austin, TX.
Chakravarthy, S. & Mishra, D. 1994. Snoop: An expressive
event specification language for active databases. Data
Knowl. Eng., Vol. 14, No. 1, pp. 126.
Chakravarthy, S., Krishnaprasad, V., Anwar, E. & Kim, S.
1994. Composite Events for Active Databases: Semantics, Contexts and Detection. In Proceedings of the
20th international Conference on Very Large Data Bases
(September 1215, 1994).
Bocca, J.B., Jarke, M. & Zaniolo, C. Eds. Very Large Data
Bases. Morgan Kaufmann Publishers, San Francisco, CA,
pp. 606617.
Dayal, U., Blaustein, B.T., Buchmann, A.P., Chakravarthy, S.,
Hsu, M., Ledin, R., McCarthy, D.R., Rosenthal, A., Sarin,
S.K., Carey, M.J., Livny, M. & Jauhari, R. 1988. The
HiPAC Project: Combining Active Databases and Timing
Constraints. SIGMOD Record, Vol. 17, No. 1, pp. 5170.
Garcia, M.L. 2001. The Design and Evaluation of Physical
Protection Systems. Butterworth-Heinemann, USA.
Gatziu, S. & Dittrich, K.R. 1994. Detecting Composite
Events in Active Databases Using Petri Nets. In Proceedings of the 4th International Workshop on Research Issues
in data Engineering: Active Database Systems, pp. 29.
Gatziu, S. & Dittrich, K.R. 2003. Events in an ObjectOriented Database System. In Proceedings of the 1st
International.
Gerani, N.H., Jagadish, H.V. & Shmueli, O. 1992a. Event
Specification in an Object-Oriented Database. In
Gerani, N.H., Jagadish, H.V. & Shmueli, O. 1992b. COMPOSE A System For Composite Event Specification and
Detection. Technical report, AT&T Bell Laboratories,
Murray Hill, NJ.

Jain, A.K., Mao, J. & Mohiuddin, K.M. 1996. Artificial Neural Networks: A tutorial. In IEEE Computer, Vol. 29, No.
3, pp. 5663.
Jones, A.K. & Sielken, R.S. 2000. Computer System Intrusion Detection: A Survey. Technical Report, Computer
Science Dept., University of Virginia.
Krishnaprasad, V. 1994. Event Detection for Supporting
Active Capability in an OODBMS: Semantics, Architecture and Implementation. Masters Thesis. University of
Florida.
LENEL OnGuard 2008. http://www.lenel.com.
Lewis, F.L. 2004. Wireless Sensor Networks. In Smart Environments: Technologies, Protocols, and Applications, ed.
D.J. Cook and S.K. Das. John Wiley, New York.
Lewis, T.G. 2006. Critical Infrastructure Protection in
Homeland Security: Defending a Networked Nation. John
Wiley, New York.
Object Management Group UML, 2008. http://www.omg.
org/uml.
OLE for Process Communication. http://www.opc.org.
Remagnino, P., Velastinm, S.A., Foresti G.L. & Trivedi, M.
2007. Novel concepts and challenges for the next generation of video surveillance systems. In Machine Vision and
Applications (Springer), Vol. 18, Issue 34, pp. 135137.
Roman, R., Alcaraz, C. & Lopez, J. 2007. The role of Wireless Sensor Networks in the area of Critical Information
Infrastructure Protection. In Information Security Tech.
Report, Vol. 12, Issue 1, pp. 2431.
Tzafestas, S.G. 1999. Advances in Intelligent Autonomous
Systems. Kluwer.

112

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Methodology and software platform for multi-layer causal modeling

K.M. Groth, C. Wang, D. Zhu & A. Mosleh
Center for Risk and Reliability, University of Maryland, College Park, Maryland, USA

ABSTRACT: This paper introduces an integrated framework and software platform that uses a three layer
approach to modeling complex systems. The multi-layer PRA approach implemented in IRIS (Integrated Risk
Information System) combines the power of Event Sequence Diagrams and Fault Trees for modeling risk
scenarios and system risks and hazards, with the flexibility of Bayesian Belief Networks for modeling nondeterministic system components (e.g. human, organizational). The three types of models combined in the IRIS
integrated framework form a Hybrid Causal Logic (HCL) model that addresses deterministic and probabilistic
elements of systems and quantitatively integrates system dependencies. This paper will describe the HCL
algorithm and its implementation in IRIS by use of an example from aviation risk assessment (a risk scenario
model of aircraft taking off from the wrong runway.

INTRODUCTION

Conventional Probabilistic Risk Assessment (PRA)

methods model deterministic relations between basic
events that combine to form a risk scenario. This
is accomplished by using Boolean logic methods,
such as Fault Trees (FTs) and Event Trees (ETs) or
Event Sequence Diagrams (ESDs). However since
with human and organizational failures are among
the most important roots of many accidents and incidents, there is an increased interest in expanding causal
models to incorporate non-deterministic causal links
encountered in human reliability and organizational
theory. Bayesian Belief Networks (BBNs) have the
capability to model these soft relationships.
This paper describes a new risk methodology known
as Hybrid Causal Logic (HCL) that combines the
Boolean logic-based PRA methods (ESDs, FTs) with
BBNs. The methodology is implemented in a software
package called the Integrated Risk Information System (IRIS). The HCL computational engine of IRIS
can also be used as a standalone console application.
The functionality of this computational engine can be
accessed by other applications through an API. IRIS
was designed for the United States Federal Aviation
Administration. Additional information on IRIS can
be found in the references (Groth, Zhu & Mosleh 2008;
Zhu et al 2008; Groth 2007).
In this modeling framework risk scenarios are modeled in the top layer using Event Sequence Diagrams.
In the second layer, Fault Trees are used to model the
factors contributing to the properties and behaviors of
the physical system (hardware, software, and environmental factors). Bayesian Belief Networks comprise

the third layer to extend the causal chain of events to

potential human, organizational, and socio-technical
roots.
This approach can be used as the foundation for
addressing many of the issues that are commonly
encountered in risk and safety analysis and hazard
identification. As a causal model, the methodology
provides a vehicle for identification and analysis of
cause-effect relationships across many different modeling domains, including human, software, hardware,
and environment.
The IRIS framework can be used to identify all risk
scenarios and contributing events and calculate associated probabilities; to identify the risk value of specific
changes and the risk importance of certain elements;
and to monitor system risk indicators by considering
the frequency of observation and the risk significance over a period of time. Highlighting, trace, and
drill down functions are provided to facilitate hazard
identification and navigation through the models.
All of the features in IRIS can be implemented with
respect to one risk scenario or multiple scenarios, e.g.
all of the scenarios leading to a particular category or
type of end state. IRIS can be used to build a single
one-layer model, or a network of multi-layer models.
IRIS was developed as part of an international research effort sponsored by the FAA System
Approach for Safety Oversight (SASO) office. Other
parts of this research created ESDs, FTs, and BBNs by
teams of aviation experts from the United States and
Europe. IRIS integrates the different models into a
standard framework and the HCL algorithm combines
quantitative information from the models to calculate
total risk.

113

The Dutch National Aerospace Laboratory (NLR)

used the NLR air safety database and aviation experts
to created a hierarchical set of 31 generic ESDs representing the possible accident scenarios from takeoff
to landing (Roelen et al. 2002).
Another layer of the aviation safety model was created by Hi-Tec Systems. Hi-Tec created a comprehensive model for the quality of air carrier maintenance
(Eghbali 2006) and the flight operations (Mandelapu
2006). NLR has also created FTs for specific accident
scenarios (Roelen & Wever 2004a, b).
The NLR and Hi-Tec models were built and analyzed in IRIS. One set of models pertains to the use
of the incorrect runway during takeoff. These models
became especially pertinent after the August 2006 fatal
Comair Flight 5191 crash in Lexington, Kentucky. The
pilot of flight 5191 taxied onto the wrong runway during an early morning takeoff due to a combination of
human and airport factors. The incorrect runway was
shorter than the minimum distance required for the aircraft to takeoff. The aircraft was less than 300ft from
the end of the runway before pilots realized the error
and attempted to takeoff at below-optimal speed. The
attempted takeoff resulted in a runway overrun and the
death of 49 of the 50 people onboard.
The NTSB (2007) cited human actions by crew
and air traffic control (ATC) contributing to the accident. The crew violated cockpit policy by engaging in
non-pertinent conversation during taxiing and by completing an abbreviated taxi briefing. Signs indicating
the runway number and cockpit displays indicating the
direction of takeoff were not mentioned by either pilot
during the takeoff. During takeoff the flight crew noted
that there were no lights on the runway as expected, but
did not double check their position as the copilot had
observed numerous lights out on the correct runway
the previous day. Pre-flight paperwork also indicated
that the centerline lights on the proper runway were
out. The flight crew did not use the available cues to
reconsider takeoff.
At the time of the accident only one of two required
air traffic controllers were on duty. According to postaccident statements, the controller on duty at the time
of the accident was also responsible for monitoring
radar and was not aware that the aircraft had stopped
short of the desired runway before he issued takeoff
clearance. After issuing takeoff clearance the controller turned around to perform administrative tasks
during take-off and was not engaged in monitoring the
progress of the flight. Fatigue likely contributed to the
performance of the controller as he had only slept for
2 hours in the 24 hours before the accident.
Impaired decision making and inappropriate task
prioritization by both crew members and ATC were
major contributing factors to this accident. The reducing lighting on both the correct and incorrect runways
at the airport contributed to the decision errors made by

crew and fatigue and workload contributed to decision

errors made by ATC. The details from the flight 5191
and the group of models for use of the incorrect runway during takeoff will be used throughout this paper
to show how the HCL methodology can be applied to
a real example.

OVERVIEW OF HCL METHODOLOGY

2.1

Overview of the HCL modeling layers

The hybrid causal logic methodology extends conventional deterministic risk analysis techniques to
include soft factors including the organizational and
regulatory environment of the physical system. The
HCL methodology employs a model-based approach
to system analysis; this approach can be used as the
foundation for addressing many of the issues that are
commonly encountered in system safety assessment,
hazard identification analysis, and risk analysis. The
integrated framework is presented in Figure 1.
ESDs form the top layer of the three layer model,
FTs form the second layer, and BBNs form the bottom
layer. An ESD is used to model temporal sequences of
events. ESDs are similar to event trees and flowcharts;
an ESD models the possible paths to outcomes, each
of which could result from the same initiating event.
ESDs contain decision nodes where the paths diverge
based on the state of a system element. As part of the
hybrid causal analysis, the ESDs define the context
or base scenarios for the hazards, sources of risk, and
safety issues.
The ESD shown in Figure 2 models the probability
of an aircraft taking off safely, stopping on the runway, or overrunning the runway. As can be seen in
the model, the crew must reject the takeoff and the
speed of the aircraft must be lower than the critical
speed beyond which the aircraft cannot stop before the

IE

PE-1

PE-2

PE-3

E
C

Figure 1.

114

Top layer: ESD

Middle layer: FT
Bottom layer: BBN

Illustration of a three-layered IRIS model.

Figure 2.

Case study top layerESD for an aircraft using the wrong runway (Roelen et al. 2002).

Figure 3. Case study middle layerFT for air traffic control

events (Roelen and Wever 2004a).

end of the runway. By the time the Flight 5191 crew

realized the mistake, the plane was above critical speed
and the runway overrun was inevitable.
The initiating event of the ESD, ATC event, is
directly linked to the top gate of the FT in Figure 3.
This FT provides three reasons an aircraft could be
placed in this situation: loss of separation with traffic,
takeoff from incorrect runway, or a bird strike.
FTs uses logical relationships (AND, OR, NOT,
etc.) to model the physical behaviors of the system. In
an HCL model, the top event of a FT can be connected
to any event in the ESD. This essentially decomposes
the ESD event into a set of physical elements affecting
the state of the event, with the node in the ESD taking
its probability value from the FT.
BBNs have been added as the third layer of the
model. A BBN is a directed acyclic graph, i.e. it cannot contain feedback loops. Directed arcs form paths
of influence between variables (nodes). The addition

Figure 4. Partial NLR model for takeoff from wrong runway. The flight plan node is fed by Figure 5, and the crew
decision/action error is fed by additional human factors.

of BBNs to the traditional PRA modeling techniques

extends conventional risk analysis by capturing the
diversity and complexity of hazards in modern systems. BBNs can be used to model non-deterministic
casual factors such as human, environmental and
organizational factors.
BBNs offer the capability to deal with sequential
dependency and uncertain knowledge. BBNs can be
connected to events in ESDs and FTs. The connections
between the BBNs and logic models are formed by
binary variables in the BBN; the probability of the
linked BBN node is then assigned to the ESD or FT
event.
The wrong runway event in the center of the FT is
the root cause of the accident. Factors that contribute
to this root cause are modeled in the BBNs in Figure 4
and 5. Figure 4 is part of the wrong runway BBN
developed by NLR (Roelen and Wever 2007); the
wrong runway FT event is linked to the output node

115

Figure 5.

Case study bottom layerBBN of flight operations (Mandelapu 2006).

of this BBN. The flight plan node in Figure 5 feeds

into the wrong runway node in Figure 4. Figure 5
is fed by the Hi-Tec air carrier maintenance model
(Eghbali 2006; not pictured) with the end node of the
maintenance model feeding information into the fleet
availability node at the top of the flight operations
model.
Since many of the casual factors in BBNs may have
widespread influence, BBN nodes may impact multiple events within ESDs and FTs. The details of the
HCL quantification procedure can be found in the
references (Groen & Mosleh 2008, Wang 2007).
2.2

Overview of HCL algorithm quantitative

capabilities

An ESD event can be quantified directly by inputting a

probability value for the event, or indirectly by linking
it to a FT or a node in a BBN. Linked ESD events take
on the probability value of the FT or node attached to it.
This allows the analyst to set a variable probability for
ESD events based on contributing factors from lower
layers of the model. Likewise, FT basic events can be
quantified directly or linked to any node in the BBN.

BBN nodes are quantified in conditional probability tables. The size of the conditional probability
table for each node depends on the number of parent
nodes leading into it. The conditional probability table
requires the analyst to provide a probability value for
each state of the child node based on every possible
combination of the states of parent nodes. The default
number of states for a BBN node is 2, although additional states can be added as long as the probability of
all states sums to 1. Assuming the child and its n parent nodes all have 2 states, this requires 2n probability
values.
In order to quantify the hybrid model it is necessary to convert the three types of diagrams into a set
of models that can communicate mathematically. This
is accomplished by converting the ESDs and FTs into
Reduced Ordered Binary Decision Diagrams (BDDs).
The set of reduced ordered BDDs for a model are all
unique and the order of variables along each path from
root node to end node is identical. Details on the algorithms used to convert ESDs and FTs into BDDs have
been described extensively (Bryant 1992, Brace et al.
1990, Rauzy 1993, Andrews & Dunnett 2000, Groen
et al. 2005).

116

BBNs are not converted into BDDs; instead, a

hybrid BDD/BBN is created. In this hybrid structure,
the probability of one or more of the BDD variables
is provided by a linked node in the BBN. Additional
details about the BDD/BBN link can be found in
Groen & Mosleh (2008).
3

HCL-BASED RISK MANAGEMENT

METRICS

In addition to providing probability values for each

ESD scenario, each FT and each BBN node, the HCL
methodology provides functions for tracking risks over
time and for determining the elements that contribute
most to scenario risk. HCL also provides the minimal cut-sets for each ESD scenario, allowing the user
to rank the risk scenarios quantitatively. Specific functions are described in more detail below. For additional
technical details see (Mosleh et al. 2007).
Figure 7 displays scenario results for the wrong runway scenario. It is clear that the most probable outcome

of using the wrong runway is a continued takeoff with

no consequences. The bottom half of the figure displays the cut-sets only for the scenarios that end with
a runway overrun. As can be seen in the figure, the
most likely series of event reading to an overrun is the
combination of using the incorrect runway, attempting a rejected takeoff, and having speed in excess of
the critical stopping speed (V1). This is the pattern
displayed by flight 5191.

3.1

Importance measures

Importance measures are used to identify the most

significant contributors to a risk scenario. They provide a quantitative way to identify the most important
system hazards and to understand which model elements most affect system risk. Importance measures
can be used to calculate the amount of additional safety
resulting from a system modification, which allows
analysts to examine the benefits of different modifications before implementation. Analysts can also use

Figure 6.

Probability values and cut sets for the base wrong runway scenario.

Figure 7.

Importance measure results for the runway overrun model.

117

importance measures to identify the elements that most

contribute to a risk scenario and then target system
changes to maximize the safety impact.
There are numerous ways to calculate importance
measures for Boolean models. However, due to the
dependencies in HCL models introduced by inclusion of BBNs, the methods cannot be applied in their
original form. Four conventional importance measures
have been modified and implemented in HCL: Risk
Achievement Work (RAW), Risk Reduction Worth
(RRW), Birnbaum, and Vesely-Fussel (VF).
The standard Vesely-Fussel importance measure
(Eq. 1) calculates the probability that event e has
occurred given that ESD end state S has occurred
(Fussel 1975).
p(e|S) =

p(e S)
P(S)

system risks and to track changes in risk over time.

Risk significance is calculated with respect to selected
ESD scenarios or end states. It can be calculated for
any BBN node, FT gate or event, or ESD pivotal event.
The risk indicator is calculated by Equation 3,
where R is the total risk, is the frequency of the
event.
R = Pr(S| f )

Pr(S| f ) is the risk weight of a BBN node or FT event or

gate ( f ) and S is the selected ESD end state or group of
end states. If S consists of an end state category or multiple end states in the same ESD Equation 3 is modified
using the same logic explained for modifying Equation 1. For multiple end states in different ESDs the
risk indicator value can be calculated using the upper
bound approximation. The procedure for performing
precursor analysis and hazard ranking follows directly
from the risk indicator procedure.
Figure 8 displays individual risk indicators and total
risk for several BBN nodes from the example model.
Frequency values are to be provided by the analyst. In
the example case, the frequency values were selected
to show how IRIS could be used to monitor the risks
before the accident; these are not values from data.
The top graph in Figure 8 shows the changing risk
values for each of the three selected indicators. The
bottom graph shows the aggregated risk value over
time. Based on the risk values obtained from the models and the hypothetical frequency data, it becomes
apparent that the risk associated with airport adequacy increased sharply between May and July. The
hypothetical risks associated with the adequacy of the
airport could have been identified in July and steps

(1)

For hybrid models, event e is a given state of a model

element, e.g. a FT event is failed or a BBN node is
degraded instead of fully functional or failed.
By addressing a particular state, it is possible to extend
importance measures to all layers of the hybrid model.
Importance measures must be calculated with
respect to an ESD end state. To ensure independence
in ESDs with multiple paths, it is necessary to treat
the end state S as the sum of the Si mutually exclusive
paths leading to it. The importance measure can then
be calculated by using Equation 2.
p(S) =



p(Si |e)
i p(e Si )

= i
p(S
)
i
i
i p(Si )

(2)

For a set of scenarios belonging to two or more

ESDs the probability can be calculated as a function of the results from each ESD or by use of the
mean upper bound approximation. Additional details
on HCL importance measure calculations can be found
in Zhu (2008).
Figure 7 provides importance measure results for
the runway overrun model. The importance measures
in the figure are arranged by the Vesely-Fussel importance measure. The items in the components column
are FT events and some selected BBN nodes. The BBN
nodes selected reflect the factors that contributed to the
runway overrun of Flight 5191. From the figure it is
obvious that take-off from incorrect runway is the most
important contributing factor to the runway overrun
end state.
3.2

Risk indicators

Risk indicators are used to identify potential system

risks by considering both the risk significance and the
frequency of the event. They are also used to monitor

(3)

Figure 8.

118

Sample risk indicators implemented in IRIS.

could have been taken to reduce these risks before

serious consequences occurred.
3.3

Risk impact

Analysts can use IRIS to visualize the change in system

risk based on observed or postulated conditions. This
can be achieved by using the set evidence function
to make assumptions about the state of one or more
BBN nodes. Once assumptions are made the model is
updated to reflect the new information, providing new
probability values for all nodes subsequently affected
by the changes.
When the BBN is linked to an ESD or FT, the new
ESD and FT models will also display new probability values. The set evidence function allows users to
see the impact of soft factors on risk scenarios. The
result is a more tangible link between the actions of
humans/organizations and specific system outcomes.
Setting evidence will provide users with a better
understanding of how low-level problems propagate
through the system and combine to form risk scenarios. Figure 9 displays updated scenario results for the
flight 5191 overrun. In this scenario, evidence was set
for three nodes in the BBN (Fig. 5). Human actions
was set to the state unsafe because of errors made by

Figure 9.

the flight 5191 flight crew. Airport adequacy was set to

the state inadequate because of the lack of proper lighting on both runways. The takeoff plan was deemed
substandard.
By comparing the results of the base case, Figure 6,
to the case updated with scenario evidence, Figure 9,
it is possible to quantify the change in risk accompany
certain behaviors. The updated probability of a runway
overrun based on human actions, airport conditions,
and the takeoff plan is an order of magnitude greater
than the probability of the base scenario. Again, the
series of events leading to the flight 5191 crash is the
most probable sequence leading to an overrun in the
model.
It is evident from Figure 10 that the three BBN
nodes strongly impact the probability of taking off
from the incorrect runway. This probability increases
by almost a factor of 2 when the model is updated with
the scenario evidence.

CONCLUSION

This paper provides an overview of the hybrid

causal logic (HCL) methodology for Probabilistic Risk

Updated scenario results for the runway overrun with information about flight 5191 specified.

Figure 10. Fault tree results showing the probability of taking off from the wrong runway for the base case (top) and the case
reflecting flight 5191 factors (bottom).

119

Assessment and the IRIS software package developed

to use the HCL methodology for comprehensive risk
analyses of complex systems. The HCL methodology and the associated computational engine were
designed to be portable and thus there is no specific
HCL GUI. The computational engine can read models from files and can be accessed through use of
an API.
The three-layer The flexible nature of the HCL
framework allows a wide range of GUIs to be developed for many industries. The IRIS package is
designed to be used by PRA experts and systems analysts. Additional GUIs can be added to allow users
outside of the PRA community to use IRIS without in
depth knowledge of the modeling concepts and all of
the analysis tool.
Two FAA specific GUIs were designed with two
different target users in mind. Target users provided
information about what information they needed from
IRIS and how they would like to see it presented. The
GUIs were linked to specific IRIS analysis tools, but
enabled the results to be presented in a more qualitative
(e.g. high/medium/low) way.
The GUIs were designed to allow target users to
operate the software immediately. Users are also able
to view underlying models and see full quantitative
results if desired.
HCL framework was applied to the flight 5191
runway overrun event from 2006, and the event was
analyzed based on information obtained about the
conditions contributing to the accident.
The three layer HCL framework allows different
modeling techniques to be used for different aspects of
a system. The hybrid framework goes beyond typical
PRA methods to permit the inclusion of soft causal factors introduced by human and organizational aspects
of a system. The hybrid models and IRIS software
package provide a framework for unifying multiple
aspects of complex socio-technological systems to perform system safety analysis, hazard analysis and risk
analysis.
The methodology can be used to identify the most
important system elements that contribute to specific outcomes and provides decision makers with a
quantitative basis for allocating resources and making
changes to any part of a system.
ACKNOWLEDGEMENT
The work described in this paper was supported by
the US Federal Aviation Administration. The authors
are indebted to John Lapointe and Jennelle Derrickson (FAAWilliam J. Hughes Technical Center) for
overall monitoring and coordination. The opinions
expressed in this paper are those of the authors and
do not reflect any official position by the FAA.

REFERENCES
Andrews, J.D. & Dunnett, S.J. 2000. Event Tree Analysis
using Binary Decision Diagrams. IEEE Transactions on
Reliability 49(2): 230239.
Brace, K., Rudell, R. & Bryant, R. 1990. Efficient Implementation of a BDD Package. The 27th ACM/IEEE Design
Automation Conference, IEEE 0738.
Bryant, R. 1992. Symbolic Boolean Manipulation with
Ordered Binary Decision Diagrams. ACM Computing
Surveys 24(3): 293318.
Eghbali, G.H. 2006. Causal Model for Air Carrier Maintenance. Report Prepared for Federal Aviation Administration. Atlantic City, NJ: Hi-Tec Systems.
Fussell, J.B. 1975. How to Hand Calculate System Reliability and Safety Characteristics. IEEE Transactions on
Reliability R-24(3): 169174.
Groen, F. & Mosleh, A. 2008 (In Press). The Quantification of Hybrid Causal Models. Submitted to Reliability
Engineering and System Safety.
Groen, F., Smidts, C. & Mosleh, A. 2006. QRASthe quantitative risk assessment system. Reliability Engineering
and System Safety 91(3): 292304.
Groth, K. 2007. Integrated Risk Information System Volume 1: User Guide. College Park, MD: University of
Maryland.
Groth, K., Zhu, D. & Mosleh, A. 2008. Hybrid Methodology
and Software Platform for Probabilistic Risk Assessment. The 54th Annual Reliability and Maintainability
Symposium, Las Vegas, NV.
Mandelapu, S. 2006. Causal Model for Air Carrier Maintenance. Report Prepared for Federal Aviation Administration. Atlantic City, NJ: Hi-Tec Systems.
Mosleh, A. et al. 2004. An Integrated Framework for
Identification, Classification and Assessment of Aviation Systems Hazards. The 9th International Probabilistic
Safety Assessment and Management Conference. Berlin,
Germany.
Mosleh, A., Wang, C. & Groen, F. 2007. Integrated Methodology For Identification, Classification and Assessment
of Aviation Systems Hazards and Risks Volume 1: Framework and Computational Algorithms. College Park, MD:
University of Maryland.
National Transportation Safety Board (NTSB) 2007. Aircraft
Accident Report NTSB/AAR-07/05.
Rauzy, A. 1993. New Algorithms for Fault Trees Analysis.
Reliability Engineering and System Safety 40: 203211.
Roelen, A.L.C. & Wever, R. 2004a. A Causal Model
of Engine Failure, NLR-CR-2004-038. Amsterdam:
National Aerospace Laboratory NLR.
Roelen, A.L.C. &. Wever, R. 2004b. A Causal Model of
A Rejected Take-Off. NLR-CR-2004-039. Amsterdam:
National Aerospace Laboratory NLR.
Roelen, A.L.C et al. 2002. Causal Modeling of Air Safety.
Amsterdam: National Aerospace Laboratory NLR.
Wang, C. 2007. Hybrid Causal Methodology for Risk Assessment. PhD dissertation. College Park, MD: University of
Maryland.
Zhu, D. et al. 2008. A PRA Software Platform for Hybrid
Causal Logic Risk Models. The 9th International Probabilistic Safety Assessment and Management Conference.
Hong Kong, China.

120

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

SCAIS (Simulation Code System for Integrated Safety Assessment):

Current status and applications
J.M. Izquierdo, J. Hortal, M. Snchez, E. Melndez & R. Herrero
Consejo de Seguridad Nuclear (CSN), Madrid, Spain

J. Gil, L. Gamo, I. Fernndez, J. Espern & P. Gonzlez

Indizen Technologies S.L., Madrid, Spain

C. Queral, A. Expsito & G. Rodrguez

Universidad Politcnica Madrid UPM, Spain

ABSTRACT: This paper surveys the current status of Spanish Nuclear Safety Council (CSN) work made to
establish an Integrated Safety Analysis (ISA) methodology, supported by a simulation framework called SCAIS,
to independently check the validity and consistency of many assumptions used by the licensees in their safety
assessments. This diagnostic method is based on advanced dynamic reliability techniques on top of using classical
Probabilistic Safety Analysis (PSA) and deterministic tools, and allows for checking at once many aspects of
the safety assessments, making effective use of regulatory resources. Apart from a theoretical approach that is at
the basis of the method, application of ISA requires a set of computational tools. Steps done in the development
of ISA started by development of a suitable software package called SCAIS that comprehensively implies an
intensive use of code coupling techniques to join typical TH analysis, severe accident and probability calculation
codes. The final goal is to dynamically generate the event tree that stems from an initiating event, improving the
conventional PSA static approach.

NEED OF RISK-BASED DIAGNOSTIC

TOOLS

Most often, in defending their safety cases within

the licensing process, industry safety analysis have
to rely on computational tools including simulation
of transients and accidents and probabilistic safety
assessments. Such an assessment capability, even if
reduced to its analytical aspects, is a huge effort
requiring considerable resources.
The increasing trend towards Risk Informed Regulation (RIR) and the recent interest in methods that
are independent on the diversity of existing nuclear
technologies motivate an even greater demand for
computerized safety case analysis. It has been further
fostered by:
new nuclear power plant designs;
the large time span and evolution of the old generic
safety analysis, that requires confirmation of its
present applicability; and
the need to extend the life of the existing plants with
associated challenges to the potential reduction in
their safety margins.

Important examples are for instance the analysis justifying the PSA success criteria and operating
technical specifications, which are often based on
potentially outdated base calculations made in older
times in a different context and with other spectrum of
applications in mind.
This complex situation generates a parallel need
in regulatory bodies that makes it mandatory to
increase their technical expertise and capabilities in
this area. Technical Support Organization (TSO) have
become an essential element of the regulatory process1 , providing a substantial portion of its technical
and scientific basis via computerized safety analysis supported on available knowledge and analytical
methods/tools.
TSO tasks can not have the same scope as their
industry counterparts, nor is it reasonable to expect

1 Examples

are GRS, IRSN, PSI, Studvik, of TSOs

supporting the regulatory bodies of Germany, France,
Switzerland, Sweden, respectively.

121

the same level of resources. Instead, in providing its

technical expertise, they shall:
review and approve methods and results of
licensees, and
perform their own analysis/calculations to verify the
quality, consistency, and conclusions of day to day
industry assessments.
The last is a difficult, different and very special regulatory task, requiring specific TSO diagnostic tools
to independently check the validity and consistency of
the many assumptions used and conclusions obtained
by the licensees in their safety assessments.
The approach and the tools shall include a sound
combination of deterministic and probabilistic single checks, pieces however of an ISA methodology,
that together constitute a comprehensive sample verifying all relevant decision making risk factors and
ensuring that the decision ingredients are properly and
consistently weighted.

ISSUES OF PARTICULAR RELEVANCE

In recent years efforts are being devoted to the clarification of the relative roles of deterministic and
probabilistic types of analysis with a view towards
their harmonization, in order to take benefit of their
strengths and to get rid of identified shortcomings,
normally related with inter-phase aspects, like the
interaction between the evolution of process variables
and its influence in probabilities. Different organizations, (Hofer 2002) and (Izquierdo 2003a), have
undertaken some initiatives in different contexts with
claims such as the need for an integration of probabilistic safety analysis in the safety assessment, up to
the approach of a risk-informed decision making process as well as for proposals of verification methods
for application that are in compliance with the state of
the art in science and technology.
These initiatives should progressively evolve into
a sound and efficient interpretation of the regulations
that may be confirmed via computerized analysis. It
is not so much a question of new regulations from
the risk assessment viewpoint, but to ensure compliance with existing ones in the new context by
verifying the consistency of individual plant assessment results through a comprehensive set of checks.
Its development can then be considered as a key and
novel topic for research within nuclear regulatory
agencies/TSO.
More precisely, issues that require an integrated
approach arise when considering:
the process by which the insights from these complementary safety analysis are combined, and

their relationships when addressing high level

requirements such as defense in depth and safety
margins.
Analysis of PSA success criteria and operating
technical specifications and its mutual consistency are
important chapters of the optimization of the protection system design encompassing, for instance,
problems like:
To ensure that the protection system is able to
cope with all accident scenarios and not only with
a predetermined set. This umbrella character is
hard to prove, particularly within an atmosphere of
reduced safety margins, (SMAP Task Group 2007).
It requires careful regulatory attention to the historic
evolution of the deterministic assessments and it is
a source of potential conflicts when risk techniques
are combined.
To ensure the adequacy of success criteria,
that become critical and sensitive, (Siu, N. &
Hilsmeier, T. 2006). Many studies demonstrating
the umbrella condition are old and perhaps unsuitable under these more restrictive circumstances.
Extension to operator actions of the automatic
protection design is one such source of potential
inconsistencies with complex aspects like available
times for operator action. Emergency procedures
verification is also worth mentioning, because they
imply longer than automatic design accident time
scales to consider and important uncertainties in the
timing of interventions, both potentially altering the
umbrella conditions of the deterministic design.
Another important extension of the scope of the
analysis is the need to consider degraded core situations to ensure acceptable residual risks, (IAEA
2007). Again consistency issues appear requiring
regulatory checks.
These consistency checks call for an appropriate
extension of the probabilistic safety metrics currently
used. Different exceedance frequency limits for different barrier safety limit indicators have been extensively
discussed and may be used as a sound risk domain
interpretation of the existing regulations. For instance,
frequency limits for partial core damage or few core
fuel failures may correctly interpret the present deterministic rules for safety margins in a way consistent
with a probabilistic approach.

DEVELOPEMENTS FOR AN INTEGRATED

PSA TO INDEPENDENT VERIFICATION OF
SAFETY CASES

The CSN branch of Modelling and Simulation (MOSI)

has developed its own ISA methodology for the above
referred purposes. This diagnostic method has been

122

designed as a regulatory tool, able to compute the

frequency of PSA sequences and the exceedance frequency of specified levels of damage, in order to
check in an independent way the results and assumptions of the industry PSAs, including their RIR
extensions/applications. This approach harmonizes
the probabilistic and deterministic safety assessment
aspects via a consistent and unified computer framework.
Apart from a theoretical approach that is at the basis
of the method, application of ISA requires a set of computational tools. A suitable software package called
SCAIS (Figure 1) at each time step couples, (Izquierdo
2003a):
simulation of nuclear accident sequences, resulting
from exploring potential equipment degradations
following an initiating event (i.e. simulation of thermal hydraulics, severe accident phenomenology and
fission product transport);
simulation of operating procedures and severe accident management guidelines;
automatic delineation (with no a-priori assumptions) of event and phenomena trees;
probabilistic quantification of fault trees and
sequences; and
integration and statistic treatment of risk metrics.
Since the final goal is to generate dynamically the
event tree that stems from an initiating event, improving the conventional PSA static approach, this simulation technique is called tree simulation. The activation
of branching conditions is referred to as stimuli activation. Stimulus Driven Theory of Probabilistic Dynamics (SDTPD), (Hofer 2002) and (Izquierdo 2003a), is
the underlying mathematical risk theory (basic concepts, principles and theoretical framework) on which
it is inspired and supported. The massive use of coupled codes has led to the definition and development
of a Standard Software Platform that allows a given
code to be incorporated quickly into the overall system

Figure 1.

SCAIS overview schema.

and to overcome difficulties derived from particular

models and computational methods.
3.1 Software package description
A consolidation and modernization program,
(Izquierdo 2003b), is currently being executed to
enhance capabilities and functionality of the software package that would also facilitate an easier
maintenance and extensibility.
The current SCAIS development includes as main
elements the Event Scheduler, the Probability Calculator, the Simulation Driver (BABIECA) and the Plant
Models (Figure 2):
1. Event scheduler (DENDROS), that drives the
dynamic simulation of the different incidental
sequences. Its design guarantees modularity of the
overall system and the parallelization of the event
tree generation, (Muoz 1999).
It is designed as a separate process that controls
the branch opening and coordinates the different processes that play a role in the generation
of the Dynamic Event Tree (DET). The idea is
to use the full capabilities of a distributed computational environment, allowing the maximum
number of processors to be active. To this end, it
manages the communications among the different
processes that intervene in the event tree development, namely, the distributed plant simulator, the
probability calculator, and an output processor.
The scheduler arranges for the opening of the
branch whenever certain conditions are met, and
stops the simulation of any particular branch that
has reached an absorbing state. The scheduler must
know the probability of each branch in order to
decide which branch is suitable for further development. Each new branch is started in a separate
process, spawning a new transient simulator process and initializing it to the transient conditions
that were in effect when the branching occurred.

Figure 2.

123

Global system architecture.

This saves a substantial amount of computation

time, since common parts of the sequences are not
recomputed. The applications of a tree structured
computation extend beyond the scope of the DETs.
In fact, the branch opening and cutoff can obey any
set of criteria not necessarily given by a probability
calculation as, for instance, sensitivity studies or
automatic initialization for Accident Management
Strategy analysis. Tasks distribution among the different processors is managed by the Parallel Virtual
Machine (PVM) interface.
2. Probability calculator module that incrementally
performs the boolean product of the fault trees corresponding to each system that intervene in the
sequence, additionally computing its probability.
The fault trees that will be used for the probability
calculations are those of PSA studies. This imposes
a strong computational demand that is optimized
by preprocessing the header fault trees as much as
possible. The current approach is trying to use fast
on-line probability computation based on the representation of fault trees using the Binary Decision
Diagram (BDD) formalism, fed from the industry
models.
3. BABIECA, the consolidated simulation driver,
solves step by step topologies of block diagrams. A
standardized linkage method has also been defined
and implemented to incorporate as block-modules
other single-application oriented codes, using parallel techniques. BABIECA driver allows also to
change the simulation codes at any time to fit the
model to the instantaneous conditions, depending on the need of the given simulation. Two
coupling approaches, namely by boundary and initial conditions, have been implemented, (Herrero
2003):
Initial Conditions Coupling. This type of coupling is used when a certain code is reaching
validity and applicability limits, and new models
and codes are necessary to further analysis. The
typical example of this type of coupling is the
transition from conventional Thermal hydraulic
(TH) codes to severe accident codes. BABIECA
accomplish that allowing parts of the model
(modules) remaining active or inactive depending on a specific variable (called simulation
mode).
Boundary Conditions Coupling. Some of the
output variables obtained at the time advancement in one of the codes are sent to the computation of the model boundary conditions in the
other code. This type of coupling is used to build
a wide scope code starting from several codes
with a more limited scope.
The synchronization points are external to the
solution advancement; this subject allows the

connection scheme to be applied to any set of codes

regardless of the particular codes. This feature
increases the power of calculation since large codes
can be split and executed in different machines.
This feature allows BABIECA to be coupled with
itself, enhancing an easier modeling and simulation
management of large model topologies.
4. Plant Models. Sequences obtained in ISA involve
very often a wide range of phenomena, not covered by a single simulation code. On the other
hand, Emergency Operation Procedures (EOPs) are
a very complex set of prescriptions, essential to the
sequence development and branching and hard to
represent in detailed codes. A plant model suitable
for ISA purposes does not comprise then a single
input deck for just one code, but several inputs
for several codes, each one being responsible for
the simulation of certain phenomena. The coupling
entails different although interrelated interfaces
(see Figure 2).
Codes as MELCOR, MAAP, RELAP, TRACE
can adapted to perform tree simulations under
control of the scheduler. At present Modular Accident Analysis Program (MAAP4) is coupled to
BABIECA to build up a distributed plant simulation. MAAP4 performs the calculation of the
plant model when the transient reaches the severe
accident conditions, being then initialized with the
appropriate transient conditions. Some parts of the
simulation (specially the operator actions, but also
control systems) may still be performed by the original code and the appropriate signals be transferred
as boundary conditions.
Consolidated versions of both driver simulator
BABIECA and DENDROS scheduler have been
designed with an object oriented architecture and
implemented in C++ language. They have been developed using OpenSource standards (Linux, XercesC,
libpq ++). The code system has been equipped
with a Structured Query Language (SQL) Relational
Database (PostgreSQL), used as repository for model
input and output. The input file has been also modernized using XML standards since it is easy to read and
understand, tags can be created as they are needed, and
facilitates treating the document as an object, using
object oriented paradigms.
3.2 Features of SCAIS coupled codes and future
extensions
In order to connect BABIECA with external codes
has been developed a wrapper communication code
based on PVM. BABIECA solves a step by step coupled blocks topology. The solution follows an Standard
Computational Scheme in a well defined synchronization points that allows an easy code coupling with other

124

codes following the Standard. Two blocks are in charge

of BABIECA communication with coupled codes:
SndCode supplies the boundary conditions; and
RcvCode. Receives messages of every spawned
code in each time step.
For any desired code to be coupled, an specific
wrapper, consistent with the Standard Computational
Scheme, must be developed and implemented. It needs
the message passing points defined by BABIECA in
each synchronization point, allowing the communication between both codes. Three wrappers have been
developed and are being tested at present:
BABIECAWrapper allows BABIECA to be connected, and used to split big topologies into smaller
ones in order to save computation time parallelizing
the processes.
ProbabilityWrapper allows BABIECA and DENDROS connections with the probability calculator
module.
MAAP4Wrapper allows BABIECA-MAAP4
connections.
This wrapper, implemented in FORTRAN77 as
MAAP4, gives us the advantage to analyze MAAP4
sequences with the data processing tools developed to BABIECA. The interface module has the
following functions:
1. To initialize the MAAP4 simulation using input
data values provided by the BABIECA model.
2. To initialize MAAP4 from a restart image data.
3. To accomplish the exchange of dynamic variables between MAAP4 and other codes.
4. To drive MAAP4 time-advancement calculation.
5. To save results and restart image data in
BABIECA Data Base.
Recent extensions
oriented to:

under

development

are

Develop the Paths and Risk assessment modules of

SCAIS system.
They implement a dynamic reliability method
based on the Theory of Stimulated Dynamics TSD
which is a variant of the more general SDTPD in its
path and sequence version, (Izquierdo 2006) and
(Queral 2006).
Development and application of advanced probabilistic quantification techniques based on Binary
Decision Diagrams (BDDs) that would release some
traditional drawbacks (e.g. truncation and rare event
approximation) of the currently used quantification
approaches.
New modeling algorithms to simulate standard
PSA, correcting for dynamic effects.
New wrapper module allowing BABIECA-TRACE
connections.

New developments in sequence dynamics. A generalization of the transfer function concepts for
sequences of events, with potential for PSA application as generalized dynamic release factors is under
investigation.
New developments about classical PSA aspects.
They include rigorous definitions for concepts like
available time for operations or plant damage states.

TEST CASE: MBLOCA SEQUENCES

A full scale test application of this integrated software

package to a Medium Break Loss of Coolant Accident
(MBLOCA) initiating event of a Spanish PWR plant
is currently under development.
The specific objective of this analysis is to demonstrate the methodology and to check the tool, focusing
on an independent verification of the event tree delineation and assessment of realistic EOPs for LOCA
plant recovery. SCAIS provides a huge amount of
results for the analyst, when unfolding the Dynamic
Event Tree (DET), even though methodology reduces
to manageable the number of sequences. The following sections outline some conclusions of the assessment, including sequence evolution and operator
actions analysis.
4.1

Sequence analysis: MBLOCA sequence with

accumulators failure

In a first step MBLOCA sequences are simulated with

MAAP code. Results show that for breaks up to 6
(inches) no action to achieve the primary depressurization is necessary and LPSI starts automatically about
3000 seconds. However, for 2 break and lower, primary system is stabilized at a higher pressure than low
pressure safety injection pump head, being necessary
the operator action to achieve low-pressure controlled
conditions, Figure 3.
Main actions of set of operator actions, corresponding to EOP E-1 (loss of reactor or secondary coolant)
and EOP ES-1.2 (cooling and decrease of pressure
after a LOCA), are the following:
1. Check the need of reactor coolant pump trip (E-1,
step 1).
2. Check the steam generators (SG) levels (E-1,
step 3).
3. Check pressure decreasing and cooling in the Reactor Coolant System (RCS) (E-1, step 11). If RCS
pressure is higher than 15 kg/cm2 a transition to
EOP ES-1.2 is required.
4. Check SG levels (ES-1.2, step 5).
5. Start the RCS cooling until cool shutdown conditions (ES-1.2, step 6).
6. Check if sub-cooling margin greater than 0 C.

125

Figure 3. MAAP4 simulation. Comparison with/without

operator actions. Primary Pressure System (MBLOCA 2 ).

Figure 4. MAAP4 simulation. Primary System Water

Temperature. Comparison with/without operator actions
(MBLOCA 2 ).

7. Check if Safety Injection System (SIS) is active.

8. Disconnect pressurizer (PZR) heaters.
9. Decrease RCS pressure in order to recover PZR
level.
These operator actions has been implemented in
MAAP. Of particular relevance is primary cool-down
process at 55 C/h by SGRV control, Figure 4.
As a second phase of the study, manual actions have
been modelled by using the available BABIECA general modules, transferring the corresponding actions
to MAAP4, by using SndCode and RcvCode blocks,
Figure 5. Simulations with MAAP4 stand-alone and
BABIECA coupled with MAAP4 give the same
results, Figure 6.
In a near future EOPs instructions will be computerized with SIMPROC that will be coupled with

Figure 5. Simplified topology scheme of BABIECAMAAP simulation.

Figure 6. Primary System Water Temperature. MAAP stand

alone and BABIECA-MAAP comparison (MBLOCA 2 ).

BABIECA and thermal hydraulic codes like MAAP

or TRACE, (Espern 2008).

4.2 Dynamic event tree simulation for MBLOCA

As a third step of the study, DET of a 6 MBLOCA is
unfolded by incorporating DENDROS to the previous
model. Each time a safety system set-point is reached,
DENDROS unfolds the tree by simulating two different branches (namely system success and failure
branch), Figure 7. In the study, the following headers
have been considered:

126

HPSI High Pressure Safety Injection.

ACCUM Safety Injection Accumulator.
LPSI Low Presure Safety Injection.
RECIR Safety Injection Recirculation Mode.

consistency of probabilistic and deterministic aspects.

It is our opinion this need could be framed on an
international cooperative effort among national TSOs.
We have also shown that SCAIS is a powerful tool
that carries out the CSN Integrated Safety Analysis. Its
versatility and extensibility makes it an appropriate set
of tools that may be used either together or separately
to perform different regulatory checks.

ACKNOWLEDGMENTS

6 .

Figure 7. MAAP4 Event Tree of MBLOCA

Automatically drawn by DENDROS simulation (branch opening time
in seconds).

This work is partially founded by Spanish Ministry of

Education & Science in the STIM project (ENE2006
12931/CON).

NOMENCLATURE

Figure 8. Safety Injection Flows for different DET

sequences (MBLOCA 6 ).

The results for some sequences show the impact

of different system failures and/or operator actions,
Figure 8.

CONCLUSIONS

We have presented an overview of CSN methods and

simulation packages to perform independent safety
assessments to judge nuclear industry PSA related
safety cases. It has been shown that this approach is
adequate for an objective regulatory technical support
to CSN decision-making.
This paper also argues on the need for development
of diagnosis tools and methods for TSO, to perform
their own computerized analysis to verify quality, consistency, and conclusions of day to day individual
industry safety assessments, in such a way that asserts

ISA Integrated Safety Analysis

SCAIS Simulation Codes System for Integrated Safety
Assessment
PSA Probabilistic Safety Analysis
CSN Spanish Nuclear Safety Council
RIR Risk Informed Regulation
EOP Emergency Operation Procedure
SAMG Severe Accident Management Guide
DET Dynamic Event Tree
PVM Parallel Virtual Machine
BDD Binary Decision Diagram
TSO Technical Support Organization
MAAP Modular Accident Analysis Program
SDTPD Stimulus Driven Theory of Probabilistic
Dynamics
RCS Reactor Coolant System
SIS Safety Injection System
PSRV Pressure Safety Relief Valve
TWPS Temperature of Water in the Primary System
XML Extensible Markup Language
MBLOCA Medium Break Loss of Coolant Accident
HRA Human Reliability Analysis
TH Thermal Hydraulic
TSD Theory of Stimulated Dynamics
PWR Pressurized Water Reactor
SQL Structured Query Language

REFERENCES
Espern, J. & Expsito, A. (2008). SIMPROC: Procedures
simulator for operator actions in NPPs. Proceedings of
ESREL 08.
Herrero, R. (2003). Standardization of code coupling for integrated safety assessment purposes. Technical meeting on
progress in development and use of coupled codes for
accident analysis. IAEA. Viena, 2628 November 2003.

127

Hofer, E. (2002). Dynamic Event Trees for Probabilistic

Safety Analysis. EUROSAFE Forum, Berlin.
IAEA (2007, September). Proposal for a Technology-Neutral
Safety Approach for New Reactor Designs. Technical
Report TECDOC 1570, International Atomic Energy
Agency, http://wwwpub.iaea.org/mtcd/publications/pdf/
te_1570 web.pdf.
Izquierdo, J.M. & Caamn, I. (2006). Status report
on dynamic reliability: SDTPD path and sequence
TSD developments. Application to the WP5.3 benchmark Level 2 PSA exercise. DSR/SAGR/FT 2004.074,
SARNET PSA2 D73 [rev1].
Izquierdo, J. (2003a). An integrated PSA Approach
to Independent Regulatory Evaluations of Nuclear
Safety Assessments of Spanish Nuclear Power Stations.
EUROSAFE Forum, Paris.
Izquierdo, J. (2003b). Consolidation plan for the Integrated
Sequence Assessment (ISA) CSN Code system SCAIS.
CSN internal report.

Muoz, R. (1999). DENDROS: A second generation scheduler for dynamic event trees. M & C 99 Conference,
Madrid.
Queral, C. (2006). Incorporation of stimulusdriven theory
of probabilistics dynamics into ISA (STIM). Joint project
of UPM, CSN and ULB founded by Spanish Ministry of
Education & Science (ENE2006 12931/CON), Madrid.
Siu, N. & Hilsmeier, T. (2006). Planning for future probabilistic risk assessment research and development. In G.S. Zio
(Ed.), Safety and Reliability for Managing Risk, Number
ISBN 0-415-41620-5. Taylor & Francis Group, London.
SMAP Task Group (2007). Safety Margins Action Plan. Final
Report. Technical Report NEA/CSNI/R(2007)9, Nuclear
Energy Agency. Committee on the Safety of Nuclear
Installations, http://www.nea.fr/html/nsd/docs/2007/csnir
2007-9.pdf.

128

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Using GIS and multivariate analyses to visualize risk levels and spatial
patterns of severe accidents in the energy sector
P. Burgherr
Paul Scherrer Institut (PSI), Laboratory for Energy Systems Analysis, Villigen PSI, Switzerland

ABSTRACT: Accident risks of different energy chains are analyzed by comparative risk assessment, based on
the comprehensive database ENSAD established by the Paul Scherrer Institut. Geographic Information Systems
(GIS) and multivariate statistical analyses are then used to investigate the spatial variability of selected risk
indicators, to visualize the impacts of severe accidents, and to assign them to specific geographical areas.
This paper demonstrates by selected case studies how geo-referenced accident data can be coupled with other
socio-economic, ecological and geophysical contextual parameters, leading to interesting new insights. Such an
approach can facilitate the interpretation of results and complex interrelationships, enabling policy makers to
gain a quick overview of the essential scientific findings by means of summarized information.

INTRODUCTION

Categorizing countries by their risk levels due to natural hazards has become a standard approach to assess,
prioritize and mitigate the adverse effects of natural
disasters. Recent examples of coordinated international efforts include studies like the World Disasters
Report (IFRC 2007), Natural Disaster Hotspots
(Dilley et al. 2005), and Reducing Disaster Risk
(UNDP 2004) as well as studies by the worlds leading
reinsurance companies (Berz 2005; Munich Re 2003;
Swiss Re 2003). Applications to the impacts of manmade (anthropogenic) activities range from ecological
risk assessment of metals and organic pollutants (e.g.
Critto et al. 2005; Zhou et al. 2007) to risk assessment of contaminated industrial sites (Carlon et al.
2001), conflicts over international water resources
(Yoffe et al. 2003), climate risks (Anemller et al.
2006), and environmental sustainability assessment
(Bastianoni et al. 2008).
Accidents in the energy sector rank second (after
transportation) of all man-made accidents (Hirschberg
et al. 1998). Since the 1990s, data on energyrelated accidents have been systematically collected,
harmonized and merged within the integrative Energyrelated Severe Accident Database (ENSAD) developed and maintained at the Paul Scherrer Institut
(Burgherr et al. 2004; Hirschberg et al. 1998).
Results of comparative risk assessment for the different energy chains are commonly expressed in a
quantitative manner such as aggregated indicators (e.g.
damage rates), cumulative risk curves in a frequencyconsequence (F-N) diagram, or external cost estimates

to provide an economic valuation of severe accidents (Burgherr & Hirschberg 2008; Burgherr &
Hirschberg, in press; Hirschberg et al. 2004a).
Recently, ENSAD was extended to enable
geo-referencing of individual accident records at
different spatial scales including regional classification schemes (e.g. IPCC world regions; subregions of oceans and seas; international organization
participation), sub-national administrative divisions
(ADMIN1, e.g. state or province; ADMIN2, e.g.
county), statistical regions of Europe (NUTS) by Eurostat, location name (PPL, populated place), latitude
and longitude (in degrees, minutes and seconds), or
global grid systems (e.g. Marsden Squares, Maidenhead Locator System). The coupling of ENSAD
with Geographic Information Systems (GIS) and geostatistical methods allows one to visualize accident
risks and to assign them to specific geographical areas,
as well as to calculate new risk layers, i.e. to produce
illustrative maps and contour plots based on scattered
observed accident data.
In the course of decision and planning processes
policy makers and authorities often rely on summarized information to gain a quick overview of a
thematic issue. However, complex scientific results
are often very difficult to communicate without appropriate visualization. Methods for global, regional or
local mapping using GIS methods are particularly useful because they reveal spatial distribution patterns
and link them to administrative entities, which allow
planning and implementation of preventive measures,
legal intervention or mitigation at the appropriate
administrative level.

129

The aim of this paper is to present and discuss

results of comparative risk assessment by means of
so-called disaster maps. The selected case studies
address different spatial scales and resolution as well as
different analytical techniques to calculate aggregated
risk indicators, based on summary statistics, multivariate statistical analyses to produce a single index, and
spatial interpolation techniques.

2.2 Geo-referencing of accident data

METHODS

ENSAD concentrates on comprehensively covering

severe, energy-related accidents, although other manmade accidents and natural catastrophes are also
reviewed, but somewhat less explicitly. The database
allows the user to make conclusive analyses and calculate technical risks, including frequencies and the
expected extent damages.
The compilation of a severe accident database is
a complex, tedious and extremely time consuming
task. Therefore, it was decided in the beginning
that ENSAD should build upon existing information
sources. Its unique feature is that it integrates data from
a large variety of primary data sources, i.e. information
is merged, harmonized and verified.
The actual process of database building and implementation has been described in detail elsewhere
(Burgherr et al. 2004; Hirschberg et al. 1998), thus
only a brief summary of the essential steps is given
here:

200 or more evacuations

a far-reaching ban on the consumption of food
a release of at least 10000 tonnes of hydrocarbons
the cleanup of a land or water surface of 25 km2 or
more
economic damages of at least 5 million USD (year
2000 exchange rates).

The plotting of accident data on maps and the

analysis of spatial distribution patterns requires specific information about the position of an event.
The geographic coordinates of accident locations
were determined using detailed accident information
stored in ENSAD as well as other resources such
as the World Gazetteer (http://world-gazetteer.com/)
or Google Earth (http://earth.google.com/). Additionally, accidents were assigned to hierarchical organization of political geographic locations, referring
to country (ADMIN0) and the first level below
(ADMIN1; e.g. state or province). All maps were
produced using ArcGIS 9.2 software (ESRI 2006b).
The basic GIS layer data for country boundaries were
based on ESRI world data and maps (ESRI 2006a),
data from the Geographic Information System of
the European Commission (GISCO) (Eurostat 2005),
and data from the U.S. Geological Survey (USGS
2004).

Selection and survey of a large number of information sources.

Collection of raw information that is subsequently
merged, harmonized and verified.
Information for individual accidents is assigned to
standardized data fields and fed into ENSAD.
The final ENSAD data is subjected to several crosschecks to keep data errors at the lowest feasible
level.
Comparative evaluations are then carried out, based
on customized ENSAD-queries.

2.3 Evaluation period

2.1

2.4 Data analysis

Severe accident definition

In the literature there is no unique definition of a

severe accident. Differences concern the actual damage types considered (e.g. fatalities, injured persons,
evacuees or economic costs), use of loose categories
such as people affected, and differences in damage
thresholds to distinguish severe from smaller accidents. Within the framework of PSIs database ENSAD
an accident is considered to be severe if it is characterized by one or more of the following consequences
(Burgherr et al. 2004; Hirschberg et al. 1998):
5 or more fatalities
10 or more injuries

The time period 19702005 was selected for analysis

of accident risks to ensure an adequate representation
of the historical experience. In the case of the Chinese
coal chain, only the years 19941999 were considered
because annual editions of the China Coal Industry Yearbook were only available for these years. It
could be shown that without access to this information
source, data are subject to substantial underreporting
(Burgherr & Hirschberg 2007).

This section describes the methodological details of

the four selected examples that are presented in this
publication. In the remainder of this paper they are
referred to as case studies 14.
Case Study 1: The total fatalities per country of
severe accidents (5 fatalities) in fossil energy chains
(coal, oil, natural gas and Liquefied Petroleum Gas
(LPG)) are plotted on a world map. Additionally, the
top ten countries in terms of number of accidents, and
the ten most deadly accidents are also indicated.
Case Study 2: Geo-referenced data for all fatal
oil chain accidents stored in ENSAD are mapped for
European Union (EU 27), candidate countries, and

130

the European Free Trade Association (EFTA). A multivariate risk score was then calculated for individual
countries to analyze their differences in susceptibility
to accident risks. The proposed risk score consists of
four indicators:

of major state-owned mines in different provinces was

investigated.

The ENSAD database currently contains 21549

accident records, of which 90.2% occurred in the
years 19702005. Within this period, 7621 accidents
resulted in at least five fatalities, of which 31.1% were
man-made, energy-related accidents. Table 1 summarizes the number of accidents and fatalities that
occurred in fossil energy chains of different country groups. Results are separated for countries of the
Organisation for Economic Co-operation and Development (OECD), European Union (EU 27), and states
that are not OECD members (non-OECD), due to
the large differences in technological development,
institutional and regulatory frameworks, and general
safety culture of OECD and EU 27 versus non-OECD.
Calculations were complemented by separate values
for the Chinese coal chain that has a significantly
worse performance than all other non-OECD countries
(Burgherr & Hirschberg 2007).

total number of accidents (accident proneness)

total number of fatalities (accident gravity)
maximum consequences (most deadly accident)
fatality rate (expectation value expressed in GWe yr
(Gigawatt-electric-year) using a generic load factor
of 0.35)

Each variable was scaled between 0 and 1 using the

equation xij = (zij min)/(max min), where zij is
the value of the jth variable for the ith country, and min
and max are the minimum and maximum values of the
jth variable. The resulting matrix was then analyzed
by means of Principal Components Analysis (PCA)
(Ferguson 1998; Jolliffe 2002). Non-centred PCA was
used because the first component (PC1) of such an
analysis is always unipolar (Noy-Meir 1973), and can
thus be used as a multivariate risk score (a lower value
indicates a lower accident risk).
Case Study 3: Tanker oil spills in the European Atlantic, the Mediterranean Sea (including the
entrance to the Suez Canal) and the Black Sea were
analyzed because these maritime areas pertain to
the geographical region addressed in Case Study 2.
Additionally, large parts of the Northeast Atlantic
(including the North Sea and the Baltic Sea), the
Canary current and the Mediterranean Sea belong to
the so-called Large Marine Ecosystems (LME) that
are considered ecologically sensitive areas (Hempel &
Sherman 2003). Distribution patterns of oil spills
(700 tonnes) were analyzed using the kriging technique, which is a geo-statistical method for spatial
interpolation (Krige 1951; Matheron 1963). In a preliminary step, locations were assigned to a Marsden
Square Chart, which divides the world into grids of
10 degrees latitude by 10 degrees longitude, because in
some cases only approximate coordinates were available. Afterwards, the number of spills per Marsden
Square was calculated, and coordinates set to the center of each grid cell. Ordinary point kriging was then
applied to compute a prediction surface for the number
of spills to be expected in a particular region, i.e. to
evaluate regional differences in susceptibility to accidental tanker spills. For methodological details on the
applied kriging procedure see Burgherr (2007) and
Burgherr & Hirschberg (2008).
Case Study 4: Severe (5 fatalities) accidents in
the Chinese coal chain were analyzed at the province
level (ADMIN1). Fatality rates were calculated for
large state-owned and local mines, and for small township and village mines, respectively. The influence of
mechanization levels in mining as well as the number

RESULTS AND DISCUSSION

3.1 Main findings and insights of case studies

Case Study 1: Figure 1 shows the worldwide distribution of total fatalities per country of severe (5
fatalities) accidents in the period 19702005. The top
ten countries in terms of total numbers of accidents
are also indicated in Figure 1, as well as the ten most
deadly accidents.
China was the most accident prone country with
25930 fatalities, of which almost 95% occurred in
1363 accidents attributable to the coal chain. However, only 15 of these resulted in 100 or more fatalities.

131

Table 1. Numbers of accidents (Acc) and fatalities (Fat)

of severe (5 fatalities) accidents in fossil energy chains
are given for OECD, EU 27, and non-OECD countries
(19702005). For the coal chain, non-OECD w/o China (first
line) and China alone (second line) are given separately.
Natural
Coal
OECD

Acc
Fat
EU 27
Acc
Fat
Non-OECD Acc

World total

81
2123
41
942
144
1363
Fat
5360
24456
Acc 1588
Fat 31939

Oil

Gas

LPG

174 103
59
3388 1204 1875
64
33
20
1236 337 559
308
61
61

Total
417
8590
158
3074
574

17990 1366 2610 27326

482 164 120 2354
21378 2570 4485 60372

Figure 1. Individual countries are shaded according to their total numbers of severe accident fatalities in fossil energy
chains for the period 19702005. Pie charts designate the ten countries with most accidents, whereas bars indicate the ten
most deadly accidents. Country boundaries: ESRI Data & Maps (ESRI 2006a).

Figure 2. Locations of all fatal oil chain accidents and country-specific risk scores of EU 27, accession candidate and
EFTA countries for the period 19702005. Country boundaries: EuroGeographics for the administrative boundaries
(Eurostat 2005).

In contrast, the cumulated fatalities of the non-OECD

countries Philippines, Afghanistan, Nigeria, Russia,
Egypt and India were strongly influenced by a few very
large accidents that contributed a substantial share of
the total (see Figure 1 for examples). Among the ten
countries with most fatalities, only three belonged

to OECD, of which the USA and Turkey have been

founder members of 1961, whereas Mexico gained
membership more than three decades later in 1994.
The LPG chain contributed almost 50% to total fatalities in Mexico because of one very large accident
in 1984 that resulted in 498 fatalities. In Turkey the

132

coal chain had the highest share with 531 fatalities

or about 55%, of which 272 fatalities are due to a single accident, i.e. a methane explosion in a coal mine in
north-western Turkey in 1992. Finally, the USA exhibited a distinctly different pattern compared to the other
countries with no extremely large accidents (only three
out of 148 with more than 50 fatalities), and over 75%

of accidents and 70% of fatalities taking place in the

oil and gas chains.
Case Study 2: The map of Figure 2 shows the
locations of all fatal oil chain accidents contained in
ENSAD for EU 27, accession candidate and EFTA
countries in the period 19702005. The calculated
risk score is a measure of the heterogeneous accident

Figure 3. Individual geo-referenced oil spills for the period 19702005 are represented by different-sized circles corresponding to the number of tonnes released. Regional differences in susceptibility to accidents were analyzed by ordinary kriging,
resulting in a prediction map of filled contours. The boundaries of the Large Marine Ecosystems (LME) are also shown.
Country boundaries: EuroGeographics for the administrative boundaries (Eurostat 2005).

133

Figure 4. For individual provinces in China, average fatalities per Mt produced coal are given for severe (5 fatalities)
accidents in large and small mines for the period 19941999. Provinces were assigned to three distinct levels of mechanization
as indicated by their shading. Locations of major state-owned coal mines are also indicated on the map. Administrative
boundaries and coal mine locations: U.S. Geological Survey (USGS 2004).

risk patterns among countries. Four countries had risk

scores greater than 0.40, namely the United Kingdom,
Italy, Norway and Turkey. This is largely due to
the substantially higher values for total fatalities and
maximum consequences, and to a lesser extent also
for total accidents compared to the respective average values for all countries. In contrast, fatality rates
(i.e. fatalities per GWe yr) of these four countries
were clearly below the overall average. These findings support the notion that besides fatality rates other
performance measures should be evaluated, particularly in the context of multi-criteria decision analysis (MCDA) because preference profiles may differ
among stakeholders (e.g. Hirschberg et al. 2004b).
Case Study 3: Geographic locations were available for 128 out of a total of 133 tanker accidents in
the years 19702005 that occurred in the European
Atlantic, the Mediterranean Sea and the Black Sea,
and each resulted in a spill of at least 700t. The severity
of the spills is divided into different spill size classes,
based on the amount of oil spilled.

Figure 3 also provides results of ordinary kriging

that are based on a spherical model for the fitted semivariogram (SV) function. Cross-validation indicated
that the model and subsequently generated prediction map provided a reasonably accurate prediction
(details on the methodology are given in Burgherr and
Hirschberg (2008) and Matheron (1963)). The prediction map based on spatial interpolation by ordinary
kriging provides useful information for identification
and assessment of regional differences in susceptibility to oil spills from tankers. Such an interpolated
surface layer is also superior to a simple point representation because it enables estimates for areas with
few or no sample points. The map clearly identifies
several maritime regions that are particularly prone
to accidental oil spills, namely the English Channel,
the North Sea, the coast of Galicia, and the Eastern
Mediterranean Sea. Finally, those spills were identified that occurred in ecologically sensitive areas
because they are located within the boundaries of the
Large Marine Ecosystems (LME) of the world. In

134

total, 82.8% of all mapped spills were located within

LME boundaries. However, tankers can generally
not avoid passing LMEs because shipping routes are
internationally regulated, and because LMEs also
include large fractions of coastal areas that ships
have to pass when approaching their port of destination. Therefore, recent efforts to reduce this share
have focused on technical measures, for example
improved navigation and the replacement of single
hull tankers by modern double hull designs. The substantial decrease in total spill volume (also in these
ecologically sensitive areas) since the 1990s reflects
the achieved improvements of this ongoing process
(Burgherr 2007).
Case Study 4: The Chinese coal chain is a worrisome case with more than 6000 fatalities every year, a
fatality rate about ten times higher than in other nonOECD countries, and even about 40 times higher than
in OECD countries (Burgherr & Hirschberg 2007).
Average values of fatalities per million tonne (Mt) of
coal output for the years 19941999 showed a distinct variation among provinces (Figure 4). The figure
also demonstrates how fatality rates are influenced by
the level of mechanization in mining. When results
of individual provinces were assigned to three groups
according to differences in mechanized levels, average
fatality rates were inversely related to mechanization
indicating that higher mechanized levels of mining do
contribute to overall mine safety. Furthermore, average fatality rates for large state-owned and local mines
were significantly lower than for small township and
village mines with very low safety standards. Finally,
major state-owned mines exhibit predominantly high
and medium levels of mechanization.

CONCLUSIONS

The coupling of ENSAD with a GIS-based approach

has proven to be useful in determining spatial distribution patterns of accident risks for selected energy
chains and country groups at global and regional
scales. Multivariate statistical methods (PCA) and
spatial interpolation techniques (kriging) have been
successfully used to extract additional value from
accident data and to produce illustrative maps of
risk scores and contour plots. The potential benefits
of such an approach are threefold: First, it allows
the identification of spatial patterns including accident hotspots and extreme events. Second, powerful,
flexible and understandable visualizations facilitate
the interpretation of complex risk assessment results.
Third, the results summarized by means of GIS provide a simple and comprehensive set of instruments
to support policy makers in the decision and planning
process.

ACKNOWLEDGEMENTS
The author thanks Drs. Stefan Hirschberg and Warren
Schenler for their valuable comments on an earlier
version of this manuscript. This study was partially
performed within the Integrated Project NEEDS (New
Energy Externalities Development for Sustainability, Contract No. 502687) of the 6th Framework
Programme of European Community.

REFERENCES
Anemller, S., Monreal, S. & Bals, C. 2006. Global Climate Risk Index 2006. Weather-related loss events and
their impacts on countries in 2004 and in a long-term
comparison, Bonn/Berlin, Germany: Germanwatch e.V.
Bastianoni, S., Pulselli, F.M., Focardi, S., Tiezzi, E.B.P. &
Gramatica, P. 2008. Correlations and complementarities
in data and methods through Principal Components Analysis (PCA) applied to the results of the SPIn-Eco Project.
Journal of Environmental Management 86: 419426.
Berz, G. 2005. Windstorm and storm surges in Europe: loss
trends and possible counter-actions from the viewpoint
of an international reinsurer. Philosophical Transactions
of the Royal Society AMathematical Physical and
Engineering Sciences 363(1831): 14311440.
Burgherr, P. 2007. In-depth analysis of accidental oil spills
from tankers in the context of global spill trends from all
sources. Journal of Hazardous Materials 140: 245256.
Burgherr, P. & Hirschberg, S. 2007. Assessment of severe
accident risks in the Chinese coal chain. International
Journal of Risk Assessment and Management 7(8):
11571175.
Burgherr, P. & Hirschberg, S. 2008. Severe accident risks
in fossil energy chains: a comparative analysis. Energy
33(4): 538553.
Burgherr, P. & Hirschberg, S. in press. A comparative analysis of accident risks in fossil, hydro and nuclear energy
chains. Human and Ecological Risk Assessment.
Burgherr, P., Hirschberg, S., Hunt, A. & Ortiz, R.A. 2004.
Severe accidents in the energy sector. Final Report to the
European Commission of the EU 5th Framework Programme New Elements for the Assessment of External
Costs from Energy Technologies (NewExt), Brussels,
Belgium: DG Research, Technological Development and
Demonstration (RTD).
Carlon, C., Critto, A., Marcomini, A. & Nathanail, P. 2001.
Risk based characterisation of contaminated industrial
site using multivariate and geostatistical tools. Environmental Pollution 111: 417427.
Critto, A., Carlon, C. & Marcomini, A. 2005. Screening
ecological risk assessment for the benthic community in
the Venice Lagoon (Italy). Environment International 31:
10941100.
Dilley, M., Chen, R.S., Deichmann, U., Lerner-Lam, A.L.,
Arnold, M., Agwe, J., Buys, P., Kjekstad, O., Lyon, B. &
Yetman, G. 2005. Natural disaster hotspots. A global
risk analysis. Disaster Risk Management Series No.
5, Washington D.C., USA: The World Bank, Hazard
Management Unit.

135

ESRI 2006a. ArcGIS 9: ESRI Data & Maps, Redlands

(CA), USA: Environmental Systems Research Institute
Inc. (ESRI).
ESRI 2006b. ArcGIS 9: using ArcGIS Desktop, Redlands
(CA), USA: Environmental Systems Research Institute
Inc. (ESRI).
Eurostat 2005. Guidelines for geographic data intended for
the GISCO Reference Database, Witney, UK: Lovell
Johns Ltd.
Ferguson, C.C. 1998. Techniques for handling uncertainty and variability in risk assessment models, Berlin,
Germany: Umweltbundesamt.
Hempel, G. & Sherman, K. (Eds.) 2003 Large Marine
Ecosystems of the world: trends in exploitation, protection, and research, Amsterdam, Elsevier B.V.
Hirschberg, S., Burgherr, P., Spiekerman, G. & Dones, R.
2004a. Severe accidents in the energy sector: comparative
perspective. Journal of Hazardous Materials 111(13):
5765.
Hirschberg, S., Dones, R., Heck, T., Burgherr, P.,
Schenler, W. & Bauer, C. 2004b. Sustainability of electricity supply technologies under German conditions: a
comparative evaluation. PSI-Report No. 0415, Villigen
PSI, Switzerland: Paul Scherrer Institut.
Hirschberg, S., Spiekerman, G. & Dones, R. 1998. Severe
accidents in the energy sectorfirst edition. PSI Report
No. 98-16, Villigen PSI, Switzerland: Paul Scherrer
Institut.
IFRC (International Federation of the Red Cross and Red
Crescent Societies) 2007. World Disasters Report 2007,
Bloomfield, USA: Kumarian Press Inc.
Jolliffe, I.T. 2002. Principal Component Analysis. Springer
Series in Statistics, 2nd ed., New York, USA: Springer.

Krige, D.G. 1951. A statistical approach to some basic mine

valuation problems on the Witwatersrand. Journal of the
Chemical, Metallurgical and Mining Society of South
Africa 52: 119139.
Matheron, G. 1963. Trait de Gostatistique Aplique. Tome
II: Le krigeage, Paris, France: Ed. Bureau du Recherches
Geologiques et Minieres.
Munich Re 2003. NatCatSERVICEA guide to the Munich
Re database for natural catastrophes, Munich, Germany:
Munich Re.
Noy-Meir, I. 1973. Data transformations in ecological ordination. Some advantages of non-centering. Journal of
Ecology 61: 329341.
Swiss Re 2003. Natural catastrophes and reinsurance, Zurich,
Switzerland: Swiss Reinsurance Company.
UNDP Bureau for Crisis Prevention and Recovery 2004
Reducing disaster risk. A challenge for development.
New York, UNDP, Bureau for Crisis Prevention and
Recovery.
USGS 2004. Open-File Report 01318: Coal geology, land
use, and human health in the Peoples Republic of China
(http://pubs.usgs.gov/of/2001/ofr-01-318/), Reston (VA),
USA: U.S. Geological Survey.
Yoffe, S., Wolf, A.T. & Giordano, M. 2003. Conflict and
cooperation over international freshwater resources: Indicators of basins at risk. Journal of the American Water
Resources Association 39(5): 11091126.
Zhou, F., Huaicheng, G. & Hao, Z. 2007. Spatial distribution
of heavy metals in Hong Kongs marine sediments and
their human impacts: a GIS-based chemometric approach.
Marine Pollution Bulletin 54: 13721384.

136

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Weak signals of potential accidents at Seveso establishments

Paolo A. Bragatto, Patrizia Agnello, Silvia Ansaldi & Paolo Pittiglio
ISPESL, Research Centre via Fontana Candida, Monteporzio Catone (Rome), Italy

ABSTRACT: At industrial facilities where the legislation on major accident hazard is enforced, near misses,
failures and deviations, even though without consequences, should be recorded and analyzed, for an early
identification of factors that could precede accidents. In order to provide duty-holders with tools for capturing
and managing these weak signals coming from operations, a software prototype, named NOCE, has been
developed. Client-server architecture has been adopted, in order to have a palmtop computer connected to an
experience data base at the central server. The operators shall record by the palmtop any non-conformance, in
order to have a plant operational experience database. Non-conformances are matched with the safety system
for finding breaches in the safety system and eventually remove them. A digital representation of the plant and
its safety systems shall be developed, step by step, by exploiting data and documents, which are required by the
legislation for the control of major accident hazard. No extra job is required to the duty holder. The plant safety
digital representation will be used for analysing non-conformances. The safety documents, including safety
management system, safety procedures, safety report and the inspection program, may be reviewed according
to the plant operational experience.

INTRODUCTION

At industrial facilities where major accident hazard is

posed, for one accident with fatalities or severe injuries
there are tens of near-misses with consequences just
on equipment and hundreds of failures that cause just
small loss of production, as well as procedures not well
understood or applied, with minor consequences. Furthermore in the life of an establishment, thousands non
conformances are usually reported both for equipment
and for procedures. The benefit of having a good program for analysing and managing near-misses, failures
and non conformances have been widely demonstrated in many papers, including Hursta et al. (1996),
Ashford (1997), Jones et al. (1999), Phimister et al.
(2003), Basso et al. (2004), Uth et al. (2004)
Sonnemans et al. (2006). There are definitely much
more failures, trivial incidents and near-misses, than
severe accidents. Furthermore failures and deviations
are very easy to identify, to understand, to control.
They are quite small in scale, relatively simple to analyze and to resolve. In near-misses, actual accident
causes are not hidden to avoid a potential prosecution,
unlike the accidents with fatalities, as demonstrated
by Bragatto et al. (2007) and Agnello et al. (2007).
At major accident hazard installations, near-misses,
failures and deviations are able to give weak signals, which have the potential of warning the operator
before the accidents happen. Therefore, by addressing

these precursors effectively, major accidents may be

prevented and consequent fatalities, severe injuries,
asset damages and production losses may be avoided.
An effective system for reporting any non conformance, even without consequence, is essential for an
efficient safety management. Furthermore, an active
monitoring of equipment and of procedures should be
planned to find out warning signals, before an incident
or a failure occurs. This monitoring strategy should
include inspections of safety critical plants, equipment
and instrumentation as well as assessment of compliance with training, instructions and safe working
practices.
The events potentially useful to detect early signals
of something potentially going wrong in the safety
system include deviations, anomalies, failures, nearmisses and incidents. These events should be recorded
in a data base. Attention to cases of near-misses and
minor incidents should include investigation, analysis, and follow-up to ensure that the lessons learnt
are applied to future operation. For failures, deviation
and minor failures, corrective and preventive measures
have to be ensured.
A rough classification of non conformances should
be useful to address the follow-up. In the case of human
or organizational failure, audits on the safety management system should be intensified and operational
procedures be reviewed. In the case of a mechanical
failure, inspection plan should be reviewed, to prevent

137

equipment deterioration and potential consequences

of the failure should be considered.
2

WEAK SIGNALS FOR POTENTIAL

ACCIDENTS

Many operators suppose that the non conformances

have to be recorded only in the cases of loss of
hazardous materials, stop of production or damaged
equipment; on the contrary every little anomaly,
defect, deviation or minor failure should be taken into
account, as even a silly event could be the very first
precursor or a potential concurrent to an accident. The
personnel should be encouraged in not discriminating
significant and silly events, as any event is potentially
useful for finding latent conditions which might lead to
an accident after a long time. That could also motivate
personnel in taking responsibility and stewardship.
The inspiration for this research comes from the
widespread use of personal wireless telecommunication equipment, such as a palmtop, even in the industrial premises. The palmtops should be exploited for
recording immediately any non conformance detected
in the plant, by writing notes, descriptions or comments, but also by capturing the event through pictures.
The analysis of non conformances recorded is very
useful for the early detection of conditions, which
might lead to an accident. For this purpose it is
essential to have sound methodologies, supported by
adequate tools, for understanding whether a single non
conformance could open a breaching in the safety barriers and for finding weak points of the safety system,
which could be improved. As the non conformance
is perturbing the safety system, it has to be quickly
notified, looking for a solution.
For the follow-up of near misses and anomalies
many models have been proposed in the literature.
The systemic approach is quite new for this matter.
An adequate model of industrial safety systems may
be found just in a few recent papers, including Beard &
Santos Reyes (2003), Santos-Reyes & Beard (2003),
Beard (2005), Santos Reyes & Beard (2008). The
development of a general systemic methodology for
finding latent accident precursors would be an overwhelming task indeed. As the scope of the research
is restricted to the facilities where major accident legislation is enforced and the goal is a demonstrative
prototype, the effort is instead affordable. The European legislation on major accident hazard (Seveso
Legislation) defines a framework, which structures the
safety system along the lifecycle of the plant, including
hazard identification and risk analysis, safety policy
and management, operational procedures, emergency
management and periodical safety audit. Furthermore,
in most plants a scrutiny of equipment is performed,
according to the best known hazard identification and

analysis methods, such as the IEC 61882 HAZOP

method (IEC 2001) or eventually the MOND Fire,
Explosion and Toxic index (Lewis 1979), which is
less probing, but also less difficult. In other words at a
Seveso facility, a structured and documented safety
system is always present. It is definitely ruled by
regulations and standards.
At Seveso establishment there is already a quite
complex system to manage risks, with its tools, its
documents and its data, which are described in many
papers, including Fabbri et al. (2005), OECD (2006).
For that reason the first objective of the research has
been to organize and to exploit in a better way the information already present, minimizing the efforts for the
operator. No new models have to be implemented but
the items already present in the safety system have to
be exploited. The results of non-conformances and
deviation analysis should be completely transferred
in the safety procedures, in the SMS manual and in
the safety report, in order to improve the overall risk
management in the establishment.
3

THE PROPOSED APPROACH

A new approach is proposed here for filling the gap

between safety documents and operational experience. According to standards and regulations, which
are enforced at major hazard facilities, the safety
system is well defined and may be digitally represented. The basic idea is to exploit this representation
for analyzing non conformances and improving the
safety system. That model is suitable for unscholarly
operators, including depot workers.
The backbone of the proposed system is the digital
representation of the equipment, as used in operation. In the present paper the equipment representation
has been integrated with a digital representation of
the safety system, tailored expressly for small sized
Seveso plants.
The potential of the digital models coming from
computer aided design system, for supporting hazard
analysis has been demonstrated by Venkatasubramanian et al. (2000), Chung et al. (2001), Zhao et al.
(2005) Bragatto et al. (2007). In our approach, the
scrutiny of the plant, required by the HAZID and
HAZAN methods, is exploited to build step by step a
plausible representation of the plant. The hierarchical
structure has basically five nested levels: Facility, Unit,
Assembly, Component, Accessory or Instrument.
Furthermore, as a result of hazard and consequences analysis, critical components and accessories
are discriminated. Components and accessories are
considered critical if their failures or anomalies are
in a chain of single events, which could lead to a
major accident. They are tagged and linked to the single event, present in the top events list, as handled in

138

the SR. In this way the top event list is embedded in

the digital representation of the plant.
At the end, the emergency plan may be included
in the net, too. For each major event found by the
hazard analysis, the plan should have an action or a
sequence of actions. In this way any emergency action
is linked to an event, which is in the top events list.
Furthermore the actions require usually operations to
be done on accessories (e.g. valves), which have to be
included in the plant digital representation. It is anyway to be stressed again that this whole representation
of the plant does not require any new duty for the plant
holder.
Just SR and safety manual, which are mandatory
according to the Seveso legislation, have been used.
The pieces of information, usually lost after the job
of preparing mandatory safety documents, are stored
in a quite structured way, in order to have a complete representation of the whole system, including the
equipment, the hazard ranked according to the Mond
index, the list of the top events, the sequences of failures that lead a top event, the sequence of actions
in the emergency plan, the SMS and the operating
manual.
In this structured digital environment, the nonconformances shall be recorded. The backward path
from the single event (failure, near miss or accident)
to the safety system is supported by the complete digital representation of the plant. When a failure or near
miss is reported, it is connected to a piece of equipment (component or accessory), present in the plant
digital representation.
There are many links between equipment and safety
documents, which may be used. Namely any single items of the Mond check list is linked to one or
more components or assembly, any event in a chain
leading to a top event is linked to a component, any
action of a safety procedure is linked to an accessory or a component. A few components could be
found without a direct link to safety documents; in
this case, the parent assembly or unit will be considered. In other words the path of the equipment
tree will be taken backward, in order to retrieve the
pieces of information about the assembly or the unit
affected by the failure. If the failed item is linked to
a chain, which could lead to a top event, it shall be
noticed.
If the non conforming item is corresponding to a
credit factor in the Mond index method, the credit
will be suspended, until the non conformance will be
removed.
The basic result of this continuous trial of reporting any non conformance, which happens somewhere
in the plant, and walks inside the plant digital safety
representation, is to have, after a short time, the
safety documents (basically Mond check list, top
event sequences and safety procedures) with a large

Figure 1. The proposed digital model for the management

of non conformances.

number of notes, which should be used for the periodic

reviewing.
4

IMPLEMENTATION OF THE PROPOSED

MODEL

In order to widespread the proposed methodology,

NOCE (NO-Conformance Events analysis), a software prototype, has been developed. Client-server
architecture has been adopted, in order to have a
palmtop computer on the field, connected to an
experience data base on the central server.
The basic objectives of the software prototype are
to keep track of the present event happened in the
plant by collecting all feasible information and taking into account all the events occurred in the past
in the same context;
to access directly into the Safety Management
System for updating the documents, having the
possibility to properly look at the risk analysis
results;
To evaluate the new event in the appropriate environment, for instance in safety management or in
risk analysis.

139

In order to define an adequate architecture for the

software, the workflow has been outlined. In the workflow two phases have been defined. The first phase is
basically the event recording and it is supposed to be
performed by each single worker, which detects the
failure or the non-conformance. The worker, or eventually the squad leader, is enabled to record directly
the event by means of a palmtop system. Immediate
corrective actions should be recorded too. The event
is associated to a component, which is supposed to be
present in the plant data base. The worker may verify
whether other failures had been recorded in the past.
The very basic data are introduced by the worker, the
follow up of this information are instead managed on
the desktop, by some safety supervisor. As the safety
manual addresses basic and general issues, this has
always to be verified, as the very first follow up. This
check is required for any type of non-conformances;
but it has to be deepened, in the case of human or organizational failure; which could require intensifying
audits on the safety management system or reviewing safety procedures. In the case of a mechanical
failure, asset active monitoring should be intensified
to prevent equipment deterioration. As the last step
the safety report has to be considered, as well as in
the study about hazard identification and risk assessment, including check list and HAZOP when present.
The workflow above described is summarized by the
flow-chart shown in figure 2.

4.1 Software architecture

The architecture of the proposed system is basically
composed by an application for workers, which have
to record events, and application for safety managers,
which have to analyze events and to update safety
documents. The application for the worker side has
been supposed running on a palmtop computer, featuring Windows CE; while the application for the safety
manager side has been running on a desktop computer
featuring Windows XP. The palmtop application is
aimed to record the non-conformances; to retrieve past
events from the database; to upload the new events in
the database. The desktop application has all the functionalities needed to analyze new recorded events and
to access to the databases, which contain equipment
digital representation, structured safety documents in
digital format and recorded non-conformances. The
following paragraphs describe the modules into more
details.

4.2 Palmtop application

4.2.1 Event information
As explained above, an event is considered something
happened in the facility, all weak signals coming from
operational experience, such as failures, incidents and
near-misses, which may involve specific components,
or a logic or process unit, but also a procedure of the
safety management system. For this reason the event
has to be associated to an item of the establishment,
which may be an equipment component or a procedure.
The worker is required to input the basic information about the event. The event is characterized
by some identifiers (e.g. workers name, progressive id number, date and time). Required information
includes
A very synthetic description
The type of cause, which has to be selected from a
predefined list
The type of consequence, which has to be selected
from a predefined list
Immediate corrective action (Of course the planned
preventive actions have to be approved in the next
phase by the safety manager)

Figure 2.

The basic architecture of the NOCE prototype.

Before recording any event, the worker is required

to connect the experience database and to look for
the precedents in order to find analogy with the
present case. The form is shown in figure 3. As
an equipment data base is supposed already present
in the establishments, the browsing is driven by the
non-conformance identifiers, which are linked to the
equipment component identifier or, eventually, to a
procedure identifier.

140

Figure 3. Consulting precedents related to a non conforming component.

Figure 4.

4.2.2 Event recording form

The recording phase corresponds to filling out the
event form, introducing the information which characterizes the event, as shown in figure 4. The prototype
requires all the fields to be filled in, in order to have a
complete description. After a new event has been sent
to the database, it will be analyzed.

4.3

Desktop application

An analysis of the event (accident or near miss) is

already contained in the recording format, which is
provided within the safety management system.
The key of these methods is basically the reconstruction of the event: the analysis goes along the steps
that allowed the event to continue, did not prevent it
from evolving or even contributed to it.
In order to learn better the lessons from the events,
it is essential to transfer into the safety documentation the new information or knowledge coming from
unexpected events.

Recording a non conformance.

The proposed extension is basically a discussion of

the event in comparison with the relevant points in the
safety documentation (SR, SMS, Active monitoring),
just where equipment or components involved in the
event are cited.
4.3.1 Safety management system
In analysis phase, the first step is to verify the safety
management system manual. The access to this module is mandatory for the next steps. It reflects the
action required by the Seveso legislation for updating the documents. The first action required is to
identify the paragraph in the SMS manual for introducing the link with the event occurred, and eventually
selecting the procedure or the operational instruction
involved. In this phase both general safety procedures
and operational procedures have to be checked. Every
operational procedure is linked to a few pieces of
equipment (e.g. valves, pumps or engines). If in the
plant digital model, the piece of equipment is linked to
a procedure, it is possible to go back from the failure
to the procedure through the equipment component.

141

In this way a detected non conformance or failure may

generate a procedure revision.
4.3.2 Safety report
The following steps are the reviewing of the safety
report and its related documents. These two modules are aimed to discuss the follow up of the event.
They exploit a hierarchical representation of the establishment, where equipment components, with their
accessories are nested in installations, which are
nested in units, which are nested in plants. In the safety
report, the point related to the failed components, or
to the affected plant unit, will be highlighted in order
to be analyzed. In many cases it will be enough to
understand better the risk documents; in a few cases a
reviewing could be required. A tag will be added in the
safety report. This tag will be considered when safety
report will be reviewed or revised.
4.3.3 Active monitoring intensification
In the case of a failure, inspection plan may be
reviewed, in order to intensify active monitoring
and prevent equipment deterioration. Affected accessories, components and units are found in the component database and proposed for maintenance. In the
case of unexpected failures on mechanical part the data
base may used to find in the plant all the parts, which
comply with a few criteria (e.g. same type or same
constructor). Extraordinary inspections may be considered as well as reduced inspection period. Potential
consequences of the failure may be considered too for
intensifying inspections.

Figure 5.

4.3.4 Session closing

Discussion session may be suspended and resumed
many times. At the end, the discussion may be definitely closed. After discussion closing, lessons learnt
are put in the DB and may be retrieved both from the
palmtop and from the desktop.
4.4

A case study

The NOCE prototype has been tested at a smallmedium sized facility, which had good historical
collections of anomalies and near misses. These
data have been used to build an adequate experience
database. The following documents have been found
in the safety reports:

Index method computation;

hazard identification;
list of top events;
Fault tree and event tree for potential major accident.

They have been used to develop the plant safety

digital model. Both technical and procedural failures
have been studied, using the prototype. The analysis
determined many annotations in the safety report and
in the related documents, as well in the procedures of
the safety management system. In figure 5 the list of
non conformances is shown, as presented in the NOCE
user interface.
In figure 6 a sample of a NOCE window is shown.
In the window each single pane is marked by a letter.
The order of the letters from A to F is according the
logical flow of the analysis. In A the digital representation of the plant is shown; the affected component

The list of near misses, failures and deviation, as presented in the NOCE graphical user interface.

142

Figure 6. A. The tree plant representation. B. The event data, as recorded via palmtop. C. The computation of Mond FE&T
index. D. Basic representation of the top event. E. The paragraph of the Safety Report that has been tagged. F. The lesson
learnt from the non-conformance.

is highlighted. In B the event data, as recorded via

palmtop, are shown. The computation of Mond FE&T
index is shown in pane C. the basic representation of
the event chain that leads to the top event is shown in D.
From left to right there are three sub panes, with the
initiating event, the top event and the event sequence,
which has the potential to give a top event. The paragraph of the Safety Report where the component is
mentioned is shown in pane E. The paragraph has been
tagged for a future review. The lessons learnt from the
non-conformance are shown in F.

The case study is quite simple. In a more complex

establishment, such as a process plant, the number
of failures and non-conformances is huge and many
workers have to be involved in the job of recording
as soon as possible the failures and the nonconformances. Their cooperation is essential for a
successful implementation of the proposed system.
Furthermore advanced IT mobile technologies could
be exploited for achieving good results even in complex industrial environment.

ACKNOWLEDGEMENTS

CONCLUSIONS

The software presented in this paper exploits the plant

safety digital representation, for reporting and analyzing anomalies and non conformances, as recorded
in the operational experience, as well as for updating safety report, safety management system and
related documents. For building the digital representation, no extra duties are required; but exploiting in
a smarter way documents that are compelled by the
major accident hazard legislation.

The authors wish to thank Mr Andrea Tosti for his professional support in implementing palmtop software.

REFERENCES
Lewis, D.J. (1979). The Mond fire, explosion and toxicity
index a development of the Dow Index AIChE on Loss
Prevention, New York.

143

Hursta, N.W., Young, S., Donald, I., Gibson, H. & Muyselaar,

A. (1996). Measures of safety management performance
and attitudes to safety at major hazard sites. J. of Loss
Prevention in the Process Industries 9(2), 161172.
Ashford, N.A. (1997). Industrial safety: The neglected issue
in industrial ecology. Journal of Cleaner Production
5(12), 115121.
Jones, S., Kirchsteiger, C. & Bjerke, W. (1999). The
importance of near miss reporting to further improve
safety performance. J. of Loss Prevention in the Process
Industries 12(1), 5967.
Venkatasubramanian, V., Zhao, J. & Viswanathan, V. (2000).
Intelligent systems for HAZOP analysis of complex
process plants. Computers & Chemical Engineering, 24,
22912302.
Chung, P.W.H. & McCoy, S.A. (2001). Trial of the
HAZID tool for computer-based HAZOP emulation
on the medium-sized industrial plant, HAZARDS XVI:
Analysing the past, planning the future. Institution of
Chemical Engineers, IChemE Symposium Series 148,
391404.
IEC (2001). Hazop and Operability Studies (HAZOP Studies) Application Guide IEC 61882 1st Geneva.
Phimister, J.R., Oktem, U., Kleindorfer, P.R. &
Kunreuther, H. (2003). Near-Miss Incident Management
in the Chemical Process Industry. Risk Analysis 23(3),
445459.
Beard, A.N. & Santos-Reyes, J.A. (2003). Safety Management System Model with Application to Fire Safety
Offshore. The Geneva Papers on Risk and Insurance 28(3)
413425.
Santos-Reyes, J.A. & Beard, A.N. (2003). A systemic
approach to safety management on the British Railway
system. Civil Eng. And Env. Syst. 20, 121.
Basso, B., Carpegna, C., Dibitonto, C., Gaido, G. &
Robotto, A. (2004). Reviewing the safety management
system by incident investigation and performance indicators. J. of Loss Prevention in the Process Industry 17(3),
225231.

Uth, H.J. & Wiese, N. (2004). Central collecting and evaluating of major accidents and near-miss-events in the Federal
Republic of Germanyresults, experiences, perspectives
J. of Hazardous Materials 111(13), 139145.
Beard, A.N. (2005). Requirements for acceptable model use
Fire Safety Journal 40, 477484.
Beard, A.N. (2005). Requirements for acceptable model use
Fire Safety J. 40(5), 477484.
Zhao, C., Bhushan, M. & Venkatasubramanian, V. (2005).
Phasuite: An automated HAZOP analysis tool for
chemical processes, Process Safety and Environmental
Protection, 83(6B), 509548.
Fabbri, L., Struckl, M. & Wood, M. (Eds.) (2005). Guidance
on the Preparation of a Safety Report to meet the requirements of Dir 96/82/EC as amended by Dir 03/105/EC EUR
22113 Luxembourg EC.
Sonnemans, P.J.M. & Korvers P.M.W. (2006). Accidents in
the chemical industry: are they foreseeable J. of Loss
Prevention in the Process Industries 19, 112.
OECD (2006). Working Group on Chemical Accident (2006)
Survey on the use of safety documents in the control of
major accident hazards ENV/JM/ACC 6 Paris.
Agnello, P., Ansaldi, S., Bragatto, P. & Pittiglio, P. (2007).
The operational experience and the continuous updating
of the safety report at Seveso establishments Future Challenges of Accident Investigation Dechy, N Cojazzi, GM
33rd ESReDA seminar EC-JRC-IPS.
Bragatto, P., Monti, M., Giannini,F. & Ansaldi, S. (2007).
Exploiting process plant digital representation for risk
analysis J. of Loss Prevention in the Process Industry 20,
6978.
Santos-Reyes, J.A. & Beard, A.N. (2008). A systemic
approach to managing safety J. of Loss Prevention in the
Process Industries 21(1), 1528.

144

Dynamic reliability

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A dynamic fault classification scheme

Bernhard Fechner
FernUniversitt in Hagen, Hagen, Germany

ABSTRACT: In this paper, we introduce a novel and simple fault rate classification scheme in hardware. It is
based on the well-known threshold scheme, counting ticks between faults. The innovation is to introduce variable
threshold values for the classification of fault rates and a fixed threshold for permanent faults. In combination
with field data obtained from 9728 processors of a SGI Altix 4700 computing system, a proposal for the
frequency-over-time behavior of faults results, experimentally justifying the assumption of dynamic and fixed
threshold values. A pattern matching classifies the fault rate behavior over time. From the behavior a prediction
is made. Software simulations show that fault rates can be forecast with 98% accuracy. The scheme is able to
adapt to and diagnose sudden changes of the fault rate, e.g. a spacecraft passing a radiation emitting celestial
body. By using this scheme, fault-coverage and performance can be dynamically adjusted during runtime.
For validation, the scheme is implemented by using different design styles, namely Field Programmable Gate
Arrays (FPGAs) and standard-cells. Different design styles were chosen to cover different economic demands.
From the implementation, characteristics like the length of the critical path, capacity and area consumption
result.

INTRODUCTION

6500

The performance requirements of modern microprocessors have increased proportionally to their growing
number of applications. This led to an increase of clock
frequencies up to 4.7 GHz (2007) and to an integration
density of less than 45 nm (2007). The Semiconductor
Industry Association roadmap forecasts a minimum
feature size of 14 nm [6] until 2020. Below 90 nm a
serious issue occurs at sea level, before only known
from aerospace applications [1]: the increasing probability that neutrons cause Single-Event Upsets in
memory elements [1, 3]. The measurable radiation on
sea level consists of up to 92% neutrons from outer
space [2]. The peak value is 14400 neutrons/cm2 /h
[4]. Figure 1 shows the number of neutron impacts per
hour per square centimeter for Kiel, Germany (data
from [5]).
This work deals with the handling of faults after
they have been detected (fault diagnosis). We do not
use the fault rate for the classification of fault types
such as transient, intermittent or permanent faults. We
classify the current fault rate and forecast its development. On this basis the performance and fault coverage
can be dynamically adjusted during runtime. We call
this scheme History Voting.
The rest of this work is organized as follows: in
Section 2, we present and discuss observations on fault
rates in real-life systems. In Section 3, we discuss

Number of neutron impacts

6300

6100
0

2000

4000

6000

8000

Time in hours

Figure 1. Number of neutron impacts per hour for Kiel,

Germany (2007).

related work. Section 4 introduces History Voting.

We seamlessly extend the scheme to support multiple operating units used in multicore or multithreaded
systems. The scheme was modeled in software and
its behavior simulated under the influence of faults.
Section 5 presents experimental results and resource
demands from standard-cell and FPGA (Field Programmable Gate Array)-implementations. Section 6
concludes the paper.

147

OBSERVATIONS

Figure 2 shows the number of transient single bit

errors (x-axis) for 193 systems and the number of
systems which shows faulty behavior (y-axis) over
16 months [7]. For many systems the number of faults
is small. Few systems encounter an increased fault
rate. From this, intermittent or permanent faults can
be concluded.
Figure 3 shows the daily number of transient single bit memory errors for a single system [7]. The
faults within the first seven months were identified as
transient. The first burst of faults appears at the beginning of month eleven, leading to the assessment of
intermittent faults.
Actual data is depicted in Figure 4. Here, the number of detected faults (y-axis) in the main memory

100

and caches of all 19 partitions for the SGI Altix 4700

installation at the Leibniz computing center (Technical University of Munich) [8] is shown. The data was
gained from the system abstraction layer (salinfo). In
the observation interval (24.07.-31.08.2007, x-axis)
two permanent faults in Partition 3 and 13 can be recognized, since we have a massive increase of errors
(y-axis). If permanent faults would be concluded from
a history, they would be recognized too late, leading to
a massive occurrence of side-effects. For the detection
of permanent faults, the fault rate in the observation
interval alone is relevant. In partitions 2, 5, 9, 10 and
14, an increase of the fault rate before a massive fault
appearance can be depicted. As much as a sudden
growth, a decrease of the fault rate can be observed.
From [7] further properties of intermittent faults
can be derived which help to classify these faults: a
repeated manifestation at the same location and that
they occur in bursts. Naturally, intermittent and permanent faults and can be recovered by replacing the
faulty component.

Number of systems

80

60

40

20

Number of transient faults

Figure 2.

Number of transient errors for 193 systems.

30

Number of faults

25

20

15

10

1Jul00

1May
00

1Mar
00

1Jan00

1Nov
99

1Sep
99

1Jul99

1May
99

Figure 3. Number of daily transient single bit errors of a

single system.

RELATED WORK

A state-comparator detects faults in a duplex-system.

For three or more results or states a majority voter is
used for the selection of a correct state, being able
to mask a fault. Besides majority voting many other
selection criteria exist [17]. The classification of faults
from their frequency was done in early IBM mainframes. One example is the automated fault diagnosis
in the IBM 3081 [14]. In [10] an analysis of faults
over time is used to forecast and separate permanent
from intermittent faults. The ES/9000 Series, model
900 [15] implements a retry and threshold mechanism
to tolerate transient faults and classify fault types. In
[12] the fault rate is used to construct groups of faults.
Detailed logs from IBM 3081 and CYPER-systems
help to detect similarities and permanent faults. In [16]
the history of detected faults within a NMR-system is
used to identify the faultiest module. The offline dispersion frame mechanism by Lin and Siewiorek [11]
diagnoses faults in a Unix-type filesystem. The heuristic is based on the observation of faults over time.
Different rules are derived, e.g. the two-in-one rule
which generates a warning if two faults occur within
an hour. Similar is the approach from Mongardi [13].
Here, two errors in two consecutive cycles within a
unit lead to the interpretation of a permanent fault. In
[9] the -count mechanism is developed. It permits to
model the above-named and additional variants. Faults
are weighted. A greater weight is assigned to recent
faults. Additionally, a unit is weighted with the value
. Faults are classified, but intermittent and permanent faults are not differentiated. Latif-Shabgahi and

148

Figure 4.

Field data for the SGI Altix 4700.

Bennett [16] develop and analyze a flexible majority voter for TMR-systems. From the history of faults
the most reliable module is determined. In [18] the
processor that will probably fail in the future is determined from the list of processors which took part in
a redundant execution scheme. Therefore, processors
are assigned weights. Like [9] we separate between
techniques which incur interference from outside and
mechanisms based on algorithms. The latter are used
in this work.
4

DIAGNOSIS AND PREDICTION

From the mentioned work and the observations, we

can conclude that a classification of faults is possible by measuring the time between them. The limits
for a classification are fluent, since the application,

technology and circuit will determine the frequency

of faults. History Voting for redundant (structural
and temporal) systems classifies fault rates during
runtime. It predicts if the fault rate will increase. Based
on this forecast, the performance of a system can be
dynamically adjusted if the prediction has a certain
quality.
We hereby exclude the possibility that the system
performance is decreased unreasonable due to a false
prediction. A trust is assigned to units. These can be
e.g. the components of a NMR-system, the cores in a
multicore system or threading units in a multithreaded
system. If the trust is zero, a faulty unit can be identified. An additional innovation is to dynamically adjust
the threshold values to the fault scenario. From this we
can exclude a false detection of a permanent fault on
an unexpected high fault rate, e.g. the bypass flight of
a space probe on a radiation emitting celestial body.

149

History Voting consists of two parts:

1. Adjustment of threshold values to the operating
environment
2. Prediction of the fault rate and calculation of trust
and prediction quality.
4.1

Adjustment of threshold values

All related works from above use fixed threshold limits

for the classification of fault types. This assumption
only matches an environment where the fault rate is
known. To flexibly classify the fault rate we introduce three threshold variables: , and . These
represent the upper borders for the rate of permanent
( ), intermittent ( ) and transient faults ( ). After
a reset the variables are set to known maximum values of the expected fault scenario. For security and
from the observations, is set to a fixed value and
is not dynamically adjusted. A counter i is needed
to measure the time (in cycles) between faults. In the
fault-free case, i is increased every cycle, else the
trend of the fault rate is classified by the function
i(a, b), defined by:

transient or intermittent (e.g. through frequent usage

of a faulty component) fault occurred. Thus, it cannot predict these faults and cannot derive the system
behavior over time. Therefore, we include a small
memory (the history). The history holds the last three
fault rates interpreted by i(a, b).
For a prediction, neuronal nets or methods like
branch prediction can be used. However, these methods are costly regarding time and area. We use a
simple pattern matching depicted in Table 2. Here,
the predicted fault rate and a symbolic representation are shown. For a time-efficient implementation,
the history is limited in size. Three recent fault rates
are considered. If the prediction matches the current
fault rate development, the prediction quality is
increased, else decremented. Here, also the field data
from Figure 4 was taken into account. The symbols
are shown in Table 3.
Table 2.

History and forecasted fault rates.

Fault rate

H[1][2][3]

Prediction

i : N N {0, 1}

0 if < (a, b) and
i(a, b) :=
1 if < (a, b)

:NNN

with
(a, b) := |a b| .

(1)

Table 1 shows the coding of i(a, b) and the consequence on the trust ().  represents the bitwise
shifting of value to the right, ++ an increase of
value . If a fault cannot be tolerated, the trust is
decremented until the minimum (null) is reached.
If i is substantially greater than , this could
mean that the diagnose unit does not respond. In this
case, tests should be carried out to prevent further
faulty behavior.
4.2

Forecast of fault rates, calculation

of trust and prediction quality

A further observation is that a fault-tolerant system

which does not know its past cannot express if a
Table 1.

Table 3.

Symbols (History voting).

Symbol

Description

Trust of unit i
Upper threshold: normal fault rate
Mid-threshold: increased fault rate
Lower (fixed) threshold:
permanent fault
Quality of prediction
If > , trust can be adjusted
(quality-threshold)
Value of cycle-counter
when detecting fault i
Entry i in the history
Maximal number of entries
in the history
Prediction of the fault rate
from the history (s. Figure 5)
Pattern matching to forecast
the fault (s. Figure 5)

i
H[i]
Entries

Coding of fault rates.

Coding i(a, b)

Fault rate

Consequence

Prediction

0
1

Normal
Increase

++


Predict

150

Figure 5.

Calculation of trust and prediction.

Figure 5 shows the algorithm for the calculation of

trust, the prediction and its quality.
First, we test if an irreparable internal (INTFAULT)
or external fault (EXTFAULT) was signaled. If so,
the fail-safe mode must be initiated. If no such
fault occurred, the cycle counter i is increased.
If a fault is detected (INT-/ and EXTFAULT are
excluded here), the forecasted fault rate i(a, b) is
entered into the history H. Over Predict, a prediction (Prediction) is made. The last fault rate is
compared with the current one. If the prediction
is correct, is increased. Else it will be decremented until the minimum is reached. Only if >
the prediction can modify the trust and thus
the system behavior. The more dense faults occur
in time, the less trust a unit gets. The greater the
trust, the higher the probability of a correct execution. A slow in- or decrease of the fault rate
signals a change within the operating environment
and threshold values are modified. i1 , the last

distance of faults in time will be compared with the

actual i . If the elevation or decrease is over 50%
((i > (i1  1)), (i (i1  1))), we have a
sudden change of the fault rate and threshold values
will not be adjusted.
Two possibilities to signal permanent internal faults
exist:
The trust i in unit i is less than the value pt (slow
degradation, not shown in Figure 5)
i is less than threshold (sudden increase)
Hereby, we assume that no unit permanently locks
resources, since then the trust of other units will
decrease disproportionately.
5

EXPERIMENTAL RESULTS

To judge the mechanism, it was modeled in software.

Figure 6 shows the successful adjustment of threshold

151

values (fault rate = 105 ). The distance of faults in

time, the threshold values and the accuracy are shown
(from top to bottom: , , and accuracy in %). We
purposely chose nearly equal (difference 100 cycles)
starting values for and , since we wanted to show
the flexibility of the mechanism. We see how the fault
rate is framed by threshold values. Threshold is set
to a value where a permanent fault can be ascertained
(100 cycles).
In the beginning, the accuracy is low due to the
initial values of and but evening out at about 98%.
If these values are correctly initialized, the accuracy
would have been 100%. Table 4 shows the resource
demands for History Voting (FPGA, Xilinx Virtex-e
XCV1000).

Table 5.

Place and route

Critical path (ps)
Area (2 )
Transistors
Capacity (pF)

Table 4.

3308
1175 1200
5784
8.8

Table 5 shows the resource demands for History

Voting for a standard-cell design by using a 130 nm,
6 metal layer CMOS technology.

Figure 6.

Resource demands (standard-cell).

CONCLUSION

In this work we presented a novel fault classification scheme in hardware. Apart from other schemes,
the developed History Voting classifies the fault rate
behavior over time. From this, a prediction is made.
Only if the prediction quality exceeds a known value,
the trust in units can be adjusted. From the implementations a proof of concept is made. From the results, we
see that the scheme is relatively slow. Since faults
apart from permanent onesoccur seldom in time,
this will not appeal against the scheme. It will easily fit even on small FPGAs. For the scheme, many
application areas exist. Depending on the size of the
final implementation, it could be implemented as an
additional unit on a space probe, adjusting the system behavior during the flight. Another application
area are large-scale systems, equipped with application specific FPGAs e.g. to boost the performance of
cryptographic applications. Here, the scheme could
be implemented to adjust the performance of a node,
e.g. identify faulty cores using their trust or automatically generate warnings if a certain fault rate is
reached.

Successful adjustment of threshold values.

Resource demands (FPGA).

Place and route

Critical path (ns)

9.962

REFERENCES
[1] E. Normand. Single-Event Upset at Ground Level.
IEEE Trans. on Nuclear Science, vol. 43, no. 6, part 1,
pp. 27422750, 1996.
[2] T. Karnik et al. Characterization of Soft Errors caused
by Single-Event Upsets in CMOS Processes. IEEE
Trans. on Dependable and Secure Computing, vol.
1, no. 2, pp. 128143, 2004.
[3] R. Baumann, E. Smith. Neutron-induced boron fission
as a major source of soft errors in deep submicron
SRAM devices. In Proc. of the 38th Intl. Reliability
Physics Symp., pp.152157, 2000.
[4] F.L. Kastensmidt, L. Carro, R. Reis. Fault-tolerance
Techniques for SRAM-based FPGAs. SpringerVerlag, ISBN 0-387-31068-1, 2006.

Energy consumption (at 200 MHz)

1,8 V voltage supply

mA
6.88

mW
12.39

Area
Slices
Slice-FFs
4-Input LUTs
IOBs
Gate Count

188
200
233
4
3661

152

[5] National Geophysical Data Center (NGDC). Cosmic

Ray Neutron Monitor (Kiel). ftp://ftp.ngdc.noaa.gov/
STP/SOLAR_DATA/COSMIC_RAYS/kiel.07. Revision 12/2007/cited 21.01.2008.
[6] The International Technology Roadmap for
Semiconductors (ITRS). Front End Processes.
http://www.itrs.net/Links/2005ITRS/FEP2005.pdf,
2005 Edition/cited 18.01.2008.
[7] C. Constantinescu. Trends and Challenges in VLSI
Circuit Reliability. IEEE Micro, vol. 23, no. 4,
pp. 1419, 2003.
[8] M. Brehm, R. Bader, R. Ebner, LeibnizRechenzentrum (LRZ) der Bayerischen Akademie der
Wissenschaften, Hardware Description of HLRB II,
http://www.lrz-muenchen.de/services/compute/hlrb/
hardware. Revision 29.03.2007/ cited 30.11.2007.
[9] A. Bondavalli, et al. Threshold-Based Mechanisms to
Discriminate Transient from Intermittent faults. IEEE
Trans. on Computers, vol. 49, no. 3, pp. 230245,
2000.
[10] M.M. Tsao, D.P. Siewiorek. Trend Analysis on System Error Files. In Proc. of the 13th Intl Symp. on
fault-Tolerant Computing (FTCS-13), pp. 116119,
1983.
[11] T.-T.Y. Lin, D.P. Siewiorek. Error Log Analysis: Statistical Modeling and Heuristic Trend Analysis. IEEE
Trans. on Reliability, vol. 39, pp. 419432, 1990.

153

[12] R.K. Iyer et al. Automatic Recognition of Intermittent

Failures: An Experimental Study of Field Data. IEEE
Trans. on Computers, vol. 39, no. 4, pp. 525537,
1990.
[13] G. Mongardi. Dependable Computing for Railway
Control Systems. In Proc. of the 3rd Dependable Computing for Critical Applications (DCCA-3),
pp. 255277, 1993.
[14] N.N. Tendolkar, R.L. Swann. Automated Diagnostic
Methodology for the IBM 3081 Processor Complex.
IBM Journal of Research and Development, vol. 26,
pp.7888, 1982.
[15] L. Spainhower et al. Design for fault-Tolerance in
System ES/9000 Model 900. In Proc. of the 22nd
Intl Symp. Fault-Tolerant Computing (FTCS-22),
pp. 3847, 1992.
[16] G. Latif-Shabgahi, P. Bennett. Adaptive Majority
Voter: A Novel Voting Algorithm for Real-Time faultTolerant Control Systems. In Proc. of the 25th
Euromicro Conference, vol. 2, p. 2113ff, 1999.
[17] J.M. Bass, G. Latif-Shabgahi, P. Bennett. Experimental Comparison of Voting Algorithms in Cases
of Disagreement. In Proc. of the 23rd Euromicro
Conference, pp. 516523, 1997.
[18] P. Agrawal. Fault Tolerance in Multiprocessor Systems without Dedicated Redundancy. IEEE Trans. on
Computers, vol. 37, no. 3, pp. 358362, 1988.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Importance factors in dynamic reliability

Robert Eymard, Sophie Mercier & Michel Roussignol
Universit Paris-Est, Laboratoire dAnalyse et de Mathmatiques Appliques (CNRS UMR 8050), France

ABSTRACT: In dynamic reliability, the evolution of a system is governed by a piecewise deterministic Markov
process, which is characterized by different input data. Assuming such data to depend on some parameter p P,
our aim is to compute the first-order derivative with respect to each p P of some functionals of the process,
which may help to rank input data according to their relative importance, in view of sensitivity analysis. The
functionals of interest are expected values of some function of the process, cumulated on some finite time interval
[0, t], and their asymptotic values per unit time. Typical quantities of interest hence are cumulated (production)
availability, or mean number of failures on some finite time interval and similar asymptotic quantities. The
computation of the first-order derivative with respect to p P is made through a probabilistic counterpart of the
adjoint point method, from the numerical analysis field. Examples are provided, showing the good efficiency of
this method, especially in case of large P.

INTRODUCTION

In reliability, one of the most common model used in

an industrial context for the time evolution of a system is a pure jump Markov process with finite state
space. This means that the transition rates between
states (typically failure rates, repair rates) are assumed
to be constant and independent on the possible evolution of the environment (temperature, pressure, . . .).
However, the influence of the environment can clearly
not always be neglected: for instance, the failure rate
of some electronic component may be much higher in
case of high temperature. Similarly, the state of the
system may influence the evolution of the environmental condition: think for instance of a heater which
may be on or off, leading to an increasing or decreasing temperature. Such observations have led to the
development of new models taking into account such
interactions. In this way, Jacques Devooght introduced
in the 90s what he called dynamic reliability, with
models issued at the beginning from the domain of
nuclear safety, see (Devooght 1997) with references
therein. In the probability vocabulary, such models
correspond to piecewise deterministic Markov processes (PDMP), introduced by (Davis 1984). Such
processes are denoted by (It , Xt )t0 in the following.
They are hybrid processes, in the sense that both components are not of the same type: the first one It is
discrete, with values in a finite state space E. Typically,
it indicates the state (up/down) for each component of
the system at time t, just as for a usual pure jump
Markov process. The second component Xt takes its

values in a Borel set V Rd and stands for the environmental conditions (temperature, pressure, . . .). The
process (It , Xt )t0 jumps at countably many random
times and both components interact one in each other,
as required for models from dynamic reliability: by a
jump from (It , Xt ) = (i, x) to (It , Xt ) = ( j, y) (with
(i, x), ( j, y) E V ), the transition rate between
the discrete states i and j depends on the environmental condition x just before the jump and is a function
x  a(i, j, x). Similarly, the environmental condition just after the jump Xt is distributed according to
some distribution (i,j,x) (dy), which depends on both
components just before the jump (i, x) and on the after
jump discrete state j. Between jumps, the discrete
component It is constant, whereas the evolution of
the environmental condition Xt is deterministic, solution of a set of differential equations which depends
on the fixed discrete state: given that It () = i for
all t [a, b], we have dtd Xt () = v(i, Xt ()) for all
t [a, b], where v is a mapping from E V to V .
Contrary to the general model from (Davis 1984),
we do not take here into account jumps of (It , Xt )t0 ,
eventually entailed by the reaching of the frontier of V .
Given such a PDMP (It , Xt )t0 , we are interested in
different quantities linked to this process, which may
be written as cumulated expectations on some time
interval [0, t] of some bounded measurable function h
of the process:

R0 (t) = E0

155

t
0


h(Is , Xs ) ds

where 0 is the initial distribution of the process.

Such quantities include e.g. cumulative availability
or production availability on some time interval [0, t],
mean number of failures on [0, t], mean time spent by
(Xs )0st on [0, t] between two given bounds. . .
For such types of quantity, our aim is to study their
sensitivity with respect of different parameters p P,
from which may depend both the function h and the
input data of the process (It , Xt )t0 . More specifically,
the point is to study the influence of variations of p P
on R0 (t), through the computation of the first-order
derivative of R0 (t) with respect to each p P. In view
of comparing the results for different p P, we prefer to normalize such derivatives, and we are actually
interested in computing the dimensionless first-order
logarithmic derivative of R0 (t) with respect to p:

ASSUMPTIONS

The jump rates a(i, j, x), the jump distribution (i,j,x) ,

the velocity field v(i, x) and the function h(i, x) are
assumed to depend on some parameter p, where p
belongs to an open set O R or Rk . All the results are
written in the case where O R but extension to the
case O Rk is straightforward. We add exponent(p) to
(p)
each quantity depending on p, such as h(p) or R0 (t).
(p)
We denote by t (i, dx) the distribution of the pro(p)
(p)
cess (It , Xt )t0 at time t with initial distribution 0
(independent on p). We then have:
R(p)
0 (t) =
=

IFp (t) =

p R0 (t)
R0 (t) p

t+

(R0 (t)/t)
p
R0 (t)/t
p

Noting that IFp (t) and IFp () only make sense

when considering never vanishing parameter p, we
consequently assume p to be positive.
This kind of sensitivity analysis was already studied
in (Gandini 1990) and in (Cao and Chen 1997) for pure
jump Markov processes with countable state space,
and extended to PDMP in (Mercier and Roussignol
2007), with more restrictive a model than in the present paper however.
Since the marginal distributions of the process
(It , Xt )t0 are, in some sense, the weak solution
of linear first order hyperbolic equations (CocozzaThivent, Eymard, Mercier, and Roussignol 2006), the
expressions for the derivatives of the mathematical
expectations can be obtained by solving the dual problem (adjoint point method), as suggested in (Lions
1968) for a wide class of partial differential equations.
We show here that the resolution of the dual problem provides an efficient numerical method, when the
marginal distributions of the PDMP are approximated
using a finite volume method.
Due to the reduced size of the present paper, all
proofs are omitted and will be provided in a forthcoming paper.

s(p) h(p) ds

  
iE

which we call importance factor of parameter p in

R0 (t). In view of long time analysis, we also want
to compute its limit IFp (), with

IFp () = lim

t
0


h(p) (i, x) ds s(p) (i, dx)

In order to prove existence and to calculate deriva(p)

tives of the functional R0 , we shall need the following
assumptions (H1 ): for each p in O, there is some
neighborhood N (p) of p in O such that, for all i,
j E E:
the function (x, p)  a(p) (i, j, x) is bounded on
V N (p), belongs to C2 (V O) (twice continuously
differentiable on V O), with all partial derivatives
uniformly bounded on V N (p),
for all function f (p) (x) C2 (V O), with all partial
derivatives uniformly bounded on V N (p), the

(p)
function (x, p)  f (p) (y)(i,j,x) (dy) belongs to
C2 (V O), with all partial derivatives uniformly
bounded on V N ( p),
the function (x, p)  v(p) (i, x) is bounded on
V N (p), belongs to C2 (V O), with all partial
derivatives uniformly bounded on V N ( p),
the function (x, p)  h(p) (i, x) is bounded on
V N (p), almost surely (a.s.) twice continuously
differentiable on V O with a.s. uniformly bounded
partial derivatives on V N ( p), where a.s. means
with respect to Lebesgue measure in x.
In all the paper, under assumptions H1 , for each p
in O, we shall refer to a N (p) fulfilling the four points
of the assumption without any further notice. We
recall that under assumptions H1 (and actually under
much milder assumptions), the process (It , Xt )t0 is a
Markov process, see (Davis 1984) e.g. Its transition
(p)
probability distribution is denoted by Pt (i, x, j, dy).
3

TRANSITORY RESULTS

We first introduce the infinitesimal generators of

both Markov processes (It , Xt )t0 and (It , Xt , t)t0 :

156

Definition 1.1 Let DH0 be the set of functions f (i, x)

from E V to R such that for all i E the function
x  f (i, x) is bounded, continuously differentiable
on V and such that the function x  v(p) (i, x)
f (i, x) is bounded on V . For f DH0 , we define
(p)

H0 f (i, x) =

a(p) (i, j, x)

Theorem 1.1 Let t > 0 be fixed. Under assumptions

(p)
H1 , the function p  R0 (t) is differentiable with
respect of p on N (p) and we have:
(p)

R0
(t) =
p

(p)

f (j, y)(i,j,x) (dy)

s(p)


0

(p)

+ v (i, x) f (i, x)
where we set a(p) (i, i, x) =
(p)


j
=i

a(p) (i, j, x) and

a(p) (i, j, x)

(p)

f (j, y, s)(i,j,x) (dy)

f
(i, x, s) + v(p) (i, x) f (i, x, s) (1)
s

We now introduce what we called importance

functions:

(p)

v(p)
(i, x) (i, x, s)
p

Formula (3) is given for one single p R+ . In case

R0 (t) depends on a family of parameters P = (pl )lL ,
we then have:
R(P)
0
pl

(p)

(t) =
0

s(P)

if 0 s t

(2)

(p)

and t (i, x, s) = 0 otherwise. The function t then

is the single function element of DH solution of the
partial differential equation
(p)

H (p) t (i, x, s) = h(p) (i, x)

for all (i, x, s) E V [0, t], with initial condition
(p)
t (i, x, t) = 0 for all (i, x) in E V .
(p)
The function t belongs to C2 (V O) and
is bounded with all partial derivatives uniformly
bounded on V N (p) for all p O.
(p)
The function t is called the importance function
associated to the function h(p) and to t.
The following theorem provides an extension to
PDMP of the results from (Gandini 1990).

(3)

for all DH and all (i, x, s) E V R+ .

Proposition 1 Let t > 0 and let us assume H1 to be

(p)
true. Let us define the function t by, for all (i, x)
E V:
t (i, x, s)
 ts
(Pu(p) h(p) )(i, x) du
=

H (p) (p)
(., ., s) ds
p t

H (p)
(i, x, s)
p

 a(p)
(p)
(i, j, x) (j, y, s)(i,j,x) (dy)
=
p
jE



(p)
a(p) (i, j, x) ( (j, y, s)(i,j,x) (dy)))
+
p
jE

s(p)

where we set:

(i,i,x) = x .
Let DH be the set of functions f (i, x, s) from E
V R+ to R such that for all i E the function
(x, s)  f (i, x, s) is bounded, continuously differentiable on V R+ and such that the function
(p)
x  f
s (i, x, s) + v (i, x) f (i, x, s) is bounded
on V R+ . For f DH , we define
H (p) f (i, x, s) =

jE

h(p)
ds
p

h(P)
ds
pl

s(P)

H (P) (P)
(., ., s) ds
pl t

(4)
(P)

for all l L. The numerical assessment of pl0 (t)

hence requires the computation of both s(P) (i, dx)
and t(P) (i, x) (independent on l L). This may be
done through two different methods: first, one may
use Monte-Carlo simulation to evaluate s(P) (i, dx) and
the transition probability distributions Pt(P) (i, x, j, dy),
from where the importance function t(P) (i, x) may
be derived using (2). Secondly, one may use the
finite volume scheme from (Eymard, Mercier, and
Prignet 2008), which provides an approximation for
s(P) (i, dx). The function t(P) (i, x) may then be proved
to be solution of a dual finite volume scheme, see
(Eymard, Mercier, Prignet, and Roussignol 2008).
This is the method used in the present paper for the
numerical examples provided further. By this method,
the computation of R(P)
0 /pl (t) for all l L requires
the solving of two dual finite volume schemes, as
well as some summation for each l L involving the

157

data h(P) /pl and H (P) /pl (see (4)), which is done
simultaneously to the solving.
This has to be compared with the usual finite differ(P)

R
ences method, for which the evaluation of pl0 (t) for
one single pl requires the computation of R(P)
0 for two
different families of parameters (P and P with pl sub(P)
R
stituted by some pl + ). The computation of pl0 (t)

for all l L by finite differences hence requires

1+card(L) computations. When the number of parameters card(L) is big, the advantage clearly is to the
present method.

(p)

H0 Uh(p) (i, x) = (p) h(p) h(p) (i, x)

(7)

for all (i, x) E V . Any other element of DH0 solution of (7) is of the shape: Uh(p) + C where C is a
constant.
The function Uh(p) is called the potential function
associated to h(p) .
The following theorem provides an extension to
PDMP of the results from (Cao and Chen 1997).
Theorem 1.2 Let us assume (i,j,x) and vs(i, x) to be
independent on p and H1 , H2 to be true. Then, the
following limit exists and we have:

ASYMPTOTIC RESULTS

We are now interested in asymptotic results and we

need to assume the process (It , Xt )t0 to be uniformly
ergodic, according to the following assumptions H2 :
the process (It , Xt )t0 is positive Harris-recurrent
with (p) as unique stationary distribution,
for each
O, there exists a function f (p) such
 +p (p)
+
that 0 f (u) du < +, 0 uf (p) (u) du <
(p)
+, limu+ f (u) = 0 and
|(Pu(p) h(p) )(i, x) (p) h(p) | f (p) (u)

differential equation:

(p)

lim

t+

(p)

H
1 R0
h(p)
(t) = (p)
+ (p) 0 Uh(p) (8)
t p
p
p

where we set:
(p)

H0
0 (i, x)
p

 a(p)
(i, j, x) 0 (j, y)(i,j,x) (dy)
:=
p
jE

(5)

for all 0 DH0 and all (i, x) E V .

for all (i, x) E V and all u 0.
In order not to give too technical details, we
constraint our asymptotic study to the special case
where only the jump rates a(p) (i, j, x) and the function
h(p) (i, x) depend on the parameter p. Assumptions H1
are then substituted by assumptions H1 , where conditions on (i,j,x) and on v(i, x) (now independent on
p) are removed.
We may now introduce what we call potential
functions.
Proposition 2 Let us assume (i,j,x) and v(i, x) to be
independent on p and assumptions H1 , H2 to be true.
Then, the function defined by:

Uh(p) (i, x) :=


0

Just as for the transitory results, the asymptotic

derivative requires, for its numerical assessment, the
computation of two different quantities: the asymptotic distribution (p) (i, dx) and the potential function
Uh(p) (i, x). Here again, such computations may be
done either by finite volume schemes (using (7) for
Uh(p) ) or by Monte-Carlo simulation (using (6) for
Uh(p) ). Also, in case of a whole set of parameters
P = (pl )lL , such computations have to be done only
once for all l L, which here again gives the advantage to the present method against finite differences,
in case of a large P.
5

A FIRST EXAMPLE

5.1 Presentationtheoretical results

+

((Pu(p) h(p) )(i, x) (p) h(p) )du

(6)

exists for all (i, x) E V . Besides, the function Uh(p)

is element of DH0 and it is solution to the ordinary

A single component is considered, which is perfectly

and instantaneously repaired at each failure. The distribution of the life length of the component (T1 ) is
absolutely continuous with respect of Lebesgue measure, with E(T1 ) > 0. The successive life lengths make
a renewal process. The time evolution of the component is described by the process (Xt )t0 where Xt

158

stands for the time elapsed at time t since the last

instantaneous repair (the backward recurrence time).
There is one single discrete state so that component It
is here unnecessary. The failure rate for the component at time t is (Xt ) where () is some non negative
function. The process (Xt )t0 is renewed after each
repair so that (x) (dy) = 0 (dy) and (Xt )t0 evolves
between renewals with speed v(x) = 1.
We are interested in the rate of renewals on [0, t],
namely in the quantity Q(t) such that:

 t
R(t)
1
Q(t) =
(Xs ) ds
= E0
t
t
0
where R(t) is the renewal function associated to the
underlying renewal process.
The function (x) is assumed to depend on some
parameter p > 0.
Assuming (x) to meet with H1 requirement, the
results from Section 3 here writes:
 
Q(p) (t) 1 t s (p)
(p)
=
(x)
s (dx)
p
t 0 0
p
(p)

(p)

(1 t (0, s) + t (x, s)) ds

where t is solution of
(x)(t (0, s) t (x, s)) +
+

t (x, s)
s

t (x, s) = (x)
x

Under assumption H2 , we get the following closed

form for Q()
p :
Q()
1
=
p
E0 (T1 )

f(p) (x) =

(p)

(p)

E(T1 )

x
0

(p) (u)du
(p)

E(T1 )

(9)

Using a result from (Konstantopoulos and Last

1999), one may then prove the following proposition, which ensure the process to be uniformly ergodic,
meeting with H2 :
Proposition 3 Let us assume that E(eT1 ) < +
for some 0 < < 1 and that T1 is new better than
+ t)
used (NBU: for all x, t 0, we have F(x
F(t),

F(x)
where F is the survival function F(t)
=
P(T1 > t)). Then, there are some C < + and
0 < < 1 such that:
(p)

|Pt h(p) (x) (p) h(p) | C t

for all x R+ .

(x)
p
0
 x 
v
(1 Q()
e 0 (u)du dv) dx
0

5.2

Numerical results

We assume that T1 is distributed according to some

Weibull distribution, which is slightly modified to
meet with our assumptions:

(,)

(x) =

x1 if x < x0
P,,x0 (x) if x0 x < x0 + 2

(x0 + 1)1 = constant if x0 + 2 x

where (, ) O = [0, +] [2 + ], x0 is chosen such that T1 > x0 is a rare event (P0 (T1 > x0 ) =

ex0 small) and P,,x0 (x) is some smoothing function which makes x  (,) (x) continuous on R+ .
For such a failure rate, it is then easy to check that
assumptions H1 and H2 are true, using Proposition 6.
Taking (, ) = (105 , 4) and x0 = 100 (which
ensures P0 (T1 > x0 ) 5 10435 ), we are now able
to compute IF (t) and IF () for t . In order to
validate our results, we also compute such quantities
by finite differences (FD) using:

for all s [0, t] and t (x, t) = 0 for all x [0, t].

Assuming E(T1 ) < +, the process is known to
have a single stationary distribution (p) which has the
following probability density function (p.d.f.):
P(T1 > x)

Q(t)
1
(Q(p+) (t) Q(p) (t))
p

for small and t . For the transitory results, we

use the algorithm from (Mercier 2007) which provides
an estimate for the renewal function R(p) (t) and hence
(p)
for Q(p) (t) = R t (t) to compute Q(p) (t) and Q(p+) (t).
For the asymptotic results, we use the exact formula
1
Q(p) () =
(p) to compute such quantities, which
E0 (T1 )

is a direct consequence of the key renewal theorem.

The results are gathered in Table 1 for the asymptotic importance factors IFp ().
The results are very stable for IF () by FD choosing different values for and FD give very similar
results as EMR. The approximation for IF () by FD
requires smaller to give similar results as EMR. Similar remarks are valid for the transitory results, which
are plotted in Figures 1 and 2 for t [0, 50] and different values of . This clearly validates the method. As
for the results, we may note that, for a Weibull distribution, the shape parameter is much more influent
on the rate of renewals than the scale parameter .

159

Table 1. IF () and IF () by finite differences (FD)

and by the present method (EMR).

FD

EMR

IF ()

IF ()

102
104
106
108
1010

4.625 103
8.212 102
2.411 101
2.499 101
2.500 101
2.500 101

2.824
2.821
2.821
2.821
2.821
2.821

-2

FD 10
-3
FD 10
-4
FD 10
-5
FD 10
-6
FD 10
-7
FD 10
-8
FD 10
EMR
IF ()

0.9
0.8
0.7
0.6
0.5

0.4

EMR
10

0.3
10

0.2
10

0.1
10

0
0

A tank is considered, which may be filled in or

emptied out using a pump. This pump may be in two
different states: in (state 0) or out (state 1). The
level of liquid in the tank goes from 0 up to R. The
state of the system tank-pump at time t is (It , Xt )
where It is the discrete state of the pump (It {0, 1})
and Xt is the continuous level in the tank (Xt [0, R]).
The transition rate from state 0 (resp. 1) to state 1
(resp. 0) at time t is 0 (Xt ) (resp. 1 (Xt )). The speed
of variation for the liquid level in state 0 is v0 (x) =
r0 (x) with r0 (x) > 0 for all x [0, R] and r0 (R) =
0: the level increases in state 0 and tends towards R.
Similarly, the speed in state 1 is v1 (x) = r1 (x) with
r1 (x) > 0 for all x [0, R] and r1 (0) = 0: the level
of liquid decreases in state 1 and tends towards 0. For
i = 0, 1, the function i (respectively ri ) is assumed to
be continuous (respectively Lipschitz continuous) and
consequently bounded on [0, R]. The level in the tank
is continuous so that (i, 1i, x)(dy) = x (dy) for i
{0, 1}, all x [0, R]. In order to ensure the process to be
positive Harris recurrent, we also make the following
additional assumptions: 1 (0) > 0, 0 (R) > 0 and

-2

10

-6

10

-7

10

-8

-5

-4

-3

10

x
20

30

40

1
du = +,
r0 (u)

1
du = +
r1 (u)

50

for all x, y, [0, R]. We get the following result:

10

Proposition 4 Under the previous assumptions, the

process (It , Xt )t0 is positive Harris recurrent with
single invariant distribution given by:

(i, dx) = fi (x) dx

-1

FD 10
FD 10 -2
-3
FD 10
-4
FD 10
IF ()

for i = 0, 1 and

f0 (x) =

(u) (u)
x
K R/2
( r 1(u) r 0(u) ) du
1
0
e
r0 (x)

(10)

Figure 1&2. IF (t) and IF (t) by finite differences and by

the present method (EMR).

f1 (x) =

(u) (u)
x
K R/2
( r 1(u) r 0(u) ) du
1
0
e
r1 (x)

(11)

where K > 0 is a normalization constant. Besides,

assumptions H2 are true, namely, the process
(It , Xt )t0 is uniformly ergodic.

-2

EMR
0

6.1

10

20

30

40

50

A SECOND EXAMPLE
Presentationtheoretical results

The following example is very similar to that from

(Boxma, Kaspi, Kella, and Perry 2005). The main
difference is that we here assume Xt to remain bounded
(Xt [0, R]) whereas, in the quoted paper, Xt takes its
values in R+ .

6.2 Quantities of interest

We are interested in two quantities: first, the proportion of time spent by the level in the tank between two
fixed bounds R2 a and R2 + b with 0 < a, b < R2 and

160

we set:
Q1 (t) =
=
=

6.3
1
E
t 0
1
t
1
t



t
0

i=0

We assume the system to be initially in state (I0 , X0 ) =

(0, R/2). Besides, we take:

1 R aXs R +b ds
2

1  t


R
2 +b
R
2 a

0 (x) = x0 ;

s (i, dx) ds

s h1 ds

(12)

with h1 (i, x) = 1[ R a, R +b] (x).

2
2
The second quantity of interest is the mean number
of times the pump is turned off, namely turned from
state in (0) to state out (1) by unit time, namely:

r0 (x) = (R x)0 ;

1 (x) = (R x)1 ;

1
Q2 (t) = E0
t

Numerical example


1{Is =0 and Is =1}

0<st


t
1
E0
0 (Xs )1{Is =0} ds
t
0
 t R
1
=
0 (x)s (0, dx) ds
t 0 0

1 t
s h2 ds
=
t 0


(13)

with h2 (i, x) = 1{i=0} 0 (x).

For i = 0, 1, the function i (x) is assumed to depend
on some parameter i (but no other data depends on the
same parameter). Similarly, the function ri (x) depends
on some i for i = 0, 1. By definition, the function h1
also depends on parameters a and b.
We want to compute the importance factors with
respect to p for p {0 , 1 , r0 , r1 , a, b} both in Q1 and
Q2 , except for parameters a and b which intervenes
only in Q1 .
As told at the end of Section 13, we have to compute
the marginal distribution (s (i, dx))i=0,1 for 0 s t
and the importance function associated to hi0 and t
for i0 = 1, 2. This is done through solving two dual
implicit finite volume schemes. A simple summation
associated to each p, which is done simultaneously to
the solving, then provides the result through (4).
As for the asymptotic results, the potential functions
Uhi0 are here solutions of

r1 (x) = x1

for x [0, R] with i > 1 and i > 1. All conditions

for irreducibility are here achieved.
We take the following numerical values:
0 = 1.05;
1 = 1.1;

0 = 1.2;
R = 1;

1 = 1.10;

a = 0.2;

b = 0.2.

Similarly as for the first method, we test our results

using finite differences (FD). The results are here
rather stable choosing different values for and the
results are provided for = 102 in case p
{0 , 1 , r0 , r1 } and for = 103 in case p {a, b}.
The asymptotic results are given in Tables 2 and 3,
and the transitory ones are given in Table 4 and 5 for
t = 2.
The results are very similar by FD and MR both for
the asymptotic and transitory quantities, which clearly
validate the method. Note that the asymptotic results
coincides by both methods, even in the case where the
velocity v(i, x) field depends on the parameter (here
i ), which however does not fit with our technical
assumptions from Section 4. Due to that (and to other
examples where the same remark is valid), one may
conjecture that the results from Section 4 are valid

d
vi (x) (Uhi0 (i, x)) + i (x)(Uhi0 (1 i, x)
dx
Uhi0 (i, x)) = Qi0 () hi0 (i, x)
for i0 = 0, 1, which may be solved analytically.
Q 0 ()
using
A closed form is hence available for ip
(1011) and (8).

161

Table 2. IFp(1) () by the present method (EMR) and by

finite differences (FD).
p

FD

EMR

Relative error

0
1
0
1
a
b

3.59 102

3.57 102

5, 40 103
3, 65 103
6, 95 103
7, 19 103
1, 06 107
1, 53 107

4.45 102
3.19 101
2.80 101
4.98 101
5.09 101

4.43 102
3.17 101
2.78 101
4.98 101
5.09 101

Table 3. IFp(2) () by the present method (EMR) and by

finite differences (FD).
p

FD

EMR

Relative error

0
1
0
1

1.81 101

1.81 101

6.22 102
6.05 102

6.19 102
6.01 102

1, 67 104
1, 30 104
5, 21 103
5, 58 103

1.71 101

1.71 101

Table 4. IFp(1) (t) for t = 2 by the present method (EMR)

and by finite differences (FD).
p

FD

EMR

Relative error

0
1
0
1
a
b

8.83 102
9.10 103
4.89 101
1.97 101
2.48 101
7.11 101

8.82 102
9.05 103
4.85 101
1.97 101
2.48 101
7.11 101

1, 08 103
5, 29 103
7, 51 103
4, 04 103
4, 89 104
7, 77 106

ACKNOWLEDGEMENT
The authors would like to thank Anne Barros,
Christophe Brenguer, Laurence Dieulle and Antoine
Grall from Troyes Technological University (Universit Technologique de Troyes) for having drawn their
attention to the present subject.
REFERENCES

Table 5. IFp(2) (t) for t = 2 by the present method (EMR)

and by finite differences (FD).
p

FD

EMR

Relative error

0
1
0
1

2.06 101
6.80 102
1.25 101
4.11 103

2.06 101
6.79 102
1.24 101
4.03 103

9, 12 104
2, 12 103
4, 27 103
2, 00 102

under less restrictive assumptions than those given in

that section.
As for the results, one may note that the importance
factors at t = 2 of 0 and 0 in Qi (i = 1, 2) are
clearly higher than the importance factors of 1 and
1 in Qi (i = 1, 2). This must be due to the fact that
the system starts from state 0, so that on [0, 2], the
system spends more time in state 0 than in state 1. The
parameters linked to state 0 hence are more important
than the ones linked to state 1. Similarly, the level is
increasing in state 0 so that the upper bound b is more
important than the lower one a.
In long-time run, the importance factors of 0 and
1 in Qi (i = 1, 2) are comparable. The same remark
is valid for 0 and 1 , as well as for a and b.
Finally, parameters 0 and 1 are more important
than parameters 0 and 1 in Q1 , conversely to what
happens in Q2 . This seems coherent with the fact that
quantity Q1 is linked to the level in the tank, and consequently to its evolution, controlled by 0 and 1 ,
whereas quantity Q2 is linked to the transition rates,
and consequently to 0 and 1 .

Boxma, O., H. Kaspi, O. Kella, and D. Perry (2005). On/off

storage systems with state-dependent input, output, and
switching rates. Probab. Engrg. Inform. Sci. 19(1), 114.
Cao, X.-R. and H.-F. Chen (1997). Perturbation realization,
potentials, and sensitivity analysis ofMarkov processes.
IEEE trans. automat. contr. 42(10), 13821393.
Cocozza-Thivent, C., R. Eymard, S. Mercier, and
Roussignol, M. (2006). Characterization of the marginal
distributions of Markov processes used in dynamic reliability. J. Appl. Math. Stoch. Anal., Art. ID 92156, 18.
Davis, M.H.A. (1984). Piecewise-deterministic Markov
processes: a general class of nondiffusion stochastic
models. J. Roy. Statist. Soc. Ser. B 46(3), 353388. With
discussion.
Devooght, J. (1997). Dynamic reliability. Advances in
Nuclear Science and Technology 25, 215278.
Eymard, R., S. Mercier, and A. Prignet (2008). An implicit
finite volume scheme for a scalar hyperbolic problem with
measure data related to piecewise deterministic markov
processes. J. Comput. Appl. Math. available online 1 Nov.
2007.
Eymard, R., S. Mercier, A. Prignet, and M. Roussignol
(2008, jun). A finite volume scheme for sensitivity analysis in dynamic reliability. In Finite Volumes for Complex
Applications V, Aussois, France.
Gandini, A. (1990). Importance & sensitivity analysis in
assessing system reliability. IEEE trans. reliab. 39(1),
6170.
Konstantopoulos, T. and G. Last (1999). On the use of
Lyapunov function methods in renewal theory. Stochastic
Process. Appl. 79(1), 165178.
Lions, J.-L. (1968). Contrle optimal de systmes gouverns
par des quations aux drives partielles. Paris: Dunod.
Mercier, S. (2007). Discrete random bounds for general random variables and applications to reliability. European J.
Oper. Res. 177(1), 378405.
Mercier, S. and M. Roussignol (2007). Sensitivity estimates
in dynamic reliability. In Proceedings of MMR 2007 (Fifth
International Conference on Mathematical Methods on
Reliability), Glasgow, Scotland.

162

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

TSD, a SCAIS suitable variant of the SDTPD

J.M. Izquierdo
Consejo de Seguridad Nuclear, Madrid, Spain

I. Caamn
Universidad Politcnica de Madrid, Madrid, Spain

ABSTRACT: The present status of the developments of both the theoretical basis and the computer implementation of the risk and path assessment modules in the SCAIS system is presented. These modules are
supplementary tools to the classical probabilistic (i.e. fault tree event tree and accident progression, PSAET/FT/APET) and deterministic (dynamic accident analysis) tools that are able to compute the frequency of
exceedance in a harmonized approach, based on a path and sequences version (TSD, theory of stimulated
dynamics) of the general stimulus driven theory of probabilistic dynamics. The contribution examines the
relation of the approach with classical PSA and accident dynamic analysis, showing how the engineering
effort already made in the nuclear facilities may be used again, adding to it an assessment of the damage
associated to the transients and a computation of the exceedance frequency. Many aspects of the classical safety assessments going on from technical specifications to success criteria and event tree delineation/
evaluation may be verified at once, making effective use of regulatory and technical support organizations
resources.

INTRODUCTION

This paper deals with the development of new capabilities related with risk informed licensing policies
and its regulatory verification (RIR, Risk Informed
Regulation), which require a more extended use of
risk assessment techniques with a significant need to
further extend PSA scope and quality. The computerized approach (see ESREL8 companion paper on
the SCAIS system) is inspired by a coherent protection theory that is implemented with the help of a
mathematical formalism, the stimulus driven theory
of probabilistic dynamics, SDTPD, (ref. 1) which is
an improved version of TPD (ref. 3), whose restrictive
assumptions are released.
Computer implementation of SDTPD makes necessary to design specific variants that, starting from the
SDTPD original formal specification as formulated
in ref. 1, go down the line towards generating a set of
hierarchical computing algorithms most of them being
the same as those in the classical safety assessments.
They aim to compute the exceedance frequency and
the frequency of activating a given stimuli as the main
figures of merit in RIR defense in depth regulation
requirements, in particular the adequate maintenance
of safety margins (ref. 2).

SDTPD PATH AND SEQUENCES

APPROACH

2.1

Background: semi-Markov path

and sequences

As it is well known, the differential semi-Markov equations for the probability j (t) of staying in state j at
time t are

d
j (t) = j (t)
pjk (t) + j (t)
dt
j

k =j

pkj (t) k (t)

(1)

k =j

where the transition rates p are allowed to be a function

of time. A closed solution is

t

exp ds A(s) k ( )
j (t) =

163


[A(s)]jk

jk

pjk (s)k  = j
j (s)
pjk (s)k = j
k =j

(2)

They may also be written as integral equations representing the solution of the semi Markov system for
the frequency of entering state j, (ingoing density)
j (t), and its probability j (t) given by
 t

[k ( ) ( ) + k ( )]qjk (t, )
d
j (t) =
0


j (t) =

that is
j (t) =

d[j ( ) ( ) + j ( )]e

ds

k =j

seq j

(t)

j

j11 (o)

k =j


d n

seq j

Qj

(t/
n )

(8)

j V ( <t)
n,j n

pjk (s)

(3)
where stands for the Dirac delta, and qjk dt is the
probability that, being in state k at time , event jk
takes place at exactly time t; it is then given by
qjk (t, ) = pkj (t)e

l =k

pkl (s)ds

(4)

The solution (eq. 2) may also be found in terms of

the iterative equations
j (t) =

jn (t)

n=1

jn (t) =


0

k =j

j1 (t)

d kn1 ( ) qjk (t, )

k (0) qjk (t, 0)

where the first term is the initiating event frequency;

seq j
Qj (t, n ) will be called the path Q-kernel.
Vector j is called a sequence and the couple of
vectors (j, ) a path. This type of solution will then
be named the path and sequence approach. In static
PSA, each component of j is a header of an event tree
and there is no distinction in principle between paths
and sequences.
We notice that in a semi-Markov system like the
one above, the factors q of the Qseq product are all the
same functions, irrespective of n. An extension of this
approach may be made, however, when these functions
depend on n and j, as will be the case with the TPD and
SDTPD theories. However, the normalization conditions of equations (5a) and (5b) above should always
be respected for all n. It is obvious that the path and
sequence approach looses interest when the number of
paths is unmanageable. Below we show cases where
this is not the case.

(5)

k =j

2.2
where n stands for the number of experienced events.
Note that the qs satisfy a strong closure relation for
all 1 < 2 < t

  2
qk,j (t, 2 ) 1
dvql,j (, 1 ) = qk,j (t, 1 )
l =j

(6a)
that ensures that at all times


j (t) = 1

(6b)

If we replace the iterations in eq. 5, we may express

j (t) in terms of an aggregate of so called path
frequencies built upon the products
seq j

Qj

(t/
n ) = qj,jn (t, n )qjn ,jn1 (n , n1 ).....qj2 ,j1
(2 , 1 )1 < < n1 < n < t
n (1 , . . . , n )

(7)

Extension to TPD dynamic reliability

In TPD, the transition rates p are also allowed to be

functions of process variables (temperatures, pressures, etc.) that evolve with time along a dynamic
trajectory. The transitions come together with a change
in the trajectory (a change in dynamics or dynamic
event).
Let x be the vector of process variables describing
the dynamic behavior of the system. We denote by i the
group of system configurations in which the dynamic
evolution is given by the equivalent explicit form
x = x(t, i) = g i (t, x0 ),

x0 = g i (0, x0 )

(9)

We will assume in the following that:

the system starts at a steady point state u0 , j0 with
given probability. A random instantaneous initiating event with a known frequency triggers the time
evolution.
process variable x can be reached from the initial point state only after one sequence of dynamic
events through a compound path defined by

164

x = x(t, j ) = gjn (t n+ , u n )

n > t > n1

+
, u n1 )
u n = gn1 (n n1

(10)

where n,+ account for the natural discontinuity in

slopes when the dynamic events take place.
Under the given assumptions, once the possible
dynamics and initial point state are defined, the paths
possible may be determined including its timing. We
will call a sequence the set of paths {j, jn } with the
same j, but differing in the timing. Then the general path and sequences approach may be applied such
that we associate the dynamic paths with the paths of
equation (6), (7) and (8) above. For every path, we
now have a deterministic transient that may be simulated with deterministic simulators, providing then
functions q va
seq j

qjjn (t, ) = pjn j (t, x(t, j ))e

l =jn

pjnl (s,x(s,j ))ds

(11)
2.3 Extension to stimuli: modeling the plant states

sub states jn of (jn , I ), associated with the different

dynamics are defining the dynamic paths as in the nostimuli case, but the stimuli state is kept in Markov
vector format.
Let Qi,j I ,J (t, )dt be the probability of the transition
(j, J ) (i, I ), a time interval dt around t after entering
dynamic substate j at time . In TSD terms, Q may be
written in the following way
J ,J

n1
Qjnn,jn1

,seq j

seq j

(n , n1 ) qjn ,jn1 (n , n1 )

(Jn , n /Jn1 , n1
) (12)

The last factor means the stimuli probability vec

tor at time t = n conditioned to be Jn1 (n1
) at

n1 . To compute this stimuli probability is then a typical problem of the classical PSA binary , Markov
problem with the stimuli matrix, STn , associated with
the (+ activated, deactivated) state of each stimuli
(i.e. A STn in eq. 2) and compound-stimuli state
vector I , once a dynamic path is selected, i.e. the
solution is

 t

J (t) =
STn (u)du
Kn (n+ )
(13)
exp

SDTPD includes an additional extension of the

state space: stimulus activation states are considered through a stimulus label indicating the activated
(IG = +) or deactivated (IG = ) state of stimulus G.
The notation I will be reserved to label these state
configurations when considering all stimuli together.
A so-called stimulus is either an order for action (in
many cases given by an electronic device or corresponding to an operator diagnosis), or the fulfillment
of conditions triggering a stochastic phenomenon. Yet
it can take many different forms, such as the crossing
of a setpoint or the entry in a region of the process variables space. In general, the term stimulus covers any
situation which potentially causes, after a given time
delay, an event to occur and subsequently a branching to take place. Thus, if Gn denotes the stimulus
that leads to transition jn jn+1 , it should be in the
activated state for the transition to take place.
SDTDP models the activation/deactivation of stimuli as events with activation/deactivation transition
rates F (t, x)/F (t, x) for stimulus F. The total system state in the course of a transient is thus defined
by the couple (jn , I ) which accounts at the same time
for the current dynamics in which the system evolves,
and for the current status of the different activations.

Matrix [+
n ] models the potential reshuffling of the
state of the stimuli that may take place as a result of
the occurrence of dynamic event n, i.e. some activate,
some deactivate, some stay as they were. It also incorporates the essential feature of SDTPD that stimuli
G n , responsible for the nth dynamic transition, should
be activated for the transition to take place. Then the
equivalent to eq (8) may be applied with:

n1
Qjnn,jn1

3.1

J ,Kn

with initial conditions at n+ with probability vector

Kn (n+ ) =


Jn
[+
n ]Kn ,Jn (n )

(14)

Jn

J
Qj,seq
) =
j (t/

J ,J ,seq j

Qj,jn n

J ,J

n1
(t, n )Qjnn,jn1

,seq j

J1 ,J2 ,...Jn
J ,J ,seq j

(n , n1 ) . . . . . . Qj22,j11

(2 , 1 ) J1 (0+ )

1 < n1 < n < t

(15)

with
J ,J

THE TSD METHODOLOGY

Kn

,seq j

(t, n )

seq j

Modeling the probability of the next event

along a dynamic path with stimuli events

The TSD method uses a specific type of path and

sequence approach to solving the SDTPD, where the

165

qjn ,jn1 (n , n1 )

   t
exp STn (u)du

Kn

Jn, Kn

[+
n ]Kn ,Jn1 (16)

becomes the extended next event frequency of the TSD

approach. This situation is very common and covers
the cases of setpoint related stimuli activation events
as well as those independent on the dynamics that may
be uncoupled, as extensively discussed1 in ref. 3. Once
again, this approach is not useful if the number of paths
is too large.
3.2

Example

I
Qj,seq
n ) = G+1 J G+ J . . . . . . G+ J ...
j (t/
2

I ,J ...
G1

G1

Gn

G1

Gn1

seq j
Qj (t/
n )

(22)

seq j

As an academic example, assume that all stimuli,

initially deactivated, only activate as a result of the
initiating event, and only deactivate again when its corresponding dynamic event do occur. For a sequence of
N headers, we label 1 to N the first N header-stimuli,
i.e. stimuli whose activation triggers the dynamic transition. Then all F (t, x), F (t, x) are zero and the
following results:
Define compound state JG , such that differs from
compound state J only in the state of stimulus G, being
it activated or deactivated. Also define
G+n I =

Repeating the process

1 if Gn I is activated
0 if Gn I is deactivated

(17)

Then



 t
STn (u)du = Identity matrix
STn = 0 exp

where Qj (t/
n ) is given in eq. (7). For n > N all
stimuli would be deactivated and then the [+
n ]I ,J contribution would be zero, as well as in any sequence with
repeated headers.
In other words, the result reduces to a standard
semi Markov case, but in the solution equation (8)
only sequences of non-repeated, up to N events are
allowed. This academic case illustrates the reduction
in the number of sequences generated by the stimuli
activation condition that limits the sequence explosion
problem of the TPD in a way consistent with explicit
and implicit PSA customary practice.
Results of this example in case of N = 2 are shown
in sections 45 below, mainly as a way of checking the
method for integration over the dynamic times, that is
the subject of the next section.

3.3 Damage exceedance frequency

(18)

I (n+ < t < n+1

) = I (n+ ) =

In TSD, additional stimuli are also defined, including

for instance stimulus Damage that, when activated,
corresponds to reaching unsafe conditions. Then, the
damage exceedance frequency is given by

J
[+
n ]I ,J (n )

G+n J I ,J J (n )
Gn

(19)

damage

Because after the initial event all stimuli become

activated

(t) =


j ,JN


jJ11 (0)

d 

J damage

N,
Qj,seq
j

(t/
)

Vn,j (
<t)


1 for J1 = (+, +, +N ) all stimuli activated
J1 +
(0 ) =
0 all other states

(23)

(20)

To obtain the damage exceedance frequency, only

the aggregate of those paths that exceed a given amount
of damage are of interest. Because they are expected
rare, it is of paramount importance to first find the
damage domains, i.e. the time combinations for which
damage may occur. This is the main purpose of the
searching approach that follows.
Next sections provide algorithms to perform this
aggregation and show results as applied to a simple
example. An air/steam gas mixture where hydrogen is
injected as initiator event and its rate suddenly changed
as a header event, (SIS), has been considered. See
Table 1 and section 5. In addition to those in table 1, a
combustion header stimulus is activated if the mixture
becomes flammable, and (section 4 only) a combustion

and
I (1+ < t < 2 ) = I (1+ ) =

=

G+1 J I ,J J (1 )

1 for I = JG1 = (, +, +N )
0 all other states

G1

(21)

1 This uncoupling allows to factor-out the traditional fault

trees when the basic events depend on the pre-accident

period time scales.

166

Table 1.
tests.

Dynamic/stimuli events used in the verification

Event Description
1

Initiating event: H2 injection into containment

air/steam mixtureresult of cold leg break with
core degradation.
H2 injection rate increase,result of core oxidation
increase after random initiation of Safety Injection
System (SIS) in the primary system.
Spray cold water injectionresult of random
initiation of the Containment Heating Removal
System (CHRS).
Gas mixture becomes flammable (damage stimulus
activated).

2
3
4

dynamic event (COMB) is included with an associated

peak-pressure damage stimulus2 .

INTEGRATION OF THE PATH TSD

FORMULA. THE SEARCH FOR DAMAGE
DOMAINS

4.1

Figure 1. Sampling domains (shadowed) corresponding to:

a) The 2D sequence [2 1], b) The 3D sequence [3 2 1].

[tini , tAD ]. Therefore, the whole squared domain will

contain two different sequences, depending on which
event occurs first. The shadowed domain corresponds
to the sequence [2 1] to be integrated where any point
fulfills that t2 t1 (the white domain being that of the
opposite sequence [1 2]). The area A of any of those
domains is given by the following expression:

Incremental volume of one path

stimuli activations for headers SIS and CHRS

(spray) follow the rules of the example 3.2.

b)

A(Xi Xj ) =

We can graphically represent the integration domain in

eq. (23) by points in an integration volume of dimension N . Each point corresponds to a specific path, and
we distinguish three types of paths: damage paths,
safe paths and impossible paths (those that do not
activate the header stimulus on time). For instance,
in the example below a sequence with two headers
have dimension 2 and impossible paths are those where
the flammability stimuli activates after the combustion
event.
Numerical integration of the exceedance frequency
eq. (23) will then consists of computing the Q-kernel
on the points inside the damage domain and adding up
all the contributions, multiplied by the incremental
volume ( V ) associated to each sampling point. The
way to perform the sampling strategy and the accurate
computation of the incremental volume associated to
each path are key points for the correct computation
of the exceedance frequency.
Figure 1a shows the domain of a 2-dimension (2D)
sequence containing stochastic events 1 and 2, where
X1 and X2 are the sampling variables of the occurrence
times of events 1 and 2 respectively. The time limits
of the sequence are tini (initiating event occurrence
time) and tAD (accident duration time), and X1 , X2

2 The

a)

(tAD tini )2
2

(i, j = 1, 2; i  = j)

Similarly, Figure 1b shows the sampling domain

of a 3-dimension (3D) sequence, where the shadowed
domain corresponds to the reduced sequence [3 2 1],
and the rest of the cubic domain to combinations of
those three stochastic events in different order. Here
the numbers refer to the events, not to the j states, as
in table 1.
The equation of the volume V for those 3D
sequences will be:

V (Xi Xj Xk ) =

(tAD tini )3
6

(i, j, k = 1, 2, 3; i  = j  = k)
and the generalization of this equation to N dimensions:
V (Xi1 Xi2 XiN ) =

(tAD tini )N
N!

(i1 , i2 , . . . , iN = 1, 2, . . . , N ; i1  = i2  = . . .  = iN )
(24)
We can associate to any damage point in the
N -dimensional sampling domain an incremental vol
ume of the form V = Ni=1 ti where the incremental times on each axis have to be determined.

167

4.2

perfectly determined by equation (21), being ti

the sampling step through the axis direction i.

Paths sampling strategy: Search of the damage

domain

As stated before, we are only interested in the damage

paths, as they are the only ones contributing to the
damage exceedance frequency. Therefore, a sampling
strategy has been designed to:
Minimize the number of non-damage paths (safe
and impossible paths) being sampled and analyzed;
Search and define the shape and size of the damage
domain within the sampling domain.
Three main sampling strategies may be considered
to do that:
1. Non-uniform Monte Carlo (MC). This sampling
strategy uses the probability density functions of
each event to perform the sampling, leading to a
non-uniform sampling distribution. This approach
can not be used here, as probability density functions are already taken into account in the TSD
formulae for the computation of the probability
Q-kernels.
2. Uniform Monte Carlo (MC). Points are sampled uniformly within the sequence time interval
(N alleatory numbers between tini and tAD ordered
in increasing order). However, even with a uniform sampling, the damage zone inside the sampling domain may be irregularly shaped, leading
again to a heterogeneous sampling density of the
damage points along the sampling domain (see
Figure 2a). We could partition the sampling domain
into smaller subdomains where a homogeneous distribution of the sampling points could be assumed,
and approximate the differential volume of each
point as the subdomain volume divided by the
number of points lying inside.
3. Meshgrid sampling. In this sampling strategy, sampling points are obtained from a cartesian meshgrid
partition of the sampling domain (see Figure 2b). In
this case, the volume associated to each sampling
point (damage, safe or impossible path-point) is

14000

12000

12000

10000

10000

8000
6000

a)

For the same number of sampling points (N = 136),

the damage domain is better determined with the
mesh grid sampling strategy.
Incremental volume associated to each damage path
is much easier and more accurate to quantify in the
mesh grid sampling strategy.
For refining purposes, neighboring points to each
sampled path are better defined in the mesh grid
sampling strategy.
We adopt then the mesh-grid sampling strategy.
4.3

Adaptive search algorithm

The ultimate goal is therefore to be able to precisely

define the size of the damage domain within the
sequence sampling domain, in order to multiply it by
the weighting factors given by the Q-kernels of each
path. An adaptive search algorithm has been designed,
based on the mesh grid sampling strategy, to refine
the search where damage paths are detected. The basic
idea is to analyze the neighbors of each damage path,
and to sample in a finer mesh grid around them until
non-damage paths (limits of the damage domain) are
found.
The algorithm is formed by one initial stage and an
adaptive search stage divided in three parts: refining
stage, seeding stage and growing stage. The adaptive
search stage is repeated at successively higher scales
(more refined mesh grids), until convergence of the
damage exceedance frequency is reached. We describe
the different stages as follows:
Initial stage: An initial mesh grid is defined and
points within that grid are sampled and analyzed,

Sequence [1 2 4]

14000

COMB

COMB

Sequence [1 2 4]

Remark from Figure 4 that:

8000
6000

6000

8000

10000
SIS

12000

14000

b)

6000

8000

10000
SIS

12000

14000

Figure 2. 2D sequence (SIS and COMB are events 1

and 2 respectively) analyzed with two different sampling
strategies: a. Uniform MC sampling strategy; b. Meshgrid
sampling strategy. Cross points are damage paths and dot
points are impossible paths. N = 136 paths.

Figure 3.

168

Neighbors of a given damage path.

via the simplified dynamic models to determine

whether they are a damage, success or impossible
path. In order to assure that the ulterior refinement
of the mesh grid does not ly at points contradicting the dynamic model time step, the time step of
this initial stage, dt, is chosen following a dyadic
scale of that one (and equal for all the axes). Additionally, information about the neighbors of each
sampled path is registered in an array, assigning a
value of 1 when the neighbor is a damage path, and
a value of 0 in any other case (success or impossible
path). Neighbors are defined here only those paths
at dt on each axis direction (see highlighted points
around the cross damage path in Figure 3), instead
of all the surrounding points (8 in 2D, 26 in 3D, etc).
The last option would lead to an unaffordable number of new sampled points as sequence dimension
increases. Figure 4a shows an example of this initial stage for the simplified benchmark sequence
[1 2 4] where 4 means the damage stimuli activation
event. This stage is defined here as scale 1.
Adaptive search stage: A loop is performed with
successively higher scales 2, 3, 4, etc. until a
stopping criterium is reached. This stage has the
following parts:
Refining stage: When entering a higher scale, the
new time step for the sampling process is half the
previous one, dt/2. Then, the algorithm samples
new points through the borders of the existing
a)

damage domain, i.e. if a neighbor of an existing damage point is not a damage one, then a
new point is sampled between them and evaluated. Figure 4b shows the result of this stage for
the same example sequence [1 2 4].
Seeding stage: An aleatory seeding of new sampling points along the whole domain has been
included here, in order to discover new damage
zones disjoint from the previous ones. Several
parameters control the stopping criteria in this
stage. In particular, we stop when a number of
seeding points proportional to the refinement of
the initial sampling mesh grid has been reached.
Figure 4c shows the result of this stage for the
actual example.
Growing stage: At this stage, the algorithm
extends the sampling through all the interior zone
of the damage domain. With the combination of
the refining stage + growing stage, we optimize
the number of new points being sampled while
refining all inside the damage domain.
Figure 4d shows the result of this stage in the
example sequence.

VERIFICATION TESTS

To verify the integration routines we took the simple case of the section 3.2 example, already used in
b)

Sequence [1 2 4]

Sequence [1 2 4]

13000

13000

12000

12000

11000

11000

10000

10000

COMB

COMB

14000

9000

9000

8000

8000

7000

7000

6000

6000

5000

5000
6000

8000

10000

12000

14000

6000

8000

SIS

d)

Sequence [1 2 4]

14000

14000

13000

13000

12000

12000

11000

11000

10000

10000

COMB

COMB

c)

10000

12000

14000

12000

14000

SIS

9000

Sequence [1 2 4]

9000

8000

8000

7000

7000

6000

6000
5000

5000
6000

8000

10000

12000

14000

6000

8000

10000
SIS

SIS

Figure 4. Stages of the adaptive search algorithm in the sequence [1 2 4]: a. Initial stage; b. Refining stage; c. Seeding stage;
d. Growing stage.

169

Table 2. State probabilities at t = tAD .

Analytical computation.

section 4, with two header events and constant, given

values of the two events transition rates, p. The initial event frequency has been set equal to 1. The
conditioned-to-the-initiating event path probability is
computed for each possible sequence built with those
two header events and the TSD integral of each analyzed sequence is then computed by aggregation of
the paths probabilities multiplied by the incremental
volume V as in section 4.
The dynamics of this section 4 and 5 example are
based on a benchmark exercise performed in the work
frame of the SARNET research network. It analyzes
the risk of containment failure in a nuclear power plant
due to overpressurization caused by hydrogen combustion (see ref. 4 for detailed specifications of the
benchmark). In the results presented in this section,
combustion has not been modeled. Instead, entering
the flammability region for H2 concentration has been
selected as the damage stimulus.
The events involved are given in Table 1. The details
about the dynamics of the transients are not within the
scope of this article, and can be consulted in ref. 4.
The cumulative probabilities q for the single dynamic
events 2 and 3 are 0.75 and 0.5 respectively, during
the accident time interval. Transition rates have been
extracted from them by integrating eq. (4) (with only
one event) within that interval and forcing q to be the
given values. In the analytical case compared below,
single event qs were additionally approximated by
functions linear with time.
As the transient ends when activating the damage
stimulus only five sequences are possible.

Probability

1
13
12
132
123

0.1250
0.1250
0.3750
0.1875
0.1875

Table 3. State probability at t = tAD . TSD computation

stopping criterium i i1 5%.
Sequence

# Total paths

Probability

1
13
12
132
123

1
16
16
136
136

0.1250
0.1231
0.3717
0.1632
0.2032

Table 4. State probability at t = tAD . TSD computation

stopping criterium i i1 1%.
Sequence

# Total paths

Probability

1
13
12
132
123

1
64
32
2080
2080

0.1250
0.1251
0.3742
0.1672
0.2091

Numerical results

The first part of the verification tests computes the

probability of being in a specific state j (t). Here no
damage stimulus is considered, and only probabilities
of the possible sequences obtained with the header
events 2 and 3 and initiator event 1 are computed.
Table 2 presents the linear qs analytical results.
Tables 3 and 4 show the results obtained without
q-approximations but integrating the time domains
with the SCAIS_PathQKernel.m module which
implements the TSD, and compare them with the
analytical case.
Firstly, the difference between linear q functions
in eq. 4 and uniform transition rates is only significant in sequences with more than one header event
(apart from the initiating event). This is because in
sequences with only one event, the cumulative survival
probabilities only include the event and are computed
up to tAD = 14220 s (accident duration time), and
those values (0.75 for event 2 and 0.5 for event 3) are
imposed the same in the two computations. Secondly,
the consistency of the SCAIS module results for different stopping criteria (convergence of the computed

1.4
[1]
[1 3]
[1 2]
[1 2 3]
[1 3 2]

1.2

0.8
pj
i

5.1

Sequence

0.6

0.4

0.2

0
0.4

Figure 5.

0.6

0.8

1
time (s)

1.2

1.4

1.6
x 10

Bar diagram of the j probabilities.

j probability) can be observed, as well as the fulfilment of the normalization condition given by eq. (6b),
further confirmed in Fig. 5, that shows a bar diagram
with the probabilities at all times.

170

new TSD methodology based on it is feasible, compatible with existing PSA and accident dynamic tools.
It is able to find the damage exceedance frequency
that is the key figure of merit in any strategy for
safety margins assessment. We have discussed some
issues associated to the development of its computer implementation, and given results of preliminary
verification tests.

Table 5. Damage exceedance relative frequency TSD computation of the verification example; stopping criterium
exc 5%.
iexc i1
Sequence

# Damage paths

# Total paths

Exceed. freq.

1
13
12
132
123

0
9
14
3155
1178

1
15
15
3308
1314

0.0000
0.1718
0.5188
0.0700
0.0350

REFERENCES

The second part of the verification tests computes the relative exceedance frequency exc /(0)
of the damage stimulus 4 (entering the flammability region for H2 combustion) for the same example.
Table 5 shows the values of exc /(0), of each possible
sequence, and the total relative exceedance frequency
will be the sum of them, totalling 0.7956.

CONCLUSIONS

We have shown a new math description of SDTPD

which is suitable for its SCAIS implementation. The

Labeau P.E., Izquierdo J.M., 2004. Modeling PSA problems.

I The stimulus driven theory of probabilistic dynamics.
II A cell-to-cell transport theory approach. NSE: 150,
pp. 115154 (2005).
NEA/CSNI/R(2007)9, Task Group on Safety Margins
Action Plan (SMAP) Safety Margins Action Plan. Final
Report. See also NEA/SEN/SIN/LOSMA(2008)1, Summary Record of the First Plenary Meeting of the CSNIask
Group on LOCA Safety Margin Assessment (LOSMA).
Izquierdo J.M., Melendez E., Devooght J., 1996. Relationship
between probabilistic dynamics and event trees. Rel. Eng.
Syst. Safety 52, pp. 197209.
Izquierdo J.M. Caamn I., Status Report on SDTPD
path and sequence TSD developments Deliverable
D-73. DSR/SAGR/FT 2004.074, SARNET PSA2 D73
[rev1]TSD.

171

Fault identification and diagnostics

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Application of a vapour compression chiller lumped model

for fault detection
J. Navarro-Esbr & A. Real
Department of Mechanical Engineering and Construction, Universidad Jaume, Castelln, Spain

D. Ginestar
Department of Applied Mathematics, Universidad Politcnica de Valencia, Valencia, Spain

S. Martorell
Department of Chemical and Nuclear Engineering, Universidad Politcnica de Valencia, Valencia, Spain

ABSTRACT: Faults in Heating Ventilating and Air-Conditioning (HVAC) systems can play a significant
role against the system in terms of energy efficiency loss, performance degradations, and even environmental
implications. Being the chiller one of the most important components of an HVAC system, the present work
is focused on it. A lumped model is proposed to predict the chiller fault free performance using data easily
obtained from an industrial facility. This model predicts the chilled water temperature, the operating pressures
and the system overall energy performance. The fault detection methodology is based on comparing actual and
fault free performances, using the proposed model, and operating variables thresholds. The technique has been
successfully applied for fault detection of a real installation in different faulty conditions: refrigerant leakage,
and water reduction in the condenser and in the secondary circuit.

INTRODUCTION

Air conditioning and refrigerating facilities based on

vapour compression systems account for a very significant portion of energy consumption in the industrial
and commercial sectors, Buzelina et al. (2005). A
system that automatically detects an incipient failure condition can help to improve better compliance
of installation objectives, to reduce energy use, to
optimize the facility maintenance and to minimize
undesirable environmental effects as those associated
with ozone depletion or greenhouse effect refrigerant
leakage.
The application of fault detection and diagnosis
(FDD) techniques to refrigerating vapour compression
systems is relatively recent, even though the existence
from the 70s of a large body of literature related to
FDD for critical processes, such as chemical process plants, Himmenlbau DM. (1978), and aerospace
related research, Potter & Suman (1978). The interest
of using fault detection methodology to refrigerating systems is a result of different studies focused on
energy saving in air conditioning and refrigerating systems and it is also interesting by the expected benefits
of applying FDD techniques, including less expensive
repairs and maintenance, and shorter downtimes.

The FDD techniques involve three main steps: fault

detection, the indication that a fault truly exists; fault
isolation or diagnosis, the location of the fault; and
fault identification, called fault evaluation according
to Isermann (1984), that consists of determining the
magnitude of the fault. The methods for fault detection and diagnosis may be classified into two groups:
those which do not use a model of the system (modelfree methods), Navarro-Esbr et al. (2006) and those
which do (model-based methods). Here we will focus
on the mode-based methods. These methods rely on
the concept of analytical redundancy, comparing the
model expected performance with the actual measured
performance to analyze the presence of faults in the
system.
The literature related to FDD applied to refrigerating vapour compression systems is relatively limited.
Several works analyze the most common faults and
their impacts, Breuker & Braun (1998) in rooftop air
conditioners or McKellar (1987) in domestic refrigerators. Stallard (1989), develops an expert system
for automated fault detection in refrigerators. Also in
this way, Rossi & Braun (1997) present a rule-based
diagnosis for vapour compression air conditioners.
Among the FDD model-based methods applied
to chillers, we will concentrate on those methods

175

The thermodynamic state of the refrigerant and

secondary fluids at each point is determined using
the measurements of 14 K-type thermocouples and
8 piezoelectric pressure transducers. The pressure and
temperature sensors are calibrated in our own laboratory using certified references, obtaining a degree of
uncertainty of 0.3 K in the thermocouples and a precision of 10 kPa in the pressure sensors. The refrigerant
mass flow rate is measured by a Coriolis-effect mass
flow meter with a certified accuracy within 0.22% of
the reading, and the secondary fluids mass flow rates
are measured with electromagnetic flow meters, introducing a maximum error of 0.25%. Furthermore, a
capacitive sensor is installed to obtain the compressor
rotation speed, with a maximum error after calibration
of 1%. The electrical power consumption of the compressor is evaluated using a digital Wattmeter (with a

T6, P6

T5, P5

T8, P8

T9

T2, P2
T3
P4

P3

T12
T13

EXPERIMENTAL TESTS FACILITY

For the development of the fault detection methodology, a monitored vapour-compression chiller, which
develops a simple compression cycle, has been used.
This facility consists of the four basic components:
an open-type compressor driven by a variable speed
electric motor, an isolated shell-and-tube (12) evaporator, where the refrigerant is flowing inside the tubes,
using a brine (water-glycol mixture 70/30% by volume) as secondary fluid, an isolated shell-and-tube
(12) condenser, with the refrigerant flowing along
the shell, where water is used inside the tubes as secondary fluid, and a thermostatic expansion valve. In
(Fig. 1) a scheme of the installation is shown.
In order to introduce modifications in the chiller
operating conditions, we use the secondary fluids
loops, which help to simulate the evaporator and condenser conditions in chillers. The condenser water
loop consists of a closed-type cooling system, (Fig. 2),
which allows controlling the temperature of the water
and its mass flow rate.
The cooling load system (Fig. 3) also regulates
the secondary coolant temperature and mass flow rate
using a set of immersed electrical resistances and a
variable speed pump.

T11

T7, P7

T4

T10
T

based on physical models. These physical models are

based on first principles involving mass, energy and
momentum balances and mechanical characteristics.
The ability of a model-based FDD technique to detect
faults during chiller operation depends on the model
performance. It is desirable that the model presents a
good prediction capability in any operating condition
and that it could be obtained in a fully automatic way,
combining simplicity with low degree of data requirement. In this way, we will present a simplified physical
model for a chiller facility and we will use it to develop
a fault detection methodology for three typical faults
in this kind of facilities, as faults in the condenser circuit, faults in the evaporator circuit and leakage in the
refrigerant circuit.
The rest of the paper is organized as follows.
Section 2 is devoted to the description of the experimental facility used to develop and to test the fault
detection methodology. In section 3, the proposed
fault detection methodology is presented and a simple model for the vapour compression chiller is briefly
described. The commonest failures of this kind of
installations are studied in section 4, including a sensitivity analysis of the most affected variables for these
failures. In section 5, the fault detection methodology
is tested with different typical faulty operation modes.
The main conclusions of the paper are summarized in
section 6.

Figure 1.

Scheme of the installation.

Figure 2.

Condenser water loop.

176

P1, T1

calibration specified uncertainty of 0.5%). Finally,

all these measurements are gathered by a National
Instruments PC-based data acquisition system (Fig. 4),
based on LABVIEW and REFPROP subroutines.

The fault detection is accomplished by evaluating the

residuals obtained from the comparison between the
actual performance of the installation, determined
from experimental measurements, and some expectation of performance obtained by using a physical
model of the system (Fig. 5). If the residuals exceed a
given threshold, then a fault is indicated.
It has to be noted that the selected output variable,
xk , must be proven to be representative of the system
operation, and its selection depends on the kind of fault
to be detected. For this reason, a previous analysis of
how a specific fault affects the main chiller operating variables is needed, being the most sensitive and
inexpensive ones selected as output variable for the
fault detection technique.
3.1

Figure 3.

Load simulation system (evaporators circuit).

VAPOUR COMPRESSION CHILLER

SCXI ACQUISITOR SYSTEM

NATIONAL INSTRUMENTS

SIGNAL CONDITIONING WITH

LABVIEW

MEASUREMENT & AUTOMATION

MATLAB

REFPROP

FAULT DETECTION METHODOLOGY

Physical model

The proposed model for the vapour compression

chiller has a general structure as the one presented
in (Fig. 6), where it can be seen that the model
inputs are the secondary fluids input variables and the
compressor speed. Using these inputs and the main
characteristics of the compressor and heat exchangers, the model predicts the fault-free performance:
operating pressures, secondary fluids output variables
(operating pressure and temperatures) and energy performance (including power consumption and energy
efficiencies).
The model computes the refrigerant properties
using dynamic libraries of Refprop, Lemmon et al.
(2002), while the secondary fluids thermo-physical
properties are evaluated by means of interpolating
polynomials calculated from the ASHRAE Handbook
(2001).
The kernel of the model consists of a set of five
equations based on physic laws, which model the main
parts of the system, shown schematically in (Fig. 7).
Comparing the measured variable with the calculated one, a residual is obtained. While this residual
remains between two confidence thresholds the operation will be considered fault-free. Some consecutive
values of the residual out of these thresholds will
generate an alarm signal and the identification of a
possible fault will proceed.

USER INTERFACE
Actual
performance

Residual
generator

System
Model (expected
performance)

Figure 4.

Scheme of the data acquistion system.

xk

Figure 5.

177

~x
k

Fault detection technique scheme.

Residuals
evaluation

A typical fault on both the evaporator and the condenser circuits may be the reduction of the mass flow
rate of the secondary fluids through the interchangers. These reductions can deteriorate the compliance
of installation objectives and, on the other hand, can
cause energy efficiency reduction due to mass flow
rate reduction of the secondary fluids at the condenser
and evaporator, which produces an increase of the
compression rate.
Thus, in the following, we will analyze these three
kinds of failures. The first step for the correct identification of each type of fault is to analyze the variations
produced bye each kind of faulty operation, followed
by the selection of a reduced group of variables that
have a significant variation during the fault that can be
used to detect it.
In order to identify these variables some variations
have been forced in the facility in such a way that they
can simulate those failures. Then, the correlation of
each variable with each fault has been calculated both
in the short term and medium term or quasi-steady
state after it.

Refrigerant thermophysical properties

Tb ,out
Tw,out

mb , Tb ,in

pe
pk

Model

mw , Tw,in
N

PC

COP
Geometric characteristics of the system
Figure 6.

Model scheme.

water

water

vapour to saturated liquid

superheated vapour

air
superheated vapour

Expansion valve

4.1 Fault in the condensers circuit

Compressor

refrigerant to saturated vapour

brine

Figure 7.

Schematic structure of the model kernel.

Special attention must be paid to the assignation of

those confidence thresholds. On one hand, too narrow
thresholds will increase the possibility of false alarms
by confusing a noisy signal with a deviation due to
a fault. On the other hand, too wide thresholds will
increase the possibility of ignoring a fault.

In order to simulate a fault in the condenser circuit, a

valve has been partially shut down to reduce water
mass flow rate. In a real facility this fault may be
caused by an obstruction in a filter of the circuit, or
a failure in the impulsion pump. Table 1 shows the
variation of each one of the main operating variables
of the facility after the shut down of the valve.
The variable that shows best conditions to be monitored is the condenser pressure, Pk . The evolution
of this variable is shown in (Fig. 8). Some temperatures also show deviations such as the condenser
outlets temperature, but they are not only smaller, but
also slower as the variation can only be seen after the
transient state.
Table 1. Deviations of the main operating variables of
the facility after a fault in the condensers circuit (short
term, ST, and medium term, MT).

COMMON FAILURES AND SENSITIVITY

ANALYSIS

The most common failures to be detected on vapour

compression chiller are: refrigerant leakage, fault in
the condenser circuit and fault in the evaporator circuit.
A refrigerant leakage may be difficult to detect since
it only affects the system operation after a large loss
of refrigerant. When it occurs, refrigerant in vapour
phase may appear in the outlet of the condenser. This
will interfere in the expansion device normal function,
it will increase the quality of the refrigerant at the evaporator inlet and thus, the efficiency of the system will
be reduced.

178

Variable

Initial

ST var (%)

MT var (%)

Pk (bar)
Pe (bar)
Tki (K)
Tko (K)
Toi (K)
T (K)
Twi (K)
Two (K)
Tbi (K)
Tbo (K)

15.18
4.00
0.068
347.63
308.61
265.09
266.51
291.00
297.68
286.89

1.46
0.22
0.08
0.02
0.04
0.02
0.00
0.01
0.11
0.02

1.96
0.22
0.73
0.14
0.23
0.01
0.49
0.05
0.15
0.00

4.2

Fault in the evaporator circuit

The simulation of this failure has been carried out in

a similar way to the one for the condensers circuit.
A valve at the secondary fluid loop has been partially
shut down in order to reduce its mass flow rate. An
obstruction in a filter or an impulsions pump failure
may be the real causes of this kind of failure. (Table 2)
shows the deviations of the main operating variables
of the facility.

As shown in (Table 2), both the pressure at the condenser and the pressure at the evaporator are likely to
be monitored. Furthermore both have a clear reaction
in the transient state and after it, however the evaporator pressure has approximately three times higher
deviation than the pressure at the condenser and thus,
it will be a more reliable indicator for the failure detection. The response of the pressure in the condenser is
shown in (Fig. 9).
4.3 Refrigerant leakage

15.7
15.6
15.5

Condenser
Pressure

15.4
15.3
15.2
15.1
15.0
14.9
1700

1900

2100

2300

2500

2700

2900

3100

Figure 8. Response of the condenser pressure when a fault

is forced in the condensers circuit.
Table 2. Deviations of the installations main variables
after a fault in the evaporators circuit.
Variable

Initial

ST var (%)

MT var (%)

Pk (bar)
Pe (bar)
Tki (K)
Tko (K)
Toi (K)
T (K)
Twi (K)
Two (K)
Tbi (K)
Tbo (K)

16.28
3.60
352.97
311.26
263.45
266.52
295.89
302.03
284.44
352.97

1.02
3.77
0.00
0.04
0.27
0.09
0.03
0.03
0.01
0.00

1.20
3.38
0.05
0.16
0.34
0.55
0.03
0.10
0.06
0.05

To simulate a refrigerant leakage in the chiller a small

valve in the refrigerant circuit has been temporally
open, thus allowing the refrigerant to flow out. The
response of the system variables is shown in (Table 3).
The analysis of variations presented in (Table 3)
shows that superheat (difference between evaporator
outlet temperature, Too , and saturation temperature,
Toi ) and pressure at the evaporator are the most sensitive parameters to a refrigerant leakage fault. However,
the pressure at the evaporator is the most sensitive
variable, having the advantage of being a directly
measured variable. So, the simplest approach consists
of using the pressure at the evaporator for the fault
detection (Fig. 10).
Table 3. Deviations of the main operating variables in a
refrigerant leakage.
Variable

Initial

Var (%)

Pk (bar)
Pe (bar)
Tki (K)
Tko (K)
Toi (K)
T (K)
Twi (K)
Two (K)
Tbi (K)
Tbo (K)

15.21
3.89
347.57
308.48
264.91
266.41
291.05
297.66
287.07
347.57

1.35
4.60
0.00
0.02
0.15
0.21
0.01
0.05
0.00
0.00

!b

3.85

4.10

3.80

4.05
Pressure [bar]

Pressure [bar]

3.75
3.70
3.65
3.60

Evaporator
Pressure

4.00
3.95
3.90

Evaporator
Pressure

3.55
3.85

3.50
3.45
17300

17800

18300

18800

3.80
7700

19300

Time(s)

7900

8100

8300

8500

8700

8900

Time (s)

Figure 9. Response of the condenser pressure after a fault

forced in the condensers circuit.

Figure 10. Response of the evaporator pressure in a

refrigerant leakage.

179

RESULTS

0.8

In this section the methodology is validated, using

the model to predict the fault free performance and
comparing measured and predicted variables. As the
evaporator and condenser pressures are the most sensitive variables to the chiller faults, the residuals of
these two variables are calculated.
To follow the behaviour of the fault-sensitive variables a threshold must be set, to decide which deviation
of the residual must be considered as a fault and which
one must be considered as noise of the signal.
The selection of those thresholds must be studied
separately in each case, as it will define the sensibility and robustness of the fault detection methodology.
In the following experiments, the criteria followed to
set the thresholds is to use a 3 limit, where represents the standard deviation of the error between the
measured and the expected value of the monitored variable. The value of is computed using an initialization
fixed time, tinit , as:
max [r (t0 ) , . . . , r (tinit )]
=
[r (t0 ) , . . . , r (tinit )]

Fault in the condensers circuit

Fault in the evaporators circuit

In a similar way, a fault in the evaporators circuit is

simulated by shutting down partially a valve in the circuit. This action causes the reduction of the secondary
fluid mass flow rate.

+3

-3

0.2
0.0
-0.2
3300

3500

3700

3900

4100

4300

4500

4700

Figure 11. Residuals of the condenser pressure due to a

fault in the condensers circuit.

0.20
0.10
0.00
-0.10
-0.20
-0.30

As has been exposed above, the best variable to predict a fault in the condensers circuit is the condenser
pressure. In order to verify if the fault would be predicted when it takes place, we will force this fault
in the facility by shutting down partially the valve at
the condenser circuit. The evolution of the residuals
associated with the condenser pressure is shown in
(Fig. 11).
As we can see in this Figure, the residuals of the condenser pressure remains inside the thresholds until the
fault takes place. That shows that the selected variable
and the methodology proposed would detect the fault,
not causing false alarms during fault-free operation.
5.2

0.4

(1)

where r(t) is the value of the residual at time t.

Thus the behaviour of the error will be considered
valid while it remains between +3 and 3 , and the
alarm will take place as the error runs out of these
limits.
The following experiments have been carried out
in order to test the functioning of the fault detection methodology. For that, some deviations have been
forced to simulate the faults analyzed in the paper.
5.1

0.6

-0.40
5000

+ 3

- 3

5500

6000

6500

Figure 12. Evaporating pressure residuals evolution when

a fault in the evaporators circuit is introduced.

As it has been shown in section 4, the most sensitive

variable of this fault is the evaporator pressure, so this
will be the monitored variable. The evolution of the
evaporating pressure residuals is shown in (Fig. 12).
As shown in this Figure, the fault is clearly detected,
since the residuals are clearly out of the thresholds
when it occurs. Furthermore, the residuals remain
inside the thresholds during the fault-free operation,
showing the ability of the methodology to avoid false
alarms.

5.3 Refrigerant leakage

To carry out the simulation of this fault, in steadystate operation a continuous small leakage from the
refrigerant circuit is introduced.
As it has been presented above, to detect this fault
the evaporator pressure is the most sensitive variable
to the refrigerant leakage. So, the residuals between
expected and actual evaporator pressure are followed.
(Fig. 13) shows the behaviour of the residuals
between measured and expected value of the evaporator pressure.

180

heat load variations, in order to see if it is possible to

distinguish this kind of transients from the ones associated with a fault in the chiller. Once a fault is detected,
an interesting subject is to couple the fault detection
technique with a classification methodology to be able
to make a diagnostic of the type of failure that is taking
place.

0.10
0.05

Residuals

0.00
-0.05
+3
-0.10
-0.15
-0.20
8000

-3

REFERENCES
8200

8400

8600

8800

Time (s)

Figure 13.
leakage.

Evaporator pressure residuals in the refrigerant

As it can be observed in this Figure, the residuals

suffer a sharp deviation out of the thresholds, showing
the capability of the methodology proposed to detect
this kind of fault.
6

CONCLUSIONS

The aim of this article has been to present a

methodology to detect common faults in a vapourcompression chiller.
Two kinds of experiments have been carried out.
A set of experimental tests used to analyze the variations of the operating variables during the faults, in
order to set the most fault-sensitive variables. Another
set of experiments or validation tests, used to test the
performance of the fault detection methodology.
In the methodology a simple steady state model has
been used to obtain the fault-free expected values of
the fault-sensitive variables. These values have been
used to obtain the residuals between the expected and
measured variables. Finally, a methodology based, on
static thresholds, has been applied. The experimental
results have shown the capability of the technique to
detect the faults.
In a future work the behaviour of the fault detection
methodology should be tested with transients associated with the normal operation of the installation, as

ASHRAE 2001. Handbook of Fundamentals, Chap. 21.

Breuker M.S., Braun J.E., 1998. Common Faults and their
Impacts for Rooftop Air Conditioners. International Journal of Heating, Ventilating, and Air Conditioning and
Refrigerating Research, 4 (3): 303318.
Buzelina L.O.S., Amico S.C., Vargas J.V.C., Parise J.A.R.,
2005. Experimental development of an intelligent refrigeration system, International Journal of Refrigeration 28:
165175.
Himmenlbau D.M., 1978. Fault Detection and Diagnosis in Chemical and Petrochemical Processes. Chemical
Engineering Monographs, Vol. 8, Elsevier.
Isermann R., 1984. Process Fault Detection Based on Modeling and EstimalionA Survey. Automatica, 20 (4):
387404.
Lemmon E.W., Mc Linden M.O., Huber M.L., 2002. Reference Fluid Thermodynamic and Transport Properties.
REFPROP v7.0. NIST Standard Reference Database 23.
McKellar M.G., 1987. Failure Diagnosis for a Household Refrigerator. Masters thesis, School of Mechanical
Engineering, Purdue University.
Navarro-Esbr J., Torrella E., Cabello R., 2006. A vapour
compression chiller fault detection technique based on
adaptative algorithms. Application to on-line refrigerant leakage detection. Int. Journal of Refrigeration 29,
716723.
Potter J.E., Suman M.C., 1978 Thresholdless redundancy
management with arrays of skewed instruments. Ingegrity
in Electronic Flight Control Systems, AGARDigraph224, 1977, 15: 125.
Rossi T.M., Braun J.E., 1997. A Statistical Rule-Based Fault
Detection and Diagnostic Method for Vapor Compression
Air Conditioners. HVAC&R Research, 3 (1): 1937.
Stallard, L.A., 1989. Model Based Expert System for Failure
Detection and Identification of Household Refrigerators. Masters thesis, School of Mechanical Engineering,
Purdue University.

181

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Automatic source code analysis of failure modes causing error propagation

S. Sarshar
OECD Halden Reactor Project/Institute for energy technology, Halden, Norway

R. Winther
stfold University College, Halden, Norway

ABSTRACT: It is inevitable that software systems contain faults. If one part of a system fails, this can affect
other parts and result in partial or even total system failure. This is why critical systems utilize fault tolerance
and isolation of critical functions. However, there are situations where several parts of a system need to interact
with each other. With todays fast computers, many software processes run simultaneously and share the same
resources. This motivates the following problem: Can we, through an automated source code analysis, determine
whether a non-critical process can cause a critical process to fail when they both run on the same computer?

code is proposed. Section 6 discusses the proposed

approach and concludes our work.

INTRODUCTION

In this paper we report on the results from the practical application of a method that was first presented
at ESREL 2007 (Sarshar et al. 2007). The presented
method identified failure modes that could cause error
propagation through the usage of the system call interface of Linux. For each failure mode, its characteristics
in code were determined so it could be detected when
analyzing a given source code. Existing static analysis
tools mainly detect language specific failures (Koenig
1988), buffer overflows, race conditions, security vulnerabilities and resource leaks, and they make use of
a variety of analysis methods such as control and data
flow analysis. An examination of some of the available tools showed that no existing tool can detect all
the identified failure modes.
The purpose of this paper is to present and evaluate a
conceptual model for a tool that allows automatic analysis of source code for failure modes related to error
propagation. Focus is on the challenges of applying
such analysis automatically. An algorithm is proposed,
and a prototype implementation is used to assess this
algorithm. In contrast to existing tools, our checks
specifically focus on failure modes that can cause error
propagation between processes during runtime. As we
show by applying the tool on a case, most of the failure modes related to error propagation can be detected
automatically.
The paper is structured as follows: Section 2
describes the background and previous work. Section 3
describes the analysis method (Sarshar et al. 2007).
In Section 4, existing tools are considered and in
Section 5 a method for automatically assessing source

BACKGROUND

We performed a study of Linux to understand how

processes can interact with the operating system and
other processes. This quickly turned our attention to
the system call interface, which provides functions to
interact with the operating system and other processes.
We analyzed these system calls in detail and identified
how these functions could fail and become mechanisms for error propagation. A system call may take
some arguments, return a value and have some systemwide limitation related to it. To identify failure modes
we applied FMEA on each of the function components. The failure effects were determined using fault
injection testing. Then, we determined the characteristics of the failure modes in code for detection when
analyzing an applications source code. Thus, the proposed methodology enabled criteria to be defined that
could make it possible to automatically analyze a given
source code for error propagation issues.
In (Sarshar 2007), the methodology was applied on
a subset of SYSTEM V APIs related to shared memory.
This target was found to be suitable because it involved
an intended channel for communication between processes through a shared resource; the memory. We also
performed FMEA on other system calls to evaluate
whether the method was applicable to a wider class of
functions and not restricted to those related to shared
memory.

183

The errors identified in this approach were erroneous values in the variables passed to the system
call interface and errors caused when return, or modified, pointer variables were not handled properly.
From the analysis we know not only which functions behave non-robustly, but also the specific input
that results in errors and exceptions being thrown
by the operating system. This simplifies identification of the characteristics a failure mode has in
source code.
Our proposed approach of analyzing error propagation between processes concerns how the process
of interest can interact with and affect the environment (the operating system and other processes).
A complementary approach could be to analyze how
a process can be affected by its (execution) environment. In (Johansson et al. 2007), the authors inject
faults in the interface between drivers and the operating system, and then monitor the effect of this faults
in the application layer. This is an example where processes in the application layer are affected by their
execution environment. Comparing this method to our
approach, it is clear that both methods make use of
fault injection to determine different types of failure
effects on user programs. However, the examination
in (Johansson et al. 2007) only concerns incorrect values passed from the driver interface to the operating
system. Passing of incorrect values from one component to another is a mechanism for error propagation
and relate to problems for intended communication
channels. Fault injection is just one method to evaluate a process robustness in regards to incorrect values
in arguments. In (Sarshar 2007), the failure effects
of several mechanisms were examined: passing of
arguments and return value, usage of return value,
system-wide limitations, and sequential issues. These
methods complement each other. Brendan Murphy,
co-author of (Johansson et al. 2007), from Microsoft
Research1 pointed out his worries at ISSRE 2007:
The driver developers do not use the system calls
correctly. They do for instance not use the return values from the system calls. It is nothing wrong with the
API, it is the developer that does not have knowledge
about how to use the system calls.
Understanding the failure and error propagation
mechanisms in software-based systems (Fredriksen &
Winther 2006) (Fredriksen & 2007) will provide the
knowledge to develop defences and avoid such mechanisms in software. It is therefore important to be aware
of the limitations for the proposed approach. This analysis only identifies failure modes related to the using
of system calls in source code. Other mechanisms for

1 Cambridge,

UK.

error propagation that do not involve usage of the system call interface will not be covered by this approach.
Eternal loop structures in code is an example of a failure mode that does not make use of system calls. This
failure mode can cause error propagation because it
uses a lot of CPU time.
3

ANALYSIS

In (Sarshar 2007) we analyzed several system calls and

identified almost 200 failure modes related to their
usage. Because these have their manifestation in the
code, it is expected that they also have characteristics
that can make them detectable when analyzing source
code. This section describes general issues regarding
the characteristics the identified failure modes will
have in code.
The identified failure modes can be categorized
depending on the object they apply to:
Arguments (syntax analysis)
Return variable (logical analysis)
The functions sequential issues and limitations
(logical analysis)
Determination of whether arguments contain a specific failure mode can be done with syntax analysis.
Syntax analysis can identify the passed arguments
variable type and value and check it against the list
of failure modes for that specific argument. For the
return variable, and for the functions sequential issues
and limits, logical analysis is required. This involves
tracking of variables and analysis of the control flow
for the code.
Failure modes related to passing of arguments can
be divided into two groups: (1) variables and (2)
pointers. A passed argument variable can have failure modes related to its type which will be common
for all arguments of that given type. In addition it may
have failure modes related to the context it is used, that
is in the context of the system call. Pointer arguments
can in a similar way have failure modes related to their
type, but will in addition have failure modes related to
its usage in the code. Pointer variables or structures are
often used to return data from a function when used as
an argument.
The following failure modes may occur for a return
variable (applies to all system calls with a return
variable):
A. The value is not used in the code
B. The value is stored in a variable of wrong type
C. The external variable errno is not used if return
value indicates an error
To determine whether failure mode A occurs in
a given source code: (1) the return variable must be
retrieved, that is, not ignored with e.g. usage of void;

184

Table 1.

Some failure mode characteristics in code for shmget().

Ref.

Failure mode

Detectability in code

Check

F.29.1.D

Parameter key is
less than type key_t
Parameter key is
greater than type key_t
Parameter key is
of wrong type
IPC_PRIVATE specified
as key when it should not
Parameter size is
of wrong type
Parameter shmflg is
not legal

Check that value is not below zero

Value

Check that value is not higher than

type key_t
Check the type of the passed variable
to be type key_t
Logical error that can be difficult
to identify without program documentation
Check the type of the passed variable
to be type size_t
Calculate the values of legal flags
(documented in man page) and verify that
the passed value is legal
Check the type of the passed variable
to be type int
Check the least 9 bits of shmflg to get hold
of permission mode for user, group and all.
If none of these are equal to 5, 6 or 7,
permission mode is not set correct
Can be detected using the same approach as for
29.3.D, but whether it is intended is
difficult to predict without documentation
Mode should have value 5 but has value 7,
but the intention is difficult to predict without
documentation
Mode should have value 7 but has value 5,
not an issue unless a write is performed on
the segment later in the code
Track the variable to the end of scope
and check whether it is used in the code

Value

F.29.1.E
F.29.1.F
F.29.1.G
F.29.2.F
F.29.3.B
F.29.3.C
F.29.3.D

F.29.3.E
F.29.3.F
F.29.3.G
F.29.4.A

Parameter shmflg is
of wrong type
Permission mode is not
set for parameter shmflg
Access permission is
given to all users instead
of user only
Permission mode is
write when it should
have been read
Permission mode is
read when it should
have been write
Return value is not used

(2) the variable must be traced to the end of its scope

to determine whether it has been used in some way.
Failure mode B can be checked with syntax analysis
by verifying that the variable storing the return value is
of the same type as described in the documentation for
that specific function. In order to determine whether
failure mode C is present in the code: (1) the return
value must be used in a statement to check whether
it indicates an error; (2) the errno variable must be
checked in a statement before the end of the function
block.
Detection of sequential issues using system calls
requires logical analysis. One must keep track of return
variables that are used as arguments for other system
calls, to determine which calls that are related. This
information can then be used to determine whether
they are run in the correct order. Use of control flow
graphs can be very helpful for analyzing source code
for such failure modes.
Failure modes related to functional and system wide
limits can be difficult to determine when analyzing
source code only. If we do not know the limitations

Type

Type
Value
Type
Value

Value
Value
Value
Variable
usage

of the underlying operating system for e.g. maximum

number of files one process can have open, it is difficult to predict such a failure mode. However, with
use of logical analysis and control flow graphs, one
may be able to determine whether the code open many
files and warn the user that such a failure mode may
be present in the code.
Though not all failure modes can be detected when
analyzing source code, a warning can be given when a
given failure mode might be of concern. For instance,
if it cannot be determined whether a return variable is
used, a warning can be issued to the programmer.
Table 1 lists an excerpt of failure modes with their
characteristics in source code for the shmget()2 system
call and Table 2 lists failure modes and their characteristics in source code related to sequential issues when
using shared memory services.

2 Allocates

185

a shared memory segment.

Table 2.

Failure mode characteristics in code related to sequential issues for shared memory.

Ref.

Function

Failure mode

Detect-ability in code

Check

F.shm.A

shmdt()

Target segment
is not attached

Line numbers

F.shm.B

shmat()

Segment to attach
is not allocated

F.shm.C

shmctl()

Target segment
is not identified

F.shm.D

shmat()

Segment is not
detached prior
to end of scope

Check line nr of shmdt() and shmat()

to verify that shmat() is run prior
to shmdt() for a specific segment
Check line nr of shmat() and shmget()
to verify that shmget() is run prior
to shmat() for a specific segment
Check line nr of shmctl() and shmat()
to verify that shmat() is run prior
to shmctl() for a specific segment
Check line nr of shmat(), shmdt() and
end of scope (return), to verify
that shmdt() is run prior to return
for a specific segment from shmat()

EXISTING TOOLS

Table 3.

A tool search was made on the internet in order to

find descriptions of existing analysis tools. We identified 27 tools related to analysis of C code based
on this search3 . These can be grouped in commercial tools, and academic research tools. Although cost
prevented actual testing of the commercial tools, the
available documentation has been examined to determine whether they can detect failure modes that can
cause error propagation. The academic and research
tools that are open source or free have been tested with
a test program to determine their capabilities regarding
the error propagation issues.
These tools were examined in (Sarshar 2007) based
on the available documentation. Table 3 lists the availability and relevance of some of the tools; whether
they might detect the kind/category of failure modes
we have unveiled in our analysis. The relevance of a
tool is divided in the following categories:
Unknown (conclusion based on available documentation)
No (does not detect any of our failure modes)
Some (detects some of our failure modes)
All (detects all of our failure modes)
As expected, we found no tools that could detect all
of our failure mode categories.

3 Available online: http://www.spinroot.com/static/,

http://www.testingfaqs.org/t-static.html, and
http://en.wikipedia.org/wiki/List_of_tools_for_static_
code_analysis.

Line numbers
Line numbers
Line numbers

Existing analysis tools for C.

Tool

Rel.

Checks

Coverity

Some

Klockwork
PolySpace
Purify
Flexelint

Some
No
Some
Some

LintPlus
CodeSonar
Safer C toolset

Unkn.
Some
Some

DoubleCheck
Sotoarc
Astree
Mygcc
Splint (LC-Lint)

Some
No
Some
Some
Some

RATS
Sparce

Unkn.
No

Passing of arguments
and return values
Ignored return values

Passing of arguments
Passing of arguments
and return values

Ignored return values

Passing of arguments
and sequential issues
Sequential issues

User-defined checks
Passing of arguments,
ignored return values
and sequential issues

ASSESSMENT OF SOURCE CODE

AND PROTOTYPE TOOL

This section describes development of a prototype

tool aimed at evaluating to what extent the process
of detecting the characteristics of the identified failure modes can be automated. The aim is to identify the
challenges involved in automating such a code analysis process and not to develop a complete tool capable
of detecting all failure modes.
For this prototype we chose to use bash scripting,
which makes use of the available utilities provided by
the Linux operating system.

186

Figure 1 illustrates the logic and architecture of the

tool where the arrows indicate control flow and the
curved-arrows indicate reporting a message to the user
without interfering with the normal control flow. The
algorithm for analyzing the code is shaded.
The prototype tool is designed in modules to allow
easy updates. The tool it self performs the preliminary part, analysis part and the report part, but the
checks that are performed to determine each failure
mode are placed in external modules. This will ease
the work of adding additional failure modes and system call checks. When a specific parameter is to be
checked, the prototype tool identifies its type and
value and sends this information to the check-module
that holds the check for that specific system call. The
module then performs all available checks for the specified parameter and reports the results of the analysis.
The return value is checked similarly. We have also
included a general module which contains checks that
apply to all system calls. This is necessary to avoid
duplication of modules containing the same checks.
The tool performs the following steps:

Compile the target source code (using gcc), exit

if any error(s)
Create a control flow graph (CFG) of the code
(using gcc -fdump-tree-fixcfg)
2. Analysis part

Identify usage of parameters and determine

their variable type and value by backtracking, then send data for each parameter to
the check-module relevant for the system call
being analyzed.
Check whether return variable is used, determine its type and name and send this data
to both the general check-module, and to
the check-module for the system call being
analyzed.
Store line number and system call info for
return value, and parameters for sequential
analysis.
Perform sequential analysis

1. Preliminary part:

Figure 1.

Search for system call usage in the CFG (using

the standard C library invoke method)
For each system call:

The logic of the prototype tool.

187

3. Report part
This part is integrated with the analysis part in
this prototype. Printed messages are in different
color depending on their contents. Warnings are
printed in orange and errors are printed in red.
As a test program, consider the code of shm.c in
Listing 1.

C. In line 22: permission mode not set for user when

passing argument shmflg (failure mode F.29.3.D)
D. In line 30: no use of the return value from shmdt()
(failure mode F.67.2.A)
E. In line 30: passing of argument workaddr which is
not attached (failure mode F.shm.A)
F. In line 37 and 41: segment referred to by workaddr
is not released prior to return (failure mode
F.shm.D)

Listing 1: The code of shm.c

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

|gcc fdump tree fixupcfgshm.c

An excerpt of the output is illustrated in Listing 2
(empty lines are removed). The notation is in GIMPLE4 which is an intermediate representation of the
program in which complex expressions are split into
a three address code using temporary variables. GIMPLE retains much of the structure of the parse trees:
lexical scopes are represented as containers, rather
than markers. However, expressions are broken down
into a tree-address form, using temporary variables to
hold intermediate values. Also, control structures are
lowered to gotos. Figure 2 illustrates the nodes of this
control flow graphically. The three nodes <bb 0>,
<L0> and <L1> are marked in the figure.

/ / C re a t e a new s h a re d memory s eg m e n t
i f ( ( shmid = s h m g e t ( key , s i z e , s h m f l g ) )
==
1) {
p e r r o r ( s hm g e t f a i l e d ) ;
re t u r n 1 ;
} else {
( vo i d ) f p r i n t f ( s t d o u t , s h m g e t
r e t u r n e d %d\ n , shmid ) ;
}

23
24
25
26
27
28
29

Listing 2: Excerpt of the control flow graph for shm.c

generated by gcc
<bb 0> :
key = 1 0 0 ;
shmflg = 512;
s i z e = 1024;
shmaddr = 0B ;
size .0 = ( s i z e t ) size ;
D. 2 0 3 8 = s h mg e t ( key , s i z e . 0 , s h m f l g ) ;
shmid = D. 2 0 3 8 ;
i f ( shmid ==
1) g o t o < L0> ; e l s e g o t o <
L1> ;
< L0 > : ;
p e r r o r (& s h m g e t f a i l e d [ 0 ] ) ;
D. 2 0 3 9 = 1 ;
go t o < bb 5> (< L4> ) ;
< L1 > : ;
stdout .1 = stdout ;
f p r i n t f ( s t d o u t . 1 , & s h m g e t r e t u r n e d %d
\ n [ 0 ] , shmid ) ;
D. 2 0 4 1 = shmdt ( wo r k a d d r ) ;
r e t = D. 2 0 4 1 ;
p e r r o r (& shmdt : [ 0 ] ) ;
D. 2 0 4 2 = s h m a t ( shmid , shmaddr , s i z e ) ;
wo r k a d d r = ( c o n s t char * ) D. 2 0 4 2 ;
i f ( wo r k a d d r == 4294967295B) go t o < L2> ;
e l s e go t o < L3> ;

/ / Make t h e d e t a c h c a l l and r e p o r t t h e
results
r e t = shmdt ( wo r k a d d r ) ;
p e r r o r ( shmdt ) ;

30
31
32
33

/ / Make t h e a t t a c h c a l l and r e p o r t t h e
results
wo r k a d d r = s h m a t ( shmid , shmaddr , s i z e ) ;
i f ( wo r k a d d r == ( char * ) ( 1) ) {
p e r r o r ( shmat f a i l e d ) ;
re t u r n 1 ;
} else {
( vo i d ) f p r i n t f ( s t d o u t , s h m a t
returned succesfully ) ;
}
re t u r n 0 ;

34
35
36
37
38
39
40
41
42

The control flow graph of this code is created using:

# i n c l u d e < s t d i o . h>
# i n c l u d e < s y s / t y p e s . h>
# i n c l u d e < s y s / i p c . h>
# i n c l u d e < s y s / shm . h>
e x t e r n vo i d p e r r o r ( ) ;
i n t main ( ) {
/ / k e y t o be p a s s e d t o s h m g e t
i n t key = 1 0 0 ;
/ / s h m f l g t o be p a s s e d t o s h m g e t
i n t shmflg = 00001000;
/ / r e t u r n v a l u e f ro m s h m ge t
i n t shmid ;
/ / s i z e t o be p a s s e d t o s h m ge t
i n t s i z e = 1024;
/ / shmaddr t o be p a s s e d t o s hm a t
char * shmaddr = 0 0 0 0 0 0 0 0 ;
/ / r e t u r n e d w o rk i n g a d d re s s
c o n s t char * wo r k a d d r ;
int ret ;

The test program contains selected failure modes,

and we will use the program to show how the prototype tool can detect failure modes. The injected failure
modes are:
A. In line 22: passing of argument key with type
different than key_t (failure mode F.29.1.F)
B. In line 22: passing of argument size with type
different than size_t (failure mode F.29.2.F)

4 This

representation was inspired by the SIMPLE representation proposed in the McCAT compiler project
at McGill University for simplifying the analysis and
optimization of imperative programs.

188

Figure 2.

Control flow graph of shm.c.

The tool then performs the described analysis.

Figure 3 illustrates the output of running the prototype
tool with shm.c as argument.
The tool identified failure modes related to the
shmget() system call and failure modes related to
sequential issues. Check modules for the shmat()
and shmdt() functions were not implemented and a
warning was therefore given by the tool.
Furthermore, the tool identified all of the injected
failure modes. It must, however, be validated through
several more examples before final conclusions can
be made to determine its ability to identify failure
modes. This is left to future work. The intention in
this paper was to demonstrate that it was possible to
detect these failure modes when analyzing source code
with an automated tool.
Figure 3.

Running the prototype tool.

DISCUSSION

In the examination of the system call interface we identified failure modes that could cause error propagation.
For each failure mode, a check was described to determine its presence in code. The focus was on detection
of these when analyzing a given source code. The analysis process was then to be automated using static
analysis. Furthermore, we created a prototype tool
that managed to detect these when analyzing source
code.
Static analysis tools look for a fixed set of patterns, or compliance/non-compliance with respect to
rules, in the code. Although more advanced tools allow
new rules to be added over time, the tool will never
detect a particular problem if a suitable rule has not
been written. In addition, the output of a static analysis tool still requires human evaluation. It is difficult
for a tool to know exactly which problems are more

or less important automatically. However, a tool can

determine presence of known errors, faults or failure
mode and can determine whether for instance error
propagation is a concern for a given program.
Another limitation using static analysis in our
approach concerns whether we can check the passed
arguments. In general, we can only control and check
the argument values that are statically set in the source
code. Dynamic values that are received from elsewhere, e.g. files or terminal input, can not be checked
statically. For such cases, dynamic analysis and testing
must be applied to the program. However, static analysis can detect whether the variables are of correct
type, and within legal range before they are passed
to a library function. To reduce such problems in runtime, one can develop wrapper functions which control
parameters against the identified failure modes before
these are sent to a system call. Such wrapper functions

189

will provide safety checks for programs that use

dynamic input as parameters for system calls.
We developed a conceptual model for a static analysis tool which we implemented as a prototype. The
tool managed to detect many of the failure modes
causing error propagation automatically, but not all.
It is difficult, if not impossible, to control and check
dynamic variables that are passed to system services
when performing static analysis.
The implemented prototype tool is not restricted to
only check for failure modes that can cause error propagation. In our analysis of system calls we identified
numerous failure modes which could cause other types
of failures than error propagation. Using the same
principles as for automatic identification of error propagation related failure modes, it is possible to extend
the tool to also identify these other types of failure
modes. Thus, the tool can determine whether error
propagation is a concern when using system calls; the
tool can indicate which part of the software code is
erroneous, and can pin-point the parts of the software
that should be focused on during testing. One should
also test the component(s) this process affects; the
tool can be used by driver and software developers
which make use of system calls, to verify that these
calls are used correctly (in regards to the documentation); and the tool can be used to identify failure modes
which needs special attention and exception handling
routines.

REFERENCES
Fredriksen R. & Winther R. 2006. Error PropagationPrinciples and Methods. Halden Internal
Report, Norway, HWR-775, OECD Halden Reactor Project.
Fredriksen R. & Winther R. 2007. Challenges Related
to Error Propagation in Software Systems. in Safety
and Reliability Conference, Stavanger, Norway,
June 2527. Taylor & Francis, 2007, pp. 8390.
Johansson A., Suri N. & Murphy B. 2007. On the
Impact of Injection Triggers for OS Robustness
Evaluation. International in Proceedings of the 18th
IEEE International Symposium on Software Reliability Engineering (ISSRE 2007). Npvember, 2007,
pp. 127136.
Koenig A. 1988. C Traps and Pitfalls. Reading, Mass.,
Addison-Wesley.
Sarshar S. 2007. Analysing Error Propagation
between Software Processes in Source Code. Masters thesis, Norway, stfold University College.
Sarshar S., Simensen J.E., Winther R. & Fredriksen R.
2007. Analysis of Error Propagation Mechanisms
between Software Processes. in Safety and Reliability Conference, Stavanger, Norway, June 2527.
Taylor & Francis, 2007, pp. 9198.

190

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Development of a prognostic tool to perform reliability analysis

Mohamed El-Koujok, Rafael Gouriveau & Noureddine Zerhouni
FEMTO-ST Institute, UMR CNRS 6174UFC/ENSMM/UTBM Automatic Control
and Micro-Mechatronic Systems Department, Besanon, France

ABSTRACT: In maintenance field, prognostic is recognized as a key feature as the estimation of the remaining
useful life of an equipment allows avoiding inopportune maintenance spending. However, it can be difficult to
define and implement an adequate and efficient prognostic tool that includes the inherent uncertainty of the
prognostic process. Within this frame, neuro-fuzzy systems are well suited for practical problems where it
is easier to gather data (online) than to formalize the behavior of the system being studied. In this context,
and according to real implementation restrictions, the paper deals with the definition of an evolutionary fuzzy
prognostic system for which any assumption on its structure is necessary. The proposed approach outperform
classical models and is well fitted to perform a priori reliability analysis and thereby optimize maintenance
policies. An illustration of its performances is given by making a comparative study with an other neuro-fuzzy
system that emerges from literature.

INTRODUCTION

The growth of reliability, availability and safety of a

system is a determining factor in regard with the effectiveness of industrial performance. As a consequence,
the high costs in maintaining complex equipments
make necessary to enhance maintenance support systems and traditional concepts like preventive and
corrective strategies are progressively completed by
new ones like predictive and proactive maintenance
(Muller et al. 2008; Iung et al. 2003). Thereby, prognostic is considered as a key feature in maintenance
strategies as the estimation of the provisional reliability of an equipment as well as its remaining useful life
allows avoiding inopportune spending.
From the research point of view, many developments exist to support the prognostic activity
(Byington et al. 2002; Jardine et al. 2006; Vachtsevanos et al. 2006). However, in practice, choosing an
efficient technique depends on classical constraints
that limit the applicability of the tools: available
data-knowledge-experiences, dynamic and complexity of the system, implementation requirements (precision, computation time, etc.), available monitoring
devices. . . Moreover, implementing an adequate tool
can be a non trivial task as it can be difficult to provide
effective models of dynamic systems including the
inherent uncertainty of prognostic. That said, developments of this paper are founded on the following
two complementary assumptions. 1) On one hand,
real systems increase in complexity and their behavior
is often non-linear, which makes harder a modeling

step, even impossible. Intelligent Maintenance Systems must however take it into account. 2) On the
other hand, in many cases, it is not too costly to equip
dynamic systems with sensors, which allows gathering real data online. Furthermore, monitoring systems
evolve in this way.
According to all this, neuro-fuzzy (NF) systems
appear to be very promising prognostic tools: NFs
learn from examples and attempt to capture the subtle
relationship among the data. Thereby, NFs are well
suited for practical problems, where it is easier to
gather data than to formalize the behavior of the system being studied. Actual developments confirm the
interest of using NFs in forecasting applications (Wang
et al. 2004; Yam et al. 2001; Zhang et al. 1998). In
this context, the paper deals with the definition of an
evolutionary fuzzy prognostic system for which any
assumption on its structure is necessary. This model
is well adapted to perform a priori reliability analysis
and thereby optimize maintenance policies.
The paper is organized in three main parts. In
the first part, prognostic is briefly defined and positioned within the maintenance strategies, and the
relationship between prognostic, prediction and online
reliability is explained. Following that, the use of
Takagi-Sugeno neuro-fuzzy systems in prognostic
applications is justified and the ways of building such
models are discussed. Thereby, a NF model for prognostic is proposed. In the third part, an illustration
of its performances is given by making a comparative study with an other NF system that emerges from
literature.

191

2
2.1

PROGNOSTIC AND RELIABILITY

From maintenance to prognostic

Maintenance activity combines different methods,

tools and techniques to reduce maintenance costs
while increasing reliability, availability and security
of equipments. Thus, one usually speaks about fault
detection, failures diagnosis, and response development (choice and scheduling of preventive and/or corrective actions). Briefly, these steps correspond to the
need, firstly, of perceiving phenomena, secondly,
of understanding them, and finally, of acting
consequently. However, rather than understanding a
phenomenon which has just appeared like a failure
(a posteriori comprehension), it seems convenient to
anticipate its manifestation in order to take adequate actions as soon as possible. This is what could
be defined as the prognostic process and which is
the object of this paper. Prognostic reveals to be a very
promising maintenance activity and industrials show
a growing interest in this thematic which becomes a
major research framework; see recent papers dedicated to condition-based maintenance (CBM) (Jardine
et al. 2006; Ciarapica and Giacchetta 2006). The relative positioning of detection, diagnosis, prognostic
and decision/scheduling can be schematized as proposed in Fig. 1. In practice, prognostic is used to be
performed after a detection step: the monitoring system detects that the equipment overpass an alarm limit
which activates the prognostic process.
2.2

From prognostic to prediction

Although there are some divergences in literature,

prognostic can be defined as proposed by the International Organization for Standardization: prognostic
is the estimation of time to failure and risk for one or
more existing and future failure modes (ISO 13381-1
2004). In this acceptation, prognostic is also called
the prediction of a systems lifetime as it is a process whose objective is to predict the remaining useful
life (RUL) before a failure occurs given the current
machine condition and past operation profile (Jardine
et al. 2006). Thereby, two salient characteristics of
prognostic appear:

Figure 1.

Prognostic within maintenance activity.

prognostic is mostly assimilated to a prediction

process (a future situation must be caught),
prognostic is based on the failure notion, which
implies a degree of acceptability.
A central problem can be pointed out from this:
the accuracy of a prognostic system is related to its
ability to approximate and predict the degradation of
equipment. In other words, starting from a current situation, a prognostic tool must be able to forecast the
future possible situations and the prediction phase
is thereby a critical one. Next section of this paper
emphasizes on this step of prognostic.
2.3 From prediction to reliability
As mentioned earlier, an important task of prognostic
is to predict the degradation of equipment. Following that, prognostic can also be seen as a process that
allows the a priori reliability modeling.
Reliability (R(t)) is defined as the probability that
a failure does not occur before time t. If the random
variable denotes the time to failure with a cumulative
distribution function F (t) = Prob( t), then:
R(t) = 1 F (t)

(1)

Let assume now that the failure is not characterized

by a random variable but by the fact that a degradation signal (y) overpass a degradation limit (ylim ), and
that this degradation signal can be predicted (y) with a
degree of uncertainty (Fig. 2). At any time t, the failure
probability can be predicted as follows:


F(t) = Pr y (t) ylim
(2)
Let note g(y/t) the probability distribution function
that denotes the prediction at time t. Thereby, by analogy with reliability theory, the reliability modeling can

Figure 2.

192

Prediction and reliability modeling.

be expressed as follows:

R(t) = 1 Pr[y(t) ylim ] = 1

g(y/t) dy

ylim

(3)
The remaining useful life (RUL) of the system can
finally be expressed as the remaining time between
the time in which is made the prediction (tp) and the
time to underpass a reliability limit (Rlim ) fixed by the
practitioner (see Fig. 2).
These explanations can be generalized with a multidimensional degradation signal. See (Chinnam and
Pundarikaksha 2004) or (Wang and Coit 2004) for
more details. Finally, the a priori reliability analysis can be performed if an accurate prognostic tool
is used to approximate an predict the degradation of
an equipment. This is the purpose of next sections of
this paper.

3
3.1

FUZZY MODELS FOR PREDICTION

Takagi-Sugeno system: A fitted prediction tool

Various prognostic approaches have been developed

ranging in fidelity from simple historical failure rate
models to high-fidelity physics-based models (Vachtsevanos et al. 2006: Byington et al. 2002). Similarly to
diagnosis, these methods can be associated with one
of the following two approaches, namely model-based
and data-driven. That said, the aim of this part is not
to dress an exhaustive overview of prediction techniques but to explain the orientations of works that are
taken.
Real systems are complex and their behavior is
often non linear, non stationary. These considerations
make harder a modeling step, even impossible. Yet, a
prediction computational tool must deal with it. Moreover, monitoring systems have evolve and it is now
quite esay to online gather data. According to all this,
data-driven approaches have been increasingly applied
to machine prognostic. More precisely, works have
been led to develop systems that can perform nonlinear modeling without a priori knowledge, and that are
able to learn complex relationships among inputs and
outputs (universal approximators). Indeed, artificial
neural networks (ANNs) have been used to support the
prediction process (Zhang et al. 1998), and research
works emphasize on the interest of using it. Nevertheless, some authors remain skeptical as ANNs are
black-boxes which imply that there is no explicit
form to explain and analyze the relationships between
inputs and outputs. According to these considerations, recent works focus on the interest of hybrid
systems: many investigations aim at overcoming the

major ANNs drawback (lack of knowledge explanation) while preserving their learning capability. In this
way, neuro-fuzzy systems are well adapted. More precisely, first order Tagaki-Sugeno (TS) fuzzy models
have shown improved performances over ANNs and
conventional approaches (Wang et al. 2004). Thereby,
they can perform the degradation modeling step of
prognostic.
3.2 Takagi-Sugeno models: Principles
a) The inference principle
A first order TS model provides an efficient and
computationally attractive solution to approximate a
nonlinear input-output transfer function. TS is based
on the fuzzy decomposition of the input space. For
each part of the state space, a fuzzy rule can be constructed to make a linear approximation of the input.
The global output approximation is a combination
of the whole rules: a TS model can be seen as a
multi-model structure consisting of linear models that
are not necessarily independent (Angelov and Filev
2004).
Consider Fig. 3 to explain the first order TS model.
In this illustration, two inputs variables are considered,
two fuzzy membership functions (antecedent fuzzy
sets) are assigned to each one of them, and the TS
model is finally composed of two fuzzy rules. That
said, a TS model can be generalized to the case of n
inputs and N rules (see here after).
The rules perform a linear approximation of inputs
as follows:
Ri : if x1 is A1i and . . . and xn is Ani
THEN yi = ai0 + ai0 x1 + + ain xn

(4)

where Ri is the ith fuzzy rule, N is the number of

fuzzy rules, X = [x1 , x2 , . . . , xn ]T is the input vector,
j
Ai denotes the antecedent fuzzy sets, j = [1, n], yi is
the output of the ith linear subsystem, and aiq are its
parameters, q = [0, n].
Let assume Gaussian antecedent fuzzy sets (this
choice is justified by its generalization capabilities and
because it covers the whole domain of the variables)

Figure 3.

193

First order TS model.

to define the regions of fuzzy rules in which the local

linear sub-models are valid:

defines himself the architecture of the model and the

antecedents parameters values (Espinosa et al. 2004).

ij = exp[4xx

Gradient Descent (GD). The principle of the GD

algorithm is to calculate the premise parameters by
the standard back-propagation algorithm. GD has
been implemented in a special neuro-fuzzy system:
the ANFIS model (Adaptive Neuro-Fuzzy Inference
System) proposed by (Jang and Sun 1995).

i 

i 2
j ]/[(j ) ]

(5)

where (ji )2 is the spread of the membership function, and xi is the focal point (center) of the ith rule
antecedent.
The firing level of each rule can be obtained by the
product fuzzy T-norm:
i = i1 (x1 ) in (xn )

(6)

The normalized firing level of the ith rule is:

i = i

N


(7)

j=1

The TS model output is calculated by weighted

averaging of individual rules contributions:

y=

N

i=1

i yi =

N


i xeT i

(8)

i=1

where i = [ai0 ai1 ai2 . . . ain ] is the vector parameter

of the ith sub-model, and xe = [1 X T ]T is the expanded
data vector.
A TS model has two types of parameters. The
non-linear parameters are those of the membership
functions (a Gaussian membership like in equation 5
has two parameters: its center x and its spread deviation ). This kind of parameter are referred to as
premise or antecedent parameters. The second type of
parameters are the linear ones that form the consequent
part of each rule (aiq in equation 4).
b) Identification of TS fuzzy models
Assuming that a TS model can approximate an inputoutput function (previous section), in practice, this
kind of model must be tuned to fit to the studied
problem. This implies two task to be performed:
the design of the structure (number and type of
membership functions, number of rules),
the optimization of the models parameters.
For that purpose, different approaches can be used
to identify a TS model. In all cases, the consequent
parameters of the system are tuned by using a least
squares approach.
Mosaic or table lookup scheme. It is the simplest
method to construct TS fuzzy system as the user

Genetic Algorithms (GAs).

GAs are well known
for their optimization capabilities. The GAs are used
by coding the problem into chromosomes and setting
up a fitness function. Since the consequent part of a
TS model can be calculated by using a least squares
method, only the premise part of the model is coded
into chromosomes and optimized by the GAs.
Clustering Methods (CMs). The basic idea behind
fuzzy clustering is to divide a set of objects into selfsimilar groups (cluster). The main interest of this type
of methods is that the user does not need to define the
number of membership functions, neither the number
of rules: CMs adapt the structure of the TS model by
the learning phase.
Evolving algorithms. These algorithms are based on
CMs and therefore, do not require the user to define the
structure of the TS model. In opposition to all previous
approaches, they do not need a complete learning data
set to start the identification process of the TS model
(start from scratch): they are online algorithms with
self constructing structure. These approaches were
recently introduced (Angelov and Filev 2003; Kasabov
and Song 2002).
3.3 Discussion: exTS for prognostic application
The selection of an identification approach for TS
model depends obviously on the prediction context.
According to the degradation modeling problem, a
prediction technique for prognostic purpose should not
be tuned by an expert as it can be too difficult to catch
the behavior of the monitored equipment. Thereby, the
first approach for identification (table lookup scheme)
should be leaved aside.
Descent gradient and genetic algorithms approaches
allow updating parameters by a learning process but
are based on a fixed structure of the model, which supposes that an expert is able to indicate the adequate
architecture to be chosen. However, the accuracy of
predictions is fully dependent on this, and such identification techniques suffer from the same problems
as ANNs. Yet, the ANFIS model is known as a fitted tool for time-series prediction and has been used
for prognostic purpose (Goebel and Bonissone 2005;
Wang et al. 2004).
In opposition, clustering approaches require less
a priori structure information as they automatically

194

determine the number of membership functions and of

rules. However, in practical applications, the learning
process is effective only if sufficient data are available. In addition to it, when trained, such a TS model is
fixed. Thereby, if the behavior of the monitored system
changes significantly (like in a degradation phase),
predictions can suffer from the lack of representative
learning data.
Considering the applicative restrictions that supposes the implementation of a prognostic tool, evolving TS models appear to be the more promising for
prognostic applications. Firstly, they are able to update
the parameters without the intervention of an expert
(evolving systems with regard to the parameters). Secondly, they can be trained in online mode as they
have a flexible structure that evolves with the data
gathered from the system: data are collected continuously which enables to form new rules or to modify an
existing one. This second characteristics is very useful to take into account the non-stationary aspect of
degradation.
According to all this, an accurate TS prediction
technique for online reliability modeling is the evolving one. A particular model is this one proposed by
(Angelov and Zhou 2006): the evolving eXtended
Takagi-Sugeno system (exTS). The way of learning
this type of model is presented in next section and the
interest of using it is illustrated in section 4.2.

3.4

Step 1. Starting from k = 2, the potential Pk of the

data point zk is recursively calculated at time k:
Pk (zk ) =

k 1+

k 1
j=1 i=1
n+m

k1

zi zk 2

(9)

Step 2. The potential of the cluster/rule centers is

recursively updated:
Pk (z ) =

k 2 + Pk

(k 1)Pk1 (z )
n+m
2

k (z )
j=1 z zk1 j
(10)

(z ) + P

Step 3. The potential of the data point (step 1) is

compared to boundaries issued from the potential of
the cluster centers (step 2):
(P Pk (zk ) P)

(11)

where (P = maxNi=1 {Pi (z )}) is the highest density/potential, (P = minNi=1 {Pi (z )}) is the lowest
density/potential and N is number of centers clusters
(xi , i = [1, N ]) formed at time k.
Step 4. If, the new data point has a potential in
between the boundaries (11) any modification of the
rules is necessary. Else, they are two possibilities:

Learning procedure of exTS

The learning procedure of exTS is composed of two

phases:

1. if the new data point is closed to an old center

i

Phase A: an unsupervised data clustering technique

is used to adjust the antecedent parameters,
Phase B: the supervised Recursive least squares
(RLS) learning method is used to update the consequent parameters.
a) Clustering phase: Partitioning data space
The exTS clustering phase processes on the global
input-output data space: z = [xT , yT ]T ; z Rn+m ,
n + m defines the dimensionality of the input/output
data space. Each one of the sub-model of exTS operates in a sub-area of z. This TS model is based on
the calculus of a potential (see after) which is the
capability of a data to form a cluster (antecedent of
a rule).
The clustering procedure starts from scratch assuming that the first data point available is a center of a
cluster: the coordinates of the first cluster center are
those of the first data point (z1 z1 ). The potential of the first data point is set to the ideal value:
P1 (z1 ) 1. Four steps are then performed for each
new data gathered in real-time.

(minNi xk xi j < 2j ), then the new data point

(zk ) replaces this center (zj := zk ),
2. else, the new data point is added as a new center
and a new rule is formed (N = N + 1; xN ).
Note that, the exTS learning algorithm presents an
adaptive calculation of the radius of the clusters (ji ).
See (Angelov and Zhou 2006) for more details.
b) RLS phase: update of the consequent parameters
The exTS model is used for on-line prediction. In this
case, equation (8) can be expressed as follows:

y k+1 =

N

i=1

i yi =

N


i xeT i = kT k

(12)

i=1

k = [1 xeT , 2 xeT , . . . , n xeT ]Tk is a vector of the

inputs, weighted by normalized firing () of the
rules, k = [1T , 2T , . . . , NT ]Tk are parameters of the
sub-models.

195

The following RLS procedure is applied:

k = k1 + Ck k (yk kT k1 );

k = 2, 3, . . .
(13)

Ck = Ck1

Ck1 k kT Ck1
1 + kT Ck1 k

(14)

with initial conditions

1 = [ 1T , 2T , . . . , NT ] = 0,

C1 = I

(15)

where is a large positive number, C1 is a R(n + 1)

R(n + 1) co-variance matrix, and k is an estimation
of the parameters based on k data samples.

4
4.1

COMPARATIVE STUDY OF ANFIS

AND EXTS
exTS versus ANFIS

To illustrate the performances of exTS, this model

is compared to the ANFIS model (Adaptive NeuroFuzzy Inference System) proposed by (Jang and Sun
1995) that emerges from literature: ANFIS shown
improved performance over conventional method and
Wang demonstrates that it is a robust machine health
condition predictor (Wang et al. 2004).
The whole learning algorithm of ANFIS can not be
fully presented here (see referenced authors for more
explanation). In a few words, ANFIS uses an hybrid
algorithm which is the combination of the gradient
descent (that enables the updating of the antecedent
parameters) and of the least squares estimate (that
optimizes the consequent parameters).
A specificity of ANFIS can be pointed out: ANFIS
is fully connected which implies that if M membership functions are assigned to each one of the
n inputs variables, then the ANFIS is composed of
N = M n rules (see Fig. 4 for an example). As
a consequence, many parameters must be updated
but, when well trained, ANFIS may perform good
predictions.
4.2 Experimental data sets
Two real experimental data sets have been used to
test the prediction performances of ANFIS and exTS.
In both cases, the aim of these predictions is to
approximate a physical phenomenon by learning data
gathered from the system. That can be assimilated to
the prediction step of the prognostic process.
Industrial dryer data set. The first data set is issued
from an industrial dryer. It has been contributed by

Figure 4.

Architecture of an ANFIS with 4 inputs.

Jan Maciejowski1 from Cambridge University. This

data set contains 876 samples. Three variables (fuel
flow rate, hot gas exhaust fan speed, rate of flow of
raw material) are linked with an output one (dry bulb
temperature).
For simulations, both ANFIS and exTS have been
used with five inputs variables. Predictions were made
at different horizons h. Assuming that t denotes the
current time, the TS models were build as follows:

input 1: x1 (t)fuel flow rate,

input 2: x2 (t)hot gas exhaust fan speed,
input 3: x3 (t)rate of flow of raw material,
input 4: x4 (t)dry bulb temperature,
input 5: x5 (t 1)dry bulb temperature,
output 1: y (t + h)predicted dry bulb temperature.

Air temperature in a mechanical system. The second data set is issued from an hair dryer. It has been
contributed by W. Favoreel2 from the KULeuven University. This data set contains 1000 samples. The air
temperature of the dryer is linked to the voltage of the
heating device.
For simulations, both ANFIS and exTS have been
used with five inputs variables. Predictions concern
the air temperature, and the TS models were build as
follows:
input 1 to 4: air temperature at times (t 3) to (t),
input 5: x5 (t)voltage of the heating device,
output 1: y (t + h)predicted air temperature.
4.3 Simulations and results
In order to extract more solid conclusions from the
comparison results, the same training and testing data

1 ftp://ftp.esat.kuleuven.ac.be/sista/data/process_industry
2 ftp://ftp.esat.kuleuven.ac.be/sista/data/mechanical

196

Table 1.

Simulation results.
ANFIS

exTS

Industrial dryer
t+1
Rules
RMSE
MASE
t+5
Rules
RMSE
MASE
t + 10
Rules
RMSE
MASE

32
0.12944
16.0558
32
0.84404
114.524
32
1.8850
260.140

18
0.01569
2.16361
17
0.05281
7.38258
17
0.18669
27.2177

Air temperature
t+1
Rules
RMSE
MASE
t+5
Rules
RMSE
MASE
t + 10
Rules
RMSE
MASE

32
0.01560
0.4650
32
0.13312
2.01818
32
0.23355
3.66431

4
0.01560
0.47768
6
0.12816
1.97647
6
0.22997
3.66373

sets were used to train and test both models. Predictions were made at (t + 1), (t + 5) and (t + 10)
in order to measure the stability of results in time.
The prediction performance was assessed by using
the root mean square error criterion (RMSE) which
is the most popular prediction error measure, and the
Mean Absolute Scaled Error (MASE) that, according
to (Hyndman and Koehler 2006), is the more adequate
way of comparing prediction accuracies.
For both data sets, the learning phase was stopped
after 500 samples and the reminding data served to test
the models. Results are shown in table 1.

4.4

Figure 5.

Predictionsindustrial dryer, t + 1.

Figure 6.

Pdf errorair temperature, t + 10.

Table 2.

Complexity of the prediction systems.

Structural properties for the air temperature benchmark

at t + 10
Criteria

nb inputs
5
nb rules
32
type of mf
Gaussian
antecedent parameters
mf/input
2
tot. nb of mf
25
parameters/mf
2
ant. parameters
2 2 5 = 20
consequent parameters
parameters/rule
6 (5 inputs +1)
cons. parameters
6 32 = 192
parameters
20 + 192 = 212

Discussion

a) Accuracy of predictions
According to the results of table 1, exTS performs
better predictions than ANFIS model. Indeed, for the
industrial dryer (data set 1), both RMSE and MASE
are minors with exTS than with ANFIS. An illustration
of it is given in Fig. 5.
However, in the case of the air temperature data
set, exTS do not provide higher results than ANFIS
(RMSE and MASE are quite the same). Moreover, as
it is shown in Fig. 6, the error spreadings of both model
are very similar. Yet, one can point out that exTS only
needs 6 fuzzy rules to catch the behavior of the studied
phenomenon (against 32 for the ANFIS model). This
lead us to consider the complexity of the structure of
both prediction systems.

ANFIS

exTS
5
6
Gaussian
= nb rules = 6
66
2
2 6 5 = 60
6
6 6 = 36
60 + 36 = 96

b) Complexity of the prediction systems

Let take the example of the last line of table 1 to compare the structures of the ANFIS and exTS models.
The number of parameters for both systems is detailed
in table 2.
As there are 5 inputs for the Air Temperature application (see 4.2), and assuming Gaussian membership

197

functions for the antecedent fuzzy sets, the ANFIS

model is composed of 212 parameters. Following that,
with a more complex application than that of the
benchmark studied in the paper, an ANFIS system can
be quickly limited by the number of inputs (because
the numbers of parameters to be updated increases).
In addition, classically, one says that the number of
learning samples for the ANFIS model must be more
than five times the numbers of parameters, which can
be critical for industrial practitioners.
In opposition, exTS evolves only if there are significant modifications on the input-output variables
as it has an on-line learning process: exTS start
from scratch with a single rule and modifications or
additions of rules are made only if relevant. As a consequence, for the same prediction purpose, an exTS
system can have the same prediction accuracy that an
ANFIS model but with less rules (6 vs 32 in the case
considered in table 2). This complexity reduction of
the prediction system can also be pointed out by considering the total number of parameters (96 vs 212).
c) Computation efficiency
Finally, although it can not be fully developed in the
paper, exTS is much more computationally effective
than the ANFIS system. This can be explained from
two complementary point of views. Firstly, as stated
before, an exTS system can perform predictions with
a slightly structure that the ANFIS, which implies
that fewer parameters have to be updated. Secondly,
when using an exTS system, all learning algorithms
are recursive ones which allows the on-line use of the
system and ensure the rapidity of treatments.

CONCLUSION

In maintenance field, prognostic is recognized as a

key feature as the estimation of the remaining useful life of an equipment allows avoiding inopportune
maintenance spending. However, it can be difficult to
define and implement an adequate and efficient prognostic tool that includes the inherent uncertainty of the
prognostic process. Indeed, an important task of prognostic is that of prediction. Following that, prognostic
can also be seen as a process that allows the reliability modeling. In this context, the purpose of the work
reported in this paper is to point out an accurate prediction technique to perform the approximation and
prediction of the degradation of an equipment.
According to real implementation restrictions,
neuro-fuzzy systems appear to be well suited for practical problems where it is easier to gather data (online)
than to formalize the behavior of the system being
studied. More precisely, the paper point out the accuracy of the exTS model in prediction. The exTS model
has a high level of adaptation to the environment

and to the changing data. It is thereby an efficient

tool for complex modeling and prediction. Moreover,
any assumption on the structure of exTS is necessary, which is an interesting characteristic for practical
problems in industry. The exTS is finally a promising
tool for reliability modeling in prognostic applications.
Developments are at present extended in order to
characterize the error of prediction at any time and
thereby provide confidence interval to practitioners.
The way of ensuring a confidence level is also studied.
This work is led with the objective of being integrated
to an e-maintenance platform at a French industrial
partner (em@systec).

REFERENCES
Angelov, P. and D. Filev (2003). On-line design of takagisugeno models. Springer-Verlag Berlin Heidelberg: IFSA,
576584.
Angelov, P. and D. Filev (2004). An approach to online identification of takagi-sugeno fuzzy models. IEEE Trans. on
Syst. Man ad Cybern.Part B: Cybernetics 34, 484498.
Angelov, P. and X. Zhou (2006). Evolving fuzzy systems
from data streams in real-time. In Proceedings of the Int.
Symposium on Evolving Fuzzy Systems, UK, pp. 2632.
IEEE Press.
Byington, C., M. Roemer, G. Kacprzynski, and T. Galie
(2002). Prognostic enhancements to diagnostic systems
for improved condition-based maintenance. In 2002 IEEE
Aerospace Conference, Big Sky, USA.
Chinnam, R. and B. Pundarikaksha (2004). A neurofuzzy
approach for estimating mean residual life in conditionbased maintenance systems. Int. J. materials and Product
Technology 20:13, 166179.
Ciarapica, F. and G. Giacchetta (2006). Managing the
condition-based maintenance of a combined-cycle power
plant: an approach using soft computing techniques. Journal of Loss Prevention in the Process Industries 19,
316325.
Espinosa, J., J. Vandewalle, and V. Wertz (2004). Fuzzy
Logic, Identification and Predictive Control (Advances
in Industrial Control). N.Y., Springer-Verlag.
Goebel, K. and P. Bonissone (2005). Prognostic information
fusion for constant load systems. In Proceedings of 7th
annual Conference on Fusion, Volume 2, pp. 12471255.
Hyndman, R. and A. Koehler (2006). Another look at
measures of forecast accuracy. International Journal of
Forecasting 224, 679688.
ISO 13381-1 (2004). Condition monitoring and diagnostics
of machinesprognosticsPart1: General guidelines.
Int. Standard, ISO.
Iung, B., G. Morel, and J.B. Lger (2003). Proactive maintenance strategy for harbour crane operation improvement.
Robotica 21, 313324.
Jang, J. and C. Sun (1995). Neuro-fuzzy modeling and
control. In IEEE Proc., Volume 83, pp. 378406.
Jardine, A., D. Lin, and D. Banjevic (2006). A review
on machinery diagnostics and prognostics implementing
condition-based maintenance. Mech. Syst. and Sign. Proc.
20, 14831510.

198

Kasabov, N. and Q. Song (2002). Denfis: Dynamic evolvinf

neural-fuzzy inference system and its application for timeseries prediction. IEEE Transaction on Fuzzy Systems
102, 144154.
Muller, A., M.C. Suhner, and B. Iung (2008). Formalisation
of a new prognosis model for supporting proactive maintenance implementation on industrial system. Reliability
Engineering and System Safety 93, 234253.
Vachtsevanos, G., F.L. Lewis, M. Roemer, A. Hess, and
B. Wu (2006). Intelligent Fault Diagnosis and Prognosis
for Engineering Systems. New Jersey, Hoboken: Wiley &
Sons.

Wang, P. and D. Coit (2004). Reliability prediction based on

degradation modeling for systems with multiple degradation measures. In Proc. of Reliab. and Maintain. Ann.
Symp.RAMS, pp. 302307.
Wang, W., M.F. Goldnaraghi, and F. Ismail (2004). Prognosis
of machine health condition using neurofuzzy systems.
Mech. Syst. and Sig. Process. 18, 813831.
Zhang, G., B.E. Patuwo, and M.Y. Hu (1998). Forecasting
with artificial neural networks: the state of the art. Int.
Journal of Forecasting 14, 3562.

199

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Fault detection and diagnosis in monitoring a hot dip galvanizing line

using multivariate statistical process control
J.C. Garca-Daz
Applied Statistics, Operations Research and Quality Department,
Polytechnic University of Valencia, Valencia, Spain

ABSTRACT: Fault detection and diagnosis is an important problem in continuous hot dip galvanizing and the
increasingly stringent quality requirements in automotive industry has also demanded ongoing efforts in process
control to make the process more robust. Multivariate monitoring and diagnosis techniques have the power
to detect unusual events while their impact is too small to cause a significant deviation in any single process
variable. Robust methods for outlier detection in process control are a tool for the comprehensive monitoring
of the performance of a manufacturing process. The present paper reports a comparative evaluation of robust
multivariate statistical process control techniques for process fault detection and diagnosis in the zinc-pot section
of hot dip galvanizing line.

INTRODUCTION

A fault is defined as abnormal process behaviour

whether associated with equipment failure, sensor
degradation, set point change or process disturbances.
Fault detection is the task of determining whether a
fault has occurred. Fault diagnosis is the task of determining which fault has occurred. Fault detection and
diagnosis is an important problem in process engineering. The normal state can be associated with
a set of desired trajectories that can be labelled as normal evolutions. According to Himmelblau (1978), the
term fault is generally defined as a departure from an
acceptable range of an observed variable or a calculated parameter associated with a process. A diagnosis
problem is characterised by a set of observations to be
explained. When these observations are different from
the expected behaviour, we defined a faulty state. The
behaviour of a process is monitored through its sensor outputs and actuator inputs. When faults occur,
they change the relationship among these observed
variables. Multivariate monitoring and diagnosis techniques have the power to detect unusual events while
their impact is too small to cause a significant deviation in any single process variable. This is an important
advantage because this trend towards abnormal operation may be the start of a serious failure in the process.
Robust methods for outlier detection in process control are a tool for the comprehensive monitoring of
the performance of a manufacturing process. Outliers
were identified, based on the robust distance. Again,
we remove all detected outliers and repeat the process

until a homogeneous set of observations is obtained.

The final set of data is the in-control set data: the
reference data. The robust estimates of location and
scatter were obtained by the the Minimum Covariance
Determinant (MCD) method of Rousseeuw.
1.1

Industrial process under study

In a continuous hot dip galvanizing line, the steel strip

is coated by passing in through a pot of molten zinc
normally between 450 and 480 C (average bath temperature of 460 C) and with-drawing it from the pot
through a pair of air-wiping jets, to remove the excess
liquid zinc, commonly known as shown schematically in Figure 1. Galvannealing is an in-line process
during which the zinc layer of the steel strip is transformed into ZnFe layer by diffusion. Depending on the
annealing temperature, annealing time, steel grade,
aluminium content in the Zn bath, and other parameters, different intermetallic ZnFe phases are formed
in the coating, which influence the formability of the
galvanized material (Tang 1999).
1.2

Data process

Data for 26 batches were available. Six variables were

selected for monitoring the process: the steel strip
velocity, four bath temperatures and bath level. The
data are arranged in a 26 6 data matrix X. The data
can be analyzed to determine whether or not a fault
has occurred in the process.

201

17

7
6
5
18
4
3
19
2

1
0
0

Figure 1. Galvannealing section of a hot dip galvanizing

line.

Figure 2.

MULTIVARIATE DETECTION: THE ROBUST

DISTANCE

Outlier identification with robust techniques is

obtained by considering two distances for each observation (Verboven and Hubert, 2005). Robust estimators of location and scale for univariate data include
the median, the median absolute deviation, and
M-estimators. In the multivariate location and scatter setting we assume that the data are arranged in a
n p data matrix X consisting of n observations on p
variables. When the number of variables p is smaller
that the data size n, robust estimates of the centre
and the scatter matrix  of X can be obtained by the
Minimum Covariance Determinant (MCD) estimator
proposed by (Rousseeuw and Van Driessen, 1999).
The aim of MCD estimator is to find a subset of data
objects with the smallest determinant of the covariance matrix. This location and covariance estimator
is very popular because of its high resistance towards
outliers. The robust distance of an observation is used
to detect whether it is an outlier or not. This robust
distance is the robust extension of the Mahalanobis
distance. The outliers are those observations having a
robust distance larger that the cutoff-value c, where c2
is the value of the 0.975 quantile of the 2 -distribution
with p degrees of freedom.
Using the robust distances and the Mahalanobis
distances (Verboven and Hubert, 2005) many outlying observations can be identified. In this plot, we
can again identify two groups of outliers. Looking at
Figure 2 we clearly see three outlying observations:
7, 17, 18 and 19. Both a classical and robust analysis
would identify these observations as they exceed the
horizontal and vertical cutoff line.
Observation 17 is away from the data majority,
whereas observations 18 and 19 are outlying to a
smaller extent.

Robust-Mahalanobis distances plot.

FAULT DETECTION AND FAULT

DIAGNOSIS

3.1 Principal components analysis

Under normal operating conditions, sensor measurements are highly correlated. Researchers have used
principal components analysis (PCA) to exploit the
correlation among variables under normal operations conditions (in control) for process monitoring
(Nomikos and MacGregor, 1995) and process fault
diagnosis (Dunia and Qin, 1998a, b). PCA captures
the correlation of historical data when the process is
in control. The violation of the variable correlations
indicates an unusual situation. In PCA, the matrix X
can be decomposed into a score matrix T and a loading
matrix P.
The score plot of the first two score vectors shows
outliers and other strong patters in the data. The score
plot in Figure 3 shows that batch 17 is above the 99%
confidence limit, while the rest of the batches are
below the 99% confidence limit. The batch 17 is abnormal and the rest of the batches are consistent (Nomikos
and MacGregor, 1995).
A T 2 Hotelling and Square Prediction Error (SPE)
control charts for reference data process are carried
out for nonitoring de multivariate process. The SPE is
a statistic that measures lack of fit of a model to data
(Jackson, 1991).
3.2 Fault detection
Normal operations can be characterized by employing
Hotellings T 2 statistic. A value for the T 2 statistic greater than the threshold given by control limit
indicates that a fault has occurred. Figure 4 shows
two complementary multivariate control charts for
batch 17. The first is a Hotellings T 2 chart on the

202

25

Contribution

95 %

15
10

90 %

t[2]

Var1

Var2

Var3

Var4

Var5

Var6

99 % limit

20

2
1
0
-1
-2

Batch Time Repeated For Each Variable

0
Integrated Contribution

-5
-10
- 15
- 20

17

70
60
50
40
30
20
10
0

Var6

Var5
Var1

Var3

Var2

Var4

-25
- 30

-20

- 10

10

20

30

t[1]

Figure 3. Batch unfolding PCA score plot for all variables

and all batches with 90%, 95% and 99% confidence limits.

Figure 5.
batch 17.

Batch unfolding PCA SPE statistic chart for

#7

12

Hotelling T 2 - statistic
20

8
95 %

14

99%

15
99%
10

10

0
95 %

6
95 %

0
0

20

40

60
# 18

12
2

20

40
Time

60

20

40

60
# 18

16
12

20

40
Time

60

8
4

Figure 4. On-line hotelling T 2 -statistic and SPE control

charts with 95% and 99% control limits for batch 17.

4
0

0
0

20

40

first two components from a PCA fit model of the

process. This chart checks if a new observation vector
of measurements on six process variables projects on
the principal component plane within the limits determined by the reference data. The second chart is a
SPE chart. The SPE chart will detect the occurrence
of any new events that cause the process to more away
from the plane defined by the reference model with
two principal components. The limits for a good operation on the control charts are defined based on the
reference good set (training data). If the points remain
below the limits the process is in control. The batch
17 is out of control.
Fault diagnosis

Once a fault has been detected, one can calculate

the contribution to the SPE-statistic for each variable,
which can be used to help classify the cause of the
faults. Contribution plots are used for fault diagnosis.
T 2 -statistic and SPE control charts produce an outof-control signal when a fault occurs, but they do not
provide any information about the fault and its cause.

60
# 19

12

3.3

#7

16
12

SPE

18

99 %

20

40

60
# 19

16
12

8
4

0
0

20

40

60

20

40

60

Figure 6. On-line Hotelling T 2 -statistic and SPE control

charts with 95% and 99% control limits for (a) batch 7,
(b) batch 18, and (c) batch 19. The symbol () show a signal
of possible fault.

Figure 5 shows the SPE statistic contribution chart for

the batch 17. Variables 5 and 6 (bath temperatures)
seem to be the major contributors in the SPE.
Figure 6 shows the T 2 -statistic and SPE control charts for batches 7, 18 and 19 with 95% and
99% control limits. These charts show that the PCA
model adequately describes the reference database:
the process is assumed to be out-of-control if three
consecutive points are out of the 99% control limit
(no batch is above the 99% control limit). The SPE

203

advantage because this trend towards abnormal operation may be the start of a serious failure in the
process. This paper show the performance monitoring potential of MSPC and the predictive capability
of robust statistical control by application to an industrial process. The paper has provided an overview of
an industrial application of multivariate statistical process control based performance monitoring through
the robust analysis techniques.
Outliers were identified, based on the robust distance.
Again, we remove all detected outliers and repeat
the process until a homogeneous set of observations
is obtained. The final set of data is the in-control set
data or reference data. The robust estimates of location and scatter were obtained by the MCD method of
Rousseeuw.
A T 2 Hotelling and SPE control charts for reference
data process are carried out for nonitoring de multivariate process. Contributions from each fault detected
using a PCA model are used for fault identification
approach to identify the variables contributing most to
abnormal situation.

Contribution Plot to the Overall SPE

70

batch # 7

30
Var4
0
-10

Var1

Var2

Var5
Var6

Var3

Contribution

50

Var2
0

Var1

Var3

Var4

Var3

Var4

Var5

Var6

-50

batch #18
-100

40

20

batch #19
Var1

Var2

Var6

0
-10

Var5

Figure 7. Contribution plots of the variables contributing to

the SPE for batch 7, batch 18, and batch 19.

control charts shows that the process itself evolves

over time as it is exposed to various disturbances such
as temperature changes; after the process instability
occurred, the projected process data returned to the
confidence limits. The disturbance will be discussed
below using the contribution plots.
Contribution plots of the variables, at the time when
the SPE limit is violated, are generated and the variables with abnormally high contribution to SPE are
identified. Figure 7 illustrates the results of fault classification in batch # 7, 18 and 19. In order to identify
the variables that caused this abnormal situation, contribution plots of the variables contributing to the
SPE at the time when the SPE violated the limits are
generated. Variables 2, 5 and 6 (bath temperatures
in strategic points of the pot) seem to be the major
contributors in the SPE.
4

CONCLUSIONS

Multivariate monitoring and diagnosis techniques

have the power to detect unusual events while their
impact is too small to cause a significant deviation
in any single process variable. This is an important

REFERENCES
Dunia, R. and Qin, S.J. 1998a. Subspace approach to multidimensional fault identification and reconstruction. The
American Institute of Chemical Engineering Journal 44,
8: 18131831.
Dunia, R. and Qin, S.J. 1998b. A unified geometric approach
to process and sensor fault identification. Computers and
Chemical Engineering 22: 927943.
Himmelbau, O.M. 1978. Fault Detection and Diagnosis in Chemical and Petrochemical Process. Elsevier,
Amsterdam.
Jackson, J.E. 1991. A users guide to principal components,
Wiley-Interscience, New York.
Nomikos, P. and MacGregor, J.F. 1995. Multivariate SPC
charts for monitoring batch processes. Technometrics 37,
1: 4159.
Rousseeuw, P.J. and Van Driessen, K. 1999. A Fast Algorithm
for the minimum Covariance Determinant Estimator.
Technometrics, 41: 212223.
Tang, N.-Y. 1999. Characteristics of continuous galvanizing baths. Metallurgical and Materials Transactions B, 30:
144148.
Verboven, S.; Hubert, M. 2005. LIBRA: a Matlab library for
robust analysis, Chemometrics and Intelligent Laboratory
Systems 75: 127136.

204

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Fault identification, diagnosis and compensation of flatness errors

in hard turning of tool steels
F. Veiga, J. Fernndez, E. Viles, M. Arizmendi & A. Gil
Tecnun-Universidad de Navarra, San Sebastian, Spain

M.L. Penalva
Fatronik. San Sebastian, Spain

ABSTRACT: The motivation of this paper is to minimize the flatness errors in hard turning (facing) operations
of hardened tool steel F-5211 (AISI D2) discs using Polycrystalline Cubic Boron Nitride (PCBN) tools with
finishing cutting conditions. To achieve this, two strategies have been developed. First an on-line Conditional
Preventive Maintenance (CPM) system based on monitoring a parameter that correlates well with the geometrical
error, namely, the passive component of the force exerted by the workpiece on the tool. The second strategy
is more sophisticated and consists of the development of an on-line Error Compensation System that uses the
error value estimated using the first strategy to modify the tool trajectory in such a way that flatness errors are
kept within tolerances. Moreover, this procedure allows the life of the tools to be extended beyond the limit
established by the first CPM system and also the reduction of part scrap and tool purchasing costs.

INTRODUCTION

The process of cutting high hardness materials with

PCBN tools is known as hard turning. This process
has been increasingly used in finishing operations and
has gained position against common grinding in recent
decades. The reason is that it presents two advantages
over grinding: a) lower process costs, and b) lower
environmental impact (Knig, 1993).
But it has been recognized that before being a
consolidated alternative to grinding, some process
outputs have to be improved, such as: part surface
integrity (absence of white layer) (Schwach, 2006.
Rech, 2002), surface roughness (Grzesik, 2006. zel,
2005. Penalva, 2002) and geometrical errors.
So far, many papers have dealt with the problem
of workpiece distortions (Zhou, 2004) induced by the
cutting and clamping forces and the heat flowing into
the workpiece from the cutting zone, the motors, the
workplace temperature and so on (Sukaylo, 2005). But
the effect of tool expansion on error (Klocke, 1998)
has yet to receive attention.
However, in hard turning, tool expansion should be
considered because when tools become worn the heat
power in the cutting zone, compared to that of conventional turning, is several times higher. In this paper,
tool expansion will be addressed for the development of a Conditional Preventive Maintenance (CPM)
system in order to determine when worn tools should

be replaced by fresh ones so that workpiece errors do

not exceed a certain value (zel 2005, Scheffer 2003).
CPM systems (Santos, 2006. Luce, 1999) have the
advantage, over the common Preventive Maintenance
systems of not requiring any knowledge about the evolution of the variable of interest, in this particular case,
the workpiece error. However, it need to find a parameter that is highly sensitivity to that variable and easy
to follow up on, so that by monitoring the parameter
the variable of interest can be estimated.
In this paper, flatness errors in a hard facing operation will be studied and a CPM that will identify when
the tool wear level is such that it can no longer machine
acceptable workpieces will be proposed.
Once the CPM system has been developed, an on
line Error Compensation System will also be developed so that tool trajectory can be modified on-line in
such a way that errors can be compensated for. This
system will allow the further use of tools that have
been classified as worn out by the CPM system.
2

FLATNESS ERROR

In Figure 1 an orthogonal model for hard turning can

be observed. In zone I the workpiece material is plastically deformed by shear. In zone II sticking and sliding
occurs between the chip and the rake face of the tool
and in zone III the workpiece transitory surface slides

205

ef (d)

Chip

Ef
Workpiece

Tool
(II)
(III)

(I)

Figure 3. Flatness error profile on a hard turned disc of

150 mm external and 50 mm internal diameters.

Workpiece
Figure 1.
model.

Heat generation zones in an orthogonal cutting

point in which it reaches a maximum value Ef . As this

error is still aggravated by the amount of tool wear,
its value has to be checked at regular times so the tool
can be removed when the total flatness error Ef value
approaches the tolerance value. In this paper, a Conditional Preventive Maintenance System (CPM) will be
developed to assess the Ef value.
This CPM requires a parameter that a) is easy
to record, and b) correlates well with flatness error.
The passive component of the force, exerted by the
workpiece on the tool, has been found to fulfill
both conditions and the experimental work leading
to the parameter selection is described in the next
paragraph.

Tool
Programed path
Workpiece

Tool
Real path

Figure 2.

Typical deviation of tool tip from programed path.

on the clearance face. All these zones generate heat

that strongly increases with tool wear. But compared
to conventional hard material turning, the overall heat
power generated in hard turning is very high as a result
of the exceptional hardness of the workpiece (around
60 HRC).
Most of this heat is removed by the chip itself, but
since the tool material in hard turning (CBN) has a
thermal conductivity that is substantially greater than
that of the chip the tool the temperature increases considerably. This rise in temperature is still greater in
hard turning due to the fact that no cutting lubricant
is used. Therefore, the tool tip volumetric expansion
is high as compared to that of conventional turning.
Of that volumetric expansion, the one in the radial
direction (perpendicular to the workpiece flank wear
contact) forces the tool tip to penetrate into the newly
formed surface and thus a deviation in the programed
tool tip path (Figure 2) is produced and, as a consequence, the local flatness error ef (d) (Figure 3)
increases from the tool entrance point to the tool exit

TEST CONDITIONS

To achieve statistically sound results 6 tools of low

CBN content (4050%), typical for continuous turning, were used and each tool made 20 facing passes to
a disk of the hardened tool steel F-5211 (AISI D2) with
150 mm of external and 50 mm of internal diameters.
All the passes started at the external diameter and no
cutting fluid was used.
During the passes the force signals were measured
by a piezoelectric dynamometer with three components. Also, after each pass the resulting face profile
was captured by a LVDT (Linear Voltage Displacement Transductor) sensor attached to the tool turret
without removing the work-piece from the machine.
A stereoscopic microscope, equipped with a digital
camera, was used to measure the maximum value of
the flank wear VB max of the tool after each pass. All
this data has been summarized in Table 1.
4

FAULT IDENTIFICATION AND DIAGNOSIS

OF FACING ERROR

Two parameters were considered for the total Ef facing error estimation (see Figure 3), a) the length cut
by the tool and b) the passive force exerted on the
tool. Since both of them are expected to be sensitive

206

Experimental work done to the analysis of flatness

Material

UNE F-5211 (62 HRC hardness)

Cutting passes

20 facing passes on each of 6

different tools
Discs of 150 mm external diameter
and 50 mm internal diameter.
4050%
Finishing
Feed rate = 0, 1 mm/rev, depth of
cut = 0, 1 mm, cutting speed
180 m/min, dry cutting.
Acquisition, during the cutting
of the three force components
(cutting, feed and passive)
Radius contour acquired by a
contact LVDT sensor
Flank tool wear measured (VB max )

Workpiece
CBN Tool content
Operation
Cutting conditions
Force acquisition
Geometrical error
Tool wear

50

Flatness error (micron)

Table 1.
error.

tool 1
tool 2
tool 3
tool 4
tool 5
tool 6

40
30
20
10
0

500

1000 1500 2000 2500 3000

Figure 4. Evolution of total flatness error Ef with the tool

life for 6 different tools.

to tool wear they should also be good facing error

estimators.
First the sensitivity of the cut length estimator will
be studied in section 4.1 and later that of the passive
force in section 4.2.
4.1

Cut length as a flatness error estimator

As was expected, tests have shown that flatness error

actually increases with tool wear and therefore with cut
length and, since the cut length is very easy and inexpensive to record, it is appealing as an error estimation
parameter. Flatness error was recorded for 6 tools and
in Figure 4 it can be observed that the error values
have a large scattering. For example, for a cut length
of about 2,700 m the scatter is of 22 m (between 18
and 40 m).
This is a consequence of the VB max values dispersion, which is also high, observed for tools with the
same cut length due to the chipping phenomena that
sometimes appears. To illustrate this, Figure 5 shows
two tools that, after having cut the same length, present
very different VB max values, as a result of the chipping
that can be observed in the tool on the right. The total
error values Ef of workpieces machined by these 2
tools have been measured and, as expected, the one
corresponding to the chipped tool presented a much
larger Ef value.
Apart from the dispersion in Ef values, Figure 4
shows that for all of the 6 tools the total flatness error
Ef sometimes goes down with respect to that of the
previous pass. For example, tool number 5 produces a
smaller Ef error in pass number 5 than in the pervious
pass number 4. This is something unexpected since
Ef should have been larger in pass 5 because a larger
cut length should have generated a greater tool wear

Figure 5. Flank wear for two different tools after having cut
the same length.

value in pass number 5 compared to pass number 4.

An explanation to this could be that Ef error is a very
complex phenomenon that depends, not only on the
maximum flank wear VB max value, but also on other
parameters like the flank wear topography. It is easy
to understand that tools with the same VB max values
but with different tool flank topographies will also create different contact conditions that generate different
amounts of heat power and therefore different error
values.
Therefore, as a result of the high dispersion values
of Ef , the cut length can not be considered a suitable estimator of the facing error, so the feasibility
of the other parameter, the passive force, will now be
considered.
4.2

Passive force as a flatness error estimator

As was commented in section 2, heat power generated

during the cutting process, brings about the expansion
of the tool tip and as can be seen in Figure 6 the
workpiece reacts to this with a passive force Fp . When
facing materials of conventional hardness, Fp is very
small, but for hard materials this component is even

207

60

Tool

Workpiece

Flatness errors (microns)

Room
Temperature
Cutting
Temperature

Fp

e f (microns)

Regression
95% CI
95% PI

40
30
20
10
0

Figure 6. Diagram of tool tip expansion and force generated

due to heat power.

50

100

150

200

Figure 8. Regression line of local flatness error as a function

of the passive force variation for 6 different CBN tools.

0
-10
-20
150

125

100

75

50

75

50

unacceptable errors, the correlation of the two curves

is good and therefore ef error can be estimated by
recording the Fp
Once it is demonstrated that a linear relationship
exists between ef (d) and Fp , the next step is to find
its arithmetic expression. To do this, 7 diameters were
randomly selected and the values of ef (d) were plotted against the corresponding passive force variations
(Fp (150) Fp (d)) and the resulting point cloud fits
the following linear expression:

Diameter(mm)
Fp (N)

50

160
140
120
100
80
150

125

100

Figure 7. Flatness error (above) and passive force (below)

during the facing of a hard steel disc with: V = 180 m/min.,
feed and a cutting depth of = 0.1 mm with a PCBN tool with
VB max = 0, 15 mm.

larger than the cutting and feeding forces and, as a

result, could be a facing error estimator although, so
far, it has not received much attention in the literature
(Lazoglu, 2006).
Figure 7 depicts the evolution of the local flatness
error ef (d) at the top and the Fp force at the bottom of
a randomly selected pass and a remarkable similarity
in their shapes can be observed. This suggests that the
correlation between the two may be high.
The correlation coefficient between ef (d) and
(Fp (150) Fp (d)) was calculated for all the tools and
results shows that it is greater than 0,95 for 90% of
all the cases. Fp (d) is the passive force in the diameter
d and Fp (150) is the force at the external diameter of
the disk where all passes begin. The remaining 10%,
(8 passes) with a value below 0.95, corresponded to
fresh tools (cut length below 500 m), but since fresh
tools always give acceptable errors, it can be concluded
that, where tool wear level is high enough to produce

ef (d) = 5, 52 + 0, 23(Fp (150) Fp (d))

(1)

with a linear regression value of 77.6% that is statistically significant.

Since experiments have shown that for any pass the
maximum value of the error variation is always located
at the minimum diameter (50 mm), equation (1) particularized for this point can identify at the pass where
the total error Ef has reached such a critical value that
the operator should proceed to replace the worn out
tool. Therefore a Conditional Preventive Maintenance
(CPM) Strategy has been developed. It indicates when
a tool should be replaced to keep errors below a critical
level.

FLATNESS ERROR COMPENSATION

STRATEGY

One step further in the direction of tool use optimization is the development of an Error Compensation
System that allows the on-line correction on the
tool tip trajectory as it deviates during the facing pass
as result of the tool tip expansion.

208

1) Signal (Fp(d))
PC
2) Flatness error
(ef (d))

Dynamometer

Z
carriage

A Conditional Preventive Maintenance (CPM)

System has been developed for the estimation of the
part error value in a hard turning (facing) operation. It
is based on the passive component of the force exerted
by the workpiece on the tool. This system presents two
advantages over the well known Systematic Preventive
Maintenance Systems: important reduction in workpiece scrap and a more efficient use of the tools.
An Error Compensation System (ECS) has also
been developed. It employs the error value that has
been estimated with the CPM mentioned above, to
compensate the tool trajectory in such a way that flatness errors are kept within tolerances even using tools
with a wear level that had been rejected by the CPM
system. Compared to the CPM system, the ECS gives
better part quality and extends the life of the tools.

Mz

Clamping
Workpiece

Z Motor

CNC

Figure 9.

On-line Compensation strategy diagram.

This is possible by implementing equation (1) in a

machine tool with an open CNC control. The system
(Figure 9) performs the following tasks:
1 Data Acquisition and analysis
2 PC-CNC Communication
3 New tool position generation

5.1

ACKNOWLEDGEMENTS
This work has been made under the CIC marGUNE
framework and the authors would like to thank the
Basque Government for its financial support.

Data acquisition and analysis

In this module, the passive force Fp (d) signal is first

obtained by a dynamometer mounted on the tool turret, then it is captured by a National Instruments DAQ
board and is then analyzed by a PC using a specific
software application that has been developed in Lab
View. In the PC the (Fp (150) Fp (d)) values are calculated for a number of positions which depend on the
required accuracy, and finally through equation 1 the
corresponding ef (d) are estimated
5.2

PC-CNC communication

Once ef (d) has been calculated in 5.1, an analog signal,

with the appropriate compensation value is generated
in the Lab View application and sent to the CNC unit
of the machine via a DAQ board analog output.
5.3

CONCLUSIONS

New tool position generation

The machine PLC (Programmable Logic Controller),

that analyses periodically all the machine parameters, registers the compensation value of the signal
input that has been received by the CNC and writes it
in the tool compensation parameter of the machine so
that the Z drive motor can modify the Z coordinate of
the tool tip to compensate for the flatness error. The
correction period can be selected by the operator but it
will always be higher than the Z axis cycle time (a few
centiseconds). For common hard facing cutting speeds
and feeds, periods between 0,1 and 1 seconds can give
decent compensation results.

REFERENCES
Grzesik W., Wanat T. 2006. Surface finish generated in hard
turning of quenched alloy steel parts using conventional
and wiper ceramic inserts. Int. Journal of Machine Tools
& Manufacture.
Knig W.A., Berktold A., Kich K.F. 1993. Turning vs
grindingA comparison of surface integrity aspects and
attainable accuracies. CIRP Annals 42/1: 3943.
Lazoglu I., Buyukhatipoglu K., Kratz H., Klocke F. 2006.
Forces and temperatures in hard turning. Machining
Science and Technology 10/2: 157179.
Luce S. 1999. Choice criteria in conditional preventive maintenance. Mechanical Systems and Signal Processing 13/1:
163168.
zel T., Karpat Y. 2005. Predictive modelling of surface
roughness and tool wear in hard turning using regression
and neural networks. Int. Journal of Machine Tools &
Manufacture 45/45: 467479.
Penalva M.L., Arizmendi M., Daz F., Fernndez J. 2002.
Effect of tool wear on roughness in hard turning. Annals
of the CIRP: 51/1: 5760.
Rech J., Lech M., Richon, J. 2002. Surface integrity in finish hard turning of gears. Metal cutting and high speed
machining, Ed. Kluwer, pp. 211220.
Santos J., Wysk Richrad A., Torres, J.M. 2006. Improving
production with lean thinking. John Wiley & Sons, Inc.
Scheffer C., Kratz H., Heyns P.S., Klocke F. 2003, Development of a tool wear-monitoring system for hard turning,
International Journal of Machine Tools & Manufacture,
Vol. 43, pp. 973985.
Schwach D.W., Guo Y.B. 2006, A fundamental study
on the impact of surface integrity by hard turning on

209

rolling contact fatigue, Int. Journal of Fatigue, 28/12:

pp. 18381844.
Sukaylo V., Kaldos A., Pieper H.-J., Bana V., Sobczyk
M. 2005, Numerical simulation of thermally induced
workpiece deformation in turning when using various cutting fluid applications, Journal of Materials Processing
Technology, Vol. 167, pp. 408414.

Zhou J.M., Andersson M., Sthl J.E. 2004, Identification

of cutting errors in precision hard turning process, Journal of Materials Processing Technology. Vol. 153154,
pp. 746750.

210

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

From diagnosis to prognosis: A maintenance experience

for an electric locomotive
O. Borgia, F. De Carlo & M. Tucci
S. Stecco Energetics Department of Florence University, Florence, Italy

ABSTRACT: Maintenance policies are driven by specific needs as availability, time and cost reduction.
In order to achieve these targets, the recent maintenance approach is based on a mix of different policies such
as corrective, planned, and curative. The significance of the predictive activities is becoming more and more
important for the new challenges concerning machines and plants health management.
In the present paper we describe our experience about the development of a rule-based expert system for the
electric locomotive E402B used in the Italian railways system, in order to carry out an automated prognostic
process. The goal of the project was to develop an approach able to improve the maintenance performance. In
particular we would like to develop an advanced prognostic tool able to deliver the work orders. This specific
issue has been identified and requested from the maintenance operators as the best and the only solution that
could ensure some results.

INTRODUCTION

Predictive and condition based maintenance policies

are intended to assess the state of health of the devices
to which they are applied. This objective is achieved
by running a continuous (sometimes might also be
on-line) monitoring of the equipments. The ultimate
purpose is to carry out the best maintenance operations at the most appropriate time, obviously before the
systems failure, in order to minimize the overall costs.
Two important phases of the predictive maintenance process are the incipient fault identification
and the subsequent life forecasting, respectively called
diagnostic and prognostic processes.
Nowadays in the industrial applications, the diagnostic process is very frequent and often well-done,
also thanks to the recent evolution of sensors technologies, data acquisition and analysis. Conversely the
prognostic approach is not so common and there arent
specific computer-based methodologies or widely
spread software tools developed for this second stage.
This paper describes an experience in the railways maintenance environment, aiming at developing a prognostic approach, starting from an existing
diagnostic system.
The case study considered is the diagnostic system
of an entire train fleet, composed by 70 locomotives
operating on the Italian railways net.
The first step of our work, as explained in the paragraph 2, was the data collection action, including extra
diagnostic data, additional maintenance reports and

information about the commercial services of the locomotive. The user of the locomotive fleet, Trenitalia
S.p.A., has potentially a lot of information but theyre
fragmented in many different databases. Therefore,
this data and knowledge, as shown in paragraph 3, is
very unproductive and inefficient.
In paragraph 4 is presented the analysis of the diagnostic data. Their great number couldnt be easily
managed. Hence, in order to keep only the relevant
data with some useful information, we adopted a
customized statistical approach.
In this manner we could remove many of the useless
records, for example the ones describing secondary
events. We could achieve our aim of fault identification thanks to the analysis of the only truthful
diagnostic records.
The prognostic skill, based on the rules logic, was
gained by the matching of maintenance activities and
of diagnostic records for a restricted time interval and
distinguishing a specific fault event.
As illustrated in paragraph 5, in the development of
the latter task we faced a lot of problems as the quantity
and reliability of the data, the information ambiguity
and the weak know-how either of the product and of
the maintenance activities.
In order to take advantage of all the available data
and information we have suggested, in paragraph 6
a new configuration for the predictive maintenance
process.
As a final point, in the last paragraph are proposed
some considerations regarding the future perspectives

211

of the true integration of the prognostic approach

within the maintenance strategies.

Table 2.
Data

Let us consider one of the electric locomotives of the

fleet investigated, as a binary system with two possible
states: operational and under maintenance.
All the data and information collected in the first
phase of the project could be ascribed to one of these
two states. All the data and information have been
classified in two clusters and the criterion for the classification has been the state of the rolling stock during
the time when the data have been generated.
Operational data & information

The first set is made by the information coming from

the operational phase. They are shown in the following
table with a brief description and with the name of the
belonging database.
The first two data types are at first stored temporarily on the memory of the trains data logger. Then,
every five minutes, they are sent, by means of a GPS
transmission, to a ground-fixed server. A specific web
application, named DRView, allows to visualize and
quickly filter the data by an advanced database query.
Table 1.
Data
Event
code

Operational data & information.

Description

Database

The locomotive is equipped by

DR View
a control system able to give a
specific alarm when a set of
operation parameters are over
their standard values. The
alarms are available on line for
the drivers, while the mission is
going on, and they are also
stored for maintenance
opportunities
DiagThe locomotive is equipped by a
DR View
nostic
control system able to give a
code
diagnostic message when a set
of abnormal parameters or event
codes have been recorded. The
diagnostic codes are stored
during the mission and they are
available for maintenance
operators
Commer- They are the daily travels with
GEST-RAZ
cial
distance information, the main
service
stations and the number of the
trains that have been carried out
by a locomotive

Description

Database

Maintenance Maintenance note contains the

SAP PM
note
information about a problem
that could be solved by a
maintenance operation
Work
Work order are generated
SAP PM
order
originating from a maintenance notice. It contains the
information for carrying out
maintenance operation (task,
skills, spare parts,). After fulfilling the job, the record is
completed with the description of the accomplished
maintenance activity
Expert
Every important fault or breakServer
opinion
down with important conseworkquences in the operational
sheet
phase are analyzed and commented from a group of experts.

DATA, INFORMATION AND KNOWLEDGE

SOURCES

2.1

Maintenance data & information.

Thanks to the GPS technology, each data is also provided with some information about the geographic
position of the place where it was generated.
At the same time, some essential information (event
code) is shown directly on the on board screen of the
drivers display in order to help managing the train
while the mission is going on.
The last data type are the daily missions carried
out from the locomotive, enriched with some related
information. These data come from a software, made
for fleets management which, at the present time, is
not integrated with the DRView tool.
The first two just mentioned kinds of information
are not available for all the fleet, since the data logger
and the transmission apparel is installed only on two
vehicles (namely number 107 and 119). The data from
the other locomotives are never downloaded into the
databases. The available data recorded on the fixed
ground server have been recorded within a time interval starting from June 06 until now. In the meanwhile
Trenitalia is developing an ongoing project for the
extension to other 10 locomotives of the on board GPS
transmitter in order to be available a wider number
of data.

2.2 Maintenance data & information

The second set of data comes from the maintenance
phase, because its collected by the maintenance crew
and managed trough a dedicated CMMS.
Maintenance notes and work orders are available
for all components of the fleet but the quality of the
included information often very poor.

212

A maintenance note can have three different origins,

each one of which its specific identification code:
ZA, if it comes from a note on the trip book of
the drivers. They are advices collected during the
mission and generally point out some abnormal
conditions;
ZB, if it comes from the inspection of the locomotive
when it enters in the maintenance workshop;
ZC, this code has been provided for a future
automated monitoring system that will be able to
diagnosis the state of the vehicles.
The opinion of the experts is simply recorded in a
worksheet. Nevertheless the information included are
very useful to understand what has really happened to
a locomotive in case of a alarming failure. Sometimes
also some suggestions for the development of a correct
approach to manage the fault are reported, and some
indications about when it should be better to intervene.
After reporting the list of data & information available into the three different not-integrated databases,
the next paragraph will explain how they are managed
and used to perform the diagnosis of the degradation
state of a locomotive and the subsequent maintenance
activities.
3

THE CURRENT MAINTENANCE

MANAGEMENT PROCESSAS IS

A diagnostic system has the main goal to execute an

on condition maintenance policy in order to prevent
failures and, when eventually faults occur, may help in
precisely identifying the root causes.
At the present time the maintenance process of the
locomotives owner provides two different procedures
for carrying out a work order. Both of them generate
either corrective or planned maintenance operations.
Usually a locomotive goes in a maintenance workshop
only if a deadline of the maintenance plan is expired
or if there is an heavy damage that causes some delays
or a reserve.
The fact remains that the suggestions of the on board
diagnostic system are currently ignored. The only way
to look at the operational data is the on board drivers
logbook. In such a document the abnormal conditions,
episodes or events (event code) are reported as long as
the on board operator considers them important for the
future maintenance activities.
3.1

Figure 1.

Work order generation process.

Moreover the maintenance notes can be generated

from a maintenance worker when the locomotive is in
maintenance operation for a planned task or for a problem detected in the general inspection when it enters
in the workshop. This general beginning inspection
includes obviously the analysis of the travel book.
This approach doesnt allow to have a clear correspondence between maintenance notes and work
orders because, as predictable, very often a work
order contains the reports of a few maintenance notes.
Sometimes it can also happen that a single work order
contains contemporarily tasks coming from the maintenance plan and from abnormal conditions signaled
in operational state.
The next image is meant to represent the different
process that can lead to a definite maintenance activity
on the vehicles.

3.2 What happens to a locomotive?

The maintenance process just described provokes a
strong lack of balance in the maintenance frequencies. For a long period (three or four weeks) a vehicle
is not maintained and it is recalled in the workshop
only if a severe fault occurs or if a planned task must
be accomplished. This approach generates a big heap
of stress due to the natural process of decay of the
locomotives main equipments. This behavior usually
involves a series of serious damages, a debacle for the
mission of the train.

Maintenance note and work order

As mentioned in paragraph 2, the maintenance warnings could be generated in the operational phase even
from the traffic control room. It happens, for instance,
when the locomotive has a problem during a mission
and the driver signals a specific problem.

3.3 Information & data quality

The most significant problem for the collected data
and information is their intrinsic reliability.
For example, all the locomotives documentation,
such as manuals, schemes, etc., have not been updated

213

for a long time and sometimes it is very difficult to

restore the right knowledge.
Since the locomotives manufacturers are not longer
traceable while the diagnostic systems manufacturer
are still in good relations, a lot of helpful information
has been lost.
On the other hand the maintenance data are very
often extremely poor in terms of information. This is
due mainly to the maintenance operators approach
which doesnt give the right value to the collection
of maintenance data. Hence the maintenance activities description doesnt allow to understand accurately
which tasks have been actually accomplished.
Although our locomotives belong to a previous
generation, without any chance of improvement, the
medium age of the means is very low. For this reason the fleet has hopefully before itself still many
years of operation, justifying the efforts, in terms of
human resources (experts) and technological investments (diagnostic system), aiming at improving the
maintenance strategy.
In the next paragraphs we will report about the
analysis and studies performed to carry out a diagnostic and prognostic approach, able to take some more
advantage from the existing data.

EXPERT SYSTEM DEVELOPMENT

The available data and their attributes suggested us

that a rule based system would have been the most
suitable approach in order to improve the maintenance
performance.
Considering Figure 2, on the one side we have
the data coming from operational phasethey are
very unmanageable, as we will show in the next
paragraphbut if properly treated, they represent
a very important information and a signal of the
incipient fault.

Figure 2.

Scheme of the available data.

Figure 3.

Rule construction approach.

On the other side of the picture we can see the maintenance data. Unfortunately their poor quality doesnt
allow us to use them to drive the rules construction
process, as expected for an ideal system configuration. At the moment the maintenance information can
be used barely to receive a confirmation after a diagnosis carried out from the diagnostic code of the on
board monitoring system.
A diagnostic rule is a proposition composed by two
arguments, the hypothesis and the thesis. The Hypothesis (XX, YY) are the events that must happen in
order to generate an event (ZZ) and therefore verify
a thesis. If the diagnostic rule is maintenance oriented,
as a replacement for the cause the rule can contain
the specific maintenance activity able to interrupt the
degradation or to resolve a breakdown if the fault is
already occurred.
In case of advanced diagnostic systems the rule
should can be prognostics oriented. This means that
the thesis statement is a prediction of the remaining
life to failure of the monitored component.
In case of an on condition based maintenance, the
thesis statement contains the prediction of the deadline
for performing the appropriate maintenance activities.
This actions should be able to cut off or at least manage
the incoming component failure. The following chart
is meant to explain these different rule methods.
In this project one of the goals is to gain a prognostic approach from the diagnostic system. Possibly
it should also be biased toward conducting an on condition maintenance. So the thesis declaration will be
generated by matching the users manual indications,
the experts suggestions and the work orders activities
descriptions.
Since the hypothesis are the most critical element in
the rule, they will be extracted from the on board diagnostic database by a significant statistical treatment.
We used the techniques developed for the process
control in order to manage a huge number of data.

214

DATA ANALYSIS AND TREATMENT

As mentioned in paragraph 2, at the moment the available data of the on board diagnostic system are only
those of locomotives number (#) 107 and 119. As
reported in the following table the data comparison
shows significant difference for each means of transportation. The locomotive #107 has nearly twice the
number of codes per kilometer of the #119. Conversely the maintenance activities are distributed with
the opposite ratio: #119 has roughly the 33% of maintenance operations more than #107. The data used
in the analysis were collected in five months, from
January to May 2007.
Although they are identical for the design, for all
subsystems characteristics and for the operational and
maintenance management, each rolling stock has its
own specific behavior.
So the diagnostic approach, basing on statistical
methods, cant be developed without considering the
specific vehicle. Aging, obviously, causes a natural
and not avoidable drift of the functional parameters.
This phenomenon is strongly amplified in complex
systems, as is a modern electric locomotives, where all
the subsystems are dependent on each other. Accordingly the time dependence of the reference parameters
will be appreciated to guarantee a dynamic diagnostic
process.
5.1

those codes whose related information are not enough.

This unpleasant circumstance is due to a bad management of the fleets documentation; many events, such
as diagnostic systems upload or modification, are not
still mapped from the user so now and then important knowledge is lost. The final number is achieved
through the cutting of the many life messages and of
some codes considered not reliable after a technical
discussion with some maintenance experts.
In Figure 4 data of locomotive #119 are shown.
Starting from this point of the paper until the end, we
will show the study results of these data. The same
process was made for the other vehicle but, for briefnesss sake it wont be described since the results are
significantly similar.

5.2 Pareto analysis and codes classification

The diagnostic system is based on about two hundred
different codes. Their great number led us to perform
a Pareto analysis in order to identify the most important code. The results in terms of codes occurrence is
reported in the following diagram.
As clearly visible the most frequent eight codes correspond to the 80% of the occurrences. Although they

Filtering

First of all the main important action of the data treatment was the filtering that allowed the management of
a suitable number of information.
As reported in Table 3, the initial number of codes
was very high, the average production is 1,8 codeskm1 and 0,8 codeskm-1 respectively for #107 and #119.
The progressive filtering procedure and its results
in terms of code identification is reported in the following table. For each step is also reported the reason
of the choice.
The first step of the filtering process is due to the
indications reported on the users manuals where it
is shown which codes are useful in support of the
driver for the management of the abnormal conditions in the mission occasions. The next removal hits
the codes whose meaning is completely unknown and
Table 3.

Figure 4.

Filtering process.

Figure 5.

Pareto analysis of diagnostic code.

Data comparison.

Data

Loco n.107

Loco n.119

Event codes
Diagnostic codes
Maintenance notes
Work orders

0, 53 km1
1, 2 km1
88
23

0, 28 km1
0, 52 km1
129
44

215

could be easily clustered in two different sets referring to two subsystems of the locomotive (engineers
cab HVAC and AC/DC power transformation), we will
treat them separately without considering any mutual
interaction.
Another important element is the logic of codes
generation by the on board diagnostic system. As
mentioned in paragraph 2 a diagnostic code is generated when a set of specific physical parameters
reach their threshold values. This means that you have
reached abnormal operating conditions. Each code
record has a start time and an end time field, representing respectively the appearance and disappearance
of the anomalous conditions. The values of these
attributes allow us to suggest a classification of codes
basing on the duration of the abnormal conditions:
Impulsive signal, code with the same value for the
start time and end time. It means that the duration
of the abnormal conditions is less than a second;
Enduring signal, code with different values for the
start time and the end time. The durations has an
high variability (from some seconds up to some
hours);
Ongoing signal, code characterize by a value for
the start time but without end time. It means an
abnormal condition still persistent.
Many codes are generated by different arrangements of signals. Nevertheless they represent an
alteration of the equipments state so they are able to
describing the states of the locomotives subsystems.

6
6.1

THE PROPOSED METHODOLOGY

Statistical process control

Production processes will often operate under control,

producing acceptable product for quite long periods.
Occasionally some assignable cause will occur and it
results in a shift to an out-of-control state, where a large
proportion of the process output does not conform anymore to the users requirements. The goal of statistical
process control is to quickly detect the occurrence
of precise causes or process shifts so that investigation of the process and corrective actions might
be undertaken before many nonconforming units are
manufactured.
The control chart is an online process-monitoring
technique widely used for this purpose. It is a tool
useful for describing what is exactly meant by statistical control. Sample data are collected and used to
construct the control chart, and if the sampled values
fall within the control limits and do not exhibit any
systematic pattern, we can say that the process is in
control at the level indicated by the chart.

Control charts may be classified into two general

types:
variables control charts (Shewhart, CUSUM,
EWMA), when the variable is measurable and its
distribution has a central tendency;
attributes control charts (p-chats, np-charts,
c-charts, u-charts), when the variable is not measurable and its occurrence is characterized by a Poisson
distribution.
Traditional control charts are univariate, for example the monitoring of an individual variable. This
implies the assumption that the variables or the
attributes used for describing the system are independent from each other. Using multivariate control
charts, all the meaningful variables or attributes are
used together. The information residing in their correlation structure is extracted and it allows a more
efficient tracking of the process over time for identifying anomalous process points. Although the multivariate control charts seem more useful for describing
complex process, they are used less than the univariate. This is due to two reasons: the univariate charts
utilization is simpler and more efficient, moreover
a successful implementation of multivariate control
chart often requires further statistical studies as the
principal components analysis.
6.2 Control chart application
A diagnostic code represents an attribute describing
the state of its corresponding subsystem. We have
applied the u-chart, an attributes dedicated control
chart, to the daily codes occurrences.
Our resulting control chart differs from the standard
ones for some important parameters differentiations.
The used parameters expressions are the following:
m
Central line, CL =

i=1 ui

= u

(1)


Upper control limit, UCL = u + 3

u
ni

(2)



Lower control limit, LCL = max u 3

 
u 
0
ni 
(3)


Upper warning limit, UWL = u + 1, 5

u
ni

(4)

where ui is the daily frequency per kilometer of the

code, m is the number of days when the process is

216

defined: UWL, UCL. Both of them correspond to a

specific deadline:
when a code occurrence gets over the UWL, the
maintenance activity should be performed within
five days;
when a code occurrence gets over the LCL, the
maintenance activity should be performed as soon
as possible.

Figure 6.

u-chart for a diagnostic code.

considered under control and ni is the daily distance

covered from the vehicle.
The central limit of the control chart has been
calculated in a time interval (m days) where the considered locomotive is under statistical control, so the
recorded diagnostic codes can be regarded just as
noise because theyre generated from not identifiable
stochastic events.
The choice of the daily frequency per km as
chart variable, forces us to introduce the variance
of the control limits depending on the daily covered
distance.
As above reported, the lower control limit is equal
to the standard expression only when it is more than
zero; otherwise its zero. This solution is adopted
because the definition of a negative threshold for an
event occurrence hasnt any meaning. Moreover we
have introduced a further limit: the upper warning
threshold. Its value is the half of the UCL.
In the following diagram an example of the application of the control chart to a diagnostic code is
reported.
6.3

The prognostic approach

In the paragraph 4, we have underlined how the expert

system should be based on a rules approach with a
prognostic focus. The data analysis process identifies the hypotheses of the rules as a threshold of the
occurrence of a diagnostic code.
On the other hand the thesis of the rule is a maintenance note for the subsystem including a deadline
warning for achieving the appropriate maintenance
activity.
Hence the rule model sounds like the following
assertion: if a diagnostic codes occurrence rises above
a specific value, a maintenance action must performed
within a fixed time.
The definition of the control chart parameter was
performed looking at the problem from this point of
view, consequently two different thresholds have been

The execution of the control charts as a monitoring

tool of the reliability performance of the train introduces another important element: the trend analysis.
Or else matching the trend occurrence with experts
experience. When a code occurrence falls for three
times consecutively between CL and UWL a degradation process must be in progress so a maintenance
activity should be scheduled inside the ten following
days.
7
7.1

FINAL CONSIDERATION
The next step of the analysis

There are a lot of opportunities for the extension of

this project, following various directions. First of all
we could investigate the likelihood to detecting specific thresholds and maintenance deadlines for each
diagnostic code.
A second step could be the investigation of the daily
codes frequency in terms of run hours by a different
attribute control chart (c-chart).
After that we could try to carry out a multivariate
statistical analysis, trying to cluster the codes belonging to the same subsystems. As foreseeable they are
conditioned by each other but in this first phase of
the analysis we didnt have enough time to study their
correlations.
Another important issue is the opportunity to investigate the daily permanence time of diagnostic codes
using the control charts as variables. Probably this data
is more helpful in terms of information than the code
occurrence.
7.2

Suggestion for the maintenance process

The project has highlighted some criticalities in the

maintenance process management. On the other side
a wide set of opportunities and possible improvements
has been outlined.
First of all, in order to return some helpful maintenance data, an important quality improvement should
be carried out. This can be obtained by two main
actions:
a strong simplification of the data entry for the maintenance operators. A computerized check list of the
possible entries for each field of the documents is

217

the only solution suggested from the maintenance

expert experience. A fixed number of choices for
field can decrease the richness of an information
but can improve the data reliability and availability;
an effective awareness campaign for the operators
that should be properly motivated by a specific
training course and, eventually offering financial
rewards.
Then we suggest the implementation of a computerbased platform able to:
integrate all the data, coming from different sources,
helpful for the prognostic approach;
carry out the analysis process that has been developed in this paper.
Another important issue of the platform could be the
chance to automatically generating the maintenance
notes, when a rule of the expert system based on the
statistical process control would be verified. Only this
type of integration and automation can be helpful to
perform a properly predictive maintenance strategy.
7.3

Overview

This paper reports the partial results of a project that

is still in progress but gives the possibility to face the
most relevant and typical problems of the maintenance
field.
As foreseeable, nowadays, the available technology guarantees any data and information without any
storage constraints or communication limits.
So the current challenges are on the one side the data
management, in terms of integration and treat-ment
from the expert system. On the other side the human
resource management in terms of expert experience
formalization and field operator activities.
These issues became necessary to implement a
maintenance strategy basing on a prognostic approach
that should be able to meet the requirements of a
competitive market.
REFERENCES
Chain, L.H., Russel, E. & Braatz, R.D. 2001. Fault detection and diagnosis in industrial systems. London, Berlin,
Heidelberg: Springer-Verlag.

Jardine, A.K.S., Lin, D. & Banjevic, D. 2006. A review

on machinery diagnostics and prognostics implementing
condition-based maintenance. Mechanical Systems and
Signal Processing 20: 14831510.
Leger, R.P., Garland, Wm.J. & Poehlmanb, W.F.S. 1998. Fault
detection and diagnosis using statistical control charts
and artificial neural networks. Artificial Intelligence in
Engineering 12: 3541.
MacGregor, J.F. & Kourtl, T. 1995. Statistical process control of multivariate process. Control engineering Practice
3(3): 403414.
Mann, L. Jr., Saxena, A. & Knapp, G.M. 1995. Statisticalbased or condition-based preventive maintenance? Journal of Quality in Maintenance Engineering 1(1): 4549.
Montgomery, D.C. & Runger, G.C. 2007. Applied Statistics
and Probability for Engineers, 4th Edition. New York,
New Jersey: John Wiley & Sons.
Niaki, S.T.A. & Abbassi, B. 2005. Fault diagnosis in multivariate control charts using artificial neural networks.
Quality and Reliability Engineering International 21(8):
825840.
Price, C. 1999. Computer-Based Diagnostic Systems.
London, Berlin, Heidelberg: Springer-Verlag.
Ryan, T.P. 2000. Statistical methods for quality improvement,
2nd edition. New York, New Jersey: John Wiley & Sons.
Schein, J. & House, J.M. 2003. Application of control
charts for detecting faults in variable-air-volume boxes.
ASHRAE Transactions.
Thomson, M., Twigg, P.M., Majeed, B.A. & Ruck, N. 2000.
Statistical process control based on fault detection of
CHP units. Control Engineering Practice 8: 1320.
Tokatli, F.K., Ali Cinar, A. & Schlesser, J.E. 2005. HACCP
with multivariate process monitoring and fault diagnosis
techniques: application to a food pasteurization process.
Food Control 16: 411422.
Vachtsevanos, G., Lewis, F.L., Roemer, M., Hess, A. &
Wu, B. 2006. Intelligent fault diagnosis and prognosis
for engineering systems. New York, New Jersey: John
Wiley & Sons.
Wikstro, C., Albano, C., Eriksson, L. et al. 1998. Multivariate
process and quality monitoring applied to an electrolysis process, Part I. Process supervision with multivariate
control charts. Chemometrics and Intelligent Laboratory
Systems 42: 221231.2.
Yam, R.C.M., Tse, P.W., Li, L. & Tu, P. 2001. Intelligent predictive decision support system for conditionbased maintenance. The international journal of advanced
manufacturing technology 17(5): 383391.
Zhang, J., Martin, E.B. & Morris, A.J. 1996. Fault detection
and diagnosis using multivariate statistical techniques.
Chemical Engineering Research and Design 74a.

218

Human factors

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A study on the validity of R-TACOM measure by comparing operator

response time data
Jinkyun Park & Wondea Jung
Integrated Risk Assessment Division, Korea Atomic Energy Research Institute, Daejeon, Korea

ABSTRACT: In this study, in order to validate the appropriateness of R-TACOM measure that can quantify
the complexity of tasks included in procedures, an operators response time that denotes the elapsed time to
accomplish a given task was compared with the associated R-TACOM score. To this end, operator response time
data were extracted under simulated Steam Generator Tube Rupture (SGTR) conditions of two reference nuclear
power plants. As a result, it was observed that operator response time data seem to be soundly correlated with the
associated R-TACOM scores. Therefore, it is expected that R-TACOM measure will be useful for quantifying
the complexity of tasks stipulated in procedures.

INTRODUCTION

For over several decades, many studies have shown

that the reliable performance of human operators is
one of the determinants for securing the safety of
any human-involved process industries and/or systems, such as nuclear industries, aviation industries,
petrochemical industries, automobiles, marine transportations, medicine manufacturing systems, and
so on (Yoshikawa 2005, Hollnagel 2005, Ghosh &
Apostolakis 2005). In addition, related studies have
commonly pointed out that the degradation of a
human performance is largely attributable to complicated tasks (Topi et al. 2005, Guimaraes et al. 1999,
Melamed et al. 2001). Therefore, managing the complexity of tasks is one of the prerequisites for the safety
of human-involved process industries and/or systems.
For this reason, Park et al. developed a task complexity measure called TACOM that consists of five
complexity factors pertaining to the performance of
emergency tasks stipulated in the emergency operating procedures (EOPs) of nuclear power plants
(NPPs) (Park and Jung 2007a). In the course of
validating TACOM measure, however, a potential
problem related to the interdependency of five complexity factors has been identified. Subsequently, in
order to resolve this problem, the revised TACOM
(R-TACOM) measure was developed based on a
theory of task complexity (Park and Jung 2007c).
In this study, in order to validate the appropriateness of R-TACOM measure, operator response
time data that were extracted under simulated steam
generator tube rupture (SGTR) conditions of two reference nuclear power plants were compared with the

associated R-TACOM scores. As a result, it was

observed that operator response time data seem to
be soundly correlated with the associated R-TACOM
scores. Therefore, it is expected that R-TACOM measure will be useful for quantifying the complexity of
tasks stipulated in procedures.

2
2.1

BACKGROUND
The necessity of quantifying task complexity

As stated in the foregoing section, a human performance related problem has been regarded as one of
the radical determinants for the safety of any humaninvolved systems. It is natural that a great deal of work
has been performed to unravel the human performance
related problem. As a result, it was revealed that the
use of procedures is one of the most effective countermeasures for the human performance related problem
(AIChE 1994, IAEA 1985, OHara et al. 2000). In
other words, procedures are very effective to help
human operators in accomplishing the required tasks
because they can use detailed instructions describing
what is to be done and how to do it. However, the use
of procedures has the nature of a double-edged knife.
That is, since procedures strongly govern the physical
behavior as well as the cognitive behavior of human
operators, it is expected that the performance of human
operators would be largely attributable to the complexity of procedures. Actually, existing literatures support
this expectation, since the performance of human operators seems to be predictable when they use procedures
(Stassen et al. 1990, Johannsen et al. 1994).

221

Figure 1.

The overall structure of TACOM measure.

In this regard, Park et al. developed a measure called

TACOM that can quantify the complexity of emergency tasks stipulated in the EOPs of NPPs (Park and
Jung 2007a). Fig. 1 briefly depicts the overall structure
of TACOM measure.
As shown in Fig. 1, TACOM measure consists of
five sub-measures that represent five kinds of factors
making the performance of procedures complicated.
Detailed explanations about these complexity factors
are provided in the references (Jung et al. 1999).
2.2

Validating TACOM measure

If the complexity of tasks stipulated procedure can

be properly quantified by TACOM measure, then it
is natural to assume that the performance of human
operators can be soundly explained by the associated
TACOM scores. Accordingly, in order to validate
the appropriateness of TACOM measure, it is crucial
to elucidate what kinds of human performance data
should be compared with TACOM scores.
In this regard, two kinds of human performance
data, response time data and OPAS (Operator Performance Assessment System) scores, were compared with the associated TACOM scores. As a
result, response time data as well as OPAS scores
showed significant correlations with the associated
TACOM scores (Park and Jung 2007a, Park and Jung
2007b).

2.3

The revision of TACOM measure

As illustrated in Fig. 1, TACOM measure quantifies

the complexity of a given task using Euclidean norm
of five sub-measures. This is based on the assumption
such that all the five complexity factors are mutually
independent. Unfortunately, in the course of comparing response time data with the associated TACOM
scores, a clue indicating that the independent assumption could be doubtful was observed from the analysis
of correlation coefficients among five sub-measures.
That is, three variables (SSC, AHC and EDC) seem
to be mutually dependent because they have relatively
strong correlations (Park and Jung 2007c). In statistics, this problem is known as multicollinearity (Cohen
et al. 2003).
Accordingly, it is indispensable to prevent against
the possibility of the multicollinearity problem among
five sub-measures.
In order to unravel this problem, creating a multiitem scale was regarded because this approach has
been frequently applied to the treatment of psychological and sociological data. In creating a new scale, the
scores of two or more variables are summated and/or
averaged to form a single scale that represents the characteristics of included variables. However, without a
systematic framework that can provide a theoretical
basis, it is likely to fail in creating an appropriate
scale. For this reason, the theory of a task complexity
is revisited.

222

R-TACOM MEASURE

On the basis of existing literatures, Harvey suggested

a generalized task complexity model that consists of
three complexity dimensions as depicted in Fig. 2
(Harvey & Koubek 2000).
In Fig. 2, it should be emphasized that this complexity model provides the three orthogonal dimensions
that affect the complexity of tasks. In other words,
although many researchers have identified various
kinds of dominant factors that can make the performance of tasks complicated, a model that provides the
overall structure as well as the dependency among task
complexity factors (e.g., the three orthogonal dimensions) seems to be very rare. From this regard, it is
expected that this complexity model can be used as a
technical basis to resolve the multicollinearity problem
of the TACOM measure.

Figure 2.

Generalized task complexity model by Harvey.

Figure 3.

The overall structure of R-TACOM measure.

In the light of this expectation, five sub-measures

were reorganized along with the definition of three
dimensions included in the generalized task complexity model. Fig. 3 shows the reorganized structure of
TACOM measure (i.e., R-TACOM measure) (Park and
Jung 2007c).
4

COMPARING RESPONSE TIME DATA

WITH THE ASSOCIATED R-TACOM SCORE

In order to investigate the appropriateness of RTACOM measure, two sets of response time data
collected from nuclear power plants (NPPs) were compared with the associated R-TACOM scores. In the
case of emergency tasks included in the emergency
operating procedures (EOPs) of NPPs, a task performance time can be defined as an elapsed time from the
commencement of a given task to the accomplishment
of it. Regarding this, averaged task performance time
data about 18 emergency tasks were extracted from
the emergency training sessions of the reference NPP
(plant 1) (Park et al. 2005). In total 23 simulations
were conducted under steam generator tube rupture
(SGTR) conditions.
Similarly, averaged task performance time data
about 12 emergency tasks under SGTR conditions
were extracted from six emergency training sessions
of another reference NPP (plant 2). It is to be noted
that, although the nature of simulated scenario is very
similar, the emergency operating procedures of two
NPPs are quite different.
Fig. 4 represents the result of comparisons
between averaged task performance time data and the
associated TACOM as well as R-TACOM scores. For
the sake of convenience, equal weights were used to
quantify complexity scores ( = = = = = 0.2

223

Figure 4.

Comparing two sets of response time data with the associated R-TACOM scores.

in Fig. 1, and = = = 1/3 1 = 2 = 1 =

2 = 0.5 in Fig. 3).
In Fig. 4, the capability of the R-TACOM measure in evaluating the complexity of tasks is improved
because the value of R 2 about the R-TACOM measure is higher than that of the TACOM measure. This
strongly implies that the R-TACOM measure can be
used to standardize the performance of human operators who carry out emergency tasks. If we adopt
this expectation, then the R-TACOM measure can be
regarded as a good starting point to explain the difference of human performance due to operating culture.
Actually, it is expected that the difference of task performance time data in Fig. 4 (i.e., Plant 1 data are
higher with similar R-TACOM scores) may come from
a different operating culture.
5

GENERAL CONCLUSION

In this study, the appropriateness of R-TACOM measure based on the generalized task complexity model
was investigated by comparing two sets of averaged task performance time data with the associated
R-TACOM scores. As a result, it was observed that
response time data obtained when human operators
accomplished their tasks using different procedures
consistently increase in proportion to the increase of
R-TACOM scores. In other words, even though human
operators used different procedures, it is expected that
the performance of human operators would be similar if the complexity of tasks they are faced with is
similar.

Therefore, although more detailed studies are

indispensable to confirm the appropriateness of the
R-TACOM measure, the following conclusion can
be drawn based on the foregoing discussionsthe
R-TACOM measure is a proper measure in quantifying
the complexity of tasks prescribed in procedures.

REFERENCES
AIChE. 1994. Guidelines for preventing human error in process safety. Center for Chemical Process Safety of the
American Institute of Chemical Engineers.
Cohen, J., Cohen, P., West, S.G. and Aiken, L.S. 2003.
Applied multiple regression/correlation analysis for the
behavioral sciences. Lawrence Erlbaum Associates, Third
Edition.
Ghosh, S.T. and Apostolakis, G.E. Organizational contributions to nuclear power plant safety. Nuclear Engineering
and Technology, vol. 37, no. 3, pp. 207220.
Guimaraes, T., Martensson, N., Stahre, J. and Igbaria, M.
1999. Empirically testing the impact of manufacturing
system complexity on performance. International Journal
of Operations and Production Management, vol. 19, no.
12, pp. 12541269.
Harvey, C.M. and Koubek, R.J. 2000. Cognitive, social,
and environmental attributes of distributed engineering
collaboration: A review and proposed model of collaboration. Human Factors and Ergonomics in Manufacturing,
vol. 10, no. 4, pp. 369393.
Hollnagel, E. Human reliability assessment in context.
Nuclear Engineering and Technology, vol. 37, no. 2,
pp. 159166.
IAEA. 1985. Developments in the preparation of operating
procedures for emergency conditions of nuclear power

224

plants. International Atomic Energy Agency, IAEATECDOC-341.

Johannsen, G., Levis, A.H. and Stassen, H.G. 1994. Theoretical problems in man-machine systems and their
experimental validation. Automatica, vol. 30, no. 2,
pp. 217231.
Jung, W., Kim, J., Ha, J. and Yoon, W. 1999. Comparative evaluation of three cognitive error analysis methods
through an application to accident management tasks in
NPPs. Journal of the Korean Nuclear Society, vol. 31,
no. 6, pp. 822.
Melamed, S., Fried, Y. and Froom, P. 2001. The interactive
effect of chronic exposure to noise and job complexity on changes in blood pressure and job satisfaction:
A longitudinal study of industrial employees. Journal
of Occupational Health and Psychology, vol. 5, no. 3,
pp. 182195.
OHara, J.M., Higgins, J.C., Stubler, W.F. and Kramer, J.
2000. Computer-based Procedure Systems: Technical
Basis and Human Factors Review Guidance. US Nuclear
Regulatory Commission, NUREG/CR-6634.
Park, J. and Jung, W. 2007a. A study on the development of
a task complexity measure for emergency operating procedures of nuclear power plants. Reliability Engineering
and System Safety, vol. 92, no. 8, pp. 11021116.

Park, J. and Jung, W. 2007b. The appropriateness of TACOM

for a task complexity measure for emergency operating
procedures of nuclear power plantsa comparison with
OPAS scores. Annals of Nuclear Energy, vol. 34, no. 8,
pp. 670678.
Park, J. and Jung, W. 2007c. A Study on the Revision of
the TACOM Measure. IEEE Transactions on Nuclear
Science, vol. 54, no. 6, pp. 26662676.
Park, J., Jung, W., Kim, J. and Ha, J. 2005. Analysis of human
performance observed under simulated emergencies of
nuclear power plants. Korea Atomic Energy Research
Institute, KAERI/TR-2895/2005.
Stassen, H.G., Johannsen, G. and Moray, N. 1990. Internal representation, internal model, human performance
model and mental workload. Automatica, vol. 26, no. 4,
pp. 811820.
Topi, H., Valacich, J.S. and Hoffer, J.A. 2005. The effects
of task complexity and time availability limitations on
human performance in database query tasks. International Journal of Human-Computer Studies, vol. 62,
pp. 349379.
Yoshikawa, H. 2005. Human-machine interaction in nuclear
power plants. Nuclear Engineering and Technology,
vol. 37, no. 2, pp. 151158.

225

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

An evaluation of the Enhanced Bayesian THERP method

using simulator data
Kent Bladh
Vattenfall Power Consultant, Malm, Sweden

Jan-Erik Holmberg
VTT (Technical Research Centre of Finland), Espoo, Finland

Pekka Pyy
Teollisuuden Voima Oy, Helsinki, Finland

ABSTRACT: The Enhanced Bayesian THERP (Technique for Human Reliability Analysis) method has been
successfully used in real PSA-studies at Finnish and Swedish NPPs. The method offers a systematic approach
to qualitatively and quantitatively analyze operator actions. In order to better know its characteristics from a
more international perspective, it has been subject to evaluation within the framework of the HRA Methods
Empirical Study Using Simulator Data. This paper gives a brief overview of the method with major findings
from the evaluation work including identified strengths and potential weaknesses of the method. A number of
possible improvement areas have been identified and will be considered in future development of the method.

1
1.1

and help establish a methodology for assessing HRA

methods using simulator data.

INTRODUCTION
HRA as part of PSA

The modeling and quantification of human interactions is widely acknowledged as a challenging task
of probabilistic safety assessment (PSA). Methods for
human reliability analysis (HRA) are based on a systematic task analysis combined with a human error
probability quantification method. The quantification
typically relies on expert judgments, which have rarely
been validated by statistical data.
1.2

1.3

Scope

This paper presents an evaluation of the enhanced

Bayesian THERP in the pilot phase of the HRA Methods Empirical Study Using Simulator Data. Major
findings from this evaluation are summarized. Presentation of the outcomes of the international study is
outside the scope of this paper.

International study

In order to compare different HRA methods an international study HRA Methods Empirical Study Using
Simulator Data has been initiated using actual simulator data as reference for the comparison (Lois et al
2007 & Dang et al 2008). The overall goal of the international HRA method evaluation study is to develop an
empirically-based understanding of the performance,
strengths, and weaknesses of the HRA methods. It is
expected that the results of this work will provide the
technical basis for the development of improved HRA
guidance and, if necessary, improved HRA methods.
As a first step in the overall HRA method evaluation
study, a pilot study was conducted to obtain initial data

2
2.1

PILOT STUDY SET-UP

Overview

The pilot study is based on a set of simulator

experiments run in the Halden Reactor Projects
HAMMLAB (HAlden huMan-Machine LABoratory)
simulator facility. Fourteen operating crews from an
operating nuclear power plant (a pressurized water
reactor) participated in a series of performance shaping factor/masking experiments. Without knowledge
of the crews performances, several HRA analysis
teams from different countries, using different methods, performed predictive analyses of the scenarios.

227

2.2

Scenarios

In the pilot study, two variants of a steam generator tube rupture (SGTR) scenario were analyzed:
1) a basic case, i.e., a familiar/routinely practiced case,
and 2) a more challenging case so called complex case.
In the complex case, the SGTR was masked by a simultaneous steamline break and a failure of all secondary
radiation indications/alarms. It could be expected that
operators have difficulties to diagnose the SGTR. The
event sequence involves several operator actions, but
this paper is restricted to the first significant operator
action of the scenarios, i.e., isolation of the ruptured
steam generator (SG).
2.3

HRA analysis teams

In order to facilitate the human performance predictions, the organizers of the experiment prepared
an extensive information package for the HRA analysis teams including descriptions of the scenarios,
description of the simulator and its man-machine interface, differences between the simulator and the home
plant of the crews, procedures used in the simulator, characterization of the crews, their work practices
and training. The task of the HRA analysis teams
was to predict failure probabilities of operator actions
defined, e.g., isolation of the ruptured steam generator,
and to qualitatively assess which PSFs affect positively or negatively to success or failure of the crew.
The members of the Enhanced Bayesian THERP team
included the authors of this paper.
2.4

Time criterion

On the empirical side, time was used as the criterion

for defining success/failure of the crew performance.
In the SG isolation case, the available time for the
operator action was considered from the radiological
consequence point of view, not from a core damage
point of view. In order to avoid opening of a SG pressure relief valve, the crew should isolate the SG before
overfilling it.
2.5

ENHANCED BAYESIAN THERP

The Enhanced Bayesian THERP (Technique for

Human Reliability Analysis) method is based on the
use of the time-reliability curve introduced in the
Swains human reliability analysis (HRA) handbook
(Swain & Guttmann 1983) and on the adjustment of the
time-dependent human error probabilities with performance shaping factors (PSFs) (Pyy & Himanen
1996). The method is divided into a qualitative and
quantitative analysis part.
3.1

Qualitative analysis

The qualitative analysis consists of a modelling of the

scenario with a block diagram and a description of
the basic information of each operator action. The
purpose of the block diagram is to define the operator actions in relation to relevant process events. The
block diagram representation is close to a PSA event
tree but is usually a somewhat more detailed model
than an event tree. The block diagram can also be used
to present the dependencies between operator actions
belonging to the same scenario. The purpose of the
description of the basic information of each operator
action is to consistently characterize main aspects of
an operator action, e.g., initiating event, scenario, time
windows, support from procedures and MMI, practical
maneuvers needed in the action and other noteworthy
information.
The block diagram is also used to show the assumed
dependencies between operator actions belonging to
the same scenario. The blocks used in the diagram
should have exact correspondence with functional
events (in event trees or system fault trees) of the
PSA-model. This is important in cases where operator
action basic events are modeled in system fault trees
so that the link to event tree branches is not obvious.
In this study, the operator actions were given, so that
the construction of the block diagrams did not serve
as defining the operator action basic events.
3.2

Quantitative analysis

The human error probability is derived using the timedependent human error probability model as follows,

Performance shaping factors

The empirical identification of PSFs was based on

a detailed analysis of simulator performances. Analysts viewed the video und transcribed key communications and events, and used also additional data
sources, such as crew interview, crew PSF questionnaire, and observer comments. Finally, the analyst
summarized the observed episode in the form of an
operational story, highlighting performance characteristics, drivers, and key problems. A specific method
was used to rate the PSFs (Lois et al 2007).


p(t) = min 1, p0 (t)

5



Ki ,

(1)

i=1

where p0 (t) is the basic human error probability taken

from Swain & Guttmann 1983, see Figure 2, t is the
time available for identification and decision making,
and K1 , . . . , K5 are the performance shaping factors.
The min-function ensures that the final probability
stays within the range 0 to 1.

228

5 min

8-10 min to go through E -0 and enter E -3

5 min for step 3 of E-3

Steam generator
tube rupture
(SGTR )

Manual scram

Feedwater to
SGs
(auto-function)
P=0

Safety injection
to primary circuit
(auto-function)

Identification and
isolation of the
ruptured SG

Valves closed in
all outlet and
inlet paths of the
ruptured SG

P=0

Sequence
continues

P=0

Automatic scram
(on low
pressurizer
pressure )
P=0
Anticipated
transient without
scram

SG dry -out ,
major SG rupture

Loss of core
cooling

Unisolated SGTR ,
contamination of the
secondary side ,
loss of primary coolant

Figure 1. Block diagram of the beginning of the SGTR basic scenario. Light grey boxes are operator actions, white boxes
process events, and dark gray boxes end states of the event sequences. E-0 and E-3 are emergency operating procedures.
Possibility of technical failures is not considered in this analysis (P = 0).

1E+0

1E-1

Probability of failure

1E-2
Swain 95%

1E-3

Swain Median
Swain 5%

1E-4

Base probability

1E-5

1E-6

1E-7
1

10

100

1000

10000

Time for identification and decision making t [min]

Figure 2.

Time-dependent human error probability curve.

The time available for identification and decision

making is shorter than the total time available for the
operator action ttot which is assumed to be composed
of three parts as follows

where tind is time for first indication, t time for identification and decision making and tact time for action.
The following performance shaping factors are used
(Pyy & Himanen 1996):

ttot = tind + t + tact ,

K1 :
K2 :

(2)

229

Quality and relevance of procedures

Quality and relevance of training

K3 :
K4 :
K5 :

Quality and relevance of feedback from process (MMI)

Mental load in the situation
Need for coordination and communication.

4.1 Human error probabilities

Each performance shaping factor can receive a

value 1/5, 1/2, 1, 2 or 5. A level above 1 means that
the action has a complicating character compared to
a nominal situation. Consequently, a level below 1
means that the action is easier than the nominal case.
Level 1 means that the factor plays no major role or
that this factor is in a nominal level.
The meaning of each value for each PSF is
explained qualitatively in the method. For instance,
regarding Quality and relevance of procedures,
K1 = 1/5 is interpreted as Very good instructions,
operators should not make any mistake, K1 = 1/2 as
Good instructions, applicable for the situation and
they support well the selection of correct actions,
K1 = 1 as Instructions play no major role in the situation, K1 = 2 as Instructions are important but
they are imperfect, and K1 = 5 as No instructions or misleading instructions, instructions would
be needed. Explanations for the other PSFs are
analogous.
The performance shaping factors will be given
independently by a number of experts, and these judgments are consolidated with a Bayesian approach.
In this approach, the performance shaping factors
are assumed to be random variables following a
multinomial probability distribution,
P(Ki = j|qj ) = qj ,
q1/5 + + q5 = 1.

j = 1/5, 1/2, 1, 2, 5,

PREDICTIONS ACHIEVED BY USING

THE METHOD

(3)

The prior distribution for the parameters of the

multinomial distribution is assumed to be a Dirichlet
distribution. The convenient feature of Dirichlet distribution is that if we assume the expert judgments as
independent observations from a multinomial distribution, the posterior distribution is also Dirichlet and
can be easily derived. The prior distribution is chosen
by maximizing the entropy function. This distribution
has an interpretation to represent maximal uncertainty.
The mathematical procedure is presented in Holmberg
& Pyy 2000.
Four experts have participated in this exercise, and
made their assessments independently of each other
based on material obtained from Halden and processed
by VTT (see e.g. the block diagrams and definitions
for the operator actions). It should be observed that
experts normally include also members from the operation crews at the actual plant, which was not possible
during this experiment.

Table 1 summarizes numerical results from the HRA

made using the Enhanced Bayesian THERP method.
Time available for identification and decision making
was taken from the information package submitted by
Halden. The prior human error probability is derived
from the Swains curve, see Figure 2. Four experts
assessed independently the performance shaping factors and the assessments were aggregated using the
Bayesian procedure. Mean values (i.e., posterior mean
values) are shown in Table 1.
According to this analysis failure probability is in
the base case 0,03 and in the complex case much higher
0,2. In the simulator experiments, 1 out of 14 crews
failed to isolate the SG within the critical time window
in the base case, and in the complex case 7 out of
14 crews failed. Numerically, the predictions and the
outcome are well in balance.
4.2 Performance shaping factors
The values of the performance shaping factors can
be interpreted so that in the base scenario the crew
should get good support from procedures, training
and process feedback to identify the situation and to
make correct decision in time. Mental load is somewhat higher than in a normal case, and there is also
some coordination and communication needs related
to the action. The experts commented in free text
that good instructions, often training, clear indications, radiation alarm gives a clear indication of SGTR,
scram and shortage of time are likely to increase
mental load. These judgments were in accordance
with empirical PSF ratings except maybe the procedural guidance where some difficulties were found
empirically.
In the complex scenario, procedures are rated to
provide good support, but the situation is now considered unfamiliar from the training point of view,
and feedback from process is considered poor or misleading. Mental stress is considered higher than in
the base case. The experts commented in free text
that good instructions, often training, clear indications, radiation alarm gives a clear indication of SGTR,
scram and shortage of time are likely to increase mental load. good instructions, situation is unfamiliar,
less trained, normal feedback missing, and mental
load is high for various reasons. Empirically, it was
judged that the procedures do not provide good support. Otherwise predictions and observations were in
line. The difference in the PSF judgments can be seen
as an expert opinion issue, and not as an HRA method
issue.

230

Table 1.

Predictions for operator failure to isolate the ruptured steam generator in time.

Time available for identification and decision making

Prior human error probability
Performance shaping factors, mean values of the expert judgments2
Quality and relevance of procedures, scale 0.25
Quality and relevance of training, scale 0.25
Quality and relevance of feedback from process, scale 0.55
Mental load in the situation, scale 15
Need for coordination and communication, scale 0.55
Posterior human error probability
Uncertainty distribution (posterior)
5%
50%
95%

Base

Complex

12 min1
7.3E-2

15 min1
0.049

0.6
0.5
0.4
1.8
1.4
2.6E-2

0.7
1.4
2.4
2.7
1.2
1.7E-1

5.8E-4
7.3E-3
9.1E-2

9.9E-4
4.0E-2
1.0E + 03

Difference in time window between the base case and complex case is that, in the complex case, the plant trip is actuated
immediately in the beginning of the scenario, while in the base scenario an alarm is received first and the first indication of
increasing level in the ruptured SG is received 3 min after the alarm.
2 Interpretation of the scale: 0.2 = very good condition, 0.5 = good condition, 1 = normal, 2 = poor condition, 5 = very
poor condition. Note that full scale is not used for all PSFs.
3 Due to the min-function in the human error probability model, see formula (1).

5
5.1

results from this pilot study did not clarify the actual
need for calibration.

EVALUATION FINDINGS
Strengths

The enhanced Bayesian THERP method seems to be

a cost-effective approach for these type of standard
PSA operator actions when the aim is to model and
quantify operator actions for a PSA model.
As mentioned above, the predictions of the method
are well in line with the outcomes of the simulator
experiments. This is true for the quantification as well
as for most of the PSFs.
5.2

Potential weaknesses

Assessment of the time window is critical in this

method. It seems, however, that the time reliability
curve is quite reasonable at least for operator actions
with a time window between 530 min. It should
be noted that the time window in this case can be
defined quite accurately since the operator action is
the first one of the event sequence. There is much
more variability in time windows for the subsequent
actions, which is a challenge for this type of HRA
models.
Another critical point in the method is the interpretation of the performance shaping factors and their
numerical rating. It is obvious that different experts
will always interpret differently the explanations given
for the scaling. As long as an expert is consistent in
his/her judgments, values given for different operator actions can be compared. From the absolute level
point of view, some calibration may be needed. So far,

5.3

Possible improvement areas

A significant empirical observation was the variability

between crews with regard to affecting performance
shaping factors which means that PSFs are not only
action dependent but also crew dependent. This variability is not explicitly accounted in the enhanced
Bayesian THERP method, even though the method
produces a probability distribution for each PSF. These
probability distributions, however, reflect variability
of expert judgements not the variability of crews.
Method development may be needed to account the
variability of the crews.
Experts should be urged to justify the rates. This
is an essential way to collect insights, e.g., for
improvements of the human factors.
Another finding was that the method could be complemented with a discussion phase after the expert
judgements where experts could jointly comment the
results and draw conclusions from the assessments.
This would facilitate the interpretation of the results
which is now based on pure interpretation of the
numbers.

CONCLUSIONS

The experiment shows that the Enhanced Bayesian

THERP method gives results in close match with

231

simulator data, at least within the experimental

limitations.
Concerning the quantitative results, no significant
deviations were identified.
For negative PSFs, there was a difference in common for both scenarios. While the Enhanced Bayesian
THERP method predicted mental load/stress and deficient feedback as important factors, the simulation
focused more on procedural feedback and task complexity. The reasons behind this might be method
related, but could also depend on limitations in expert
selection and/or differences in stress level between real
operation and simulator runs.
The comparison of empirical observations with predictions was found as a useful exercise to identify areas
of improvements in the HRA method. An aspect not
covered by the method is the variability between the
crews with regard to importance of different PSFs.
Also explanations for numerical scales for PSFs could
be improved to harmonize the way experts interpret
the scales. In this way, empirical tests are necessary to
validate an HRA method.
Otherwise the evaluation gives confidence that the
time reliability curve is a feasible and cost effective
method to estimate human error probabilities, at least
when the time window is well defined and relatively
short.

REFERENCES
Lois, E. et al 2007. International HRA Empirical Study
Description of Overall Approach and First Pilot Results
from Comparing HRA Methods to Simulator Data. Report
HWR-844, OECD Halden Reactor Project, draft, limited
distribution.
Dang, V.N. et al 2008. Benchmarking HRA Methods Against
Simulator DataDesign and Organization of the Halden
Empirical Study. In: Proc. of the 9th International Conference on Probabilistic Safety Assessment and Management
(PSAM 9), Hong Kong, China.
Swain, A.D. & Guttmann H.E. 1983. Handbook of Human
Reliability Analysis with Emphasis on Nuclear Power
Plant Applications. NUREG/CR-1278, Sandia National
Laboratories, Albuquerque, USA, 554 p.
Pyy, P. & Himanen R. 1996. A Praxis Oriented Approach
for Plant Specific Human Reliability AnalysisFinnish
Experience from Olkiluoto NPP. In: Cacciabue, P.C., and
Papazoglou, I.A. (eds.), Proc. of the Probabilistic Safety
Assessment and Management 96 ESREL96PSAMIII
Conference, Crete. Springer Verlag, London, pp.
882887.
Holmberg, J. & Pyy, P. 2000. An expert judgement based
method for human reliability analysis of Forsmark 1
and 2 probabilistic safety assessment. In: Kondo, S. &
Furuta, K. (eds.), Proc. of the 5th International Conference on Probabilistic Safety Assessment and Management
(PSAM 5), Osaka, JP. Vol. 2/4. Universal Academy Press,
Tokyo, pp. 797802.

232

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Comparing CESA-Q human reliability analysis with evidence

from simulator: A first attempt
L. Podofillini
Paul Scherrer InstitutPSI, Villigen PSI, Switzerland

B. Reer
Swiss Federal Nuclear Safety InspectorateHSK, Villigen HSK, Switzerland
(until July 2007 with the Paul Scherrer Institut)

ABSTRACT: The International Human Reliability Analysis (HRA) Empirical Study started in late 2006, with
the objective of assessing HRA methods based on comparing their results with data. The focus of the initial phase
is to establish the methodology in a pilot study. In the study, the outcomes predicted in the analyses of HRA teams
are compared with the findings obtained in a specific set of simulator studies. This paper presents the results of
one HRA analysis team and discusses how the predicted analysis compares to the observed outcomes from the
simulator facility. The HRA method used is the quantification module of the Commission Errors Search and
Assessment method (CESA-Q), developed within the HRA research project at the Paul Scherrer Institut (PSI). In
this pilot phase, the main focus of the comparison is on qualitative results: the method predictions, the scenario
features and performance factors that would mostly contribute to failure (or support success). The CESA-Q
predictions compare well with the simulator outcomes. This result, although preliminary, is encouraging since
it gives a first indication of the solidity of the method and on its capability of producing founded insights for
error reduction. Also, the comparison with empirical data provided input to improve the method, regarding
the treatment of the time factor and of knowledge- and training-based decisions. The next phases of the HRA
Empirical Study will address also the quantitative aspects of the HRA. It is planned to use further insights from
the next phases to a) refine the CESA-Q guidance and b) evaluate the method to see whether additional factors
need to be included.

INTRODUCTION

The aim of the Human Reliability Analysis (HRA)

empirical study is to assess the quality of HRA methods based on simulator data (Dang et al., 2007; Lois
et al. 2008).
Fourteen operating crews from a nuclear power
plant participated in a series of scenarios in the
Halden Reactor Projects HAMMLAB (HAlden
huMan-Machine LABoratory) simulator facility, in
Halden, Norway. Without knowledge of the crews
performances, HRA analysis teams performed predictive analyses of the scenarios.
The first pilot phase of the project run until beginning of 2008, with the aim of establishing the methodology for comparing the predictions of the HRA teams
with the experimental outcomes of the simulator runs.
The comparison addressed two Human Failure Events
(HFEs), out of the nine on which the HRA predictive analyses were made. The focus in this phase of
the project is on the qualitative results of the HRA,

i.e. their predictions, identification of the scenario

features and performance factors that would most contribute to the failures of these actions (or support
success). The results of this first pilot phase will be
published in a forthcoming report (Lois et al., 2008).
The next phases of the project will address the
remaining HFEs (as well as additional HFEs) and
the quantitative results of the HRA, the values of the
human error probabilities (HEPs).
This paper presents the results of one HRA analysis
team and discusses how the predicted analysis compares to the observed outcomes from the simulator
facility. The focus is on the qualitative HRA results.
The HRA method used is the quantification module
(CESA-Q) (Reer, 2006a) of the Commission Errors
Search and Assessment method (CESA), (Reer et al.,
2004; Reer & Dang, 2007).
The application of CESA-Q in this study has been
exploratory. The methods development and previous
applications have focused on errors of commission,
while this study addresses errors of omissions, i.e. the

233

non-performance of required actions. CESA-Q is still

under development. Although preliminary, the evaluation of the method against empirical data has been
very informative at this stage and provided insights
(minor, but worthwhile) for further improvement.
There are still methodological aspects to be resolved
on how to compare an HRA analysis with evidence
from the simulator. Together with the HRA teams, the
steering committee of the international HRA empirical
study is currently working on resolving these.
The paper is organized as follows. Section 2 gives
more details on the HRA empirical study. Section 3
presents CESA-Q as it has been applied in the study,
describing also how the method was adjusted to
address the EOOs. Section 4 describes the HFEs
addressed in the comparison. The CESA-Q HRA of
the considered HFE is presented in Section 5. Section 6
presents the comparison between the method predictions and the outcome from the simulator runs. Derived
insights for improvement of CESA-Q are discussed in
Section 7. Conclusions are given at closure.
2

THE HRA EMPIRICAL STUDY

The motivations for the Empirical Study are the differences in the scope, approach, and models underlying
the diversity of established and more recent HRA
methods (Dang et al., 2007; Lois et al., 2008). These
differences have led to a significant interest in assessing the performance of HRA methods. As an initial
step in this direction, this international study has been
organized to examine the methods in light of data,
aiming to develop an empirically-based understanding
of their performance, strengths, and weaknesses. The
focus of the study is to compare the findings obtained
in a specific set of simulator studies with the outcomes
predicted in HRA analyses.
Hosted by the OECD Halden Reactor Project, the
Empirical Study has three major elements:
predictive analyses where HRA methods are applied
to analyze the human actions in a set of defined
scenarios,
the collection and analysis of data on the performance of a set of operator crews responding to
these scenarios in a simulator facility (the Hammlab
experimental simulator in Halden),
and the comparison of the HRA results on predicted
difficulties and driving factors with the difficulties
and factors found in the observed performances.
The tasks performed in 2007 aimed a) to establish
the methodology for the comparison, e.g. the protocols for interacting with the HRA analyst teams, the
information exchanged, and the methods for the data
analysis and comparison; and b) to test the comparison
methodology with expert teams submitting predictive

HRA analyses for evaluation against the data. In this

way, initial results concerning the HRA methods as
well as feedback on the comparison methodology itself
were obtained.
In this phase of the Empirical Study, two Steam
Generator Tube Rupture scenarios were defined,
specifically a straightforward or base scenario and
a more difficult or complex scenario. The base scenario includes four main operator actions while the
complex scenario included five operator actions.
The HRA analysis teams performed predictive analyses with their chosen HRA method on the basis of
reference inputs (an information package) prepared
by the assessment group. Further details on the overall
study methodology are presented in (Dang et al., 2007;
Lois et al., 2008) and in other related papers from this
conference.
In the pilot phase, the HRA analysis teams analyzed the failure of the nine actions or human failure
events. The qualitative results of their analyses,
identify the scenario features and performance factors that would most contribute to the failures of
these actions (or support success); these make up
the teams predicted outcomes. Their quantitative
results are the estimated human error probabilities. On
the empirical side, the Halden staff and the Studys
assessment group analyzed the data collected on the
crews performance in these scenarios to identify the
scenario features and factors that were observed or
inferred to cause difficulties for the operators, leading to delays in completing the actions or to failures
to complete the actions in time. At the qualitative
level, the predicted features and factors from each
HRA analysis of an action are compared with the
features and factors observed in the data for that
action.
In view of the ambitious schedule and the desire to
obtain feedback from the HRA analyst teams early in
the study, the comparison in this first pilot phase was
limited to two actions defined in the scenarios, the
identification and isolation of the faulted steam generator in the base and complex scenarios respectively.
For each action, the prediction results are compared
with the results from the observations.
The simulator studies with the operator crews were
carried out in late 2006. To avoid biasing the comparison, the assessment group and HRA teams were
not provided information of any kind on the simulator
observations until after the review of the HRA team
submissions were completed.
The second pilot phase planned for 2008 will
address the remaining actions in the two SGTR scenarios and include the comparison in quantitative terms,
in other words, it will address how well the HEPs
estimated by an HRA method correlate with the level
of difficulty observed in the empirical data for these
actions.

234

THE CESA-Q METHOD AS APPLIED

IN THE HRA EMPIRICAL STUDY

Table 1. Steps of the CESA-Q analysis as applied in the

HRA empirical study.

The initial development of PSIs CESA method

focused on the identification and prioritization of
aggravating operator actions in post initiators scenarios, i.e. errors of commission (EOCs). CESA is
described in (Dang et al., 2002; Reer et al., 2004;
Reer & Dang, 2007). Subsequently, work was started
to address quantification, again focusing on EOCs,
resulting in an outline of a method for EOC quantification: CESA-Q (Reer, 2006a). The quantification
module CESA-Q was applied in the HRA empirical study. Identification and prioritization, addressed
by CESA, were not relevant for the empirical study
since the HFEs were given as part of the study
design.
Some features of CESA-Q are as follows (refer
to Reer (2006a) for a complete description of the
method). The EOC is analyzed in terms of the plantand scenario-specific factors that may motivate inappropriate decisions. The focus of CESA-Q is therefore
in the analysis of inappropriate decision-making. As
discussed later, this focus makes the application of
CESA-Q to EOOs plausible. Two groups of factors
are introduced: situational factors, which identify
EOC-motivating contexts, and adjustment factors,
which refine the analysis of EOCs to estimate how
strong the motivating context is. A reliability index
is introduced which represents the overall belief of the
analyst regarding the positive or negative effects on the
EOC probability (defined from 0 for strongly errorforcing contexts to 9 for contexts with very low EOC
probabilities). Quantification is done by comparing
the pattern of the factors evaluations with patterns
of catalogued reference EOCs (identified from 26
operational events, previously analyzed qualitatively
and quantitatively in Reer and Dang (2006) and Reer
(2006b).
The HFEs addressed in the HRA empirical study are
EOOs, while the development of CESA-Q and its previous applications have focused on EOCs. Therefore,
the application of CESA-Q has been explorative.
In particular, the EOOs (i.e. the non performance
of the required action) were interpreted as the final
result of inappropriate actions or decisions made by
the crew while proceeding along their path to the
required action, with the result that the required action
is not performed. Therefore, CESA-Q could be used
to analyze these inappropriate decisions.
The HFEs analysis by CESA-Q was based on the
identification of the decision points encountered by
the operators while proceeding with their response.
Then, the decision points were analyzed to determine
if inappropriate decisions made by the crew at the decision points could result in the required action not being
performed within the required time window (i.e. in the

Step
#
1

4
5

Description
List the decision points that introduce options for
deviating from the path (or set of paths) leading
to the appropriate response, and select relevant
deviations, contributing to the human failure
event (HFE) in question.
For each decision point, evaluate whether a
situational factor (Reer, 2006a) motivates the
inappropriate response (deviation from the
success path). If not proceed with step 3. If yes
proceed with step 4.
If no situational factor applies, estimate a reliability
index (i) in the range from 5 to 9. Guidance is
provided in (Reer, 2007).
Proceed with step 6.
For the EFC case, evaluate the adjustment factors
which mediate the impact.
Refer to the guidance presented in (Reer, 2006a).
For the EFC case, estimate a reliability index (i)
in the range from 0 to 4.
Refer to the reference cases (from operational
events) summarized in (Reer 2006a).
a) Assign an HEP (pF|i value) to each decision
point (CESA-Q associates each reliability index
with an HEP value, Reer 2006a), and b) determine
the overall HEP (as the Boolean sum
of the individual HEPs).
Evaluate recovery (the option for returning to the
correct response path) for the most likely error
assigned in step 6a. Apply the recovery HEP
assessment guidance in Reer (2007).

EOO). Table 1 gives details on all steps of the CESA-Q

analysis.
For example, the decision by the crew to transfer to
an inappropriate procedure (where the required action
is not instructed) may result in the required action not
being performed, or, if the decision is revised too late,
not being performed in the available time.
4

HFE SCENARIO DESCRIPTION

AND SUCCESS CRITERION

The results reported in this paper relate to one of the

two HFEs addressed in the first pilot phase: HFE #1B,
the failure to identify and isolate the ruptured SG in
the complex SGTR scenario.
The complex SGTR scenario is designed such that
important cues for the diagnosis of the SGTR event,
which is needed for HFE #1B, are missing. In the
HAMMLAB simulator, this was accomplished by
defining hardware failures additional to the SGTR. In
particular, a main steam line break, SLB, precedes the

235

SGTR and leads to closure of the main steam isolation

valves, with the consequence (in the reference plant)
that most of the secondary radiation indications read
normal. The remaining available secondary radiation
indications is also failed as part of the scenario design.
The Emergency Operating Procedures (EOPs)
guiding the crews are of the Westinghouse type. The
immediate response is guided by procedure E-0. The
response to the SGTR event is guided by procedure E-3. There are a number of opportunities to
enter E-3:
step 19 of E-0, with instructions based on secondary
radiation indications;
first transfer to ES-1.1 (Safety injection termination) at step 21 of E-0 and then transfer to E-3
based on the ES-1.1 fold-out page, based on the
check in the fold-out of secondary radiation conditions and whether level in any intact SG is
increasing in an uncontrolled manner;
step 24b in E-0, based on whether the level in SG
#1 cannot be desirably controlled; or
step 25a in E-0, based on secondary radiation
indications.
The challenge connected with HFE#1B is the lack
of secondary radiation cues, which would be expected
in case of a SGTR and on which a number of the
Table 2.

procedural transfers to the applicable EOP E-3 are

based. Indeed, the applicable EOPs for the reference plant rely strongly on radiation indications. It
should be noted that this may not be the case in other
plants. Therefore the operators have to diagnose the
SGTR event based only on the level indications in the
ruptured SG.
Success in performing #1B requires that the crew:
enters procedure E-3 (based on the various opportunities to enter E-3), and
has closed/isolated all steam outlet paths from the
ruptured SG (SG #1), and
stopped all feed to the ruptured SG as long as the
ruptured SG level is at least 10% as indicated on the
narrow range SG level indications (to ensure the SG
U-tubes will remain covered).
perform the above by 25 minutes once the steamline
break occurs (which is the start of the event) constitutes failure as this would be a slower response
than expected/desired.

THE CESA-Q ANALYSIS OF HFE #1B

This Section reports the qualitative aspects of the

HRA. As mentioned, these are the aspects on which the

Critical decision points (selected) identified in the CESA-Q analysis of HFE #1B.

Decision point

Description of the inappropriate

decision

Consequence

#1B.1 (E-0, step 1)

Operators erroneously transfer from

E-0 step 1 to procedure FR-S.1 response
to nuclear power generation/ATWS.

Potential delay (not within 25 min) to accomplish

#1B. If error is not recovered, operator could go
as far as to inject boron, but this action has
consequences on the safe side.

#1B.4 (E-0, step 16)

Operators erroneously transfer

from E-0, step 16 to E-1 Loss of
reactor or secondary coolant.

Most likely consequence is potential delay (not

within 25 min) to accomplish #1B. In addition,
depending on how far operators go into E-1 they
may risk overfeeding the RCS or the SG.

#1B.6 (E-0, step 18)

Operators erroneously transfer

from E-0, step 18 to E-2 Isolation
of faulted SG.

Potential delay (not within 25 min) to accomplish

#1B.

#1B.7 (E-0, step 21,

and fold out page
of ES-1.1)

Operators erroneously do not transfer to E-3

(transferring from E-0, step 21 to ES-1.1
first and then from the fold out page
of ES-1.1 to E-3).

Operators stay in ES-1.1 and become involved in the

steps required to control SI. Potential delay to
accomplish #1B, i.e. not within 25 min.

#1B.8 (E-3, steps 2

and 3)

Operators fail to identify and isolate the

ruptured SG (E-3 step 3).

Primary to secondary leak is not controlled.

#1B.12 (E-3,
step 3b)

Operators isolate PORV of ruptured SG

when SG erroneously perceived below
70.7 bar).

Isolation of SG A PORV may lead to challenge of the

SG A safety valves which may stick open and
thus disable isolation of SG A. It is assumed here
that the SG safety valves are not qualified for
water discharge and thus certainly fail under SG
fill-up conditions.

236

Table 3. Analysis of the situational factors for decision

point #1B.6 at step 18 of E-0Operators erroneously transfer
to procedure E-2 (instructing isolation of faulted SG).
Situational
factor

Evaluation
Comment

Misleading
No
indication
or instruction (MI)

Adverse
exception
(AE)

Table 4. Analysis of the situational factors for decision

point #1B.7Operators do not transfer to E-3 (first from
E-0, step 21 to ES-1.1 and then from fold-out page of ES-1.1
to E-3).

No

Indications are not misleading:

pressure in the SG is not
decreasing in uncontrollable
manner as required in
step 18b for transfer to E-2.
Procedures do not support
transfer to E-2 in this scenario.
Transfer to E-2 is inadequate for
this scenario. It is not made
inadequate by an exceptional
condition.

Adverse
Yes
distraction
(AD)

There is an initial drop in all SG

pressures due to the main SLB.
Although pressure recovers very
fast upon main SL isolation,
operators may fix on this initial
cue, quickly reach step 18 and
enter E-2.

Risky
incentive
(RI)

There is no credible reason to

transfer to E-2 in order to follow
conflicting goals.

No

current phase of the HRA empirical study is focused

on. The steps of the analysis related to quantification
have been skipped in the following description.
5.1.1 Step 1list decision points
Procedures were analyzed in order to identify the
decision points that may contribute to HFE #1B.
13 decision points were found, identified as #1B.1,
#1B.2, . . . , #1B.13. Selected decision points are
reported in Table 2. The consequence of the inappropriate decisions can be a delay in performing #1B
(compared to the time window of 25 minutes) as well
as aggravation of the plant condition.
5.1.2 Step 2evaluate the situational factors
The 13 decision points were analyzed in terms of
the four CESA-Q situational factors (Reer, 2006a),
which may motivate an inappropriate response. Potential error-forcing conditions (EFCs) were identified at
two decision points, #1B.6 and #1B.7 (Table 2), the
analyses of which are reported in Table 3 and Table 4,
respectively.
5.1.3 Step 4evaluate adjustment factors
The decision points #1B.6 and #1B.7, for which error
forcing conditions were found, were further analyzed
in terms of adjustment factors (Reer, 2006). Table 5

Situational
factor

Evaluation

Misleading
indication
or Instruction (MI)

Yes

Cues on high secondary radiation

levels are missing. This is
a consequence of the exceptional
condition of SLB and SGTR
combined.

Adverse
exception
(AE)

Yes

The exceptional condition of

combined SLB and SGTR
results in the lost of important
cues: high secondary radiation
levels. These are the first cues
mentioned in E-0 as indications
of a SGTR (E-0, step 19). Cues
related to differences in the SG
levels come up later than expected for a normal SGTR. Under
this error-forcing condition, the
operators are expected to enter
ES-1.1. At this point it is possible
that they get involved in terminating SI and overlook the transfer
E-3 in the foldout page of ES-1.1.

Adverse
distraction
(AD)

Yes

Lack of relevant cues (high radiation levels in the secondary) is a

distraction for the operators.

Risky
incentive
(RI)

No

Operators would not miss the

transfer to E-3 in order to follow
conflicting goals.

Comment

reports the analysis for decision point #1B.7, which,

according to the analysis, dominates #1B.6. Note that
dominating decision points result from the quantitative
analysis, which is not reported in this paper.
5.1.4 Step 7evaluate recovery
Recovery analysis is carried out for the most likely
decision errors identified in the previous step 6, i.e.
#1B.6 (E-0 step 18) and #1B.7 (E-3 step 2, step 3); see
Table 6.
6

COMPARISON OF THE CESA-Q ANALYSIS

TO OPERATING CREW DATA

The CESA-Q analysis of HFE 1#B was submitted

by the HRA team to the assessment and comparison
group. The assessment and comparison group compared the HRA predictions to the simulator outcomes.
This Section highlights some points of the comparison,
which were used to derive the insights discussed in
Section 7. The detailed comparison will be published
in the forthcoming study report (Lois et al., 2008).

237

Table 5. Analysis of the adjustment factors for decision

point #1B.7Operators erroneously do not transfer to E-3
(transferring from E-0, step 21 to ES-1.1 first and then from
the fold out page of ES-1.1 to E-3).
Adjustment
factor
Evaluation

Table 6. Recovery analysis for decision point

#1B.7Operators do not transfer toE-3 (first from E-0, step
21 to ES-1.1 and then from fold-out page of ES-1.1 to E-3).
Recovery
factor

Evaluation

RTP
Recovery
Timely
Possible

Yes

A time reserve of 5 min is defined

as a boundary condition for the
HRA. This reserve is deemed
as sufficient for returning to the
appropriate path.

RCA
Recovery
Cue
Available

Yes

Step 24b in E-0 provides

alternative path from E-0 to E-3,
with cues based on the SG level.

ST
Shortage
of Time

Yes

Time reserve of 5 min (HRA

boundary condition) indicates
shortage of time.

MC
Masked
Cue

Probably

Although the main indication is

masked, difference in the SGs
level becomes progressively
evident.

Comment

Comment

Verification 0.8 (slightly

The foldout page of
hint
error-forcing)
ES-1.1 gives indication that the adequate
to transfer to E-3. The
rate response is of 0.8
it is not known how has
frequently the operators
the foldout page.
Verification 1 (success
Level indications are
means
forcing)
available and clearly
visible.
Verification 0.8 (slightly
Cognitive requirement been
difficulty
error-forcing)
slightly increased, since
lack of radiation indications the given since
alrepresents a deviation
from the base case of
trained rule application.
In addition, the level
indications are available.
with some delay
compared to the
expectation.
Verification 1 (success
Negligible physical effort
effort
forcing)
required for verification.
Time
0.5 (modTime for taking decision
pressure
erately error
is around 10 minutes.
forcing)
Benefit
1 (success
No particular benefit
prospect
forcing)
to stay in ES-1.1 in this
scenario.
Damage
0 (not success
No particular damage
potential
forcing)
potential is implied to
stay in ES-1.1.

As a general statement, this prediction was well in

line with the observations. Indeed, the dominant crew
behaviors were as follows.
Six crews entered ES-1.1, thus passing the transfer
to the SGTR procedure at E-0 step 19. All crews
eventually transferred to E-3, based on the instructions in the fold-out page in ES-1.1, or on their
knowledge that increasing level in one SG provides
a cue for SGTR (although half of the crews did not
do so in the required time of 25 minutes).
Five crews transfer directly from E-0 to E-3, without passing through ES-1.1. This transfer was
knowledge-based as well.

According to the CESA-Q predictions, the way the

human failure would most likely develop is as follows.
As the result of the error-forcing condition of missing
radiation indications and (probably) delayed SG levels
indications, the operators are expected to pass the EOP
step transferring to the SGTR EOP (E-0, step 19) and
enter ES-1.1, as instructed by later steps of E-0 (step
21). At this point it is possible that they get involved
in performing the necessary steps to terminate safety
injection as instructed by ES-1.1 and overlook the
transfer to E-3 in the foldout page of ES-1.1. The
decision to transfer to E-3 in this case is not straightforward due to the EFCs. However, it is expected that
they would at some point transfer to E-3, realizing
that the increasing level in one SG is a cue for SGTR
(procedures have additional later transfers to E-3).

The CESA-Q analysis predicted that the crews

would have eventually managed to transfer to E-3
using the fold-out page of ES-1.1. Actually, in the
observations, 2 crews did so, while many of the crews
that entered ES-1.1 decided to transfer to E-3 from
knowledge-based diagnosis. Indeed, as predicted by
the CESA-Q analysis, there was no particular benefit
to stay in ES-1.1 too long in this scenario.
The CESA-Q analysis predicted that some shortage of time may have been experienced by the crews if
they eventually enter ES-1.1 and become involved in
the ES-1.1 procedural steps. This would result in a difficulty for the crews to meet the time requirement of 25
minutes. Indeed, the observations from the simulator
confirmed that the time of 25 minutes available for the
response was limited and this had an important impact
on the crews performance. Although all of the 14
crews managed to enter the SGTR procedure E-3 and

238

isolate the ruptured SG, 7 out of the 14 crews did not do

so within the 25 minutes (but 13 crews did it within 35
minutes and the last one did it after 45 minutes). Compared to the simulator outcome, the CESA-Q analysis
did not recognize that the situation would slow the
crews response such that as many as half of the crews
did not meet the time requirement of 25 minutes (see
success criteria in Section 4). The implication of this
will be discussed in the next Section 7.1.
However, it must be noted that the 25 minutes
criterion is not a typical PSA success criterion for
identification and isolation of the ruptured SG in a
SGTR event. A typical PSA success criterion would
be to respond in time to avoid SG overfill or to avoid
damage of the SG safety valves due to flooding of
the steam line. Although 25 minutes is about the time
after which it is expected that the SG level will reach
100% on the wide range indicators, the operators are
aware that still some time is left before the SG overfills. Indeed in their response, the goal the operators
have in mind is to avoid or limit overfill rather than
respond in 25 minutes (indeed, except for one, all the
other crews were just late up to 10 minutes).

7
7.1

INSIGHTS FOR IMPROVEMENT OF CESA-Q

FROM THE COMPARISON
Influence of the time available on the HEP

CESA-Q treats the time factor by focusing on the

effect of time pressure, intended as urgency to act,
on the quality of decision-making, and accounts for
shortage of time in decision error recovery. Time
pressure impacts the ability of the operators to think
straight about alternative decisions and to possibly
revise an inappropriate decision. In CESA-Q, these
aspects enter into the evaluations of the situational
factors Adverse distraction and Risky incentive,
in the adjustment factor Time pressure and in the
evaluation of the time available for recovery.
The evidence from the simulator showed that the
failure in HFE #1B was influenced by the adequacy
of time, among other factors, rather then by time pressure. The time frame of 25 minutes resulted to be short
to reach consensus (or for the crew leader to reach
enough initiative) to enter E-3 in the complex scenario.
But there was no evidence that the crews performance
was negatively influenced by time pressure.
Therefore, it seems that the method, in its current
version, does not give proper credit to the effect of
running out of time while making correct decisions
as guided by the procedure, which seems to be one of
the drivers for HFE 1B. Indeed, CESA-Q was developed based on the analysis of catalogued reference
EOCs (identified from 26 operational events in Reer
& Dang, 2006), where the success of the operators

entailed to re-establish safety functions and running

out of time was not a concern.
It is planned to use the successive phases of the HRA
empirical study to gather additional insights on how
CESA-Q should address the adequacy of time factor
and to provide guidance to the users as necessary.
7.2

Focus of CESA-Q on aspects

of knowledge-based behavior required
for EOC quantification

CESA-Q accounts for operator behavior based on

knowledge (or training) in the following factors:
in the evaluation of EFCs represented by situational
factors AD and RI. For example, concerning AD:
the distraction caused by an indication not referred
to in the nominal path through the procedures may
suggest a response (inappropriate), followed by
knowledge or training; and
in the evaluation of adjustment factors (Table 5) and
recovery (Table 6), e.g. regardless of the guidance
in the procedure, an abnormal SG level indication
may be credited as a hint to verify if an evaluation
of operator training concludes that the SG level is
in the focus of the operators attention.
A thorough analysis of operational events with
EOC involved has shown that account of these aspects
of knowledge-based behavior is required for EOC
quantification (Reer & Dang, 2006).
However, some of the experimental results from this
HRA empirical study suggest that additional consideration of knowledge-based behavior may be required,
especially for EOO quantification. The observation
of the crews behaviors has shown that, especially in
the case in which the EOP guidance is not optimal
like in the HFE#1B case, knowledge-based as well as
training-based decisions become important drivers for
successful performance, therefore for not committing
the EOC or the EOO. Indeed, for many of the crews,
the decision to transfer to E-3 was made based on their
knowledge of the SGTR symptoms, after realizing that
the EOPs were not conducting them to E-3.
Guidance may therefore be helpful on when the
CESA-Q analysis should extend the consideration
of knowledge-based and training-based decisions in
the scope of the analysis. Indeed, it should keep
into account that how much the crews adhere to the
procedures or integrate them with knowledge-based
and training-based decisions may vary depending on
plant-specific or country-specific work processes.
Consequently, also guidance should be included on
when the analysis should consider multiple success
paths, for example based on the likelihood of the crews
taking the different paths and in the identification of
the single, or eventually multiple, dominant path.

239

CONCLUSIONS

This paper has concerned the results of the HRA performed by the CESA-Q team on one of the nine HFEs
addressed by the HRA empirical study. The paper has
discussed how the predicted analysis compares to the
observed outcomes from the simulator facility.
It should be emphasized that the application of
CESA-Q in this study is explorative: the methods
development and previous applications have focused
on errors of commission, while this study addresses
errors of omissions.
PSIs CESA-Q method performed well on the qualitative aspects of the exercise, i.e. how well the
methods predicted what elements of the actions may
be challenging. This result, although preliminary, is
encouraging since it gives a first indication of the solidity of the method and on its capability of producing
founded insights for error reduction. These qualitative
aspects were the main emphasis in this phase of the
study; currently, the assessment group is planning to
treat the more quantitative aspects in the next phase.
The empirical data, consisting of systematic observations of the performances of multiple crews on the
same scenarios, have been useful in deriving insights
on potential improvements of the CESA-Q method.
For instance, in treating time, CESA-Q focuses on
the effect of time pressure on the quality of decisionmaking and accounts for shortage of time in decision
error recovery. It seems that the method, in its current
version, does not give proper credit to the effect of
running out of time while making correct decisions
as guided by the procedure.
It should be also investigated whether guidance
should be added to base the analysis on multiple
expected response paths and to consider knowledgebased and training-based decisions in the definition of
the expected response paths and of the critical decision
points.
An aspect that makes CESA-Q well suited for
comparison against simulator data is that it produces
detailed descriptions of crews behaviors, in the form
of paths of response actions and critical decisions
taken along the response. These paths and decisions
could be indeed observed in the simulators. CESA and
CESA-Q shares this characteristic with some other
recent HRA methods like EDFs MERMOS and US
NRCs ATHEANA.
Finally, the study can provide an opportunity not
only to compare CESAs predictions with empirical
data but also to compare HRA methods and their
resulting analyses on the same set of actions. In particular, when performed in the context of empirical data,
a method comparison has the added value that there is
a shared basis (the data) for understanding the scope
of each factor considered by a method and how the
method treats these in detail.

The next steps planned for the HRA empirical study

are the comparison of the remaining HFEs in the Steam
Generator Tube Rupture scenarios and another comparative analysis on a Loss of Feedwater scenario.
Generally, it is planned to use the insights to a) refine
the CESA-Q guidance and b) evaluate the method to
see whether additional factors need to be included. In
this regard, empirical data provide invaluable input.
Although an empirical model of performance needs to
be based on far more than one scenario (two variants in
this case), this data contributes to such a model. This
should lead to improvements in CESA-Q as well as
other HRA methods.

ACKNOWLEDGEMENTS
This work is funded by the Swiss Nuclear Safety
Inspectorate (HSK), under DIS-Vertrag Nr. 82610.
The views expressed in this article are solely those
of the authors.
B. Reer contributed to this work mostly while he
was with the Paul Scherrer Institut (since July 2007,
he is with the HSK).

LIST OF ACRONYMS
CESACommission Errors Search and Assessment
CESA-QQuantification module of CESA
EFCError Forcing Condition
EOPEmergency Operating Procedures
HAMMLABHalden huMan-Machine LABoratory
HFEHuman Failure Event
HRAHuman Reliability Analysis
MSLBMain Steam Line Break
PSFPerformance Shaping Factor
PSAProbabilistic Safety Assessment
PSIPaul Scherrer Institut
SGTRSteam Generator Tube Rupture

REFERENCES
Dang, V.N., Bye, A. 2007. Evaluating HRA Methods in Light
of Simulator Findings: Study Overview and Issues for
an Empirical Test. Proc. Man-Technology Organization
Sessions, Enlarged Halden Programme Group (EHPG)
Meeting, HPR-367, Vol. 1, paper C2.1, 1116 March
2007, Storefjell, Norway.
Dang, V.N., Bye, A., Lois, E., Forester, J., Kolaczkowski,
A.M., Braarud, P.. 2007. An Empirical Study of HRA
MethodsOverall Design and Issues Proc. 2007 8th
IEEE Conference on Human Factors and Power Plants
(8th HFPP). Monterey, CA, USA, 2631 Aug 2007,
CD-ROM, (ISBN: 978-1-4244-0306-6).
Dang, V.N., Reer, B., Hirschberg, S. 2002. Analyzing Errors
of Commission: Identification and First Assessment for a

240

Swiss Plant. Building the New HRA: Errors of Commissionfrom Research to Application, NEA OECD report
NEA/CSNI/R(2002)3, 105116.
Forester, J., Kolaczkowski, A., Dang, V.N., Lois, E. 2007.
Human Reliability Analysis (HRA) in the Context of HRA
Testing with Empirical Data. Proc. 2007 8th IEEE Conference on Human Factors and Power Plants (8th HFPP).
Monterey, CA, USA, 2631 Aug 2007. CD-ROM, (ISBN:
978-1-4244-0306-6).
Lois, E., Dang, V.N., Forester, J., Broberg, H., Massaiu, S.,
Hildebrandt, M., Braarud, P.., Parry, G., Julius, J., Boring, R., Mnnist, I., Bye, A. 2008. International HRA
empirical studydescription of overall approach and first
pilot results from comparing HRA methods to simulator
data. HWR-844. OECD Halden Reactor Project, Norway
(forthcoming also as US NRC report).
Reer, B., 2006a. Outline of a Method for Quantifying Errors
of Commission. Paul Scherrer Institute, Villigen PSI,
Switzerland, Draft.

Reer, B. 2006b. An Approach for Ranking EOC Situations

Based on Situational Factors. Paul Scherrer Institute,
Villigen PSI, Switzerland, Draft.
Reer, B. 2007. Notes on the Application of the CESA Quantification Method for Scenarios Simulated in the Halden
Reactor Experiments. Paul Scherrer Institute, Villigen
PSI, Switzerland, Draft.
Reer, B., Dang, V.N. 2006. Situational Features of
Errors of Commission Identified from Operating Experience. Paul Scherrer Institute, Villigen PSI, Switzerland,
Draft.
Reer, B., Dang, V.N. 2007. The Commission Errors
Search and Assessment (CESA) Method. PSI Report
Nr. 07-03, ISSN 1019-0643, Paul Scherrer Institut,
Switzerland.
Reer, B., Dang V.N., Hirschberg S. 2004. The CESA method
and its application in a plant-specific pilot study on errors
of commission. Reliability Engineering and System Safety
83: 187205.

241

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Exploratory and confirmatory analysis of the relationship between

social norms and safety behavior
C. Fugas & S.A. da Silva
ISCTE/CIS, Lisbon, Portugal

J.L. Meli
University of Valencia, Valencia, Spain

ABSTRACT: Although several researchers have argued that social norms strongly affect health behaviors,
the measurement of health and safety norms has received very little attention. In this paper, we report the
results of our study designed to: 1) test the reliability and construct an effective questionnaire devoted to
the measurement of social influences on safety behavior; 2) assess the predictive validity of supervisor and
coworkers descriptive and injunctive safety norms on safety behavior; 3) test a Four-Factor CFA model of social
influence on safety behavior (confirmatory factor analysis). The questionnaire has 11 items four-scales and
used a 7-point Likert-type scale of safety behavior. A self-reporting scale of safety behavior was also included.
A sample (N = 250) of operational team workers from a Portuguese company participated voluntarily and
anonymously in the study. Overall results from this study (EFA and CFA) confirmed the questionnaire structure
and provided support for a correlated, Four-Factor model of Safety Group Norms. Furthermore, this study
had demonstrated that coworkers descriptive and injunctive safety norms were a strong predictor of safety
behavior.

INTRODUCTION

Work accidents are still a problem in the world, far

from being solved, even in some developed countries.
According to Hmlinen, Takala & Saarela (2006),
the fatality and accident rates per 100 000 workers
in Canada, United States, Ireland, Italy, Portugal and
Spain are higher than average.
The understanding of the antecedents and determinants of safety behavior can be an important contribution to the prevention of work accidents. Consistent
with the common focus of safety research at the
individual level of analysis, the vast majority of behaviorally based safety interventions (e.g. safety training,
feedback) have focused on individual factors. However, there is a widespread view that group norms
represent an important contextual influence on health
and safety behaviors (e.g. Conner, Smith & McMillan,
2003; Linnan, et al., 2005) and interventions aimed at
changing behaviors by reshaping peoples normative
beliefs may be promising. Thus, understanding the
pressure points, whereby group and normative variables shape workers health and safety, may contribute
to maximize the effectiveness of health promotion
programs.

Evidence from workplace safety (safety climate)

research reveals that groups emerge as a critical unit
analysis in safety task accomplishment and safety
performance (e.g. Zohar, 2002) and it is suggested
that they will become more important in the evolving
modern workplace (Tesluk & Quigley, 2003). Nevertheless, the role of group and team factors in safety,
relative to other criteria, has not been extensively
studied.
On the other hand, the measurement of health and
safety norms has also received little attention. Some
recent applications of the theory of planned behavior use single-item or one-dimensional measures (see
for an exception Fekadu & Kraft, 2002; Conner &
McMillan, 1999) and have ascribed only crossover
effects. These findings reflect the lower importance
given to normative factors in the study of behaviors.
A meta-analytic review of the efficacy of the theory of planned behavior (Armitage & Conner, 2001)
found that more reliable multiple-item measures of
subjective norm and normative beliefs had significantly stronger correlations with intention than any
of the other measures.
Thus, multidimensional measures, that might be
consistently applied to quantify and characterize the

243

structure of norms, can be valuable to the understanding of the contextual normative influences on workers
health behaviors.
2

OVERVIEW OF THEORETICAL
FRAMEWORK

The proposed conceptual framework in this paper

relies on workplace safety literature (e.g. Zohar &
Luria, 2005) and on social cognitive theories, namely
the theory of planned behavior (e.g. Ajzen, 2005) and
focus theory of normative conduct (e.g. Cialdini, et al.,
1990, 1991) to understand how group norms impact
safety behavior.
According to Cialdini & Trost (1998) group norms
are guidelines for acceptable and unacceptable behavior that develop through interactions among group
members and are informally agreed upon by group
members. They may be actively/verbal (e.g. explicit
statements) or passively transmitted/non-verbal (e.g.
modeling). Any social punishments for not complying with norms come from social networks and not
from formal systems established by the organization.
Following these assumptions, it is our view that safety
group norms are internalized informal safety rules that
work groups adopt to regulate and regularize group
members behavior.
Reviews of previous research indicate that, typically, normative influence on behavior has been
studied in terms of subjective norm or persons perceptions of whether specific salient others think he/she
should engage in the behavior and the motivation
to comply with such pressure (cf. Ajzen, 1991,
2005). For instance, an application of the theory of
planned behavior to the prediction of safe behavior (see
Johnson & Hall, 2005) used three items to measure
subjective norms (e.g., Most people who are important to me would strongly encourage/discourage me to
lift materials within my strick zone).
But, there is an important distinction in the literature on social influence between injunctive norms
and descriptive norms. Cialdini and colleagues (e.g.
Cialdini et al., 2006) call the subjective norms, in
the usual applications of the theory of planned behavior, injunctive social norms, as they concern others
social approval or disapproval. They argue that when
considering normative influence on behavior, it is crucial to discriminate between the is (descriptive) and
the ought (injunctive), because each refers to a separate source of human motivation (Deutsch & Gerard,
1955) (cit. Cialdini et al., 1990, p. 1015). In other
words, whereas descriptive norms provide information about what is normal, what most people do and
motivate human action providing evidence of what is
likely to be effective and adaptative action, injunctive norms provide information about what ought to

be done and motivate by promising social rewards and

punishments.
It is usually that descriptive and injunctive norms
are mutually congruent. According to Cialdini et al.
(1990) Because what is approved is often actually
what is done, it is easy to confuse these two meanings of
norms . . . it is important for a proper understanding of
normative influence to keep them separate, especially
in situations where both are acting simultaneously
(p. 1015).
Although the discriminant and convergent validity of descriptive and injunctive norms (or subjective
norms) constructs have been supported by literature
(see for a review, Rivis & Sheeran, 2003) it remains
unclear, in the specific domain of safety, how both
normative components predict safety behaviors.
Finally, in the present study, the definition of safety
behavior comprises not only compliance behaviors
such as properly using personal protective equipment
and engaging in work practices that reduce risk, but
also more proactive safety behaviors or safety citizenship behaviors, such as helping teach safety procedures
to new crew members, assisting others to make sure
they perform their work safely, making safety-related
recommendations about work activities (see Burke,
Sarpy, Tesluk, & Smith-Crowe, 2002 and Hoffmann,
Gerras & Morgeson, 2003).

METHOD

3.1 Participants
Participants in this study were operational workers,
members of work teams employed in a Portuguese
company (passenger transportation company with
high safety standards). The sample consisted of
250 workers, who provided anonymous ratings of
the descriptive and injunctive safety norms of their
coworkers and supervisors and of their own individual
safety behavior. All participants were male and most
respondents (83,2%) were not supervisors and had heterogeneous jobs in the company. Almost half of the
participants (48,8%), were between 31 e 40 years old
and the other half (42,5%) were over 41 years old.
Finally, 38,4% had tenure on the job between 6 and 15
years and 41,6% tenure superior to 15 years.
3.2 Material
In this study, data was collected from individual group
members, using a quantitative methodology (survey
administered in the form of a questionnaire).
Four 11-item scales measured descriptive and
injunctive safety group norms and we also considered
the referent implied in the survey items. Therefore,
descriptive and injunctive norms were assessed in

244

relation to coworkers and supervisors. The scale items

(descriptive vs. injunctive) were based on the assumptions of the focus theory of normative conduct (e.g.
Cialdini et al., 2006) concerning the measurement
of social norms, as prior applications to the study
of health behaviors in other contexts (e.g. Conner &
McMillan, 1999). We should note that the four scales
included generic items that can be applicable to many
types of jobs involving safety issues. To begin scale
development, we departed from Zohar and Lurias
(2005) group-level safety climate scale, we created a
pool of 11 items for the individuals perceptions about
the supervisorss descriptive safety norms (e.g. My
direct supervisor checks to see if we are obeying safety
regulations). Then, each item was adapted to the individuals perceptions of the descriptive safety norms of
their coworkers (e.g. My team members check to see
if other team members obey the safety rules) and that
we follow the same procedure for injunctive safety
norms (e.g. My direct supervisor thinks that compliance to safety rules should be checked and My
team members think that compliance to safety rules
should be checked out). This procedure resulted in an
initial 44-item Safety Group Norms questionnaire.
Respondents rated the frequency (using a 7-point scale
ranging from never to always) with which coworkers and direct supervisors performed/think should be
performed each of the 11 indicated behaviors.
The Safety Behavior self-report was measured
using a revised and updated version of the original General Safety-Performance Scale (Burke et al.,
2002), and Safety Citizenship Role Definitions and
Behavior Items (Hofmann et al, 2003). Because our
interest was in the overall safety behavior, we combined these two measures, resulting in a 12-item Safety
Behavior Scale (e.g. I properly performed my work
while wearing personal protective equipment and
I have made suggestions to improve safety). We
assume that safety behaviors can be assessed with
respect to the frequency that employees engage in
the behaviors, using a 7-point Likert-type scale ranging from never (1) to always (7) with high scores
representing more positive safety behaviors.
Given the limitations associated with the use of
self-reports of behavioral safety, the present study also
used microaccidents1 as outcome criterion of behavioral safety. Participants were also asked to self-report
the number of microaccidents they were involved in
the last 6 months.

1 In

this study microaccidents were conceptualized

according to Zohar (2000) as on-job behavior-dependent
minor injuries requiring medical attention, but not incur
any lost workdays.

RESULTS

This Two-step approach was comprised of exploratory

analysis and confirmatory factor analysis.
In the first step, exploratory analysis included factor analysis, reliability and predictive validity analysis.
Data was analyzed using the Statistical Package for the
Social Sciences (SPSS). The method of factor extraction was principal component (varimax rotation).
In the second step, a confirmatory factor analysis was performed with Structural Equation Modeling
(SEM) using the Maximum Likelihood Estimation
(MLE) method. All estimations were made with
AMOS 7.0 software. This multivariate technique was
used to test (confirm) the measurement construct
validity of a Four-Factor CFA model of safety group
norms, and to compare this model with One-Factor
and Two-Factor CFA models.

4.1 Exploratory factor analysis

Exploratory Factor Analysis (EFA) played an important complementary role in evaluating the proposed
dimensionality of constructs. The initial 44-item
Safety Group Norms Questionnaire was submitted
for two independent principal components analysis
(rotation of varimax) with the purpose of identifying the underlying structure of the variables and
also to create a subset of variables representative of
the factors, much smaller in number (four 5-item
scales) for inclusion in subsequent analysis (confirmatory factor analysis). The overall results of reliability analysis of this smaller version is presented
in Table 1. Four dimensions reflected by a separate
factor were identified (SDSNSupervisors Descriptive Safety Norms; SISNSupervisors Injunctive Safety Norms; CDSNCoworkers Descriptive
Safety Norms; CISNCoworkers Injunctive Safety
Norms). All factor loadings were above .65, all Cronbachs alpha coefficients exceed .80, corrected itemto-total correlations were all above .60 and inter-item
correlations were greater than .40. In brief, exploratory
factor analysis results confirm the proposed multidimensionality of constructs and the appropriateness of
the variables.
The 12-item Safety Behavior Scale was also submitted to factorial analysis (principal components analysis) and two factors were extracted. Only 9 items were
retained for further analysis. The first factor, was
denominated proactive safety practices (PASB -5
items) and explained 52,60% of variance; the second factor was designated compliance safety practices (CPSB4 items) and explained 11,48% of total
variance. Both scales had Cronbachs alphas higher
than .70, coefficients indicative of a good internal
consistency.

245

Table 1.

and injunctive norms accounted for 31,7% of the

variation in compliance safety behavior (Adjusted
R 2 = .32). However, as found previously, compliance safety behavior was only significantly associated
with coworkers descriptive safety norms (Beta = .37,
p < .0001) and coworkers injunctive safety norms
(Beta = .23, p < .0001).
Microaccidents did not predict safety behavior
(Adjusted R 2 = .02, p > .05).

Exploratory factor analysis results.

1A. Supervisors
safety norms.

1B. Coworkers
safety norms.

Items

Items

P1g
P1h
P1i
P1j
P1k
P2b
P2c
P2d
P2e
P2f

.84
.87
.86
.83
.79
.42
.42
.35
.37
.37

.38
.38
.38
.45
.35
.79
.82
.84
.85
.86

P3a
P3c
P3d
P3e
P3k
P4f
P4g
P4h
P4i
P4j

.22
.27
.15
.24
.39
.81
.87
.90
.83
.87

.81
.81
.75
.78
.65
.26
.26
.24
.30
.24

Alpha

.96

.95

Alpha

.86

.94

Total variance 84,87%

4.2 Confirmatory factor analysis

With the purpose of confirming the dimensionality
of safety group norm constructs, a Four-Factor CFA
model was tested (Model 1). This model was compared
with a Two-Factor CFA model (Model 2) and OneFactor CFA model (Model 3). Model fit was assessed
considering the Chi-square, overall goodness-of-fit
statistics, an analysis of residuals and the magnitudes
of item factor loadings (see Table 3 and Fig. 1).

Total variance 72,54%

Table3.

Table 2. Multiple regression analysis to predict safety

behavior (variable criteria).
PASB
B

Model
1
2*
3

CPSB
SE B B

SDSN 0.02 0.06

SISN
0.05 0.06
CDSN
0.27 0.07
CISN
0.30 0.07

.03
.07
.29
.29

SE B

0.05
0.08
0.28
0.20

0.05
0.05
0.06
0.06

Summary of model fit statistics.

2/DF
2.16
3.75
14.7

GFI

CFI

RMSEA

AIC

.90
.93
.48

.97
.98
.63

.07
.11
.24

287.8
105.5
1595.95

ECVI
1.19
.44
6.62

* This values refers to supervisors scale

.09
.13
.37
.23

err1

p1g

err2

p1h

err3

p1i

err4

p1j

,91
,94
,93

Note: Adjusted R 2 = .28 (p < .0001) for Proactive Safety

Behavior; Adjusted R 2 = .32 (p < .0001) for Compliance
Safety Behavior. p < .0001, p < .001.

Regression analysis was used to assess the predictive validity of safety group norms on safety behavior
(variable criteria). The principal results of this multiple
regression are reported in Table 2.
The results of the statistical test of the regression
models ability to predict proactive safety behaviors reveals that the model is statistically significant
(F = 24, 201; p < .0001); supervisor and coworkers
descriptive and injunctive safety norms accounted
together for 28,1% of the variation in proactive safety
practices (Adjusted R 2 = .28). Nevertheless, proactive safety practices were only significantly associated
with coworkers descriptive safety norms (Beta =
.29, p < .0001) and coworkers injunctive safety
norms (Beta = .29, p < .0001).
The model for the prediction of compliance
safety behaviors was also confirmed (F = 28, 819;
p < .0001); supervisor and coworkers descriptive

SDSN

,94

,79
err5

p2b

err6

p2c

err7

p2e

err8

p2f

,90
,94
,86

,62
SISN

,92

,49

,55
err9

p3a

err10

p3c

err11

p3e

err12

p3k

err13

p4g

err14

p4h

err15

p4i

err16

p4j

,89
,91
,59

CDSN

,51

,69

,55
,89
,95
,84

CISN

,85

Figure 1. Standardized maximum likelihood solution for

the hypothesized four-factor CFA model (all p s. < .0001).

246

The whole set of fit statistics confirms the FourFactor CFA model (according to Hu and Bentler,
1999 criteria for fit indexes). The examination of the
first set of model fit statistics shows that the Chisquare/degrees of freedom ratio for the Four-Factor
CFA model (2.16) is adequate; the GFI index (.90),
the CFI (.97) and the RMSEA (.07) are consistent
in suggesting that the hypothesized model represents
an adequate fit to the data; finally, the AIC and
the ECVI (criteria used in the comparison of two or
more models) presents smaller values than in the OneFactor CFA model, which represents a better fit of the
hypothesized Four-Factor CFA model.
The standardized path coefficients are portrayed
in Figure 1. The parameters relating items to factors
ranged between .59 and .95 (p < .0001). Correlations between the four factors were also significant
(p < .0001) and ranged between .49 and .79.
5

CONCLUSIONS

The results of the current study adds to the growing

body of literature for the valid measurement of safety
group norms. Results of EFA and CFA confirms the
questionnaire structure. Supervisors and coworkers
descriptive and injunctive safety norms are not isomorphic constructs, they refer to separate dimensions of social influence on safety behavior. Overall
results provide support for a correlated, four-factor
model.
In addition, this study further examines the predictive power of safety group norms on safety behavior.
In line with expectations, multiple regression results,
have demonstrated that coworkers descriptive and
injunctive safety norms are a strong predictor of proactive and compliance safety behavior. Surprisingly, in
this company, supervisors safety norms didnt predict
individual safety behavior. These results propose that
coworkers safety behaviors are mostly influenced by
their peers, and less by their supervisors, suggesting
that coworkers can mediate the effect of supervisors safety norms on individual safety behavior at
work. This is an important finding given that very
little research has examined how peers can impact
safety behavior. These results also suggest that organizational safety initiatives should be aware of the
important role of fellow team members on individual
attitudes and safety behaviors at work. New research
should be conducted in order to replicate the factorial
structure and contrast if the pattern of safety influences is an idiosyncratic characteristic of the structure
of teamwork and supervision in this company. Also,
the validity of this model for predicting behavior will
further include the test of a more complete sociocognitive model with SEM, ideally using a multilevel
approach.

ACKNOWLEDGEMENTS
This research was supported by Fundao para a Cincia e TecnologiaPortugal (SFRH/BDE/15635/2006)
and Metropolitano de Lisboa.

REFERENCES
Ajzen, I. 1991. The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50:
179211.
Ajzen, I. 2005. Laws of human behavior: Symetry, compability and attitudebehavior correspondence. In A. Beauducel, B. Biehl, M. Bosniak, W. Conrad, G. Shnberger e D.
Wagener (eds.), Multivariate Research Strategies: 319.
Aachan, Germany: Shaker Verlag.
Armitage, C.J. & Conner, M. 2001. Efficacy of the theory of planned behaviour: a meta-analytic review. British
Journal of Social Psychology, 40: 471499.
Burke, M.J., Sarpy, S.A., Tesluk, P.E. & Smith-Crowe, K.
2002. General safety performance: A test of a grounded
theoretical model, Personnel Psychology, 55.
Cialdini, R.B., Kallgren, C.A. & Reno, R. 1991. A focus
theory of normative conduct. Advances in Experimental
Social Psychology, 24: 201234.
Cialdini, R.B., Reno, R. & Kallgren, C.A. 1990. A focus theory of normative conduct: Recycling the concept of norms
to reduce littering in public places. Journal of Personality
and Social Psychology, 58(6): 10151026.
Cialdini, R.B., Sagarin, B.J., Barrett, D.W., Rhodes, K. &
Winter, P.L. 2006. Managing social norms for persuasive
impact, Social Influence, 1(1): 315.
Cialdini, R.B. & Trost, M.R. 1998. Social influence: Social
norms, conformity and compliance. In D.T. Gilbert,
S.T. Fiske, & G. Lindzey (Eds.), The Handbook of
Social Psychology (4th ed., Vol. 2): 151192. New York:
McGraw-Hill.
Conner, M. & McMillan, B. 1999. Interaction effects in
the theory of planned behavior: Studying cannabis use,
British Journal of Social Psychology, 38: 195222.
Conner, M., Smith, N. & McMillan, B. 2003. Examining
normative pressure in the theory of planned behaviour:
Impact of gender and passengers on intentions to break
the speed limit, Current Psychology: Developmental,
Learning, Personallity, Social, 22(3): 252263.
Deutsch, M. & Gerard, H.B. 1955. A study of normative
and informational social influences upon individual judgment. Journal of Abnormal and Social Psychology, 51:
629636.
Fekadu, Z. & Kraft, P. 2002. Expanding the Theory
of Planned Behavior: The role of social Norms and
Group Identification. Journal of Health Psychology, 7(1):
3343.
Hmlinen, P., Takala, J. & Saarela, K.L. 2006. Global
estimates of occupational accidents. Safety Science, 44:
137156.
Hofmann, D.A., Morgeson, F.P. & Gerras, S.J. 2003. Climate as a moderator of the relationship between LMX
and content specific citizenship behavior: Safety climate
as an exemplar. Journal of applied Psychology, 88(1):
170178.

247

Hu, L.-T. & Bentler, P.M. 1999. Cutoff criteria for fit indexes
in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A
Multidisciplinary Journal, 6: 155.
Johnson, S. & Hall, A. 2005. The prediction of safe lifting
behavior: An application of theory of planned behavior.
Journal of Safety Research, 36: 6373.
Linnan, L., Montagne, A., Stoddard, A., Emmons, K.M. &
Sorensen, G. 2005. Norms and their relationship to behavior in worksite settings: An application of the Jackson
Return Potential Mode, Am. J. Health Behavior, 29(3):
258268.
Rivis, A. & Sheeran, P. 2003. Descriptive norms as
an additional predictor in the Theory of Planned
Behaviour: A meta-analysis. Current Psychology:
Developmental, Learning, Personality, Social, 22(3):
218233.

Tesluk, P. & Quigley, N.R. 2003. Group and normative influences on health and safety, perspectives from taking a
broad view on team effectiveness. In David A. Hofmann
e Lois E. Tetricck (Eds.), Health and Safety in Organization. A multilevel perspective: 131172. S. Francisco:
John Wiley & Sons.
Zohar, D. 2000. A group-level model of safety climate:
Testing the effect of group climate on microaccidents in
manufacturing jobs. Journal of Applied Psycholog, 85:
587596.
Zohar, D. 2002. The effects of leadership dimensions, safety
climate, and assigned priorities on minor injuries in work
groups. Journal of Organizational Behavior, 23: 7592.
Zohar, D. & Luria, G. 2005. Multilevel model of safety
climate: Cross-level Relationships between organization
and group-level climates. Journal of Applied Psychology,
9(4): 616628.

248

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Functional safety and layer of protection analysis with regard

to human factors
K.T. Kosmowski
Gdansk University of Technology, Gdansk, Poland

ABSTRACT: This paper is devoted to selected aspects for knowledge-based layer of protection analysis (LOPA)
of industrial hazardous systems with regard to human and organizational factors. The issue is discussed in the
context of functional safety analysis of the control and protection systems to be designed and operated according
to international standards IEC 61508 and IEC 61511. The layers of protection can include, for instance, the
basic process control system (BPCS), human-operator (HO) and safety instrumented system (SIS). Such layers
should be independent; however, due to some factors involved the dependencies can occur. Thus, it may result in
the risk level increasing of accident scenarios identified in the process of risk analysis. The method is illustrated
on example of the control system (TCS) and protection system (TPS) of a turbo-generator set, which has to
be shut-down in some situations of internal or external disturbances. The required risk reduction is distributed
among TCS and TPS to be designed for appropriate safety integrity level (SIL).

INTRODUCTION

The intensive development of advanced technologies

is observed in many domains, such as manufacturing,
electric power sector, process industry, transport, shipping etc. However, the organisational and management
structures in these domains are often traditional. Some
researchers express opinion that second generation
management is applied to fifth generation technology
(Rasmussen & Svedung 2000).
The research works concerning causes of industrial accidents indicate that broadly understood human
errors, resulting from organisational inefficiencies and
management inadequacies, are determining factors in
7090% of cases, depending on industrial sector and
the system category. Because several defences against
foreseen faults and potential accidents are usually used
to protect people and installations plants, multiple
faults have contributed to many accidents.
It has been often emphasized that disasters arise
from a combination of latent and active human errors
committed during the design, operation and maintenance (Embrey 1992). The characteristic of latent
errors is that they do not immediately degrade the
functions of systems, but in combination with other
events, such as random equipment failure, external
disturbance or active human errors, can contribute to
a catastrophic failure or major accident with serious
consequences.
Embrey (1992) distinguishes two categories of
latent errors: operational and organizational. Typical

operational latent errors include maintenance errors,

which may make a critical system unavailable or leave
it in a vulnerable state. Organizational latent errors
include the design errors, which give raise to intrinsically unsafe systems, and management or safety
policy errors, which create conditions that induce
active human errors. Some categorizations of human
errors have been proposed, e.g. by Swain & Guttmann
(1983) and Reason (1990).
Traditionally, the human and organisational factors
are incorporated into the probabilistic models through
the failure events and their probabilities evaluated
using relevant methods of human reliability analysis
(HRA) (Gertman & Blackman 1994). In some HRA
methods it is possible to introduce also the organisational factors, related in principle to the safety culture
aspects. Correct evaluation of human behaviour and
potential errors is a prerequisite of correct risk assessment and rational safety-related decision making in the
safety management process (Swain & Guttmann 1983,
Dougherty & Fragola 1988, COA 1998, Hollnagel
2005).
Lately some approaches have been proposed by
Carey (2001), Hickling et al. (2006) and Kosmowski
(2007) how to deal with the issues of human factors in
the functional safety management. The human actions
and related errors can be committed in entire life cycle
of the system, from the design stage, installation,
commissioning, its operation to decommissioning.
During the operation time the human-operator interventions include the control actions during transients,

249

disturbances, faults and accidents as well as the diagnostic activities, the functionality and safety integrity
tests, maintenance actions and repairs after faults.
The operators supervise the process and make decisions using some alarm panels within the operator support system (OSS), which should be designed carefully
for abnormal situations and accidents, also for cases of
partial faults and dangerous failures within the electric,
electronic and programmable electronic (E/E/PE) systems (IEC 61508) or the safety instrumented systems
(SIS) (IEC 61511). The OSS when properly designed
will contribute to reducing the human error probability
and lowering the risk of potential accidents.
The paper outlines the concept of using a
knowledge-based method for the layer of protection
analysis (LOPA) of industrial hazardous systems with
regard to the influence of human and organizational
factors (H&OF). Various layers of protection can
be distinguished in the context of identified accident scenarios, including e.g. basic process control
system (BPCS), human-operator (HO) and safety
instrumented system (SIS) designed according to
requirements and probabilistic criteria given in functional safety standards: IEC 61508 and IEC 61511.
The protection layers should be independent. However, due to some factors involved the dependencies
can occur. It can result in a significant increase of
the risk level of accident scenarios identified in the
risk evaluating process. The problem should be carefully analyzed at the system design stage to consider
relevant safety related functions of appropriate SILs.
To reduce the human failure probability (HEP) an
advanced OSS should be designed. For dynamic processes, with a short HO reaction time permitted, the
HEP can be high, close to 1. The paper emphasizes the
importance of the context-oriented human reliability
analysis (HRA) within the functional safety management and necessity to incorporate in a systematic way
more important influencing factors.

2
2.1

FUNCTIONAL SAFETY ANALYSIS

AND HUMAN FACTORS
Safety integrity levels and probabilistic criteria

control and/or protection systems. Human-operator

(HO) contributes to realization of a SRF according
to the technical specification. There are still methodological challenges concerning the functional safety
management in life cycle (Kosmowski 2006).
An important term related to the functional safety
concept is the safety integrity (IEC 61508), understood
as the probability that given safety-related system will
satisfactorily perform required SRF under all stated
conditions within a given period of time. The safety
integrity level (SIL) is a discrete level (1 4) for
specifying the safety integrity requirements of given
safety function to be allocated using the E/E/PE system
or SIS (IEC 61511). The safety integrity level of 4
(SIL4) is the highest level, which require a redundant
architecture of the E/E/PE system with diagnosing and
testing of subsystems.
For consecutive SILs two probabilistic criteria are
defined in IEC 61508, namely:
the average probability of failure to perform the
safety-related function on demand (PFDavg ) for the
system operating in a low demand mode, and
the probability of a dangerous failure per hour PFH
(the frequency) for the system operating in a high
demand or continuous mode of operation.
The interval probabilistic criteria for the safetyrelated functions to be implemented using E/E/PE
systems are presented in Table 1. Similar interval criteria also used in assessments of SIS (IEC 61511)
and Safety-related Electrical, Electronic and Programmable Electronic Control Systems for Machinery
(SRECS) (IEC 62061).
The SIL for given SRF is determined in the risk
assessment process for defined risk matrix, which
includes areas for distinguished categories of risk, e.g.
unacceptable, moderate and acceptable (IEC 61508).
Verifying SIL for given safety-related function to
be implemented using the E/E/PE or SIS system is
usually a difficult task due to lack of reliability data and
other data used as parameters in probabilistic models
of the system in design. In such situation, a qualitative
method for crude verifying of SIL is permitted in IEC
61508 for the system architectures considered at the
design stage.

Modern industrial systems are extensively computerised and equipped with complex programmable control and protection systems. In designing of the control
and protection systems a functional safety concept
(IEC 61508) is more and more widely implemented
in various industrial sectors, e.g. the process industry
(IEC 61511) and machine industry (IEC 62061).
The aim of functional safety management is to
reduce the risk of hazardous system to an acceptable or tolerable level introducing a set of safetyrelated functions (SRFs) to be implemented using the

250

Table 1. Probabilistic criteria for safety functions to be

allocated using E/E/PE systems.
SIL

PFDavg

PFH [h1 ]

4
3
2
1

[105 , 104 )
[104 , 103 )
[103 , 102 )
[102 , 101 )

[109 , 108 )
[108 , 107 )
[107 , 106 )
[106 , 105 )

However, the rules proposed in this standard for

qualitative evaluation of the E/E/PE block diagrams
should be used with caution, because they do not
cover fully cases of potential dependent failures of
subsystems. The problem is more challenging if it is
necessary to include potential human errors, which
are related to the human and organisational factors
(Kosmowski 2007).

2.2

Functional safety management with regard

to human factors

All relevant human and organizational factors

issues,
Procedural faults and human errors,
Abnormal and infrequent modes of operation,
Reasonably foreseeable misuse,
Claims on operational constraints and interventions.
It is emphasized that the operator interface analysis
should:

Lately, a framework was proposed (Carey 2001) for

addressing human factors in IEC 61508 with consideration to a range of applications of the E/E/PE systems
in safety-related applications. It has highlighted the
diversity of ways in which human factors requirements
map on to various E/E/PE systems in different industries and implementation contexts. Some conclusions
were formulated as follows:
Determination of the safety integrity level (SIL)
for an E/E/PE system requires careful consideration
of not only of the direct risk reduction functions
it is providing, but also those risk reduction functions performed by staff that interact with it; this
requires addressing in the step of Hazard and Risk
Analysis of the IEC 61508 lifecycle;
Having determined the safety integrity of the
E/E/PE system, it was suggested that the effort that
needs to be placed into operations and maintenance
in relation to human factors should be greater as the
SIL level increases, especially for solutions of SIL3
and SIL4;
The types of human factors issues that need to be
addressed vary between the classes of systems discussed; therefore, the framework is not specific in
terms of the technology or other aspects related to
human factors.
Some general remarks were made for addressing
human factors (HFs) within IEC 61508 that include:
Incorporation of human tasks and errors into the
Hazard and Risk Assessment process;
Use of the tables to define the human factors
requirements for a given safety integrity level.
In the paper by Hickling et al. (2006) publishing
a Guidance for Users of IEC 61508 was announced.
The guidance is designed to respond to requirements
laid down in this standard. They fall into two broad
categories of those associated with:
1. the hazard and risk analysis, and
2. the interface between human operator and technology (process).

The hazard and risk analysis has to include

(Gertman & Blackman 1994, Hickling et al. 2006,
IEC 61508):

Be covered in safety requirements,

Take account of human capabilities and limitations,
Follow good HF practice,
Be appropriate for the level of training and awareness of potential users,
Be tolerant of mistakessee classification of
unsafe acts and human errors by Reason (1990).
Thus, the scope of analyses should include relevant
human and organizational factors (H&OFs) aspects
to be traditionally included in the HRA methods
used in Probabilistic Safety Analysis (PSA) (Swain
& Guttmann 1983, Humphreys 1988, COA 1998).
2.3 Human reliability analysis in probabilistic
modeling of safety-related systems
The human reliability analysis (HRA) methods are
used for assessing the risks from potential human
errors, and for reducing the system vulnerability, operating in given environment. However, some basic
assumptions in HRA methods used within probabilistic safety analysis of hazardous systems are the
subjects of dispute between researchers (Hollnagel
2005).
Practically all HRA methods assume that it is meaningful to use the concept of human errors and it is
justified to estimate their probabilities. Such point of
view is sometimes questioned due to not fully verified assumptions about human behavior and potential
errors. Hollnagel (2005) concludes that HRA results
are of limited value as an input for PSA, mainly
because of oversimplified conception of human performance and human error.
In spite of this criticism, waiting for next generation of HRA methods, the human factor analysts use
for PSA several exiting HRA methods. Below selected
HRA methods are shortly characterized, which can be
applied for HRA within the functional safety analysis. The rough human reliability assessments based on
qualitative information with regard to human factors
can be especially useful for initial decision making
at the designed stage of the safety-related functions

251

and systems (Kosmowski 2007). It will be demonstrated that a functional safety analysis framework
gives additional insights in HRA.
Several traditional HRA methods have been used
in PSA practice, including THERP method (Swain
& Guttmann 1983), developed for the nuclear
industry, but applied also in various industrial sectors.
Other HRA methods, more frequently used in practice are: Accident Sequence Evaluation ProcedureHuman Reliability Analysis Procedure (ASEP-HRA),
Human Error Assessment and Reduction Technique
(HEART), and Success Likelihood Index Method
(SLIM)see description and characterization of HRA
methods (Humphreys 1988, COA 1998).
Two first mentioned methods (THERP and ASEPHRA) are the decomposition methods, based on a
set of data and rules for evaluating the human error
probabilities (HEPs). HEART consists of generic
probabilistic data and a set of the influence factors
for correcting nominal human error probability. SLIM
enables to define a set of influence factors, but requires
data for calibrating the probabilistic model.
In the publication by Byers et al. (2000) five HRA
methods were selected for comparison on the basis
of either relatively widespread usage, or recognized
contribution as a newer contemporary technique:

made to the base error weights. Each task no longer

has to be rated on both processing (diagnosis) and
response (action) components, only if the task contains diagnosis does diagnosis get rated, and similarly
for action.
Changes were also made to the worksheet to
enhance usability and to gather more information
when non-nominal ratings are made. The overall range
of possible HEPs has been expanded.
The final conclusion is that the enhanced SPAR-H
methodology is useful as an easy-to-use, broadly
applicable, HRA screening tool. The comparisons
and enhancements allow the SPAR-H methodology to
maintain the strengths of the original ASP HRA (1994)
methodology, while taking advantage of the information available from user feedback and from other HRA
methods and sources.
HEP is evaluated when the human failure event is
placed into the structure of probabilistic model of the
system. In the HRA within PSA only more important human failure events are considered (Kosmowski
2004). Then the context related PSFs are specified and
determined according to rules of given HRA method.
As the result the particular value of HEP is calculated.
Different methods are used for evaluating HEP with
regards to PSFs, e.g. assuming a linear relationship for
each identified PSFk and its weight wk , with constant
C for the model calibration

Technique for Human Error Rate Prediction

(THERP) (Swain and Guttmann, 1983);

Accident Sequence Evaluation Program (ASEP)
wk PSFk + C
(1)
HEP = HEPno min al
(Swain, 1987);
k
Cognitive Reliability and Error Analysis Method
(CREAM) (Hollnagel, 1998);
or nonlinear relationship, as in the SPAR-H method Human Error Assessment and Reduction Technique
ology (Gertman et al. 2005)
(HEART) (Williams, 1988);
A Technique for Human Event Analysis (ATHEANA)
NHEP PSFcomposite
(USNRC, 1998).
(2)
HEP =
NHEP(PSF
composite 1) + 1
In addition to these methods, other sources of
information were also examined to provide insights
where NHEP is the nominal HEP. NHEP equals 0.01
concerning the treatment and evaluation of human
for diagnosis, and NHEP equals 0.001 for action.
errors. Comparisons were also made with regard to
the SPAR-H HRA method (Gertman et al. 2005).
The team members were asked to construct more
2.4 Including human failure events in functional
detailed summaries of their methods and sources,
safety analysis and probabilistic modeling
including specific lists of: 1) error types; 2) any base
rates associated with the error types; 3) performance
In probabilistic modeling of the E/E/PE safety-related
shaping factors (PSFs); 4) PSF weights; and 5) depensystem the human failure events and their probabilities
dency factors. Obviously, not all methods and sources
are elements of a subsystem model as explained below.
contained all of this information, and certainly not all
For instance, PFDavg of a E/E/PE subsystem (SUB)
used the same terminology of error types and PSFs
operating in the low demand mode is calculated from
(Byers et al. 2000).
the formula (Kosmowski et al. 2006):
In summary the authors conclude that a result of
the comparison of the ASP HRA methodology to
SUB
FT
AT
+ PFDavg
+ PFDHE
(3)
PFDavg
= PFDavg
other methods and sources, enhancements were made
as regards: error type names, error type definitions,
FT
is the average probability of a subwhere: PFDavg
PSFs, PSF weights, PSF definitions, dependency conditions and dependency definitions. No changes were
system failure on demand, detected in periodical

252

AT
functional test (FT); PFDavg
the probability of subsystem failure on demand, detected in automatic test
(AT); PFDHE the probability of failure on demand
due to human error (HE). Depending on the subsystem and situation considered the human error can be a
design error (hardware of software related) or an operator error (activities of the operator in the control room
or as a member of maintenance group).
The E/E/PE safety-related system in designing consists of subsystems: sensors/transducers/converters
(STC), programmable logic controllers (PLC) and
equipment under control (EUC). Each of these subsystems can be generally treated as KooN architecture,
which is determined during the design. Each PLC
comprises the central unit (CPU), input modules (digital or analog) and output modules (digital or analog).
The average probability of failure on demand PFDavg
of the E/E/PE safety-related system (SYS) is evaluated as the sum of probabilities for these subsystems
(assuming small values of probabilities) from the
formula
STC
PLC
EUC
SYS
+ PFDavg
+ PFDavg
PFDavg
= PFDavg

(4)

The PFDs in formula (4) can be evaluated

qualitatively (indicating SIL with relevant interval
PFDavg see Table 1) or quantitatively (results of
probabilistic modeling process) with regard to potential dependent failures and relevant human errors.
For complex KooN architectures suitable methods of
probabilistic modeling are used (Kosmowski 2007).

independent of the initiating event and the components of any other IPL already claimed for the same
scenario,
auditable, i.e. the assumed effectiveness in terms
of consequence prevention and PFD must be capable of validation (by documentation, review, testing,
etc.).
An active IPL generally comprises following components:
Aa sensor of some type (instrument, mechanical,
or human),
Ba decision-making element (logic solver, relay,
spring, human, etc.),
Can action (automatic, mechanical, or human).
Such IPL can be designed as a Basic Process Control System (BPCS) or SIS (see the layers 2 and 4
in Fig. 1). These systems should be functionally and
structurally independent; however, it is not always possible in practice. Figure 2 illustrates the functional
relationships of three protection layers: 2, 3 and 4
shown in Figure 1. An important part of such complex
system is the man-machine interface (MMI) (GSAChP
1993, Gertman & Blackman 1994). Its functionality
and quality is often included as an important PSF in
HRA (Kosmowski 2007).

6. Plant / community emergency response

5. Relief devices / physical protection

HUMAN FACTORS AND LAYERS

OF PROTECTION ANALYSIS

4. Safety instrumented systems (SISs)

3. Critical alarms and operator interventions

3.1

Defining layers of protection

2. Control and monitoring (BPCS)

Industrial hazardous plants are nowadays designed

according to a concept of defense in depths, distinguishing several protection layers. The design of
safety-related systems is based on the risk assessment
for determination of required SIL and its verification in the process of probabilistic modeling (GSAChP
1993, IEC 61511).
Figure 1 shows typical protection layers of a hazardous industrial plant. An interesting methodology
for preliminary risk analysis and safety-related decision making is the Layer of Protection Analysis
(LOPA) methodology (LOPA 2001).
The layers shown in Figure 1 can be dependent. The
layer (a device, system, or action) to be considered
as the Independent Protection Layer (IPL) should be
(LOPA 2001):
effective in preventing the consequence when it
functions as designed,

1. Installation /
PROCESS

Figure 1.
tion.

Typical protection layers of hazardous installa-

PLANT / PROCESSES
Equiment under
Control(EUC)
Safety-related
system (SRS)

EUC (SIS)

STC (SIS)

Safety
Instrumented
System (SIS)

State control,
Testing, Supervision

STC

EUC

Basic Process
Control System
(BPCS)
Information /
interface

Decisons /
Control

SRS

Sensors /
Transducers /
Converters
(STC)

Information /
Interface

INFORMATION / ALARMS > HUMAN OPERATORS > DECISIONS

Figure 2. Components of functional safety systems for

control and protection of hazardous plant.

253

PL1
BPCS

PL2
HO

3.3

PL3
SIS

Figure 3. Protection layers for reducing the frequency of

accident scenarios.

PL1
TCS

PL2
HO

PL3
TPS

Figure 4. Protection layers reducing the frequency of accident scenario.

3.2

Dependency issue of protection layers

Protection layers (PLs) 2, 3 and 4 from Figure 1 are

shown in Figure 3. They include:
PL1basic process control system (BPCS),
PL2human-operator (HO),
PL3safety instrumented system (SIS).
If these layers are treated as independent the frequency of i-th accident scenario FiIPLs is calculated
form the formula
FIIPLs = FiI PFDiIPL1 PFDiIPL2 PFDiIPL3

As it was mentioned the international standard IEC

61508 emphasizes in many places the significance of
human failures and their potential contribution to the
probabilistic evaluations, however, there is no direct
indication how this problem should be solved. Some
general proposals are mentioned in the report (Carey
2001). One of the methods for treating dependent
failures is proposed in THERP technique (Swain &
Guttmann 1983).
THERP offers a dependency model for potential human failure events to be considered in complex situations distinguishing: ZDzero dependence,
LDlow dependence, MDmoderate dependence,
HDhigh dependence, and CDcomplete dependence. This model assumes fixed levels of the dependency factors, equivalent to beta-H factors: H = 0 for
ZD, H = 0.05 for LD, H = 0.14 for MD, H = 0.5
for HD and H = 1 for CD.
This dependency model is illustrated on an example
of two dependent events of potential human errors:
A (previous) and B (consecutive). The probability to
make an error A and error B (potentially dependent)
is evaluated as follows
P(A B) = P(A)P(B|A)

(5)

where: FiI is the frequency of an initiating event (I)

and PFDiIPL1 , PFDiIPL2 , PFDiIPL3 are probabilities of
failure on demand for protection layers shown in Fig. 4,
treated as independent layers.
If these protection layers are dependent the frequency of i-th accident scenario FiPLs is calculated taking into account the conditional probabilities resulting
in a higher value of the frequency will be obtained
FIPLs = FiI PFDiPL1 PFDiPL2 PFDiPL3 = d FiIPLs

Treating of dependent failure events

in human reliability analysis

(6)

The value of coefficient d can be much higher than

1 (d >> 1). Its value, depending on situation analyzed, can be higher even an order of magnitude. The
value of d is significantly influenced by assumptions
concerning the probabilistic modeling of dependent
failures due to equipment failures and/or human errors
(Kosmowski 2007).
An important problem is to model the dependencies
on the level of accident scenario (between protection
layers). It is especially challenging for the layer 2 (HO)
which is usually significantly dependent on the layer 1
(BPCS). These cases are considered below with regard
to potential influence of human and organizational
factors.

= (1 H )QA QB + H QA

(7)

where: P(A) = QA and P(B) = QB are probabilities

of relevant failure events. For H = 0 (independence
of events/errors) the result is P(AB) = QA QB , but for
H = 1 (complete dependence of errors) P(A B) =
QA = P(A). When QA = QB << 1 then a simplified version of formula (7) can be used: P(A B)
=
H QA . The value of H depends on various factors
to be considered in probabilistic modeling of given
safety-related situation (Swain & Guttmann 1983).
3.4 An example of protection layers analysis
including human failure event
Two safety-related systems are considered below: a
turbine control system (TCS) and a turbine protection
system (TPS). These systems perform an important
function to switch off the relevant valves and shut down
a turbine set of the electrical power unit in situations
of internal or external disturbances. Failures of these
systems can lead to very serious consequences.
The TSC operates in a continuous mode of operation and TPS is the system operating in a low demand
mode (see item 2.1). These systems are designed
with regard to required SIL, determined in the process of risk assessment. Based on these assessments it
was evaluated that the risk measure should be lowered by factor of 104 (equivalent to SIL4) using

254

TSC and TPS, treated as first and third barrier (see

Fig. 3). Between these layers there is a second barrier:
Human-Operator (HO) (see Fig. 4).
Preliminary results of the functional safety analysis
of TCS and TPS indicate that assumed architectures
TCS ensures the safety integrity level SIL1. This low
level is mainly due to the safety integrity of subsystem C (a control valve V-TCS) evaluated on the level
of SIL1 (Fig. 5). The subsystem A (sensors, transducers and converters) contributes less significantly
to the result obtained (SIL2) and the contribution to
PFDavg of subsystem B (the E/E/PE system) is very
low because of SIL3.
The TPS ensures the safety integrity level 2 (SIL2),
as it can be seen in Figure 6. It is mainly due to SIL
of subsystem C (a control valve V-TPS) evaluated as
SIL2. Subsystem A (sensors, transducers and converters) contributed less significantly (SIL2) and the
contribution of subsystem B (E/E/PE system) to the
PFDavg is low thanks to SIL3.
Resulting safety integrity levels of protection layers
are shown in Figure 7. Due to significant dynamic of
the system for most initiating events and high dependency of human actions the human error probability
was evaluated as very high (HEP 1). Therefore,
the risk reduction by TCS and TPS is on the level of
103 (SIL1 + SIL2).
Thus, the risk reduction required at level of 104
is not met and the design improvements have to be
considered. The analysis has shown that it is not possible at reasonable costs to make a higher SIL of TCS
(from SIL1 to SIL2). Therefore other design option
was taken into account to make higher SIL for TPS
(from SIL2 to SIL3see right block in Fig. 7).

A.
STC

B.
E/E/PES

SIL2

SIL3

Figure 5.

SIL1

Subsystems of the turbine control system (TCS).

A.
STC

B.
E/E/PES

SIL2

SIL3

Figure 6.
(TPS).

C. EUC
V-TCS

C. EUC
V-TPS
SIL2

Subsystems of the turbine protection system

PL1
TCS

PL2
HO

PL3
TPS

SLI1

HEP~1

SLI2->SIL3

Figure 7. Final safety integrity levels in protection layers of

the turbine shut-down system.

It is proposed to increase SIL of TPS increasing

SIL of subsystem A from SIL2 (see Fig. 6) to SIL3
and of subsystem C from SIL2 to SIL3. In subsystem
A the architecture 2oo3 is proposed to lower PFDavg
and to lower frequency of TPS spurious operation. For
increasing SIL in subsystem C it is not justified to
implement 1oo2 structure due to too very high costs.
Therefore, the regulatory aspects (the valve selection
of a high quality) and testing strategy of the shut-down
valve must be adapted to meet requirements for SIL3
(see interval criteria in Table 1).

CHALLENGES IN SAFETY MANAGEMENT

REGARDING HUMAN FACTORS

4.1 Advanced solutions influencing

human-operator reliability
As it has been mentioned the hazardous industrial
plants require integrated safety management system
oriented not only on technical aspects, but first of all on
human and organizational factors. Below two aspects
are emphasized related to: (1) the dependability of
safety-related systems including human-operator reliability, and (2) the quality and safety aspects of
computerized systems used for monitoring, control
and protection.
Nowadays more and more designers and safety
managers are convinced that knowledge concerning
the human performance capabilities as well as the
human and organizational factors should be a basis for
development of advanced technology for monitoring,
control and protection of hazardous systems. There
are new ideas how to design an advanced control room
concentrating on human factors.
From the safety point of view of hazardous plants
it is crucial to support the operators using advanced
tools for safety-related decision making oriented on
reducing probabilities of human errors. The idea of an
intelligent operator support system (OSS) is usually of
interest, integrated with an on-line diagnostic system
and an on-line dynamic risk assessment system. Such
OSS could be also used in situations of partial faults
and dangerous failures within E/E/PE safety-related
systems or SISs, contributing to reducing the probability of human errors and risks of abnormal states.
The OSS should provide intelligent advice during accidents based on advanced procedures that combine the
symptom-oriented procedures and the dynamic risk
oriented procedures.
4.2 Towards integrated knowledge-based approach
in managing safety including human factors
Current approaches used for the safety management of programmable systems are not coherent.

255

The research works have been undertaken to develop

advanced methods, data/knowledge bases and tools
for knowledge-based computer aided functional safety
management (Kosmowski 2007) oriented on:
Determining required SIL for safety-related functions with using (a) quantitative risk models or
(b) the risk graphs defined using qualitative information with relevant knowledge bases and interactive graphical support,
Verifying SIL for E/E/PE or SIS using (a) quantitative probabilistic models or (b) block diagrams defined using qualitative information with
relevant data/knowledge bases and interactive
graphical support for the design and operation
phases,
Knowledge bases for supporting evaluation of
dependent failures and human errors on the level of
safety-related functions and systems, and accident
scenarios,
Knowledge bases for supporting designing and
managing an integrated safety and security strategy
in the life cycle for systems operating in network.
The quantitative risk model can be a basis for developing an on-line dynamic risk assessment module
within OSS to reduce human error probabilities and
risks of abnormal states.

CONCLUSIONS

The paper outlines the knowledge-based methodology

of layer of protection analysis (LOPA) for industrial hazardous systems with regard to human and
organizational factors (H&OF). The layers of protection include basic process control system (BPCS),
human-operator (HO) and safety instrumented system
(SIS) that can be dependent due to influencing factors
involved. The human and organizational factors are
incorporated into the probabilistic models using a rulebased human reliability analysis taking into account
some existing HRA methods.
The functional safety oriented approach offers
a framework for more extensive HRA analysis with
emphasis on context-oriented human-operator behavior in situations of danger and safe failures (to be
potentially partly detected and partly undetected) with
regard to the safety-related functions of control and
protection systems. They should be designed with
regard to knowledge concerning the human behavior and potential errors. Some additional research is
needed to provide more comprehensive insights concerning the contextual influence of human factors to be
included in HRA in the context of safety-related functions to be designed using the programmable control
and protection systems.

ACKNOWLEDGMENTS
The author wish to thank the Ministry for Science
and Higher Education in Warsaw for supporting the
research and the Central Laboratory for Labour Protection (CIOP) for co-operation in preparing a research
programme concerning the safety management of
hazardous systems including functional safety aspects.
REFERENCES
Byers, J.C., Gertman, D.I., Hill, S.G., Blackman, H.S.,
Gentillon, C.D., Hallbert, B.P. & Haney, L.N. 2000.
Simplified Plant Risk (SPAR) Human Reliability Analysis (HRA) Methodology: Comparisons with Other HRA
Methods. INEEL/CON-00146. International Ergonomics
Association and Human Factors & Ergonomics Society
Annual Meeting.
Carey, M. 2001. Proposed Framework for Addressing Human
Factors in IEC 61508. Prepared for Health and Safety
Executive (HSE). Contract Research Report 373. Warrington: Amey Vectra Ltd.
COA 1998. Critical Operator ActionsHuman Reliability Modeling and Data Issues. Nuclear Safety,
NEA/CSNI/R(98)1. OECD Nuclear Energy Agency.
Dougherty, E.M. & Fragola, J.R. 1988: Human Reliability Analysis: A Systems Engineering Approach with
Nuclear Power Plant Applications. A Wiley-Interscience
Publication, New York: John Wiley & Sons Inc.
Dougherty, Ed. 1993. Context and human reliability analysis. Reliability Engineering and System Safety, Vol. 41
(2547).
Embrey, D.E. 1992. Incorporating Management and Organisational Factors into Probabilistic Safety Assessment.
Reliability Engineering and System Safety 38: 199208.
Gertman, I.D. & Blackman, H.S. 1994. Human Reliability
and Safety Analysis Data Handbook. New York: A WileyInterscience Publication.
Gertman, D., Blackman, H., Marble, J., Byers, J. &
Smith, C. 2005. The SPAR-H Human Reliability Analysis Method. Idaho Falls: Idaho National Laboratory.
NUREG/CR-6883, INL/EXT-05-00509.
GSAChP 1993. Guidelines for Safe Automation of Chemical
Processes. New York: Center for Chemical Process Safety,
American Institute of Chemical Engineers.
Hickling, E.M., King, A.G. & Bell, R. 2006. Human Factors
in Electrical, Electronic and Programmable Electronic
Safety-Related Systems. Warrington: Vectra Group Ltd.
Hollnagel, E. 2005. Human reliability assessment in context. Nuclear Engineering and Technology, Vol. 37, No. 2
(159166).
Humphreys, P. 1988. Human Reliability Assessors Guide.
Wigshaw Lane: Safety and Reliability Directorate.
IEC 61508:2000. Functional Safety of Electrical/ Electronic/
Programmable Electronic Safety-Related Systems, Parts
17. Geneva: International Electrotechnical Commission.
IEC 61511:2003. Functional safety: Safety Instrumented
Systems for the Process Industry Sector. Parts 13.
Geneva: International Electrotechnical Commission.
IEC 62061:2005. Safety of MachineryFunctional Safety of
Safety-Related Electrical, Electronic and Programmable

256

Electronic Control Systems. Geneva: International Electrotechnical Commission.

Kosmowski, K.T. 2004. Incorporation of human and organizational factors into qualitative and quantitative risk
analyses. Proceedings of the International Conference on
Probabilistic Safety Assessment and Management (PSAM
7ESREL 04). Berlin: Springer, Vol. 3: 20482053.
Kosmowski, K.T. 2006. Functional Safety Concept for
Hazardous System and New Challenges. Journal of Loss
Prevention in the Process Industries 19: 298305

Kosmowski, K.T., Sliwi

nski, M. & Barnert T. 2006. Methodological Aspects of Functional Safety Assessment. Journal of Machines Operation and Maintenance (ZEM, Polish
Academy of Science), Vol. 41 (148): 158176.

Kosmowski, K.T. (ed.) 2007. Functional Safety Management in Critical Systems. Gdansk: Gdansk University of
Technology.
LOPA 2001. Layer of Protection Analysis, Simplified Process Risk Assessment. New York: Center for Chemical
Process Safety, American Institute of Chemical Engineers.
Rasmussen, J. & Svedung, I. 2000. Proactive Risk Management in a Dynamic Society. Karlstad: Swedish Rescue
Services Agency.
Reason, J. 1990. Human Error. Cambridge University Press.
Swain, A.D. & Guttmann, H.E. 1983. Handbook of Human
Reliability Analysis with Emphasis on Nuclear Power
Plant Application. NUREG/CR-1278.

257

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

How employees use of information technology systems shape reliable

operations of large scale technological systems
Thale K. Andersen, Pl Nsje, Hans Torvatn & Kari Skarholt
SINTEF Technology and Society, Trondheim, Norway

ABSTRACT: Modern large scale technology systemslike power gridsreliance on information systems is
considerable. Such systems employ not only one, but a range of different information systems. This creates three
important, interdependent challenges for safe and reliable operations. The first is the sheer volume of systems,
which tie up organisational members in bureaucratic work, removing them from operational tasks and thus
introduce additional stress. This implies that the employees must be willing to speak out their previously tacit
knowledge, rules of thumb and know-hownot written in formal job instructionsand enter this information
into the systems, risking to loose personal assets relating to career and workplace identity, being thus a second
challenge. The third problem relates to data quality. Without valid and reliable data the systems will not have
any real value. The systems rely on the quality of key information entered by organisational members.

INTRODUCTION

Since the early 1990s Norwegian electrical power

grid companies have focused on developing, implementing and upgrading their Network Information
Systems (NIS)systems containing information on
the condition and functioning of the electrical network
and its components, as well as geographic position.
There is increased focus on economic performance and
more rigorous demands from governmental regulating
institutions in terms of reporting. The decision to introduce digital NIS was typically based on an expected
profitability from cost-benefit analysis. Earlier systems were costly in terms of man hours, with poor
data quality, and not supporting geographical information. Furthermore, they would in most cases offer
no functionality beyond documentation purposes, thus
not supporting new requirements for decision-making.
The cost-benefit analysis would in most cases identify
several areas, not only reporting, where substantial
savings could be obtained. More than a decade later
the results prove to be modest, especially regarding
economic gains. The NIS remain underutilized, data
quality is poor, new work practices are not fully integrated, and old systems and work practices run in
parallel with new ones.
This story is not unique to the Norwegian context. While the exact details in the development and
use of information and communication technology
(ICT) will vary between countries and sectors depending on a range of characteristics and conditions, this
is the story of work life in the nineties: how large

scale technological organizations introduce large scale

technological information systems in order to improve
organizational efficiency and profit. The outcomes of
these processes are not specific to a Norwegian context, either; international research on information systems reveals many similar examples. As argued; The
review highlights the relative lack of overall principles
or frameworks for improving information management per se and the overall information management
infrastructure including people, practices, processes,
systems and the information itself (Hicks 2007:235).
1.1

Three challenges for safe and reliable

operations

The introduction of ICT systems creates three important interdependent challenges tied to work practices
and processes for safe and reliable grid operations. The
first is the understanding of the interaction between
technology and organisation, identifying actual as
opposed to formal work practice in the adaptation and
implementation of ICT. Second, employees must be
willing to change their occupational identity and share
their tacit knowledge, rules of thumb and know-how
and enter this information into the systems, which
might have consequences for their careers. The third is
the shift in balance between operational and administrative tasks for the individual. This shift forces the
individual to work more with computers. ICT systems rely on input from people codifying and storing
information. Thus the systems tie up employees in
bureaucratic work, removing them from operational

259

tasks and adding strain. Research has shown negative

effects on health and safety from all this paperwork
in the offshore sector (Lamvik & Ravn, 2004). These
challenges all relate to data quality. Without highly
valid and reliable data the ICT systems are not functional. And data validity and reliability depend on the
precision and accuracy of the employees entering key
information. Earlier research indicates that employees
are not necessarily trained, motivated or equipped to
handle these tasks (Nsje et al. 2005).
1.2

Outline of the paper

This paper will mainly look at the installers role in

power grid safety, and starts with a study of grid operating companies current situation based on a variety
of empirical data collected over a two-year period in
two different companies located in two major cities in
Norway.
Theoretically, our research-in-progress is inspired
by the works of Barley and Kunda, Orr, Zuboff and
Orlikowski on work practice (Barley, 1996; Orr, 1996;
Barley and Kunda, 2001; Orlikowski, 2002; Zuboff,
1988). The underlying idea is to map and describe
how work is carried out in order to get to a better
understanding of how to use ICT in grid maintenance
organizations in order to ensure safe grid operations.
As argued by Barley and Kunda (2001) a shift is needed
in organizational thinking where a focus on the processual aspects of ICT-based work is more central as the
nature of work is changing. Tasks and roles must be
redistributed, the transfer of knowledge and technology need to be rethought, and routines and workplace
values should be renegotiated.

The qualitative data has been stored an analyzed in

Nudist 6. The analysis has been following explorative
analytic methods. First, single interesting observations have been tracked and identified in the program.
Second, these observations have been interpreted
with the goal of identifying generic issues in the
material (i.e. knowledge sharing, translation, competence). Third, the generic issues have been paralleled with relevant background variables in the data,
such as age, ICT-proficiency, function, company and
so on.

METHOD AND DATA

This is a qualitative study, funded by the Norwegian

Research Council and the two grid companies
involved. Our project group has been working closely
with two large power grid companies since 2006. Over
the two years we have conducted 52 semi-structured
interviews (mostly face to face, but also some via
telephone) with employees in all positions in the companies, from top management to installers. We have
carried out four field visits/observations, each of them
lasting several days. We have had access to a wide
range of information and documentation describing
rules and procedures. During the project period we
have also reported and discussed preliminary findings
with the two grid companies on several occasions.
Feedback and discussions of findings have involved
personnel in all positions, but focused on management.
All interviews have been taped and transcribed.
Field notes have been written, and photos taken.

FINDINGS

The two companies we have studied have both chosen to invest heavily in the same large scale Net
Information System (NIS). They have the same goal
for their investment, more cost-efficient and reliable
power grid operations. In this chapter we will describe
the two companies, the work practice of their planners and installers, as well as their ICT reality and
implementation strategy.
3.1

Description of companies and company groups

The major task of a power grid company is to develop

and maintain its part of the national power grid,
so electrical power can be delivered to all its customers. Since electrical power is vital to a modern
society the sector is carefully and strictly regulated,
and any failure to deliver will result in sanctions
from the national authorities. Power grid companies
today usually employ an organisational model called
planner and doer. The planners analyse and plan
tasks, deciding when, where and how grid maintenance should take place. They sketch development
projects, bill customers, report to authorities and so
on. The planners are usually educated as BSc or MSc
in electricity.
The doershere after called installerscarry out
the manual work designed by the planners, and report
both how they carried out their work and the state of
the grid after they have performed their tasks. Both
these reports (work conducted and state of grid) are
fed into NIS for the planners to use in their further
work. Unlike the planners, most of the installers do
not have a college or university degree; rather they
have vocational education, with various certifications
which allows them to carry out a range of physical
labour grid operations.
This has to do with what we will call the dilemma
of data use: The total dependence from the side of the
planners on data reported from the part of the installers,
and the lack of perceived usefulness of the ICT from
the side of the installers, which does not motivate them
to use and report to the systems.

260

The two companies are similar in several respects:

Both are among the largest grid companies in
Norway
Both operate complex grids in big cities (by Norwegian standards) and surrounding areas with a high
number of customers
They both employ the same NIS
Both have a history of growth by acquisition the last
decade
Both started implementing NIS in the mid-90s, and
carried out huge data capture projects to get NIS
data initially
Both acknowledge that gains from NIS has been
smaller and slower forthcoming than initially calculated, but management maintains that NIS is crucial
for their operations
The two companies differ in one important dimension. One of the companies, Company X, has chosen
to outsource its installation work, and hence does not
itself employ any installers. Rather, it hires all its
installers, project by project. Company Y, on the other
hand, have kept their installers, and is in this respect an
integrated company with planners and installers in the
same organisational unit. While there are important
differences between outsourced and integrated companies our experience tells us that there are also lots of
similarities. In this paper we focus on the work practice
and role of the installer, and from what we have seen
in our interviews and fieldwork the similarities among
installers clearly outweigh the dissimilarities when it
comes to the handling and use of NIS. However, it
should be noted from an occupational health point of
view that the outsourced installers complained about
deteriorated working conditions in a number of areas
after being outsourced. A survey where this issue is
investigated further is in progress.
3.2

Integration of ICT in the grid companies

Modern power grid operations are completely dependent on complex and secure ICT-solutions as tools for
safe and reliable functioning, and for becoming more
cost-efficient. The quality of information is vital in
order to ensure safe operations, involving trustworthiness, relevance, clarity, opportunities and availability.
The quality of the information flow is equally important, including the attitude to learning, cooperative
climate, process and competencies. Data quality thus
has prime priority and is a condition for ICT-based
cooperationthe data systems in use must be correct
and continually updated. ICT makes this multi-actor
reality possible.
NIS is installed and in use today among planners and management in the companies, but a key
challenge is the installers lack of competencies
in computers and data systems. Another important

challenge is the installers attitude to the use of ICTbased tools as an inherent part of their routines and
work practicesshowing the identity issues lying
between working physically on the grid versus increasingly working with a system. The question is which
organisational elements need to be considered in order
to facilitate this change in work practice, intensifying
the interplay between technology and organisation.
3.2.1 The planners use of ICT
The planners and the management of the two enterprises use NIS in their daily work. An area where
use of NIS has been successful is authority reporting.
NIS has simplified and automated several of the steps
in yearly reports. Such reporting was a process that
earlier was manual, cumbersome and time consuming
(Nsje et al. 2005).
NIS is also used as a tool for early stage projecting when a new project is started, usually before it is
decided when or by whom the project will be carried
out. When determined, a more detailed planning is
done. However, for this more detailed planning work
NIS is considered to be awkward and not up to date.
The planners state that: We use NIS as a first step for
simple and overall calculations. We dont use NIS for
detailed project planning. To do so is inconvenient;
NIS lacks basic features, and data is unreliable. Once
you get under the ground everything looks different
anyway.
Thus, the project planners want to get out in the
field, to have look at it and discuss it with the doers.
While the goal of top management is that planners
should be able to do all the planning without leaving
their office desks the planners themselves believe in
visiting the project site, discussing with the installers,
and rely on the installers local knowledge as far as
possible. They simply do not trust data in NIS to
be good enough to carry out project work, without
calibrating with the physical reality.
All though NIS is in use among the planners, they
are aware of the systems shortcomings regarding data
quality, and in order to carry out projects smoothly
and establish reliable grids they prefer hands on local
knowledge instead of decontextualised information in
an ICT system.
3.2.2 The installers use of ICT
Most installers do not use NIS other than in the
most limited way possible. Some of them manage to
print out topographic instructions entered by planners,
while most get a printed copy from his project leader
which is used as a reference out in the field. Where
the NIS print-outs do not concur, they deal with the
unforeseen problem physicallydigging holes in the
ground checking the situation themselves, and calling
their employers, contractors and subcontractors.

261

What regards the use of computers as the installers

main tool for ICT-based operations; while the companies has provided most installers with a personal
computer, during our visits we only rarely saw one of
them opened. Most of the installers have 20 years of
experience or more, and do not see the computer as a
primary work tool. Another drawback is that most of
the installers have very little knowledge of computers,
let alone data systems. And as the data systems are
not sufficiently updated, they often encounter unforeseen challenges in the field. As said by two of the
installers, respectively 39 and 36 years old: We didnt
become installers because we liked going to school or
because we were into computers, and Now I just
got a live-in partner, and she has started to show me a
thing or two about computers and the internet. The
installers have a work environment characterised by
strong social relations and a high degree of professional pride. They express regret for the decreasing
focus on professional development and training, and
the unpredictable working situation makes it necessary
to have several ongoing projects at the same time.
4

DISCUSSION

The discussion will focus on important organisational

processes necessary to be explicit on in order to make
installers use NIS more actively. First in this part we
will look at how ICT shape safe and reliable operations
today, and then at how to understand the interaction
between ICT systems like NIS and the installers. It will
further be outlined how ICT may represent a threat to
the occupational identity and career development of
installersmeaning that work identity can constitute
a considerable barrier to use and thus to more reliable operations. Finally, as bureaucratisation of work
imposes an important additional challenge, it is important to consider its implications for the use of NIS
among the installers. A summary of barriers to good
implementation of ICT systems as well as some guidelines on how to integrate NIS in installers working
reality will conclude the discussion.
4.1 Man-machine interaction and the social
construction of technology
The multitude and variance of the challenges tied to
the interaction between ICT-based grid operations and
installers outlined in the previous section show that the
relationship between human factors and use of technology is complex. Installers are dependent on high
quality data from updated databases and reliable information in order to reach a high degree of accuracy
in their operations. In order to handle the ICT-based
systems, the installers need to be skilled not only in
ordinary computer technology, but also in the actual

software. They need the ability to enter and transmit

new information, as well as the ability to find, extract,
interpret, mediate and use data.
Understanding the organisation as not only defined
through its operations and products but as an interpretative entity, is to see that organisational processes, including the implementation of ICT, take
place through communication; we create and share
meaning through dialoguedialogues mediate learning and is a collective tool. It is thus important to
establish a culture for learning that goes beyond formal coursing. Not least because Formal descriptions
of work and of learning are abstracted from actual
practice. They inevitably and intentionally omit the
details. In a society that attaches particular value to
abstract knowledge, the details of practice have come
to be seen as nonessential, unimportant and easily
developed once the relevant abstractions have been
grasped (Brown & Duguid, 1991, p. 40).
Conventional learning theorywhich most studies
part fromemphasise abstract over practical knowledge, and thus separate learning from performance,
believing that learning constitutes the bridge between
working and innovating. And thus emerges an understanding of complex work processes as decomposable
in smaller and less complex parts where it is not necessarily a precondition that the worker understands what
he or she is doing (Brown & Duguid, 1991). This
makes it seem like installers perform their tasks consistent with a formal job description, which is also
used as a decision-making tool for diagnosing and
problem-solving. But when something goes wrong the
installers seldom will look at the formal job description, but use his or hers long experience and will
eventually develop a non-sequential practice in order
to complete the non-covering formal job description.
Howeverthe installers will not use NIS properly if
they do not understand its purpose; e.g. if it does not
correspond to their working reality.
Levin (1997) forward the view that both how the
technology works and how to work it in a particular setting is something that needs to be learned, understood
as knowledge that is socially constructed in the particular company. In this social process all types of knowledge interact, formal knowledge also, but especially
various forms of tacit and operational knowledge, not
least cultural skills. An organisation is always contingent of its organisational culture and procedures;
. . . routines are the factual organisation. . . [. . . ]. . .
The core essence of an organisational development
process will thus be the facilitation of learning to
shape a new understanding and the skills necessary to
change the organisation. (Levin, 1997, p. 301). Thus,
work practices or work routines must be transformed
and adapted to the new working reality, through negotiation between all employees affected by the change.
This might be especially useful in implementing the

262

understanding of how the system work, as well as why

it is introduced, for the use of the new ICTin this
case NIS.
4.2

Occupational identity and tacit knowledge

as threats to the integration of NIS

When using a computer, installers do not get the instant

satisfaction of seeing the outcome, and thus ICT might
not be perceived as particularly meaningful regarding
the outspoken demandsgetting the job done. On the
contraryin stead of being given the possibility of
extended interaction and integration with the administrative personnel as all employees now are supposed to
operate on and from the same platform (NIS), installers
feel increasingly alienated not only from their original
work tasks but also from the organisation itself, as the
new company policy implies a greater physical separation of the technical installers and the planners and
management. New ICT systems are thus implemented
physically, but not integrated in a meaningful way in
the installers working situation, as they correspond
on a very limited basis to their working reality.
By offering their installers two sessions 45 minutes in order to learn the new systems, which first of all
must be said to be too little, the management implicitly assumes that 1) their installers are familiar with
working with computers and ICT systems in the first
place, and 2) their installers automatically understand
the usefulness and impact of using the new systems, or
worse; that they need not know what they are doing at
work. This means that the companies take for granted
employees ability and willingness to learn and use the
new technology, with minimal effort from the side of
the organisation. Rather, the organisation must move
from a focus on the performance of physical work to
cognitively operated tasks, and these different ways of
work need to be learned in different waysthey are by
no means directly transferable to each other (Zuboff,
1988).
Change in work practice implies a change in routines. For true ICT-based change to happen, installers
need to change their work practice, which is something that will not easily occur unless they get a good
reason to do so, other than extrinsic rewards or sanctions (Nelson & Winter, 1982). Routinesby some
also called norms or institutionalised behaviour, processes and structuresbecome important in defining
and expressing employees occupational identities. If
installers shall be willing to change their routines, they
will need a meaningful replacement of the occupational identity built by these routines. Installers may be
less receptive to use new ICT as they have been doing
perfectly fine up to now without it. When their manual tasks are automated, they might have difficulties in
perceiving the change as valuable. From working in a
reality where productivity outcomes are concrete and

immediate, the installers are expected to operate more

through computers at the same time as they are told
to lessen the physical contact with the other main part
of the organisation, namely the part that might have
helped them create references with regard to the new
technology.
Occupational identity may thus be a potential source
of resistance against ICT-based change, as installers
experience a devaluation of what is their professionalism, and thus occupational identity, without a proper
replacement. Experiencing ambiguity related to the
use of ICT implies a feeling of insecurity vis--vis the
new technology as the informating side has not been
implemented together with the actual computers and
software. Installers are told to let go of their previous
work identitybeing valuable in that it represent the
profession of being a technical installer, but without an
equivalent valuable identity offered as a replacement
it will not be perceived as interesting to do so.
An individual asset that may hamper the proper
implementation and integration of NIS is the installers
tacit knowledge tied to work tasks and organisational
functioning. In order to develop and use NIS, individuals and teams must be willing to share their
knowledge. In times of organisational change based
on ICT, the changes often imply such a great redefinition of boundaries tied to expertise and responsibilities that employees may feel that theiroften
hard-builtposition in the formal or informal hierarchy is threatened. A way to obtain protection is
not to share ones specific knowledge and perhaps
even consciously produce counter-activities towards
the systems proper functioning (Schn, 1983; Argyris
& Schn, 1978). This behaviour may be perceived by
the individual installer as the protection of his career.
Knowledge and authority are redistributed and forms
a new hierarchy through the implementation of ICT;
those above the technology and those understanding
and thus handling it best are those with the highest
status. According to Zuboff (1988, p. 391), those . . .
who must prove and defend their own legitimacy do not
easily share knowledge or engage in inquiry. Workers who feel the requirements of subordination are
not enthusiastic learners. New roles cannot emerge
without the structure to support them.
4.3

The bureaucratisation of grid operations

The strict meaning of the word bureaucracy is civil

service, and is traditionally associated with paperheavy office-work. The bureaucracy exerts a specific
type of authority and power, with a great number of
rules, regulations and routines (Weber, 1974). The
rationality behind the bureaucracy is standardisation,
legitimacy and easy error-detection. In the power grid
industry the new bureaucracy to an increasing extent
are the total of the ICT-based systems which needs to

263

be handled in the correct manner by employees, so they

can perform their work tasks as stated in the company
policy. ICT-induced bureaucratisation involves, as the
findings in this study show, a strong focus on data.
High quality data further implies a vigorous emphasis on the collecting of the correct information, the
correct encoding and storage of informationthat the
collected data are placed in the correct systems and
databases, and in the correct categories, and the correct
use of stored information, entailing correct interpretation of these data. This illustrates that the quality
and reliability of data to a high degree is dependent
of the installers handling it, and thus their subjective
judgment, abilities, willingness and motivation.
For teams, which often find themselves faced with
new tasks due to automation and the elimination of
routine jobs, it implies a simplification of communication and infrastructural support. But, communication
is more than the transmitting of symbols from one person to another. It also includes absorption and understanding as well as practical use of the transmitted
informationthe significance can be difficult to keep
through coding and decoding. Information technology
makes it possible to operate quantitative information
that earlier was too complex to handle, because of the
calculation-load and too many causal relationships.
The limiting factors are the costs of the information
and our ability for understanding and using it.
The challenges imposed by the inherent characteristics of ICT are above all again connected to the
interaction between the installers and the technology,
showing that the bureaucratic effect it imposes separate the employees from each other, contributing to
the disruption of social relations in the workplace,
and gives negative effects on health and safety of
employees, in struggling with the extra work this
bureaucratisation represents. Trist (1981) refers to
what he calls the primary work system, which is the
concrete work group an employee belongs to and the
framework for the performance of his or hers daily
work tasks. According to Zuboff (1988), technological
change creates limiting conditions for what is possible,
as technology is not neutral but specific. This gives a
new angle vis--vis Groths (1999) theoretical standpoint that humans are the main restricting agents for
the possibilities for technological and organisational
development. Both might be right. But technology
does not exist in a vacuum. In order to be useful it
needs to interact with both human and organisational
systems. Bureaucratisation is an abstraction of work
that must be explicitly considered.
4.4

Barriers to use and some guidelines for the

successful integration of ICT systems

We have seen that the challenges to the use of NIS in

our two grid companies are numerous, heterogeneous,

and complex. Organisational challenges are general

un-preparedness when it comes to abilities, motivation and perceived usefulness of the systems from
the part of the installers, the lack of a learning
culture including awareness of the systems contextspecific conditions for use (local informal knowledge
and actual work practice), failure to recognise the
implications of ICT-based organisational change for
installers occupational identity and career development, as well as little insight in the consequences of
the bureaucratisation of work.
In addition to these more universal challenges,
empirically based on a range of different studies, we
identified, in accordance with other empirical findings
(Venkatesh et al. 2003), also a set of specific factors
being especially relevant for the grid companies in
question. The first is work experienceit seems clear
that among the installers having worked 20 years or
more (and most have) in the same field of work with
approximately the same methods as when they started,
the motivation is not very high when it comes to integrating manual and digital operations. This finding is
of course tightly bound to the next; the lack of competence and thus self-confidence among the installers
when it comes to the handling of computers. The transition from action-oriented to cognitive-based skills
may be seen as unattainable. A third barrier is earlier
and existing work practice. Existing work practice is
to a large degree informal and characterised by a number of heuristic and ad hoc methods. And finally, the
physical equipment must be easy and convenient to
use. In order to be able to use NIS and other systems
while working out in the field, the installers need to
have optimal network availability. For example, loosing network connection when driving through a tunnel,
and having to restart the computer in order to get back
on track again, is not optimal. And likewise, it must be
easy and convenient to install the personal computer
in the car.
In order to reduce the gap between formal and informal theory, and thus between NIS and the installers,
the organisations management need to engage in an
inquiry to make explicit the underlying assumptions in
the installers theory-in-use. Schn (1983) launches
organisational inquiry as a tool for organisational
change and development. An important element in
this organisational inquiry is to raise consciousness
of practice, values and underlying assumptions tied to
actual performance and then compare it to the written and formal procedures. Inconsistencies need to
be identified and corrected in order to secure safe
operations. Moreover, routines must be established
for information sharing, relating both to tacit knowledge concerning work tasks as well as competencies
when it comes to computer handling. These routines
must be perceived as necessary and beneficial by the
installers in order to work properly, which is of crucial

264

importance for data quality and reliability. In this

respect it is important to be aware of possible obstacles
to organisational learning.
4.5

How ICT shape safe and reliable operations

As for today ICT systems like NIS seem to have a

smaller impact on safe and reliable operations than we
might have expected after almost two decades of use.
While planners use NIS for reporting and doing some
project work, they also supplement information from
NIS with on-the-spot observations and discussions
with those who actually implement their planning.
Installers to a large degree avoid using NIS. They
have neither training, nor motivation, and does not
need to use NIS. For them NIS is not a support tool
but rather a control mechanism which they are supposed to report to, in order to document their work
progress. However, also installers are given their work
orders and maps from NIS, and hence to some degree
use NIS information as their point of departure for their
work. The understanding of the usefulness of NIS on
an abstract level is not self-evident, and reporting is
something they try to do as little of as possible. In
the field they prefer to rely on their own and their colleagues local knowledge. Thus, installers and planners
avoid relying entirely on a system without high quality
data. This is a two-sided dilemma, since high quality
data is dependent on high quality reporting and use of
the system.
Still NIS is being used in decision making. It is
used in buying services (in the tender processes), when
it comes to fixing the priorities of maintenance and
development, it is used to prepare work orders for the
installers, and it is used for reporting purposes. And
use and reliance on NIS is increasing as more and more
data is entered, and more and more functions are built
into the system. While detailed project planning is difficult today, maybe the next version or the version after
will be better. Top management is still likely to pursue
the goal of the planner who does not leave his desk.
While both installers and planners today to a varying extent ignore and distrust NIS, in the future more
and more decisions, work and operations will depend
on the ICT systems. In this perspective it is necessary
to get installers to use and report to NIS.

CONCLUSION

Grid operating companies after market deregulation

embraced ICT systems in order to become increasingly
effective and fast-responding.
The companies have focused on information systems in general, and NIS in particular for more than
a decade. The maximum efficiency and cost-saving

aspects of NIS were seen as attainable through an overall level implementation of the systemfrom management in order to analyse the reliability and efficiency
of grid operations to installers, using data to locate
exactly where they should operate geographically out
in the field, reporting back standard work data, as well
as inconsistencies between the information systems
and the actual field in order to continuously update
the systems. ICT is thus significant both as a grid
controlling and as an operating device.
As the two companies see it, they also need to
become efficient ICT-based organisations so as to be
able to respond to the increased complexity of their
working environment and organisational structures, in
order to continue to carry out reliable grid operations.
But because of too little focus on strategy and necessary conditions for continuous use and functioning,
the implementation process has more or less stagnated. Results show that while planners are coping
fairly well based on previous experience with computers and constant interaction with management and
support functions, the installers only have been given
between one and two hours of training in using their
new equipment, which basically has been the only
interaction they have had with other organisational layers regarding the implementation and use of the new
technology.
Recent research on organisational change put
emphasis on the importance of well conducted change
processes, important elements are taking existing
social norms into account, being aware of workforce
diversity, making conflicts constructive, clear distribution of roles and responsibilities, and the managements availability (Saksvik et al. 2007). Orlikowski
(1996) views organisational change as enacted through
the situated practices of organisational actors as they
improvise, innovate and adjust their work routines.
As ICT engender automating as well as informating capacities (Zuboff, 1988), implementation of NIS
requires a process where dialogue and negotiations are
key. Technology transfer as the learning of new ICT
must be recognised as a collective achievement and
not top-down decision-making, or grid reliability will
remain random and difficult to control.
The studied companies learned the necessity of
rethinking several aspects of information handling
in the process of becoming ICT-based somewhat the
hard way. The systems are implemented but remain
disintegrated, and thus data quality is poor. Some problemsintrinsically tied to the lack of focus on systems
specifications before the start of the implementation
processes, were the non-compatibility of the old systems with the new ones, as well as poor routines for
data quality assurance and data maintenance. The lack
of synchronisation between the development of the
software and which functionalities that are actually
needed in order to use it as operating device, is also a

265

major reason for this underutilisation, and the use of

the system is not equally rooted in all organisational
layers and subgroups.
Among management and project leaders the necessity of the systems seems clear, while installers
have not been sufficiently explained the relationship
between the systems and their specific tasks, and thus
they dont see the relevance of adjusting their work
practices to system requirements. As an earlier study
showed; Data will not come before use, and use will
not come before there is data. Thus, data collection
must be supported or pushed by stakeholders, especially management and different operational units.
(Nsje et al. 2005, p. 5). The analyses show that it
is crucial to focus more strongly on aspects tied to
the collection and the management of data, as this to
a large degree was neglected in the selection of software. The grid operating company must on one hand
decide to what extent the individual installer shall be
given access to the systems, and if, whether read-only
or editing. On the other hand, the company needs
to make sure that the intended use of these systems
gets incorporated into the routines and work practices of their installers; The lack of closeness between
information and operational tasks is a major barrier,
hindering efficient completion. (Nsje et al. 2005,
p. 5). The shift from local knowledge on how to solve
taskspresent in the head of the installers, to centralised knowledge leaning on integrated data in a
common database, is an important challenge for safe
grid operations.

REFERENCES
Alvesson, M. (1993). Organizations As Rethoric: KnowledgeIntensive Firms and the Struggle with Ambiguity. Journal
of Management Studies, 30(06), 9971015.
Argyris, C. & Schn, D. (1978). Organizational learning:
A theory of action perspective. USA: Addison Wesley
Publishing.
Barley, S.R. (1996). Technicians in the workplace: Ethnographic evidence for bringing work into organization
studies. Administrative Science Quarterly 41(3), 404.
Barley, S.R. & Kund G. (2001). Bringing work back in.
Organization Science 12(1): 76.

Brown, J.S. & Duguid, P. (1991). Organizational Learning

and Communities of Practice: Toward a Unified View
of Working, Learning, and Innovation. Organization
Science, 2(1), 4057.
Groth, L. (1999). Future organizational design. Chichester,
UK: John Wiley and sons.
Hicks B.J. (2007). Lean information management: Understanding and eliminating waste International Journal of
Information Management 27(4), 233249.
Lamvik, G. & Ravn, J.E. (2004). Living safety in drilling:
how does national culture influence HES and working
practice? (No. STF38 A04020). Trondheim: SINTEF
Industrial Management, New Praxis.
Levin, M. (1995). Technology Transfer is Organizational DevelopmentAn investigation in the relationship between technology transfer and organizational
change. International Journal of Technology Management, 14(2/3/4), 297308.
Nelson, R.R. & Winter, S.G. (1982). An evolutionary Theory of Economic Change. Cambridge: Belknap Press of
Harvard University Press.
Nsje, P.C., H.Y. Torvatn, et al. (2005). Strategic Challenges in Implementing NIS: Investigations on Data
Quality Managment. IPEC, 7th Int Power Engineering
Conference, Singapore, IEEE Conference Proceedings.
Orlikowski, W.J. (2002). Knowing in practice: Enacting a collective capability in distributed organizing.
Organization Science 13(3), 249.
Orr, J.E. (1996). Talking about machines: an ethnography of
a modern job. Ithaca, N.Y., Cornell University Press.
Saksvik, P.., Tvedt, S.D., Nytr, K., Buvik, M.P., Andersen,
G.R., Andersen, T.K. & Torvatn, H. (2007). Developing
criteria for healthy organizational change. Work & Stress.
21: 243263.
Schn, D. (1983). Organizational Learning. In G. Morgan
(Ed.), Beyond Method: Strategies for Social Research
(pp. 114128). Beverly Hills, Ca: Sage.
Trist, E.L. (1981). Evolution of sociotechnical systems. In
A.H. van de Ven & W.F. Joyce (Eds.), Perspectives on
organization design and behaviour (pp. 1975): John
Wiley & sons.
Venkatesh, V., Morris, M.G., Davis, G.B., & Davis, F.D.
(2003). User acceptance of information technology:
Toward a unified view. MIS Quarterly, 27(3), 425478.
Weber, M. (1974). Makt og byrkrati. Oslo: Gyldendal
Zuboff, S. (1988). In the Age of the Smart Machine. New
York: Basic Books.

266

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Incorporating simulator evidence into HRA: Insights from the data analysis
of the international HRA empirical study
S. Massaiu, P.. Braarud & M. Hildebrandt
OECD Halden Reactor Project, Norway

ABSTRACT: The observation of nuclear power plant operating crews in simulated emergencies scenarios
reveals a substantial degree of variability in the timing and execution of critical safety tasks, despite the extensive
use of emergency operating procedures (EOP). Detailed analysis of crew performance shows that crew factors
(e.g. leadership style, experience as a team, crew dynamics) are important determinants of this variability.
Unfortunately, these factors are problematic in most Human Reliability Analysis (HRA) approaches, since most
methods do not provide guidance on how to take them into account nor on how to treat them in predictive
analyses. In other words, factors clearly linked to the potential for errors and failures, and information about
these factors that can be obtained from simulator studies, may be neglected by the HRA community. This paper
illustrates several insights learnt from the pilot phase of the International HRA Empirical Study on analysis,
aggregation and formatting of the simulator results. Suggestions for exploiting the full potential of simulator
evidence into HRA are made.
1

INTRODUCTION

This article is based on insights from the empirical data

analysis of the first phase of the International HRA
Empirical Study (Lois et al. 2008). The project is an
international effort coordinated by the OECD Halden
Reactor Project, aimed at assessing the strengths and
limitations of existing HRA (Human Reliability Analysis) methods. The study is based on a comparison of
observed performance in HAMMLAB (HAlden Man
Machine LABoratory) simulator trials with the outcomes predicted in HRA analyses. The project goal is
to develop an empirically-based understanding of the
performance, strengths, and weaknesses of a diversity
of methods currently available. This article focuses on
empirical aspects of the project methodology.
2

THE INTERNATIONAL HRA EMPIRICAL

STUDY

The Pilot study (2007) was designed around four sets

of participants: 1) the operator crews, 2) the Halden
experimental staff, 3) the HRA teams, and 4) the study
assessment group.
In the period from October to December 2006, 14
crews of three licensed PWR operators per crew participated in an experiment performed in HAMMLAB
(Braarud et al. 2007). All crews responded to two versions (a base case and a more challenging complex

case) of two scenarios, a steam generator tube rupture

(SGTR) and a total loss of feed water (LOFW). This
article relates to the results of the analysis and comparisons of the so called pilot phase of the International
HRA Empirical Study, which is based on only the two
variants of the SGTR scenario. Further it is limited
to only the first and most important phase of this scenario type, consisting of identification and isolation of
the ruptured steam generator. These two scenarios are
two versions of a SGTR: one base case where there are
no other complications than the tube rupture, and one
complex case where the radiation indications to diagnose the tube rupture are lacking. The HAMMLAB
pressurized water reactor (PWR) emergency operating
procedures (EOPs) used in the study were based on the
Emergency Response Guidelines (ERGs) developed
by the Westinghouse Owners Group.
The Halden experimental staff conducted the simulator sessions in the HAMMLAB facility and had the
main responsibility for the analysis of the experimental
data. Regulators, utilities and research institutions provided the HRA teams. The teams applied one or more
HRA methods to obtain predictions for the Human
Failure Events (HFEs) in the scenarios defined for the
study. Table 1 summarizes the methods tested and the
teams applying them.
The Study assessment group acted as the interface between the HRA teams and the experimental
staff. As a starting point, and together with Halden
staff, it provided an information package (analysis

267

Table 1.

Methods tested and HRA teams.

Method

Team

Country

ATHEANA

NRC staff +
consultants
EPRI (Scientech)
PSI
NRI

USA

CBDT
CESA
CREAM
Decision
Trees + ASEP
HEART
KHRA
MERMOS
PANAME
SPAR-H
THERP
THERP (Bayesian
enhanced)
Simulations
IDAC
MICROSAINT
QUEST-HP

NRI
Ringhals +
consultants
KAERI
EDF
IRSN
NRC staff +
consultants, INL
NRC staff +
consultants
VTT/TVO/Vattenfall

University of Maryland
Alion
Riso
Politecnico di Milano

USA
Switzerland
Chech
Republic
Chech
Republic
Sweden/
Norway
South Kores
France
France
USA
USA
Finland/
Sweden
USA
USA
Denmark
Italy

inputs) to the HRA teams and answered requests

from the HRA teams for additional information and
questions concerning ambiguities in the instructions
and assumptions. The information package included,
among other, instructions to the HRA teams, scenario
descriptions and HFEs, characterization of the crews
(e.g. their work practices and training), the procedures used in Hammlab, and forms for the responses
of the HRA teams. Finally, the study assessment group
reviewed the HRA team responses and performed the
assessment and comparison of the predicted outcomes
vs. the experimental outcomes, together with input
from experimental staff from Halden.
2.1

Focus on the qualitative outcomes: driving

factors and PSFs

At a high level, HRA methods have the same purpose

(or aims) due to the role of the HRA within the PRA.
These common aims are: 1) an identification of the
HFEs to be included in the PRA accident sequence
model, 2) the qualitative analysis of the HFEs, and
3) the quantification of the probability of these HFEs.
The study focused mainly on qualitative analysis
and to a lesser degree on the quantitative analysis.
The HFEs were defined for the HRA analysts. Definition of the HFEs was needed in order to control for
variability in the interpretation of the various teams

as to what are the actions to be analysed. It should be

noted that defining the HFEs for the HRA teams did
not eliminate the qualitative analyses to be performed,
since the HFEs were defined on a functional level, i.e.
fails to perform X within Y minutes.
One of the most important aspects of HRA methods
is the identification and evaluation of the factors driving performance of the HFEs, commonly referred to
as performance shaping factors (PSF) or the driving factors of performance. Comparing the specific
factors identified as driving factors by the HRA teams
for the defined HFEs with those observed in Hammlab
was the main focus of the comparison.
The HRA teams typically provided their analyses and the outcomes they predicted in several ways.
They filled out one form (Form A ) that emphasized identifying the main drivers of performance in
terms of PSFs, causal factors, and other influence
characterizations explicitly identified through the
HRA method they were using. In addition, in Form
A, they were also asked to describe how and why a
scenario might be difficult or easy for the crews in
terms of the specifics of the scenarios, i.e. in more
operational terms. They were also asked to provide an
HEP for each HFE based on the application of their
method. Additionally the teams were asked to provide a Form B, a closed-form where the responses
had to be structured according to a modified version of the HERA taxonomy (Halbert et al. 2006).
Finally, the HRA teams were asked to provide their
method specific analysis documented according to
PRA (good) practices, which included the derivation
of the HEPs.
2.2

Comparison of predicted and experimental

outcomes

The core of the HRA empirical study is the evaluation

of HRA methods by means of comparing their predictions with observed performance. Comparisons were
performed on several levels:
On the factors that most influenced the performance
of the crews in the scenarios (driving factors).
On the level of difficulty associated with the operator actions of interest (the HFEs). For the HRA
predictions, the level of difficulty was mainly
represented by the HEP.
On the reasons for the difficulties (or ease) with
which the crews performed the tasks associated with
each HFE, and how these difficulties were expressed
in operational and scenario-specific terms (operational expressions).
In addition, several other factors were evaluated,
like the insights given by the HRA method for error
reduction, sensitivity issues such as the impact of

268

qualitative choices on the HEP, and issues of guidance

and traceability.

performers). The criteria used in the selection process

were:

2.3

SGTR isolation time

Ruptured SG level at isolation.

Experiment analysis methodology

As the qualitative HRA methods predictions typically

refer to the operational requirements and to the driver
factors that an idealised crew would face in the scenarios of the study, the main challenge of the analysis was
how to aggregate the performance of 14 different crews
into one average or exemplary operational description,
as well as into a general assessment of the driving
factors. In other terms, while there were observed 14
operational stories for each scenario, together with
as many constellations of factors affecting the crews,
the HRA predictions were (mainly) on one set of factors only and on one or few operational options (see
Figure 1).
2.4

Quantitative screening and crew selection

The starting point for the experimental analysis was

to look at the quantitative data, namely performance
figures generated from simulator logs (e.g. performance times), OPAS data (Skraaning 1998), expert
and observer performance ratings, and crew PSF ratings. This was a necessary step for assessing the
crews performance of the HFE under consideration (e.g. time used for identification and isolation,
SG1 level at isolation). This screening also provided
information which was later used for the writing of
summary operational stories (i.e. typical or average
crews progressions through the scenarios), by e.g. providing the execution times of important procedural
steps.
However, a thorough qualitative analysis was necessary to derive the required insights into drivers of
performance. The time schedule of the study and
the resource limitations conducted to the selection of
a subset of crews for in-depth study. The selection
was aimed at identifying a mixture of crews at both
ends of the performance spectrum (best and worst

Observations
(raw data)
- Audio / video

- Operational story

- Simulator logs

- Observed
influences
(PSFs)
- Observed
difficulties
(interpreted in
light of other
types of raw
data)

- On-line performance
ratings
- On-line comments
- Crew self-ratings
- Observer ratings

Figure 1.

2.5 In-depth analyses

The bases for the in-depth qualitative analysis were the
audio-video recordings, the recorded on-line expert
comments, the simulator logs, and the crew interviews. The core of the analysis process was the detailed
review of the video recordings of the scenario phases
corresponding to HFE 1. These reviews were structured so as to be useful and relevant for comparison to
the HRA analysis submissions.
The analysts viewed the video and transcribed key
communications and events. They also wrote comments about salient aspects of crew performance.
Immediately after the viewing, they completed a simplified version of the HERA system worksheets in
order to record the PSF details identified during the
video review in a common format. In completing
HERA, the analysts also drew on additional data
sources, such as the crew interviews, crew PSF
questionnaires, and observer comments. Finally, the
analysts summarised the observed episode in the form
of an operational story, highlighting performance characteristics, drivers, and key problems into so called
crews summaries.
The format of the crew summaries was based on
the format for the reporting of the HRA methods
assessment, as follows:
1. Short story of what happened in the selected part
of the scenario (written after reviewing DVD,
logs and interview) for identification and isolation
separately, which consisted of:
Extracts of crew communications, including
time stamps
A short summary of the observed episode in a
free form (not chronologically) including comments on crew performance.

14 Crew-level
performances

- Interviews
- OPAS

The criteria led to the selection of 9 crews, 3 base

cases (2 successes, 1 failure) and 6 complex
cases (3 successes, 3 failures). Other crews were
also analysed in-depth, and this information was used
to confirm and/or extend the tendencies identified
from the analysis of the best and the worst performers.

2 scenario-level
performances (over
all crews)
- 2 operational
expressions
- 2 sets of driving
factors with
ratings

2. Summary of the most influencing factors affecting

performance:
The PSFs were categorised as direct negative influences, negative influence present,
neutral influence, or positive influence. In
analyzing a performance, a PSF is a direct
negative when its negative influence on the

Data integration and analysis.

269

crews performance of the HFE was observed.

In some cases, factors were identified as negative, but only as present, meaning there was
no clear evidence that they significantly affected
the performance of the HFE.

Table 2.

Performance times in the two scenarios.

SceSG
SG
Crew nario Time1 level2 Crew Scenario Time1 level2
M
H
L
B
A
I
E
K
D
J
G
F
C
N

3. Summary of the observed difficulty or ease the

crew had in performing the HFE and of its assumed
causes.
The Summary of the most influencing factors
affecting performance was for each crew a combination of variable PSFs and constant PSF. The constant
PSFs were assessed by an expert panel and are constant for all crews, e.g. the quality of the interface.
The constant PSFs have then a fixed assessment that
is applied to all crew stories to keep overview of all
PSFs assessed when looking at each individual story.

Base
Base
Base
Base
Base
Base
Base
Base
Base
Base
Base
Base
Base
Base

10:23
11:59
13:06
13:19
13:33
13:37
14:22
15:09
16:34
17:38
18:38
18:45
18:53
21:29

20
10
6
21
17
31
40
39
55
44
39
73
57
75

L
B
I
M
G
N
H
K
D
A
C
F
J
E

Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex
Complex

19:59
21:10
21:36
22:12
23:39
24:37
24:43
26:39
27:14
28:01
28:57
30:16
32:08
45:27

78
1003
70
81
88
86
91
64
100
100
99
100
100
98

1 From

to tube rupture to isolation; 2 At isolation; 3 simulator

problem.

EXPERIMENT RESULTS

This paper concentrates on the empirical side of the

pilot project. The results of the HRA methods-to-data
comparisons are reported extensively on the pilot study
final report (Lois et al. 2008). By results here we mean
the simulator trials results.
The most striking result of the simulations was the
presence of a large degree of variability on the crew
performance of the first scenario phase identification
and isolation. Even a visual inspection of Table 2,
where the times the crews used to isolate the ruptured
steam generator are displayed, shows that the scenario
variant (base/complex) accounts for a large amount of
the performance time differences. There is no overlap
between the two groups, with the exception of the base
case run of crew N.
Also, within both cases there is large variability in
completion times. The range of performance times for
the base case is 10:23 min to 21:29 min, with the range
for the complex runs being 19:59 min to 45:27 min.
While it is clear that the experimental manipulation
had a significant effect, this alone does not reveal how
complexity translates into performance difficulties or
work styles which, in turn, produce the within-group
time differences.
One sense in which complexity translates in operational behaviour is the way the crews progressed
throughout the procedures. In the base case scenario,
all crews took the same way through the procedures
(E-0 SGTR identification stepE-3) and transferred to the isolation procedure by following an easily
recognizable transfer condition (radiation indication).
In the complex case, on the contrary, there were 8 different paths that were used among the 14 crews. In
Table 3 below all paths and the grounds for transfer
are displayed. In most instances the transfer to the isolation procedure was knowledge based, i.e. the crew

did not find an explicit transfer condition but judged

the plant state to require the implementation of the
ruptured SG isolation procedure (E-3).
A closer analysis at the two tables shows that there
does not exist an ideal way through the procedures
that will automatically lead the crew to success (fast
transfer and isolation). Fast and slow crews often share
the same procedure paths.
In both scenario variants, the analysis identified
crew factors as the primary determinants of withingroup performance time variability. Drivers categorized as team dynamics and work processes, appeared
to have an impact on performance time in the investigated emergency situation. In particular, the shift
supervisors situation understanding, his or her ability to focus on the main goal and to give directions to
the crew, as well as good procedures reading seemed
to be key factors to success (Broberg et al. 2008a).
3.1 Crew factors
In both scenario variants, we have attributed the variability of performance times to crew factors. Crew
factors can be distinguished in two categories (this
categorization was not followed in the pilot study):
1) crew characteristics, and 2) crew mechanisms (also
crew interaction).
Crew characteristics are systematic, stable factors
of the crews being investigated. Examples are: frequency and recency of training, the average experience
of the teams, the degree of stability in the crew set-ups
at the plant-unit level, the work practices at the plant
such as the organization of work, shifts, and national
characteristics. All these factors are needed in order to
describe the average crew to be modeled.

270

Table 3.

Procedure progressions and transfer grounds in complex scenario.

Crew

Procedure progression

Ground for transfer to E-3

C
G
L
N
A
M
E
F
I
H
B
D
J
K

E-0 step 21
E-0 step 21
E-0 step 21
E-0 step 21
E-0 step 21ES-1.1 foldout page
E-0 step 21ES-1.1 foldout page
E-0 step 21ES-1.1E-0 step 19
E-0 step 21ES-1.1E-0 step 19
E-0 step 21ES-1.1E-0 step 19
E-0 step 21ES-1.1FR-H5E-0 step 19
E-0 step 24
E-0 step 2425
E-0 (second loop) step 14E-2 step 7
E-0 step 19

Knowledge based (level)

Knowledge based (level)
Knowledge based (level) + ES-1.1 foldout
Knowledge based (level)
SG Level
SG Level
SG1 gamma level 1 and 2 (slow crew)
Knowledge based (level)
Knowledge based (level)
Knowledge based (level)
SG level
Knowledge based (level)
Knowledge based (level)
Gamma radiation. The crew manually trips
the reactor. This causes lower level in SG1
(less scrubbing effect) and high pressure in
SG1 early, which likely leads to the PORV
opening on high pressure and the consequent
activation of the total gamma sensor.

Decision guided or confirmed by procedure transfer point

Crew mechanisms are the elements and functions

that describe how a team works (and, in HRA terms,
possibly fail). Examples are leadership and roles, communication, openness, coordination, adaptability and
prioritization.
The crew mechanisms, in addition to explaining
how the crews perform, determine both crew-to-crew
and crew-to-scenario variability, and should ideally
be represented in the (crew) reliability model of an
HRA method. In contrast, crew characteristics would
determine systematic crew-to-crew variability (they
generally refer to the preparedness of the crew to the
task) and are, or should be, considered as context
factors in HRA.
To understand the point, we could see traditional
HRA as built around reliability models of individual cognition, where personal level variability is the
irreducible random variability across people (this is
assumed to be small compared to the variation caused
by the context on the cognitive functions). Were the
reliability model to include crew cognition (as well
as non-cognitive team interactions) the systematic
crew-to-crew variability would be expected to be larger
than the person-to-person variability, in part due to the
fact that the crew would not be modeled on micro-tasks
in laboratory settings, and in part because the status of
knowledge about team interaction does not account
for many general laws. The existence of a considerable crew-to-crew systematic variability is already
reflected by the fact that many HRA methods possess
crew characteristics PSFs. However, the lack of
guidance on the use of this type of PSFs testifies the

little integration of crew characteristics in most methods, integration that should ideally also account for the
relations between the two kinds of crew factors: crew
characteristics and crew interaction.

THE CHALLENGE: HRA PREDICTIONS VS.

SIMULATOR RESULTS

The main effort of the data analysis process has been in

finding a presentation format compatible with outputs
obtainable from HRA applications (section 2 above).
Overall, the study methodology have been judged as
adequate to the purposes of the study by all concerned
parties, HRA-teams, Steering group, and Experimentalist. However, on specific aspects, potential
for improvements have been individuated (Lois et al.
2008). This paper concentrates on areas of improvement relative to empirical data integration, formatting,
and use. The main issue here is that crew factors
and other dynamic factors, which were observed to
influence performance, could be better represented
and used in the method-to-data comparison, and, in
general, to inform HRA.
To begin with, let us remind that the experimental
results were summarized in three main parts:
1. Response times for identification/isolation and ruptured SG levels.
2. Aggregated operational stories for the two scenario
variants.

271

3. Aggregated driving PSFs (based on driving factors summaries for best and worse performing
crews and operational stories).
These presentation formats were chosen to allow the
comparison of HRA method predictions with observed
simulator performance. The response times were necessary in order to assess the performance of the HFEs
of the study. The aggregated stories were written in
order to summarize the performance of 14 different
crews (in the two scenarios), into single operational
expressions, which are the typical level of representation of HRA analyses (as a discretization of all possible
scenario variants). The same goes for the summary
of the driving PSFs, which could be considered as
the PSFs of the aggregated stories, as opposed to
the various configurations of context in the individual
scenario runs.
Concerns could be raised about the level of accuracy and completeness of the empirical information
reached trough this process of aggregation and formatting, concerns which regard all three types of results
output.
4.1

Failure vs. performance

Crew performance in the pilot study was operationalized as crew performance of the HFEs, in terms of
completion time and ruptured SG level at isolation
(the lowest the better). When the crews are evaluated
on the performance of the HFE a strong emphasis was
laid on time, with best crews being the fastest to
isolate, and the worst the slowest. This is a consequence of defining the HFE on a time criterion,
although the time criterion has a strong relation to several functional goals, including the PSA relevant one
of avoiding filling up the steam generators (Broberg
et al. 2008b).
On the fine-grained level of a simulator trial,
however, the speed of action can only be one of
several indicators of good performance, and one
that can never be isolated from the other indicators. For instance, a crew can act very fast when
a shift supervisor takes an extremely active role,
decides strategies without consultation and orders the
crew to perform steps from procedures, although the
latter is a reactor operator responsibility. This performance would not be optimal in terms of other
indicators: first, such behavior would not be considered in accordance to the shift supervisor function
and the training received. Further, it would reduce the
possibility for second checks, with one person centralizing all diagnosis and planning functions. Third,
it may disrupt successive team collaboration as the
reactor operator would feel disposed of his/her functions and could assume either a passive or antagonistic
position.

In such cases there is a disjunction between PSFs for

the HFE, those that influence the speed of actions and
hence quick success, and the factors which would normally be considered influences on good performance
(like good consultations) but which could in the same
cases slow down the performance of the HFE. In other
words, there is a mismatch between the categorization
required by the PRA HFEs representation and the one
implicit in the reliability models of HRA methods. The
pilot study testifies this mismatch: the PSF profile of
the second fastest crew has many similarities to the
profiles of the slow performing ones in the complex
scenario.
In general terms, it is difficult to see how HRA
can be directly informed by focusing on HEPs. In
the first place it was recognized that HEPs could not
be easily derived from observed errors in the trials, even given the large number of runs. HFEs were
therefore adapted to the constraints of a simulator exercise (Broberg et al. 2008b). Yet, they still bring the
theoretical framework of the PRA, superimposing it
on the crew performance. In other words, while the
HFEs focus on what particular goals are not achieved
(i.e. what prescribed human-initiated recovery functions are not performed), this cannot be a viable
perspective for understanding crew performance. In
fact, human and crew behavior in emergency scenarios are not directed by the discrete temporal succession
of PRA goals (which moreover are unknown to the
agents until their course of action is aligned to the
right event response), but by the dynamic goals
which derive from the interaction of contextual inputs
with their interpretation and the responses to them.
As a consequence, if the empirical basis of HRA is
information on the reliability models and their parameters, this has to be derived from studying what was
done and why, rather than focusing on what design
based actions were not performed (failures and errors).
The approach of focusing on the reliability models and
their parameters, rather than failures and errors, would
have consequences on both the criteria for crew selection (best and worst crew would not be defined
foremost from completion times), and on the PSFs
profiling (the best/worse crews PSFs profiles would
have more consistency).
4.2 The difficult treatment of variable PSFs
The derivation of the driving PSFs for the base and
complex scenarios are based on the driving factors
identified during the DVD reviews and summarized
in the crew summaries. The DVD review of individual
scenario runs (as well as their final aggregation and
evaluation), were influenced by the HERA terminology and other HRA-specific documents. This has been
challenging for the experimentalists, since such classifications and their definitions are not observational

272

tools, and since they incorporate context-performance

models not necessarily meant for fine-level analysis
of crew behaviour and interaction.
Further, a distinction was made between constant
PFS and variable PSFs. Constant PSF were considered the same for all crews, and, in part, could be
determined before the actual runs (e.g. indication of
conditions) based on:

unresolved issue in the pilot phase of the study. On the

other hand, it is also true that the scenario-based differences (the manipulation of the study), complexity and
procedure-situation fit, which were captured by the
constant PSFs, did have strong and observed effects
as testified by the performance differences between
base and complex conditions.

scenarios descriptions
the nature of the simulated plant responses, procedures, and interface
the plant specific work practices of the participating
crews.

4.3

Variable PSFs are those factors not supposed to be

the same for all crews, and which had to be evaluated for each crew after the scenario run. Many of
these PSFs have a dynamic nature in that they could
be evaluated only as the result of their interaction with
other context factors and of the interaction of these
with crew behaviour. For instance, stress levels can
vary across crews: a late identification could create
high stress levels during isolation for a crew with little experience in working together, but not in a more
experienced one. Most variable PSFs identified turned
out to relate to crew characteristics/mechanisms (e.g.
leadership style, accuracy of procedure reading) and
as such were classified under work practices, crew
dynamics and communication.
This classification prompts three orders of related
problems. The first is that the variable, crewinteraction PSFs do not fit most of the current HRA
methods, since those methods do not incorporate reliability models of crew interaction and functioning (at
best they model the crew as a second level of information processing), and, foremost, cannot analytically
treat crew-to-crew variability (they can at best accommodate it mathematically, in sensitivity analysis).
The second problem is that there is little guidance
for most HRA methods and tools (e.g. HERA) on how
to determine the presence and appropriate level of constant (systematic) crew-characteristics PSFs (e.g. work
practices, differences in experience and cohesion). In
HERA for instance both work practices and crew
dynamics have sub-items on supervisor behaviour,
so it is not to clear where to classify that dimension.
It is therefore hard to compare simulator results with
predictions on such factors.
The third problem is that the results of the empirical study regarding main drivers and PSFs might have
overweighed the importance of the constant/crewindependent PSFs at the expense of the variable and
crew-level PSFs, because they better fit the methodology of the study followed to produce the results. This
is reflected by the fact that the treatment of team
dynamics, work processes, and communication
in the identification of main drivers, was admittedly an

The interaction between PSF: observed vs.

modelled models

For the identification of the driving factors from the

crew summaries, the scenario events analyzed were
evaluated against a list of PSFs: for each item in the
given set the presence, direction and effect was determined, as well as a description of its manifestation. For
instance, in one crew summary communication was
rated as negative influence present and described in
the following way: While working on the isolation,
RO and ARO talk past each other, and the orders to the
field operator are initially not what they intended.
This format is consistent with the conventional
modeling of performance and PSFs for HRA in PRA,
where the assessment problem can be formulated as
follows:
Pf (Ti ) = f (wi1 v(F1 ), . . . , win v(Fn ), ei )

(1)

where Pf (Ti ) is the probability of failure of task

Ti in a particular event sequence, F1 , . . . , Fn are
PSFs that influence human performance of the given
task, v(F1 ), . . . , v(Fn ) are their quantified values,
wi1 , . . . , win are weighting coefficients representing
the influence of each PSFs in task Ti and ei is an error
term representing model and data uncertainty. f represents the function that yields the probability estimate,
which together with the parameters of the expression
above could be called the reliability model, (i.e. a
model of human performance). Different HRA methods incorporate different reliability models. Strictly
speaking, in the context of PRA, HRA empirical data
is information about the parameters of the reliability
models.
For several methods the reliability model or function f is one of independent factors, e.g. in SLIM:
Pf (Ti ) = wi1 v(F1 ) + wi2 v(F2 ), . . . , + win v(Fn ) + ei
(2)
This type of models treats the PSFs as orthogonal,
direct influences on the probability of task failure.
Even leaving aside the issue of failure, this kind
of modeling is generally not adequate for describing
task performance in simulator trials. In the first place,
the assignments and ratings cannot be done one-byone, as the PSFs are not independent (HRA itself

273

recognizes them to be correlated or overlapping). In

addition, the categorization in terms of presence
and directness does not exhaust the range of possible interactions. An alternative and more realistic
modeling would have to detail the entire assumed
influence set, by specifying all direct and indirect
links (and possibly reciprocal effects). Hypothetical,
observed structural models would be a much closer
approximation to the operational events as described
and explained in the narratives of the crew summaries. With such models, the process of aggregation
across crews would then be a process of generalizing
influence patterns across crews and events.
It must be added that, although the drivers were
presented and aggregated in tabular form, the methodto-data comparisons in the pilot study have been
performed by using all available information: by
combining information on PSFs to operational data,
descriptions, and evaluation of difficulties, as well as
by interacting with the experimentalists. Also, some
more recent methods as ATHEANA and MERMOS do
not foremost adopt the factorial model illustrated
above, but develop operational stories/deviation scenarios, which allowed for a more direct comparison to
observed operation.

CONCLUSIONS

The pilot study has shown that it is possible to inform

HRA by empirical evidence, provided that a sound
methodology for comparing HRA methods predictions and simulator results is followed. At the same
time, issues for improvement have been identified.
One central idea of this paper is that HRA
empirical content is information about the reliability
models (i.e. models of crew performance) incorporated by the methods. In this view, HEPs are theoretical
entities generated by the methods and their assumptions, which cannot be the object of observation,
and therefore cannot be empirically informed independently of the process, methods and assumptions
used to produce them. As a consequence, special
care has to be taken when defining the objects of
comparisons between HRA predictions and simulator
evidence, as illustrated in the discussion of failure vs.
performance.

Also, detailed analyses of crew performance show

that crew factors are important determinants of
performance variability. If HRA wants to exploit the
full potential of empirical information, we suggest
crew factors (and especially crew mechanisms) as the
centre of research for future HRA developments, to
the same extent that individual cognition has been so
far. Models of individual human cognition are unable
to explain the full spectrum of crew interactions, even
when a second (crew) level information processing is
modeled. Important aspects of crew behavior would
be missed by such modeling.
Finally, the well-known issue of PSFs interaction
cannot any longer be avoided by HRA methods and
empirical analysis alike, as complex patterns of causal
factor are the essence of observed operating crews
behavior.
REFERENCES
Braarud, P.., Broberg, H. & Massaiu, S. (2007). Performance shaping factors and masking experiment 2006:
Project status and planned analysis, Proceedings of the
Enlarged Halden Programme Group meeting, Storefjell,
Norway, 1116 March 2007.
Broberg, H., Hildebrandt, H., Massaiu, H., Braarud, P.. &
Johansson, B. (2008a). The International HRA Empirical
Study: Experimental Results and Insights into Performance Shaping Factors, proceedings of PSAM 2008.
Broberg, H., Braarud, P.. & Massaiu, S. (2008b). The International Human Reliability Analysis (HRA) Empirical
Study: Simulator Scenarios, Data Collection and Identification of Human Failure Events, proceedings of PSAM
2008.
Hallbert, B., Boring, R., Gertman, D., Dudenhoeffer, D.,
Whaley, A., Marble, J., Joe, J. & Lois, E. (2006).
Human Event Repository and Analysis (HERA) System,
Overview, NUREG/CR-6903, U.S. Nuclear Regulatory
Commission.
Lois, E., Dang, V.E., Forester, J., Broberg, H., Massaiu,
S., Hildebrandt, M., Braarud, P.., Parry, G., Julius, J.,
Boring, R., Mnnist, I. Bye, A. (2008). International
HRA Empirical StudyDescription of Overall Approach
and First Pilot Results from Comparing HRA Methods
to Simulator Data, HWR-844, OECD Halden Reactor
Project, Norway.
Skraaning, G. (1998). The operator performance assessment system (OPAS), HWR-538, OECD Halden Reactor
Project.

274

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Insights from the HRA international empirical study: How to link data
and HRA with MERMOS
H. Pesme, P. Le Bot & P. Meyer
EDF R&D, Clamart, France

ABSTRACT: MERMOS is the reference method used by Electricite de France (EDF) for Human Reliability
Assessment (HRA), to assess the operation of nuclear reactors during incidents and accidents in EDF Probabilistic and Safety Assessment (PSA) models. It is one of the second generation HRA methods that have participated
in the HRA international empirical study organised by the NRC and the Halden Reactor Project. This international study is not finished but has been already an opportunity to debate on relevant HRA issues during the
workshops in Halden and in Washington in 2007.
In this paper we will focus on the nature and meaning of predictive HRA, compared to the nature of data
(from observations on simulators or small real incidents). Our point of view on this subject will be illustrated
with an example of a MERMOS analysis implemented for the international study. Predictive HRA exists when
failure cannot be observed: it is a way to explore and reduce uncertainty, regarding highly reliable socio-technical
systems. MERMOS is a method which is supported by a model of accident that enables to describe the risk
and to link it to data. Indeed failure occurs when a way of operating (that usually leads to success) happens to
be inappropriate to a very specific context. Then data, in fact knowledge, is needed to describe two things: the
different operating ways (for example focusing for a while on the recovery of a system), and specific situations
(a serious failure of this system with a problem for identifying it). The HRA analyst has then to find which
combinations of operating habits and very specific situations could mismatch and lead to a serious failure of the
human action required to mitigate the consequences of the accident.
These links between operating habits, small incidents and big potential accidents that we will try to describe
in this paper should be understood for decision making in the field of safety, human factors and organisation;
indeed for example changing a working situation might be very risky regarding the whole panel of situations
modelled in a PSA. HRA should thus help the decision making process in the Human Factors field, besides
ergonomic and sociological approaches.

INTRODUCTION

An international empirical HRA (human reliability

analysis) study has been launched in 2007, with the
Halden Research Institute and several HRA experts
participating and debating [1]. The method designed
and used for Human Reliability Analysis in EDFs PSA
(probability and safety assessment) is MERMOS (in
French: mthode dvaluation de la ralisation des
missions oprateurs pour la sret); it is a second
generation HRA method [2] [3]. A team from EDF
has participated to this international study with the

MERMOS method: indeed it was an opportunity for

EDF to make MERMOS more widely understood and
to compare different HRA methods through the implementation of common examples. Some first insights
have been presented in a paper for PSAM 9: how to
describe potential failure and its influencing factors in
a given situation [4].
This paper focuses on the use of simulator data
for HRA: indeed this international study has chosen
a specific process to use simulator data, whereas the
MERMOS method follows a very different one: we
will try to explain these differences and describe how

275

data and HRA can be linked using the MERMOS

method, giving examples taken from the analyses
implemented for the international study.

2
2.1

HOW TO DEFINE HRA RESULTS

AND HRA DATA
Nature and meaning of HRA

HRA is helpful when failure cannot be observed: in

fact it is a way to explore and reduce uncertainty,
regarding high reliable socio-technical systems. Thus
HRA results intend to describe how failure can occur,
regarding PSAs criteria: they describe big failures,
not small ones that we can see on simulators from time
to time.
2.2

Nature of HRA data

What we call data in HRA are data from simulations

or from small real incidents. In the frame of the international study, we focus on the use of simulator data,
more precisely of data from dedicated simulations on
Hammlab (the Halden simulator).
From the international empirical study point of
view, the data collected on simulator can be compared to HRA results, and this is even one of the
main objective of the study to be able to show how the
HRA methods could predict what could be observed
on simulator.
On the other hand from our point of view, what
could be observed on simulator are only small errors
which never lead to failure in the sense of PSA, given
the high level of safety of a nuclear powerplant. So
such data cannot be compared to HRA results; however
there is of course a link between those data and HRA.

PROCESS OF COMPARISON OF HRA

RESULTS TO HALDEN DATA
IN THE INTERNATIONAL STUDY

The purpose of the international study is to compare

the results of the methods to data to see how well they
could predict the factors that drive performance. We
are very cautious with this purpose because we think
that this process is not justified: simulator data are
inputs for HRA (small failures, . . .) and cannot be
compared to HRA results (failure at a safety mission
level).
But, as quoted in the summary of the workshop,
some models use operational stories, like MERMOS: the advantage of such a method is that some
expressions of these operational stories can be easily compared to observed operational expressions,
and we think surely more easily than Performance

Shaping Factors (PSF) can. We would give some

examples further.
MERMOS is not focused on general micro individual performance (including success) prediction, but
on macro collective and systemic failure at a safety
mission level. However, systemic failures occur only
in very specific contexts, which include some operational expressions that we can observe on simulator,
as the comparison done in the international study
shows well. Moreover those MERMOS operational
stories are a way to express how different factors can
combine to lead to failure, in opposition to considerations on PSF which does not take into account their
combinations.
However it is not sufficient to compare the MERMOS operational expressions with the operational
expressions of failure that has been observed during
the dedicated simulations on Hammlab: what is interesting is on the one hand to observe the MERMOS
operational expressions in any of the teams test (successful or not), and then on the other hand to check that
the combination of the items in the MERMOS scenarios leads to failure (even if the whole MERMOS failure
scenario has not been observed).

LINKING DATA WITH MERMOS

ANALYSES: EXAMPLES TAKEN
FROM THE INTERNATIONAL STUDY

4.1

How to link data and MERMOS analyses

MERMOS is a method which is supported by a model

of accident that enables to describe the risk and to link
it to data. Failure occurs when a way of operating (that
usually leads to success) proves inappropriate to a very
specific context.
We consider as data any information which is useful to the MERMOS application and that involves
the entire operating system (interactions between
team, interface and procedures), not only individual
behaviour.

C
success

C context A
CICA 1

C context B

failure
CICA 1

Non conceivable scenarios

Figure 1. The MERMOS model of accident (from P. Le Bot,

RESS 2003).

276

There are three types of information:

1. Information interpreting a particular operating system context: Situation features describing the
structural and contextual characteristics of the situation. It is at this level in particular that individual
errors are found.
Example: The reactor operator makes a test
error and is oriented to the ECP1 procedure
This observation reflects the variations in human
behavior which can create uncommon, particular
contexts in which the resulting collective behavior
may fail.
2. Information interpreting a configuration or orientation of the entire system: the CICA (Important
Characteristics of Accident Management) allow the
operation of the operating system to be described
over time. They interpret a system configuration or
orientation. They are rarely observed directly during the test, but are the result of the interpretation
of these observations.
Example: the system follows procedures step
by step
3. Information which feeds expert judgments (for the
evaluation of situation features and CICA).
Example (context: loss of electric board LLE
and overabundant safety injection): document
complexity is delaying the diagnosis
An expert who needs to analyze a human factor
which occurs in this context will be able to take its
specificity into account and imagine failure scenarios where the complexity of documents and a
delayed diagnosis influence the failure.
This data, gathered in a particular context (for this
simulation, with the team being observed and for a
particular task), is reusable in order to imagine failure scenarios for a similar HF (Human Factor) task,
and also for a different HF task: many observations
can in fact be generalized. Some can directly be used
statistically: for example graphs showing the distribution of the observed times needed for carrying out
an action can be drawn up, however their creation and
use will need to be refined. However it is very difficult
to decide what to incorporate into the samples, as the
situations are never completely reproducible: in our
experience, the data taxonomy and other data categorizations are by nature reductive and dependent upon
models which can quickly become obsolete.
What is more, their extrapolation in conservative PSA contexts is delicate and must be carried
out with prudence. It is mainly through expert judgment that we could reuse this data in a quantitative
manner, taking into account these limits in modeling both observations and predictions. In addition,
observing simulations gives information which allows
us to increase each analysts knowledge of accident
management (operation of the operating system, how

procedures are carried out etc.) and contributes to

building up their expertise. The data gathered thus constitutes a knowledge database of qualitative dataas
opposed to a database of uniquely statistical data or
probabilities- which can be used more easily for expert
judgment.
Given these data, the HRA analyst has then to find
which combinations of operating habits and very specific situations could mismatch and lead to a serious
failure, as considered in PSAs [5]. Let us take an
example to illustrate it.
4.2

Examples of analyses from the international

study: HFE1A & HFE 1B

Nine HFE has been analysed for the international

study. Let us take the two first ones as examples:
HFE 1A:
The initiating event is: steam generator tube rupture
(SGTR, base scenario).
In this situation the mission is: to identify and
isolate the ruptured steam generator 20 minutes after
SGTR during full power, in order to prevent overfilling the ruptured steam generator. These 20 minutes
has been chosen by adding 5 minutes to the mean time
and does not correspond to a PSA criteria; however
they approximately correspond to the overfilling of
the steam generators.
HFE 1B:
The initiating event is: SGTR + major steam line
break quickly isolated (so that the secondary radiation
indications are shown normal).
In this situation the mission is: to identify and
isolate the ruptured steam generator 25 minutes after
SGTR during full power, in order to prevent overfilling the ruptured steam generator. These 25 minutes
has been chosen by adding 5 minutes to the mean time
and does not correspond to a PSA criteria; however
they approximately correspond to the overfilling of
the steam generators.
4.3 Examples of data observed on Hammlab
and found in the MERMOS analyses
In the MERMOS analyses of HFE 1A and HFE 1B, we
can recognize some elements, that we can call operational stories that could have been also observed on
the Halden simulator. This is not surprising because
MERMOS describes qualitatively the ways to fail,
specifying precisely the situations that could partially
be observed during simulations.
Here are some examples, gathering the comparison
work from the expert board of the international study
and some complements by EDF.

277

Operational expressions predicted in

MERMOSHFE 1A

Operational expressions
observed during the dedicated experiments in
Halden (HFE 1A)

not act with a sense

of urgency

Several crews decided

to take a meeting to
assess status and develop
a strategy before transferring to E-3 (based on
radiation indication)

trained transient

Good to very good

training and experience

easy to diagnose
transient

HMI and indication of

conditions: very good

Operational expressions
Operational expressions observed during the
predicted in
dedicated experiments in
MERMOSHFE1B
Halden (HFE 1B)
The system does not
Supported by the evidence:
perform the procedural
A crew was delayed
steps fast enough and
in E-0 due to their
does not reach the
early manual steam line
isolation step within
break identification and
the allotted time
isolation
Also: They use some time,
check if sampling is
open
Identification of the
SGTR by checking
steam generator levels
can cause problems or
time wastage.

Crew cannot explain SG1

level without radiation
and lack of level in
PRZ. Long discussions
and unstructured
meetings

The absence of
This is strongly supported
radioactivity does not
by the empirical evidence
facilitate diagnosis or
and is in fact a dominant
enable other hypotheses
operational issue
to be developed for the
(when combined with
event in progress.
the procedural guidances
reliance on this indication
and an apparent lack of
training on the alternative
cues).
ARO takes time to
check that the level in
SG#1 is rising uncon-

This is similar to
the second operational
expression above. Note

(continued)

Operational expressions
predicted in
MERMOSHFE1B
trollably. This is probable (assigned p = 0.3).
The ARO will not be
fast because this check
is not often trained
during SGTR scenarios
[which rely more
strongly on other cues]

Operational expressions
observed during the
dedicated experiments in
Halden (HFE 1B)
that the lack of training
on checking of alternative
cues for SGTR is
supported strongly by
the empirical data.

The operators follow

the instructions
cautiously

2 crews read carefully

the foldout page

Working through F0, the SS wishes to

quickly orientate the
team towards FR-H.5

SS abruptly to RO:
you can go to FR-H5
(one is not allowed to
enter this procedure
before step 22)

Waiting for feedback

This refers to E-3 Step 3.
from local actions leads
The evidence indeed
to delays (isolation not
shows that the crews
completed in time
need a fair amount of
window)
time to complete this
step due to the local
actions mixed in with
the control room actions.

4.4 Examples of data observed on Hammlab

and not found in the MERMOS analyses
One important operational story that has been
observed on the Halden simulator and that has caused
some delay does not take place in any MERMOS failure scenario. Our explanation is that we did not find
there could be a problem of transfer to procedure E3 for HFE 1B. Indeed theoretically, as mentioned in
the package, there are several paths to enter E-3 by
following the procedures. From our experience with
French teams, operators should have no difficulties to
enter E3 if they follow strictly the procedure, without
using their knowledge to transfer to E3 directly from
step 19. Because of that, we imagined that the main
reason to fail was a too strict following of the procedures that leads to spend too much time. In fact, it has
been observed exactly the opposite: Haldens teams
spent time to transfer directly from step 19 to E3, by
communicating and concerting.

278

Only a good knowledge of the habits of the Haldens

teams could have alert us and imagine that difference. We can explain that difference by the fact that
today, French procedures (state based procedures) are
designed to provide the operators with a solution in any
case. Then the operators trust the procedure and follow
them first. We do not agree with the fact that an important PSF for HFE1B is the inadequate procedure guidance: it is more likely the lack of trust of the operators
in the procedures, but this has to be deeper analyzed.
This difference between our predictions and observations illustrates well one of the main objections we
raised when the international study was launched: it is
groundless to aim at predictive analyses without knowing the operators way of working; the best way to
achieve this goal is to observe simulator tests. The purpose of HRA is not to predict what can be observed but
to predict from observations what cannot be observed.
4.5

Examples of data observed on EDF simulators

and found in the MERMOS analyses

Last, it is important to underline that some of the meaningful operational stories in MERMOS could not be
observed on the Halden dedicated simulations. Indeed
the MERMOS analyses take the most of all the simulations that we know (from our EDF simulators) and
that could be extrapolated for this study.
Here are some examples of those data:

CONCLUSION

In this paper we have focused on the nature of predictive HRA, compared to the nature of data (from
simulations or from small real incidents). Predictive
HRA exists when failure cannot be observed: it is
a way to explore and reduce uncertainty, regarding
high reliable socio-technical systems. MERMOS is a
method which is supported by a model of accident that
enables to describe the risk and to link it to data, as
we could see through examples from the international
study; the HRA analyst has then to find which combinations of operating habits and very specific situations
could mismatch and lead to a serious failure of the
human action required to mitigate the consequences
of the accident, as considered in PSAs.
These links between operating habits, small incidents and big potential accidents that we have tried
to describe in this paper should be understood for
decision making in the field of safety, human factors and organisation; indeed for example changing
a working situation might be very risky regarding the
whole panel of the situations modelled in a PSA. HRA
should thus help the decision making process in the
Human Factors field, besides ergonomic and sociological approaches, even if it still needs research to
push away its boundaries.
REFERENCES

Operational
expressions predicted in
MERMOSHFE 1A
and HFE 1B
local actions may
cause delay
run through the procedures
step by step
the SS does not incite
the operators to accelerate
the procedural path
the SS does not worry
about the effective performance of the action
delegation of . . . to the
other operator
the operator makes a
mistake in reading . . .
the shift supervisor
leads or agrees the strategy
of the operators
suspension of operation

Operational
expressions observed
on EDF simulators
Yes (and this is
simulated by the
trainers)
Yes
Yes
Yes
Yes
Yes
Yes
Yes (in order to
gain a better
understanding of
the situation)

279

[1] Lois E., Dang V., Forester J., Broberg H., Massaiu
S., Hildebrandt M., Braarud P. ., Parry G., Julius J.,
Boring R., Mnnist I., Bye A. International HRA
Empirical StudyDescription of Overall Approach and
First Pilot Results from Comparing HRA Methods
to Simulator Data, HWR-844, OECD Halden Reactor Project, Norway (Forthcoming also as a NUREG
report, US Nuclear Regulatory Commission, Washington, USA), 2008.
[2] Bieder C., Le Bot P., Desmares E., Cara F., Bonnet
J.L. MERMOS: EDFs New Advanced HRA Method,
PSAM 4, 1998.
[3] Meyer P., Le Bot P., Pesme H. MERMOS, an extended
second generation HRA method, IEEE/HPRCT 2007,
Monterey CA.
[4] Pesme H., Le Bot P., Meyer P., HRA insights from the
International empirical study in 2007: the EDF point of
view, 2008, PSAM 9, Hong Kong, China.
[5] Le Bot P., Pesme H., Meyer P. Collecting data for MERMOS using a simulator, 2008, PSAM 9, Hong Kong,
China.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Operators response time estimation for a critical task using the fuzzy
logic theory
M. Konstandinidou & Z. Nivolianitou
Institute of Nuclear Technology-Radiation Protection, NCSR Demokritos, Aghia Paraskevi, Athens, Greece

G. Simos
Hellenic Petroleum S.A., Aspropyrgos, Athens, Greece

C. Kiranoudis & N. Markatos

School of Chemical Engineering, National Technical University of Athens, Zografou campus, Athens, Greece

ABSTRACT: A model for the estimation of the probability of an erroneous human action in specific industrial
and working contexts based on the CREAM methodology has been created using the fuzzy logic theory. The
expansion of this model, presented in this paper, covers also operators response time data related with critical
tasks. A real life application, which is performed regularly in a petrochemical unit, has been chosen to test
the model. The reaction time of the operators in the execution of this specific task has been recorded through
an indication reported in the control room. For this specific task the influencing factors with a direct impact
on the operators performance have been evaluated and a tailored made version of the initial model has been
developed. The new model provides estimations that are in accordance with the real data coming from the
petrochemical unit. The model can be further expanded and used in different operational tasks and working
contexts.

INTRODUCTION

In Human Reliability Analysis the notion of human

error does not correspond only to the likelihood that
an operator will not perform correctly the task that he
has been assigned to do but also (among other things)
to the likelihood that he will not perform the assigned
task within the required time. Most of the critical
tasks include the concept of time in their characterization as critical and most of the error taxonomies
developed specifically for human reliability analysis include errors like too early/too late, action
performed at wrong time, delayed action, operation incorrectly timed, too slow to achieve goal
and inappropriate timing (Embrey 1992, Hollnagel
1998, Isaac et al. 2002, Kontogiannis 1997, Swain &
Guttmann 1983). What is thus important for Human
Reliability Analysis is the identification and quantification of human error and at the same time the
estimation for the response time of the operator in
the performance of a critical task. In modeling human
performance for Probabilistic Risk Assessment it is
necessary to consider those factors that have the
biggest effect on performance. The same is also valid
for factors that influence operators response time.

Many factors influence human performance in complex man-machine systems like the industrial context
but not all of them influence the response time of
operators, at least not with the same importance.
Many studies have been performed to estimate operators response time mainly for Nuclear Power Plants
(Boot & Kozinsky 1981, Weston & Whitehead 1987).
Those were dynamic simulators studies with the objective to record response time of operators under abnormal events (Zhang et al. 2007) and also to provide
estimates for human error probabilities (Swain &
Guttmann 1983). A fuzzy regression model has also
been developed (Kim & Bishu 1996) in order to assess
operators response time in NPP. The work presented
in this paper is an application for the assessment of
operators response times in the chemical process
industry.
A model for the estimation of the probability
of an erroneous human action in specific industrial
and working contexts has been created using the
fuzzy logic theory (Konstandinidou et al. 2006b).
The fuzzy model developed has been based on the
CREAM methodology for human reliability analysis and includes nine input parameters similar to the
common performance conditions of the method and

281

one output variable: the action failure probability of

human operator. Validation of the model and sensitivity analysis has already been performed (Konstandinidou et al. 2006b, 2006a).
The expansion of this model, presented in here,
covers also operators response time data related with
critical tasks. The model disposes now of a second output variable that calculates the estimated response time
of the operator performing a specific task in a specific
industrial context. For the estimation of the response
time the model takes into account factors (common
performance conditions) that influence the reaction of
the operator during this specific task.
Section 2 of this paper gives a quick overview of the
model developed for the estimation of human error
probabilities while section 3 presents the expansion
of the model to cover also operators response time.
Section 4 describes the real industrial task which will
be used for the application of the model. Section 5
presents the application of the model in this task and
a shorter version of the model which is more tailored
made to include only those input parameters that affect
operators response time. Section 6 makes a comparison of the results between the two models while
section 7 presents the conclusions of this paper.
2

FUZZY MODEL FOR HUMAN RELIABILITY

ANALYSIS

A fuzzy logic system for the estimation of the probability of a human erroneous action given specific
industrial and working contexts has been previously
developed (Konstandinidou et al. 2006b). The fuzzy
logic modeling architecture has been selected on
account of its ability to address qualitative information and subjectivity in a way that it resembles the
human brain i.e. the way humans make inferences and
take decisions. Although fuzzy logic has been characterized as controversial by mathematician scientists,
it is acknowledged that it offers a unique feature: the
concept of linguistic variable. The concept of a linguistic variable, in association with the calculi of fuzzy
ifthen rules, has a position of centrality in almost all
applications of fuzzy logic (Zadeh, 1996).
According to L. Zadeh (2008) who first introduced
fuzzy logic theory, today fuzzy logic is far less controversial than it was in the past. There are over 50,000
papers and 5,000 patents that represent a significant
metric for its impact. Fuzzy logic has emerged as a very
useful tool for modeling processes which are rather
complex for conventional methods or when the available information is qualitative, inexact or uncertain
(Vakalis et al. 2004).
The Mamdani type of fuzzy modeling has been
selected and the development of the system has been
completed in four steps.

i.
ii.
iii.
iv.

Selection of the input parameters

Development of the fuzzy sets
Development of the fuzzy rules
Defuzzification

The fuzzification process was based on the

CREAM methodology and the fuzzy model included
nine input variables similar to the common performance conditions of the same methodology namely:
Adequacy of organization, Working conditions,
Adequacy of man-machine interface and operational
support, Availability of procedures, Number of simultaneous goals, Available time, Time of day (circadian
rhythm), Adequacy of training and Crew collaboration quality. The fuzzy logic system has as output
parameter the Action Failure Probability of Human
Operator.
For the development of the fuzzy sets and the fuzzy
rules the phrasing and the logic of CREAM has been
used. According to CREAM a screening of the input
parameters can give an estimation of the mode in which
an operator is acting (based on his Contextual Control
Mode). The rules are constructed in simple linguistic
terms and can be understood at a common sense level.
At the same time these rules result in specific and
reproducible results (same inputs give same output).
The defuzzification process is performed through
the centroid defuzzification method (Pedrycz 1993),
where an analytical calculation of the gravity center
produces the final result. The output fuzzy sets cover
the interval from 0.5 105 to 1 100 (corresponding
to the values of probability of action failure defined in
CREAM).
The system has operated with different scenarios
and the results were very satisfactory and in the range
of the expectations (Konstandinidou et al. 2006b).
These results can be used directly in fault trees and
event trees for the quantification of specific undesired
events, which include in their sequences failures of
human factors.
Another use of the model, which can be compared
to sensitivity analysis, deals with the input parameters
of the model as influencing factors in Human Reliability. Factors which influence human reliability play
a very important aspect in the quantification of human
error. The context in which the human action will take
place is defined by these factors. These are the factors
that usually have the name of Performance Shaping
Factors, or Common Performance Conditions or
Performance Influencing Factors. Obviously those
factors, as their name indicates, influence the action
failure probability of the human operators, by increasing it when they have a negative effect on it or by
decreasing it when they support the action and the
operator. What is common knowledge (but not common practice) is that the better the quality of these
factors the more reliable the operator behavior.

282

The fuzzy model has been used in order to detect

the critical transitions in the optimization of human
reliability, in a form of sensitivity analysis, for corresponding changes in each of the nine input parameters,
tested in specific contexts. Changes in the working
context have been evaluated through their relevant
reduction to the action failure probability of the human
operator (Konstandinidou et al. 2006a). As concluded
from the application of the model, the input parameters
that induce the highest variations in the action failure probability are the adequacy of training and the
crew collaboration quality. For the parameter time
of the day the model shows that operators are more
prone to errors during the night hours (20:004:00) and
also during shift ending and turnovers. The results of
this application are in the form of percentages. These
percentages represent the variations induced on the
output result, namely the action failure probability,
from the variations in the input parameters. The meaning of these percentages is that with an improvement in
the training of operators and in the ability to collaborate with each other, the level of human reliability will
increase significantly. The values of these percentages
are not so important; the most important is that with the
use of the fuzzy model the critical intervals are defined
within which the significant variations are located. The
determination of the critical transitions depicts in this
way the points in which the analyst should focus and
the areas of improvement which are meaningful and
essential. With the ability of having numerical values of human error probabilities, the analyst is able to
prioritize the possible improvements in elements that
affect operators reliability. These results can be used
furthermore in costbenefit analysis with the objective to compare the parameters adjustments cost to the
impact they induce on the performance and reliability
of the human operator.

The new model disposes of a new output parameter namely operators response time. The output
parameter provides the needed estimations for operators response time. In order to maintain the connection with the initial model the same names and
notions in the output parameters were used. The output fuzzy sets correspond to the four control modes
of the COCOM model that is the cognitive model
used in CREAM (Hollnagel 1998). Those modes are:
the strategic control mode; the tactical control
mode; the opportunistic control mode; and the
scrambled control mode.
For the application of the ORT fuzzy model the
four control modes were used to define the time intervals within which the operator would act to complete
a critical task. Hence quick and precise actions that
are completed within very short time are compatible
with the strategic control mode; tactical control mode includes actions within short time intervals
slightly more broad than the previous one; opportunistic control mode corresponds to slower reactions
that will take longer time while scrambled control mode includes more sparse and time consuming
reactions.
The relevant time intervals as defined for the four
control modes in the ORT fuzzy model are presented in table 1. A graphical representation of the four
fuzzy sets is given in figure 1. The range of the four
fuzzy sets is equivalent to the range used in the probability intervals of action failure probabilities in the
initial model (Konstandinidou et al. 2006b) expressed
in logarithmic values.
Table 1.

Operators response time

(minutes)

FUZZY MODEL WITH RESPONSE

TIME ESTIMATIONS

In order to produce estimates for response time of operators in industrial context the fuzzy model for Human
Reliability Analysis has been used. With this model
as a basis the fuzzy model for Operators Response
TimeORT estimation has been built.
The functional characteristics of the initial model
remained as they were defined. That means that the
same nine input parameters with the same defined
fuzzy sets have been used. The phrasing and the linguistic variables have remained the same too. This was
very helpful in order to have a correspondence between
the two models. For more details concerning the functional characteristics of the initial model please refer
to (Konstandinidou et al. 2006b).

Control modes and response time intervals.

Control mode

Min

Strategic
Tactical
Opportunistic
Scrambled

0
0.01
0.1
1

Max
<t<
<t<
<t<
<t<

0.1
1
5
10

Operator Response Time

1
Strategic
Tactical
Opportunistic
Scrambled

0
0

10

Time interval

Figure 1. Fuzzy sets representation for the Operator

response time output variable.

283

A crucial step in the development of the model

is the development of the fuzzy rules. A cluster of
fuzzy rules to include all the possible combinations of
the input parameters fuzzy sets has been developed in
(Konstandinidou et al. 2006b). 46656 rules have been
defined, taking into consideration the multiple fuzzy
sets of each input parameter and using the logical AND
operation as the building mode.
The fuzzy rules for the extension of the model
retained the if part of the initial model and the
when part was changed accordingly to include the
time notion.
An example (i.e. the first rule) is the following:
If the adequacy of organization is deficient AND
the working conditions are incompatible AND the
availability of procedures and plans is inappropriate
AND the adequacy of man-machine interface and
operational support is inappropriate AND the number of simultaneous goals is more than actual capacity
AND the available time is continuously inadequate
AND the time of the day is night AND the adequacy of
training and experience is inadequate AND the crew
collaboration quality is deficient THEN the operator would act in a SCRAMBLED way. Acting in a
SCRAMBLED way means that the response time for
the operator is between 1 and 10 minutes.
In this way all the possible combinations of the input
fuzzy sets correspond to one (and only one) output
fuzzy set and to the relevant control mode with the
associated time interval.
In order to have a crisp number as output variable
(and not an output set) the centroid defuzzification
method (Pedrycz 1993) has been used as in the initial
model. In this way the model comes up with specific estimates for operators response time expressed
in minutes.

SPECIFIC APPLICATION FROM

THE PROCESS INDUSTRY

In order to test the model a real life application has been

chosen. A specific task, which is the opening/closure
of a manual valve in order to maintain a desired pressure drop, is performed regularly in a petrochemical
unit. This task may be performed at least twice a day
during normal operation in order to unclog the drain
channel. The same task is performed during maintenance operation in order to shut down or start up the
unit. In case of an abnormality that leads to the trip
of the unit or in case of equipment malfunction the
operators are called to act immediately and perform
the same task in order to maintain the desired pressure drop so that the unit is not jeopardized. This is
equivalent to emergency response situations.

The required time frame for the specific task is very

tight. Operators must complete their actions within
12 minutes. Otherwise pressure may rise or may drop
beyond the safety limits and disturb the operation of
the whole unit or even worse (in case of extreme variations) result in equipment failure. Pressure rises and/or
drops in few seconds in the specific node so operators response is crucial and should be prompted. For
the completion of the task one operator is needed.
The reaction time of the operators in the execution
of this task has been recorded through the pressure
drop indication reported in the control room. Data
concerning the specific infield task of the petrochemical unit has been gathered during a whole year
period. From those data it was noticed that normal
reaction time is within 1015 seconds (when performing the normaldrain operation), reaction time during
maintenance was around 1 minute, while reaction time
in emergency situations was between 1 to 10 minutes
depending on the case.
After discussion with the key personnel of the unit
on the specific events that took place during the one
year period the conclusions were that the elements that
differentiate the reaction time of the operators is the
level of experience each operator has and the number of tasks he is assigned to do in the same time.
This number varies between normal operation, maintenance and emergency response situations. What has
also been observed through the collected data is that
the time of the day plays also an important role in
some situations: operators response time is different
between day and night shifts.
Hence for this specific task the influencing factors
that have a direct impact on the operators performance
are: the circadian rhythm of the operator, expressed in
terms of the hour of the day that he/she is requested
to perform the task; the experience and the training
he/she obtains, expressed in years of presence in the
specific unit (and the petrochemical plant); the number
of simultaneous goals, expressed in terms of parallel
tasks to be performed during normal operation, maintenance (task performed in order to shut down or to
start up the unit) or emergency situations (equipment
malfunction, trip of the unit).
The conclusions of our observations were the basis
for the development of a shorter version of the fuzzy
model, a model that would include only the influencing factors of this application with the relevant fuzzy
sets. This is meaningful since all the nine parameters that are included in the full version of the ORT
model do not affect response time in this particular
application and the computational cost of the model is
significantly decreased with the use of only three input
parameters. Additionally by building a newtailored
made model for the specific application new fuzzy
sets for the output parameter operators response

284

The observation data and the expertise of the key

personnel were the knowledge base for the development of the fuzzy rules. The following observations
determined the definition of the fuzzy rules:

time can be used and adjusted according to real

data.
5

FUZZY MODEL FOR THE SPECIFIC

APPLICATIONSHORT ORT MODEL

For the development of this tailored made Operators Response TimeORT short model the Mamdani
type of fuzzy modeling has been selected and the
development of the system has been completed in four
steps.
i. Selection of the input parameters
Three input parameters have been chosen
according to the conclusions stated in the previous
section. These input parameters are:
a. The number of simultaneous goals
b. The adequacy of training and experience
c. The time of the day
As unique output parameter was defined the
Operators Response Time.
ii. Development of the fuzzy sets
In the second step, the number and characteristics
of fuzzy sets for the input variables and for the output
parameter were defined. The definition of the fuzzy
sets was made according to the observations from the
real data and the comments of the key personnel as
stated previously.
Number of simultaneous goals: for the first input
parameter three fuzzy sets were defined namely Normal operation, Maintenance and Emergency
Situation.
Adequacy of training and experience: for the
second input parameter two fuzzy sets were defined
namely Poor Level of Training and Experience and
Good Level of Training and Experience.
Time of the day: for the last input parameter two
fuzzy sets were distinguished corresponding to Day
and Night.
Operators response time: The output parameter
had to cover the time interval between 0 and 10 minutes. Five fuzzy sets were defined to better depict small
differences in reaction time and the equivalent time
range was expressed in seconds. The fuzzy sets with
the time intervals each of them covers are presented
in table 2. More precisely operators response time is
Very good from 0 to 20 seconds, Good from 10
to 110 seconds, Normal from 60 to 180 seconds,
Critical from 120 to 360 seconds and Very critical
from 270 to 1170 seconds. A graphical representation
of the five fuzzy sets is given in figure 2 in order to
visualize the range of each time set.
iii. Development of the fuzzy rules

a. Time of the day (day/night) does not affect operators response time during normal operations
b. Time of the day (day/night) does not affect operators response time for operators with good level of
training and experience
According to the observed data and by taking into
account the above mentioned statements 8 fuzzy rules
were defined for the short ORT fuzzy model:
Rule 1: If number of goals is equivalent to normal
operation and adequacy of training and experience
is good then operators response time is very good.
Rule 2: If number of goals is equivalent to normal
operation and adequacy of training and experience
is poor then operators response time is good.
Rule 3: If number of goals is equivalent to maintenance and adequacy of training and experience is
good then operators response time is good.
Rule 4: If number of goals is equivalent to maintenance and adequacy of training and experience is
poor and time is during day shift then operators
response time is normal.
Rule 5: If number of goals is equivalent to maintenance and adequacy of training and experience is
poor and time is during night shift then operators
response time is critical.
Rule 6: If number of goals is equivalent to emergency
and adequacy of training and experience is good
then operators response time is normal.
Rule 7: If number of goals is equivalent to emergency
and adequacy of training and experience is poor and
time is during day shift then operators response
time is critical.
Rule 8: If number of goals is equivalent to emergency
and adequacy of training and experience is poor and
time is during night shift then operators response
time is very critical.
iv. Defuzzification
Since the final output of the fuzzy system modeling
should be a crisp number for the operators response
time, the fuzzy output needs to be defuzzified. This
is done through the centroid defuzzification method
(Pedrycz 1993) as in the previously developed fuzzy
models.
The fuzzy logic system has been built in accordance
with the real data coming from the petrochemical unit.
The testing of the model and its comparison with
the full version will be presented in the section that
follows.

285

Operators Response Time

1

Very Good
Good
Normal
Critical
Very Critical

0
0

100

200

300

400

500

600

Time Interval

Figure 2. Fuzzy sets representation for the Operator

Response Time output variable of the short model.
Table 2.

Output fuzzy sets for operators response time.

Fuzzy set

Time interval (in seconds)

Very good
Good
Normal
Critical
Very Critical

0
10
60
120
270

6
6.1

<t<
<t<
<t<
<t<
<t<

20
110
180
360
1170

RESULTS FROM THE APPLICATION

OF THE MODELS
ORT fuzzy model general application

The results from the general application of the full

version of ORT fuzzy model are presented in table 3.
Details for better understanding of these results are as
following.
First row (first run) represents the worst case scenario, which means a context where all input parameters are judged as inadequate and are given the
minimum possible value they can be assigned to. In
this scenariocase the ORT estimated a value of
6 minutes (480 seconds) for operators response time.
The second row (second run) is still a worst case scenario but with slightly improved parameters value. In
this case the ORT produced a slightly improved
response time of 476 seconds.
Situation 3 (third rowrun) is a best case scenario
with all parameters assumed to be in the most efficient
level. In this case the ORT estimated a time of 8 seconds for operators response time while in the fourth
situation which represents the best case scenario with
top values for all parameters the estimated time was 7
seconds.
With the above runs the sensitivity of the model was
tested. Indeed the model depicts differences in its input
and the calculated result is alternated respectively.
Fifth row represents a mediumnormal case. In
this situation all parameters have a medium level of
quality that corresponds to usual working contexts. In
this case the operators response time was estimated
at 59 second (1 minute).

The rest of the runs represent situations where all

input parameters except of Number of simultaneous
goals, Time of the day and Adequacy of training
and experience have medium level values (equal to
a value of 50). Changes in values were made only for
the three input parameters that affect the specific task
which was examined for the application of the model.
For these three parameters the values that were used
with their correspondence with operational situations
(expressed in linguistic variables) are the following:
Number of simultaneous goals: A value of 15
was chosen to represent emergency situations, while
a value of 50 was attributed to maintenance and 90 to
normal operation.
Time of day: Value 12 was assigned to day time
while value 0 to night time.
Adequacy of training and experience: A value of
0 depicts the inexperienced operator, a value of 50 is
the operator with adequate training but with limited
experience while a value of 100 corresponds to the
very experienced operator.
What can be seen from the results is that the model
only depicts differences in time response during night
of an inexperienced operator for an emergency situation (estimated reaction time 276 seconds). For the
rest of the modeled situations (very and limited experienced operator, normal operation, maintenance and
emergency situation, day and night time) the model
estimates the same reaction time for all operators (59
second approx. 1 minute).
The reaction time estimated is in accordance with
the real time of operators response. The inflexibility
of the model to provide different estimates according
to different situations is explained by the fact that for
the six out of the nine inputs medium values were chosen. This affects the inference of the model by leaving
a smaller number of rules to interfere for the calculation of the final result. From the total number of
46656 rules only 18 rules are activated with the use of
medium values.
Additionally from the three parameters whose values are alternated two (number of simultaneous goals
and time of the day) comprise fuzzy sets that have a
neutral influence on the final output (Konstandinidou
et al. 2006b). That means that even by improving
the value of those inputs the estimated response time
would remain the same since the improvement has no
effect on the output result.
6.2 ORT fuzzy model specific application
In order to overpass the observed shortages a new
group of runs has been selected. This group comprises runs that correspond to the specific context
under which the selected critical task is performed.
For the specific application of the model the following
input values were used.

286

Table 3.

Results from the application of the ORT fuzzy model.

Number of
Adequacy Working
MMI and simultaTraining Crew
Operators
of organiz- condi- Procedures operational neous
Available Time and expe- collabora- response
ation
tions
and plans support
goals
time
of day rience
tion
time (sec)
0
10
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90

0
10
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20

0
10
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90

0
10
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50

10
15
90
90
50
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90
15
50
90

10
20
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20

Adequacy of organization: A value of 90 was used

for this input parameter since the overall organization
of the specific unit in safety issues is judged to be
excellent.
Working conditions: A value of 20 was given
in this parameter since the task is performed in a
petrochemical unit with unadvantageous working conditions (noise, poor ergonomy, bad lighting, smoke
and bad odors).
Adequacy of procedures and plans: A value of 90
was used for this input parameter because the complete

0
2
12
13
12
0
0
0
12
12
12
0
0
0
12
12
12
0
0
0
12
12
12
0
0
0
12
12
12
0
0
0
12
12
12
0
0
0
12
12
12

10
10
90
100
50
0
0
0
0
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
0
0
0
0
100
100
100
100
100
100
50
50
50
50
50
50

0
10
90
100
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50

480
476
8
7
59
276
59
59
59
59
59
59
59
59
59
59
59
59
59
59
59
59
59
294
294
294
276
59
59
59
59
59
59
59
59
276
59
59
59
59
59

sequence of actions for the specific task (as well as for

the rest of the tasks) is provided in well written and
updated procedures.
Adequacy of MMI and operational support: A
value of 50 was assigned to this input as the whole task
is performed on a specific valve so MMI is always the
same and does not interfere in operators action.
Number of simultaneous goals: As previously a
value of 15 was chosen to represent emergency situations, while a value of 50 was attributed to maintenance
and 90 to normal operation.

287

Available time: A value of 20 was assigned for this

input as the available time for performing the specific
task is always very little (12 minutes).
Time of day: A value of 12 was used for day time
while a value of 0 for night time.
Adequacy of training and experience: A value of
0 depicts the inexperienced operator, a value of 50 the
operator with limited experience while a value of 100
corresponds to the very experienced operator.
Crew collaboration quality: A medium value of 50
was assigned to this parameter since no crew is needed
to perform this task (task performed by one operator
only).
The results from this application of the model are
presented in second part of table 3.
The estimated times with input expressed in linguistic variables for the three parameters which influence
the operators response time in the specific application
are presented in table 4 in order to facilitate the reader
in the comparison of results later on.
As expected, estimated times are greater for inexperienced operators during night (294 sec) and day
(276 sec) and between emergency (276 sec) and normal operations or maintenance (59 sec). However for
a well trained operator the estimated times are always
the same (59 seconds 1 minute).
This is due to the same reason as explained in the
previous section: number of simultaneous goals and
time of the day have fuzzy sets with neutral effect on
the output result. Thus, changes in the specific input
parameters do not affect output.
In order to overcome this problem the fuzzy model
has to be modified. This was not feasible since a
change in the effect of the fuzzy sets would disturb the
initial structure of the model and make it incompatible
with the estimations of the action failure probabilities which are already validated and in accordance

Table 4. Results from the application of the two versions

of ORT fuzzy model.
Number of
simultaneous goals

ORT ORT
Time model short
Training of day (sec)
(sec)

Normal operation
Maintenance
Emergency situation
Normal operation
Maintenance
Emergency situation
Normal operation
Maintenance
Emergency situation
Normal operation
Maintenance
Emergency situation

Good
Good
Good
Good
Good
Good
Poor
Poor
Poor
Poor
Poor
Poor

Day
Day
Day
Night
Night
Night
Day
Day
Day
Night
Night
Night

59
59
59
59
59
59
59
59
276
294
294
294

13
60
120
13
60
120
60
120
240
60
240
570

with CREAM (Konstandinidou et al. 2006b). Instead

the team decided to produce a shorter version of the
model that would be more prone to adjustments and
correspond better to the specific task.
The development of the short version was also
advantageous from the computational point of view
as explained previously in section 4. The results from
the application of the ORT short fuzzy model are
presented in the following paragraph.
6.3 ORT-short fuzzy model
The results from the application of the short version
of ORT fuzzy model are presented in table 4. With
the use of the ORT short model the estimates for
operators response time differ when the input parameters change in all possible combinations. In this way
improvements in operators level of experience are
depicted as well as differences in time shifts.
According to the estimates of the ORT short
model a well experienced operator will react in 13 seconds during normal operation in day and night shift, in
60 seconds during maintenance in day and night shift
and in 120 seconds in emergency situations during day
and night shift. This is in accordance with observation
a) that the time of the day does not affect the response
time of an operator with good level of training and
experience. Subsequently an inexperienced operator
will react in 60 seconds during normal operation in day
and night shift, in 120 seconds during maintenance in
day time and 240 in night time shifts, and in 240 seconds in emergency situations during day shift and 570
in night shift. This is in accordance with observation
b) that the time of the day does not affect the response
time in normal operations.
The fuzzy logic system estimations are in accordance with the real data coming from the petrochemical unit. Indeed, observation data showed very slow
and very critical response of inexperienced operators
during night shifts and in emergency situations. In
fact the registered response time in one such case has
reached the period of 10 minutes (600 seconds).
The model can be further expanded and used in different tasks and contexts e.g. maintenance tasks, other
in-field actions or control room operations in the running of a petrochemical unit or more generally of a
chemical plant. The only constraints for the application of the model are the knowledge of the influencing
factors for the specific tasks and the availability of real
data.
7

CONCLUSIONS

The criticality of certain tasks in industrial contexts

deals not only with the right performance of the task
but also with the correct timing of the performance.

288

A critical task performed too late may be in certain

cases equal to an erroneously performed task. Thus
for the purposes of Human Reliability Analysis not
only the Human Error Probability of a task would be
necessary but also the response time of the operator
under specific conditions.
A model for the estimation of the probability of
an erroneous human action in specific industrial and
working contexts has been previously created based on
CREAM methodology and using the fuzzy logic theory. The model has been expanded in order to produce
estimates for operators response time. This version
was called the ORT model.
In order to test the model real data have been gathered form a petrochemical unit concerning a specific
in-field task: the opening/closure of a manual valve
in order to maintain a desired pressure drop. The
recorded reaction time for the specific task is about
1 minute in certain conditions. In extreme cases the
recorded time reached the period of 10 minutes. For
this specific task the influencing factors that have a
direct impact on the operators performance were: the
time of the day when the task is performed; the level
of training and experience of the operator and the
number of simultaneous goals the operator is asked
to perform differentiating between tasks during normal operation, during maintenance or in emergency
situations (equipment malfunction, trip of the unit).
Three different applications of the model were made
for the specific task. In the first application of the
model (original version with second output parameter
response time) with medium values for all parameters
except from the three influencing ones the estimated
response time was within the range of expectations
(1 minute). Although the model was sensible in input
variations, when medium values were used for most of
the input parameters the model was rather inflexible.
This was due to the fact that medium values deactivated
most of the fuzzy rules of the original model (99% of
the total fuzzy rules were not used).
When the same model was used with values representing the real life situation the output resulted in
more sensitive values. Estimates were in the range
of expectations and variations in input induced variations in the output results. However variations in
number of simultaneous goals and time of the day
were still not depicted (i.e. operators response time
does not alter for day and night changes neither for
different operations) since those two input parameters
have fuzzy sets with neutral influence on the output
result in the original structure of the model.
Therefore and since only three parameters were
acting as influencing factors in the specific task, an
ORT short model was developed to include only
these input parameters. In this way the application of
the model is more tailored made to the specific task
and a lot of saving in computational cost is achieved.

Indeed the ORT short fuzzy model came up with

estimates in operators response time that are in accordance with the observed data and additionally more
sensible to input variations.
Differences between day and night shifts as well as
task performed during normal operation, maintenance
and emergency situation from experienced and inexperienced personnel are well depicted with relevant
differences in operators response times. In fact in the
extreme situation of an emergency during night shift
where an inexperienced operator is called to act the
estimated response time from the model is 570 seconds which is in accordance with the observed data of
10 minutes (600 seconds).
Following the steps described in the present paper
ORT short tailored made models based on fuzzy
logic architecture can be developed for different tasks
and contexts e.g. maintenance tasks, other in-field
actions or control room operations in the running of a
chemical plant.

REFERENCES
Bott, T.F. & Kozinsky, E. 1981. Criteria for safety-related
nuclear power plant operator actions. NUREG/CR-1908,
Oak Ridge National Lab, US Nuclear Regulatory Commission.
Embrey, D.E. 1992. Quantitative and qualitative prediction of human error in safety assessments, Major hazards
Onshore and Offshore, Rugby IChemE.
Hollnagel, E. 1998. Cognitive reliability and error analysis
method (CREAM). Elsevier Science Ltd.
Isaac, A., Shorrock, S.T. & Kirwan, B. 2002. Human error
in European air traffic management: the HERA project.
Reliability Engineering & System Safety 75 (2): 257272.
Kim, B. & Bishu, R.R., 1996. On assessing operator response
time in human reliability analysis (HRA) using a possibilistic fuzzy regression model. Reliability Engineering &
System Safety 52: 2734.
Kontogiannis, T. 1997. A framework for the analysis of cognitive reliability in complex systems: a recovery centred
approach. Reliability Engineering & System Safety 58:
233248.
Konstandinidou, M., Kiranoudis, C., Markatos, N. &
Nivolianitou, Z. 2006a. Evaluation of influencing factors transitions on human reliability. In Guedes Soares
& Zio (eds). Safety and Reliability for Managing Risk.
Estoril Portugal, 1822 September 2006. London: Taylor
& Francis.
Konstandinidou, M., Nivolianitou, Z., Kyranoudis C. &
Markatos, N. 2006b. A fuzzy modelling application of
CREAM methodology for human reliability analysis.
Reliability Engineering & System Safety 91(6): 706716.
Pedrycz, W. 1993. Fuzzy control and Fuzzy Systems Second
extended edition, London: Research Studies Press Ltd.
Swain, A. & Guttmann, H. 1983. Handbook on Human Reliability Analysis with Emphasis on Nuclear Power Plant
Application NUREG/CR-1278 US Nuclear Regulatory
Commission.

289

Weston, L.M. & Whitehead, D.W. 1987. Recovery actions in

PRA for the risk methods integration and evaluation programme (RMIEP) NUREG/CR-4834, Vol. 1. US Nuclear
Regulatory Commission.
Zhang, L., He, X., Dai, L.C. & Huang, X.R. 2007. The
simulator experimental study on the operator reliability
of Qinshan nuclear power plant, Reliability Engineering
and System Safety 92: 252259.

Vakalis D., Sarimveis H., Kiranoudis C., Alexandridis A.

& Bafas G. 2004. A GIS based operational system for
wild land fire crisis management, Applied Mathematical
Modelling 28 (4): 389410.
Zadeh L.A. 1996. Fuzzy logic and the calculi of fuzzy rules
and fuzzy graphs, Multiple-Valued Logic 1: 138.
Zadeh L.A. 2008. Is there a need for fuzzy logic? Information
Sciences doi: 10.1016/j.ins.2008.02.012.

290

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The concept of organizational supportiveness

J. Nicholls, J. Harvey & G. Erdos
Newcastle University, UK

ABSTRACT: This paper describes the process of investigating, defining and developing measures for
organizational supportiveness in employment situations. The methodology centres on a focus group of people of diverse age, gender, grade and commercial and industrial disciplines that met many times over a period
of several weeks. The focus group contribution was developed into a large questionnaire that was pilot tested
on a general population sample. The questionnaire was analysed using factor analysis techniques to reduce it
to a final scale of 54 items, which was evaluated by a team of judges, and was then field tested in a nuclear
power station. The analyses revealed a supportiveness construct containing eight factors, being: communication,
helpfulness, empowerment, barriers, teamwork, training, security and health and safety. These factors differ
from other support-related measures, such as commitment, by the inclusion of a barrier factor. The findings
are evaluated with an assessment of the host company results and opportunities for further research.

INTRODUCTION

This paper aims to identify behaviours that contribute

to the harmonisation of collective goals between
employers and employees and develop a measure of
these behaviours that can be used to identify strategies
to promote the willing adoption of such behaviours
within business organizations. To help achieve this,
a construct of organizational supportiveness is developed that can be readily measured in organizations
and will indicate areas where management can alter,
introduce, or delete, practices and policies, which
will assist in the development of more positive internalised attitudes in their staff and consequently bring
added value to their business (Baker, 1996). Employee
expectations of their employers and their working
environments can be measured through such a scale.
The objects of this paper are to:
Define a construct of organizational supportiveness
and demonstrate its construct validity
Develop a measure of the construct and demonstrate
its internal consistency
Test the measure and demonstrate its generalisability.
2

with our family, friends, neighbours and work colleagues (Palmer, 2001). In the work context, new
technology and increasing competition have forced
businesses to examine their business strategies and
working practices and evolve new and innovative
means of maintaining their competitive advantage in
the face of stiff competition, particularly from emerging economies. This has resulted in the globalisation
of markets, the movement of manufacturing bases,
the restructuring of organizations, and changes to the
employer/employee relationship. Also, the changing
nature of traditional gender roles that has occurred in
the latter half of the 2th century (Piotrkowski et al.,
1987; Moorhead et al., 1997) and an increase in popularity of flexible working policies (Dalton & Mesch,
1990; Friedman & Galinski, 1992) are impacting significantly upon the traditional psychological contract.
Wohl (1997) suggests that downsizing and the associated reduction in loyalty and secure employment has
caused people to re-evaluate their priorities in respect
of their working and private lives, whilst Moses (1997)
states:

BACKGROUND

The last 50 years has witnessed enormous changes

in the social, political, cultural and working environments around the world. These changes have slowly,
but progressively, impacted on the lifestyles of the general population and the ways in which we all interact

291

Providing career support for staff has never been

more challenging for managers, given the diverse
demographic landscape and the complexity of
both the new work configurations and peoples
lives. Managers must help staff navigate in a jobless, promotionless, securityless environmenta
challenge that becomes even harder when much
of the workforce may be temporary, contract or
part time. The changes occurring in todays workplace are confusing to employees. Managers can

facilitate a sense of reassurance and renewal,

recognising that as leaders and representatives
of their organizations, their participation is
not only desirable but critical (Moses, 1997:
page 44).
So it is clear that supportiveness can be critical to
well-being of the workforce; it can also be stated that
the implications of getting this wrong can have serious
impact upon the reliability and safety of an organizational system or its processes. But supportiveness can
mean a lot more than social support, and can be related
to commitment, trust and the psychological contract,
amongst its many possible correlates. Consequently,
the literature on supportiveness is diverse and varied. It includes the above issues and the work ethic,
teamworking, organizational citizenship behaviours,
expectations and aspirations, and may be seen in
terms of theories of motivation, attribution, cognitive
dissonance etc.
In order to develop an understanding of what supportiveness an organization might provide, we propose
that the literature and research into supportiveness
can be summarised into five main thematic areas of
influence:
Supportiveness related constructs: psychological
contract, organizational citizenship behaviours,
organizational commitment and social support
Attitudes: trust, prevailing management culture,
beliefs, values, work ethic and motivation
Lifestyle variables: domestic arrangements such
as working partner, childcare or care of relatives,
environment and non-work interests
Aspirational variables: promotion, development,
transferable skills, job satisfaction, and flexible
hours
Business issues: welfare, pensions, and safety at
work, unions, legislation, and security.
For each of these areas, a body of literature existstoo much for one paper to cover, but exemplified by
the following.
The rapid changes to working life and societal values in the latter half of the 20th century, along with
changing economic circumstances and globalisation
have impacted on the traditional psychological contract and new working practices are evolving to meet
the demands of the changing work environment. There
is some evidence to suggest that traditional values of
work ethic and organizational commitment may be
waning in favour of more personal fulfilment outside
of work. However, it is not clear whether the pace of
change of peoples attitudes is less than or equal to
the pace of change of an emerging new psychological contract and the effect of a dynamic workplace
environment.

The ability to trust enables people to interact,

to build close relationships, is important for social
exchange and is essential for psychological health
and development (Asch, 1952; Erickson, 1959; Blau,
1964; Argyle, 1972; Kanter, 1977; Barber, 1983). Not
only that, but positive relationships have been found
between HR strategies and high organizational performance and commitment (Arthur, 1994; Huselid, 1995;
Tsui et.al. 1997). However, Koys, (2001) researched
the relationship between HR strategies and business
performance and contends that HR outcomes influence business outcomes rather than the other way
round. HR by its very title is concerned with people more than processes and suggests that it is in fact
people that make the difference.
The prevailing management culture can influence
support for alternative working arrangements differentially according to the supervisors perception of the
subordinates career success Johnson et al, 2008); this
may be consistent with a good or poor culture and prevailing beliefs about trust. Leung et al, (2008) have
shown that informal organizational supports, particularly those concerning relationship conflict, work
underload and lack of autonomy are more effective than formal supports in reducing stress; again
these may also reflect issues of trust, management
culture etc.
3

DEFINING A SUPPORTIVENESS
CONSTRUCT

Whilst some measures of near concepts, such as

social support, trust and commitment exist, none of
these are particularly relevant to the notion of the supportiveness that might provided by the organization.
Thus, it was decided to establish the construct from
first principles by asking people what they considered
a supportiveness construct would be.
The construct was developed through an iterative
process of:
Extracting variables from established constructs and
concepts within existing literature.
Holding informal interviews and conversations with
individuals and small groups from diverse employment situations to collate their ideas about what
they considered that a good supportive organization
would be.
Establishing a focus group of selected people who
would brainstorm the potential variables of their
perception of a supportive organization without the
influence of any prior input.
Establishing a number of informal ambassador
groups whose collective output was ultimately fed
into the formal focus group.
Collating all inputs and conducting a ranking and
rating exercise with the focus group.

292

Table 1.

The focus group constitution.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

4
4.1

Gender

Age

Industry

Job type

Female
Male
Female
Female
Female
Male
Male
Male
Male
Male
Female
Female
Male
Female
Male

3140
5160
5160
5160
<20
2130
2130
3140
4150
4150
2130
3140
>61
3140
5160

Power industry
Media
Leisure
Education
Commerce
Student
Construction
Research/development
Engineering
Local Govt
Retailing
Health
Shipping
Power industry
Off-shore

Senior manager
Employee
Employee
Middle manager
Clerical employee
n/a
Employee
Professional
Technician
Accountant
Assistant
Nursing sister
Employee
Clerical
Technician

METHODOLOGY
The focus groups

The sample of people consulted in the various groups

was chosen to be representative on the basis of age,
gender, job grade, industrial sector and whether
employed or unemployed. This was to provide confidence that all of the bases, including outliers, had been
considered and there was no inherent bias towards people currently in work. The constitution of the primary
focus group is shown in Table 1.
The ambassador groups consisted of a network of
independent and informal focus groups that were formulated and led by friends, colleagues and associates
(ambassadors) who had received the same brief as the
primary focus group.

4.2

The brief

The following brief was given to each member of

the primary focus group and each ambassador and
provided the remit for the groups contribution.
The purpose of this research is to identify
behaviours, which contribute to the harmonisation of
collective goals between employers and employees and
develop strategies that promote the willing adoption of
such behaviours within business organizations. I hope
to develop a concept of Organizational Supportiveness
and ultimately a measure of this concept that can be
operationalised into a questionnaire that can be used
to assess the degree of supportiveness that is given by
organizations to the employees of that organization.
To assist me in defining this concept, I am asking you,
both individually and collectively, to consider and list
all of the means by which organizations do, or could,

demonstrate their support to their workforce. In considering this task I would like you to include examples
from your own experience of:
Company policies, Temporary initiatives, Manager
and co-worker traits, Personal aspirations that you may
have held but not directly experienced. Include also
any other factor that you consider could have an impact
on this concept. It is equally important that you include
both positive and negative examples and ideas of organizational supportiveness in the workplace in order to
assess not only what a supportiveness concept is, but
also, what is unsupportive.
4.3

To develop a measure and establish

its construct validity and reliability

The focus group process yielded 51 separate elements

that the group rated in terms of their strength on a
subjective scale of high, medium and low strength
that reflected the relevance of that element to an
overall supportiveness concept. The 51 elements covered many diverse issues and include elements such
as allowing personal creativity through providing
medical support and pension schemes etc. Construct validity may be defined as the degree to
which inferences can legitimately be made from the
operationalisation in your study to the theoretical
constructs on which those operationalisations were
based (Trochim, 2002).
The theoretical construct in this study may be
defined as the mutual harmonisation of the human
and business imperatives of organizations. In order
to assess construct validity, a second focus group who
were experienced in questionnaire design and followed
Oppenheims (1994) principles, formulated the 51 elements into a 517 item questionnaire of approximately

293

10 Likert-style items per element with a 7-point scale

ranging from strongly disagree to strongly agree.
An initial pilot on 10 PhD management students was
followed by a larger pilot in cooperating organisations
in the general industrial and commercial domain, generating 103 responses (47 male, 32 female and 24
gender undisclosed). The questionnaires were factor
analysed and rationalised into seven factors that were
represented in a 54 item questionnaire; an eight factor,
health and safety, was omitted from this final questionnaire as it was felt that this would lengthen the
questionnaire unnecessarily when such issues could be
better tested using the organizationss own health and
safety inspection procedures. An established Organizational Commitment measure (Meyer & Allen 1997)
was added for the field test in a single organisation (a
nuclear power station) about which a great deal was
known i.e. the people, processes, operational culture,
corporate history etc. Consequently, it was possible to
take the mean scores of the returned questionnaires and
ask three inter-rater judges to give their opinion of the
mean score, between 1 and 7, for each question. This
process, whilst not fitting the classical application of
inter-rater reliability, is consistent with the approach
described by Marques & McCall, (2005) and yielded
general consensus between the judges that the measure
derived from the questionnaire was a true reflection
of the actual level of organisational supportiveness
present at that time.

coefficients for job security and teamwork. The results

of this analysis is shown in Table 3
One explanation for this may be that teamwork
relies on some type of inter-personal bonding between
co-workers that takes time to establish and that security cannot be felt during a period of probationary
employment. The standard terms and conditions of
employment at the nuclear power station required all
employees to satisfactorily complete a one/two year
probationary period before being confirmed in their
post. This observation may be consistent with the
measure being a reliable indicator of perceived organizational supportiveness if it is accepted that UK
employment legislation permits unaccountable dismissal within the first 12 months of employment
without any recourse to appeal, then it is reasonable for
new employees not to feel secure in any organisation.
As a validity and benchmarking exercise to compare
our factors with commitment, the commitment measure of Meyer and Allen (1997) was correlated with
F1 communications and F4 barriers, with the results in
Table 4. The results suggest that affective commitment
is strongly associated with both factors, as indeed it
often is with other word attitudes such as trust and job
satisfaction. For the normative commitment measure
that relates more to obligation, this is more strongly
related to F4 barriers than to communications, which
would be consistent in terms of attitude theory.
5.2 Sample size issues

5
5.1

RESULTS
The 8-factor solution

The results from the two independent datasets, the initial pilot test in the general population (N = 103)
and the field test in a host organisation (N = 226),
were compared and found to be statistically in 91%
agreement. So the two datasets were combined into a
one (N = 329) and subjected to another exploratory
factor analysis where they demonstrated a 96% agreement with the field test result. The factors of this
second exploratory factor analysis with alpha coefficients for the host organisation (N = 226) are shown
in Table 2. The eighth factor was identified as health
& safety and conditions of employment). Since this
involves clear lines of statutory enforcement and corporate responsibility, it is omitted henceforth from the
analysis since these are largely outwith the control of
the local workforce.
Further analysis of each of the 7 factors in respect
of: age, gender, grade, length of service and shiftwork/daywork effects showed no significant difference from the generalised test results with the exception of length of service effects for employees with
less than 2 years service who demonstrated lower alpha

294

The two datasetsfrom the pilot test in the general

population and the field test at the nuclear power
station- were gathered approximately 1214 months
apart and the closeness with which these two independent surveys agree with each other suggests that the
construct has both internal consistency and construct
validity. However, neither survey produced the very
high ratios of cases to variables (N:P ratio) that factor analysis ideally requires to have a high confidence
in the outcome of the factor analysis. The first survey had a 2:1 N:P ratio and the second a 4.5:1 N:P
ratio. As the questionnaires were identical, their data
can be combined into a single dataset giving a 6.5:1
N:P ratio. This not only increases the cases to variables ratio but it mixes one set of data that could be
focused on and therefore skewed to a particular organization (the field test data) with data that were in
effect randomly gathered from a number of different
organizations (the general population data).
There is some debate about what constitutes a suitable sample size for factor analysis. In general, it
is reasonable to say that the larger the sample the
better the results. Small samples are likely to yield
spurious results where factors splinter and cannot
be replicated in subsequent analyses or may contain unrepresentative bias (Froman 2001). Minimum

Table 2. The 8-factor solution of the field test in the host organization, excluding factor 8 (N = 226).
Factor 1 COMMUNICATION ( = .866)
1 My Manager communicates too little with us.
2 I am told when I do a good job.
3 I can speak to my Manager at anytime.
4 My Manager operates an open door policy.
5 My Manager is always open to my suggestions.
6 My Manager will always explain anything we need
to know.
7 My Manager only tells me what I have done wrong,
not what should be done to improve things.
8 I sometimes feel I am wrongly blamed when things
go wrong.
9 I can speak to my Manager and tell him/her what I
would like.

Factor 4 BARRIERS ( = .912)

31 I trust this organization.
32 Managers here do not appreciate the work that I do.
33 Only the in crowd get promotion here.
34 Management here do not care about the workers.
35 At my place of employment, it is every man for himself.
36 The culture at work does not foster loyalty.

Factor 2 HELPFULNESS ( = .923)

10 My Manager will assist me with difficult tasks.
11 If I need advice and support, it is available here at
work.
12 My Manager is loyal to me.
13 I feel I can rely on my Manager defending my
groups interests.
14 My Manager always looks after him/herself first.
15 My Manager always plays down the good things
that I do.

37 Management here are manipulative.

38 Vindictiveness among managers is common at work.
39 Generally, Management here are supportive of the
workflow.
40 This organization has business integrity.
41 The reasons given for decisions are not usually the real.
42 This organization always does the morally right things.
43 Managers here put their own agendas before those of my
group.
44 In this Company everybody is out for themselves.

16 My Manager is only interested in his own career.

17 My Manager does not care about my domestic
problems.
18 My Manager always sets unrealistic targets.
19 My Manager encourages me to continually develop
my skills and knowledge.
20 Sometimes I am given a task that the Manager
knows will probably fail.
21 My Managers behaviour is very varied and
inconsistent.
22 My Manager has their favourite people.
23 How my Manager treats me changes from day to day.
day.

Factor 5 TEAMWORK ( = .789)

45 Interaction with others in my group is an important part of
what I do.
46 I generally enjoy working as part of a team.
47 To be successful, my job depends on the work of colleagues.
48 We depend on each other in my workgroup.
Factor 6 TRAINING ( = .708)
49 This organization provides training to maintain existing
skills and to develop new skills.
50 There is a good induction training programme here now.
51 We are given the opportunity to get re-training whenever
we want it.

Factor 3 EMPOWERMENT ( = .903)

24 My workplace is one where I can be creative or
flexible.
25 My Manager seeks my views.

Factor 7 SECURITY ( = .721)

52 I feel relatively secure here.
53 My job will still be here in 5 years time.
54 I feel my job is very safe.

Table 3.

26
27
28
29
30

I have choices and decisions to make in my job.

I am given freedom to implement my ideas at work.
My Managers listens to my ideas.
My Manager encourages me to think of new ideas.
I am allowed to make improvements to my daily
routine.

Factor means and Cronbachs alphas by length of service in years.

Factor

Mean score

Cronbach coefficient

Length of service

<1

12

25

510

1020

>20

<1

12

25

510

1020

>20

Communication
Helpfulness
Empowerment
Barriers
Teamwork
Training
Security
N

5.097
4.908
4.304
4.634
5.000
4.917
3.958
8

5.232
4.948
4.335
4.438
5.630
5.014
3.174
23

4.421
4.350
3.929
4.104
5.792
4.625
3.625
24

4.222
3.978
4.190
3.881
5.583
3.444
3.333
3

5.003
4.711
4.777
4.110
5.666
4.617
4.042
80

4.833
4.600
4.758
4.097
5.599
4.920
4.102
88

.790
.900
.929
.803
.880
.786
.730
8

.861
.921
.764
.887
.317
.840
.341
23

.907
.953
.920
.946
.837
.831
.795
24

.967
.986
.841
.849
.793
.627
.60
3

.844
.909
.909
.893
.722
.600
.738
80

.855
.933
.905
.926
.852
.681
.740
88

295

Table 4. Pearsons correlations between F1, F4 & the Meyer

and Allen (1997) commitment scales.
Commitment
scores

F1
Communications

F4
Barriers

Normative
Continuance
Affective

.282
.066
.545

.498
.041
.592

Key: = p < .0001.

sample sizes suggested range from 3 to 10 subjects

per item with a minimum of between 100 and 300
subjects regardless of the number of items (Gorsuch 1983; (Cattell 1978; Tinsley & Tinsley 1987;
Nunnally & Bernstein 1994). Higher estimates range
from a large sample of several hundred (Cureton &
DAgostino 1983) to 20 subjects per factor (Arindell
& Van der Ende 1985). The sample sizes used in
this analysis are for the pilot test, 103 subjects which
satisfy Gorsuchs (1983) criteria; for the field test,
226 subjects, which satisfy Gorsuchs (1983), Cattells
(1978) and Tinsleys & Arindells (1987) criteria. For
the combined sample of 329, all of the criteria are
satisfied.
5.3

Ordinal scale

From the results that were obtained, a subjective

supportiveness scale was developed for the host
organisation; this can be used for each subscale and
its linguistic anchors for each scale point are: 7 =
exceptionally supportive, 6 = very supportive, 5 =
supportive, 4 = neither and so on to 1 = very poor.

A means of measuring, particularly in terms relative

to a previous or subsequent measure, the strength or
level of supportiveness that currently exists within any
organization has been developed.
The construct validity and internal consistency of
the measure has been demonstrated.
We have shown that organizational supportiveness
is an independent construct compared to organizational commitment and not a direct correlate of OC.
Trust and loyalty appear consistently as important throughout the management literature, and their
absence appears regularly to be associated with negative consequences. The ability to trust enables people
to interact, to build close relationships, is important
for social exchange and is essential for psychological health and development (Erickson, 1959; Blau,
1964; Argyle, 1972; Kanter, 1977; Barber, 1983).
A supportiveness construct such as this one with a
focus on communication can build emotional trust
and an emphasis on developmental training can mitigate the uncertainty of the current business environment. In asking where the supportiveness construct fits into the body of theory and constructs that
exist to describe and explain workforce attitudes to
their employers and to business organisations, there
is evidence that positive support through good HR
policies are associated with high organizational performance as well as employee commitment and trust
(Koys, 2001; Arthur, 1994; Huselid, 1995; Tsui
et.al. 1997). In addition, the evidence that employee
assistance programmes support systems more than
pay for themselves demonstrates the usefulness of
supportiveness.

REFERENCES
6

CONCLUSION

In summary, the following outcomes have been

achieved:
The important philosophical and practical elements
that define an organizational supportiveness construct
have been identified.
These elements have been rationalised into seven
factors that can be measured and operationalised into
policies and strategies in any organization.
An eighth important factor was identified (Health
& Safety and Conditions of Employment) that may not
be within the direct control of most people within an
organization, but one for which clear lines of statutory
enforcement and corporate responsibility exists. This
factor has been omitted from the construct on the basis
that health & safety is bounded by statute and terms
and conditions of employment are very often outwith
the control of the local workforce other than during
periods of formal collective bargaining.

Arindell, W.A. & Van der Ende, J. (1985). An empirical

test of the utility of the observation-to-variables ratio in
factor and components analysis. Applied Psychological
Measurement, Vol. 9: 165178.
Argyle, M. (1972). The Social Psychology of Work. Harmondsworth: Penguin.
Arthur, J.B. (1994). Effects of human resource systems on
manufacturing performance and turnover. Academy of
Management Journal, Vol. 37: 670687.
Baker, N. (1996). Lets talk about relationships. workplace &
family www.findarticles.com/cf_0/m0FQP/n4301_v125/
1871 7517/print.jhtml
Barber, B. (1983). The Logic and Limits of Trust. New
Brunswick, NJ: Rutgers University Press.
Blau, P.M. (1964). Exchange and Power in Social Life. New
York: Wiley.
Briner, R.B. (1996), Making Occupational Stress Management Interventions Work: The Role of Assessment.
Paper to the 1996 Annual BPS Occupational Psychology
Conference.
Cattell, R.B. (1978). The scientific use of factor analysis.
New York: Plenum.

296

Cronbach, L.J. & Meehl, P.E. (1955), Construct validity

in psychological tests. Psychological Bulletin, Vol. 52:
281302.
Cureton, E.E. & DAgostino, R.B. (1983). Factor analysis:
An applied approach. Hillsdale, NJ: Erlbaum.
Dalton, D.R. & Mesch, D.J. (1990). The impact of flexible scheduling on employee attendance and turnover.
Administrative Science Quarterly, Vol. 35: 370387.
Erikson, E.H. (1959). Identity and the life cycle. Psychological Issues, Vol. 1: 1171.
Friedman, D.E. & Galinski, E. (1992). Work and family issues: A legitimate business concern. In S. Zedeck
(Ed.), Work Families and Organizations. San Francisco:
Jossey-Bass.
Froman, R.D. (2001). Elements to consider in planning
the use of factor analysis. Southern Online Journal
of Nursing Research, Vol. 2 No. 5: 122 www.
snrs.org/publications/SOJNR_articles/iss05vol02_pdf
Gorsuch, R.L. (1983). Factor Analysis (2nd Ed.). Hillsdale,
NJ: Erlbaum.
Huselid, M.A. (1995). The impact of human resource management practices on turnover, productivity, and corporate
financial performance. Academy of Management Journal,
Vol. 38: 635672.
Johnson, E.N., Lowe, D.J. & Reckers, P.M.J. (2008). Alternative work arrangements and perceived career success:
Current evidence from the big four firms in the U.S.
Accounting Organizations and Society. Vol. 33 No. 1:
4872.
Kanter, R.M. (1977). Men and Women of the Corporation.
New York: Basic Books.
Koys, D.J. (2001). The effects of employee satisfaction,
organization citizenship behavior, and turnover on organizational effectiveness: A unit level longitudinal study.
Personnel Psychology, Vol. 54 No. 1: 101114.
Leung, M.Y., Zhang, H. & Skitmore, M. (2008) Effects of
organisation supports on the stress of construction estimation participants. Journal of Construction Engineering
and Management-ASCE Vol. 134 No. 2:8493.

Marques, J.F. & McCall, C. (2005). The Application of

Interrater Reliability as a Solidification Instrument in a
Phenomenological Study. The Qualitative Report, Vol. 10
No. 3: 438461.
Meyer, J.P. & Allen, N.J. (1997). Commitment in the
Workplace. Sage.
Moses, B.(1997). Building a life-friendly culture. Ivey
Business Quarterly, Vol. 62 No. 1: 4446.
Moorhead, A., Steele, M., Alexander, M., Stephen, K. &
Duffin, L. (1997). Changes at work: The 1995 Australian
workplace and industrial relations survey. Melbourne:
Longman.
Nunnally, J.C. & Bernstein, I.H. (1994). Psychometric
Theory (3rd. Ed.). New York: McGraw Hill.
Oppenheim, A.N. (1994). Questionnaire Design, Interviews
and Attitude Measurement. (2nd Edition) Pitman.
Palmer, M.J. (2001). Why do we feel so bad? (personnel management in the health care industry) http://
www.findarticles.com/cf_0/m0HSV/6_14/79788282
/print.jhtml
Piotrkowski, C.S., Rapoport, R.N. & Rapoport, R. (1987).
Families and work. In M.B. Sussman & S.K. Steinmetz (Eds.), Handbook of marriage and the family (pp.
251283). New York: Plenum Press.
Tinsley, H.E. & Tinsley, D.J. (1987). Uses of factor analysis in
counselling psychology research. Journal of Counselling
Psychology, Vol. 34: 414424.
Trochim, M.K. (2002). Construct Validity. http://www.
socialresearchmethods.net/kb/constval.htm
Tsui, A.S., Pearce, J.L., Porter, L.W. & Tripoli, A.M. (1997).
Alternative approaches to the employee-organization relationship: Does investment in employees pay off? Academy
of Management Journal, Vol. 40: 10891121.
Wohl, F 1997. A panoramic view of work and family. In S.
Parasuraman & J.H. Greenhaus (Eds.), Integrating work
and family: Challenges and choices for a changing world.
Wesport, CT: Greenwood Publishing Group.

297

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The influence of personal variables on changes in driver behaviour

S. Heslop, J. Harvey, N. Thorpe & C. Mulley
Newcastle University, Newcastle, UK

ABSTRACT: There are several theories that relate to drivers risk-taking behaviour and why they might choose
to increase the level of risk in any particular situation. Focus groups were run to identify transient factors that might
affect driver risk-taking; these were recorded, transcribed and content-analysed to obtain causal attributions. Five
main themes emerged, which could then be sub-divided; these themes are discussed in light of existing theories.
It was found that the attribution to self was the most frequent, but that causal explanations were consistent with
the theories of zero-risk, risk homeostasis and flow.

1
1.1

INTRODUCTION AND BACKGROUND

Introduction

Driver behaviour has, for many years, been recognised

as the main cause of death and injury on the roads.
Sabey and Taylor (1980) for example, showed that
road users were a causal factor in 94% of collisions.
The road environment and the vehicle were causal factors in 28% and 8.5% of collisions (total greater than
100% because crashes can be attributable to multiple
causes.) The driver has repeatedly been shown to play
a dominant role in collision causation (Department for
Transport, 2007).
The research reported here investigates which motivations and other transient factors are most salient with
respect to affecting driver behaviour.
1.2

Behavioural adaptation

In a seminal paper Taylor (1964) measured the galvanic skin response (GSR) of drivers in the following
road types or conditions: urban shopping streets; winding country roads; arterial dual carriageways; peak and
off-peak; day and night and found that GSR, taken to
be a measure of subjective risk or anxiety, was evenly
distributed over time over the range of roads and conditions studied. His results suggest that driving is a
self-paced task governed by the level of emotional tension or anxiety that drivers wish to tolerate (Taylor,
1964). It is now accepted that drivers adapt to conditions and state of mind: Summala (1996, p.103) stated
that the driver is inclined to react to changes in the
traffic system, whether they be in the vehicle, in the
road environment, or in his or her own skills or states.
Csikszentmihalyis (2002) flow model represents
the following two important dimensions of experience: challenges and skills. When individuals are

highly skilled in an activity and are faced with low

challenges, the individuals experience a state of boredom. Conversely, when individuals are not skilled in
an activity and are faced with significant challenges,
they experience a state of anxiety. When individuals
skills and the challenges posed by an activity are evenly
matched, the individuals experience a pleasurable flow
state. Csikszentmihalyis model (2002) suggests that
under normal driving conditions drivers remain within
a flow channel, balancing challenges, ultimately controlled by speed and attention (Fuller, 2005). Should
the driving environment become more challenging, a
reduction in speed or an increase in attention, perhaps achieved by aborting secondary tasks can bring
the driver back into a flow state. If it became less
challenging, an increase in speed or reduction in attention, perhaps achieved by adopting secondary tasks
can similarly ensure that the driver remains in a flow
state. Thus by varying their speed or attention, drivers
can remain in flow when challenges and or skill levels
change.
In the context of road safety, an understanding of the
circumstances under which drivers might be motivated
to accept (knowingly or otherwise) increased risks or
challenges is of particular interest. The role of motives
in driver behaviour has been recognised in two main
theories: the zero-risk theory (Ntnen and Summala,
1974) and the theory of task difficulty homeostasis
(Fuller, 2000).
The zero-risk theory suggests that driver behaviour
is a function of mobility, extra motives and safety
(Summala, 1998). On the one hand mobility and
extra motives such as self-enhancement, time goals,
thrill-seeking, social pressure, competition, conservation of effort, pleasure of driving and maintenance of speed and progress push drivers towards
higher speeds (Summala, 2007). On the other hand,

299

warning signals control risk taking if safety margins

that are learned through experience are compromised
(Summala, 1998).
The theory of task difficulty homeostasis argues
that drivers monitor task difficulty and aim to drive
within an acceptable range of difficulty (Fuller, 2000).
Safety margins may be affected by motives, through
their influence on the upper and lower limits of acceptable task difficulty. Fuller (2000) lists motives that
push drivers towards higher speeds and reduced safety
margins in three categories: social pressure; critical
threshold testing; and other motives. Social pressure
includes: maintaining the speed of flow in a traffic
stream; pressure from passengers; fear of arriving late
for an appointment; a desire to drive like others; and
a wish to avoid sanctions. Critical threshold testing
includes a desire to test the limits, perhaps attributable
to sensation seeking needs, and a drift towards the limits, due to a lack of negative feedback associated with
previous risky driving practices. Other motives include
time pressure, time-saving and gaining pleasure from
speed.
Research has been carried out in order to establish how drivers are affected by motivation in terms
of either increased or reduced safety margins: when
drivers are motivated by time saving, they drive faster
thus taking increased risks (see for example: Musselwhite, 2006; McKenna and Horswill, 2006); similarly,
when angry, they adopt aggressive driving patterns
and drive faster, thus reducing safety margins (see for
example: Mesken, Hagenzieker, Rothengatter and de
Waaard, 2007; Stephens and Groeger, 2006).
Whilst theories of driver behaviour acknowledge
that motivations play an important role in terms of
driver risk taking, there has been no research to determine which are the most salient motives and transient
factors of influence other than in relation solely to
speed (Silcock et al., 2000).
2

METHODOLOGY

Eight individuals with full UK driving licences who

drove on a frequent basis participated in the study.
The study group comprised 6 males and two females,
with ages spread across the range 1765. Data were
recorded using a portable cassette recorder with microphone.
2.1

Design and procedure

Two unstructured focus group discussions were held.

Participants were encouraged to talk about their own
driving behaviour, and the circumstances under which
they felt that their driving behaviour changed, giving examples where appropriate. Both focus groups
were transcribed for content analysis, including

attributional analysis. Attributional statements have

been defined as any statement in which an outcome is indicated as having happened, or being
present, because of some event or condition (Munton,
Silvester, Stratton and Hanks, 1999). Statements meeting this definition were extracted from the transcript.
Where several different causes or effects were given for
a particular relationship, attributions were extracted
and coded separately.
3

RESULTS

1026 attributional statements were extracted from the

transcripts and were reduced into major causal areas
(attribution frequency in parentheses):
self (1831)
physical environment (660)
other road users (251)
vehicle (105)
road authority (105)
The self theme can be extended or broken down into
the following:
extra motives (879)
flow state (332)
perception (233)
journey type (142)
capability (116)
time pressure (103)
experience (26)
Data were then coded within each of these themes.
A selection of the attributional statements of other road
users, vehicle and the roads authority as causal themes
are shown in Table 1.
Self and physical environment themes had to be
further divided in order to preserve the richness of
the data. The physical environment theme was divided
into the following sub-themes: physical environment,
road condition, and road familiarity. To illustrate the
presence of these causal themes, a selection of attributional statements are shown in Table 2. The self
theme was divided and is shown in Figure 2. To
illustrate the presence of self in the sub-themes,
a selection of attributional statements are shown in
Table 3.
4

DISCUSSION

We have shown from this research that there are five

main categories of causal areas, being the driver him
or herself; the physical environment; other road users;

300

Table 1.

Statements where the cause was coded under other road users, vehicle and roads authority.

Attributional statement
[a] For other road users as a cause
Id accept a smaller gap, if I was waiting to get out
(the old bloody biddies driving like idiots) really, really gets up my nose
there was no traffic, at all (on M1) so you just put your foot down
then I saw a straight and took a risk that I wouldnt usually have taken
If Im in a traffic jam I turn the radio up
[b] For vehicle as a cause
driving the big car is helped because Ive got a much better view
they all assume that Im a builder or plasterer and therefore behave in a lewd way
[c] For roads authority as a cause
I would definitely stick to the speed limit if it was wet
Im unwilling to overtake on those roads because partly they remind you how many
people have been killed there
Beach Road (I think thats a forty limit) is one of my favourites for (breaking speed
limits); Its . . . in effect, a three lane but its configured as a two-lane road

Table 2.

Identity code

Freq.

High traffic levels

Incompetence
Low traffic levels
Slow car in front
Traffic jam

68
17
12
54
22

Large vehicle
White van

45
19

60 mph limit
Road safety
measures

22
25

Limit too slow

20

Statements where the cause was coded under physical environment, road condition, and road familiarity.

Attributional statement
[a] Physical environment theme
I fell asleep on the motorway
I drive with smaller safety margins in London
You just go round them faster (as you drive along the route)
In Newcastle old ladies are terribly slow and that irritates me
Trying to overtake on those two stroke three lane roads is quite nerve racking
Id accept a smaller gap, if I was in a hurry
(I can enjoy the driving when on) familiar rural roads (in Northumberland)
Driving round streets where people live, (Im much more cautious)
I got to the end of the Coast Road, and thought, I dont actually remember going
along the road
Put me on a bendy country road: Ill . . . rise to the challenge
I was a bit wound up so I was sort of racing, traffic light to traffic light
youve got 3-lane 30 mph roads; I think it is safe to drive there at 60 mph)
[b] Road condition theme
(there was black ice) and I went up the bank and rolled the car
[c] Road familiarity theme
(the old people that come into the city centre here) changes the way I drive

the roads authority and the vehicle being driven. Of

these, by far the most frequently occurring was self.
Each of these is considered below in relation to existing
theory and evidence.
Self contained seven sub-themes: perception;
capability; journey type; time pressure; extra motives;
flow state; and experience; each of these is considered
below.
In terms of perception, whether the situation felt
safe or felt risky were frequently reported. Perception is a stage inherent within most models of driver
behaviour (Fuller, 2000; Summala, 1998 and 2007;
Ntnen and Summala, 1974). This study has found

Code identity

Freq.

Dual/motorway
London
Many roundabouts
Newcastle centre
Open rural roads
Priority junction
Rural roads
Slow urban roads

123
18
15
10
55
28
33
24

Tyne and Wear

Tight rural roads
Urban roads
Wide urban roads

40
98
26
14

Adverse

25

familiar

88

that people adjust their speed upwards or downwards

according to the weather or their knowledge of the
road; thus zero-risk is an inappropriate explanation
prima facie as the situation does not involve the opposing motives proposed by Summala (1998). However
the idea that drivers adapt to the environment in order
to remain within a flow state or avoid boredom might
be a better explanation, as would the homeostasis
theory (Csikszentmihalyi, 2002; Fuller, 2000). Capability was also a frequently reported cause: it was
often perceived to have increased (such as being in
a larger vehicle) or reduced (through fatigue). If capability increased, drivers are expected to adapt by either

301

Table 3.

Statements where the cause was coded under the umbrella theme, self .

Attributional statement
[a] Perception theme
In adverse weather I drive more slowly
because I knew it really well, I would drive faster
[b] Capability theme
you drive differently in the small car
(I can drive faster when alone in the car because no distractions)
[c] Journey type theme
I find I can push out into traffic when driving a mini bus
I actually unwound driving home
When Im on a long journey I drive faster
I can enjoy the driving when its pretty rolling countryside
[d] Time pressure theme
you have to (drive more aggressively (in London))
(I was driving calmly) partly because there wasnt a set time I had to be home
[e] Extra motives theme
Everyone expects you to abuse a minibus a bit, like U-turns in the road)
I drive my fastest when Im on my own
(GPS) encourages me to keep the speed up
Yes, (I drive at a speed Im comfortable at regardless of the speed limit)
if Im following someone who is more risky (I will raise my risk threshold)
If somebodys driving really slowly Ill judge whether I can get past safely
(Sometimes I find if you turn it up) youre just flying along
I was really upset, and I drove in a seriously risky manner
I know I can get distracted if Ive got passengers in the car
I only drive slowly with you (because) Id want you see me as a good driver
Ill just tear down there (in order to impress my friend)
I saw the traffic jams, as a means of winding down and actually unwound
I would hang back (if I saw risky driving in front of me)
I drive fast on bendy country roads
I just wanted to show him he was an idiot (so chased him)
I was starting to get really sleepy and I blasted the radio up really loud
[f] Flow state theme
In adverse weather I drive more slowly
(If Im in a traffic jam and its boring) I annoy other drivers by . . . (singing)
[g] Experience theme
Almost hitting the slow car in front makes me concentrate more for a while

increased speed or reduced concentration; again these

ideas are consistent with both flow and homeostasis
explanations (Csikszentmihalyi, 2002; Fuller, 2000).
Extra motives, illustrated in Table 3, include:
boredom relief, flow-seeking, maintaining progress,
negative and positive emotions, penalty avoidance,
relaxation, risk-aversion, risk and thrill-seeking and
tiredness. The theory of flow (Csikszentmihalyi, 2002)
would predict boredom relief motives to result in either
faster driving or adoption of secondary tasks, in either
case reducing safety margins; flow-seeking motives
should encourage faster driving when the challenges
are low but slower driving when the challenges are
high. The findings here would support this explanation, but so they also would support the homeostasis
and zero-risk theories, both of which emphasise the
motivation tendencies propelling drivers to drive faster

Code identity

Freq.

Feels risky
Feels safe

102
123

Reduced
increased

59
57

Club trip
Commute home
Long journey
Leisure trip

46
15
20
38

Yes
No

62
41

Act as 44 van driver

Alone in car
Boredom relief
Flow seeking
Following someone
Maintaining progress
Music
Negative emotions
Passengers
Show off safe driving
Show off risky
relaxation
Risk aversion
Risk seeking
Road rage
fatigue

32
13
40
52
17
222
21
17
11
31
18
57
110
103
10
11

Anxious
Bored

108
224

Accident or near miss

16

as sensation-seeking or to avoid lateness (Summala

2007; Fuller 2000; Ntnen and Summala, 1974). In
our study, negative emotions such as anger resulted in
drivers effecting more aggressive behaviour and hence
reduced safety margins whereas positive emotions had
the reverse effect and resulted in increased safety margins; this is consistent with other findings (Underwood
et al., 1999; Stephens and Groeger, 2006). Other
motives such as penalty avoidance, relaxation, tiredness and risk-aversion all led to reduced speeds but
probably increased boredom and took drivers out of
their flow channel or band of acceptable task difficulty; again this is consistent with other findings
(van der Hulst et al., 2001). Risk-seeking motives
are likely to encourage driving at the upper bounds
of the flow channelband of acceptable task difficulty
(Csikszentmihalyi, 2002).

302

Journey type was an important causal theme in our

findings and included leisure trips, long journeys and
commuting. Motivations related to different journey
types have associated and different effects on driver
behaviour: for example, club trips (found in this study)
might encourage risky driving amongst peers (Arnett
et al., 1997) whilst leisure trips may be expected to
encourage a generally calmer and more relaxed driving manner. In general for any journey type, whether
the driver was under time pressure or not (including time-saving) were frequently nominated factors
of influence. Motorists are likely to drive faster and
more aggressively when under time pressure, and by
implication in a more relaxed fashion when not, as
was found here (Musselwhite, 2006; McKenna and
Horswill, 2006).
Flow state was found to be a key theme: whether the
situation resulted in the driver feeling bored or anxious
was reported frequently. The flow model (Csikszentmihalyi, 2002) suggests that a state of boredom or
anxiety is not pleasant, and as such drivers are expected
to look for opportunities to achieve a flow state either
by adjusting task demands or attention, such as our
examples illustrate.
Even car insurance premia reflect experience, so
unsurprisingly it was also found to be a causal influence, but only represented in our data by whether the
driver had experienced an accident or near miss. This
also fits with the commonly observed phenomenon of
drivers slowing down after observing the evidence of
an accident, only to speed up again later. Initial findings support the suggestion that drivers adopt more
risk-averse attitudes after such events and as such are
expected to increase safety margins for a period, driving within the lower limits of their flow channels; this
explanation fits all three theories.
Other road users are also cited frequently, with high
traffic levels the most frequently nominated cause.
Predictably, high traffic levels were found to affect
anger, violations and accidents; being held up by
a slower car in front and responding aggressively
were also causes cited in our study, again consistent with other research (Underwood et al., 1999;
Bjorklund, in press). Additional causes within this
theme included other drivers incompetent or risky
behaviours and traffic jams affecting anger and low
traffic levels encouraging higher speeds via reduced
challenges (Bjorklund, in press; Csikszentmihalyi,
2002). That other road users were found to be a main
causal theme in the determination of driver behaviour
supports both the zero-risk and risk homeostasis theories where perception of the environment (including
other road users behaviour) are key stages within the
models (Fuller, 2000; Summala, 2007; Ntnen and
Summala, 1974).
The vehicle being driven, in terms of size and
image, were factors of influence in our study. Driving

a large vehicle appears to result in faster driving,

presumably from increased visibility and reduced
challenges, whilst the opposite appears to be the
case when driving a small vehicle (Csikszentmihalyi, 2002). Vehicle image appeared to influence driver
behaviour if the vehicle being used wasnt the drivers
usual vehicle. In light of work by Newman et al. (2004)
drivers could be expected to drive at reduced speeds
in non-private vehicles, but our findings indicate the
opposite to be true.
The road authority can self-evidently be an influence, for example in terms of safety measures, traffic
calming measures, chevrons and speed limits and these
in fact were frequently nominated causal factors. The
fact that drivers nominated these as causal factors
suggests that they are likely to slow down in these situations (McKenna and Horswill, 2006), possibly because
of fear, responsibilities, correction of risk misperceptions and associations between warnings and increased
challenges. Drivers also stated as a causal factor that
the speed limit is too low; flow theory would suggest that they would weigh up the costs and benefits
of speeding versus experiencing a state of boredom
(Csikszentmihalyi, 2002).
The physical environment theme was divided into
the physical environment, road condition and road
familiarity. Road type was frequently reported as an
influence, particularly motorways or dual carriageways, followed by tight rural roads and open rural
roads; these different road types have different design
speeds to which drivers are expected to adapt- for
example we found drivers saying they drove faster
on open roads with high visibility, and slower on
tight roads where it is reduced (Csikszentmihalyi,
2002). Within the road condition sub-theme, the only
factor cited in this study was adverse conditions,
such as water or snow on the road, where drivers
would be expected to slow down, taking more care
(Csikszentmihalyi, 2002; Stradling, 2007). In terms
of road familiarity, drivers said they would drive faster
on familiar roads, consistent with all of the three main
theories which suggest that challenges are reduced
with familiarity (Stradling, 2007); this can be accommodated by an increase in speed and/or a reduction
in attention. We also found road familiarity may be
associated with increased confidence and a sense of
ownership; an example of this being the old people
that come into the city centre here change the way I
drive.
To summarise, the findings of self as the dominant
cause is an interesting attribution, not entirely consistent with findings elsewhere that most drivers make the
attribution of blame to others; this is consistent with
self-serving attributions and the fundamental attribution error in attribution theory (Hewstone, 1989).
Most of our findings are consistent with all three
theories of driver behaviour and flow (Fuller, 2000;

303

Summala 2007; Csikszentmihalyi, 2002; Ntnen

and Summala, 1974).

CONCLUSIONS

This research forms the preliminary base for a longer

and more detailed study, but this required the identification of themes from which to proceed, and these
have been identified here. In all, five major themes
have been identified which sub-divide into 13 smaller
themes, which can form the basis for evaluation and
judgement in further interviews and surveys. However, even with these themes and our data supporting
them as often very frequently occurring, it is not possible to identify which of the three main explanatory
frameworks-flow theory, zero-risk theory or homeostasis offers the best explanations for either our
themes or their frequency of occurrence. Somehow, all
these theories seem to tell us that drivers may be motivated to drive faster, but whether this is to maintain
flow, or for some form of extra motive, or as critical
threshold testing, cannot easily be ascertained, even
with the causal reasons provided in our attributional
analyses.
The study so far has some limitations- we need
more focus groups, but directed more towards causal
factors particularly for self. Once these have been
conducted and analysed, then questionnaires can be
built to test some of the more pertinent aspects of
self and other causes. Additionally, we have focussed
solely on drivers in one part of the United Kingdom,
so more work will be done in relation to other driving
areas, such as capital cities and rural environments.

REFERENCES
Arnett, J.J., D. Offer, et al. (1997). Reckless driving in adolescence: State and trait factors. Accident Analysis &
Prevention 29(1): 5763.
Bjorklund, G.M. Driver irritation and aggressive behaviour.
Accident Analysis & Prevention In Press, Corrected Proof.
Csikszentmihalyi, M. (2002). The classic work on how to
achieve happiness, Rider.
DfT (2007). Road Casualties Great Britain: 2006Annual
Report, The Stationary Office.
Fuller, R. (2000). The task-capability interface model of the
driving process. RechercheTransportsSecurite 66:
4757.

Fuller, R. (2005). Towards a general theory of driver

behaviour. Accident Analysis & Prevention 37(3):
461472.
Hewstone, M. 1989. Causal Attribution: from cognitive
processes to collective beliefs. Oxford: Blackwell.
McKenna, F.P. and M.S. Horswill (2006). Risk Taking From
the Participants Perspective: The Case of Driving and
Accident Risk. Health Psychology 25(2): 163170.
Mesken, J., M.P. Hagenzieker, et al. (2007). Frequency,
determinants, and consequences of different drivers emotions: An on-the-road study using self-reports, (observed)
behaviour, and physiology. Transportation Research Part
F: Traffic Psychology and Behaviour 10(6): 458475.
Munton, A.G., J. Silvester, et al. (1999). Attributions in
Action: A Practical Approach to Coding Qualitative Data,
John Wiley & Sons Ltd.
Musselwhite, C. (2006). Attitudes towards vehicle driving behaviour: Categorising and contextualising risk.
Accident Analysis & Prevention 38(2): 324334.
Ntnen, R. and H. Summala (1974). A model for the role of
motivational factors in drivers decision-making. Accident
Analysis & Prevention 6(34): 243261.
Newnam, S., B. Watson, et al. (2004). Factors predicting intentions to speed in a work and personal vehicle.
Transportation Research Part F: Traffic Psychology and
Behaviour 7(45): 287300.
Sabey, B.E. and H. Taylor (1980). The known risks we run:
the highway. Crowthorne, TRL Ltd.
Silcock, D., K. Smith, et al. (2000). What limits speed? Factors that affect how fast we drive, AA Foundation for Road
Safety Research.
Stephens, A.N. and J.A. Groeger (2006). Do emotional
appraisals of traffic situations influence driver behaviour?
Behavioural Research in Road Safety: Sixteenth Seminar,
Department for Transport.
Stradling, S.G. (2007). Car driver speed choice in Scotland.
Ergonomics 50(8): 11961208.
Summala, H. (1988). Risk control is not risk adjustment: The
zero-risk theory of driver behaviour and its implications,
United Kingdom: Taylor & Francis.
Summala, H. (1996). Accident risk and driver behaviour.
Safety Science 22(13): 103117.
Summala, H. (2007). Towards understanding motivational
and emotional factors in driver behaviour: comfort
through satisficing. Modelling Driver Behaviour in Automotive EnvironmentsCritical Issues in Driver Interactions with Intelligent Transport Systems. P.C. Cacciabue,
Springer Verlag: 189207.
Underwood, G., P. Chapman, et al. (1999). Anger while driving. Transportation Research Part F: Traffic Psychology
and Behaviour 2(1): 5568.
van der Hulst, M., T. Meijman, et al. (2001). Maintaining task
set under fatigue: a study of time-on-task effects in simulated driving. Transportation Research Part F: Traffic
Psychology and Behaviour 4(2): 103118.

304

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The key role of expert judgment in CO2 underground storage projects

C. Vivalda
Schlumberger Carbon Services, Clamart, France

L. Jammes
Schlumberger Carbon Services, Paris La Dfense, France

ABSTRACT: Carbon Capture and Storage (CCS) is a promising technology to help mitigate climate change
by way of reducing atmospheric greenhouse gas emissions. CCS involves capturing carbon dioxide (CO2 ) from
large industrial or energy-related sources, transporting, and injecting it into the subsurface for long-term storage.
To complete the limited knowledge about the site and experience about the operational and long-term behavior
of injected CO2 , a massive involvement of expert judgment is necessary when selecting the most appropriate
site, deciding the initial characterization needs and the related measurements, interpreting the results, identifying
the potential risk pathways, building risk scenarios, estimating event occurrence probabilities and severity of
consequences, and assessing and benchmarking simulation tools. The paper sets the basis for the development
of an approach suited for CCS applications and its role in the overall framework of CO2 long-term storage
performance management and risk control. The work is carried out in the frame of an internal research and
development project.

INTRODUCTION

Carbon Capture and Storage (CCS) is a promising

technology to help mitigate climate change by way of
reducing atmospheric greenhouse gas emissions. CCS
involves capturing carbon dioxide (CO2 ) from large
industrial or energy-related sources, transporting, and
injecting it into the subsurface for long term storage.
At the current state of the art, the behaviour of
CO2 and injection-related effects are still not fully
understood (IPCC, 2005). Extensive research activity is underway aiming at bridging these knowledge
gaps, for instance characterizing the short and longterm chemical interactions between CO2 , formation
fluids, and the rock matrix, or the mechanical effects
due to increasing the formation pressure (Berard & al.,
2007). Research and demonstration projects, along
with more field data and history matching, will provide
more accurate characterization of these physical processes and will permit continuous updates and tuning
of initial evaluations.
Natural analogs, know-how from other industrial
sectors, and the experience gained in pilot/demonstration projects where data gathering and measurement
campaigns are being carried out, are the references
today to gain better understanding of the behavior of
the injected CO2 at a storage site.
To complete the current limited knowledge of
the site and of the behaviour of CO2 underground,

a massive involvement of expert judgment is necessary when selecting the most appropriate site, deciding
the initial characterization needs and related measurements, interpreting results, identifying potential risk
pathways, building risk scenarios, estimating event
occurrence probabilities and severity of consequences,
and assessing and benchmarking simulation tools.
The study of CO2 injection and fate currently has
very few technical experts, and there is a need to capitalize on the expertise from other domains, such as
oil and gas, nuclear waste disposal, chemical, etc. to
obtain relevant and valuable judgments.
Formalized methods able to infer the searched
information and data from individual expertise in a
well-defined field, and to analyze and combine the
different experts opinions in sound, consistent final
judgments seems the best way to achieve this.
Several bibliographic references concerning expert
judgment elicitation processes exist in literature
(Simola & al., 2005), as well as practical applications in some industrial sectors, such as oil and gas,
nuclear, aerospace, etc. The approaches presented,
quite generic and in principle almost independent from
the industrial area, represent a valuable background
for the development of expert judgment elicitation
processes dedicated to CO2 storage studies.
These general methodologies need to be made more
specific to be applied in the domain of CO2 geological storage, whenever relevant data are lacking or

305

unobtainable, while issues are very serious and/or very

complex.
The paper sets the basis for the development of
an approach suited for CCS applications and its role
in the overall framework of CO2 long-term storage
performance management and risk control. The work
is carried out in the frame of an internal research
and development project and examples of potential
applications are presented.
2

and documented process should add value to the

current practice and provide justified assumptions and
results for external audit.

CHALLENGES

Using expert judgment to help overcome the insufficient, lack of, or poor, data is challenged by the
following:

STATE OF THE ART

At the time the authors are writing (2008) there are no

published formalized methods for experts judgment
elicitation in the domain of CO2 storage. Nevertheless,
expert judgment is tacitly or explicitly employed in
studies and projects to complement lack of or poor
information and data.
Experts are currently required to assess existing
data related to a site, such as geophysics data, geological reports, petro-physical properties . . . and
decide for further data acquisition and site characterization when a site is screened for suitability to store
CO2 . They are also asked: to estimate physical quantities, such as rock permeability from measurements
and provide degree of confidence or probability density functions of the estimated quantities; to intervene
in the qualitative and semi-quantitative steps of risk
assessment to identify risk pathways and scenarios,
and when not otherwise possible, to assess their probability of occurrence and their potential consequences.
They are expected to make choices among models to
represent the actual CO2 behavior, for example during
injection and storage in the underground formation
for a long time period, and to assess and benchmark
simulation tools.
CCS projects in Australia (Bowden & Rigg, 2004)
are an example. Experts have been involved in the
process of selecting the most suitable storage sites.
The expert panel was formed by individuals carrying the relevant competencies, such as geomechanics,
geochemistry, geology, project management, etc. and
led by normative experts. A pool of experts has also
filled and is maintaining a risk register for one selected
site (Otway). Weyburn project in Canada is another
example (Whittaker, 2004). In this project experts
have been questioned to identify the list of failures,
events and processes (FEPs) relevant to the project
(Savage & al., 2004) and to build risk scenarios. Other
examples, such as Sleipner in Norway or Illinois in
United States can also be mentioned.
In all cases, no formal process was followed but the
quality of the results was assured by the excellence and
common sense of the people in the panels.
Notwithstanding the importance of these first references and applications, it seems that a well established

The experts are questioned on a domain that is being

discovered;
CCS experts are contributing to the discovery of the
domain and can be biased by their own understanding;
There are no available statistics (e.g. on unwanted
events such as leakage from the formation to the
atmosphere; on fault and fracture activation; etc.).
However new data from research and demonstration
projects are recorded: they have to be continually
mined to update first subjective estimations;
Experimental results cover a limited number of specific phenomena as, for example, wellbore cement
degradation (Kutchko & al., 2007; Barlet-Goudard
& al., 2007). Extrapolations to the overall site
behavior are not yet viable and their limited number
as well as confidentiality makes it difficult to build
statistics. In spite of that, experts should be able
to benefit from this information when providing
judgments;
Numerical modeling of physical phenomena is
underway and is rapidly progressing, despite the
limited feedback from the field necessary to calibrate and history match the models. This implies
that current simulation models should consider
the uncertainty of the input parameters and make
explicit the degree of confidence of the results. The
role of experts to assess the models as well as to
provide estimations of input parameters is crucial;
Generic risk pathways have been imagined, such
as CO2 escape into potable aquifers due to wellbore failure, vertical migration and leakage through
abandoned or closed wells in the vicinity (Benson
& al., 2002) but not feed-backed and quantified by
experience. Starting from these generic pathways,
experts may be required to build site specific scenarios and assess their probability of occurrence and
the magnitude of their consequences;
Each site is unique (not reproducible). Knowledge about the structure and the properties of the
geological system is usually obtained by interpreting existing relevant geophysical data such as
seismic, and measuring underground strata characteristicsmechanical, petro-physical, fluid flow
properties, chemical composition. . . Experts will

306

always be required to interpret and assess site specific data. Extrapolations to or from other sites
are normally risky and shall be appropriately documented.
These challenges influence the way the formal
process for expert judgment elicitation is built.

EXPERTS JUDGMENT ELICITATION

PROCESS

The internal project is currently in the phase of setting the requirements for the development of an expert
judgment elicitation process able to:
Address the challenges described above;
Capture the most relevant information and data from
the experts;
Be well balanced between time and costs and quality
of captured information;
Address specific needs and objectives of expert
judgment use;
Be reproducible.
The formal process will be based on existing practices and methods (Cooke & al., 1999; Bonano
& al., 1989), which will then be customized to be
implemented on CO2 storage projects.
While building the process, it is important to define
the domains it will be applied to, and ensure it covers
all of them.
The following domains are identified:
Measurements interpretation (e.g. CO2 plume size,
mass in place, etc.);
Physical parameters estimation (e.g. permeability,
fault/fracture extent, etc.);
Risk related parameters estimation (e.g. scenario
identification, likelihood and severity estimation,
etc.);
Conceptual and mathematical models selection and
validation (e.g. CO2 transport model, mechanical
model, etc.).
All along these domains, the experts are also
required to qualify or quantify uncertainty, e.g. in
qualitative assessment, in numerical values for key
parameters, in predictions, etc.
The process will be deployed into the following four
main steps.
4.1

Experts selection

There is a need to find a balanced way to select the

experts knowing that the expertise on CO2 geological storage is being built in parallel to the increase
in knowledge. Three main external areas can be
mentioned where useful know-how exists.

The first is represented by natural and industrial

analogs. They have the advantage of presenting some
characteristics similar to CO2 storage, such as leakage/seepage mechanisms for natural analogs, long
term repository for wastes, gas injection into formations for gas storage; but also important differences,
such as CO2 already in place for natural analogs, short
storage period (few months to years) and reversible
process for gas storage sites, and dangerous substances
for waste repositories. The second area is represented
by other industrial applications, such nuclear, aeronautics, chemical, oil and gas. . . The experts in these
two areas need accurate training on the system characteristics, performance, and challenges to avoid underestimation of the differences, such as peculiar CO2
physical and chemical properties, long-term storage
requirements, etc. and draw biased conclusions.
The third area is represented by expertise outside
industrial applications, such as policy and regulations,
insurance, sociology, environmental protection. . .
This contribution is very valuable for example when
assessing the consequences of a given scenario and
their impact on population, the environment, etc. In
this case the experts need a full understanding of the
overall system and its challenges, spanning from the
CO2 storage site to the legal framework, the local
population habits, the flora and the fauna, etc.
On top of these three areas, a valuable set of experts
is found among the CCS practitioners, i.e. those who
have or are participating in CCS research, demonstration and future large-scale projects. Their know-how
and feedback is very precious.
Depending on the objectives of the elicitation process, a balanced mix of the above expertise needs to
be reached. The presence of experts outside the CCS
community already accustomed to giving judgments
should positively contribute to keep the approach upto-date and smooth the way towards implementation
of the elicitation process in this new field.
Following the classical approach for expert judgment elicitation (Bonano, 1989) three main kinds of
experts will be selected:
Generalists: knowledgeable about various aspects
of the storage site and its performance;
Specialists: at the forefront of specialties relevant to
the storage site;
Normative experts: trained in probability theory,
psychology and decision analysis. Assist generalists and specialists with substantive knowledge in
articulating their professional judgments.
Typical specialties necessary to create a panel
for CO2 storage projects is geology, geomechanics,
geochemistry, geophysics, petro-physics, monitoring
techniques and instrumentations, materials, etc.
The size of the expert panel varies according to the
domain and objects to be elicited.

307

4.2

Experts elicitation instruments and modes

The mode of carrying out the experts elicitation

depends on the elicitation purposes, the domains it
addresses, the level of general knowledge about the
objects to be questioned, and the available expertise
relevant to the object to be questioned.
It is envisaged to use at least two approaches to
elicit knowledge. The first one is by setting up expert
panels, to have all the experts gathered together at least
once and encourage discussion and confrontation. The
second one is by survey, where the experts judgment is
elicited independently but not confronted. The results
will then be analyzed by the normative expert(s).
4.3

Experts judgments analysis and combination

The judgments provided by the experts will then be

analyzed and combined.
Methods to combine expert judgments exist in literature. They vary from very simplistic methods, based
on averaged values of experts estimations, giving
each expert the same weight, to more complicate and
informative (Gossens & al., 1998), where the experts
are weighted and their judgments probabilistically
aggregated.
The choice of the most appropriate method to combine the different judgments will vary depending upon
the final use of the estimations. Three examples are
presented below.
The first example considers a pool of experts gathered at the initial phase of a project to compile a risk
register. Once the list of hazards, their potential causes,
and consequences are identified and recorded, elicited
outcomes will be the likelihood and severity of the
identified hazard. These two parameters, at this stage
of the project, are poorly known and reference statistics
do not exist, so they will be classified using indexes
(e.g. from 1 to 5) running for example from improbable to probable for likelihood and from negligible to
catastrophic for severity. Each expert is asked to index
likelihood (IL ) and severity (IS ) of each hazard and
provide a credal interval (e.g. [IL 1; IL + 1]). The
judgments are then combined on the indexes and, due
to the broad level of the analysis, a simple combination method, like averaging the individual indexes, will
be sufficient to obtain an integer mean index Im . The
aggregated credal intervals can also be evaluated.
The second example concerns a project where an
existing well is being evaluated for potential re-use as
an injection well (Grard & al., 2006; Van Der Beken,
A. & al., 2007), and where cement evaluation logs are
available. In this case, the permeability of the cement is
one of the most significant factors to drive the choice
of re-using the well, because the probability of having CO2 leaking through the cement depends on it.
In this case, more precise judgments are required and

experts can be asked to estimate the physical quantity,

and provide the degree of confidence or the probability density functions of this parameter in different
sections of the well. More refined methods, such as
the Classical Model suggested by Cooke & al. (1999)
are better suited to deal with this case to find a rational
consensus on the final distributions.
The third example addresses the estimation of the
probability of a fault opening under certain pressure
conditions, when data from site seismic measurements
are very poor and mechanical properties uncertain. In
this case, the experts may give very different judgments and seeking consensus is less informative than
estimating the degree of variability among experts
best estimates. This allows getting a measure of the
overall uncertainty around that variable and then deciding to take actions, such as asking for additional
measurement campaigns, to reduce the uncertainty
bounds.
4.3.1 Uncertainty measures
Uncertainty associated to the judgments has to be measured and considered when the results of the elicitation
process are analyzed. Efficient measures to look at are
the following (Hoffmann & al., 2006):
Extent of agreement between experts and external attributions (e.g. analogues): use of relevant
analyses (e.g. regression) to explain differences;
Degree of variability among experts best estimates:
statistical measure of consensus of expert opinion;
Individual uncertainty: measure of the certainty
experts have about their own attribution judgment.
The uncertainty is described by probability distributions, credal intervals, when numerically quantifiable,
or by qualitative statements, such as large, small . . .
when not. The treatment of quantified uncertainties
is done using traditional probabilistic and statistical
methods.
4.4 Updating initial estimations
An important feature of a geological site is that
its description becomes more and more accurate by
acquiring new data, by monitoring, and by running calibrated simulation models. This implies that some of
the objects initially evaluated by experts are now measured or calculated, and also that some judgments have
to be updated because of more available information.
A positive outcome is that the uncertainty surrounding
the overall CO2 storage system is narrowed.
Updating sessions are therefore necessary and valuable. They may require the involvement of only one or
a few experts to inform the system of new consolidated
data. Or they may want a pool of experts to re-run an
elicitation process that will consider all the new information at hand. The composition of the pool may be

308

A methodology for performance management and risk

control of CO2 storage sites and related models and
tools are being developed within Carbon Services, and
it includes a formal process to elicit experts judgment and extract knowledge in terms of qualitative or
quantitative appraisals (Vivalda & al., 2008).
The core of the methodology is the identification
of the possible short-, mid-, long-term losses in system performance and their description in terms of
reference risk pathways, relevant scenarios and their
probabilistic evaluation. A risk pathway is defined as
the mechanism of exposure of a receptorpotable
aquifers, soils, living areasto a stressorCO2
leakage and related events. A scenario is defined
as hypothetical sequences of events constructed for
the purpose of focusing attention on causal process.
Figure 2 shows an example of CO2 storage site and
a leakage pathway, i.e. CO2 vertical migration from
a deep saline formation to a potable water aquifer
through an abandoned well.
Experts are expected to lead the identification of
the risk pathways and to build representative scenarios. In addition, for each representative risk pathway/scenario, the likelihood/probability of occurrence
and its impact (i.e. the severity of its consequences)
are sometimes estimated using expert judgment. This
occurs quite often when the project is in its early
phases, and available data and statistics are very poor.
Another step where the experts are required is the
identification and the selection of mitigation measures
(preventive or corrective) when there is the necessity
to lower or control some risks. Selecting the measures
is a decision making process where the advantages of
the introduction of the measure should be balanced
with its effectiveness.

Figure 1. Excerpt of decision diagram for structured expert

elicitation process selection.

the same as at the former elicitation or may include

different experts. It can even be modified in size and
types of expertise when deemed necessary.
4.5

EXPERT JUDGMENT AND CO2 STORAGE

PERFORMANCE MANAGEMENT

Structure

The process for expert elicitation is structured starting

from an inventory of available practices and methods,
properly combined in a decision diagram to give indication on how and when they can be used with respect
to the site life cycle and the corresponding path to
follow.
The purpose of structuring the process in this way is
to give flexibility to each specific project to select the
most appropriate path according to the degree of detail,
precision expected by the elicitation process and the
project data available at the moment of the assessment.
An excerpt of decision diagram is provided in Figure 1.
The decision diagram is deployed all along the relevant phases of CO2 storage site life cycle, which are the
following: a) site selection, b) site screening, c) characterization, d) design, e) injection, and f ) long-term
storage/surveillance.

CO 2

Monitoring
well

Potable Water
Injector
well

Abandoned
well

Fault
Fractures

Caprock
Deep Saline Formation

Figure 2.

309

Example of CO2 storage site and risk pathway.

The overall approach for CO2 performance

management and risk control considers uncertainties,
including those introduced by the judgments.
5.1

Expert judgment and modeling

Modeling is the core of the CO2 storage site performance management and risk control methodology.
Its main purpose is to capture the main features of
the overall site and simulate the most important processes induced by CO2 injection, to represent the
reality in the most truthful way. Simulation models are used, for example, to quantitatively evaluate
the risk pathways/scenarios. To take into account the
limited knowledge of today, the models have to be
probabilistic and include uncertainty analysis. Experts
are involved in the process of selecting the most
appropriate and representative models.
The quality of the results depends on the quality of
the simulation models and tools, and the role of experts
in benchmarking the tools is significant. The decision
diagram for structured expert elicitation process selection, as it is conceived today, does not address methods
for model assessment because of a different nature and
a dedicated approach will be investigated.

EXAMPLES OF POTENTIAL
APPLICATIONS

A few examples of potential application of the process

during the CO2 storage site life cycle are given below.
6.1 Site selection
Site selection corresponds to the phase were the sites
suitable for CO2 storage are selected among a number of candidates against a list of comparison criteria.
During this phase, the experts assess the available data,
and they are explicitly questioned about the weights of
comparison criteria.
6.2 Site screening
Site screening corresponds to the phase where the site
is described on the basis of the available information.
The experts interpret existing data and make an initial understanding of the earth model. As the domain
is new, the skill of domain experts has to be developed and it seems advisable that, instead of a single
expert, a panel of domain experts is formed and their
assessments aggregated. The experts semi quantitatively assess the likelihood and the severity of the
hazards identified through e.g. a tabular process (e.g.
Risk register, . . . ) as typical of the initial performance
analysis.

6.3 Site characterization

Site characterization corresponds to the phase where
additional measurements are made following the
directions given during the site screening phase on
the base the preliminary performance assessment. In
this phase experts interpret measurements. To carry
out a detailed performance assessment, they identify
risk pathways and scenarios. Experts are questioned
about the probability of occurrence and magnitude of
consequences of risk pathways/scenarios.
During this phase a static model can be built and
dynamic simulations run to quantitatively assess the
risk pathways or the scenarios. The involvement of
a group of experts to create the model is crucial to
take maximum profit of the various expertises and to
solve conflicts when modeling choices are made, such
as the size of the 3D area to be described, the inclusion
of surface properties, the inclusion of distant wells, etc.
6.4 Long-term storage
Long-term storage corresponds to the phase where
the stored CO2 is expected to stay in place and the
site is monitored to confirm this expectation. Experts
interpret monitoring data. Most of the interpretations
are derived from indirect measurements and need to
be understood. Experts update predictions as far as
pathway/scenarios are concerned.

CONCLUSIONS

The paper underlines the main features of the use

of expert judgment within CO2 geological storage
projects. It shows the need of defining a justified and
accepted method able to structure experts judgment
use and related elicitation processes and to combine the
outcome to reach objective appraisals even if uncertain. Such a method does not exist today in the CCS
domain.
The work carried out within the project at the basis
of this paper aims at bridging this gap by the definition of a structured approach able to capture expert
knowledge, elicit judgments, treat uncertainties and
provide sound estimation when there is a lack of field
and operational data.

REFERENCES
Barlet-Goudard, V. Rimmel, G. Goff, B. & Porcherie, O.
2007. Well Technologies for CO2 Geological Storage:
CO2 -Resistant Cement. Oil & Gas Science and TechnologyRev. IFP, Vol. 62, No. 3, pp. 325334.
Benson, S.M. Hepple, R. Apps, J. Tsang, C.-F. &
Lippmann, M. 2002. Lessons Learned from Natural

310

and Industrial Analogues for Storage of Carbon Dioxide in Deep Geological Formations. Lawrence Berkeley
National Laboratory LBNL-51170.
Brard, T. Jammes, L. Lecampion, B. Vivalda, C. &
Desroches, J. 2007. CO2 Storage Geomechanics for Performance and Risk Management, SPE paper 108528,
Offshore Europe 2007, Aberdeen, Scotland.
Bonano, E.J. Hora, S.C. Keeny, R.L. & von Winterfeldt, D.
1989. Elicitation and use of expert judgment in performance assessment for high-level radioactive waste repositories. NUREG/CR-5411; SAND89-1821. Washington:
US Nuclear Regulatory Commission.
Bowden A.R. & Rigg, A. 2004. Assessing risk in CO2
storage projects. APPEA Journal, pp. 677702.
Cooke, R.M. & Gossens, L.J.H. 1999. Procedures guide
for structured expert judgment. EUR 18820. Brussels,
Euratom.
Grard, B. Frenette, R. Auge, L. Barlet-Goudard, V.
Desroches J. & Jammes, L. 2006. Well Integrity in CO2
environments: Performance & Risk, Technologies. CO2
Site Characterization Symposium, Berkeley, California.
Gossens, L.H.J. Cooke, R.M. & Kraan, B.C.P. 1998. Evaluation of weighting schemes for expert judgment studies.
In Mosleh, A. & Bari, A. (Eds.) Probabilistic safety
assessment and management. Springler, Vol. 3, pp.
19371942.
Helton. J.C. 1994. Treatment of Uncertainty in Performance
Assessments for Complex Systems. Risk Analysis, Vol.
14, No. 4, pp. 483511.
Hoffman, S. Fischbeck, P. Krupnik, A. & McWilliams, M.
2006. Eliciting information on uncertainty from heterogeneous expert panels. Discussion Paper. Research for the
future. RFF DP 06-17.

IPCC, 2005. IPCC Special Report on Carbon Dioxide

Capture and Storage. Intergovernmental Panel on Climate
Change, Cambridge University Press, Cambridge, UK.
Kutchko, B.G. Strazisar, B.R. Dzombak, D.A. Lowry, G.V. &
Thaulow, N. 2007. Degradation of well cement by CO2
under geologic sequestration conditions. Environmental
Science & Technology, Vol. 41, No. 13, pp. 47874792.
Oberkampf, W.L., Helton, J.C., Joslyn, C.A.,
Wojtkiewicz, S.F. & Ferson, S. 2004. Challenge Problems: Uncertainty in System Response Given Uncertain
Parameters. Reliability Engineering and System Safety,
Vol. 85, pp. 1119.
Savage, D., Maul, P.R., Benbow, S. & Walke, R.C. 2004.
A generic FEP database for the assessment of long-term
Performance and Safety of the geological storage of CO2 .
Quintessa, QRS-1060A-1.
Simola, K., Mengolini, A. & Bolado-Lavin, R. 2005. Formal expert Judgment an overview. EUR 21772, DG JRC,
Institute for Energy, The Netherlands.
Van Der Beken, A., Le Gouvec, J., Grard, B. & Youssef, S.
2007. Well Integrity Assessment and Modelling for CO2
injection. In Proceedings of WEC07, Alger, Algeria.
Vivalda, C. & Jammes, L. 2008. Probabilistic Performance
Assessment Methodology for long term subsurface CO2
storage. In Proceedings of PSAM 9 Conference, Hong
Kong.
Whittaker, S.G. 2004. Investigating geological storage of
greenhouse gases in southeastern Saskatchewan: The IEA
Weyburn CO2 Monitoring and Storage Project. In Summary of Investigations, Misc. Rep. 2004-4.1, Vol. 1, Paper
A-2, Saskatchewan Geological Survey, Sask. Industry
Resources, Canada.

311

Integrated risk management and riskinformed decision-making

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

All-hazards risk frameworkAn architecture model

Simona Verga
DRDC CSS (Centre for Security Science), Ottawa, Canada

ABSTRACT: This work advocates an architecture approach for an all-hazards risk model, in support of
harmonized planning across levels of government and different organizations. As a basis for the architecture, a taxonomic scheme has been drafted, which partitions risk into logical categories and captures the
relationships among them. Provided that the classifications are aligned with the areas of expertise of various departments/agencies, a framework can be developed and used to assign portions of the risk domain to those
organizations with relevant authority. Such a framework will provide a structured hierarchy where data collection
and analysis can be carried out independently at different levels, allowing each contributing system/organization
meet its internal needs, but also those of the overarching framework into which it is set. In the end, the proposed
taxonomy will provide a blueprint for the all-hazards risk domain, to organize and harmonize seemingly
different risks and allow a comparative analysis.

INTRODUCTION

Terrorism has become a major concern in recent times,

but the potential severity of health and natural disasters
has been demonstrated in Canada, as well as worldwide, throughout history in one form or another. While
vaccination campaigns and public education have led
to the eradication of many of the past killer diseases,
and while science and engineering advances have led
to more resilient human settlements and infrastructures, at least in our part of the world, we are far
from invulnerable. The emergence of new diseases and
the possibility of new pandemics; the exacerbation of
weather events by global warming; unforeseen effects
of emerging technologies; the asymmetric distribution
of wealth and resources leading to increased demographic pressures; the increased complexity of our
infrastructures stressing control mechanisms closer
and closer to the breaking point; and, not in the least,
the greater and greater expectation of protection by
the population from its government; these are examples that underscore the need for a robust, all-hazards
approach to public safety and security.
In Canada, all levels of Governmentfederal, provincial/territorial and municipalshare the
responsibility to protect Canadians and Canadian society. Within each jurisdiction, the governments public
safety and security functions are shared among many
departments and agencies. Hence, preparedness at the
national level depends on synchronized efforts among
many partners. For practical reasons, any attempt to
formalize an overarching risk model needs to respect
existing structures for ownership and responsibility in

managing risks. As an example, in Canada, public

safety and security functions are shared horizontally,
within one jurisdiction, among several departments
and agencies, and they cross jurisdictional boundaries based on severity of consequences of the realized
risk event. This adds to the complexity of planning
and managing even single emergencies that escalate
across jurisdictions and organizational boundaries.
The added challenge of an all-hazards planning comes
from the lack of a coherent picture on the relative severity of the risks associated with various threats/hazards.
Taking such an approach is, nonetheless, important.
2

UNDERSTANDING RISK

Risk is an intellectual construct, contingent on the

belief that human intervention can influence the outcome of future events, as long as an effort is made to
anticipate these events. Oftentimes risk has a negative
connotation and is associated with the possibility of
future harm or loss (Rowe 1977).
2.1

Riskthe what, why and how

In simple terms, risks are about events that, when triggered, cause problems. Because risks refer to potential
problems in the future, often there is a great deal of
uncertainty with regard to how and to what degree
such events may be realized. If an organizations interests are potentially affected, processes are set up to
manage the uncertainty and strategies are developed
to minimize the effect of such future events on desired

315

Figure 1. Generic risk management process, showing a typical sequence of steps commonly followed, as well as the review
and feedback processes.

outcomes. Risk management is the structured approach

to set up such processes and develop such strategies
(Haimes 2004).
Although the details of the risk management
approach vary widely across risk domains/organizations, the following steps are commonly followed: (1)
risk identification; (2) risk assessment; (3) identification/analysis and implementation of risk treatment
options; and (4) monitoring and evaluation. In most
cases, risk management is a cycle that incorporates
review and feedback loops. Figure 1 shows a generic
risk management process.
Invariably, risk management starts with risk identification. An important part of risk identification is
establishing the context. This implies selecting the
domain of interest, establishing the identity and objectives of risk managers and other interested parties
(stakeholders), the scope of the process, and the basis
upon which risks will be evaluated (i.e., assumptions,
constraints). After establishing the context, the next
step in the process is to identify potential risks. Risk
identification may be approached from different ends:
it may start with identifying potential sources of the
risk event, or with the risk event itself. More will be
said on this essential step in risk management in subsequent sections, as risk identification and classification
constitutes the main topic of this paper.
Once a risk has been identified, it must be assessed.
Risk is the product of Likelihood and Consequences,
where likelihood refers to the risk events chance of
occurrence and Consequences measures the severity and extent of the effects associated with the event.
Indeed, any risk assessment methodology is likely to

start with a risk equation along these lines. But

while this formula is sufficiently general to grant
acceptance with a broad audience of risk professionals, it must be refined further in order to be of
practical value to anyone tasked with a specific assessment. The major difficulty facing an all-hazards
methodology is defining and finding good measures
for Likelihood and Consequences consistently
across risks of very different nature.
2.2

Classification of risksvarious perspectives

Risks must be identified and described in an

understandable way before they can be analyzed and
managed properly. Risk identification should be an
organized, thorough approach to seek out probable
or realistic risks. Among existing methods, the use
of taxonomies meets and promotes the objectives
listed before. Establishing appropriate risk categories
provides a mechanism for collecting and organizing
risks as well as ensuring appropriate scrutiny and
attention for those risks that can have more serious
consequences.
Taxonomies are meant to classify phenomena with
the aim of maximizing the differences among groups.
The term taxonomy refers to the theory and practice
of producing classification schemes. Thus, constructing a classification is a process with rules on how to
form and represent groups (Greek word: taxisorder),
which are then named (Greek word: nomoslaw, or
science). Taxonomies are useful, if they are able to
reduce the complexity of the domain studied into more
tractable macro-classes (Coccia 2007).

316

The design of taxonomies can be a useful first

step in theory building. Classification as an output
deals with how groups and classes of entities are
arranged, according to the taxonomic approach used.
The numerous paradigms advanced by specialists in
various risk fields may be distilled into a number of
general perspectives, or facets, on how to structure
the risk domain. Each of these perspectives offers certain advantages for bringing forward a particular facet
of a risk event. The multi-faceted approach is also
respectful of the inevitable dose of ambiguity regarding the classification criteria, which may be more
significant for certain classes of risks than for others.
The blurring of traditional categories related to environmental and technical risk events is a case in point
(Cohen 1996).
The above considerations illustrate the fact that the
construction of taxonomy inevitably confronts limitations and requires execution of somewhat arbitrary
decisions. Nonetheless, even an imperfect structured
approach is preferred, if it enables a comparative
analysis of broad risk categories, instead of punctual
treatment of an arbitrary set of risks in isolation.
There are several ways of classifying risks.
From a time evolution perspective, there are two
main sequences of events when talking about risks.
The first type of event sequence is a sudden occurrence
that brings immediate consequences. Examples are
earthquake, structural collapse, or terrorist attack. The
second type of event sequence happens gradually, and
the consequences may become apparent after a long
period of time. Examples here are the risk to human
health posed by the agricultural use of pesticides, or
that of emerging nanotechnologies.

Figure 2.

Impact taxonomy.

Another approach is to slice risk with respect to

the domain on which the risk event has an impact, as
illustrated in Figure 2: (1) the natural environment:
earthquake, flood, storm; (2) the built environment:
structures, transportation, IT networks; and (3) the
social environment: people, communities. We will
come back to this view in a later section, when we
discuss building methods for assessing consequences
of risk events.
Last but not least, the classification of risks based on
the nature of risk sources provides a basis to systematically examine changing situations over time. The next
section presents a risk taxonomy based on the source
of the risk event. The classification exercise provides
a powerful way to visualize and understand a complex
domain.
3

THE ALL-HAZARDS RISK ASSESSMENT

TAXONOMY

This section proposes a taxonomic approach that

breaks down all-hazards risk by trying to identify risk
sources, and discusses the advantages of using this
approach, as well as the shortcomings of the proposed
scheme.
3.1 Structuring and organizing existing
knowledge
One of the main objectives of the AHRA taxonomy exercise is directed primarily at retrieving and
organizing risk-related knowledge that exists diffusely
throughout various organizations and levels of the
Canadian government, focusing here on the federal
level. This knowledge constitutes the basis on which a
harmonized methodology for assessing national risks
will be built. Access to the most updated and relevant
information across risk domains under government
responsibility represents an essential enabler for risk
assessment methodology development.
As suggested in (Lambe 2007), an effective taxonomy has three key attributes: it provides a classification scheme; it is semantic; and it can be used as a map
to navigate the domain. The proposed scheme has been
designed with those key attributes in mind.
First of all, the proposed scheme classifies risks
by grouping related risk events together, in categories
and subcategories structured in a way that reveals the
nature of the underlying relationships. The organization of classes is respectful of multiple types of
relationships; for example, the two major classes of
risk events are separated based on whether malicious
intent plays a role in the realization of any given risk
within the class.
Within each major category, the subcategories are
decided based on similarity of attributes, although

317

Figure 3.

The all-hazards risk taxonomy.

the classification principle is clearer for some than

for others (e.g., weather-related events and geological
event are clear examples of Natural Disasters, while
ecological disaster may legitimately fall under both
Natural Disaster and Unintentional events resulting
from human action).
Second, the AHRA taxonomy is semantic; it functions within a controlled vocabulary that describes the
content of the whole domain and each class and subclass; the agreed definitions for key risk terms are
documented in the AHRA lexicon (Verga 2007).
To ensure the success of the taxonomy building
exercise, the team of analysts at the Centre for Security Science has undertaken a collection of evidence
from stakeholdersCanadian federal organizations
with a public safety and/or national security risk mandatewith a focus on establishing common ground,
mapping activities to risk categories and uncovering
critical gaps.
Figure 3 shows the most current version of the
AHRA Risk Event Taxonomy. It must be noted that the
categories are not fully stabilized and they may evolve
as more parties become engaged and provide input; in
fact, as the global context evolves continuously (World

Economic Forum 2008), which is most certainly going

to affect local conditions, so do potential sources of
risk and it is desirable that the taxonomy goes through
periodic reviews and updates in order to stay adequate.
3.2

Taxonomy as basis for data collection

Based on the initial categories included in the risk taxonomy, the team of analysts proceeded to building
a Risk Domain Architecture. To that end, a survey was prepared and administered in order to collect
relevant information from various risk communities
within the Canadian federal government. The survey
was structured around three main parts:
1. A standard first part allowed the respondents to
self-identify and communicate their role as risk
practitioners;
2. A second part consisted of a suite of questions
developed to elicit specific information for use
in the risk architecturelargely based on the risk
categories in the taxonomy;
3. The survey ended with a series of open questions
aimed at gaining an overall impression of the stateof

318

Figure 4.

OV-05 operational activity model.

risk assessment at the federal level, also designed

to elicit additional input to further improve the
taxonomy.
The information collected through the survey was
used to construct a systems architecture, which helped
illustrate how people and organizations were involved
in risk activities, the relationships among them, and
how their mandated responsibilities aligned with categories proposed in the risk taxonomy. The survey also
served to identify those participants willing to share
risk-related tools, methods and assessments. More
details about how the survey was administered, how
the data was analyzed and the architecture built can be
found in (Keown 2008). The next section summarizes
the results that are most relevant to the purpose of this
paper.
3.3

Resultsa federal risk community architecture

Using the information captured through the survey,

an architecture was constructed. The U.S. Department of Defense Architecture Framework (DoDAF).
DoDAF represents a standard way to organize a systems architecture and is constructed using a number
of complementary and consistent views, enabling both
an operational and a systems perspective on the built
architecture. A full discussion of DoDAF is beyond the
scope of this paper and can be found in (DoDAF 2004).
The AHRA architecture was built and visualized using
Telelogic System Architect Version 10.6. The diagrams created are too detailed to be viewed clearly, but

a brief description of the results is included, together

with magnified insets to showcase the main point.
Figure 4 shows the AHRA taxonomy translated
into an operational viewOV-05, or Operational
Activity Model. In this model, each risk event in the
taxonomy is represented as an activity that needs
to be completed for an all-hazards risk assessment
to be thoroughly conducted. The figure shows an
enlarged portion, for clarity. It should be noted that
risk events such as Ecological Disasters or Emerging Technologies, which in the AHRA taxonomy
connect to more then one column, are not properly
represented in this Operational Activity Model, since
this type of relationship is not logical in an activity
hierarchy.
In the survey, participants indicated their sections,
the larger organizations they belonged to, and each
of the taxonomic events they were actively reviewing
through a risk assessment. This information was visualized using an Operational Activity Node, or OV-02,
diagram. Activity Nodes represent sections within participating departments and agencies that conduct risk
assessment activities. Figure 5 shows an examplea
node representing the Canadian Ice Service section
within the Department of Environment Canada.
Based on these results, it was possible to return to
the activity hierarchy and indicate which nodes are
responsible for each activity, as shown in Figure 6.
Although not shown, the information collected
through the survey also allowed mapping communications or other types of relationships between
organizations.

319

Finally, based on the survey responses, it was possible to identify risk practitioners who volunteered
to provide information on tools, methods or specific
assessments. Based on the material collected from
these respondents, and an extensive review of the
research literature on the topic, the team of analysts
at CSS hopes to develop a harmonized methodology capable of sustaining an all-hazards scope. The
next section goes back to the AHRA taxonomy and
shows how it can be used to guide methodology
development.

3.4 Taxonomy as a basis for methodology

development
Based on the discussion in the last paragraph of
section 2.1, as well as extensive research on methodology employed within the global risk management
community, the universal, widely-accepted risk
assessment principle is illustrated by the equation:
Risk
Likelihood
magnitude of
=

Magnitude
of occurrence
Consequence
This is the fundamental equation in risk management, and it would be difficult to carry out authentic
risk assessment without basing the overall process on
this relationship. And while the choice of assessment
processes, scaling/calibration schemes, or even the set
of parameters must accommodate specific risks, as it
will be further discussed, this basic equation provides
the common framework required for comparative analysis of different risks, albeit at a fairly high level.
This being said, the next paragraph will discuss how
the taxonomy can be used to understand the different requirements in treating each of the major risk
categories.
The bottom part of Figure 3 showcases the discriminating principle that separates between the two main
classes of risk events: under Malicious Threats,
risk events originate in malicious intent acts of enemy
actors, who can be individuals, groups, or foreign states. The division hints at one major difference in methodological approaches employed in the

Figure 5.

Sample of OV-02 activity node.

Figure 6.

OV-05 operational activity diagram linked to activity nodes.

320

respective domains: adaptability of the sourcethreat

actorin order to avoid risk treatment measures.
For this reason, estimating the likelihood of a malicious risk event requires a different approach and
different metrics than estimating the likelihood of a
non-malicious event.
For the latter type, at least in principle, likelihood is
a more straightforward combination of the frequency
of occurrence for that event type, based by historical data, and vulnerability/exposure of the human,
natural and build environment exposed to the given
hazard/threat. Although variations exist within each of
the three major branches: unintentional (man-made);
natural disasters; and health disasters, the calculation algorithm can be sufficiently consistent for those
subcategories with a clear relationship to the class.
Specialization of Consequence assessment may
also be necessary. However, it is conceivably possible
to establish a common set of consequence parameters,
along with adequate metrics, of which for each risk
category only the relevant parameters will be selected
and included in the assessment. The set of Consequence parameters may be organized around the three
major domains in the Impact Taxonomy illustrated in
Figure 2, which also includes overlap of the domains,
allowing for hybrid parameters.
The rationale for having a common set of Consequence parameters stems from the aim of the AHRA
methodology, which, as stated in earlier sections of
this document, is to enable a comparative analysis
across broad categories of different risks, in order to
inform decision-making at the highest level. The final
output may be assumed to consist of a set of Risk
Magnitude assessmentsrelating to the top risks of
the daywhich will need to be viewed both in absolute magnitude terms and in relative magnitude terms.
In order that the relative assessment of Risk Magnitudeleading to priorities for action and/or allocation
of resourcescan work, there will have to be a common set of parameters for the understanding of the
various consequences.
For example, it may be necessary to prioritize
resources between impending natural disasters on
the one hand and anticipated terrorist activity on
the other. Common ways of understanding the total
consequences for Canada of each Risk on the list
will be required. For decision-making at the highest
level the risk magnitude assessment will probably
need to be supported by high-level information on
Consequences.
Thus, a logical way to proceed with the AHRA
methodology is to develop a number of modules
that reflect variations in the way in which the Risk
Magnitude calculation needs to be performed for the
different classes of risk events. The calculation of
Likelihood could be rather different for Malicious
Threats, for Natural Disasters and for Unintentional

Figure 7.

All-hazards risk assessment modules.

Figure 8. Risk assessment tetrahedrona graphical representation of the common risk equation.

man-made disasters. Also, the calculations of Consequences could be quite different, although this
paper strongly advocates a common approach for Consequence assessment. A modular AHRA would,
however, need to provide commonality in the way
in which the final results are presented to decision
makers. Figure 7 illustrates a possible breakdown in
risk assessment modules, while Figure 8 represents
a graphical representation of risk equation, showing
how the modules can be brought together to provide a
common picture of different assessments.
To end this section, a last look at the AHRA taxonomy is in order. Figure 3 also illustrates a challenge in
finding the right treatment for risks that do not seem
amenable to the same kind of tidy subdivision. The
sub-categories Ecological Disasters and Emerging Technologies remain somewhat ill-defined and
unwieldy. The difficulty with these two sub-categories
originates in one shortcoming of the current scheme:
the classification principle does not consider the time
sequence in the realization of risk events, as discussed
in section 2.2. From a time perspective, these two
groups of risk would naturally fall under the gradual type; many of the other sub-categories belong
to the sudden occurrence category, although some
of the boxes in Figure 3 will break under the new
lens. This last point highlights the challenges in the
ambitious enterprise of tackling all-risk in one battle, particularly the difficulty of bringing the time

321

dimension into the equation. Developing the AHRA

methodology is work in progress in the Risk group at
DRDC Centre for Security Science and this document
is meant to illustrate the problem, the approach and
initial results as well as an awareness of the challenges
associated with this exciting project.
4

CONCLUDING REMARKS

This paper proposes a taxonomic scheme that partitions the All-Hazards Risk Domain into major event
categories based on the nature of the risk sources, and
discusses the advantages of using this approach, as
well as the shortcomings of the proposed scheme. The
taxonomy enables an architecture approach for an allhazards risk model, in support of harmonized planning
across levels of government and different organizations. Provided that the classifications are aligned with
the areas of expertise of various departments/agencies,
a framework can be developed and used to assign portions of the risk domain to those organizations with
relevant authority. Essential actors, who are active
in conducting assessments and/or performing functions that need to be informed by the assessments,
are often invisible from the point of view of authority structures. Such a framework provides a structured
hierarchy where data collection and analysis can be
carried out independently at different levels, allowing each contributing system/organization to meet
its internal needs, but also those of the overarching
framework into which it is set.
Finally, based on the survey responses, it was possible to identify risk practitioners who volunteered
to provide information on tools, methods or specific
assessments. Based on the material collected from
these respondents, and an extensive review of the
research literature on the topic, the team of analysts
at CSS hopes to develop a harmonized methodology
capable of sustaining an all-hazards scope.
The paper also shows how the AHRA taxonomy
can be used to guide methodology development. The
taxonomy can be used to understand the different
requirements in treating distinct risk categories, which
in turn guides the choice of assessment processes,
scaling/calibration schemes, or the set of parameters in order to accommodate specific risks. As a
consequence, a national all-hazards risk assessment
needs a fold-out, modular structure, that reflects and
is able to support the different levels of potential

management decisions. At the same time, the taxonomy pulls together the different components and
provides the common framework required for harmonized assessment and comparative analysis of different
risks, albeit at a fairly high level.
A noteworthy quality of the risks identified in the
taxonomy is that they do not exist, and cannot be
identified and assessed, in isolation. Many are interconnected, not necessarily in a direct, cause-and-effect
relationship, but often indirectly, either through common impacts or mitigation trade-offs. The better the
understanding of interconnectedness, the better one
can design an integrated risk assessment approach and
recommend management options. But this remains
methodologically and conceptually difficult, due to
the inherent complexity of the domain and our limited
ability to represent it adequately.
The above considerations add to the methodological
hurdles around the representation of interconnectedness inter- and intra- risk domains. In addition, one
cannot leave out the global risk context, which is both
more complex and more challenging than ever before,
according to the World Economic Forum. (Global
Risks 2008).
REFERENCES
DoD Architecture Framework Working Group. 2004. DoD
Architecture Framework Version 1.0, Deskbook., USA.
Department of Defence.
Global Risks 2008. A Global Risk Network Report, World
Economic Forum.
Coccia, M. 2007. A new taxonomy of country performance
and risk based on economic and technological indicators.
Journal of Applied Economics, 10(1): 2942.
Cohen, M.J. 1996. Economic dimensions of Environmental and Technological Risk Events: Toward a Tenable
Taxonomy. Organization & Environment, 9(4): 448481.
Haimes, Y.Y. 2004. Risk Modeling, Assessment and Management, John Wiley & Sons.
Keown, M. 2008. Mapping the Federal Community of Risk
Practitioners, DRDC CSS internal report (draft).
Lambe, P. 2007. Organising Knowledge: Taxonomies,
Knowledge and Organisational Effectiveness. Oxford:
Chandos Publishing.
Rowe, W.D. 1977. An anatomy of Risk, John Wiley & Sons.
Verga, S. 2007. Intelligence Experts Group All-Hazards
Risk Assessment Lexicon, DRDC CSS, DRDC-Centre for
Security Science-N-2007-001.

322

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Comparisons and discussion of different integrated risk approaches

R. Steen & T. Aven
University of Stavanger, Norway

ABSTRACT: There exist many discipline oriented perspectives on risk. Broadly categorised we may distinguish between technical, economic, risk perception, social theories and cultural theory. Traditionally, these
perspectives have been viewed to represent different frameworks, and the exchange of ideas and results has been
difficult. In recent years several attempts have been made to integrate these basic perspectives to obtain more
holistic approaches to risk management and risk governance. In this paper we review and discuss some of these
integrated approaches, including the IRGC risk governance framework and the UK Cabinet office approach.
A structure for comparison is suggested, based on the attributes risk concepts and risk handling.

INTRODUCTION

To analyze and manage risk an approach or framework is required, defining what risk is, and guiding
how risk should be assessed and handled. Many
such approaches and frameworks exist. To categorize
these, Renn (1992, 2007) introduce a classification
structure based on disciplines and perspectives. He
distinguishes between:
Statistical analysis (including the actuarial approach), Toxicology, Epidemiology, Probabilistic risk
analysis, Economics of risk, Psychology of risk, Social
theories of risk, Cultural theory of risk.
To solve practical problems, several of these perspectives are required. Consider a risk problem with
a potential for extreme consequences and where the
uncertainties are large. Then we need to see beyond
the results of the probabilistic risk analyses and the
expected net present value calculations. Risk perception and social concerns could be of great importance
for the risk management (governance).
But how should we integrate the various perspectives? In the literature, several attempts have been
made to establish suitable frameworks for meeting
this challenge, integrating two or more of these perspectives. In this paper we restrict attention to four of
these:

The HSE framework. Reducing risks, protecting

people (HSE 2001)
The UK Cabinet office approach (Cabinet office
2002)
The risk governance framework IRGC (Renn 2005)
The consequence-uncertainty framework introduced by Aven (2003, 2007).

The aim of this paper is to compare these integrated perspectives, by looking at their differences and
communalities. The comparison is based on an evaluation of how risk is defined and how risk is handled
(managed).

AN OVERVIEW OF SELECTED
FRAMEWORKS

The main purpose of this section is to summarize the

main ideas of the different integrated risk approaches
mentioned above.
2.1

The HSE framework

The main purpose of this approach is to set out an overall framework for decision making by HSE (Health
and Safety Executive) and explain the basis for HSEs
decisions regarding the degree and form of regulatory
control of risk from occupational hazards. However
the framework is general and can also be used for
other type of applications. This approach consists of
five main stages; characterising the issue, examining
the options available for managing the risks, adopting
a particular course of action, implementing the decisions and evaluating the effectiveness of actions. Risk
is reflected as both the likelihood that some form of
harm may occur and a measure of the consequence.
The framework utilises a three-region approach
(known as the tolerability of risk, TOR), based on
the categories acceptable region, tolerable region and
unacceptable region. Risks in the tolerable region
should be reduced to a level that is as low as reasonably
practicable (ALARP). The ALARP principle implies
what could be referred to as the principle of reversed

323

onus of proof. This implies that the base case is that

all identified risk reduction measures should be implemented, unless it can be demonstrated that there is
gross disproportion between costs and benefits.
2.2

The UK cabinet office approach

The UK Cabinet office approach (Cabinet office 2002)

sets out how government should think about risk, and
practical steps for managing it better. It proposes principles to guide handling and communication of risks
to the public. Risk refers to uncertainty of outcome, of
actions and events, and risk management is about getting the right balance between innovation and change
on the one hand, and avoidance of shocks and crises on
the other. The approach is based on the thesis that the
handling of risk is at heart about judgement. Judgement in the context of government decision making
can, and should, be supported by formal analytical
tools which themselves need enhancing. But these
cannot substitute for the act of judgement itself. The
approach frames how far formal risk analysis can be
usefully enhanced and made systematic, so that there is
greater clarity about where analysis endsand judgement begins. It also explores and suggests what else
we need to do to enhance our handling of risk and
innovation.
2.3

The risk governance framework IRGC

The risk governance framework (Renn 2005) has been

developed to provide structure and guidance on how
to assess and handle risk on the societal level. The
framework integrates scientific, economic, social and
cultural aspects and includes the effective engagement
of stakeholders. The framework is inspired by the conviction that both the factual and the socio-cultural
dimension of risk need to be considered if risk governance is to produce adequate decisions and results.
It comprises five main phases; pre-assessment, risk
appraisal, tolerability and acceptability judgment, risk
management and communication. Risk is defined as
an uncertain consequence of an event or an activity
with respect to something that humans value. The
framework gives importance to contextual aspects
which, either, are directly integrated in a risk management process, or, otherwise, form the basic conditions
for making any risk-related decision. The framework
also introduces a categorisation of risk problems which
is based on the different states of knowledge about each
particular risk.
2.4

The consequence-uncertainty framework

In the consequence-uncertainty framework introduced

by Aven (2003, 2007), risk is defined as the twodimensional combination of events/consequences and

associated uncertainties. Focus is on the events and

consequences (referred to as observables quantities),
such as the number of fatalities and costs, and these are
predicted and assessed using risk assessments. Probabilities and expected values are used to express the
uncertainties, but it is acknowledged that they are
not perfect tools for expressing the uncertainties. To
evaluate the seriousness of risk and conclude on risk
treatment, a broad risk picture needs to be established,
reflecting also aspects such as risk perception and societal concern. The analyses need to be put into a wider
decision-making context, which is referred to as a
management review and judgment process.

COMPARISON OF THE APPROACHES

This section compares the four frameworks, with

respect to the risk perspectives and risk handling.
3.1

The risk concept

We introduce a context and terminology as follows:

We consider an activity, from which events A may
occur leading to consequences C. The occurrence of A
and the consequences C are subject to uncertainties U.
The likelihoods and probabilities associated with the
events and possible consequences are denoted L and P,
respectively. Using these symbols, the risk definitions
can be summarised in the following way:

HSE framework: (L, C)

Cabinet Office framework: U and (L, C)
IRGC framework: C
Consequence-uncertainty framework: (A, C, U).

Likelihood is normally understood as the same as

probability, and in this paper we do not distinguish
between these two concepts. Note however that some
see likelihood as a more qualitative description than
probability which is restricted to a number in the
interval [0, 1].
The Cabinet Office defines risk by uncertainty.
Often the uncertainty is seen in relation to the expected
value, and the variance is used as a measure of risk. As
an example, consider the problem of investing money
in a stock market. Suppose the investor considers two
alternatives, both with expectation 1, and variances
0.16 and 0.08, respectively. As alternative 1 has the
lowest risk (uncertainty), expressed by the variance,
this alternative would normally be chosen. As another
example, consider the number of fatalities in traffic
next year in a specific country. Then the variance is
rather small, as the number of fatalities shows rather
small variations from year to year. Hence according
to this definition of risk, we must conclude that the
risk is small, even though the numbers of fatalities

324

Uncertainty

Activity

Values at stake

Figure 1.
2008a).

Risk

Eventsand
consequences
(outcomes)
Values at stake

Risk defined as a consequence (Aven & Renn

are many thousands each year. Clearly, this definition of risk fails to capture an essential aspect, the
consequence dimension. Uncertainty cannot be isolated from the intensity, size, extension etc. of the
consequences. Take an extreme case where only two
outcomes are possible, 0 and 1, corresponding to 0
and 1 fatality, and the decision alternatives are A and
B, having uncertainty (probability) distributions (0.5,
0.5), and (0.0001, 0.9999), respectively. Hence for
alternative A there is a higher degree of uncertainty
than for alternative B, meaning that risk according to
this definition is higher for alternative A than for B.
However, considering both dimensions, both uncertainty and the consequences, we would of course judge
alternative B to have the highest risk as the negative
outcome 1 is nearly certain to occur.
The IRGC framework defines risk by C. See
Figure 1.
According to this definition, risk expresses a state
of the world independent of our knowledge and perceptions. Referring to risk as an event or a consequence,
we cannot conclude on risk being high or low, or compare options with respect to risk. Compared to standard
terminology in risk research and risk management, it
lead to conceptual difficulties that are incompatible
with the everyday use of risk in most applications, as
discussed by Aven & Renn (2008a) and summarised
in the following.
The consequence of a leakage in a process plant
is a risk according to the IRGC definition. This
consequence may for example be expressed by the
number of fatalities. This consequence is subject to
uncertainties, but the risk concept is restricted to the
consequencethe uncertainties and how people judge
the uncertainties is a different domain. Hence a risk
assessment according to this definition cannot conclude for example that the risk is high or low, or that
option A has a lower or higher risk than option B, as it
makes no sense to speak about a high or higher consequencethe consequence is unknown. Instead the
assessment needs to conclude on the uncertainty or the
probability of the risk being high or higher. We conclude that any judgement about risk needs to take into

account uncertainties/likelihoods, so why not include

this dimension into the risk concept?
Also the (L, C) and (P, C) definitions can be challenged. A probability is not capturing all aspects of
concern. To explain this we need to first introduce
the two common ways of interpreting a probability:
the classical relative frequency interpretation and the
subjective Bayesian interpretation.
According to the classical relative frequency
paradigm, a probability is interpreted as the relative
fraction of times the events occur if the situation analyzed were hypothetically repeated an infinite number of times. The underlying probability is unknown,
and is estimated in the risk analysis. Hence if this interpretation is adopted in the above definitions of risk, we
have to take into account that the risk estimates could
be more or less accurate relative to the underlying true
risk. The uncertainties in the estimates could be very
large, and difficult to express.
The alternative (the Bayesian perspective) considers probability as a measure of uncertainty about
events and outcomes (consequences), seen through the
eyes of the assessor and based on the available background information and knowledge. Probability is a
subjective measure of uncertainty, conditional on the
background information. The reference is a certain
standard such as drawing a ball from an urn. If we
assign a probability of 0.4 for an event A, we compare our uncertainty of A to occur with drawing a red
ball from an urn having 10 balls where 4 are red. True
probabilities do not exist.
However, a probability is not a perfect tool for
this purpose. The assigned probabilities are conditional on a specific background information, and they
could produce poor predictions. Surprises relative
to the assigned probabilities may occur, and by just
addressing probabilities such surprises may be overlooked (Aven 2008a, b).
The consequence-uncertainty definition (A, C, U)
may be rephrased by saying that risk associated with an
activity is to be understood as (Aven and Renn 2008):
Uncertainty about and severity of the consequences
of an activity (I), where severity refers to intensity,
size, extension, and so on, and is with respect to
something that humans value (lives, the environment,
money, etc). Losses and gains, for example expressed
by money or the number of fatalities, are ways of
defining the severity of the consequences.
The main features of the definition are illustrated
in Figure 2.
The uncertainty relates to both the event and the
consequences given that this event occurs.
3.2

Risk handling (management)

The risk handling covers all activities to direct and

control an organisation with regard to risk. It typically

325

Uncertainty

Likelihood increasingly uncertain

Risk
Severity
Events and
consequences
(outcomes)

Activity

Values at stake

emphasis on consequences
eg if serious/irreversible or
need to address societal
concerns

towards
ignorance

rely on past
experience
of generic hazard

consider putative
consequences
and scenarios

conventional
risk assessment

Values at stake

Consequences increasingly uncertain

Figure 2.
and (I).

Illustration of the risk definition (A, C, U)

Figure 3. Procedures for handling uncertainty when assessing risks (HSE 2001).

covers risk assessments, risk avoidance, risk reduction,

risk transfer, risk retention, risk acceptance and risk
communication (ISO 2002).

Identify risks and

define framework
Embed and
review

HSE framework
The main steps of the risk management process are presented in section 2.1. The steps follows to large extent
the standard structure for risk management processes,
see e.g. AS/NZS 4360 (2004). However, the framework has some genuine features on a more detailed
level, of which the following are considered to be of
particular importance:
Weight to be given to the precautionary principle in
the face of scientific uncertainty. The precautionary principle describes the philosophy that should
be adopted for addressing hazards subject to high
scientific uncertainties. According to the framework the precautionary principle should be invoked
where:
a. There is good reason to believe that serious harm
might occur, even if the likelihood of harm is
remote.
b. The uncertainties make it impossible to evaluate the conjectured outcomes with sufficient
confidence.
The tolerability framework (the TOR framework)
and the ALARP principle. This framework is based
on three separate regions (acceptable region, tolerable region and unacceptable region).
An acknowledgment of the weaknesses and limitations of risk assessments and cost-benefit analysis.
These analyses could provide useful decision support, but need to be seen in a wider context, reflecting that that there are factors and aspects to consider
beyond the analysis results. For many type of situations, qualitative assessment could replace detailed
quantitative analysis. In case of large uncertainties,
other principles and instruments are required, for
example the precautionary principle.
The procedures (approaches) for handling uncertainties are illustrated in Figure 3. The vertical axis

Evaluate risks

Assess risks
apetite

Gain assurance
about control
Identify suitable
reponse to risks

Communicating about risk and uncertainty

Figure 4.

Framework for handling risk and uncertainty.

represents increasing uncertainty in the likelihood that

the harmful consequences of a particular event will be
realised, while the horizontal axis represents increasing uncertainty in the consequences attached to the
particular event.
At the lower left corner, a risk assessment can be
undertaken with assumptions whose robustness can be
tested by a variety of methods. However, as one moves
along the axis increasingly assumptions are made that
are precautionary in nature and which cannot be tested
(HSE 2001).
Cabinet Office framework
The main features of the framework are shown in
Figure 4.
The framework is to large extent based on the HSE
framework. For example it is referred to the approach
described by Figure 3, and the TOR framework as
well as the ALARP principle constitutes fundamental building blocks of the Cabinet Office framework.
There are also several original aspects, and we would
like to point at some of these:
The definition of three different levels of decision making: Strategic, programme and operational.

326

Table 1. Risk problem categoryuncertainty induced exampleimplications for risk management (Aven and Renn 2008b,
adapted from Renn (2005)).
Risk problem
category
Uncertainty induced
risk problems

Management strategy

Appropriate instruments

Risk informed and

Caution/Precaution based
(risk agent)

Risk assessments. Broad risk

characterisations, highlighting uncertainties and features like
persistence, ubiquity etc.
Tools include:
Containment
ALARP (as low as reasonably possible)
BACT (best available control technology), etc.

Risk informed.
Robustness and Resilience
focused (risk absorbing system)

Risk assessments. Broad risk characterizations.

Improving capability to cope with surprises
Diversity of means to accomplish desired benefits
Avoiding high vulnerabilities
Allowing for flexible responses
Preparedness for adaptation

At the strategic level decisions involve the formulation of strategic objectives including major external
threats, significant cross-cutting risks, and longer term
threats and opportunities. At the programme level, the
decision-making is about procurement, funding and
establishing projects. And at the project and operational level, decisions will be on technical issues,
managing resources, schedules, providers, partners
and infrastructure. The level on uncertainty (and hence
risk) will decrease as we move from strategic level to
the programme and then operational level.
In addition, the focus is on risk appetite, i.e. the
quantum of risk that you are willing to accept in pursuit
of value. There is a balance to be made between innovation and change on the one hand and, and avoidance
of shocks and crises on the other. Risk management
is often focused on risk reduction, without recognition
of the need for taking risks to add values.
IRGC framework
On a high level the framework is similar to the two
other frameworks presented above. However on a more
detailed level, we find several unique features. One
is related to the distinction between different type
of situations (risk problems) being studied, according to the degree of complexity (SimpleComplex),
Uncertainty and Ambiguity (Aven & Renn 2008b):
Simplicity is characterised by situations and problems
with low complexity, uncertainties and ambiguities.
Complexity refers to the difficulty of identifying
and quantifying causal links between a multitude
of potential causal agents and specific observed
effects.
Uncertainty refers to the difficulty of predicting the
occurrence of events and/or their consequences

based on incomplete or invalid data bases, possible changes of the causal chains and their context
conditions, extrapolation methods when making
inferences from experimental results, modelling
inaccuracies or variations in expert judgments.
Uncertainty may results from an incomplete or inadequate reduction of complexity, and it often leads to
expert dissent about the risk characterisation.
Ambiguity relates to i) the relevance, meaning and
implications of the decision basis; or related to
ii) the values to be protected and the priorities to
be made.
For the different risk problem categories, the IRGC
framework specifies a management strategy, appropriate instruments and stakeholder participation, see
Table 1 which indicates the recommendations for the
category uncertainty.
The consequence-uncertainty framework
The framework follows the same overall structure
as the other frameworks and is characterised by the
following specific features:
It is based on a broad semi-quantitative perspective on risk, in line with the perspective described
in section 3.1, with focus on predictions and highlighting uncertainties beyond expected values and
probabilities, allowing a more flexible approach
than traditional statistical analysis. It acknowledges
that expected values and probabilities could produce
poor predictionssurprises may occur.
Risk analyses, cost-benefit analyses and other types
of analyses are placed in a larger context (referred
to as a managerial review and judgment), where the

327

limitations and constraints of the analyses are taken

into account.
The cautionary and precautionary principles as considered integrated features of risk management. The
cautionary principle is a basic principle in risk management, expressing that in the face of uncertainty,
caution should be a ruling principle, for example
by not starting an activity, or by implementing measures to reduce risks and uncertainties (Aven &
Vinnem 2007, HSE 2001). The precautionary principle is considered a special case of the cautionary
principle; its definition is discussed in Section 4.
A risk classification structure is suggested by the
combination of expected consequences EC and an
assessment of uncertainties in underlying phenomena
and processes that can give large deviations compared
to the expected values.
Starting from the classification based on the traditional risk description using the expected consequences, we may modify the classification based on
the uncertainty assessments: For example, if a system
is classified to have a medium risk according to the
expected consequences criterion, we may reclassify it
as having high risk, if the uncertainties in underlying
phenomena and processes are very large. The uncertainties may be related to for example new technology,
future use and demand for the system, and political
events.

DISCUSSION

The four frameworks have similarities as shown in

Section 3. The overall principles are to large extent
overlapping:
Risk perspectives highlighting events, consequences, probability (likelihood) and uncertainties.
Risk management processes along the lines shown
in Figure 4, and the use of risk reduction processes
such as ALARP.
A risk-informed use of risk analysis. It is acknowledged that risk analyses may provide useful decision
support, but they need to be placed in a wider context, where their scope and limitations are taken into
account. Quantitative risk analyses cannot replace
sound judgments.
Different terminology is used, and different aspects
are highlighted in the four frameworks. The terminology is to varying degree consistent with the
intentions and ambitions of the frameworks, as shown
in Section 3.1. For example, in the IRGC framework
risk is defined as an uncertain consequence of an
event or an activity with respect to something that
humans value. However, the framework in general
does not restrict risk to a consequence. Rather the

framework stresses the importance of reflecting consequence, likelihoods and uncertainties. The adjusted
definition uncertainty about and severity of the consequences of an activity with respect to something that
humans value (Aven and Renn 2008a), can be seen
as reformulation of the original one to better reflect
the intention.
As another example, the Cabinet office (2002)
refers to risk as uncertainty, which means that risk
is considered low if one expects millions of fatalities
as long as the uncertainties are low. Risk management certainly needs to have a broader perspective on
risk, and this of course also recognised by the cabinet
office framework. The terminology may however be
challenged.
When referring to the likelihood of an event we
mean the same as the probability of the event. However, the term probability can be interpreted in different ways as discussed in Section 3.1 and this would also
give different meanings of likelihood. With exception
of the consequence-uncertainty framework none of
the frameworks have specified the probabilistic basis.
In the consequence-uncertainty framework probability means subjective probabilities. Hence there is
no meaning in discussing uncertainties in the probabilities and likelihoods. If such a perspective is
adopted, how can we then understand for example
Figure 2, which distinguishes between uncertainties about likelihoods (probabilities) and uncertainties
about consequences?
The former types of uncertainties are referred to
as epistemic uncertainties and are also called secondorder probabilities. It is based on the idea that there
exists some true probabilities out there, based on
the traditional relative frequency approach, that risk
analysis should try to accurately estimate. However
this view can be challenged. Consider for example the
probability of a terrorist attack, i.e. P(attack occurs).
How can this probability be understood as a true probability, by reference to a thought-constructed repeated
experiment? It does not work at all. It makes no sense
to define a large set of identical, independent attack
situations, where some aspects (for example related
to the potential attackers and the political context)
are fixed and others (for example the attackers motivation) are subject to variation. Say that the attack
probability is 10%. Then in 1000 situations, with
the attackers and the political context specified, the
attackers will attack in 100 cases. In 100 situations
the attackers are motivated, but not in the remaining
900. Motivation for an attack in one situation does not
affect the motivation in another. For independent random situations such experiments are meaningful,
but not for more complex situations as for example
this attack case.
Alternatively, we may interpret the likelihood
uncertainties in Figure 2 by reference to the level of

328

consensus about the probabilities, or by reference to

the amount and quality of the background data and
knowledge for the probabilities.
Or we may use other classification structures, for
example the one which is based on expected consequences and uncertainties in underlying phenomena
and processes (refer to the consequenceuncertainty
framework).
The use of the precautionary principle needs to
be based on an understanding of the risk and uncertainty concepts. The precautionary principle applies
when there are scientific uncertainties about the consequences, but are also uncertainties of the likelihoods
and probabilities included? This is discussed by Aven
(2006).
The level of uncertainties would affect our management policies and strategies. We will always give
weight to the cautionary and precautionary principles in case of large uncertainties. All frameworks
acknowledge this, although the terminology varies.
The HSE framework and the consequence-uncertainty
framework highlight both the cautionary principle,
and not only the precautionary principle. The cautionary principle means that caution, for example by not
starting an activity, or by implementing measures to
reduce risks and uncertainties, shall be the overriding
principle when there is uncertainty about the future
occurrence of events and their consequences. The precautionary principle is a special case of the cautionary
principle used when there are scientific uncertainties
about the consequences. In practice many refer to precautionary principle in the meaning of the cautionary
principle. However, this could be considered unfortunate as it would mean a reference to this principle too
often. Aven (2006) among others prefer to restrict the
precautionary principle to situations where there is a
lack of understanding of how the consequences (outcomes) of the activity are influenced by the underlying
factors.
To manage risk it is common to use a hierarchy of
goals, criteria and requirements, such as risk acceptance criteria (defined as upper limits of acceptable
risk) or tolerability limits, for example the individual probability of being killed in an accident shall not
exceed 0.1%. The use of such criteria constitute an
integrate part of the various frameworks. However, the
weight given to the criteria varies. In the consequenceuncertainty framework these criteria do not play an
important role. According to this framework such criteria should be used with care, and avoided if possible
in particular on a high system level, for example a
plant or a an industry (Aven & Vinnem 2007, Aven
et al. 2006). It is argued that principally speaking, a
requirement (criterion) related to risk and safety cannot be isolated from what the solution and measure
mean in relation to other attributes, and in particular costs. It is impossible to know what should be the

proper requirement without knowing what it implies

and what it means when it comes to cost, effect on
safety etc. If such criteria are defined, they give a focus
on obtaining a minimum safety standardand there is
no drive for improvement and risk reduction. If a high
level of safety is to be obtained, other mechanisms
than risk acceptance criteria need to be implemented
and highlighted, for example ALARP processes. Furthermore, no method has a precision that justifies a
mechanical decision based on whether the result is
over or below a numerical criterion.
HSE (2001) sets the value of a life (the cost one
should be willing to pay for reducing the expected
number of lives by 1) equal to 1 million and proposes that the risk of an accident causing the death of
50 people or more in a single event should be regarded
as intolerable if the frequency is estimated to be more
than one in five thousand per annum. HSE believes
that an individual risk of death of one in a million per
annum for both workers and the public corresponds
to a very low level of risk and should be used as a guideline for the boundary between the broadly acceptable
and tolerable regions.
For the offshore industry a value of a life of
6 million is considered to be the minimum level,
i.e. a proportion factor of 6 (HSE 2006). This value
is used in an ALARP context, and defines what is
judged as grossly disproportionate. Use of the proportion factor 6 is said to account for the potential
for multiple fatalities and uncertainties. Hence the
base case is that a risk reducing measures should be
implemented, and strong evidence (costs) is required
to justify no-implementation.
To verify these criteria expected value based
approaches such as cost-benefit analyses and cost
effectiveness analyses are used, calculating ICAF values (implied cost of averting one fatality, i.e. the
expected cost per expected reduced number of fatalities). This approach is indeed questionable; as the
expected values do not take into account the risks and
uncertainties. A main objective of a safety measure
is to reduce risk and uncertainties, but then we cannot use a principle based on expected values which to
large extent ignores the risk and uncertainties (Aven &
Abrahamsen 2007).
All risk perspectives and frameworks considers in
this paper acknowledges the need for taking into the
account the risk and uncertainties beyond the expected
values, but the practice of using expected value based
approaches is in direct conflict with this recognition.
5

FINAL REMARKS

Our analysis demonstrates that the four frameworks

are based on the same type of fundamental ideas
and principles. However the terminology and practical

329

approaches and methods adopted differ substantially.

This can be partly explained by different scientific
traditions, as the frameworks have been developed in
different scientific environments, and partly explained
by different needs and objectives. The foundations of
all the frameworks have not been clarified. Processes
need to be initiated to strengthen the theoretical basis
of the frameworks. An example in this direction is
Aven and Renn (2008b).
REFERENCES
AS/NZS 4360 2004. Australian/New Zealand Standard: Risk
management.
Aven, T. 2003. Foundations of Risk AnalysisA Knowledge
and Decision Oriented Perspective, Wiley, NY.
Aven, T. 2006. On the precautionary principle, in the context
of different perspectives on risk. Risk Management: an
International Journal, 8, 192205.
Aven T. 2007. A unified framework for risk and vulnerability analysis and management covering both safety and
security. Reliability Engineering and System Safety, 92,
745754.
Aven, T. 2008a. A semi-quantitative approach to risk analysis, as an alternative to QRAs. Reliability Engineering &
Systems Safety, 93, 768775.
Aven, T. 2008b. Risk Analysis, Wiley, NJ.
Aven, T. and Abrahamsen, E.B. 2007. On the use of costbenefit analysis in ALARP processes. I. J. of Performability. 3, 345353.

Aven, T. and Renn, O. 2008a. On risk defined as an event

where the outcome is uncertain. Submitted.
Aven, T. and Renn, O. 2008b. Determining the right level
of investments in societal safety and securitythe role
of quantitative risk assessments. Submitted for possible
publication.
Aven, T. and Vinnem, J.E. 2005. On the use of risk acceptance
criteria in the offshore oil and gas industry. Reliability
Engineering and System Safety, 90, 1524.
Aven, T. and Vinnem, J.E. 2007. Risk Management, with
Applications from the Offshore Oil and Gas Industry.
Springer Verlag, NY.
Aven, T., Vinnem, J.E. and Red, W. 2006. On the use
of goals, quantitative criteria and requirements in safety
management. Risk Management: an International Journal.
8, 118132.
Cabinet Office 2002. Risk: improving governments capability to handle risk and uncertainty. Strategy unit report. UK.
HES 2001. Reducing risk, protecting people. HES Books,
ISBN 0 71762151 0.
HSE 2006. Offshore installations (safety case) regulations
2005 regulation 12 demonstrating compliance with the
relevant statutory provisions.
ISO 2002. Risk management vocabulary. ISO/IEC Guide 73.
Renn, O. 2005. Risk Governance: Towards an Integrative Approach. White Paper No. 1, written by Ortwin
Renn with an Annex by Peter Graham (International Risk
Governance Council: Geneva 2005).

330

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Management of risk caused by domino effect resulting

from design system dysfunctions
S. Sperandio, V. Robin & Ph. Girard
IMS, Department LAPS, University Bordeaux 1, France

ABSTRACT: The aim of this paper is to ensure the development of design projects in an environment with
limited resources (material, human, know-how, etc.) and therefore satisfy the strategic performance objectives
of an organization (cost, quality, flexibility, lead time, deadlines, etc.). The matter of this paper is also to set
down the problems of a real integration between risk and design system managements. With this intention, a
new paradigm for Risk Management Process (RPM) is proposed then illustrated via an industrial case. Such
a RMP includes the tasks of establishing the context, identifying, analysing, evaluating, treating, monitoring
and communicating risks resulting from design system dysfunctions. It also takes into account risks caused
by domino effect of design system dysfunctions and completes risk management methodologies provided by
companies that dont consider this aspect.

INTRODUCTION

In the design, the organization and the planning phases

of services or manufacturing system, the capacity
of a system to maintain its performances in spite
of changes related to its environment or an uncertainty of the data used, constitutes a new issue which
has arisen to complicate an already complex situation. Therefore, in response to the evolutions of its
environment and its possible internal dysfunctions, a
system evolves thanks to actions on its constitution
and/or on the products manufactured and/or the network to which it belongs (suppliers, subcontractors,
customers, etc.). The design system can be defined
as the environment where design projects (product,
service, system or network design) take place. Its
management requires to control each phase of the
product/service/system/network lifecycle, the design
methodologies, the process and associated technologies. The aim of this paper is to ensure the development
of design projects in an environment with limited
resources (material, human, know-how, etc.) and
therefore satisfy the strategic performance objectives
of an organization (cost, quality, flexibility, lead time,
deadlines, etc.). The propagation of accidental scenarios and the comtempory increase in the severity
of event are usually named domino. However, risk
management methodologies provided by companies
dont take into account risks caused by domino effect
of design system dysfunctions. The matter of this
paper is also to set down the problems of a real integration between risk and design system managements.

With this intention, a new paradigm for Risk Management Process (RPM) is proposed then illustrated via
an industrial case. Such a RMP includes the tasks of
establishing the context, identifying, analysing, evaluating, treating, monitoring and communicating risks
resulting from design system dysfunctions.
This paper is organized as follows. In section 2, general points of risk management are introduced. Then,
the RMP relative to design projects is constructed.
Sections 3, 4 and 5 detail the different phases of the
RMP. Risks are examined according to three points of
view (functional, organic, and operational) making it
possible to determine their impacts on the objectives
of the organization. These impacts are evaluated and
quantified thanks to the FMECA methodology, which
allows to identify potential failure modes for a product or process before the problems occur, to assess the
risk associated with those failure modes and to identify and carry out measures to address the most serious
concerns. An industrial case enables to illustrate the
methodology in section 6. Some conclusion remarks
and discussions are provided in the last section.

RISK MANAGEMENT: GENERAL POINTS

Project success or failure depends on how and by

whom it is determined (Wallace et al., 2004). Many
design projects deviate from their initial target due
to modifications or changes to customer needs as
well as internal or external constraints. Other projects

331

Definition of the design project

Monitoring
and
Review

Modeling of the design system

Context Analysis

Risk identification
Monitoring
and
Review

Risk analysis
Risk evaluation
Risk assessment

Corrective actions
No corrective action

Monitoring
and
Review

Risk treatment
Choice of design scenarios
Communication and consulting
Figure 1.

Risk Management Process relative to design projects.

may be simply stopped due to uncontrolled parameters and unidentified constraints arising from various
processes within the project such as the product design
process or the supply chain design (identification of
suppliers and subcontractors for example). The consequence: no risk, no design. Risk processes do not
require a strategy of risk avoidance but an early diagnosis and management (Keizer et al., 2002). Nevertheless, most project managers perceive risk management
processes as extra work and expenses. Thus, risk management processes are often expunged if a project
schedule slips (Kwak et al., 2004). In a general way,
main phases of risk management are (Aloini et al.,
2007): context analysis (1), risk identification (2), risk
analysis (3), risk evaluation (4), risk treatment (5),
monitoring and review (6) and communication and
consulting (7). In agreement with such a methodology,
we propose to use the following process to manage the

risks relative to design projects (Fig. 1). The different

phase of the process are detailed hereafter.

3
3.1

RMP: CONTEXT ANALYSIS

Definition of the design project

The definition phase aims to determine the contents

of the project: product, service, system and/or network design. Design process is the set of activities
involved to satisfy design objectives in a specific
context. The design objectives concern the product/service/system/network definition. They are constrained by the enterprise organization and the design
steps, and are influenced by technologies or human and
physical resources (Wang, 2002). Design is mainly
a human activity and it is also very complex to

332

Actor Axis

Actor
Link 3

Enterprise

Link 2

Process

Organisation

Link 5
Link 4
ic
log
no
h
c
Te

Design
system

is
Ax

Scientific and
Technological
Knowledge

Product

En
vir
on
me
nt
Ax
is

Link 6

External and
Internal
Environments

Link 1

Figure 2.

Design system modeling, interactions between factors influencing the design system (Robin et al., 2007).

understand the tasks carry out by designers (Gero,

1998). When the resolution steps are known (routine
design process), the project is structured according to different activities which transform the product/service/system/network knowledge. Then, the
project is defined like the intention to satisfy a design
objective (a technical function, a part, an assembly or
a complex mechanism). The project manager decomposes the project according to the identified activities
and the actors tasks are very prescriptive. In this
case the respect of the delay is the main performance
objective. So, the project manager decides on the
synchronisation of the human and material resources
availability with the activities needs.
In the other cases, design cant be considered as
a solving problems process, as a creative or innovative design process, and activities dont structure
the project. Design must be identify as a process
which support emergence of solutions. In this case,
the project is organised to favour the collaboration
between actors of the process and the project manager
searches to create design situations which facilitate
the emergence of solutions. He decides on the adapted
organization to favour collaborative work. Innovation
corresponds to the application of new and creative
ideas. Therefore, implement an innovation project
leads to investing in a project by giving up the idea of
an immediate profitability and by accepting a certain

risk. Different types of innovation exist (organizational

change, process or technological innovations, product
development, etc.), which do not rest on the implementation of similar knowledge and means and thus do not
generate the same risks. For example, a survey done
by the Product Development and Management Association (PDMA) reveals that more than 50% of the
sales in successful companies were coming from new
products and that the percentage was even over 60% in
the most successful overall company (Balbontin et al.,
2000).
3.2

Modeling of the design system

Modeling contributes to ideas development and structuring, and can be used as a support of reasoning
and simulation. Design management requires understanding of design process context in order to adapt
actors work if it turns out that it is necessary. Hence,
in their generic model of design activity performance, ODonnell and Duffy insist on the necessity
to identify components of the design activity and their
relationships (ODonnell et al., 1999).
The design system can be defined as the environment where design projects (product or system or
network design) take place. We have identified three
factors influencing the design system and which have
to be considered to follow and manage suitably the

333

design system evolution and the design process (Robin

et al., 2007):
The context in which the design process takes
place. It includes natural, socio-cultural and econoorganizational environments (external and internal
environments). External environment is the global
context in which enterprise is placed (its market,
its rivals, its subcontractors . . .). Internal environment describes the enterprise itself: its structure, its
functions and its organization. Internal environment
is an exhaustive description of the system in order
to take into account all the elements which could
have an influence on the decision-making at each
decision-making level.
The technological factor that concerns the technophysical environment (scientific and technological
knowledge). Scientific knowledge regroups the natural science and the engineering sciences. Technological knowledge concerns the manufacturing
practices and the technology. Interest is to have a
global vision of the knowledge possessed and usable
by the enterprise and to identify a potential lack of
knowledge in some design tasks.
Human and his different activities during design
process (actor). Actor aspects have to consider
multi-facets of the designers. Human resources will
be described with classical indicators (availability of
a resource, hierarchical position, role in a project,
training plan . . .). But factors very close to the
actors personality have to be taken into account too.
These factors influence the design process and more
generally the design system.
These three global factors are global performance
inductors for design system. These factors and their
interactions are integrated in a model composed with
a technological axis, an environment axis and an actor
one (Fig. 2). Then specific objectives, action levers
and performance indicators, dedicated to the design
system, have to be identified according to elements of
this model.
To identify and manage relationships between
global factors influencing performance of the design
process, we propose to use product, process and organizational models (Fig. 2). Product model acts as a
link between knowledge and external/internal environments (link 1, Fig. 2). Product is the expression of the
scientific and technological knowledge of an enterprise and permits to evaluate its position on a market.
Its a technical indicator which allows to make evolve
a firm or to identify a possible lack of competency to
be competitive on a market. As process corresponds
to the place where the knowledge is created and used
by the actors to develop the product, it connects actor
and knowledge (link 2, Fig. 2). Finally, influences of
environments on actors are taking into account by the
mean of an organizational model. This model joins

up actor and external/internal environments (link 3,

Fig. 2). Organization has to favour allocation of
adapted human resources to a specific situation in
a particular context. These models are local performance inductors for design system and interactions
between them provide a dynamic vision of the design
system evolution (links 4 to 6, Fig. 2). In this model,
the description of factors influencing the design system, at each decision-making level provides a global
vision of the design context. Hence, thanks to such
a representation of the design context, the decisionmaker can analyse the design situation and identify
particularities of each project. He is able to observe
evolution of each component (environment, technological and actor one), interactions between them and
consequently to adapt his project management method
by taking the right decision of management to satisfy
objectives.

RMP: RISK ASSESSMENT

4.1 Risk identification

Risk can be planned or unexpected, from external
or internal origin, linked to market trends, eruption
of new technologies, strategic or capitalistic decisions, etc. Risk Identification is discovering, defining,
describing, documenting and communicating risks
before they become problems and adversely affect
a project. There are various techniques that can be
used for risk identification. Useful techniques include
brainstorming methods as well as systematic inspections and technological surveys.
4.2 Risk analysis
According to Lemoigne (Lemoigne, 1974), complex
system analysis refers to three points of view: a functional point of view, i.e. the description of system
functionality and behaviour, an ontological or organic
point of view, i.e. the description of resources used
(human or technical), materials and information, and
related control structures, and a genetic point of view,
which renders system evolutions and development.
Consequently we have developed a modeling approach
to provide analysts with the appropriate view of a
system. Approach regroups a functional model to situates the system within its environment, an organic
view to depict the physical organization and resources
which achieve functions previously identified and
an operational view which stipulates the ways the
organic system is exploited. The application of such
a decomposition to the global and local factors of the
design system is proposed hereafter (Fig. 3). Examples of potential evolutions of these factors are also
presented.

334

Functional

Global factors

Internal
environment

External
environment

Scientific and
technological
knowledge

Actor

Local factors

Product

Process

Organization

Organic

4.3

Evolution of the company structure

Operational

Evolution of the company operating modes

Functional

Evolution of the company place in the network

Organic

Evolution of the partners in the network (subcontractors, collaborators,)

Operational

Evolution of the partnerships (new contracts, new collaborations,)

Functional

Company knowledge and know-how evolution (fundamental sciences)

Organic

Evolution of the methods and techniques associated to this knowledge

Operational

Evolution of the operating modes

Functional

Evolution of the actors role

Organic

Evolution of the actors positioning in the structure

Operational

Evolution of the actors knowledge and competencies (training period,)

Functional

Evolution of the product service function

Organic

Evolution of the product structure

Operational

Behavioral evolution of the product

Functional

Evolution of the design activities

Organic

Evolution of the design activities sequencing (anteriority, parallelism,)

Operational

Evolution of the resources that realize the activities

Functional

Evolution of the design system role

Organic
Operational

Figure 3.

Evolution of the company role

Evolution of the design system structure (design centres)

Evolution of the design centres processes

Potential modifications of local and global factors of the design system.

Risk evaluation

When a risk is identified, its impact or effect on the

design system (and therefore on the design project
since the design system can be seen as the environment where design projects take place) is evaluated
thanks to the FMECA methodology. Such a methodology allows to highlight potential failure modes for a
product or a process before the problems occur, to
assess the risk associated with those failure modes
and to identify and carry out measures to address the
most serious concerns (Kececioglu, 1991). In accordance with the FMECA methodology, the criticality of
events is valued by the Risk Priority Number (RPN)
(see Equation 1 below), which is linked to three criteria: the occurrence frequency of the event O, the
gravity degree of this event G and the risk of its
non-detection D.

Case 4:
Fatal

Initialization

Functional
definition

Case 3: Strongly
disturbing

Risk
Risk
management

2
Risk

Case 1: Slightly
disturbing

2
Case 2: Fairly
disturbing

Organic
definition

Operational
definition

Project Reengineering

Figure 4.
system.

RPN = O G D

Project Engineering

Criticality classes and domino effect on design

(1)
5

The RPN allows to organize eventual corrective

actions. Criteria weighting results from the expertise
of leaders and/or analysts and/or operators, on the basis
of experiment.

RMP: RISK TREATMENT

A good risk treatment requires a decisional framework

in order to visualize the different risk criticalities. We
consider that a risk is (Sperandio et al., 2007):

335

Criticality classes

Level of risk

C1
(case 1, Fig. 4)

Acceptable in the present state

C2
(case 2, Fig. 4)

Tolerable under regular control

C3
(case 3, Fig. 4)

Difficult to tolerate

C4
(case 4, Fig.4)

Unacceptable

Figure 5.

Decision
No action or modification at the operational level.
Follow-up, monitoring and review.
Risk assessment.
Modification at the organic level.
Follow-up, monitoring and review.
Risks assessment.
Modification at the functional level.
Follow-up, monitoring and review.
Risk assessment.
Change of strategy.
Total reorganization of the project.

Criticality scale and decisional framework.

Definition of the design project

Modeling of the design system

Monitoring
and
Review

Risk identification
Monitoring
and
Review

Risk analysis
Risk evaluation

Corrective actions
No corrective action

Monitoring
and
Review

Choice of design scenarios

Communication and consulting
Figure 6.

Criticality classes and domino effect on the Risk Management Process.

Slightly disturbing if it has no impact on the

structure of the design system (case 1, Fig. 4):
changes of resources capacities, modifications of
operational constraints, for example. There are several types of operational constraints: constraints of
precedence, temporal constraints, cumulative constraints, disjunctive constraints, etc. A slightly risk
belongs to the criticality classe number 1 (C1),
which includes a risk level acceptable in the present
state.
Fairly disturbing if it acts upon the organic definition of the design system (case 2, Fig. 4): capacities

and/or competences of human and technical

resources, legislatives constraints, etc. Such a risk
belongs to the criticality classe number 2 (C2),
which includes a risk level tolerable under regular
control.
Strongly disturbing if it requires strategic adjustments of the design system, impacting its functional
characteristics (case 3, Fig. 4): change of industrial
activity, integration of new technologies, modification of the supply chain, etc. Such a risk belongs to
the criticality classe number 3 (C3), which includes
a risk level tolerable with difficulty.

336

 Fatal if it makes the design system obsolete (case 4,

Fig. 4): bad investments, for example. Such a risk
belongs to the criticality classe number 4 (C4),
which includes an unacceptable risk level.
Determination of these criticality classes depends
on specificities of each system. When all the
risks impacting design system are identified and
all the corresponding RPN are calculated, a criticality scale could be created and valuation of
classes could be done.
Reengineering the project during its life-cycle is
a procedure triggered each time significant events
which impact the project occur. The functional,
organic and operational definitions of the project
should then be tuned accordingly to support reengineering reasoning.
A decisional framework enables to visualize risk
criticalities. According to their value, the generic criticality of risks consequences entitles to define actions
to be carried out (see Figure 5). Obviously, such
actions are dependant on the companys intention of
doing risk management. Therefore, an effort combined with the actions of risk management can be
defined according to a qualitative scale: for example,
no action, vigilance or selective action, vigilance or
periodic action, vigilance or continuous action, etc.
Finally, the propagation of accidental scenarios and its
impact on the Risk Management Process are presented
hereafter (Figure 6). Such an approach enables to take
into account risks caused by domino effect of design
system dysfunctions.

6
6.1

INDUSTRIAL CASE
Introduction

The company produces ice creams for the great distribution. Characteristic of this branch of industry lies in
the fact that it is a seasonal production. The off season corresponds to a relatively reduced production. At
the height of the season, the company is required to
have a perfect control of its production equipments
(very high rate of production), because any matter
loss or overconsumption could have a strong impact on
the productivity. Therefore, the appearance of events
susceptible to modify the functionality, the structure
and/or the operational scenarios of the production system is extremely prejudicial for the company, and it is
imperative to analyze the criticality of these events
in order to launch rapidly adequate and corrective
actions.
Innovation has a key role to play in the performance
of such a firm. Considered as a solution for growth and
competitiveness, it is used by managers to create new
sources of value.

In this example, we decide to analyze two possible

evolutions (evolutions of the company or of its environment) which can constitute a risk for the design
system. Our analyse is limited to the investigation of
the global inductors impacts on the local inductors,
but it is important to also consider that the modification of these global inductors can influence other
global inductors. Let us consider two events E1 and E2.
E1 corresponds to a breakdown in the process and E2
is a lack of natural aroma of chocolate.
6.2

Risk management

Criticality analysis of events susceptible to have an

impact on the design system refers to a grid (Figure 7),
in which the quotation scales of the occurrence frequency of the event (O), the gravity degree of the
same event (G) and the risk of its non-detection (D)
illustrate the expectations of the various actors in the
company:
The occurrence scale (O) is defined by the appearance frequency of a dysfunction. An exceptional
event will be credited with 0.25 point; a rare event
with 0.50 point, a recurrent event (weekly, daily)
with 0.75 point and a permanent event with 1 point.
The gravity scale (G) considers the sum of the quotations obtained for the criteria previously defined:
organization, process and product. Criteria are
marked out of 25. Therefore, a criterion of minor
gravity (operational consequence on the exploitation on the system) is credited with 5 points, while
the criterion considered as the most significant criterion, i.e. the criterion which has big consequences
on the exploitation of the system (functional definition), is credited with 25 points.
The non-detection scale (D) measures the probability of not detecting a potential failure when the
cause exists. The term probability used here is
not standard, since the values corresponding to these
probabilities are not contained between 0 and 1. Voluntarily, the graduation is bounded by 10 and 20,
in order to clearly indicate the importance of the
various causes at the origin of the dysfunctions.
It is then necessary to identify the origin of the risk
and to quantify the impact of this risk on the elements
of the design system (product, process, organization).
Such a work is carried out by the FMECA group: one
or more experts who have knowledge on the tool and
conduct the discussions, actors of the design who provide essential knowledge relating to the design system,
and persons in charge of other services (production,
marketing, etc.) who evaluate the impact of the risk
in their own fields. In accordance with the FMECA
methodology, the criticality of events is valued by the
Risk Priority Number (Figure 7).

337

R2

Figure 7.

Recurrent

Permanent

0.50

0.75

Risk Priority Number

Rare

Product
Organic

0.25

Functional

Operational

5 15 25

412.5

325

RPN

Impact of enterprise evolution on the design system.

The Risk Priority Number of E1 is given by:

RPN 1 = 15 (15 + 25 + 15) 0.5 = 412.5

(2)

The Risk Priority Number of E2 is given by:

RPN 2 = 20 (25 + 25 + 15) 0.25 = 325

(3)

6.3 Conclusion
The Risk Priority Number of E1 is higher than the Risk
Priority of E2. Therefore, in order to launch the corrective actions efficiently, it will be necessary to initially
treat dysfunctions due to E1. Moreover, impacts of the
risks on the design system have been quantified, which
will enable to adjust the design strategy. Events leading
to operational modifications of the company are common, and are an integral part of everyday life. With
this intention, the company launched out in a step of
continuous improvement (preventive and autonomous
maintenance) in order to, on the one hand, preventing slightly events and, on the other hand, solving all
dysfunctions which can exist in production workshops
(dysfunctions related to workers environment. Such a
step aims at extending the life span of equipments and
decreasing times of corrective maintenance.
7

Organic

5 15 25

Exceptional

5 15 25

Functional

Operational

Process

Organization
Organic

Functional

Operational

Organic

Functional

Functional

Operational

10 15 20 10 15 20 10 15 20
R1

Occurrence :
Appearance frequency
of a dysfunction

Consequences :
Local inductors evolution

Actors

Scientific and
technological
knowledge
Organic

Functional

Operational

Organic

Operational

Internal / external
environments

Causes:
Global inductors evolution

CONCLUSION

Reengineering the design project during its life-cycle

is a procedure triggered each time significant events

occur. The functional, organic and operational models (or definitions) of the design system should then
be tuned accordingly to support reengineering reasoning. The methodological guidelines are based on
event criticality analysis. A classification of events
was made to guide the analysts towards appropriate model tuning, such that the representation of the
system be permanently in conformity with the system despite the continuous modifications encountered
by the system during its life-cycle. A risk management methodology is also provided in order to take
into account risks caused by domino effect of design
system dysfunctions. The Risk Management Process
includes the tasks of establishing the context, identifying, analysing, evaluating, treating, monitoring
and communicating risks resulting from design system
dysfunctions.

REFERENCES
Aloini, D., Dulmin, R., Mininno, V. (2007). Risk
management in ERP project introduction: Review
of the literature, in: Information & Management,
doi:10.1016/j.im.2007.05.004.
Balbontin, A., Yazdani, B.B., Cooper, R., Souder, W.E.
(2000). New product development practices in American
and British firms, in: Technovation 20, pp. 257274.
Gero, J.S. (1998). An approach to the analysis of design
protocols, in: Design studies 19 (1), pp. 2161.
Kececioglu, D. (1991). Reliability Engineering Handbook,
Volume 2. Prentice-Hall Inc., Englewood Cliffs, New
Jersey, pp. 473506.

338

Keizer, J., Halman, J.I.M, Song, X. (2002). From experience:

applying the risk diagnosing methodology, in: Journal
Product Innovation Management 19 (3), pp. 213232.
Kwak, Y.H., Stoddard, J. (2004). Project risk management:
lessons learned from software development environment,
in: Technovation 24, pp. 915920.
Lemoigne, J.L. (1974). The manager-terminal-model system is also a model (toward a theory of managerial
meta-models).
ODonnell, F.J.O., Duffy, A.H.B. (1999). Modelling product
development performance, in: International Conference
on Engineering Design, ICED 99, Munich.
Robin, V., Rose, B., Girard, Ph. (2007). Modelling collaborative knowledge to support engineering design project
manager, in Computers in Industry 58, pp. 188198.

Sperandio, S., Robin, V., Girard, Ph. (2007). PLM in the

strategic business management: a product and system
co-evolution approach, in the proceedings of the Internal Conference of Product Lifecycle Management, July
1113 2007, Milan, Italy.
Wallace, L., Keil, M., Rai, A. (2004). Understanding software project risk: a cluster analysis, in: Information &
Management, 42, pp. 115125.
Wang, F., Mills, J.J., Devarajan, V. (2002). A conceptual approach managing design resource, Computers in
Industry 47, pp. 169183.

339

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On some aspects related to the use of integrated risk analyses

for the decision making process, including its use
in the non-nuclear applications
D. Serbanescu, A.L. Vetere Arellano & A. Colli
EC DG Joint Research Centre, Institute for Energy, Petten, The Netherlands

ABSTRACT: Currently, several decision-support methods are being used to assess the multiple risks faced
by a complex industrial-based society. Amongst these, risk analysis is a well-defined method used in the
nuclear, aeronautics and chemical industries (USNRC, 1998; Haimes, 2004). The feasibility of applying
the Probabilistic Risk Assessment approach (USNRC, 1983) in the nuclear field (PRA-Nuc) for some new
applications has been already demonstrated by using an integrated risk model of internal and external events
for a Generation IV nuclear power plant (Serbanescu, 2005a) and an integrated risk model of random technical and intentional man-made events for a nuclear power plant (Serbanescu, 2007). This paper aims to
show how such experiences and results can be extended and adapted to the non-nuclear sectors. These extensions have been shown to trigger two main methodological novelties: (i) more extensive use of subjective
probabilities evaluations, in the case of non-nuclear applications and (ii) inclusion of hierarchical systems
theory in the PRA modelling. The main aspects of the results and conclusions of the above-mentioned
cases, along with insights gained during this analysis are presented and discussed in this paper. In particular, this paper is a synthesis of insights gained from modelling experiences in extending PRA-Nuc to new
applications.

INTRODUCTION

Risk analysis methods are used to support decisionmaking for complex systems whenever a risk assessment is needed (USNRC, 1983). These methods are
well-defined and routinely employed in the nuclear,
aeronautics and chemical industries (USNRC, 1998;
Haimes, 2004; Jaynes, 2003). This paper will describe
experiences gained and lessons learnt from four applications: (i) use of PRA-like methods combined with
decision theory and energy technology insights in
order to deal with the complex issue of SES (Security
of Energy Supply); (ii) use of PRA for new applications in the nuclear field (PRA-Nuc); (iii) use of PRA
for modelling risks in non-nuclear energy systems,
with special reference to hydrogen energy chains and
(iv) use of PRA for modelling risks in photovoltaic
energy production. An important experience gained
in this research is the discovery that the problems
encountered in the use of this approach in nuclear and
non-nuclear applications have common problems and
solutions that can be grouped into a set of paradigms,
which are presented below.

CONSIDERATIONS ON PRA
EXTENTED USE

The next paragraphs describe the seven problemsolution paradigms common for all four applications.
2.1

Suitable systems for PRA modeling

The systems that could be modelled using PRA-Nuc

should be a special type of complex systems, i.e.
complex, self-regulating, self-generating and hierarchical [Complex Apoietic Systems-CAS] (Serbanescu
2007a; Serbanescu et al, 2007a, 2007b).
The model to be built and the method to be used
for the evaluation of the risk induced by CAS have to
comply with seven requirements: (i) To be systemic
(i.e. use of systems theory in modelling), systematic
(i.e. use the same approach throughout the entire PRA
process) and structured (i.e. consider the model hierarchical and evaluate each of the levels one by one
either by using top-down or bottom-up approach;
(ii) To be able to define structures of dynamic cybernetic interrelations between components with both

341

random and intentional types of challenges to CAS,

and to solve nonlinear dynamic models by defining
what linearity means for a CAS; (iii) To define the
system as a whole, as being a result of the synergetic
interfaces of its components and to define also the
CAS interface with the environment; (iv) To have a
solution for the system control (like for instance distributed control, hierarchical control and/or external
to the CAS unitary control); (v) To have a system
management based on predefined objectives, such as
energy/substance balance or risk impact, including
validation and verification processes. (vi) To solve the
specifics of the cause-effect issue for a CAS, which is
connected in its turn to other issues like definition of
linearity, uncertainty and system structure modeling;
(vii) To be dynamic and highly flexible in defining
initial and boundary conditions.
Feasibility of extension of PRA-Nuc application
was previously demonstrated for other CAS-like Generation IV nuclear power plants (Serbanescu, 2005a,
2005b) and an integrated risk model of random technical and intentional man-made events for a nuclear
power plant was described in Serbanescu 2007a,
2007b.
2.2

Figure 2. Representation of a CAS model to be implemented in PRA-Nuc codes (e.g. Risk Spectrum).

System description

Figure 1 shows the general process of PRA analysis

for CAS systems: for all the Initiating Events (IE),
the reaction of CAS is modeled considering also the
expected End States (ES) of each scenario, while the
PRA model for a CAS is represented in Figure 2.
Three levels of modeling can be identified: level 1 is
related to the CAS reaction after all the functional barriers have reacted; level 2 represents the CAS model
after the last physical barrier separating the installation

Figure 3. Sample representation of the main systems of a

hydrogen distribution installation.

from environment/population has been considered

and level 3 represents the final-total risk impact of
the challenges to the installation.
As modeling requires also a representation of the
interconnections between the systems, a set of Connecting End States (CES) is defined. CES are used
to assure both horizontal (at the same level of the
CAS model) and vertical (between levels of CAS)
connections between scenarios.
As an example, Figure 3 summarizes the CAS
model being developed for modeling of the hydrogen distribution station. Similar models were also built
for security of energy supply evaluation (Serbanescu
et al, 2008) and for photovoltaic manufacturing process (Colli et al, 2008). Abbreviations codes for each
of the elements are shown in the Figure 3; abbreviations IV1, IV2 and IV3 are used for isolating valve
systems.
2.3 Scenarios and faults (barriers) description

Figure 1. General representation of the PRA process for a

Complex Apoietic System (CAS).

Figures 4 and 5 illustrate the general process of generating ES by combining the scenarios defined in the

342

Figure 4.

Sample representation of a CAS fault tree model.

Figure 5.

Sample representation of CAS event tree model.

Event Trees with the Fault Trees of the systems, which

are assumed to operate as barriers against various
challenges. Such a process is known in PRA-Nuc as
integration of Fault Trees into Event Trees.
In these figures, the representation technique shown
is similar to the graph theory approach and is in accordance with the existing standards (USNRC, 1983).
Nodes and basic events (i.e. final failures of a given
chain of failures) represent failures in both figures.
However, in Figure 4, nodes are showing possibility
of failures for entire systems of the CAS, which were
designed to act as barriers, while in Figure 5, nodes are
showing the possibility of failure of given components
in a given system. In Figure 4, ES(1-n) represent end
states of a given scenario, which indicate certain levels
of damage after the scenario took place, while ES(1-n)
are used as CES. The TOP nodes from Figure 5 are

connected with the SC and EN type nodes in Figure 4.

The calculation of scenarios leading to an end state
in Figure 4 will include combination of failures at the
level of components of the CAS. Please note that the
notations shown in these figures follow the PRA-Nuc
principles and are used during the implementation of
the model into the computer codes.
It is possible to show that a CAS model of challenges
is generating a -algebra over the set of all possible ES
(Serbanescu 2005a, 2007a). In such a structure risk
is defined as a norm in the measurable vector space of
the sets of scenarios. The risk is defined as the distance
between normal state and an altered end state. This
distance is usually calculated, as a product between
the probability of a given sequence of scenarios and
the damages associated with the endstates after the
scenario will take place.
The scenarios resulting from Boolean combinations
of failures differ in many aspects, as for instance the
combination of failed barriers or the ES (final conditions) after the scenario came to an end for the given
level of PRA model.
Figure 6 shows an example of a CAS scenario definition. End states for levels 1 and 2 are also known as
Release Categories (RC) while ES of level 3 are known
as Risk Categories (RK). More details of these and representative examples of these entities are presented in
(Colli et al, 2008).
In any case, the ES identification is subject to continuous iterations and sensitivity analyses following
the process previously represented in Figure 1. Table 1
is a sample of endstates for the hydrogen installation shown in Figure 3. The Code column is only
indicative to show that a code is needed for the implementation of the model into PRA-Nuc computer codes.
A sample of a typical IE list for the hydrogen installation presented in Figure 3 is: (i) Breaks (e.g. break

Figure 6.
tion.

343

Sample representation of CAS scenario defini-

Table 1.

between the acceptable threshold for risk and the

frequency of a given event (IE) to happen.
From a topological point of view, this threshold in
a 3-dimensional subset (defined by the risk, probability of an IE and the dominant parameter of the CAS)
splits the space of possible solutions into two areas separating the acceptable solutions from the unacceptable
ones. The threshold line is linear if the axes of risks and
event probability/frequency are represented in the logarithmic scale. This line is another representation (in
a 3D form) of Pareto sets of solutions as presented in
(Serbanescu 2005a, 2007a), and for this reason, PRA
can be also used to define the Pareto set of acceptable
solutions for a CAS.
More in detail, (Serbanescu 2005a) one can consider a system as a cybernetic hierarchical one (e.g. of
a CAS type as explained before).
As it was shown in Serbanescu 2005a for such
systems the distance from the best solution is quantitatively described by a specific type of risk measure
calculated as a loss of information entropy (Jaynes,
2003, Smithson, 2000). Consequently, the optimum
for a certain problem is achieved when the Shannon
information entropy1 reaches a minimum, given the
variables constraints. In mathematical terms, this leads
to the task of finding the optimum for (1) with the limit
conditions (2).

Example of End States for Hydrogen installation.

End State

Parameter(s)

Code

10 mbar
500 mbar

EFOV4
EFOV4

Thermal

3 kW/m2
35 kW/m2

FTH1
EFTH3

Gas Cloud

Not dangerous
Dangerous in size

EFGC1
EFGC2

Window break
Building collapse
Eardrum rupture
Fatalities 1%
Fatalities 100%

HOVDG1
HOVDG2__
HOVINJ__
HOVFAT1__
HOVFAT3__

Thermal

Burns degree 1
Burns degree 2
Glass/window fail
Fatalities 1%
Fatalities 100%

HTHBU1
HTHBU2__
HTHFIRE__
HTHFAT1__
HTHFAT3__

Fire

Fatalities 1%
Fatalities 100%

HFIFAT1
HFIFAT3

Explosions

Fatalities 1%
Fatalities100%

HEXFAT1
HEXFAT3

Effect levels
Overpressure

Harm effects
Overpressure

Sinf = XI ln XI ,
of vessels/cylinders in storage cabinet unit); (ii) Leaks
(e.g. leak from the hydrogen inlet ESD); (iii) Overpressurization (e.g. overpressurization of 10 mbar in
PD Unit); (iv) Thermal challenges to installation (e.g.
thermal impact of 3 kW/m2 in PD Unit); (v) Fire,
explosions, missiles (e.g. fire generated by installation failures in PD unit); (vi) Support systems failure
(e.g. loss of central control); (vii) External event (e.g.
external Flood in UND); (viii) Security threats (e.g.
security challenge in area close to units).
A specific feature of CAS model level 3 for security
of energy supply (SES), is that ES are SES risk related
criteria.
In this case, scenarios leading to various SES categories are ranked on the basis of their risk and importance, in accordance with PRA-Nuc methodology. It
is also important to note that for the SES case, a set
of economical (e.g. energy price variations) and sociopolitical initiators (e.g. result of referendum for a given
energy source) were also defined (Serbanescu 2008).

2.4

xi ,

(1)
(2)

In formulas (1) and (2), X stands for any risk metric of a given PRA level (e.g. probability of failure
or risk) and is used for the limitations imposed to
those parameters (e.g. optimization target for acceptable probability of a given ES or risk values). The
solution of equations (1) and (2) represented in the
form of a Lagrangean Function (LF).
In order to analyse possible changes to the risk metrics, a tool is needed to evaluate the resulting impacts
so that decisions can be taken on what course of action
to follow. The proposed tool to carry this out is based
on Perturbation Theory (Kato, 1995). This approach is
used for matrices built with event trees and fault trees
in PRA specialized computer codes. Therefore, in
order to evaluate the impact of challenges on the CAS
models and/or modifications and/or sensitivity analyses related changes, linearity is assumed within the
perturbated limits of the CAS model. This linearity
is related to the logarithmic risk metrics set of results.

Objective function and risk metrics

For any type of CAS (nuclear, chemical etc), constraints exist in terms of risk. These constraints are
usually represented in the form of a linear dependency

1 Shannon

information entropy is defined for the set of

CAS risk metrics solution.

344

The verification of the impact of this assumption is

one of the main targets of the sensitivity cases.
As a consequence of such an assumption the building of the model and the calculation of solutions is
guided by formulas of type (3), in which R represent the risk metrics function, NL index is related
to non-perturbated initial model/solution, HC is
related to the induced perturbation in a format of a
Hazard Curve, R 0 is the non perturbated solution
for risk metrics (reference solution), R the modification induced in the results by a given assumption and
error is the tolerance.
Figure 8. 3-dimensional representation of the risk metrics
results for CAS level 3.

RNL = FNL (PM , HC) = F0 HC

= R 0 + R + error

(3)

As it is shown in Figure 7 and in formula (3), F is

the function that builds the solutions of the PRA-Nuc
model and it is defined based on the sets of end states
PM and the hazard curve of a given new phenomenon
to be modeled (HC), and with values in the sets of values obtained for risk calculations of all ES as defined
by the set RNL . (Serbanescu 2007a).
An example of risk metrics results for a Generation IV nuclear power plant has been explained in
more detail in Serbanescu 2005a, where different
approaches to evaluate risk metrics were discussed
regarding their use in decision-making related to optimized solutions from a risk perspective. Figure 8
shows two approaches to presenting risk metric results.
A2 is the classical approach, as defined by USNRC
(1983). B2 is the approach based on the Lagrangean
Function (LF) approach mentioned in Section 4 of this
paper. If we would like to use these approaches in
decision-making, it is important to note the subtle difference between A2 and B2. In the classical approach
(A2), the user has to bear in mind both the space
of acceptable solutions from a risk perspective (i.e.

paraboloid inside the cone) and the limitations (the

cone). In the LF approach (B2), the limitations are
already embedded in the space of acceptable
solutions defined by the internal paraboloid (i.e. the
cone is already embedded within the paraboloid). The
external paraboloid in B2 indicates the external upper
bound of the solutions and not the limits defined by
the cone in A2.
In both cases the acceptable solutions from a risk
perspective indicate scenarios for acceptable design
solutions with various degree of feasibility, providing also important information for the CAS risk
optimization process.
As shown in Figure 8 there are 10 types of solutions identified in the solutions space. For A2 these
are: 1-low risk impactnot difficult design solutions;
2-low risk impactsolutions defined mainly by active
systems features; 3-medium risk impactachievable
design solutions; 4-medium to high risk impact
solutions supported by passive features; 5-high but
still acceptable risk impactdifficult design solutions. For B2 these are 610, where 6 corresponds
to 5, i.e. high but still acceptable risk impact
difficult design solutions, and 10 corresponds to 1
(i.e. low risk impactnot difficult design solutions).
It is also worth noticing that results shown in Figure 8
are typical for all the CAS mentioned in this paper
(details for other cases are in (Colli et al, 2008) and
(Serbanescu et al, 2008)).
2.5 Risk calculation and decision making
processes

Figure 7. Sample representation of the definition of the

solutions for PRA-Nuc in new applications based on the
old ones.

One of the main uses of the risk analyses is to support the decision making process in various areas. The
graphical representation of the commonly used decision analysis problems methods was introduced by
Howard (1984).
PRANuc has been used to support the decision making process both as a guiding information (Risk Informed Decision MakingRIDM) or

345

to assess risk based criteria (Risk Based Decision

MakingRBDM). The experience gained in its application to the nuclear system led to identify a set of
areas of applicability for both the deterministic and
probabilistic tools, as shown in Figure 9.
In this classification the area of applicability is
defined by two criteria: (i) credibility in the assumed
degree of conservatism of the model and (ii) credibility
in the uncertainty level of the built model and applied
method to evaluate risk.
Formula (4) can be applied in order to find out how
certain could one be on a risk analysis based on probabilistic and deterministic set of results and/or how to
combine them (if possible).
O=

(P U(P) )

RP

RG1

(D U(D) )

RD

RG2

(F U(F) )
(4)

RF

The function O (Objective of the decision process)

is a result of a combination using a series of logic
operators ():
RP,D for reasoning on credibility of probabilistic
and respectively deterministic results, and RF for
the reasoning on the credibility of reasoning based
on feedback from experiments/real cases;
RG1 and RG2 for connecting results on reasoning
based on:
probabilistic evaluations (for the terms noted with
Pprobabilistic statements and UP probabilistic statements uncertainties),
deterministic evaluations (for the terms noted
with
Ddeterministic
statements
and
UD uncertainties of deterministic statements),

Figure 9. Sample representation of areas of applicability for decision making of deterministic and probabilistic
approaches (Serbanescu 2007b).

feedback review statements (for the terms noted

with Fstatements based on feedback review
and UF uncertainty of the statements from
feedback review).
It is also important to notice that risk analyses
results are fundamentally divided in deterministic oriented statements and probabilistic oriented
statements.
For deterministic judgments the result is composed
of the criteria value D and the level of uncertainty
in this values (UD ); for the probabilistic results the
components of the results are P and UP . There is
also a component of results given by feedback from
real object while compared to the model (F set of
statement).
Operator will have various impact on the final
function (with low-L, medium-M or high-H) impact)
as shown in Table 2 depending on the type of judgment cases, in which the decision maker positions
himself (which could be optimistic, pessimistic etc.).
The result of how the final credibility should be considered given a set of deterministic results is illustrated
in the Table 2, which shows that the role of the decision maker can be also modeled and considered a priori
so that variations in the conclusions of the same risk
results used by various interest groups could be predicted and understood. Understanding risk results is
one of the main conditions of assuring a good risk governance process and maximizing the use and impact
of the risk evaluations.
2.6 Understanding errors and judgment biases
In the search for solutions on the task of self-assessing
the credibility of risk analyses results, two groups of
systematic biases are identified. The first is related
to the generic scientific method and its drawback, for
which a series of systematic errors (called scientific
myths) exist as defined in (Mc Comas, W. 1996):
(i) Hypotheses become theories which become laws;
(ii) Hypothesis is an educated guess; (iii) A general
and universal scientific method exists; (iv) Evidence
accumulated carefully will result in sure knowledge;
(v) Science and its methods provide absolute proof;
(vi) Science is procedural more than creative; (vii)
Science and methods can answer all questions; (viii)
Scientists are particularly objective; (ix) Experiments
are the principle route to scientific knowledge.
The second group of biases is related to the myths
generated by risk analyses for CAS, for which a list
as defined by (Hanson S.O., 2000): (i) Risk must
have a single, well-defined meaning; (ii) The severity of risks should be judged according to probability
weighted averages of the severity of their outcomes;
(iii) Decisions on risk should be made by weighing
total risks against total benefits; (iv) Decisions on risk

346

Table 2.

Sample representation of the reasoning operators from formula (4) used in decision making statements.
Case 1
Optimistic
trust in risk
results

Case 2
Pessimistic
trust in risk
results

Case 3
Neutral trust
attitude on risk
results

Case 4
Overoptimistic
trust of risk
results

Case 5
Overpessimistic trust
attitude in risk
results

L
L
H
L

L
L
H
L

M
M
H
M

L
L
H
L

M
H
H
H

H
H
L
M

H
L
L
M

M
L
L
L

H
H
L
M

L
L
L
L

H
H
L
H

H
L
L
M

M
L
L
L

H
H
L
M

L
L
L
L

R G1
R G2
O
Total objective
function

M
H

M
M

L
M

L
H

L
M

Impact Function/
Cases
RP
U(P)
TOTAL P
RP
U(D)
TOTAL D
RF
U(F)
TOTAL F

should be taken by experts rather than by laymen; (v)

Risk-reducing measures in all different sectors of society should be decided according to the same standards;
(vi) Risk assessments should be based only on wellestablished scientific facts; (vii) If there is a serious
risk, then scientists will find it if they look for it.
2.7

Solutions to paradoxes

Identification of existence of possible systematic

errors induced by various pre-concepts usually generated in general science and in risk analyses of CAS,
in particular, is only the first step in an attempt to find a
solution for continuous improvement of the credibility
of results.
In order to search possible solutions for the systematic errors in the models and in the methods some
criteria were identified. One of them is related to the
knowledge acquisition process. If the theory of triadic approach as defined in (Peirce, C.S., 2008) is
applied as one possible interpretation of the process,
then the process of developing CAS models could be
represented as in Figure 10.
The process of finding solutions for the CAS model
using an adequate method can then be represented
as in Figure 11, where the areas of acceptable solutions for the CAS model is defined by the intersection
between the areas of the set of model representation and the set of solutions obtainable with the CAS
method described before.

Figure 10.

Sample representation of the reasoning triads.

The model is therefore defined by a set of three

relations between the set of real objects (R), for which
the set of CAS (S) models are built, while generating
the set of systematic biases (M).
As it is shown in Figure 11; the triad of models
is represented by KNW-SOC-ENV (being equivalent
to the triad in Figure 10) and the triad of methods is
represented by RSK-DEC-CAS, which is built based
on PRA-Nuc with the specifics presented previously
in this paper (cf. Sections 1,3 and 4).
The intersection of the two triads (models and methods) generates four areas of solutions represented by
the concentric circles (0, I, II and III) shown in Fig. 10.
These circles are indicating in increasing order the
level of uncertainty and knowledge gained by a certain
risk evaluation process.

347

Table 3. Representation of the reasoning process as per

(Descartes, 1637).
GOALS AND CRITERIA
R1
R2
R3

Goal of knowledge is to formulate judgments (true

and fundamental) on Objects
Objects are targets, for which knowledge is
indisputable and real
The main criterion for sound knowledge is to get
yourself intuition of the objects

THE TOOL IS DEFINED BY THE RULES

R4
R5

Figure 11. Sample representation of the process of building

methods in order to find solutions to the CAS type models.

R6

Another important aspect is related to the search

of systematic errors in CAS risk analysis. This search
is done using the Cartesian approach in each triad of
the model or method. This search, which is shown
in Table 3, is used for all steps in CAS modelling
and solution-finding processes, independent of the
risk evaluation approach type, i.e. deterministic or
probabilistic.
In (Serbanescu 2005a) a suggested series of steps
was identified in all the CAS modelling performed so
far for new PRA applications.
In order to define the specific actions needed to
avoid loss of control over the process of credibility selfassessment of results, three preliminary steps were
performed (Serbanescu 2007b):
First step: identify the detailed governing principle of each of the phases described below:
P1 Unique source modeled with integrated
method leads to the need to diversify definitions of
objective functions.
P2 Deterministic versus probabilistic approaches
as complementary approaches seem to cope with the
dual features of CAS risk models, but their use creates
paradoxes if the applicability for each of them is not
defined.
P3 Combine deterministic and probabilistic by
using weights in judgments on random variables generating numerical results for which credibility depends
on a solution not at the level of a simple combination between the two types of data (deterministic and
probabilistic).
P4 Test the model in order to check its stability
for paradoxes identified up to this phase, which results
in a closed logic loop that defines the reference case
while remaining in the limits of the model itself.
P5 A CAS model could be in one of nine stages.
They follow each other, and, their duration, intensity
and degree of challenge of paradoxes at each phase
differ from one CAS type to another. Furthermore,

The Tool is the Method

The tool features are to aim at Organizing and
Developing Gradually Increased Complexity
Hierarchy of the ObjectsGICHO
In order to fulfill rule R5 a Complexity Level
Criteria (CLC) derivation process is used so that
the following are performed:
To define what is simple and hence the reference CL0
To define the departure from CL0 of other
objects C = CL CL0

EVALUATE COMPLETENESS OF KNOWLEDGE

R7

Evaluate of the objects in the interaction (synergetic

mode).
The condition for a successful method is
to consider knowledge (KNW) as a process
(Process)KNWP
R8
The application of the method has a necessary and
sufficient condition, i.e. the existence of and
intuitive understanding of each object at all levels
R9
The KNWP has to start from CL0 and to get
confirmation at each step that the subject can
understand the object
R10 Subject improves himself by practicing on CL0
RESULTS OF KNWP
R11

Subject has to use all mind tools to achieve KNW:

i) Reasoning, ii) Imagination, iii) Senses and
iv) Memory
In order to:
Perceive CL0 by intuition
Build KNW starting from CL0

the causal approach of identification of beliefs behind

the paradoxes for each phase results in challenges to
the cause-effect approach.
P6 The CAS model is built for a given set of
objective functions and the metrics of this model is
expected to include risk, which usually is not one of
the intended goals for the user.
P7 Merging deterministic and probabilistic
models forces the modeller and user to have common
statements about combinations of probabilistic (which
are by default uncertain) and deterministic events.
However, this can sometimes lead to conflicting
conclusions with problems of logic inconsistence.

348

P8 Management of risk model leads to managerial/procedural control in order to limit the uncertainty
in the real process of CAS evaluation. However this
action is in itself creating new systematic assumptions
and errors and is shadowing the ones accumulated up
to this phase.
P9 The completion of a nine-cycle phase CAS
and its implementation reveal the need to restart the
process for a better theory.
Second step: defines believes identified to be
the main features of each of the steps
P1 It is assumed that there is a unique definition
for risk and the risk science has a unique and unitary
approach to give all the answers.
P2 It is assumed that the well-established scientific facts showing both random and deterministic data
used in a scientific manner could provide support for
certitude by using risk assessments.
P3 It is assumed that in the case of risk analyses
a scientific method of universal use exists to evaluate
severity of risks by judging them according to their
probability and the outcomes/damages produced.
P4 It is assumed that by using carefully chosen
experience and model results one can derive objective
results proving the validity of results for the given CAS
model.
P5 It is assumed that by using educated guesses
and experiments scientists can find and evaluate any
significant risk due to the objectivity and other specific
features of science.
P6 It is assumed that based on the objectivity of
science and the approach in risk analyses to evaluate
risks against benefits; the results could be used as such
in decision-making process.
P7 It is assumed that by using a scientific method,
which is honest and objective and by using risk reducing measures in all sectors of society (any type of
CAS model) the combined use of well proven tools
in all science of analysis/deterministic and synthesis/probabilistic approaches assures success in CAS
modeling.
P8 It is assumed that science is more procedural
than creative (at least for this type of activity) and the
decisions themselves have to be made by trained staff
and scientists.
P9 It is assumed that science is evolving based on
laws which appear by the transformation of hypotheses
into theories, which become laws and for any CAS, in
this case, if there will be a real risk then the scientists
will find it.
Last step: a set of actions to prevent generation
of paradoxes is defined:
P1 Model diversity of objective functions for
CAS metrics and use hierarchy for its structure looking
for optimum at each Hierarchical level.

P2 Applicability areas of deterministic and

probabilistic parts in the CAS model and in the
decision module to be clearly defined and used.
P3 Use numerical and logical functions as
switches and connectors between the deterministic and
probabilistic part of CAS models and their metrics.
P4 Use special sensitivity analyses phase to
define the sensitivity to unseen/not clearly formulated
assumptions embedded in the model based on their
reflection in the paradoxes. In the screening process
of those issues, use diverse methods not included so
far in CAS model.
P5 Perform a full inventory of identified paradoxes in CAS model and the possible beliefs generating them in order to have a better understanding of the
CAS model bias and define the CAS reference model
for further analysis.
P6 Identify the rules for the post processing of
risk analyses results in order to be used in the decision
making process, by defining the place and desirability
for the user of Risk Informed Decision Makinga
module to be added to the actual results from risk
analyses for CAS models.
P7 The merging action of deterministic and probabilistic approaches in CAS risk models is accompanied by an improved set of logical construction of the
formulations of results and modules merging the two
approaches.
P8 Continuously check the objectivity of the CAS
model through risk managerial actions, including their
potential distortion into the model.
P9 Restart the cycle of modeling even if there is
no user request for it since there is always a need to have
full coherent answers to all paradoxes encountered,
even if they seem of no interest to the user and/or scientific community. The expected effect of the actions
implemented in the CAS risk analysis process was represented and discussed in (Serbanescu 2005a). The
main conclusion of this representation showed (with
examples from Generation IV modeling) the importance of the feedback modeling in CAS risk analyses
in order to assure a convergent set of solutions by identifying and managing mainly the systematic possible
errors.
3

CONCLUSIONS

The paper presents the main issues identified during

the extended use of PRA-Nuc for new non-nuclear
applications. The results concluded so far identify
some common generic aspects, which need to be considered in order to implement properly and/or extend
existing PRA-Nuc models and tools to new applications. More pilot cases and applications are also under
way in order to confirm further the results obtained so
far and/or to identify new issues to be considered.

349

REFERENCES
Colli A & Serbanescu D., 2008. PRA-Type Study Adapted
to the Multi-crystalline Silicon Photovoltaic Cells Manufacture Process, ESREL 2008 under issue.
Descartes R., 1637, Discours de la mthode, Paris,
GarnierFlammarion, 1966, edition. Discourse on
Method (1637), Haimes, Yacov Y., 2004, Risk Modeling, Assessment and Management, 2nd Edition, Wiley &
Sons, New Jersey.
Hansson S.O., 2000, Myths on Risk Talk at the conference
Stockholm thirty years on. Progress achieved and challenges ahead in international environmental co-operation.
Swedish Ministry of the Environment, June 1718, 2000
Royal Institute of Technology, Stockholm.
Howard et al, 1984, Howard R.A. and Matheson J.E., (editors), Readings on the Principles and Applications of
Decision Analysis, 2 volumes (1984), Menlo Park CA:
Strategic Decisions Group.
Jaynes E.T., 2003, Probability ThoeryThe Logic of Science, Cambridge University Press, Cambridge, UK.
Kato, Tosio, 1995. Perturbation Theory for Linear Operators, Springer Verlag, Germany, ISBN 3-540-58661.
Mc Comas, W. 1996, Ten Myths of science: Reexamining
what we know, vol 96, School Science & Mathematics,
01-01-1996, p. 10.
Peirce C.S., 1931, Collected Papers of Charles Sanders
Peirce, 8 vols. Edited by Charles Hartshorne, Paul Weiss,
Arthur Burks (Harvard University Press, Cambridge,
Massachusetts, 19311958, http://www.hup.harvard.edu/
catalog/PEICOA.html

Serbanescu D., 2005a. Some insights on issues related to

specifics of the use of probability, risk, uncertainty and
logic in PRA studies, Int. J. Critical Infrastructures, Vol. 1,
Nos. 2/3, 2005.
Serbanescu D, 2005b. Integrated Risk Assessment, ICRESH
2005, Bombay, India.
Serbanescu, D. & Kirchsteiger C., 2007a. Some methodological aspects on a risk informed support for decisions on
specific complex systems objectives, ICAP 2007.
Serbanescu D., 2007b. Risk Informed Decision Making,
Lecture presented at VALDOC Summer School on Risk
Issues, Smoegen (Sweden), Vol. Karita Research Sweden
(Organiser), JRC PB/2007/IE/5019.
Serbanescu D. &Vetere Arellano A.L. et al., SES RISK a new
method to support decisions on energy supply, ESREL
2008 under issue.
USNRC 1983, PRA Procedures Guide. A guide for
the performance of Probabilistic Risk Assessment for
Nuclear power Plants (1983) USNRC, NUREG/CR2300, February 1.
USNRC 1998, Regulatory Guide 1.174, An approach for
using PRA in Risk Informed Decisions on plant specific
changes to the licensing basis, July 1998.
Smithson, Michel J., 2000, Human judgment and imprecise probabilities, web site of the imprecise probabilities
project http://ippserv.rug.ac.be 19972000 by Michel J.
Smithson and the Imprecise Probabilities Project.

350

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On the usage of weather derivatives in AustriaAn empirical study

M. Bank & R. Wiesner
alpSCentre for Natural Hazard Management, Innsbruck, Austria, University of Innsbruck,
Department of Banking and Finance, Innsbruck, Austria

ABSTRACT: This paper introduces to the fundamental characteristics of weather derivatives and points out the
relevant differences in comparison to classical insurance contracts. Above all, this paper presents the results of a
survey conducted among Austrian companies which aims at investigating the objectives of weather derivatives
usage and at analysing concerns regarding their application. The survey was conducted via face-to-face interviews
among 118 firms from different sectors facing significant weather exposure such as energy and construction
companies, beverage producers and baths. As no other survey has put a focus on weather derivatives so far,
this paper aims at filling a lack of relevant information regarding weather derivative practices. The results will
grant a deeper insight in the risk management practices and the needs of potential costumers. This facilitates
the development of target group specific weather risk management solutions which may enhance the usage of
weather derivates in various industries.

INTRODUCTION

Many industries and accordingly their sales and

income are at risk due to weather fluctuations. It is
estimated that nearly 2030 percent of the U.S. economy is directly affected by the weather (CME 2006).
Other estimations assume that 70 percent of all businesses face weather risk in some way (Jain & Foster
2000). Weather fluctuations still cannot be controlled,
but with weather derivates a relatively new financial
tool has been created, which offers protection against
weather-related risks.
Weather derivatives are believed to have substantial
market potential in weather affected industries but so
far research has primarily been focused on pricing and
valuation issues. Hence, it is the aim of this paper
to investigate reasons and concerns regarding the use
of weather derivatives as this topic was not covered
before. However, this is of fundamental interest for
the application of these instruments.
The remainder of this paper is structured as follows. Section two provides the theoretical background
of weather derivatives and points out the main differences in comparison to classical insurance contracts.
The third section focuses on the descriptive analysis of the survey results regarding risk management
practices and weather derivatives usage among Austrian companies. The conclusions are presented in
section four.

2.1

THEORETICAL BACKGROUND
Weather risk and weather exposure

Weather risks refer to weather events which affect the

revenues and earnings of a company in a negative
way. Type and quantity of relevant weather parameters depend on the business area and can include
events such as temperature, precipitation or humidity
(Schirm 2001). These kinds of weather characteristics
are called noncatastrophic events which typically
have a high-frequency but low severity. They are contrasted with catastrophe-related low frequencyhigh
severity risks, e.g. hurricanes, tornados, etc. Weather
derivates have been developed to facilitate protection against profit impacts given adverse weather
conditions but not for property or catastrophic risk
protection (Clemmons 2002).
Weather risks belong to the category of operational
risks. In particular they refer to the danger of losses
due to external events, as weather conditions cannot
be influenced by an enterprise. Furthermore, weather
risks are considered to be volumetric risks which can
affect both supply as well as demand.
The sensitivity to weather conditions can be defined
as weather exposure. It quantifies the dependency of
operating figures on adverse weather conditions and
is the first condition for a comprehensive weather
risk management. The determination of the weather

351

exposure usually requires a detailed analysis of the

companys data and potential influencing factors on a
companys business success. Weather derivatives then
serve as an efficient instrument to reduce weather risks
and respectively weather exposure, and hence minimize their potential negative influence on companys
success. A systematic weather risk management process therefore assures competitiveness and presumably
causes positive effects (Jewson & Brix 2005):
Reduces the year-to-year volatility of profits
Profit smoothing leads to constant year-to-year tax
burdens
Low volatility in profits often reduces refinancing
costs
In a listed company low volatility in profits usually
translates into low share price volatility, and less
volatile shares are valued more highly
Low volatility in profits reduces the risk of
bankruptcy and financial distress
A sample of weather risk and corresponding financial risks faced by various industries is given in
Table 1.
Table 1. Illustrative links between industries, weather type
and financial risks.
Risk holder

Weather type Risk

Energy industry Temperature

Energy
consumers
Beverage
producers
Construction
companies
Ski resorts
Agricultural
industry
Municipal
governments
Road Salt
companies
Hydro-electric
power
generation

Lower sales during warm

winters or cool
summers
Temperature Higher heating/cooling
costs during cold
winters and hot
summers
Temperature Lower sales during
cool summers
Temperature/ Delays in meeting
Snowfall
schedules during
periods of poor
weather
Snowfall
Lower revenue during
winters with belowaverage snowfall
Temperature/ Significant crop losses
Rainfall
due to extreme
temperatures or rainfall
Snowfall
Higher snow removal
costs during winters
with above-average
snowfall
Snowfall
Lower revenues during
low snowfall winters
Precipitation Lower revenue during
periods of drought

Source: Climetrix 2006.

2.2 Characteristics of weather derivatives

A weather derivative is a financial contract between
two parties with a payoff which is derived from the
development of an underlying index. They differ from
common financial derivatives as their underlying is
not a tradable asset but meteorological variables.
As we already noted weather affects different companies in different ways. In order to hedge these
different types of risk, weather derivatives can be based
on a variety of weather variables. These include temperature, wind, rain, snow or sunshine hours. The
only condition is that the weather variables can be
transformed into an index and that they are measured
objectively without the influence of any counterparties
(Cao et al. 2003, Dischel 2002).
The pay-off of these financial instruments is derived
from a weather index, not from the actual amount of
money lost due to adverse weather conditions. Therefore, it is unlikely that the pay-off will compensate
the exact amount of money lost. The potential difference between an actual loss and the received pay-off
is known as basis risk. Generally, basis risk is reduced
when the companys financial loss is highly correlated
with the weather and when contracts of optimum size,
structure and location are used for hedging (Jewson &
Brix 2005).
Typically, a standard weather derivative contract is
defined by following attributes:
Contract period: defines a start date and an end date,
usually a month or a season
Measurement station: preferably close to the companys location to reduce geographical basis risk
Weather variable: corresponding to weather exposure and hedging needs
Underlying: index which aggregates the weather
variable over the contract period
Pay-off function: determines the cash-flows of the
derivative
Strike Level: value of the index at which the pay-off
changes from zero to a non-zero value
Tick and Tick Value: defines how much the pay-off
changes per unit of the index
The pay-off functions of weather derivates rely on
the pay-off functions of traditional derivative instruments, e.g. options, collars, straddles or swaps (Hull
2006). Because weather is not a tradable asset, the
exercise of weather derivatives always results in cash
settlement. In addition, contracts may involve financial limits in the maximum pay-off.
2.3 Weather derivatives in comparison
to weather insurance contracts
Traditional insurance companies already offer protection against weather risks. So, is there really a

352

need for weather derivatives or can weather risks be

hedged by means of traditional insurance contracts?
To answer this question the key characteristics of
weather derivatives and weather contracts have to be
contrasted.
The main difference between the two instruments
is that the holder of an insurance contract has to
prove that he actually suffered a financial loss due
to adverse weather conditions in order to be compensated. Hence, weather insurance contracts consist of
two triggers: a weather event plus a verifiable financial
loss. If the insured is not able to prove this, he will not
receive payments from the contract (Becker & Bracht
1999). In contrast, the payments of weather derivatives
solely rely on the weather index value, an actual loss
does not have to be demonstrated. Weather derivates
offer the option holder the advantage of receiving
prompt compensation without the risk of potentially
needing long lasting proof of the actual loss (Rasp
2002).
In conjunction with this feature another distinctive
criterion has to be considered. The buyer of a weather
derivate does not need to have any weather sensitivity
or intention to protect himself against adverse weather
conditions, i.e. he does not need to show an insurable interest. On the other hand, insurance contracts
are based on the concept of insurable interest, which
means a direct relationship between the insured risk
and the insurance purchaser has to exist, i.e. a natural
exposure is required (Culp 2002). Weather derivatives
can also be bought for mere speculation (Alaton et al.
2002).
An additional important feature of weather
derivates compared to insurance is that two counterparties with opposed risk exposures can enter into a
contract to hedge each others risk, e.g. via a swap
structure. This is usually not possible in the insurance
market (Alaton et al. 2002). Furthermore, swap contracts allow protection at no upfront costs, whereas a
premium must always be paid for insurance contracts
(Becker & Bracht 1999).
The typical usage of weather derivates and insurance contracts can be seen as another difference
between the two. As stated before, weather derivates
are mainly constructed for protection against highfrequency/low-severity risks. Insurance contracts usually refer to risks of extreme or catastrophic nature but
with low occurrence probability (Alaton et al. 2002).
A possible explanation for these typical application
fields can be found in the instrument specific basis
risk. The pay-off of weather derivates shows no dependency on the actual loss which results in a natural
basis risk and implies the risk of an insufficient compensation in extreme events. Insurance solutions are
preferred in this situation as they are indemnity based
and do not contain basis risk. In contrast, insurance
contracts do not function well with normal, frequent

weather risks as all losses have to be proven. This is

time-consuming and costly. The straightforward compensation process of weather derivates justifies the
acceptance of basis risk. Furthermore, the existence
of basis risk has the beneficial effect that weather
derivates are lower-priced than insurance contracts.
Other aspects in which derivatives and insurance
differ include legal, tax and regulatory issues. A
comprehensive discussion of these topics is given in
Rasp (2002), Edwards (2002), Ali (2004) and Kramer
(2006).
To summarize, weather derivates and insurance
contracts show significant differences but are not to be
considered as exclusively competing concepts. Rather
they should be used as complementary instruments
in weather risk management as both offer application
specific advantages.
3
3.1

THE SURVEY
Research methodology and respondents
profile

The literature lacks relevant information regarding

actual weather derivatives usage so a main objective of
the survey was to fill this research gap. It was also considered useful to investigate the broader issues related
to that topic. Therefore, questions regarding general
risk management practices were included in the questionnaire. Following research ideas served as starting
point:
How do companies perceive climatological risks?
What reasons for the use and concerns regarding
weather derivatives are named by active users?
Which are the main factors that hinder companies
to use weather derivatives?
The survey was conducted via face-to-face interviews among 118 Austrian firms from different sectors
facing significant weather exposure such as energy
and construction companies, beverage producers and
baths. The breakdown of respondents per sector is as
follows: 37.3% energy, 25.4% baths, 20.3% beverages
and 16.9% construction sector. The sample consists to
3.4% of micro-, 26.4% of small-, 23.1% of mediumand 13.2% of large sized enterprises. In comparison to
the average percentage of enterprise sizes in Austria
a relatively large part of the sample is represented by
medium- and large-sized companies. This is mainly
attributed to the energy and construction sector. The
main findings of the survey will be highlighted in this
paper.
3.2

Weather risk and weather exposure

First, respondents were asked to specify on a five point

scale (1 for no weather risk and 5 for very high weather

353

respondents face a high exposure as up to 50% of their

revenues are affected. Finally, a large group of respondents (21%) face a very high exposure, as weather
risks threaten more than 50% of their revenues. The
high proportion of firms with medium, high and very
high exposure indicates that there is definitely a need
for a sophisticated weather risk management (Fig. 2).

no

low
12%
5%
very high

49%
21%
moderate

3.3
13%

high

Figure 1.

Weather risk faced by respondents.

> 50 Percent

21%

41 - 50 Percent

11,5%

31 - 40 Percent

10%

21 - 30 Percent

11%

11 - 20 Percent

14%

1 - 10 Percent

26%

unknown

7%

0%

Figure 2.

5%

10%

15%

20%

25%

Weather exposure faced by respondents.

risk) to what extent their company faces weather risk.

Almost the half of firms (49.2%) stated that they face
very high weather risks and another 13% indicated
high weather risks. The results show that the majority of firms in the analyzed sectors are significantly
exposed to weather risks. Only 12% of firms said that
they face no weather risks at all. Figure 1 presents a
distribution of the answer by weather risk category.
After the majority of firms face considerable
weather risks we were interested to analyze whether
they can quantify their exposure. Only a more detailed
knowledge of the weather exposure facilitates an effective risk management process. Therefore, the companies were asked to indicate the percentage of operating
revenues at risk due to adverse weather conditions.
Only a low number of 7% of the respondents stated
that the exposure is unknown, which highlights the
importance of weather risk in these business sectors.
26% of the firms have a moderate weather exposure
influencing up to 10% of their revenues. Further, 14%
state an exposure between 1120% of their revenues.
Aggregated 21% face a medium weather exposure
affecting up to 40% of their revenues. 11.5% of

Risk perception and risk management

Given the relatively high exposures across the sample,

it seems interesting to analyse whether an institutionalised risk management process exists. This topic was
covered in the questionnaire by asking which department or person is conducting the risk analysis and
within which frequency risk analysis is carried out.
The results show, that in the majority of the cases
(62%) the management board is performing these analyses and that special departments are assigned only
to some extent (15%). On the one hand, it can be
positive that the executives are in charge of the risk
management process as it signals its importance. On
the other hand, specialised departments could focus
and explore in more detail the potential risks as general
managers probably do. Therefore, designated departments or employees seem to be beneficial from the
companys perspective.
Another question of the survey asked respondents to
indicate the frequency at which they analyse their risks.
This could serve as indicator how important risks are
perceived and whether a functioning risk management
exists. 58% of respondents to this questions stated that
they never assess their risks. Nearly 12% indicate to
analyse risks at least sometimes. 30.5% of respondents
admit regular risk analysis.
These findings are worrying if we consider the significant weather risks and exposures stated by the
companies. Therefore, it was also analyzed how the
specified exposure corresponds with the regularity of
risk analysis. It was expected that the frequency of risk
analysis declines with diminishing weather exposure,
i.e. that firms with a high exposure are performing
a regular risk assessment in opposite to low exposure firms. The relationships between the level of
weather exposure and the frequency of risk analysis
are displayed in Figure 3.
As expected, a large number of companies with
low exposure are never performing risk analysis at
all (52%). 33% of respondents indicate regularly risk
assessments and other 15% state to analyse risks at
least sometimes. These findings are in line with the relatively low exposure. In contrast to this, the frequency
of risk analysis does not improve with increasing exposure, instead the proportion of firms with regular risk
assessments even declines. Figure 3 displays, that a
large amount of firms facing medium or very high
exposures are not controlling their risks on a regular

354

> 50 Percent

68%

Exposure

41 - 50 Percent

25%

32%

17%

31 - 40 Percent

58%

60%

21 - 30 Percent

10%

55%

27%

60%

11 - 20 Percent

1 - 10 Percent

20%

15%

40%
Never

18%

20%

52%

0%

30%

60%
Sometimes

20%

33%

80%

100%

Regularly

Figure 3. Cross tablefrequency of risk analysis and

weather exposure.

basis. 55% to 68% of respondents never perform risk

analysis. Also the percentage of firms with risk assessments on a regular basis remains stable in-between
18% to 32% across the different exposure classes. An
exception seem to be respondents belonging to the
exposure class of 41% to 50%. The majority of this
group states to asses risk regularly (58%) or at least
sometimes (17%). Only 25% of these firms do not
asses their risks profile.
A reduction of exposure classes into three categories corresponding to low (120%), medium
(2140%) and high (>40%) exposure yields some
improvement. 41% of respondents belonging to the
high exposure category indicate that they regularly
analyse risks. But, still 53% of these firms do not
evaluate risks. The proportion of risk analysis in low
and medium exposure classes remains the same.
Overall, these findings suggest that the majority of
firms do not have an institutionalised risk management as a significant percentage of respondents do
not analyse their risks at all despite a considerable
weather exposure. Unexpectedly, the results remain
relatively stable across exposure classes meaning that
also the majority of firms with high weather exposure
do not evaluate the riskiness of their business regularly.
Accordingly, the results imply that firms underestimate or misinterpret their actual weather exposure and
are presumably not able to control their weather related
risk efficiently.

3.4

Reasons and concerns regarding the use

of weather derivatives

One objective of this study was to gain a deeper insight

into the use of weather derivatives. The survey asked
firms to indicate the importance of different factors
in the decision of using weather derivatives. Eleven
reasons were listed, and users were asked to rank them
on a five-point scale (1 for not important and 5 for
very important).

Three respondents actually use weather derivatives,

two belonging to the energy and one to the construction sector. The small number of users was expected
due to the fact that weather derivatives are a relatively new financial hedging tool. Besides that, todays
weather derivative contracts are usually directed at
large firms as contract volumes need to be relatively
high, given the fixed costs incurred in transactions, to
allow reasonable premiums.
Respondents primary reason to use weather derivatives is earnings stabilization. Two respondents rank
this factor as very important, respectively one respondent as important. Another important factor in the
decision of using weather derivatives is the absence
of an actual loss proof to receive a payout. Accordingly, the transparent and unproblematic clearing of
weather derivatives is named as major factor as compensation only relies on the development of objective
weather parameters. Moreover, respondents indicate
that traditional insurance products do not offer comparable protection regarding weather related risks or
are not available.
The results show that classical risk management
aspects, such as the reduction of earnings volatility,
are the primary reasons for the application of weather
derivatives. Besides that, instrument specific advantages over traditional insurance are also perceived
as important. Respondents also indicate that weather
derivatives are not used for mere speculation. On
the contrary, users of weather derivatives seem to be
aware of their exposure and are searching for specific
hedging tools to protect their profits.
The major concerns regarding the use of weather
derivatives were also ascertained to understand possible reasons for the limited use of weather derivatives.
A list of eleven possible concerns was presented in the
questionnaire to be ranked by respondents again on a
five-point scale (1 for no concern and 5 for very high
concern).
Tax, legal and accounting issues appear to be
the companies major concerns as all users indicate
very important, important or at least moderate concern. This can be attributed to the fact that weather
derivatives are relatively new financial instruments
which still results in some uncertainty regarding their
accounting and legal treatment (Mojuy 2007). Difficulty of quantifying the firms exposure and a general
lack of knowledge in derivatives are also major concerns to the respondents. Further, difficulties in evaluating the hedge results are indicated. These results
suggest clearly that, where companies have concerns
regarding weather derivatives, this arises to a great
extent from a lack of their expertise.
Also, the potentially negative image of derivatives
is a serious concern for users. Weather derivatives
are often presented in the media as bets or gambling
against the weather instead of focusing on their role in

355

risk management. This kind of press coverage could

also contribute to the fact that potential end-users are
reluctant in their decision to apply weather derivatives.
Interestingly, basis risk is not concerned as overly
important by respondents. One companies ranks this
fact as important, the others state moderate and no
importance. These results are to some extent in contradiction to the expectations as literature often names
basis risk as a major problem in the application
of weather derivates. But, as these respondents are
actually using weather derivatives, basis risk seems
acceptable or they simply have not had any problems
regarding this issue.
Overall, the results imply that lack of knowledge
seems to be the firms major concern by the use of
weather derivatives.

Mean
Mode

3
2

3,56

3,52

3,21

2,99

2,78

2,52

2,5

2,42

2,26

2,23

2,15

2,12

1,95

Lack of expertise in derivatives

Instrument does not fit firms needs

Exposure effectively managed by other means

Difficulty of quantifying the firms exposure

Difficulty of pricing and valuing derivatives

Insufficient weather exposure

Costs of hedging exceed the expected benefits

Uncertainty about accounting treatment

Difficulty of evaluating hedge results

Uncertainty about tax and legal treatment

Company policy not to use derivatives

Concerns about perception of derivative use

3,62

We ather derivatives unknown

3,92

No benefits expected

1
0

Reasons for not using weather derivatives

In the following, the analysis will focus on non-users

and their reasons why they choose not to apply them.
Therefore, non-users were asked to rank the importance of different factors regarding the decision against
the usage of weather derivatives. A total of 16 reasons were listed and ranked by the respondents on a
five-point scale (1 for not important and 5 for very
important). The responses to this question are shown
in Figure 4.
The most important reason for the decision against
weather derivatives is that the respondents simply
never considered using them (54% of answers in the
category very important/ mean 3.92/ mode 5). This
reason is closely followed by the reason that firms do
not know weather derivatives (45 % very important/
mean 3.56/ mode 5). Summarized 7080% of respondents rank this reason as moderate, important or very
important. These findings suggest that the majority of
potential user simply does not know that an instrument
for the hedging of weather related risks exists.
Respondents also indicate to a large proportion
that they do not expect benefits from weather derivatives (44% very important/ mean 3.62/ mode 5). The
perception of an insufficient benefit may also be
attributed to the limited knowledge of weather derivatives. This seems to be confirmed as lack of expertise
in derivatives is ranked in fourth position (39% very
important/ mean 3.52/ mode 5). Another concern
receiving some attention was that weather derivatives
may not fit the firms needs (Mean 3.21/ mode 5).
Difficulties of quantifying the weather exposure
and valuing weather derivatives were also ranked
as relatively important. They are related to limited
expertise. Some respondents also indicate that they
potentially face a too small weather exposure.
All other reasons are perceived as relatively unimportant by respondents given means below 3. Also the

Median

Never considered using weather derivatives

3.5

Figure 4.

Reasons for not using weather derivatives.

reason image of derivatives, which was important for weather derivative users, has a subordinate
position. A potentially negative image of weather
derivatives or their promotion as bets on the weather
seem to have no major influence on the decision not
to use them.
In summary, results indicate that the majority of
companies are not aware of weather derivatives as
risk management tool, as they either do not know the
instrument or never considered using them. Another
important factor seems to be a general lack of knowledge regarding derivatives and their application. This
includes the instrument itself but also areas such as
the quantification of weather exposure. Because of a
lack of knowledge, companies seem to be also very
sceptical in appreciating potential benefits of weather
derivates.
The findings imply that there are some predominant
reasons which are mainly responsible in the decision
of not using weather derivatives. Therefore, a factor
analysis was conducted to identify the significant factors. Factors which strongly correlate with each other
will be reduced to one factor. A Kaiser-Meyer-Olkin
test yields 0.84. It is above the eligible level of 0.8 and
therefore qualifies the sample as suitable for factor
analysis.

356

Table 2.

identified factors seem plausible given the findings

before. As weather derivatives are relatively new
instruments and many firms never had experience
with derivatives before, they lack expertise in evaluating and handling these instruments. Additionally,
as nearly 60% of companies do not analyze their
risks on a regular basis it seems reasonable that
they are not able to quantify their exposure correctly.
Further, they probably cannot appraise the potential
benefits of weather derivatives as they hardly know
the instruments. This leads to the third factor that
weather derivatives are widely unknown. It is entirely
possible that factor 1 and 2 also arise from a general unawareness of hedging measures with weather
derivatives. Therefore, lack of knowledge regarding
weather derivatives and their application seems to be
the most important reason for not using them so far.
Taken this into consideration, weather derivates
should be actively promoted in the public and to
potential end users to increase the awareness level.
Furthermore, the functioning and characteristics of
weather derivatives as risk management tool should
be highlighted to demonstrate their benefits and to
build up some basic expertise within the firms. The
promotion of risk analysis may also contribute to a
success of weather derivatives as the firms get aware
of so far neglected weather risks. Finally, if potential
users know the instrument and its role in weather risk
management they will more likely apply them.

Rotated component matrix.

Component
1

l
m
n
k
o
j
d
e
a
p
f
h
i
c
b

Difficulty of evaluating
hedge results
Uncertainty about
accounting treatment
Uncertainty about tax and
legal treatment
Difficulty of pricing and
valuing derivatives
Concerns about perception
of derivative use
Difficulty of quantifying
the firms exposure
No benefits expected
Instrument does not fit
firms needs
Insufficient weather
exposure
Company policy not to
use derivatives
Exposure effectively
managed by other means
Costs of hedging exceed
the expected benefits
Lack of expertise in
derivatives
Never considered using
weather derivatives
Weather derivatives
unknown

,902
,872
,869
,833
,678
,551

,455
,742
,716
,662
,641
,639

,516

,616
,820
,752

3.6

,743

A factor analysis yields three significant factors

which explain up to 66% of the total variance. Using
the factor loadings given in the rotated component
matrix the questions can be assigned to the three
extracted factors. The factor loadings shown in Table 2
are already sorted for easier interpretation as well as
factor loadings below 0.4 are not displayed.
Factor 1 appears to be related to lack of knowledge (l, m, n, k, j and less so for o) as it mainly
encloses reasons in the context of pricing, valuation,
hedging and taxation uncertainties. Factor 2 could be
named no benefits expected (d, e, f, h, a and less
so for p) as the containing reasons focus on the fact
that there might be better alternatives or that weather
derivatives do not fit the firms needs. Finally, factor 3 includes questions regarding unawareness of
weather derivatives in general (c, b, i) which leads
to its assignment of instrument unknown.
The factor analysis indicates that the reason not to
use weather derivatives can be mainly attributed to
three factors: lack of knowledge in derivatives, no
benefits expected and instrument unknown. The

Potential usage of weather risk management

instruments

Expecting the low number of active weather derivative users we also asked the respondents to indicate
whether they can imagine using weather derivatives
or comparable instruments in the future. The results
show that roughly one-fourth (26.7%) of firms are
generally willing to apply weather derivatives. Given
the actual state this proportion indicates a significant
market potential for weather derivatives.
The question demonstrates that many respondents
are open-minded about weather risk management with
weather derivatives. Of course, it has to be investigated
which firms finally can use these instruments as basis
risk and large contract sizes may have unfavourable
impacts. High cost for small weather derivative transactions could be reduced via bundling schemes in
which different companies located in one region are
sharing a derivative contract to reduce the costs and to
achieve reasonable contract sizes.
Another possibility is the integration of weather
derivatives in loans. This structure enables a company
to enter into a loan agreement with a higher interest
rate that already includes the weather derivatives premium which the bank pays to the counterparty. In case
of an adverse weather event, the company only pays a

357

fraction or nothing of the usual loan due, thus receiving a financial alleviation in an economical critical
situation. Moreover, weather-indexed loans would be
less likely to default which is also favourable for the
bank itself as it strengthens the banks portfolio and
risk profile (Hess & Syroka 2005). Weather-indexed
loans seem especially applicable for SMEs as they are
financed to a large extent by loans. This facilitates the
access for banks and offers considerable cross selling
potential. The potential use of a weather-linked credit
by companies was tested in question 24.
The results show that 17% of respondents can imagine to use a weather-indexed loan. In comparison to
question 23, it does not seem as attractive as weather
derivatives for potential users. On the one hand, this
could be attributed to the more complex product structure. On the other hand, potential users could simply
prefer a stand-alone product instead of buying some
sort of bundling schemes. Further, the comments of
respondents on questions 23 and 24 indicate that there
is a general interest in weather derivatives but additional information is requested. This highlights again
that instrument awareness has to be improved as well
as lack of knowledge has to be reduced.
4

from a lack of knowledge regarding derivatives.

Furthermore, the factors no benefits expected and
instrument unknown seem to restrict the usage of
weather derivatives.
Considering these findings, the main challenge
will be to increase the acceptance, awareness and
knowledge towards weather risk management and corresponding tools. The results show, that a significant
proportion of companies not using weather derivatives
so far may begin to use them as knowledge of these
instrument increases.
Finally, the survey confirmed that many firms
are interested in weather risk management, but the
process is still in its infancy. Firms should put
more emphasis on investigating their weather exposure and potential hedging measures to reduce the
impacts of adverse weather conditions on operating
figures. Especially weather derivatives offer the possibility to manage weather risks actively and flexible.
On the other hand, todays large contract sizes hinder mainly SMEs from the application of weather
derivatives. Hence, future research will focus on contract structuring issues, such as bundling schemes or
weather-indexed loans, to facilitate the use of weather
derivatives.

CONCLUSION
REFERENCES

The main objectives of this work were to give an

introduction to weather derivatives and to investigate weather derivative practices. A survey, conducted
among 118 Austrian companies, confirmed that the
respondents face a significant weather exposure and
that they are also to a large extent aware of it. On
the other hand, the survey revealed that the majority
of firms are lacking a sophisticated risk management
concept as they hardly control their risks on a regular basis. This seems worrying given the general
knowledge of their exposure.
Respondents already using weather derivatives
indicated that they generally consider them as a
useful tool in weather risk management. Primarily,
weather derivatives are applied for earnings stabilization. Instrument specific advantages, such as the lack
of damage proofs and prompt compensation payments,
are emphasized. Typical concerns regarding the usage
of weather derivatives are pricing, valuation, tax and
legal issues. Further, the image of weather derivatives
is perceived relatively critically. The study implies that
also companies already using weather derivatives may
need some additional assistance in the weather risk
management process.
The relevance of different reasons in the decision of companies not to use weather derivatives was
also investigated. A factor analysis on the survey
results produced evidence that the reluctance of firms
to use weather derivatives arises to a great extent

Ali, P.U. 2004. The Legal Characterization of Weather

Derivatives. Journal of Alternative Investments 7(2):
7579.
Becker, H.A. & Bracht, A. 1999. Katastrophen- und
Wetterderivate: Finanzinnovationen auf der Basis von
Naturkatastrophen und Wettererscheinungen. Wien.
Cao, M. et. al. 2003. Weather derivatives: A new class
of financial instruments. www.rotman.utoronto.ca/wei/
research/JAI.pdf, 16.06.2007.
Clemmons, L. 2002. Introduction to Weather Risk Management. In Banks, E. (ed), Weather risk management:
markets, products, and applications: 313. Basingstoke.
Climetrix 2006. ClimetrixWeather Derivatives Software.
http://www.climetrix.com, 15.08.2006.
CME 2006. CME Weather Products, http://www.cme.com/
trading/prd/weather/index14270.html, 09.08.2006.
Culp, C.L. 2002. The ART of risk management. New York.
Dischel, R.S. 2002. Climate Risk and the Weather Market.
London.
Edwards, S. 2002. Accounting and Tax Treatment. In Banks,
E. (ed), Weather risk management: markets, products, and
applications: 246261. Basingstoke.
Hess, U. & Syroka. J. 2005. Weather-based insurance in
Southern Africa. The International Bank on Reconstruction and Development, Washington.
Hull, J. 2006. Options, futures, & other derivatives, Upper
Saddle River, NJ.
Jain, G. & Foster, D. 2000. Weather RiskBeyond
Energy. Weather Risk supplement to Risk magazine August
2000, http://www.financewise.com/public/edit/energy/
weather00/wthr00-beyondenergy.htm, 10.08.2006.

358

Jewson, S. & Brix, A. 2005. Weather derivative valuation:

the meteorological, statistical, financial and mathematical foundations. Cambridge.
Kramer, A.S. 2006. Critical Distinctions between Weather
Derivatives and Insurance. In Culp, C.L. (ed). Structured
finance and insurance: the ART of managing capital and
risk: 639652. Hoboken.

Mojuy, B. 2007. Does the hot dog vendor lose out?

International Financial Law Review 26(11): 2729.
Rasp, A. 2002. Legal and Regulatory Issues. In Banks, E.
(ed), Weather risk management: markets, products, and
applications: 224245. Basingstoke.
Schirm, A. 2001. Wetterderivate: Einsatzmglichkeiten und
Bewertung. Working Paper, University of Mannheim.

359

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Precaution in practice? The case of nanomaterial industry

H. Kastenholz & A. Helland
Technology and Society Lab, Empa, St. Gallen, Switzerland

M. Siegrist
Institute of Environmental Decisions, Consumer Behavior, ETH, Zurich, Switzerland

ABSTRACT: Nanoparticulate materials (NPM) pose many new questions on risk assessment that are not
completely answered and concerns have been raised of their potential toxicity and life cycle impacts. Voluntary
industrial initiatives have been often proposed as one of the most promising ways to reduce potential negative
impacts on human health and the environment from nanomaterials. We present a study which had the purpose to
investigate how NPM industry in general perceives precaution, responsibility and regulations, how they approach
risk assessment in terms of internal procedures, and how they assess their own performance. The survey shows
that industry does not convey a clear opinion on responsibility and regulatory action, and that the majority of
companies do not have standardized procedures for changes in production technology, input substitution, process
redesign, and final product reformulation as a result of a risk assessment. A clear majority of the companies
found their existing routines regarding these procedures to be sufficient.

INTRODUCTION

Nanotechnologythe art and science of manipulating

matter at the nanoscale to create new and unique materials and productsis a recently emerging and rapidly
growing field whose dynamics and prospects pose
many challenges to scientists and engineers as well
to society at large. This technology has a major potential to generate new products with numerous benefits.
There are already many products containing nanoparticulate materials on the market today, and there are
expectations for more applications ranging from lightweight materials, drug-delivery systems, and catalytic
converters to food, cosmetics and leisure products.
However, the unique properties of nanomaterials
and the rapid development of nanomaterial based products have also brought up many concerns over its
consequences for the human and environmental health
(Nowack & Bucheli 2007; Nel et al. 2006; Oberdrster
et al. 2006). Especially the impact of intentionally produced nanomaterials with targeted properties fulfilling
specific functions is recently in discussion. It has
been shown that some nanomaterials may have damage
potential if they are exposed to humans or the environment. Against this backdrop, different stakeholders
have called for action to ensure the workplace, consumer and environmental safety of NPM production
and products (Helland et al. 2006).
NPM may fall under different regulations depending on the application, but the regulations are currently

found to be inadequate in dealing with the novel properties of NPM (Davis 2007). Therefore there is an
ongoing discussion regarding assessing and managing
the risks derived from NPM properties, the methodological challenges involved, and the data needed
for conducting such risk assessments (EPA 2007;
Morgan 2005). However, given that NPM may cause
harm and that there are currently no regulations that
take the specific properties on NPM into account,
the responsibility for safe production and products is
mostly left with industry. Risk assessment procedures
and precautionary measures initiated by industry are
therefore vital to managing the environmental health
and safety of nanomaterials (Som et al. 2004). These
aspects have been reflected in previous cases involving
other emerging technologies as well.
The objectives of this paper were to investigate
how NPM industry in general perceives precaution,
responsibility and regulation, how they approach risk
assessment in terms of internal procedures, and how
they assess their own performance. To this end, we
first introduce the survey methodology. Secondly the
results of the survey are presented and finally we will
discuss the implications of the results.
2

METHODS

In order to investigate how companies deal with the

risks of nanomaterials, we conducted a written survey.

361

The data was collected in Germany and Switzerland

between December 2005 and February 2006. The
sample consisted of a total of 135 companies, 48 of
them from Switzerland and 87 from Germany. The
companies were identified through websites, literature reviews and personal contacts. A prerequisite for
company selection was that companies should have
NPM-based products available on the market. A total
of 40 companies filled out the questionnaire, which
represents a response rate of 29.6%. Before sending
out the questionnaire, we contacted each company
by phone and requested the person in charge of risk
assessment procedures to fill out the questions.
The two largest industrial sectors were chemicals and materials and consumer goods, and the
most common application fields for NPM within these
industrial sectors were coating and thin films for
different materials (e.g. glass, wood and textile), medical applications and electronic products. Twenty-five
companies had less than 100 employees, 8 companies
had between 100 and 1000 employees, 6 companies
had more than 1000 employees and 1 company did
not answer this question. Fourteen companies reported
that they were primary producers of NPM, 21 companies were downstream users working with NPM
purchased for their applications, 2 companies produced and purchased NPM for their applications, and
3 companies did not answer this question. In order
to address the objectives we developed a number of
questions regarding (1) industrial interpretations of the
precautionary principle, life cycle responsibility and
regulation, (2) industrial procedures, and (3) industrys own assessment of these respective procedures
and areas.
3

RESULTS

Regarding the industrial interpretations of the precautionary principle, a clear majority of the responders
found that all emissions should be kept As Low As
Reasonably Achievable (ALARA) and that measures
should be taken if specific criteria of potential irreversibility are fulfilled. However no majority opinion
was found regarding whether the burden of proof
should be on the proposing actor. A principle component analysis identified only one factor that explained
a total variance of approximately 87.55% and one can
therefore conclude that the respondents answered all
the questions in a consistent manner.
In the production phase of the life cycle, most of the
companies felt responsible for potential environmental
health impacts that may occur, whereas 2 thought this
responsibility should be shared with the government.
In the use phase, 24 companies opined that the responsibility should be borne mainly by industry, whereas
only 8 thought the government or the consumer should

be responsible and 4 thought it should be shared with

other actors. In the end-of-life, the industrial opinion
is divided: 17 companies argued that the responsibility should be taken by the industry alone, 10 thought
the government or the consumer should be responsible, whereas 9 thought it should be shared with other
actors. The responsibility can be seen as externalized
throughout the life cycle stages.
In research and development 19 companies considered that no regulations are needed, whereas 17
found that industrial standards should be established
and 2 companies preferred governmental standards. In
the production stage a clear majority of the companies
found that NPM should be regulated by industrial standards, whereas in the usage stage and the end-of-life
stages the company opinions were divided as to how
to best regulate this area. In general, the industry can
therefore be seen as divided on whether industrial standards or governmental regulations comprise the most
appropriate form of regulation.
On the question: If risk assessment reveals a lack
of knowledge and there is a possibility of harmful
effects, does your company have standardized criteria or procedures for a) raw material substitution
b) final product reformation c) process change?
The majority, 21 companies, did not indicate any
standardized procedures following a risk assessment.
Eleven companies were found to be promoting risk
research, whereas 21 companies, a majority, did
not promote any such research. Furthermore, a total
of 15 companies had procedures involving different
stakeholder concerns in product development (e.g.
public consultations, hearings, scenario analysis).
In general, companies were quite satisfied with
their current performance. All companies found that
best available information acquisition was important,
although the routines used for it could be improved.
However, it was particularly in terms of promoting
risk research, sharing knowledge with other organizations such as safety knowledge, and risk assessment
methodology, that the companies saw the greatest
potential for improvement.
4

DISCUSSION

Industries perceive themselves as clearly responsible

for potential impacts to human health and environment in the research, development and production
stages, but this responsibility is gradually being externalized to other stakeholders throughout the life cycle.
This clear acknowledgement of industrial responsibility is in sharp contrast to a less uniform perception
of regulation, where 38 companies wanted industrial standards or no regulations in the production
stage, 25 companies wanted the same in the usage
phase and 18 companies in the end-of-life. The

362

combination of increasingly industrial externalization

of responsibility and regulations throughout the life
cycle may be problematic as current regulations generally do not cover NPM. Thus we may have a situation
where there is a vacuum concerning the question who
should monitor that NPM are being developed and
used in a safe manner.
The prevention of the possibility of harm arising
from production or products by eliminating problems at the source may involve changes in production
technology, input substitution, process redesign and
re-engineering, and final product design and reformulation. About two thirds of the companies did not have
such procedures in place, but at the same time most
companies assessed their existing routines regarding
these respective procedures and found them to be
sufficient with only a few companies thinking these
procedures could be improved. This may imply that
there are no high incentives for considering alternative technology options and thus a business-as-usual
scenario seems the most likely path that industry would
chose in this respect. One exception was the procedure
of risk assessment. It may be argued that companies need assistance in this respect as a majority of
companies do not have risk assessment procedures
in place, but do express a wish to improve their
performance. It is obvious that the current lack of
established risk assessment and management frameworks for NPM is problematic for the companies
(Reijnders 2006).
The awareness among key actors in a company
(managers, engineers, researchers, safety experts,
product designers etc.) is heavily influenced by social
factors such as communication and cooperation with
other stakeholders in addition to the existing corporate
culture, the core values of the company (Ashford &
Zwetsloot 2000). Increasing awareness among companies key actors through training and outreach
activities can therefore comprise effective measures
to improve the safety culture of the companies. The
surrounding network of agents may therefore provide important contributions to industry for governing
risks, but nonetheless greater industry contributions to
the public knowledge base on NPM have been called
for (Helland & Kastenholz 2007). We may conclude
from our results that most companies have an established network in which risk information is exchanged.
This information exchange usually takes place among
companies and between companies and universities,
although very few actively involve themselves in
funding university research.
The majority of companies also found that information exchange was an area in which there was room for
improvement. A further aspect resulting from insufficient communication with external stakeholders may
be the consumer response. Nanotechnology experts
may not be inclined to initiate the risk assessment

expected by the public, as they perceive fewer risks

associated with nanotechnology than lay people do
(Siegrist et al. 2007a; Siegrist et al. 2007b).

CONCLUSIONS

Do the companies have to be pushed or pulled to

improve their precautionary measures or is the current
situation satisfactory? In the case of NPM, industries
find that regulation should be evidence-based, but the
fate of NPM throughout the life cycle receives little
industrial attention (Helland et al. 2008). Thus, we
have a situation of interpreting whether the state of scientific evidence warrants regulatory action. We have
shown that the industrial opinion is not clear on how to
interpret the state of evidence, but that two things do
stand out: emissions should be kept as low as possible
and specific characteristics of irreversibility are reason
enough for taking precautionary measures. However,
who should monitor and demonstrate such specific
characteristics as industry does not necessarily see
that as their responsibility? How could a common
approach be defined in order to create standards of
risk assessment and other quality standards?
Building on the consensus opinion of the industry,
one may draw two specific priorities for governmental
bodies to investigate in collaboration with industry:
What are the possible sources of NPM emissions
throughout the product life cycle, how high are such
emissions and what is the environmental fate of such
emissions?
What NPM characteristics may serve as signs of
potential irreversibility for ecosystemic interactions
(e.g. persistency, bioaccumulation, etc.) and which
NPM specifically inhibit such characteristics?

REFERENCES
Ashford, N.A. & Zwetslot, G. 2000. Encouraging inherently
safer production in European firms: a report from the
field. Journal of Hazardous Materials, 78: 123144.
Nowack, B. & Bucheli, T.D. 2007. Occurrence, behavior and
effects of nanoparticles in the environment. Environmental Pollution., 250: 522.
Davis, J.M. 2007. How to assess the risks of nanotechnology:
learning from past experience. Journal of Nanoscience
and Nanotechnology, 7: 402409.
Helland, A., Kastenholz, H., Thidell, A., Arnfalk, P. &
Deppert, K. 2006. Nanoparticulate materials and regulatory policy in Europe: An analysis of stakeholder perspectives. Journal of Nanoparticle Research, 8: 709719.
Helland, A., Scheringer, M., Siegrist, M., Kastenholz, H.,
Wiek, A. & Scholz, R.W. 2008. Risk assessment of engineered nanomaterialsSurvey of industrial approaches.
Environmental Science & Technology, 42(2): 640646.

363

Helland, A. & Kastenholz. H. 2007. Development of Nanotechnology in Light of Sustainability. J. Clean. Prod,
online: doi:10.1016/j.jclepro.2007.04.006.
Morgan, K. 2005. Development of a preliminary framework
for informing the risk analysis and risk management of
nanoparticles. Risk Anal., 25(6): 16211635.
Nel, A., Xia, T., Mdler, L. & Li, N. 2006. Toxic potential of
materials at the nanolevel. Science, 311: 622627.
Oberdrster, G., Oberdrster, E. & Oberdrster, J. 2005.
Nanotoxicology: An emerging discipline evolving from
studies of ultrafine particles. Environmental Health Perspectives, 113(7): 823839.
Reijnders, L. 2006. Cleaner nanotechnology and hazard
reduction of manufactured nanoparticles. Journal of
Cleaner Production, 14: 124133.

Siegrist, M., Wiek, A., Helland, A. & Kastenholz, H. 2007a.

Risks and nanotechnology: the public is more concerned
than experts and industry. Nature Nanotechnology, 2: 67.
Siegrist, M., Keller, C., Kastenholz, H., Frey, S. &
Wiek, A. 2007b. Lay peoples and experts perception of
nanotechnology hazards. Risk Analysis, 27(1): 59.
Som, C., Hilty, L.M. & Ruddy, T.F. 2004. The precautionary principle in the information society. Human and
Ecological Risk Assessment, 10(5): 787799.
U.S. EPA 2007. Nanotechnology White Paper; U.S. Environmental Protection Agency: Washington, DC.

364

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk based maintenance prioritisation

Gunhild Birkeland & Siegfried Eisinger
DNV, Hvik, Norway

Terje Aven
University of Stavanger, Stavanger, Norway

ABSTRACT: Maintenance prioritisation is about management of all pending maintenance activities of a

system. Due to lack of spare parts, personnel or access to the component to be maintained, the maintenance
activities are often delayed or deferred. This creates a maintenance backlog. To handle this backlog, prioritisation
of the maintenance work is required, and a common approach for doing this is to introduce a criticality measure
based on the consequences of failures. However, the adequacy of this approach can be questioned as it does not
reflect risk. This paper proposes a new prioritisation method meeting this critique. The method is based on the
use of SIL (Safety Integrity Levels) requirements. The method also produces risk monitoring indices useful for
management and authorities.

INTRODUCTION

On any industrial plant, the preventive and corrective

maintenance activities must be managed and adapted
to the available resources. It is necessary to keep a
to-do list of maintenance activities and prioritise at
any time which tasks are to be performed first. The
prioritisation is affected by external constraints like the
availability of spare parts and access to the component
to be maintained.
It is observed that maintenance prioritisation on
many industrial plants is currently performed through
a predetermined equipment criticality list which, in
combination with failure impact, sets the priority on
a given maintenance activity. Based on this priority
the maximum time until the maintenance activity shall
be performed is set. Criticality takes only account of
the consequences of a failure (see NORSOK (2001)).
Since the frequency part is not taken into account
the approach is not risk based. It is believed that
a fully risk based approach will lead to better prioritisation decisions and better monitoring of risk
associated with delayed maintenance. Risk indices can
be defined for monitoring the risk level associated with
the backlog.
In this paper a risk based prioritisation method
is described. The method is based on SIL (Safety
Integrity Levels) which refers to the reliability of safety
functions. Simplicity is emphasised, in order to ensure
that the method can and will actually be used by
operators on industrial plants.

To test the developed methodology a pilot case

study was run on a part of a plant. The main results
from this case study are reported in the paper.

THE METHOD

In the Oil & Gas Industry there is established a

risk evaluation and acceptance scheme on component
level. This scheme is related to IEC 61508 and IEC
61511 (see IEC (2000) and IEC (2003)). Since 2001
all new electronic safety-related systems on Norwegian platforms must comply to IEC 61508 (see also
OLF (2000)). IEC 61508 only relates to electronic
safety-related systems, but the concept of the Safety
Integrity Level (SIL) within IEC 61508 can be used
in an analogous way also for non-electronic systems.
The SIL defines a maximum allowable unreliability of
a safety function, see Table 1.

365

Table 1. Relation between SIL and Probability of Failure on

Demand (PFD) (low demand mode) as given in IEC (2000).
SIL

PFD for low demand mode of operation

4
3
2
1

105 104
104 103
103 102
102 101

If the safety function reliability is below the SIL

value, the total system risk might become unacceptable. The SIL is thus a risk based measure and is as
such a good candidate for the evaluation of risk related
to deferred maintenance. Note however that if a safety
function is less reliable than required according to the
SIL, the exact increase of total system risk cannot be
directly inferred.
The SIL value represents an established risk based
measure for the maximum allowable unreliability of a
safety function and is used as a basis for the following
risk based prioritisation method.
Now suppose a safety function has failed. The problem is to determine the maximum time dm before the
maintenance activity for this unit should start (referred
to as the maximum delay time). Suppose the unit
belongs to SIL category 2, i.e. the probability of failure on demand (PFD), p, should be between 103 and
102 . To simplify we require p 0.005. For a specific
period of time (T), say one year, this corresponds to
an expected downtime of 1.8 days. The interval limits 103 and 102 correspond to 0.365 days/year and
3.65 days/year, respectively. To determine the maximum acceptable delay time, we need to define what is
small compared to 1.8 days. The actual delay time
should not significantly increase this downtime. Obviously if the maximum delay time is a factor 10 higher
than 1.8 days, i.e. 18 days, it is too large. A factor 5
is smaller but still it gives a major change in the total
downtime. We see that we have to reduce the factor to
about 1 to not significantly change the downtime, and
the proposed method uses this number as a criterion
for determining the maximum delay time.
We formalise this by introducing the delay time factor D, expressing the ratio between the additional delay
time d due to deferred maintenance and the expected
downtime due to the SIL requirement, i.e.
D=

d
pT

(1)

The criterion used to determine the maximum delay

time can then be written
D 1,

(2)

i.e. dm p T .
Next we address the case of redundant safety systems, i.e. safety functions consisting of two or more
components in parallel such that one failure does
not immediately cause the safety function to fail. It
is observed that the dominant reason for failures of
redundant safety functions is Common Cause Failures
(CCF). To incorporate these failures into the method,
we use a modifying factor , the so-called -factor
(see IEC (2000) and IEC (2003)). Typical values for
range from 1% to 10%. For the present study, a

-factor of 10% was chosen and the degree of redundancy is taken into account as 0r where 0 = 10%
and r is the degree of redundancy. For example, r = 1
means that the safety function tolerates one failure,
and 0r = 0.1. It follows that the expected additional
delay time due to the deferred maintenance equals
d 0r , and hence Equation (1) can be written
D=

d
r
pT 0

(3)

A system might for example include a safety function rated at SIL 2 with redundancy of 1. The function
contains two sensors which can both detect a high
pressure. Over one year the delayed maintenance of
the sensors has amounted to 8.7 days. Using again
p = 0.005 the delay factor becomes D = 0.48. The
maintenance delay should be acceptable.
As will be demonstrated in Section 3, observed
delay time factors Di can be readily calculated from
maintenance data for all safety functions i. Based on
these we can produce overall indices and diagrams
providing the management and the authorities with
a picture of the maintenance backlog status. These
indices and diagrams can be used as a basis for
concluding on whether the maintenance of safety functions is going reasonably well or whether safety is
eroding away. An example of an approach for visualising the status is the following: Calculate all delay
time factors Di for all safety functions i which have
failed in the time interval T and present a graph showing the associated cumulative distribution. At D = 1
a dividing line is drawn which marks the maximum
delay time. The number of safety functions to the right
of this line should be small, i.e. centre of the histogram
should be well to the left of this line.
If several systems are to be compared it would
be interesting to calculate a single number which
summarises the information on the maintenance performance with respect to deferred maintenance of
safety systems. One could readily calculate the mean
value D of all Di* , with respect to the safety functions
which have failed in the time interval T . The mean
value calculates the centre of gravity of the distribution of the Di s, and as will be shown in Section 3,
the mean value provides a good overview over the
maintenance backlog status. Moreover, the mean value
D can easily be plotted as a function of time, which
provides an overview over the evolution of the maintenance backlog in time. The mean value, D , could
also be plotted as an accumulated function of time,
which provides an overview of total delay time factor
over a period of time. Care has however to be shown
when using D as a basis for judgments about the
acceptability of the maintenance backlog. For example, requiring D 1, would allow many large delay
times (far beyond the limit 1), as long as the average

366

is within the criterion 1. Hence if a criterion should be

based on D it should be considerable lower than 1.
Note that potential effects from compensating measures for failed safety functions is only included as a
black and white model: if a compensating measure
exists, is perfectly effective and is in fact used, the
respective failures of the safety function can be taken
out of the list. If not all conditions are fulfilled, the failures are left in the list conservatively assuming that the
compensating measure is totally ineffective.

Cumulative distribution of D*i

1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
0

0,5

1,5

2,5

3,5

4,5

Delay time Factor

PILOT CASE STUDY

A pilot case study has been performed for a

Norwegian offshore installation based on information from the computerised maintenance management system (CMMS). Notifications (failure reports)
for this installation were extracted for the period
01.01.200731.12.2007. The total numbers of analyzed notifications are 1402. The pilot case study
covers only corrective maintenance of safety functions.
The maintenance delay times could be extracted
from the CMMS with reasonable accuracy. Unfortunately, the safety rating of the safety functions were
not given as SIL values, but as a safety criticality with
3 categories. For the purpose of the case study, a simple
mapping between these categories and the SIL rating
was performed. The mapping is considered reasonable for some functions where experience exists, but
the mapping must be validated for real-life usage of
the method. The redundancy level was also readily
available in the CMMS.
3.1 Distribution of delay time factor
As pointed out in Section 2, the distribution of the
delay time factor over all safety functions which
have failed in the time interval T will give valuable
information on the maintenance backlog status.
In Figure 1 the cumulative distribution is shown for
the pilot case. For about 70% of all failed safety functions the maintenance delay time factor was below 1,
but there is also a considerable number of maintenance
delays with a delay time factor well above 1the maximum delay time factor in the period is in fact 162. The
average delay time factor is D = 4.3. As shown in
Figure 1, the contribution from delay time factors >1
is rather high.
3.2

Maintenance performance as function of time

Next we study how the backlog evolves over time.

Figure 2 presents the accumulated annual delay time
factor D as a function of time. Starting from a zero
delay time factor, the figure shows how D evolves

Figure 1.

Cumulative distribution of delay time factor.

4.5

Accumulated annual delay time factor D*

All SILs

SIL1

3.5

SIL2
3
2.5
2
1.5
1
0.5
0
jan

feb

mar

apr

mai

jun

jul

aug

sep

okt

nov

des

Time

Figure 2. Accumulated annual delay time factor D as a

function of time.

through the year to arrive at the above reported D =

4.3 at the end of the year.
The different safety levels (SILs) are kept separate
and are summed in the curve All SILs. The mean
delay time factors D are computed on the basis of the
actual delay times d at the specific point in time; for
example on 1. April a safety function that failed four
days earlier is assigned a delay time d equal to 4 days
(refer Equation (3)).
Figure 2 illustrates clearly that there are periods
with considerable backlog, i.e. the periods with high
slope. During the end of 2007 the maintenance delays
seem to be better under control than in the months
February and March.
Figure 2 also illustrates how this method can be used
for active maintenance prioritisation. At any point in
time a number of notifications are active. In order to
minimise the maintenance delay factor (and thus the
risk), focus should be put on the safety functions contributing to the highest slopes, i.e. the safety functions
for which the factor 0r /p is highest among the active
notifications. For example on 26. January a notification with SIL = 2 and r = 0 (no redundancy) is

367

created and it takes 51 days of delayed maintenance,

contributing to a clearly visible slope in Figure 2. Other
active notifications during this time contribute to considerably less delay time. In general attention should
be placed on non-redundant SIL2 functions.
From Figure 2 one may conclude that the accumulated risk from delayed maintenance is too high, as
there are many safety functions for which the PFD is
more than doubled by delayed maintenance.

performance through time, find important contributors to delayed maintenance, prioritise maintenance
from day to day basis or to compare the maintenance
performance of several systems.
As part of future work it is planned to include preventive maintenance delays in the methodology, i.e.
an analogous delay time factor originating from preventive maintenance being performed later than the
planned time.

REFERENCES

CONCLUSION AND FUTURE WORK

A new method for prioritisation of maintenance backlog with respect to maintenance of failed safety
functions has been proposed. The proposed method
is linked to the established Safety Integrity Level
regime from IEC 61508 which specifies the reliability requirements of safety functions. The method is
thus inherently risk based. It furnishes a variety of risk
indices to gain an overview over the status of the maintenance backlog of safety functions. The indices can be
used to determine acceptance levels which tell the user
whether the maintenance of safety functions is going
reasonably well or whether safety is eroding away.
The indices can be used to monitor the maintenance

IEC. 2000. IEC 61508Functional safety of electrical/electronic/programmable electronic safety-related

systems. First edition. IEC.
IEC. 2003. IEC 61511Functional safetySafety instrumented systems for the process industry sector. First
edition. IEC.
NORSOK. 2001. NORSOK standard Z-008Criticality
analysis for maintenance purpose. Revision 2. Norwegian
Technology Centre.
OLF. 2001. OLF 070Recommended Guidelines for the
application of IEC 61508 and IEC 61511 in the Petroleum
Activities on the Norwegian Continental Shelf. Revision
0.1 The Norwegian Oil Industry Association.

368

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Shifts in environmental health risk governance: An analytical

framework
H.A.C. Runhaar, J.P. van der Sluijs & P.P.J. Driessen
Copernicus Institute for Sustainable Development and Innovation, Utrecht University,
Utrecht, The Netherlands

ABSTRACT: Environmental Health Risks (EHRs) traditionally have been dealt with in a hierarchical and
technocratic manner. Preferably based on scientific expertise, standards set are uniform and based on legal
regulations. However, this approach has encountered implementation problems and deadlocks, particularly in
cases where scientific knowledge is at best incomplete and interests of powerful stakeholders conflict. Many
new approaches to manage EHRs have been implemented, which share two characteristics: an increased
integration of (a) cost-benefit and other considerations; (b) the public and other stakeholders; and (c) of EHR
objectives in other sectoral policies, and an increased differentiation of EHR standards (partly as a consequence
of the former characteristic). Still little systematic empirical research has been conducted on the experiences
with these shifts in EHR governance, in particular in the light of the shortcomings of the traditional approach
to EHR governance. This paper proposes an analytical framework for analyzing, explaining and evaluating
different categories of, and shifts in, EHR governance regimes. We illustrate our paper with the trends in EHR
governance described above.

INTRODUCTION

The physical environment poses many different risks

to human health. In many countries worldwide governments attempt to regulate environmental health
risks (EHRs) by avoiding or reducing their occurrence, reducing or avoiding exposure to these risks,
or by mitigating the effects of exposure to these risks.
A wide variety of approaches to dealing with EHRs
can be observed. Not only are similar risks regulated differently in an international perspective as
regards standards, principles for standard-setting or
instruments implemented for reaching these standards,
national governments appear not to have consistent
regimes for dealing with different risks, either (Hood
et al., 2004; ORiordan, 1985; Zoetbrood, 2005).
To date, only limited research has been conducted
on the reasons for the observed heterogeneity in EHR
governance and the relative contribution of the various
EHR governance regimes to the reduction of EHRs
(Hood et al., 2004). The few publications in this area
are primarily concerned with classifying EHR governance regimes and elaborating on the appropriateness
of these regimes in distinct contexts (see Section 2).
Yet, these papers are mainly conceptual of nature;
empirical evaluations in this area are scarce.
In this paper we develop an analytical framework
for addressing the above knowledge gap and show how

this framework can be used for analyzing some trends

in EHR governance that we observed in a recently
conducted empirical quick-scan survey (Soer et al.,
2008). We conclude the paper with some suggestions
for further research in this area.

2
2.1

EHR GOVERNANCE: KEY CONCEPTS

What are EHRs?

In the risk research community, a substantial body

of work has been carried out on how the concept
risk should be defined. The International Risk Governance Council (IRGC) defines risk as an uncertain
(generally adverse) consequence of an event or an
activity with respect to something that humans value
(Bunting, 2008). Analytically, risks have two main
components: the probability of an event and its consequences for health, environment and goods (cf.
Neumann and Politser, 1992).
EHRs are risks stemming from the physical environment in which humans live. Distinction can be
made between health risks stemming from natural
events (e.g. floods, earthquakes) and man-induced
EHRs (e.g. by the emission of toxic substances or
the production of noise). In this paper we focus on
the latter type of EHRs, as we expect that for these

369

type of risks there are more opportunities to influence

them (in the case of natural events EHR management
will mainly be focused on adaptation and mitigation
of adverse health effects and less with fundamentally
addressing their causes).

The regulatory and decision-making style (in part

based on ORiordan, 1985);
The requirements with respect to organizational
and institutional capabilities for assessing, monitoring and managing risks (including emergency
management).

2.2

Other relevant elements of risk governance regimes

found in the literature include culture, principles
and uncertainty and ambiguity. Rayner and Cantor
(1987) discuss four different institutional cultures
that may underlie approaches to risk management.
These cultures reflect distinct worldviews, based
among other things on particular principles of social
justice and perceived economic interests. The importance of culture and cultural values in the perception and selection of risks is also recognized by,
among others, Douglas and Wildavsky (1982). It
is not clear however whether culture is part of
governance regimes or an important context variable of influence on the way in which governance
regimes are shaped and function. Related to cultural aspects are the principles behind standard setting; both elements are based on some core values.
As regards principles for EHR standard-setting, De
Hollander and Hanemaaijer (2003) for instance distinguish between four approaches, reflecting different
principles:

EHR governance regimes

EHR governance regimes can be defined as the complex of institutional geography, rules, practice, and
animating ideas that are associated with the regulation
of a particular risk or hazard (Hood et al., 2004: 9).
Only limited research has been conducted on the classification of risk regimes (Hood et al., 2004). Below
we discuss a few of such attempts.
In comparing the ways in which risks were regulated
in different countries, ORiordan (1985) identified
four (partly overlapping) styles of risk regulation that
can be considered as four risk governance regimes.
These regimes primarily differ in the way in which
decision-making is organized (top-down, interactive
etc.) and the extent to which consensus among stakeholders is possible.
Hood et al. (2004) consider risk governance regimes
as systems consisting of interacting or at least related
parts and identify nine distinct configurations or risk
governance regimes, which they identify by means of
two dimensions:
Basic control system components: ways of gathering information; ways of setting standards; and
ways of changing behavior in order to meet the
standards (i.e. policy instruments);
The instrumental and institutional elements of
regulatory regimes: the regulatory regime context
(different types of risks at hand, the nature of public
preferences and attitudes over risk and the ways in
which stakeholders are organized) and the regime
content (the policy setting, the configuration of
state and other organizations directly engaged in risk
regulation and their attitudes, belief and operating
conventions).
The IRGC takes a more normative approach. It
advocates a framework for risk governance that aims
at, among other things, enhanced cost-effectiveness,
equal distribution of risks and benefits, and consistency in risk assessment and management of similar
risks (Bunting, 2008; Renn, 2006). The IRGC framework takes into account the following elements:
The structure and function of various actor groups in
initiating, influencing, criticizing or implementing
risk decisions and policies;
Risk perceptions of individuals and groups;
Individual, social and cultural concerns associated
with the consequences of risk;

Right-based approach: an equal protection for all

above a certain risk level;
Utility-based approach: maximization of benefits
for society as a whole (i.e. achieving the greatest
increase in overall public health at the lowest cost);
Technology-based approach: using the best available techniques;
The precautionary principle: a cautious approach
in the light of complexity, uncertainty and irreversibility.
The EEA (2001) emphasizes the importance of
uncertainty in risk governance and suggests a framework for classifying risks regarding different forms
of uncertainty (risk, uncertainty and ignorance) and
actions (and guiding principles) for handling these
situations. Klinke and Renn (2002) also discuss ambiguity: the presence of contested views about the
desirability or severity of a given hazard (Klinke
and Renn, 2002: 1092). UNESCO COMEST (2005)
addresses the issue of risk governance under uncertainty and explores the implications of the precautionary principle for risk governance. Van der Sluijs
(2007) reviews the challenges to risk governance of
uncertainty and precaution, focusing on the interface
between science and policy, and concludes that a precautionarypost normalstyle of risk governance
has slowly but steadily started to invade traditional
approaches. Methodological challenges remain as

370

to how to further integrate different approaches to

deal with uncertainty, in particular the more technical approaches and the approaches from the policy
perspective.
The above discussion suggests that there is no such
thing as a generally accepted framework for identifying and classifying risk governance regimes, but
provides us with what seem to be relevant elements
in such a framework:
EHR type in terms of (perceived) seriousness, certainty of the knowledge basis and the prevalence of
conflicting interests among stakeholders as regards
EHR handling;
Governance style, referring to the plurality and
diversity of actors allowed to be involved in
decision-making, the moment when they are
involved in decision-making processes, and their
role in the implementation of EHR strategies;
The role of science and other knowledge sources in
decision-making on EHR objectives;
The principles behind EHR objectives;
Dominant instruments in the implementation of
EHR strategies (e.g. hierarchical or more voluntary
forms of regulation);
Culture and values;
The institutional context;
The interlinkages between the above elements.

2.3

Evaluating EHR governance regime

performance

Although many criteria can be employed for assessing

the performance of EHR governance regimes (see for
instance those employed by the IRGC in Section 2.2),
one of the basic criteria would seem to be the extent
to which they succeed in reducing EHRs to levels that
are acceptable to decision-makers, the public, scientists and other stakeholders. At least two interrelated
dimensions are relevant here. The first is the speed
with which EHRs are identified, decision-making on
their management is completed, and implementation is
organized. The second is the outcome in terms of (perceived) risk reduction and its evaluation by relevant
actors.
Yet, it is also interesting to move beyond the
level of individual EHRs, e.g. by looking at how
EHR regimes impact upon other EHRs. Relevant in
this light is the concept of risk migration: the situation where one risk is reduced but another one
created or enhanced. An example is polybrominated
diphenyl ethers (PBDEs) based flame-retardant compounds that accumulated in the environment and in
humans and in particular caused societal anxiety when
they were found in breast milk (Alcock and Busby,
2006).

2.4

Explaining EHR governance regime selection

and performance

What explains the presence of particular EHR governance regimes and how can we understand the
contribution of these regimes to the reduction of
EHRs? Hisschemller and Hoppe (2001) and Hoppe
(2002) adopted a useful theoretical model that links
governance regimes to two specific characteristics of
the policy problem at issue: certainty of the knowledge basis and the extent to which norms and values
converge (see Figure 1). In contrast to the authors discussed in the preceding Section, Hisschemller and
Hoppe (2001) and Hoppe (2002) explicitly consider
these characteristics as independent variables relevant
to the appropriateness and performance of governance
regimes. Interactions and negotiations with, and input
from, stakeholders are assumed to be necessary when
stakes of the various actors involved are high, norms
and values diverge, and when there is high uncertainty about causes of the policy problem or impacts of
alternative policy programs i.e. when unstructured
policy problems are at issue. This unstructured problem category is similar to the post normal science type
of risk assessment proposed by Funtowicz and Ravetz
(1993). In these situations stakeholder involvement is
required in all stages of policy-making, including analysis, both in order to get access to relevant information
and to create support for policy-making. Examples
of this mode of problem solving are discussed in
Kloprogge and Van der Sluijs (2006) for the risks of
climate change. Structured policy problems, on contrary, can be solved in a more hierarchical way. Here,
policy can be left to public policy-makers; involvement
of stakeholders is not needed for analysis or successful problem solving. In this case policy-making has
a technical character and is often heavily based on
scientific knowledge. In the case of moderately structured, means problems stakeholder involvement is
not required for recognition of the problem at issue,

Structured problems

Moderately structured
/ means problems

e.g. road maintenance

e.g. traffic safety

Moderately structured
/ goals problems

e.g. abortion

Unstructured
problems
e.g. car mobility

Figure 1. Types of policy problems (with examples).

Source: Hoppe, 2002 (adapted).

371

but mainly for the selection of the means by which this

goal is to be reached. Since there is high uncertainty
about the effectiveness of and stakeholder preferences for various solutions, policy-makers together
with stakeholders search for adequate problem-solving
activities. Finally, in the case of moderately structured,
goals problems there is substantial agreement on certain knowledge but sometimes intense disagreement
about norms and values at stake and about the goals
that should be set. Interaction with stakeholders is
required in order to identify and streamline stakeholder
preferences.
Other bodies of literature focus on specific aspects
of governance regimes, in particular the role of science in policy (e.g. Funtowicz and Strand, 2008;
Gibbons et al., 1994). Many publications in that area
start emphasize that science often cannot be the only
knowledge producer, in particular in the case of
uncertainties or moral issues where perceptions of
stakeholders diverge and where consensus is required
for successful policy (Funtowicz and Strand, 2008).
This is in line with Hisschemller and Hoppe (2001)
and Hoppe (2002).
The above literature is useful as a starting point
for searching for explanations for EHR governance
regimes found in practice and their performance, but
there are two aspects that deserve attention. One, the
literature discussed is in part normative; it prescribes
ideal governance responses in particular situations.
However, in practice problems that can be characterized as unstructured are not always dealt with
according to what the literature suggests. For finding
explanations for the presence of certain governance
regimes therefore also other aspects may be relevant (e.g. the existence of diverging interests and
power relations; see also Flyvbjerg, 1998). Two,
the literature starts from problems that have been
accepted on societal and political agendas and does
not explicitly address the issue of when problems
become accepted. In the case of EHRs (and risks in
general) calamities often are followed by intense public and political reactions irrespective of the chance
of such events. Risk awareness can also be created by policy entrepreneurs with an interest to
bring together problem-owners, decision-makers and
solutions (Kingdon, 1995).
2.5

Figure 2. Analytical framework for characterizing,

explaining and evaluating EHR governance regimes.
Source: Authors.

UNDERSTANDING EHR GOVERNANCE

IN PRACTICE: ILLUSTRATION

Rather than predefining distinct configurations of

EHR governance regimes and analyze and evaluate
these, we suggest taking the elements from the analytical framework presented above to identify distinct
EHR governance regimes and developments in these
regimes from observations in practice. In this Section
we will illustrate the value of the analytical framework
by analyzing and evaluating EHR governance regimes
at the macro level encompassing shifts in EHR regimes
in general. Our framework however can also be used
for micro level analyses focusing on distinct regimes
for a particular EHR in a particular country and time
period.
In a quick-scan survey conducted in 2007 Soer et al.
(2008) compared trends in EHR governance in the
Netherlands with those in nine European countries,
the U.S. and Australia. The authors observe shifts that
they summarize as a turn from a traditional, specialized approach to a more integrated and differentiated
approach. How can this trend be analyzed, understood and evaluated from our analytical framework?
Below we will first describe and evaluate the traditional approach to EHR governance and then try to
explain shifts and make an ex ante evaluation of them.

Towards an analytical framework

for evaluating EHR governance regimes

Figure 2 brings together the elements of our analytical

framework that we discussed above. (Block) arrows
indicate the types of relationships between the elements that we expect to find in empirical analysis.
Feedback loops from EHR governance regimes and
their outcomes to their context are possible, but for
the sake of this paper ignored in the figure.

3.1 Specialized EHR governance regimes

3.1.1 Characterizing EHR governance regimes
The traditional way of dealing with EHRs as could be
observed in many western countries as from the Industrial Revolution can be summarized as follows (e.g.
Fiorino, 1990; Heriard-Dubreuil, 2001; ORiordan,
1985; Sunstein, 2002):

372

 Regarding the role of stakeholders: usually government agencies set standards in hierarchical ways.
(Formal) public participation is limited or absent.
Informal influences by stakeholder groups however
do exist (see below);
Regarding the role of science: scientists are the
logical providers of knowledge on the nature and
severity of EHRs and levels at which health risks
are acceptable;
Regarding principles underlying EHR standards:
standards are based primarily on estimated health
impacts, preferably based on scientific risk assessments. Cost-benefit and other considerations usually do not play a major role in an early stage of
EHR governance, i.e. in the definition of standards;
Regarding standards: Environmental health risks
are dealt with on an individual basis. Where possible
quantified standards are formulated that generic, i.e.
apply to all of the situations specified and do not discriminate between groups of people. This results in
detailed standards, which is one of the reasons why
this approach is often considered technocratic.
Compensation of health risks geographically or
between groups of people (either concerning one
specific risk type or between different risk types)
usually is not allowed;
Regarding policy instruments: a linear approach to
EHR governance is taken: implementation and the
selection of instruments (of which legislation and
licenses are typical) follows after risk assessment
and standard-setting and does not play a large role
in earlier stages of EHR policy.
Today this approach, which can be called specialized due to its focus on individual EHRs, still can
be observed worldwide. In Europe, for instance, this
approach is visible in the EU Directives regulating
concentrations of particulate matter and other EHRs.
3.1.2 Explaining specialized EHR regimes
The technocratic and hierarchical character of EHR
regimes reflect wider ideas on the role of governments in society, which were dominant in Western,
liberal societies until some two decades ago. (Central) governments were considered to have a strong and
leading role in addressing social problems and there
was much faith in the contribution that science could
have in enhancing the effectiveness and efficiency of
government policy (Fischer, 1997; Van de Riet, 2003).
This has become institutionalized in very specialized
(and growing) bureaucracies and an important role of
(institutionalized) science-based knowledge providers
(policy analysis).
3.1.3 EHR governance outcomes and explanations
The specialized approach has resulted in the reduction of various health risks. For instance, in the

Netherlands air and water quality improvements were

realized and smog levels were reduced dramatically,
whereas the noise problem was stabilized (Keijzers,
2000). A similar picture emerges if we take a European perspective (e.g. EEA, 2003, 2007 regarding air
and water quality). Yet, the traditional approach to
EHRs also encountered various (often interrelated)
problems, which at least in part can be explained
by characteristics of the EHR governance regime
configuration.
First, various examples are known of EHRs with
severe health consequences that have not been regulated at all. An important explanation is found in
limitations in scientific knowledge, hindering the setting of clear EHR standards (Health Council of the
Netherlands, 1995; Open University, 1998, U.S. EPA,
2004; VROM, 2004). However another explanation is
found in powerful lobbies of interest groups that succeed in lowering EHR standards (see for instance Hood
et al., 2004 for the U.K. situation). Here strong economic interests and weak(ly presented) public (health)
interests result in deadlock situations; EHRs not being
actively dealt with. A third explanation lies in the
principles behind EHR standard setting, where costbenefit considerations usually are not of primary
importance. This has resulted in, what is perceived
as, excessive costs of meeting EHR standards (e.g.
Health Council of the Netherlands, 1995). An example is soil quality; in many western countries including
for instance Denmark and the Netherlands there are
vast amounts of land polluted by industrial activities that need to be cleaned up but for which a lack
of funding is available (Soer et al., 2008). Another
example is radon gas in buildings, causing lung cancer. At least in the Netherlands, the costs of reducing
this risk are perceived as excessive (VROM, 2004).
A fourth explanation of EHRs not actively being taken
up is that it is not always clear which societal actor
should take responsibility of dealing with EHRs, in
particular when those causing risks or in control of
instruments for reducing EHRs can be distinguished
clearly or when governments are expected to deregulate (Renn, 2006; U.K. H.M. Treasury, 2005; VROM,
2004). It should be noted that also examples are found
of EHRs policies that haven been criticized due to
a limited scientific basis or low cost-effectiveness.
In the Netherlands for instance millions of euros
were made available for research on health effects
of radiation from mobile phone transmitting stations
in response to public concerns, while there was little scientific evidence of such effects. This illustrates
an important other shortcoming of the specialized
approach: difficulties in reconciling objective scientific EHR assessments and subjective EHR perceptions by the public. Perceptions of stakeholders are
considered to have become more important due to scientific limitations, democracy concerns and a growing

373

dependence of governments on stakeholders. Even

though knowledge has been built on variables that
affect the perceived seriousness of risks (e.g. voluntariness and personal control, the catastrophic potential,
emotional associations with the risk, trust in regulatory agencies etc.) and the availability of guidelines
for taking public perceptions of EHRs explicitly into
account (e.g. Klinke and Renn, 2002), this issue is
thus far only has been addressed marginally in EHR
governance regimes (e.g. in the Netherlands and elsewhere for external safety a distinction is made between
individual risk and group risk based on the societal
impact of a group of people killed or injured in comparison to individual casualties. Yet an overall framework
for incorporating public concerns is lacking.
A second and related problem is that realizing EHR
standards or the setting of stricter EHR standards is
often hindered by conflicts between EHR objectives
and those of other policy domains, in particular spatial planning. In particular in intensively used urban
areas, further spatial developments are often hindered
by strict EHR norms (e.g. Glasbergen, 2005; Wheeler
and Beatley, 2004). Yet, the institutional context that
lies responsibilities for EHR governance mainly with
state agencies that, in addition, often prescribe detailed
(minimum) standards for EHRs, combined do not
provide an incentive for actors involved (manufacturers, planners, etc.) to reduce EHRs beyond the legal
minimum standards or to find innovative ways for
reducing EHRs (Health Council of the Netherlands,
1995).
A third problem with the specialized approach relates to difficulties in dealing with cumulative effects
of EHRs (e.g. Health Council of the Netherlands,
1995; U.S. EPA, 2004). On the one hand, such effects
are overlooked in the case of source-based regulations
(in contrast to effect-based regulations, such as the EU
Directives for air pollution that prescribe maximum
concentrations of for instance particulate matter). On
the other hand, problems exist because of a lack of
knowledge on the combined effect of different EHRs
(e.g. Robinson and MacDonell, 2006, VROM, 2004).
Apart from that, the specialized approach sometimes
results in risk migration(see Section 2.3).

3.2

Integrated and differentiated EHR

governance regimes

3.2.1 Characterizing EHR governance regimes

A recent quick-scan survey of trends in EHR governance in ten European countries, the U.S. and
Australia (Soer et al., 2008) revealed various changes
in EHR governance regimes in Australia, the U.S. and
in ten European countries. It should be noted that
these trends are not found in all of these countries
in the same intensity. Nevertheless the study points

to the following changes in the elements of EHR

governance regimes outlined in our analytical framework (see Figure 2):
Regarding the role of stakeholders: the public and
other stakeholders are increasingly involved in the
formulation and implementation of EHR governance regimes. Not only is a more participatory
approach to EHRs increasingly considered necessary in order to create support for EHR policies,
it also stems from a desire to make other stakeholder co-responsible for preventing and reducing
EHRs (Renn, 2006; U.K. H.M. Treasury, 2005;
Soer et al., 2008; VROM, 2004). Participation
sometimes serves to create legitimacy and trust,
for instance in the U.K. where public protest arose
after the public was exposed to unacceptable risks
relating to among other things BSE (Fisher, 2000);
Regarding the role of science: a few efforts have
been made to reconcile scientific and stakeholder
risk perceptions. The U.K. H.M. Treasury (2005)
for instance has proposed to make concern assessment and integral part of risk management. Since
2006 there is a legislative requirement to attempt
incorporating public opinion in the decision-making
process by means of such an assessment, although
it is recognized that public attitudes to risk are difficult to measure; and that public concerns may
be founded on ignorance, false information, unreliable media reporting, etc. (U.K. House of Lords,
2006: 6). Eventually it is up to decision-makers to
determine the relative weight of such factors;
Regarding principles underlying EHR standards:
cost-benefit, economic and other concerns are
more explicitly (and perhaps also more intensively)
included in EHR assessments (e.g. U.K. H.M. Treasury, 2005, VROM, 2004). In this way also a better
integration of EHR standard setting and EHR implementation is strived after. There are however also
countries that explicitly do not consider costs or feasibility in the risk assessment and standard setting
stage (e.g. Australia and the U.S.; Soer et al., 2008);
Regarding standards: an increased differentiation
can be discerned. In the Netherlands for instance
some experiments have been conducted with providing urban planners more policy freedom in the
formulation of area-specific environmental ambitions (Runhaar et al., 2009). In this way a better
coordination and local optimization of environmental (health) planning and spatial planning is
facilitated. These experiments allow for a limited
form of differentiation of environmental objectives
(in a spatial sense). Compensationa lower environmental quality in one domain being offset by an
improvement in anotherwas also envisaged, but
appeared impossible (also in light of EU Directives).
Also in Scotland it is legally possible to generate

374

area-specific environmental standards (Soer et al.,

2008). Obviously differentiation of EHR standards
is not a new phenomenonhowever in the past
differentiation usually related to predefined situations. Differentiation is also envisaged in another
forman explicit and deliberate (rather than an
ad hoc) differentiation of EHR governance regimes
for distinct types of EHRs. Up to now several
countries including the Netherlands, Malta and the
United Kingdom have announced to adopt such an
approach (Soer et al., 2008). Yet implementation is
still in an early stage. It is therefore still uncertain
whether or not this type of differentiation is feasible,
realistic, or even desirable;
Regarding policy instruments: risk assessment and
management are more integrated. Also, EHR objectives are increasingly integrated in other sectors, in
particular in spatial planning. Examples of countries are Australia, Germany and the Netherlands.
Regularly used tools for better incorporating EHRs
in spatial planning are Health Effect Screenings,
Health Impact Assessments and Strategic Environmental Assessments. In Malta integration between
policy sectors is strived after through the formation
of an inter-departmental committee on environment
and health (Soer et al., 2008).

societal context. The trend towards more public and

stakeholder participation and integration of policy
sectors form part of a broader shift towards more
integrated approaches to dealing with societal issues
in Western, liberal societies, which are commonly
referred to a change in planning from government
to governance. There is no agreed-upon definition
of what governance entails (Jordan et al., 2005). Yet,
often-mentioned characteristics are: a decline of central governments ability to regulate society in a topdown fashion, resulting in more horizontal forms of
steering in cooperation with other government ties and
stakeholders outside the government; a more direct
and early inclusion of stakeholders in the formulation
and implementation of policies; shared responsibilities between state, civil society and market actors in
the formulation and implementation of policies; (as
a consequence) a blurring of the traditional boundaries between state, civil society and market (Van
Kersbergen and Van Waarden, 2001). Two important
causes of this trend towards governance are a (perceived) ineffectiveness of central government policy
due to, among other things, limits to the extent in
which society can be known and steered, and the plea
for more direct democracy.

On a more abstract level, these trends reflect a

change to more integration (of cost-benefit and other
considerations in EHR standard-setting; of stakeholders in EHR in the formulation and implementation
of EHR governance regimes; and of EHR objectives
in other policy sectors) and more differentiation of
EHR standards (partly as a consequence of the former
shifts).

3.2.3 Specialized EHR regimes: Outcomes,

explanations and challenges ahead
The shift towards more integration and differentiation is observed in several countries but is by far
crystallized. In many of the countries examined,
plans are in immature stages or even only under
consideration. Not much experience has been built
with increased integration and differentiation in EHR
policies. Yet a few new dilemmas were identified,
including the weighing of stakeholder opinions and
scientific inputs and the weighing of health and other
concerns (see also Glasbergen, 2005 and Runhaar
et al., 2009). In addition EHR planners state that
they lack a clear framework for systematically dealing with EHRs. Old problems include a lack of
(scientific) data (e.g. on cumulative and risk migration effects), insufficient funding and problems in
communication between EHR planners, other sectoral planners and stakeholders. Finally a potential
risk of more integration is compromising of EHR
standards in favor or other ambitions (Soer et al.,
2008).
It is far from certain that more integrated and differentiated governance regimes will completely replace
more traditional EHR regimes. In Europe, EHRs are
increasingly targeted at a supranational level (EU)
by means of Directives that prescribe strict and uniform standards for acceptable EHRs. It is interesting
to examine how these various forms of EHR regimes
interact.

3.2.2 Explaining specialized EHR regimes

The shifts in EHR governance regimes discussed
above in part stems from the problems with the
specialized approach. In the Netherlands for example EHR planners have differentiated norms for soil
quality according to land use functions. Soil norms
for houses are stricter than for those for industrial
areas. A same approach has been followed as regards
noise nuisance risks (Glasbergen, 2005). This differentiated approach is based on a desire to realize
EHR reductions in a more cost-effective manner and
to better reconcile EHR policies and spatial policies. Setting EHR standards based on best available
technology (BAT) or on what is reasonably practicable (ALARP) is another well-known approach
to reconciling EHR reduction and economic concerns. In this context these shifts can be considered as
ways to overcome the outcomes of more specialized
regimes.
However, the above shifts can also be explained
by more general developments in the broader

375

CONCLUSIONS

Given the limited attention being paid to EHR governance regimes in risk research, our aim was to develop
an analytical framework for characterizing, explaining
and evaluating such regimes. Based on a review of relevant literature we developed a framework, which we
illustrated by means of some macro trends we observed
in some Western countries.
The framework seems to be useful for guiding
research into the above area as it allows for a systematic examination of relevant elements and possible
relationships. In the analysis of recent shifts in EHR
governance we discussed as an illustration of our
framework not all of elements were elaborated in much
detail. Cultural influences on governance regimes
for instance may be identified more explicitly in an
international comparative study.
We suggest that further research is conducted in
order to classify EHR governance regimes with the aid
of our framework. What distinct configurations can be
found in practice, next to the (perhaps artificially constructed) specialized approach we discussed? And
what relations between the various elements exist? The
shifts towards integration and differentiation identified in the quick-scan survey conducted by Soer et al.
(2008) may act as a starting point. Three remarks
should be made here. One, in this study only a
(probably non-representative) sample of countries was
included. Two, the shifts identified were not observed
in each of the sample countries in the same intensity.
Three, the survey focused on the main characteristics of national EHR governance regimes; no in-depth
research was conducted into specific EHRs. However,
many of the trends that we observed are similar to
those reported by for instance Amendola, 2001, De
Marchi, 2003, Heriard-Dubreuil, 2001 and Rothstein
et al., 2006.
REFERENCES
Alcock, R.E. & Busby, J. 2006. Risk migration and scientific
advance: the case of flame-retardant compounds. Risk
Analysis 26(2): 369381.
Amendola, A. 2001. Recent paradigms for risk informed
decision making. Safety Science 40(14): 1730.
Bunting, C. 2008. An introduction to the IRGCs risk
governance framework. Paper presented at the Second RISKBASE General Assembly and Second Thematic
Workshop WP-1b. May 1517. Budapest: Hungary.
De Marchi, B. 2003. Public participation and risk governance. Science and Public Policy 30(3): 171176.
Douglas, M. & Wildavsky, A. 1982. How can we know the
risks we face? Why risk selection is a social process. Risk
Analysis 2(2): 4951.
EEA. 2001. Late lessons from early warnings: the precautionary principle 18962000. Copenhagen: European
Environmental Agency.

EEA. 2003. Europes water: An indicator-based assessment.

Copenhagen: European Environmental Agency.
EEA. 2007. Air pollution in Europe 19902004. Copenhagen: European Environmental Agency.
Fiorino, D.J. 1990. Citizen participation and environmental risk: a survey of institutional mechanisms. Science,
Technology & Human Values 15(2): 226243.
Fischer, F. 1997. Evaluating public policy. Chicago: NelsonHall Publishers.
Fisher, E. 2000. Drowning by numbers: standard setting
in risk regulation and the pursuit of accountable public
administration. Oxford Journal of Legal Studies 20(1):
109130.
Flyvbjerg, B. 1998. Rationality and power. Democracy in
practice (translated by S. Sampson). Chicago/London:
University of Chicago Press.
Funtowicz, S.O. & Ravetz, J.R. 1993. Science for the postnormal age. Futures 25(7): 739755.
Funtowicz, S. & Strand, R. 2008. Models of science and policy. In T. Traavik & L.C. Lim (eds), Biosafety first: holistic
approaches to risk and uncertainty in genetic engineering
and genetically modified organisms. Trondheim: Tapir
Academic Press (forthcoming).
Gibbons, M., Limoges, C., Nowotney, H., Schwarzman, S.,
Scott, P. & Trow, M. 1994. The new production of
knowledge: the dynamics of science and research in
contemporary societies. London: Sage.
Glasbergen, P. 2005. Decentralized reflexive environmental
regulation: opportunities and risks based on an evaluation of Dutch experiments. Environmental Sciences 2(4):
427442.
Health Council of the Netherlands. 1995. Niet alle risicos
zijn gelijk: kanttekeningen bij de grondslag van de risicobenadering in het milieubeleid (Not all risks are equal,
a critical approach to environmental health risk policy; in
Dutch). The Hague: Gezondheidsraad.
Heriard-Dubreuil, G.F. 2001. Present challenges to risk
governance. Journal of Hazardous Materials 86(13):
245248.
Hisschemller, M. & Hoppe, R. 2001. Coping with
intractable controversies: the case for problem structuring in policy design and analysis. In: M. Hisschemller,
R. Hoppe, W.N. Dunn & J.R. Ravetz (eds), Knowledge, power, and participation in environmental policy
analysis: 4772. New Brunswick/London: Transaction
Publishers.
Hollander, A. de & Hanemaaijer., A. 2003. Nuchter
omgaan met risicos (Dealing sensibly with risks; in
Dutch). Bilthoven: Netherlands Environmental Assessment Agency.
Hood, C., Rothstein, H. & Baldwin, R. 2004. The government
of risk. Oxford: Oxford University Press.
Hoppe, R. 2002. Cultures of public policy problems. Journal
of Comparative Policy Analysis: Research and Practice
4(3): 305326.
Jordan, A., Wurzel, R.K.W. & Zito, A. 2005. The rise of
new policy instruments in comparative perspective: has
governance eclipsed government? Political Studies 53(3):
477496.
Keijzers, G. 2000. The evolution of Dutch environmental policy: the changing ecological arena from
19702000 and beyond. Journal of Cleaner Production
8(3): 179200.

376

Kersbergen, K. van & Waarden, F. van. 2001. Shifts in governance: problems of legitimacy and accountability. The
Hague: Social Science Research Council.
Kingdon, J.W. 1995. Agendas, alternatives, and public
policies. New York: Harper Collins College.
Klinke, A. & Renn, O. 2002. A new approach to risk evaluation and management: risk-based, precaution-based,
and discourse-based strategies. Risk Analysis 22(6):
10711094.
Kloprogge, P. & Sluijs, J.P. van der. 2006. The inclusion
of stakeholder knowledge and perspectives in integrated
assessment of climate change. Climatic Change 75(3):
359389.
Neumann, P. & Politser, R. 1992. Risk and optimality. In
F. Yates (ed). Risk-taking Behaviour: 2747. Chicester:
Wiley.
Open University. 1998. Risicos: besluitvorming over veiligheid en milieu (risks: decision-making on safety and
the environment, in Dutch; course book). Heerlen: Open
University.
ORiordan, T. 1985. Approaches to regulation. In:
H. Otway & M. Peltu. Regulating industrial risks. Science, hazards and public protection: 2039. London:
Butterworths.
Rayner, S. & Cantor, R. 1987. How fair is safe enough?
The cultural approach to societal technology choice. Risk
analysis, 7(1): 39.
Renn, O. 2006. Risk governance. Towards an integrative approach. Geneva: International Risk Governance
Council.
Riet, O. van de. 2003. Policy analysis in multi-actor policy settings. Navigating between negotiated nonsense
and superfluous knowledge. Ph.D. thesis. Delft: Eburon
Publishers.
Robinson, P. & MacDonell, M. 2006. Priorities for mixtures health effects research. Environmental Toxicology
and Pharmacology 18(3): 201213.
Rothstein, H., Irving, Ph., Walden, T. & Yearsley, R. 2006.
The risks of risk-based regulation: insights from the
environmental policy domain. Environment International
32(8): 10561065.

Runhaar, H.A.C., Driessen, P.P.J. & Soer, L. 2009. Sustainable urban development and the challenge of policy integration. An assessment of planning tools for
integrating spatial and environmental planning in the
Netherlands. Environment and Planning B, 36(2) (forthcoming).
Sluijs, J.P. van der. 2007. Uncertainty and precaution in environmental management: insights from the UPEM conference. Environmental Modelling and Software 22(5):
590598.
Soer, L., Bree, L. van, Driessen, P.P.J. & Runhaar, H.
2008. Towards integration and differentiation in
environmental health-risk policy approaches: An international quick-scan of various national approaches to
environmental health risk. Utrecht/Bilthoven: Copernicus Institute for Sustainable Development and Innovation/Netherlands Environmental Assessment Agency
(forthcoming).
Stoker, G. 1998. Governance as theory: five propositions.
International Social Science Journal. 50(155): 1728.
Sunstein, C.R. 2002. Risk and reason. Safety, law, and the
environment. New York: Cambridge University Press.
U.K. H.M. Treasury. 2005. Managing risk to the public:
appraisal guidance. London: H.M. Treasury.
U.K. House of Lords. 2006. Economic AffairsFifth Report.
London: House of Lords.
UNESCO COMEST. 2005. The precautionary principle.
Paris: UNESCO.
U.S. EPA. 2004. Risk assessment principles and practices.
Washington: United States Environmental Protection
Agency.
VROM. 2004. Nuchter omgaan met risicos. Beslissen met
gevoel voor onzekerheden (Dealing sensibly with risks;
in Dutch). The Hague: Department of Housing, Spatial
Planning, and the Environment.
Wheeler S.M. & Beatley T. (eds.) (2004). The sustainable urban development reader. London/New York:
Routledge.

377

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

What does safety margin really mean?

J. Hortal, R. Mendizbal & F. Pelayo
Consejo de Seguridad Nuclear, Madrid, Spain

ABSTRACT: The term safety margin has become a keyword when discussing about the safety of the
nuclear plants, but there is still much confusion about the use of the term. In this paper the traditional concept
of safety margin in nuclear engineering is described, and the need is expressed of extending the concept to
out-of-design scenarios. A probabilistic definition of safety margin (PSM) is adopted for scalar safety output at
a scenario-specific level. The PSM is easy to generalize (to initiating events, multiple safety outputs, analytical
margins) and, combined with frequencies of initiators and accidents, makes up the plant risk. Both deterministic
and probabilistic approaches to safety assessment found easy explanation in terms of PSM. The role of the
probabilistic margins in the safety assessment of plant modifications is discussed.

INTRODUCTION

In recent years, the international nuclear community is

becoming more and more concerned by the possibility
that significant changes in plant design or in operation
strategies result in adverse side effects usually referred
to as erosion of safety margins. A number of initiatives have been launched to address this problem due
to the increasing number of plants applying for power
uprates, life extensions, increased fuel burn-up, etc.,
where some voices claim that, even complying with
applicable regulations, there could be an unacceptable
loss of safety margins. Moreover, the development
of new designs for nuclear power plants where the
existing technical regulations for LWRs are not necessarily applicable rises the need to establish criteria for
determining what is an acceptable level of safety.
A second reason for the discussion about safety
margins is the increasing trend to apply the so-called
Risk-informed Regulation (RIR). From the pioneering Regulatory Guide 1.174 of the USNRC, most
safety standards and guides on this matter ask for
maintaining enough safety margins as a condition
for acceptability of any change being licensed in the
framework of RIR.
For these and other reasons, the term safety margin has become a keyword when discussing about the
overall safety of the plants but there is still much confusion about the use of this term. This paper is intended
to propose a consistent concept of safety margin that
could be helpful in the discussions about overall plant
safety without losing its traditional meaning.

THE TRADITIONAL CONCEPT

OF SAFETY MARGIN

The introduction of safety margins in traditional engineering is a protection design technique aimed at providing some additional protection capability beyond
the one that is considered strictly necessary. The benefit of using safety margins is two-fold. On one hand,
they allow to accommodate tolerances for little known
phenomena, uncertainties in model data, variabilities
in initial or boundary conditions, and so on. On the
other, they result in a significant simplification of
the design methods as they allow to split the design
in several decoupled stages where the applicable criteria are not too closely linked to the details of the
phenomenology considered in the design analyses.
This approach, essentially deterministic, is considered conservative in most cases and, therefore,
provides confidence that the protection is able to cope
with or to mitigate challenging situations, including
some that were not considered in the design analyses.
In addition, it is convenient for developing safety regulations. This idea was also applied from the beginning
in the nuclear industry and, in particular, in the analysis
of Design Basis Transients and Accidents (DBT&A,
from now on referred to as DBT) where the capabilities
of the protection are assessed. A set of well defined,
enveloping scenarios, classified into a few frequency
classes, are taken as design basis for the protection
and a set of safety variables are used as damage indicators or as indicators of challenges to the protection
barriers. For this limited set of design basis scenarios

379

Figure 1.

Safety margins in the analysis of Design Basis Transients and accidents.

it is possible to define class-specific acceptance criteria in terms of extreme allowed values of the safety
variables, also called safety limits.
In this context, the concept of safety margin is
applied on a scenario-specific basis and its meaning
can be agreed without much difficulty. However, even
at single scenario level, a great variety of margins
appear and all of them can be properly called safety
margins. Figure 1 tries to represent these margins and
how they relate to each other.
In this figure, the two left-most columns represent
the barrier analysis and the two on the right side represent the analysis of radiological consequences. In the
first column, a particular safety variable in a particular
DBT is represented. Since the DBT is an enveloping
scenario, the extreme value of the safety variable in the
enveloped transients will stay below the value of the
same variable in the DBT which should, indeed, stay
below the acceptance criterion or safety limit. There
will be as many left-most columns as the number of
safety variables times the number of DBT. This is indicated in Figure 1 by the dashed ellipse entitled Other
S.V. and Acc. Crit. In every one of these columns
there will be an Analytical Margin and a Licensing
Margin.
Each safety variable and its corresponding safety
limit are selected to prevent a particular failure mode of
a protection barrier. However, the safety limit is not a
sharp boundary between safety and failure. Overpassing the safety limit means that there are non-negligible
chances for a given failure mode but, in most cases,
there is a margin (the Barrier Margin in figure 1)
between the safety limit and the actual failure. A given

failure mode of a particular barrier can result from a

variety of transients, as indicated by the converging
arrows linking the two first columns of figure 1.
As in the previous case, there are several possible
modes of failure of each barrier, as indicated by the
dashed ellipse Other barrier failures. Each combination of barrier failures and type of accident gives rise
to a particular release of radioactive products (source
term), as indicated by the converging arrows linking the second and third columns. Again, a limited
set of enveloping DBT is selected in order to perform the source term analysis. These DBT will be,
in general, fewer and different from those used in
the barrier analysis. The selection of DBT for source
term analysis and the analysis of these DBT to confirm that they remain below the Source Term Reference
Limit introduce two new margins, identified as Source
Term Analytical Margin and Source Term Margin in
figure 1.
Finally, the radiological effects of the source term
are calculated in terms of doses. The use of the Source
Term Margin allows to decouple the dose calculations
from the source term analysis. If the doses resulting
from a release equal to the Source Term Reference
Limit are lower that the Authorised Dose Limit, any
change in the source term analysis does not force a
recalculation of doses, provided that the new source
term remains below the reference limit. The drawback
of this approach is that it could result in a more difficult
application of the ALARA principles. In any case, the
difference between the calculated dose and the Authorised Dose Limit is an additional margin, identified as
Dose Margin in figure 1.

380

A Global Plant Margin is indicated in figure 1.

However, this is only a qualitative concept. Note that
each column of figure 1 corresponds to different physical magnitudes and, therefore, they cannot simply be
summed-up. In addition, we have concurrent margins
that cannot be easily combined into a single margin
measurement. It is clear that an adequate Global Plant
Margin can only exist if all the partial margins exist.
Moreover, the larger the partial margins, the larger the
plant margin. However, a quantification of the Global
Plant Margin is not possible.

THE NEED TO EXTEND THE CONCEPT

OF SAFETY MARGIN

The difficulty to quantify the Global Plant Margin

resulting from the analysis of DBT is not the only
limitation of this concept of plant margin. Worldwide
experience on the operation of nuclear power plants
showed soon that the exclusive use of the analysis
of DBT to assess the plant safety could be insufficient. Some events, especially the TMI accident,
showed that more complicated scenarios, resulting
from out-of-design sequences of events needed to be
addressed. The question of how to deal with so many
possibilities made it inevitable to better evaluate their
frequencies in order to weight their relative importance. This gave rise to the incorporation of system
reliability engineering techniques, as it had been advocated by some precursor studies, like WASH-1400 in
USA or the Deutsche Risikostudie Kernkraftwerke in
Germany. Among other important lessons learned
from this experience was that operators and their
actions were needed but not necessarily beneficial, so
their impact should be taken into account.
Probabilistic Safety Assessment (PSA) techniques
implement these new aspects of the safety analysis
but they have been applied only to the assessment of
severe accidents and their consequences. Other types
of accidents, more likely to occur but resulting in lower
consequences, were left out of scope of PSA. As a
consequence, the rules on the use of PSA for licensing
purposes, when existing, often include a requirement
to demonstrate that enough safety margin is maintained. What is, then, enough safety margin? This
question cannot be answered neither with the analysis
of DBT only nor with PSA only. The need to answer
this question has given rise to some international initiatives. In particular, the NEA Committee on the Safety
of Nuclear Installations (CSNI) promoted in 2003 an
Action Plan on Safety Margins (SMAP 2007). A working group was established that developed a framework
for integrated assessments of the changes to the overall
safety of the plant as a result of simultaneous changes
in plant operation or design.

The analysis framework proposed by the SMAP

group consists of an adequate combination of traditional deterministic (analysis of DBT) and probabilistic (PSA) techniques. The use of event and fault trees,
similar to those of PSA, applied to the assessment
of any safety objective, is combined with simulation techniques, typical of DBT analyses, in order
to quantify the exceedance frequencies of the limits
that define the safety objectives being assessed. These
exceedance frequencies are then used as risk indicators
that characterize the overall plant safety.
The concept of safety margin proposed in this paper
is an extension of the traditional concept of safety margins (as depicted in figure 1) and is especially suited
for methodologies aimed, like SMAP, to answer the
question on the sufficiency of safety margins.
4

DEFINITION OF SAFETY MARGIN

According to our previous discussion, the safety margin should be defined as a measure of the distance
from a calculated safety output to some limit imposed
on it. In general, both elements are calculated variables, which are uncertain magnitudes because they
inherit the uncertainty of the inputs to the calculations
and moreover they incorporate the uncertainty from
the predictive models being used. We will adopt in
this paper the classical representation of uncertain
magnitudes as random variables.
Let us consider a transient or accident A in an industrial or technological facility, and a scalar safety output
V calculated for A. V will be, in general, an extreme
value (maximum or minimum) of a safety variable in
a certain spatial domain during the transient. We will
symbolize by V the random variable representing such
safety output as calculated with a realistic model M.
Let us also suppose an upper safety limit L for V, and
assume that L is time-independent. In this case, V will
be the maximum value of a safety variable. This simple setting will be considered through the paper, the
treatment for other settings (e.g. when L is a lower
limit and V is a minimum during the transient) being
completely analogous.
In general L can be considered a random variable
as well, because it can have some uncertainty. For
instance, L can be a damage threshold from some
safety barrier, obtained directly from experimental
measures or from calculations with a model M. The
variable
D LV

(1)

is sometimes called safety margin in the Reliability

Engineering realm. But our intention is to reserve such
denomination for a conservative function of D.

381

When V < L, the acceptance criterion for V is

fulfilled, and the result is successful or acceptable.
On the contrary, when V > L, the limit is violated
and the result is failed or unacceptable. When L has
no uncertainty, the region V < L is the acceptance
region for V and the complementary set V L is
the rejection or forbidden region for V. Whenever L
is uncertain such regions can still be defined, being
aware that they have random or uncertain boundaries.
With this proviso, the safety margin can be defined
as a measure of the distance from the safety output to
the forbidden region V L. Uncertainties are to be
considered in the safety margin definition.
The view of V and L as random variables is analogous to the load-capacity formulation in Classical
Reliability (figure 2). V can be viewed as the load
on a static component, whose capacity or resistance
is L. The component fails whenever its capacity is
overcome by the load.
A classical election for the safety margin of V
in A is:
SM (V ; A) Max{k D , 0}

PSM (V ; A) PR {V < L/A}

PSM (V ; A) PR {D > 0/A}

(3)

(5)

namely, it is the probability of V being under its limit,

conditioned to the occurrence of the accident A. The
probabilistic safety margin (PSM) is the probability
that the calculated V during A is in the safe region.
Such probability is implicitly conditioned to the use
of the predictive model M.
A general definition of the PSM is:
PSM (V ; A) PR {V RV /A}

(6)

where RV is the acceptance region of V, i.e. the

region where V must stay as a safety requirement.
The complementary of RV is the forbidden region. As
already stated, the boundaries of these regions (e.g.
safety limits) can be uncertain.
The expression (4) can be written in terms of the
probability distributions of both V and L. In the sequel
we will suppose that both V and L are continuous random variables with probability density functions (pdf),
and that they are independent, as it most commonly
occurs in safety analysis. Then


V = Vi with probability

(4)

or, from (1):

(2)

D being the -quantile of the random variable V

and a small number in the range [0,1]. This means to
choose as safety margin a low quantile of the difference
D, provided it is nonnegative. k is a constant chosen
so that the margin is nondimensional.
The problem of a definition like (2) is that it does not
provide an easy way of combining safety margins. Let
us as consider an initiating event IE and let A1 , . . ., AS
be the accident sequences deriving from it (i.e. they
are the combination of IE with one or more concurrent failures, including single and coincident failures).
The safety output V for IE is defined as the following
random variable:

pi PR(Ai /IE)

for every sequence Ai , i = 1, . . ., S and the problem

is posed of how to combine them in order to obtain
the safety margin for the initiator IE. With a definition such as (2) the combination of safety margins in
not straightforward. Therefore, a better definition of
margin should be found.
A probabilistic definition of safety margin has been
proposed (Martorell et al. 2005, Mendizbal et al.
2007). For our setting, we define:

PR {V < L/A} =


fL (s)


fV (z; A) dz ds
(7)

where Vi is the output as calculated for the i-th

sequence. Definitions (1) and (2) can then be applied to
IE. But let us suppose that we know the safety margins

or, in a more compact form


PR {V < L/A} =

fL (s)FV (s; A) ds

(8)

where fL is the pdf of L and fV , FV are respectively

the pdf and cumulative distribution function (cdf)
of V conditioned to the occurrence of A. This is a
convolution integral, extended to the range of L.
The advantages of defining the safety margin as a
probability are:
Figure 2.

Load-capacity formulation of safety margins.

PSMs are nondimensional and ranging in the

interval [0,1]

382

 Its use can be extended to other situations, for

instance to analytical safety margins
PSMs can be generalized to multidimensional
safety outputs and to combinations of accident
sequences (e.g. initiating events)
PSMs combine according to the laws of probability
The probability in (4) and (5) corresponds to the
calculation uncertainty of V and L, sometimes called
epistemic uncertainty, as opposed to the aleatory
uncertainty arising from the unpredictability of the
accidents (i.e. related to the occurrence of initiators,
additional failures, some phenomena, etc). In some
sense, thus, the PSM for an accident sequence reflects
the lack of knowledge about the behaviour of safety
variables.
The probability of V exceeding the limit is one
minus the PSM:
1 PSM (V ; A) = PR{V L/A}

(9)

Let us now recall the initiator IE with the derived

sequences A1 , . . ., AS . The definition (4) can be
extended to IE:
PSM (V ; IE) PR {V < L/IE}

(10)

which, according to the law of total probability, can be

expressed as:

PSM (V ; IE) =

S


pj PSM (V ; Aj )

(11)

PSM (B; A) =

PR{V k < Lk /A}

k=1

PSM (V k ; A)

(13)

k=1

This is another example of how probabilistic margins can merge into another margin.
Now let us focus again on the scalar safety output V
and consider all the initiating events IEi , i = 1, . . . M
that can start accidents challenging V. The frequency
of V exceeding the limit L is:
(V > L) =

S


i (1 PSM (V ; IEi ))

(14)

i=1

where i is the frequency of IEi . In (14) the frequencies of initiators combine with the exceedance
probabilities:
1 PSM (V ; IEi ) =

S


pij (1 PSM (V ; Aij ))

(15)

j=1

Aij being the j-th sequence evolving from the i-th

initiator.
The epistemic uncertainty about V, represented by
the PSM, merges in (14) with the aleatory uncertainty
represented by the frequencies i .
We conclude that probabilistic safety margins can
combine with initiator frequencies and produce
exceedance frequencies, which constitute the plant risk.

j=1

5
That is, the margin for the initiator is a weighted
average of the margins for sequences, the weight being
the conditional probability. This is an example of how
probabilistic margin combine. The same expression
holds for the exceedance probabilities.
Now, let us suppose a safety barrier B having several
failure modes, the i-th failure mode being typified by
a safety output Vi with an upper safety limit Li , i =
1, . . ., F. A safety margin can be assigned to the barrier,
conditioned to the accident A:

PSM (B; A) PR

F



(V < L )/A
k

(12)

k=1

which is the probability of no failure conditioned to A.

It is a generalization of the PSM for a multidimensional
safety output. Whenever the random variables Vi are
independent, the probability in (12) factorizes and

CALCULATION OF PROBABILISTIC
SAFETY MARGINS

If the probability distributions of V and L are known,

the PSM is calculated as a convolution integral. But
such situation is rare. Very often, the calculation of
V is difficult and time-consuming, so that the large
random samples needed to confidently calculate the
probability distribution are almost unthinkable.
The PSM can be calculated by means of Monte
Carlo methods. Random values of V are obtained by
randomly sampling the inputs to the model M and performing calculations for the accident A. The same
procedure may yield random values for the limit L.
Then, PSM can be estimated by means of statistical methods. For instance, the same methods used in
Reliability for the estimation of component failureon-demand probabilities can be applied to the PSM;
a survey of such methods can be found in (Atwood
2003). For an exposition about the estimation of PSM
we refer to (Mendizbal et al. 2007, 2008). Strategies

383

FV is an increasing function, and hence

of variance reduction (e.g. importance sampling) can

be implemented in the Monte Carlo method.

6

Conservative methodologies

In the frame described in section 4, let us consider

a DBT calculated with a conservative methodology,
producing a conservative value Vb of the safety output
V. This means that the probability of V being less than
Vb conditioned to IE is very close to 1. Notice that
such probability is formally a PSM of V with respect
to Vb , and therefore has the sense of an analytical
safety margin.
There is a quite interesting property of the probabilistic safety margin concerning the DBTs. The convolution integral (8) can be split around the constant
value Vb :


Vb

fL (s)ds

Vb

= FV (Vb ; IE) [1 FL (Vb )]

Conservative, wherein the DBT consequences are

calculated with conservative models and hypotheses
Realistic or best-estimate, wherein the DBT consequences are calculated with realistic models and
hypotheses, and its uncertainty is evaluated and
included in the comparison with the safety limits. They are also called BEPU (best-estimate plus
uncertainty) methodologies.

PR {V < L/IE} =


fL (s) FV (s; IE) ds > FV (Vb ; IE)

Vb

PROBABILISTIC SAFETY MARGINS AND

DETERMINISTIC SAFETY ASSESSMENT

The deterministic safety assessment (DSA) evaluates

the response of a facility against initiating events using
the so called design basis transients (DBTs) (Pelayo &
Mendizbal 2005). For each initiating event IE one
or more DBTs are outlined, defined as enveloping or
bounding events. This means that the consequences
of a DBT are worse than those of the great majority of transients deriving from the initiating event. An
obvious way of building up a DBT is by adding conservative assumptions to the initiator. Typically, some
additional system failures are assumed (e.g. the single
failure criterion).
Once established the DBTs, their consequences are
calculated and compared to the safety limits in order
to prove that the probability of limit violation is low
enough.
The deterministic methodologies can be classified
as:

6.1

(17)

Introducing (17) in (16) and expressing the cdfs as

probabilities, the inequality
PR {V < L/IE} > PR {V Vb /IE} PR {Vb < L/IE}
(18)
is derived, that can be expressed in terms of probabilistic margins:
PSM (V ; IE) > PR {V Vb /IE} PSM (V ; DBT )
(19)
(19) says that the probabilistic safety margin for V
is higher than the PSM for the DBT multiplied by the
probability of V being lower than Vb , that can be called
the analytical PSM for V. This is the product of the two
shaded areas in Figure 3.
As we pointed out, the DBT is chosen so that the
analytical PSM is close to 1. Then, (19) states that
a sufficient condition for the PSM of V to be high
enough is that the PSM for the DBT is close to 1.
The inequality (19) contains the essence of the
DSA. If a clearly conservative transient is outlined
and it is shown to be clearly under the safety limit
(high PSM), then it can be assured that the PSM for
the initiating event is high as well.
Let us now recall the expression (14) for the exceedance frequency of the safety limit L. If an enveloping transient or DBT is set up for each initiating

fL (s) FV (s; IE) ds

+

Vb

fL (s) FV (s; IE) ds

(16)
Figure 3.

384

Lower limit to the probabilistic safety margin.

event, the inequality (18) can be introduced in (14)

to yield:
(V > L) <

M


k [1 PR(V Vbk )PR(Vbk < L)]

k=1

(20)
where Vbk is the value of V calculated for the k-th
design basis transient. (20) trivially transforms into
(V > L) <

M


k PR(V Vbk )PR(Vbk L)

k=1

M


k PR(V > Vbk )

(21)

k=1

(21) provides an upper bound to the exceedance

frequency of the limit L. The first addend in the right
hand side represents the contribution of the transients
enveloped by the DBTs (those within the design basis)
and the second one is the residual contribution stemming from the not enveloped fraction (beyond design
basis sequences). The right hand side of (21) gives
the exceedance frequency of L supposing that the
not enveloped fraction gives rise to a limit violation
(i.e. progress beyond design basis assumptions) and
assigning the probabilistic margin of the DBTs to the
enveloped fraction of transients. If the DBT is adequately chosen, the residual term can be neglected
against the main one, and the maintenance of the safety
margin is assured through the enveloping character and
the safety margin of the DBTs. This is the basis of the
deterministic design approach.
6.2

Realistic methodologies

In the realistic methodologies of DSA, the value Vb

of the output V in the enveloping transient is no
longer a constant value, but an uncertain variable, its
uncertainty coming from the calculation process. But
the cornerstone of the method is still valid: the high
safety margin is assured through the definition of an
enveloping transient with a safety margin high enough.
The DBT in realistic methodologies produces less
conservative values of the safety outputs than those
found in the conservative methodologies, because realistic models and hypotheses are used throughout and
few pessimistic assumptions (e.g. single failure criterion) are maintained. The uncertainties of the outputs
are estimated instead.
The probabilistic definition of safety margins is
implicit in the nuclear regulation which refers to realistic methodologies applied to DSA. In 1989, the
USNRC issued the Regulatory Guide 1.157 in order

to give guidance to loss-of-coolant accident (LOCA)

analyses performed with best-estimate models. This
guide was the realistic counterpart to the overconservative guidance provided in 10 CFR 50.46 and the
Appendix K to 10 CFR 50, where the acceptance criteria for a LOCA analysis were spelled out in terms
of several safety outputs (peak cladding temperature,
local maximum oxidation of the cladding, core-wide
oxidation) and the limit that they could not surpass. In
RG 1.157, the requirement was that during a LOCA
the limits are not to be violated with a high probability.
The ordinary deterministic criteria were transformed
into probabilistic criteria, the probability being related
to the uncertainty of the calculated safety outputs.
The BEPU methodologies, supplemented by statistical procedures as those hinted in section 5, can be
used to estimate the PSM for DBTs. It is important to
point out that the estimated PSM has statistical uncertainty, stemming from the finite size of the random
samples. Therefore, an acceptance criterion for the
PSM of the enveloping transient should read:
PR{PSM (V ; DBT ) > M0 } 1

(22)

that is, the margin must be higher than M0 , a value

close to 1, with a high statistical confidence ( is a
low value, say 0.05). When the statistical sample is
large enough (a possibility if the calculations with M
are not time-consuming) (22) simplifies to:
PSM (V ; DBT ) > M0
7

(23)

THE ROLE OF PSM: PLANT

MODIFICATIONS AND PSM AFFECTATION

So far we have defined probabilistic safety margins as

building blocks of the NPP risk. It is on the regulatory
side to decide the magnitude of the tolerable damage
exceedance frequency as a function of the damage,
making up the risk curve against which the design
should be checked and a safety margin preserved.
It should be noted, see (14), that the exceedance
frequency of a safety limit is not conditional to any particular event or sequence but cumulative to all of them.
The design involves the characterisation of the initiating events covering all plant conditions and modes
of operation (ANS 1983, IAEA 2001) and the delineation of the bounding cases, not necessarily only one
case per IE. It is the core of the designers job the building of the bounding case and the methodology used to
obtain the exceedance probability of the safety limit,
and a description exceeds the target of this paper.
The classification of the postulated events (within
the design basis) is made according to the expected
frequency of the initiating event or of the accident
sequence. This classification approach should be

385

made consistent with the above mentioned assertion

regarding the cumulative nature of the exceedance frequency criterion. No such classification is made within
the realm of PSA, where design basis postulates break
down.
In this section we will tackle the role of safety margins when modifications in a plant are being assessed.
It is common practice to evaluate the impact in safety
of plant modifications with a double perspective: i)
deterministic (preservation of the design bases) and
ii) probabilistic, estimating the impact of the change
on exceedance frequencies of some safety limits.
Concerning i), the impact is customarily assessed
against DBTs and verified that specific limits are
not violated. Current regulation requires that for a
plant modification being acceptable it should be verified that the frequency of initiating events is not
affected and that enough safety margin is preserved
for the DBTs. Both elements appear in the right hand
side of (21). A thorough evaluation of the new plant
conditions is required to confirm that the new plant
operating conditions do not increase the expected frequency of initiating events or create new ones, that
equipment needed to cope with events will do according to design and that the magnitude of the damage
is kept low. Due to the limited number of events and
the bounding approach used in the sequences delineation, PSM estimation is straightforward, although
not an easy task, by means of the BEPU methodologies (Boyack 1989). The frequencies of accident
sequences should be calculated by means of probabilistic techniques, including operational experience
and precursory analyses. It is important to note that, in
standard conservative methodologies, where the safety
limits are given without uncertainty, the main term in
(21) has only contributions from the DBTs producing
limit violations.
As an example, following a power plant increase,
it has been observed that the operating conditions of
plant equipment will become more demanding of current intensity on the main plant transformers implying
that the frequency of load rejection events may notably
increase. Similarly the new enveloping case assuming
a new power rate will result in a decrease of the PSM.
Both effects tend to augment the upper bound in (21).
As a result a plant modification will require design
changes in a way as to preserve the frequency of initiating events (modifications on the main transformer)
and the preservation of the safety margin (limiting
fuel peaking factors, new fuel designs, . . .). These
results can also be achieved by reducing the epistemic
uncertainty and the conservative bias of the simulation
models used in the safety outputs calculation (hidden
safety margins).
The evaluation of PSM affectation from the probabilistic perspective is nowadays subject to intense
research (SMAP 2007). Several approaches are being

explored. Following the standard PSA formalism, the

possibility to make use of a limited number of bounding cases like those presented before, breaks down due
to the fact that sequence delineation becomes more
complex as the number of headers of the associated
event tree increases. On the contrary, thermalhydraulic
calculations within PSA are not making use of the
BEPU methods and uncertainty only stems from considerations linked to the failure rates of event tree headers (uncertainty on the frequency of the sequence).
Within Consejo de Seguridad Nuclear two convergent approaches are being studied. One of them rests
on the concept of dynamic event trees where sequence
delineation is stochastic but conditioned by the past
history being generated (Izquierdo & Caamn 2008).
An alternate short term approach makes use of the
BEPU methodologies approach and of current PSA
results (Martorell et al. 2005). Such method focuses on
the analyses of standard success sequences where
an aggregated PSM and frequency of exceedance are
generated. Event tree failed sequences are assumed to
have PSM = 0 and a contribution to the frequency
of exceedance comes from the success sequences
once BEPU methods are applied to the most relevant
sequences. In both cases the incorporation of uncertainty stemming from the simulation (mainly thermohydraulic) has profound implications. For instance, the
concept of cut set should be revised in the presence of
such TH uncertainty, and interpreted as a collection of
basic events whose simultaneous occurrence implies
the violation of the safety limit with a high probability.

CONCLUSIONS

The role of safety margins in Safety Analysis of

nuclear plants is examined through the present paper.
The need of a general definition of safety margin is pointed out, and a probabilistic definition of
safety margin is proposed. Probabilistic margins can
combine with other probabilities and frequencies to
make up the exceedance frequencies of safety limits.
Both the deterministic and probabilistic approaches
to safety analysis found easy explanation in terms
of probabilistic safety margins, as well as the safety
assessment of plant modifications.

REFERENCES
ANS (American Nuclear Society) 1983. Nuclear Safety Criteria for the Design of Stationary Pressurized Water
Reactor Plants. ANSI/ANS-51.11983.
Atwood, C.L. et al. 2003. Handbook of Parameter
Estimation for Probabilistic Risk Assessment. Sandia
National LaboratoriesU.S. Nuclear regulatory Commission. NUREG/CR-6823.

386

Boyack, B. et al. 1989. Quantifying Reactor Safety Margins. Prepared for U.S. Nuclear Regulatory Commission.
NUREG/CR-5249.
IAEA. 2001. Safety Assessment and Verification for Nuclear
Power Plants. Safety Guide. Safety Standard Series No.
NS-G-1.2 , 2001.
Izquierdo, J.M. & Caamn, I. 2008. TSD, a SCAIS suitable
variant of the SDTDP. Presented to ESREL 2008.
Martorell, S. et al. 2005. Estimating safety margins considering probabilistic and thermal-hydraulic uncertainties.
IAEA Technical Meeting on the Use of Best Estimate
Approach in Licensing with Evaluation of Uncertainties.
Pisa (Italy), September 1216, 2005.
Mendizbal, R. et al. 2007. Calculating safety margins
for PSA sequences. IAEA Topical Meeting on Advanced

Safety Assessment Methods for Nuclear Reactors. Korea

Institute of Nuclear Safety, Daejon, October 30th to
November 2nd 2007.
Mendizbal, R. et al. 2008. Probabilistic safety margins:
definition and calculation. Presented to ESREL 2008.
Pelayo, F. & Mendizbal, R. 2005. El carcter determinista del anlisis de accidentes en centrales nucleares.
Seguridad Nuclear 35: 2029.
SMAP (Task Group on Safety Margins Action Plan). 2007.
Safety Margins Action PlanFinal Report. Nuclear
Energy Agency. Committee on the Safety of Nuclear
Installations, NEA/CSNI/R(2007)9.

387

Legislative dimensions of risk management

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Accidents, risk analysis and safety managementDifferent perspective

at a Swedish safety authority
O. Harrami
Swedish Rescue Services Agency, Karlstad, Sweden
Department of Fire Safety Engineering and Systems Safety, Lund University, Lund, Sweden

M. Strmgren
Swedish Rescue Services Agency, Karlstad, Sweden
Division of Public Health Sciences, Karlstad University, Karlstad, Sweden

U. Postgrd & R. All

Swedish Rescue Services Agency, Karlstad, Sweden

ABSTRACT: This study gives an overview of approaches for risk management used within the Swedish
Rescue Services Agency (SRSA). The authoritys commission covers a broad spectrum of safety issues. Group
interviews were performed within different sectors of the organisation. The results show that several perspectives
on accidents and different understandings of safety terms exist within the SRSA. How the organisation uses risk
analyses and carries out risk evaluations differs among the sectors. The safety work includes various types
of accidents, injuries and incidents. The SRSA also use a variety of strategies for safety based on tradition,
legislation and political direction. In such an extensive safety authority, it is not unproblematic to coordinate,
govern and regulate safety issues. Different safety paradigms and risk framings have created problems. But these
differences can also give opportunities to form new progressive strategies and methods for safety management.

INTRODUCTION

The Swedish Rescue Services Agency (SRSA) is a

governmental authority that works with public safety,
accident prevention and risk management. The authority is a result of the merging of a handful former
authorities and fields of activities. The SRSA has a
central role in the management of public safety, which
makes it interesting to study how the authority handles
different kind of safety issues and which principles that
govern the operation. Few studies have been performed
on how Swedish national authorities take decisions
and operate. Lundgren & Sundqvist (1996) studied
how the Swedish Environmental Protection Agency
take decisions when handling permission and authorisation cases. Abrahamsson & Magnusson (2004)
studied nine national authorities work with risk and
vulnerability analysis (SRSA was one of the studied
authorities).
This aim of the study is to give an overview of
perspectives on, and approaches to, accident prevention and risk management used within the SRSA.
The authority has around 800 employees and is a
governmental safety organisation that handles a large

variety of accidents and safety related subjects within

different jurisdictions. The SRSA is active in many
areas of expertise, for example, fire safety, chemical safety, natural hazards and safety for children
and elderly. It covers four main legislations: the Civil
Protection Act, the Transport of Dangerous Goods
act, the Law on Measures to Prevent and Limit the
Consequences of Serious Chemical Accidents (Seveso
Directive) and the Flammable and Explosive Goods
Act. The authority also has commissions in education and international humanitarian relief and disaster
operations.
In 2003 the deputy director-general of the SRSA
commissioned the authors of this paper, who also
were employed at the agency, to obtain an overview
of the use of risk analysis within the SRSA, and
to investigate the need for coordination and education. It was stated that knowledge of risk analysis
methods should be regarded as a strategic area for the
authority. At the same time, expertise on risk analysis was split into several parts of the authority and
the coordination among the experts was poor. Some
other circumstances that have contributed to this study
being initiated are: a generally increase in the use

391

of risk analysis, a development of competences and

methodologies regarding risk management, an extension of the authoritys commission to new risk areas
and a recent reorganisation.
The commission was the start of work that in time
expanded to a more comprehensive study, which also
included subjects such as: the use of safety management terms, perspectives on accidents, strategies and
processes for risk management, actors, tools for analysis of accident risks (e.g. risk analysis), risk evaluation
and safety regulations. The study is described in-depth
and published in a Swedish report (All et al. 2006).
The objective of this paper is to describe and
discuss some themes and underlying principles for
risk management in a multifaceted national safety
authority.

within that sector, and be aware of the traditions and

views prevalent within the sector.
The interviews were performed from October 2004
through March 2005. Each group interview took
approximately three hours to perform. The same structured question guide was used for all group interviews.
The question guide consisted of open questions and
was structured around different themes such as: the
phenomenon of accident and injury, actors, safety
management, common terms, safety legislation and
the use of risk analysis.
After the interviews the preliminary results were
presented to the studied sectors both as a report and
at a workshop. The feedback received was included in
the final results.
In the analyses comparisons were made of terms,
strategies, and how risk analysis and risk evaluation
were performed.

METHOD

Aspects that this paper studies are how the Swedish

Rescue Service Agency works with safety, risk analysis, risk assessment and risk evaluation. Structured
group interviews were performed with 12 different
sectors within the organization. The reason why interviews were preferred to questionnaires was that many
of the questions were difficult to describe with clear
and simple questions. More correct interpretations
were also facilitated since different parts of the organization use terms and definitions in different ways.
In order to structure the application sectors, several logics and methods to divide the organisation
were considered. The final division was based on a
combination of legislations used, traditions, branches
and contexts. The studied sectors are those working
with:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

Fire protection in buildings

Transportation of dangerous goods
Flammables
Explosives
Land use planning
Natural disasters
Environmental and risk appraisal for industry,
including Seveso-establishments
Emergency response support
Safety management in municipalities
National injury prevention program
Emergency preparedness for chemical, biological,
radiological and nuclear (CBRN) incidents
Supervision, in four different jurisdictions

From each sector two or three key persons were

selected to the group interview. The selections of keypersons were made in cooperation between the analysis
group and the department director of each sector. The
key person should have had long experience in working

3
3.1

RESULTS AND ANALYSIS

Safety related terms

How central safety terms are used and understood

within the different sectors has been studied.
Many sectors explain accidents with words like
suddenly, undesired, unintentional, negative
event, event leading to injuries and damages etc.
Several sectors express that an accident is an event
that will initiate a rescue operation. Some sectors only
include unintentional events in the term accident while
others also regard intentional events, like arsons, as
accidents. For some sectors injuries are emphasized
while others focus on damage on property or the environment. There are also examples of sectors where
definitions of the term accident are influenced by, or
stated, in different legislations.
Risk and safety are very central terms for the SRSA
but still they seem to be hard to explain and define.
These terms are often used as technical terms but also
in a more general sense like we have to reduce the
risks and for a safer society. The sector Environmental and risk appraisal for industry refers to official
documents e.g. the Seveso II directive and in national
guidelines when explaining the term risk. These are
examples of technical definitions of risk. Almost all
sectors the studied described that risk, in one way or
another consists of two parameters: probability and
consequences. There are, however, variations in how
the sectors emphasise these two parameters. Three
divergent groups of explanation of the term risk were
found:
1. Probability (P)
2. Consequence (C)
3. A combination of probability and consequence

392

In addition there was a difference in what way the

two parameters should be combined in group 3. Three
types of mathematical functions were proposed: a sum
of P and C, a product of P and C and finally other
relations between P and C. Another interesting finding
is that many of the interviewed sectors did not feel
comfortable when using the term risk.
If risk is a difficult word to explain, the term safety
seems even more challenging to grasp. Many of the
answers were quite vague and general. Only one sector,
the National injury prevention program, gave a more
academic and detailed explanation of the term safety
by defining it as: . . . a circumstance characterised
by an adequate control of physical, material and moral
threats, which contribute to the perception of being
protected from hazards. Safety thus cannot be defined
in an absolute meaning. Safety is a dynamic status.
Safety is not only lack of injuries and hazards. Safety
cannot be restricted to injury prevention. Many of
the other sectors gave explanations like safety is the
opposite of risks, accidents or injuries. From the
broad spectrum of received explanations on the term
safety, four types of meanings were identified:
1. Control of risk
a. Systems and organization to prevent accidents
b. Measures that are taken to prevent accidents
2. Absence or low levels of
a. Accidents
b. Injuries and damage
3. Risk-free or without any risk
4. Feeling of safety and security
The meanings 1 and 2 are in some respect related to
objective measures, quantifications and observations.
The meanings 3 and 4 are more abstract and subjective
than the first two.
The concepts of risk management and safety management (safety work) are both quite common in
Sweden. The concepts have similar meanings both
in a practical and a theoretical sense. How the sectors apprehend these two concepts have been studied.
The answers show that most of the sectors either use
the term risk management or the term safety management in everyday work. Although all sectors work
thoroughly with risk and safety issues, many of them
had difficulties to give a comprehensive explanation of
these two concepts. Common explanations of the two
concepts are: a process, steps, systematic activities and
measures taken in order to improve safety. Risk management is sometimes perceived to be more systematic
and more scientific than safety management. A more
in-depth analysis of the answers and actual practices
reveals that only a minority of the sectors sees any
real differences between the two concepts. A couple

sectors indicate in their answers that they saw the two

concepts as dissimilar.
3.2 Structure and coordination of the work
with public safety
This section starts out with an analysis of accidents that
the sectors work with. It continues with a presentation
of analyses made from four theories and models, in
order to illustrate different views on and structures
of operations at the SRSA. The analyses deal with
questions about the constitution and definition of an
accident, strategies to control safety and strategies for
countermeasures.
An analysis was made of some accidents and types
of events that the sectors work with (Tab. 1). The analysis illustrates the complexity of operations within
the SRSA. Two sectors, Emergency response support
and Safety management in municipalities, work with
almost all of the listed events. On the other hand there
are some sectors that only handle a few types of events.
Some types of events are handled only by a few of
the sectors while other events, e.g. fire and contact
with hazardous substances, are handled by almost all
studied sectors. When several sectors work with the
same type of event, they deal with different aspects
of the event. These aspects could for example depend
on: where the event take place, if certain substances
are involved in the event, which legislations govern
the event, the events origin etc. To exemplify this the
event type fire, which is handled by several sectors
from different aspects, is analysed. The sector Fire
protection in buildings is concerned with this event
if it takes place in buildings. The sectors Dangerous
goods, Flammables, Explosives, Environmental and
risk appraisal for industry and Emergency preparedness for CBRN incidents are concerned with the event
if the fire includes or affects hazardous materials. The
sector Dangerous goods is also concerned with fires
from a transportation point of view. The sector Land
use planning is concerned with this event both from a
hazardous materials and fires in buildings point
of view. The sector Natural disasters is interested in
fire if it takes place in forests or other vegetation. The
sector Emergency response support is concerned with
the event from a fire extinguishing point of view. The
sector Safety management in municipalities has a more
comprehensive approach to the event type fire.
An analysis was made on what constitutes and
defines accidents. There are various literatures on the
nature of accidents (Andersson 1991, Campbell 1997,
Hollnagel 2004, Kjelln 2000). Despite all complex
accident models the results in this study have been
analysed against a simplistic accident model, which
states that an accident is constituted by a combination
of an event and an injury/damage. It was found that
the interviewed sectors focused on the two parts in a

393

Table 1. The table shows which types of event the different sectors work with. Note that the sector 12 (Supervision) has been
excluded in this table.

Storm

Contact with hazardous substances

Explosion

Fire

Arcing and electric shock

Exposure to heat

Exposure to cold

Drowning

Suffocation

Mechanical & construction collapse

Landslide

Flooding

Accidental release of substances

Collision, crash

Trapped in

Lightning

1. Fire protection in buildings

2. Dangerous goods
3. Flammables
4. Explosives
5. Land use planning
6. Natural disasters
7. Environ. and risk appraisal for industry
8. Emergency response support
9. Safety management in municipalities
10. National injury prevention program
11. Emergency prep. for CBRN incidents

Cuts

Sector

Fall

Types of events

Legend: = primarily focus, = semi-focus.

varying extent, and hence a supplement to the model

was made. The supplement displays the sectors views
on accidents based on the main focus, which could be
either:
Event-based
Damage-based
Event- and damage based
Accidents that are event-based will be regarded as
accidents, independent of the magnitude and extent
of the damages. A typical event-based accident is an
unintentional explosion. Damage-based accidents on
the other hand, have the magnitude and extent of the
damage as a starting-point for defining an accident.
A typical damage-based accident is a natural disaster.
Regarding natural disasters, Carr argued that a (natural) disaster is defined by human beings and not by
nature i.e. it should be defined and understood by its
damage. The event itself is not an accident or a disaster
not every windstorm, earth-tremor, or rush of water is
a catastrophe (Carr 1932 in Furedi 2007). For the last
group (event- and damage base), no clear emphasis is
seen on the event or the damage. Both are required
for defining it as an accident. Most of the sectors
have an event-based view on accidents i.e. Fire protection in buildings, Dangerous goods, Flammables,

Explosives and Emergency response support. The

sectors Natural disasters and National injury prevention program have a damage-based view on accidents.
The sectors Environmental and risk appraisal for
industry and Safety management in municipalities
have an event- and damage based view. The remaining sectors used more than one type of the views on
accidents.
Next the operation of the selected sectors were
analysed and categorised in relation to Rasmussens
three types of control strategies: empirical safety control, evolutionary safety control and analytical safety
control (Rasmussen 1994). These strategies describe
the control of safety from the perspective of available information and suitable analysis method. The
findings show that the sectors Emergency response
support, National injury prevention program and
Safety management in municipalities, have their focal
point in the empirical control strategy. Dangerous
goods, Explosives and Flammables have their focal
point in evolutional control strategy. The sectors Land
use planning, Natural disasters and Emergency preparedness for CBRN incidents primarily utilize the
analytical control strategy. Two of the sectors utilize two strategies. Fire protection in buildings uses
both empirical and evolutional safety control, and

394

Environmental and risk appraisal for industry uses

both evolutional and analytical safety control.
The sectors operation were also analysed using a
model that is commonly used at the SRSA describing the safety management process in five stages: 1)
Prevent accidents, 2) Take mitigation actions before
accidents occur, 3) Prepare rescue operations, 4) Carry
out rescue operations, and 5) Take actions after rescue operations. These five strategies to control risk,
described in the model, are divided up both from
the perspective of time and type of countermeasure.
The first three stages are measures that are taken
before an accident, while stage 4 and 5 are measures
taken during an accident and after an accident, respectively. Note that the model has some resemblance
with the generic model Federal Emergency Management Agency (FEMA) uses for disaster and emergency
management, which has four stages: Mitigation, Preparedness, Response, and Recovery (FEMA 1996).
The five-stage model above has been used to determine where in the process the different sectors focus
their work. Most of the interviewed sectors work with
all of these five stages. The work of nine of the twelve
selected sectors have their main focus in stages 1 and 2.
The sectors Emergency response support and Emergency preparedness for CBRN incidents, have their
primarily focus on stages 3, 4, and 5. Only the sector Safety management in municipalities has the same
focus on all stages.
Finally, an analysis was done of the different sectors utilization of Haddons ten strategies for countermeasures. A revised version of the ten strategies based
on Folkhlsoinstitutet (1996), Gunnarsson (1978) and
Haddon (1970) has been used in this analysis:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.

Eliminate the risk

Separate from the risk
Insulate the risk
Modify the risk
Equip to handle the risk
Train and instruct to handle the risk
Warn about the risk
Supervise the risk
Rescue if accident happens
Mitigate and restore

The twelve sectors utilized in some way most of the

strategies for countermeasures presented above. The
dominant strategies for countermeasures are number
3, 4, 5 and 6. Strategies 3 and 4 are focused on actions
that reform and reduce the hazard (risk source), while
strategies 5 and 6 focuses on preparing actions that can
avoid and manage incidents.
3.3

Risk analysis and other similar tools

In this section the term risk analysis is used as a generic

term for all types of investigations, examinations and

analysis that are concerned with risk and safety issues.

This could be e.g. regulation impact assessments,
cost-effect analyses, risk inventories, risk evaluations,
risk assessments, incidents and accident outcomes,
environmental impact assessments, quantitative risk
analysis (QRA), flood inundation mapping and fire
spreading analysis.
A description has been made of how risk analyses
are used in different sectors in the authority. The results
show that risk analyses focuses on very different parts
of the socio-technical system (Rasmussen 1997). The
main part of the work with risk analyses that SRSA
does, that is reviews of conducted risk analyses and
issuing directions on how to perform risk analyses,
is primarily focused on a narrow part of the sociotechnical system. These analyses are delimited to an
operation (site specific), organisation or geographic
sector (often local). The SRSA (i.e. the interviewed
sectors) also conduct their own safety investigations
and risk analyses. Many of these analyses focus on a
comprehensive part of the socio-technical system e.g.
from a national or regional point-of-view. Some analyses were also made from a narrow socio-technical
system point-of-view e.g. in connection with authorisations for an operation. An example of risk analysis
that has been made from a comprehensive sociotechnical point-of-view is risk analysis of the system transportation of dangerous goods in Sweden or
Europe. Other similar examples are the analysis of
expected effects on fire safety and societal costs, conducted in connection with introducing guidelines on
compulsory smoke detectors in residences (Sweden),
and the ongoing investigations about possible requirements on self-extinguishing cigarettes (Sweden and
Europe).
The purpose and aim of the risk analysis was discussed with the sectors. The results show that they
differ between sectors and are unclear in some cases.
The results also show that there is on the one hand an
official purpose with a risk analysis and on the other
hand a more pragmatic (non-stated) purpose. At the
same time as the purpose and aim of risk analysis is
only briefly and often vaguely described in Swedish
regulation, the regulations and guidelines often state
that the purpose and aim of a risk analysis has to be
made clear early in the process. This transfer of the
issue from central to a local and operation-specific
context, makes it possible to direct the shaping of the
risk analysis in the assignment or the actual risk analysis. This is unfortunately not done to any great extent.
The vague specification of the purpose in the regulation also creates insecurity among the stakeholders
and supervising authorities about how the risk analyses
should be conducted and reviewed. The most common purposes or aims found in the actual risk analysis
usually quote what is stated in the regulation, or establish that the purpose is to fulfil the requirement in

395

certain regulation. Fulfilling requirements from insurance companies or from the organisation itself have
also been stated. Another common category of purpose is that the risk analyses shall constitute some
kind of basis for decisions for different stakeholders
and public authorities. Yet another common category
is the purpose to verify or show something. Examples
of this are to verify that the regulation is fulfilled, the
safety level is adequate or that the risk level is acceptable. Other examples are to show that it is safe, more
actions do not have to be taken, the risk is assessed in
a proper manner or that a certain capacity or skill is
obtained. A less common category of purpose is to display a risk overview or a comprehensive risk picture.
An outline for a generic structuring of the different purposes of a risk analysis based on the received answers
could hence be:
1. Formal requirements
a. Legal based
b. Non-legal based
2. Basis for decision-making
a. Public decisions
b. Non-public decision
3. Verification
a. Design of detail, part of a system or a system
b. Risk level
4. Risk overview/comprehensive risk picture
3.4

Evaluation of risk and safety issues

thought that adequate safety was not primarily a question of achieving a certain safety level or safety goal;
instead the focus should be on the characteristics of
the safety work.
The principles and directions used for evaluation of
risk in the authoritys operation were surveyed. Fifteen
types of principles and direction were referred to by the
sectors as starting-points when doing risk evaluation.
Some examples are: risk comparison, economical considerations, zero risk target (vision zero), experience
and tradition (practice), lowest acceptable or tolerable risk level, national goals, the principle of avoiding
catastrophes, a third party should not be affected by
the accidents of others, and balanced consideration
and compromises between competing interests. Some
of the sectors did not have a clear picture of how the
principles were utilized in real cases or situations. Also
in a couple of cases the sectors did not know which
principles that governed the evaluation of risk within
their domain.
Based on some of the results presented above, an
outline for classification of risk criteria was made. The
intention with the outline is to display some different aspects that the criteria focus on and hence also
the evaluations. The outline has five main classes
of criteria: 1) Safety actions and system design, 2)
Rights-based criteria, 3) Utility-based criteria, 4)
Comparisons, 5) Comprehensive assessments. More
about the outline for classification of risk, and other
results on how SRSA conceived and do evaluation is
found in Harrami et al. (in press).
4

An analysis was made on how the evaluation of risk

and safety are conceived within different sectors of
the SRSA. The results show that even though the
sectors daily take decisions and standpoints that incorporate evaluation of risk, some of them do not consider
themselves doing evaluations of risk.
Many of the sectors had difficulties to describe
how they did the assessment of adequate or satisfactory safety. Ten clusters of answers were identified
in the answers given by the sectors on the question
When does the SRSA know that adequate safety is
achieved? A couple of these clustered answers were:
when no accidents, injuries or deaths occur; it can be
shown that a product or an operation does not have a
higher risk level than other similar products or operations; it has been shown that the quantitative risk levels
is lower than established criteria; it has been shown that
certain required equipment or functions exist; a decision becomes a legal case, goes to trial and become a
precedent; all the parties are pleased and nobody complains about the safety level or actions taken. A couple
of the sectors answered that it is not possible to determine if adequate safety is achieved. Also one sector

DISCUSSION

The extent of this study forces the presentation and

analysis of the result in this paper to be quite general. Four experts employed by the authority have
carried out the study. This may have influenced the
results, even if the authors have tried to be balanced
and objective.
The use and understanding of several central safety
terms seems sometimes considerably divergent within
the authority. Some potential problems that have been
identified are that the ability to collaborate, both with
other organisations and within the authority, may be
negatively affected. Poor internal co-ordination could
lead to unnecessary fractionising and bureaucratisation of the work with safety. Possible ways to tackle
these problems could be to discuss terminology issues
continuously and to use terms in a more precise and
distinct manner i.e. define and explain the used terms
better.
The authority uses a wide range of different theories, models and strategies for safety work. However,
a development towards fewer standardised strategies
is not advocated. Instead it is important to increase

396

the understanding about different theories, models and

strategies, and the implications that there might have
on the authoritys work. If used correctly, the diversity
of different theories, models and strategies for safety
work could strengthen the authority and contribute to
develop new ways to work with risk management.
The uses of and knowledge about risk analyses have
always been considered to be crucial for the SRSA.
Even so, most sectors had difficulties in describing
and specifying in what way risk analyses affect safety
and the significance of risk analysis in safety work.
The authors had expected more comprehensive and
profound discussions on the relevance, use, utilization,
content of the risk analyses and how the reviewing
and evaluation of the results was done. If risk analyses
shall continue to be regarded as important and strategic
subject for the organisation in future, there is a need
to specify what function and importance risk analyses
should have in risk management regulated by SRSA.
There is also a need for the sectors to increase their
competence as well as the focus on these issues.
The different sectors had difficulties in expressing
and describing how they assess adequate, sufficient,
acceptable or tolerable safety. They were also uncertain about which values, directions and principles
that governed their operation and how these should
be used. These insufficiencies regarding knowledge
that has a general and non-technical nature may be
an obstacle for good transparency, the understanding
of decisions, communication of decisions and standpoints and support to decision-makers. These findings
are most probably the result of deficient reflections
and discussions about risk evaluation issues within the
authority. In order to develop a more robust assessment
there is a need for SRSA to develop a more substantial foundation for the evaluation of risk and safety
issues.
Harms-Ringdahl & Ohlsson (1995) carried out a
study among eleven Swedish safety authorities. They
found that there were major differences between the
authorities according to their areas of responsibilities,
traditions and operating conditions, manifested by differences in terminology, views on accidents and how
they should be prevented. The results in this paper
show that major differences also exist within a single
authority.
The SRSA is the result of the merging of several
operations of a handful of authorities and fields of
activities. This has resulted in a governmental authority with a widespread operation that holds very different approaches, theories, strategies and practices. The
width and diversity of the operation can be explained
to a large extent by different safety paradigms based
on tradition, legislation, sector boundaries and political directions. The diverse approaches, theories and
models focus on different aspects of safety and risk,
and will result in different analyses and explanations

of causes, effects, relations, correlations and solutions.

Hence, which theory and model that is used probably
affects both the organisation (the structuring of the
sectors) and the operation at the authority. The coordination of the work, e.g. the delimitation of which
aspects of an accident/event different sectors shall handle also depend on which fundamental model and
theory that is applied. Therefore the choice and use
of approaches, theories, strategies and practices must
be conscious and deliberate. According to Hovden
(2004) for over 100 years there have been muddling
through processes to form the safety and rescue institutions and regulatory regimes, and this has resulted in
an over-complex jungle of safety institutions. This
is in some respects the case for the SRSA. The muddling through process within the SRSA has been
ongoing for over two decades and has resulted in an
authority with a very scattered commission and operation. On the other hand there are also advantages with
an integrated management of issues concerning safety,
risk, accidents, disasters, injuries, damage and rescue operations. An authority that handles accidents in
a more comprehensive way can co-ordinate the work
better.
Finally the authors also believe that many of the
issues discussed above will become even more significant in the future. Especially since a new national
authority for safety, security and crisis will be established in January 2009. This authority will take
over the tasks of the SRSA, the Swedish Emergency
Management Agency and The National Board of
Psychological Defence.
5

CONCLUSIONS

The divergence in theories, practice and use of terms

within the SRSA, together with a lack of understanding, complicates the co-operation and coordination
within the SRSA and with other actors within the field
of safety. These differences can also give opportunities
for advancements in safety management. To achieve
this it is necessary for more profound discussions and
reflections on how to perform safety management in
such a diversified safety agency. Subjects that ought
to be discussed are for example:
The basis on which events, accidents, injuries, crisis
and catastrophes should be included in the work at
the authority
Which risk management strategies should be used
The importance of risk analysis and how it should
be utilised in the safety work
Standpoints on fundamental ethics, values and
principles for evaluating risk.
Several of the issues that the authority handle are
complex and do not have any simple and obvious

397

answers. Therefore these issues need to be handled

with humility and caution. It is also important that
the employers have understanding and respect for how
other sectors work in the authority, and that they question and critically examine their own work with safety.
The co-operation within the organisation and with
other authorities would be facilitated if the authority
develops a strategy or a policy on how to handle central
terms.
6

FUTURE WORK

The SRSA has initiated several activities based on

the study, e.g. a project that aims to develop methods and strategies for municipal risk assessment, a
development study regarding risk evaluation. It has
also initiated a project group that co-ordinates risk
communication and central terms and definitions.
ACKNOWLEDGEMENTS
We would like to thank the officials participating in the
interviews for the openness and the Swedish Rescue
Service Agency for making it possible to conduct this
study. We also would like to thank Prof. Lars HarmsRingdahl and an anonymous referee for constructive
comments on the conference paper.
REFERENCES
Abrahamsson, M. & Magnusson, S.-E. 2004. Anvndning av risk- och srbarhetsanalyser i samhllets krishanteringdelar av en bakgrundsstudie (in Swedish).
LUCRAM report 1007. Lund: Lund University Centre
for Risk Analysis and Management.
All, R., Harrami, O., Postgrd, U. & Strmgren, M. 2006.
Olyckor, riskanalyser och skerhetsarbetengra olika
perspektiv inom Rddningsverket (in Swedish). Report
P21-480/07. Karlstad: The Swedish Rescue Service
Agency.
Andersson, R. 1991. The Role of Accidentology in Occupational Injury Research. Ph.D. thesis. Arbete och hlsa,
vetenskaplig skriftserie 1991:17. Stockholm: Karolinska
Institutet.
Campbell, R. 1997. Philosophy and the accident. In R. Cooter
& B. Luckin (eds), Accidents in History: Injuries, Fatalities and Social Relations. Amsterdam: Rodopi.

Carr, L.J. 1932. Disaster and the Sequence-Pattern Concept

of Social Change. American Journal of Sociology 38(2):
207218.
FEMA 1996. Guide for All-Hazard Emergency Operations
Planning. State and Local Guide (SLG) 101. Washington:
Federal Emergency Management Agency.
Folkhlsoinstitutet 1996. P vg mot ett skadefritt Sverige
(in Swedish). Report 1996:117. Stockholm: Swedish
National Institute of Public Health.
Furedi, F. 2007. The changing meaning of disaster. Area 39
(4): 482489.
Gunnarsson, S.O. 1978. Strategies for accident prevention. In R. Berfenstam, L.H. Gustavsson & O. Petersson
(eds), Prevention of accidents in childhood; A symposium
in the series of congresses and conferences celebrating
the 500th anniversary of Uppsala University, held at
the Department of Social Medicine, University hospital,
October 57, 1977, Uppsala. Uppsala: Uppsala University
Hospital.
Haddon, W.J. 1970. On the escape of tigers: an ecologic
note. American Journal of Pubilc Health (December):
22292234.
Harms-Ringdahl, L. & Ohlsson, K. 1995. Approaches
to accident prevention: A comparative study of eleven
Swedish authorities. Safety Science 21(1): 5163.
Harrami, O., Postgrd, U. & Strmgren, M. (in press).
Evaluation of risk and safety issues at the Swedish Rescue Services Agency. In Proceedings of ESREL 2008
AND 17th SRA EUROPE CONFERENCEAnnual Risk,
Safety and Reliability Conference, Valencia Spain, 2225
September 2008. Rotterdam: Balkema.
Hollnagel, E. 2004. Barriers and accident preventionor
how to improve safety by understanding the nature of
accidents rather than finding their causes. Burlington:
Ashgate.
Hovden, J. 2004. Public policy and administration in vulnerable society: regulatory reforms initiated by a Norwegian
commission. Journal of Risk Research 7(6): 629641.
Kjelln, U. 2000. Prevention of Accidents Through Experience Feedback. London: Taylor & Francis.
Lundgren, L.J. & Sundqvist, G. 1996. Varifrn fr
miljvrdsbyrkraterna sin kunskap? In Lars J. Lundgren
(ed.), Att veta och att graOm kunskap och handling inom miljvrden (in Swedish): 129171. Lund:
Naturvrdsverket Frlag.
Rasmussen, J. 1994. Risk management, adaptation, and
design for safety. In: B. Brehmer & N.-E. Sahlin (eds),
Future risks and risk management. Dordrecht: Kluwer
Academic Publishers.
Rasmussen, J. 1997. Risk management in a dynamic society:
A modelling problem. Safety Science 27(23): 183213.

398

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Evaluation of risk and safety issues at the Swedish Rescue Services Agency
O. Harrami
Swedish Rescue Services Agency, Karlstad, Sweden
Department of Fire Safety Engineering and Systems Safety, Lund University, Lund, Sweden

U. Postgrd
Swedish Rescue Services Agency, Karlstad, Sweden

M. Strmgren
Swedish Rescue Services Agency, Karlstad, Sweden
Division of Public Health Sciences, Karlstad University, Karlstad, Sweden

ABSTRACT: This study investigates how evaluation of risk and safety are conceived and managed in
different parts of the Swedish Rescue Services Agency (SRSA). Group interviews were performed within
twelve different sectors of the organisation. The results show that some of the representatives do not consider
themselves doing evaluations of risk, even though they take daily decisions and standpoints that incorporate
evaluation of risk. In most sectors profound reflection and discussion about these issues had only been carried
out to a very limited extent. The different sectors had great difficulties to express or describe how to assess
adequate, sufficient, acceptable or tolerable safety. There is a need for SRSA to develop a more substantiated
foundation for evaluation of risk and safety issues to receive better internal and external understanding of the
decisions, a more transparent process, easier and clearer communication of decisions and standpoints, and better
support to decision-makers.

INTRODUCTION

Most national authorities have an important role in

interpreting legislation in order to: support political
and legislative decisions, issue regulation and guidelines and balance conflicting interest in decisions. At
the same time Reid (1999) have suggested Conflict
arising from conflicting evaluations of given information generally cannot be resolved and This sort
of conflict commonly arises in relation to the sociopolitical context of risk, and it involves competing
interests and incompatible value systems. So how
does a national safety authority handle these difficult
issues?
Lundgren & Sundqvist (1996) is one of the few
studies that have investigated how Swedish national
authorities operate. The study describes how the
Swedish Environmental Protection Agency take decisions when handling permission and authorisation
cases. The Swedish government appointed in 2002 a
commission on public safety in traffic tunnels, which,
amongst other things, investigated the view on risk
evaluation among the concerned national authorities

i.e. the Swedish Road Administration, the Swedish

Rail Administration, the National Board of Housing,
Building and Planning and the Swedish Rescue Services Agency (SRSA). One of the conclusions of the
investigation was that while the competent authorities
agreed on the risk picture, they did not agree on
how the risk evaluation should be made and hence
how risk treatment (safety actions and safety design)
should be assessed for new road and rail tunnels
in Sweden (Boverket 2005). This is an example of
what Reid (1999) describes as conflicting evaluations of given information (e.g. conflicting views of
the acceptability of estimated risks). The two other
types of conflict situations, described by Reid, different information and conflicting information are
more connected to the limitation and estimation of the
risk. The authors earlier experience from risk evaluation in various fields (such as land use planning, large
scale hazardous materials, dangerous goods, natural
hazards, infrastructure risks and fire protection) is that
the stakeholders could have diverging views on early
steps of the risk management process e.g. risk identification or risk analysis. However situations where the

399

stakeholders agree on the risk picture, but disagree

on how to evaluate and assess the risk are even more
common.
The following study on how evaluations of risk and
safety issues are done within the authority, was made as
a part of a larger investigation of how the SRSA works
with accidents, risk analysis and safety assessment.
More findings from the study and a more comprehensive description of how the study was initiated are
found in All et al. (2006) and Harrami et al. (in press).
The SRSA is a governmental organisation working
with miscellaneous types of accidents, safety work and
risk management in many arenas. The authority has
around 800 employees and is active in many areas of
expertise and the legal competence covers four main
legislations: the Civil Protection Act; the Transport of
Dangerous Goods Act, the Law on Measures to Prevent and Limit the Consequences of Serious Chemical
Accidents (Seveso Directive) and the Flammable and
Explosive Goods Act. The authority also has commissions in education and international humanitarian
relief and disaster operations.
The objective of this paper is to describe and discuss
how evaluation of risk and safety issues is carried out
at the SRSA, and what the basis is for the evaluation.

METHOD

This paper studies how the Swedish Rescue Service

Agency works with risk evaluation in its operation.
Structured group interviews were performed with 12
different sectors within the organization. The reason why interviews were preferred to questionnaires
were that many of the questions were difficult to
describe with clear and simple questions. More correct interpretations were also facilitated since different
parts of the organization use terms and definitions in
different ways.
In order to structure the application sectors, several logics and methods to divide the organisation
were considered. The final division was based on a
combination of legislations used, traditions, branches
and contexts. The studied sectors are those working
with:
1.
2.
3.
4.
5.
6.
7.

Fire protection in buildings

Transportation of dangerous goods
Flammables
Explosives (including LPG)
Land use planning
Natural disasters
Environmental and risk appraisal for industry,
including Seveso-establishments
8. Emergency response support
9. Safety management in municipalities
10. National injury prevention program

11. Emergency preparedness for chemical, biological,

radiological and nuclear (CBRN) incidents
12. Supervision in four different jurisdictions
From each sector two or three key persons were
selected for the group interview. The selections of keypersons were made in cooperation between the analysis
group and the department director of each sector. The
key person should have had long experience in working
within that sector, and be aware of the traditions and
views prevalent within the sector.
The interviews were performed from October 2004
through March 2005. Each group interview took
approximately three hours to perform. One structured
question guide was used for all group interviews.
The question guide consisted of open questions and
was structured around different themes such as: the
phenomenon of accident and injury, actors, safety
management, common terms, safety legislation and
the use of risk analysis.
This paper focused on risk evaluation and the
questions used were:
When does the SRSA know that adequate safety
is achieved in the work with decisions, regulations, directions, handbooks, information, guidance, education etc.? How have you decided on
these actions?
Which principles and directions govern the decisions in your sector? Describe the principles.
Are any legal or recommended risk or safety levels
used? How are these levels described? Is it possible
to assess if the levels are achieved?
After the interviews the preliminary results were
presented to the studied sectors both as a report and
at a workshop. Received feedback was included in the
final results.

THE DIFFERENT SECTORS VIEW ON

EVALUATION OF RISK AND SAFETY
ISSUES

Below is a summary of the answers, for each one of

the twelve sectors, on the questions described in the
method section above. The summaries are presented
in the same order and have the same numbering as
the presentation of the sectors in the previous method
section.
The extent and depth of the answers varied a lot
which is apparent from the summaries. One has to
bear in mind that the scope and focus of the commission vary between the studied sectors. There are also
variations in other contextual conditions such as the
characteristics of the risk, who the stakeholders are
and how adjacent legislation is formulated.

400

3.1

Fire protection in buildings

The work of this sector is made within the framework

of the Civil Protection Act that is applicable when
assessing fire protection in existing buildings. The
Building Code regulates the construction period up to
the approval of the building. The legal requirements on
safety actions, shall express the lowest level accepted
by society. The requirements shall also be economically reasonable i.e. the action must give a protection
that is proportional to the cost. It is possible to require
new actions to a certain extent, but considerations
have to be taken to the regulation that existed when
the building was built. These circumstances make
each case unique, which requires profound knowledge
about different parts of the system as well as flexibility when assessing the buildings safety and risk level.
The principles used in the assessment are focused on
saving lives and preventing injuries. A key principle
is that people have to be able to evacuate the building before critical conditions arise. Another important
principle is to protect a third party from injuries and
damages. The interpretation of the regulation and the
assessment of safety levels, are to a certain degree
also dependent on the political climate and on to what
degree recent fires have become a media event.
3.2

Dangerous goods

The objective of the regulatory work for the transport of dangerous goods is to ensure safe transport.
The regulatory work is done within the framework of
the UN Economic & Social Council (ECOSOC) and
is based on negotiations between the member states.
39 countries are involved in the work with the rules
for road transport, and 42 countries work with the
rules for rail transport. There is no set safety target
level but a general principle that is used in the assessment is that the more dangerous the substance is the
higher safety the level is required. The rules are a
balance between enabling transportation and ensuring
safety. This means that the diverse conditions in the
member-states (weather, density of population, road
quality, economy etc) may play an important role in
the negotiations. Since the regulatory work is based
on negotiation the political aspects are dominant and
this may also be reflected in the outcome. Cost-effect
analyses are utilized to some extent. The rules mainly
address the design of the technical system and to some
extent organisational issues. It is difficult to assess if
adequate safety is achieved.
3.3

Flammables

Adequate safety is assessed in official inspections

at the site before permissions are issued. Aspects
that are assessed are the design and performance of

details, subsystems and systems. The assessment is

performed based on experience from normal operations and incidents, as well as continuous dialog and
discussions with companies and the industry. Fulfilment of requirements is primarily done by living
up to accepted practice, and only to a minor extent
by satisfying a certain safety criterion. If the practise is not satisfied, the risk owner has to be credible
when asserting that the activity still fulfils the requirements. There are three main principles that are used in
the assessment: First, as few as possible should be
exposed to risk. Second, the control-room shall be
safe in order to handle control emergencies. Third,
people outside the plant shall not be affected. The
assessment of damages and injuries is based using
a design scenario that is more reasonable than a
worst-case scenario. The design scenario is based on
experience within the industry, such as information
about incidents and similar systems. The rupture of
a tank/cistern is for example not considered to be a
reasonably empirical damage scenario while a broken coupling is. Effects of organisational changes are
very difficult to assess compared to changes in the
technological system. The cost for an action shall be
(economically) reasonable. There are no legal or recommended levels for safety but the regulation states
that as few as possible should be injured. The industry calculates probabilistic risk, e.g. individual risk
(IR) and societal risk (SR), in their risk analyses
and wants the SRSA to use probabilistic risk criteria. The risk analyses that calculate IR and SR are
difficult to assess, primary since the analyses are very
uncertain but also because no probabilistic risk criteria have been set. The sector thinks that the focus
should be on actions and measures for improving
safety instead of focusing on the figures and results
in risk analysis.
3.4

Explosives

A standard model e.g. safety distances is used when

applicable, otherwise a risk analysis should show that
the (process) method doesnt cause larger risk than
similar alternative standard methods. The criteria for
liquefied petroleum gas (LPG) pipes are based on
long traditions of manufacturing steel and on experience from incidents. The principles used for LPG
constructions are: Firstly, the gas has to remain in
the containment. Secondly, in case of leakage the gas
has to be ventilated. Thirdly, ignition has to be prevented. Finally a single mistake or error should not
in it self result in a dangerous situation. The safety
criteria for explosion protection is based on international figures and information (e.g. from NATO) and
adapted for domestic conditions. Three principals are
used for assessing the safety of explosives: An initiation of explosion should be avoided. The explosives

401

should be protected from the surroundings. The surroundings should be protected from the explosives. It
is very difficult for the sector to assess if the desired
safety level is achieved. The absence of accidents is
interpreted as an indication that the safety level probably is good. The applied criteria are absolute but some
consideration of the cost of actions may be taken.
3.5

Land use planning

Most cases are unique. It is therefore important to

collect all possible documents, information and analysis. Usually a comprehensive assessment made within
the frame of the Planning and Building Act, balances
different values and makes assessment of the reasonableness. To a certain extent this assessment also has
to take the Environmental Code into account, which
assumes that certain codes of consideration have been
followed. E.g. that sufficient knowledge and bestavailable-technology have been used, chemicals have
been replaced with less hazardous ones and that the
location of the activity in question is the most appropriate. In local regulations safety distances are utilized
as well as prohibition to establish new hazardous operations. It is not possible to assess if adequate safety is
achieved. One way to approach this issue is to put
safety issues on the agenda. The more safety issues
that are discussed and managed in the planning process the better, and in some sense the closer you get
to adequate safety. Some principles that are used in
the application of risk criteria (Davidsson 1997) are:
(1) the principle of reasonableness i.e. the activity in
question should not imply risks that might reasonably
be avoided or reduced, (2) the principle of proportionality i.e. the risks that an activity gives rise to may not
be disproportionate in relation to the benefits, (3) the
principle of distribution i.e. no individuals or groups
should be put at a risk that far exceeds the benefits
they derive from the activity and (4) the principle of
avoiding catastrophes i.e. manageable accidents with
limited consequences are more preferred than ones
with consequences of catastrophic magnitude.
3.6

Natural disasters

When assessing adequate safety with respect to

flooding, the effects of calculated or estimated
flows is compared to buildings, public functions and
other properties and values that have to be protected.
There are no general recommendations for how
the municipalities should assess flooding risks. The
municipalities decide permissible water levels for
new settlements. The evaluation of landslides is
mainly assesses through safety factor calculations and
national recommended safety factors are utilised.
Much of the safety work performed in this sector is based on guidelines from the Commission on

Slope Stability (1995) and the Swedish Committee

for Design Flood Determination (Fldeskommittn
1990). The latter have recently been updated and
replaced by new guidelines (Svensk Energi, Svenska
Kraftnt & SveMin 2007).
3.7

Environmental and risk appraisal for industry

Most authorizations are new and unique cases. The

work in this sector is therefore similar to the one
described for land use planning (section 3.5). The
authorization decision is based on risk analysis and
on the operators answers to complementary questions by SRSA. It is very difficult to know if the
safety level is adequate. Lately the regulation has been
evolving towards more functional regulation. A couple of precedents exist that contains interpretations
of how safety and risk evaluations within the jurisdiction of the Civil Protection Act and the Law on
Measures to Prevent and Limit the Consequences of
Serious Chemical Accidents (Seveso Directive) should
be assessed. These precedents constitute directions
for future assessments. The sector also refers to the
same principles used in the application of risk criteria
presented earlier in section 3.5 (Land use planning).
3.8

Emergency response support

This sector is both comprehensive and situationdependent, and focuses on two aspects of risk evaluation: general planning of the fire and rescue work in
the municipalities (task, size, location, equipment etc.)
and safety for the personnel during a rescue operation.
The evaluations done in connection with the general
planning vary between different municipalities and are
based on different information and statistics. Statistics
on incidents, injuries and accidents are common information sources. During the last years SRSA has been
developing and promoting other methods which some
municipalities have adopted e.g. cost-effect methods
as well as different measures, indicators and keyratios. Evaluations done during a rescue operation are
to a certain degree directed by regulation e.g. prerequisites for conducting a rescue operation and directions
for certain activities (e.g. fire fighting with breathing apparatus and diving). The rescue commander has
to make fast and sometimes difficult assessments and
evaluations based on limited information.
3.9

Safety management in municipalities

The work of this sector is largely done in the municipalities, and is based on political decisions that give
directions for the safety work. The priority given to
safety issues differs between municipalities. A common guiding principle is that saving life is prioritised
compared to saving property, and saving property is

402

prioritised compared to saving the environment. There

are also continuous discussions about how resources
should be allocated between small and large accidents.
Usually the assessment is done by comparing safety
levels and accident statistics with similar municipalities as well as with the nation as a whole. Target
levels for the citizens protection are in some cases
expressed. Cost-effect methods are used to a certain
extent.
3.10 National injury prevention program
The Safe Community concept promotes (Welander
et al. 2004) systematic safety work at the municipal level. The vision is zero injuries but in reality
technological and other changes cause new types of
accidents and injuries, and this impedes the achievement of safety goals set within the frame of a vision.
The Swedish Public Health Policy gives general directions that guide the assessments (Folkhlsoinstitutet
2003). Each municipality sets its own levels based
on the guidelines for public health. A common way
to direct safety work is to identify the most costeffective actions. Priority is given to saving lives.
Childrens safety is often prioritised when establishing
new housing estates.
3.11 Emergency preparedness for CBRN incidents
Adequate safety is generally very difficult to assess
for CBRN incidents since the uncertainties are large.
The response time for the national oil protection supplies is considered to be sufficient. The emergency
protection and preparedness for nuclear incidents are
also considered to be sufficient, even though the prerequisites for these incidents are continually updated.
The assessments of the safety levels for these incidents are based on relevant scenarios. The preparedness has to be flexible in order to adapt to different
situations.

4
4.1

ANALYSIS AND DISCUSSION

The view of sectors on risk evaluation issues

At first during the interviews some of the sectors

did not consider themselves doing evaluations of risk,
even though they take daily decisions and standpoints
that in one way or another incorporate the evaluation of risk. These tasks include issuing permits and
authorisations, issuing regulation and other guidelines,
developing and providing methods and tools for safety
work, promoting safety in certain lines of business or
arenas, giving comments on submitted proposals etc.
Evaluation seems to be a very integrated, natural and
unconscious part of the everyday work done by the
different sectors.
The reason for the sectors views, not considering
themselves doing evaluation, have not been analysed
thoroughly. However two possible explanations have
been seen, and they are both connected to the background and the role of the officials. Most officials
have a technical or natural science background even
though this is gradually changing. This means that
they are used to working with proposals and solutions
that are expressed in figures and charts, and based on
calculations and statistical methods. Inspired by the
work of Lupton (1999), Summerton & Berner (2003)
and Renn (1992), Hultman (2004) describes seven perspectives on risk within socio-technical systems. Most
SRSA-officials have, and work within, what Hultman
describes as, an engineering perspective on risk.
One explanation could be that the term evaluation
has been interpreted as using a certain systematic, scientific or quantitative method. Their background may
have lead their thoughts to associations about scientific evaluation, which they do not consider themselves
doing. Another explanation could be the indistinct role
of the officials. Most of them are scientific experts by
profession. They also form a link between the legal
world and other experts. If evaluation was interpreted as a legal issue they probably did not consider
themselves doing that either.

3.12 Supervision
Supervision is made within the framework of four legislations, and is based on general information about
the operation, risk analysis and other documents. It
is not possible to process and consider everything in
an operation. Strategic choices are made, on what
to study. With improved experience the supervisors
learn how to receive a good picture of the status of
different operations and how to assess the safety
level. Supervision is made up of two main parts: the
administrative organisation and the hands-on operation. The main starting-point for the assessments is
the intentions of the legislation. Hence the interpretations of the legislation become crucial for the safety
assessment.

4.2

Assessment of adequate or satisfactory safety

Most of the respondents thought that the question of

how they determine if adequate safety is achieved
was relevant. At the same time many of the sectors
had difficulties describing how they did the assessment. Below is a summary of different views on how
adequate or satisfactory safety is assessed. Adequate,
satisfactory, sufficient, acceptable or tolerable safety
is attained when:
no accidents, injuries or deaths occur
tests and examinations show that characteristics
for a product or an activity meet predetermined
thresholds

403

 it can be shown that a product or an activity/operation does not have a higher risk level than
other similar products or activities/operations
all experiences acquired by the involved parties are
used as far as possible
it has been shown that the quantitative safety/risk
level is lower than established criteria
it has been shown that certain required equipment
or functions exist
a process involving certain parties and including
certain stages, has preceded a safety decision
a decision becomes a legal case, goes to trial and
become a precedent
a decision on safety is determined in relation to other
factors (adequate safety is relative)
all the parties are content and nobody complains
about the safety level or actions taken.
Some of the sectors answered that it is not possible
to know or determine if adequate safety is achieved.
One sector thought that adequate safety was not primarily a question of achieving a certain safety level or
safety goal. Instead the focus should be on the characteristics of the safety work. An operation or activity
that works systematically and strives for continuous
safety improvements should be considered to have
adequate safety.
The results show that the sectors have interpreted
the question in several different ways. The sectors
have either answered how requirements in regulation
are satisfied (legal view) or how adequate safety is
assessed from a scientific or from a philosophical
point of view. Also the answers given on this question
could be summarised in three types of main discussions (discourses), where discussion type 1 and 2 were
the most common (Fig. 1).
Type 1: Quantitative risk measures (outcomes and
expectations) were discussed and contrasted to each
other. In most case the discussion also included some
questions or criticisms of these risk measures (to some
referred to as the reality). Type 2: Different safety
actions, necessary functions and ways of managing
safety were discussed and contrasted. In most cases
the discussion also included issues on how to assess
the reasonableness of required safety actions. Type 3:
In a couple of cases the two earlier types of discussions
(type 1 and type 2) were contrasted to each other.
4.3

Principles used for evaluation of risk

and safety issues

All sectors gave example of directions and principles

that were used as starting-points in the evaluation of
risk and safety issues. Below is a summary:
Risk comparison
Ethics and values
Economic considerations

2
3
Outcomes
&
Expectation
values

Actions
&
Functions
&
Way of work

The reality

Reasonableness

Figure 1. Three types of discourses on adequate safety were

identified during the interviews. Most common was the type 1
and 2 discourses.

The precautionary principle

Zero risk target (Vision zero)
Political decisions and directives
Experience and tradition (established practice)
Lowest acceptable or tolerable risk level
National goals (e.g. Public Health Policy)
The principle of avoiding of catastrophes
Limitation and mitigation of consequences
The evaluation is transferred to others
A third party should not be affected by the accidents
of others
Balanced consideration and compromises between
competing interests
Saving human lives are prioritised, in relation to
saving property and the environment
As seen from the results, many sectors refer to
more than one principle. There is sometimes no clear
boundary between two principles. Also some principles may include (parts of) other principles. It was
often unclear to what extent the principles were used
and what importance the different principles had. For
some of the sectors it was not clear how the principles were utilized in real cases or situations. In a few
cases the sectors did not have a clear picture of which
principles governed the evaluation of risk within their
domain. The reason for the uncertainty about the principles, their importance and how they are utilized may
be due to the fact that almost no sector had been carrying out any profound reflections and discussions about
these issues.
4.4 An outline for classification of risk criteria
An outline for classification of risk criteria has been
made based on how the sectors assess adequate safety
and what principles they use in the evaluation process.
The outline is based on the classifications presented by

404

Mattsson (2000) and Morgan & Henrion (1990). These

classifications have been expanded and modified to
better fit the results found in this study, which have
resulted in five groups of criteria:
1.
2.
3.
4.
5.

Safety actions and system design criteria

Rights-based criteria
Utility-based criteria
Comparison criteria
Comprehensive assessments

Group 1 consists of four types of criteria: (a)

Detail regulation, commonly regulates actions that
target technology, operation and to some extent the
organisation; (b) Functional regulation, expressed as
a certain function that has to be fulfilled in a given
situation; (c) Measures, indicators and key-ratios,
which could be a safety distance, an index or a ratio;
(d) Technology-based criteria, that states the best
available technology should be used.
Group 2 include criteria that are rights-based, quantitative and express the highest permitted risk level:
(a) Zero-risk and similar goals for outcomes,
expressed as the accepted level of deaths or injuries;
(b) Individual risk; (c) Societal risk criteria. The criteria (b) and (c) express the highest accepted level for
expected values of deaths.
Group 3 consists of three types of criteria that all
come from utility theory and are variants of utilitybased criteria (Mattson 2000): (a) Cost-benefit: assess
if the sum of all benefits of an action shall exceed
the costs; (b) Cost-effect: assess which action meets
the safety goal to the lowest cost; (c) Multi-attributive
utility criteria: assess different alternatives based on
a combination of several types of preferences with
different weights.
Group 4 consists of criteria with three types of comparisons: (a) Between the risk/activity in question and
similar risks/activities; (b) Between the risk in question and dissimilar types of risks e.g. smoking, driving
and climbing; (c) Between different design alternatives, which is common in environmental impact
assessment.
Group 5: The assessment of safety is done together
with many other factors and the assessment does not
have to be quantitative. The assessment processes do
not have to be as systematic as multi-attributive analysis. The outcome of such an assessment may be
difficult to predict since these assessments are unique
for each case.
In some cases hybrid criteria are used (Mattsson
2000). These combine different criteria e.g. first societal or individual risk analysis is assessed followed by
a cost-benefit analysis. Also many of the presented
criteria may be expressed as guideline values/levels
and not absolute values/levels. The difference is that
the guideline values are not unconditional and can be
seen as something to strive for.

4.5

Assessment of reasonableness and the use

of quantified risk criteria

The issue of reasonableness was common in the

sectors discussion on risk evaluation. The sectors discussed reasonableness both regarding the evaluation
process and the requirements on safety actions. Most
probably this is due to that the applicable legislations
state that requirements on safety actions shall be reasonable. These legislations may have been influenced
by utility theory, but the assessments of the reasonableness made by the courts have an even more comprehensive scope. The few existing precedents (stated by
courts) have interpreted the meaning of reasonable
as: the costs for an action shall be proportional to the
safety turnout of the required action. Also, practical
feasibility as well as the question of existing versus
new objects/activities is taken into consideration when
assessing if an action is reasonable.
In some of the interviewed sectors it is common that
the industry use quantitative risk analyses (QRA) that
present the result as individual and societal risk (i.e.
rights-based criteria). In some European countries,
such as the Netherlands and Great Britain, quantitative risk criteria are used to assess such risk analyses
(Ale 2005). There are no set quantitative risk criteria
in Sweden that can be applied when assessing quantitative risk analysis. Therefore the industry (or a single
risk owner), either set their own criteria or refers to
levels used in other countries.
This situation is perceived as problematic by some
sectors, since the Swedish legislation generally promotes risk decisions based on overall assessments of
different aspects, standpoints and interests. Also most
of the sectors made a distinction between calculated
risk (expectation values) and actions to manage risk,
which is also exemplified by the two most common
types of discussions displayed in Figure 1. Some sectors also interpreted that the legislation is more focused
on risk reducing actions, than on risk as a level or a
condition. The divergent view between the industry
and some of the sectors, on how to use and assess
quantitative risk analysis arises several questions:
Why do the industry calculate individual and societal risk, even though no quantitative criteria exist?
Do the industry focus more on risk levels than on
safety actions?
Is it possible for authorities to evaluate the results
of QRA, even though no quantitative criteria exist?

CONCLUSIONS

In this study we have striven to give a picture of how

the assessment of risk evaluation has been carried out
within the SRSA and to a certain extent also why.

405

The findings in this study show that the evaluations

of risk and safety issues are done in many different
ways, due to several context-dependent factors that
influence the assessment. A study that investigates if it
is possible to find more universal foundations for evaluations of and decisions on risk within the authority is
desired. Such foundations could perhaps supplement
or replace some of the existing principles.
The evaluation component in the authoritys processes that include decisions about risk has to be
elucidated. The benefits of such elucidation could be
among other things better internal and external understanding of the decisions, a more transparent process,
easier and clearer communication of decisions and
standpoints, better support to decision-makers such as
other officials and politicians, and a reduction of the
influence of individual preferences linked to certain
officials.
A more systematic exploration and discussion of
these questions is needed within the SRSA if we are
to develop these issues and get a better general view
and co-ordination of risk evaluation.
6

FUTURE STUDIES

Based on the study, the SRSA has concluded that

improvements are needed and initiated a development study regarding risk evaluation at the authority.
The first ongoing step in the three-year project is to
carry out a review of research done on the subject.
The review will combine four knowledge compilations done by researchers from the fields of sociology,
economics, philosophy and technology.
The results from the ongoing review could together
with the findings from the study presented in this paper
constitute a starting-point for a more substantial foundation for the evaluation of risk and safety issues at
the authority.
ACKNOWLEDGMENTS
We would like to thank the officials participating in the
interviews for the openness and the Swedish Rescue
Service Agency for making it possible to conduct this
study. We also would like to thank Prof. Lars HarmsRingdahl and an anonymous referee for constructive
comments on the conference paper.
REFERENCES
Ale, B.J.M. 2005. Tolerable or acceptable: A comparison
of risk regulation in the United Kingdom and in the
Netherlands. Risk Analysis 25 (2): 231241.

All, R., Harrami, O., Postgrd, U. & Strmgren, M. 2006.

Olyckor, riskanalyser och skerhetsarbetengra olika
perspektiv inom Rddningsverket (in Swedish). Report
P21-480/07. Karlstad: The Swedish Rescue Service
Agency.
Boverket 2005. Riskvrderingdelprojekt 2.1, bilaga till
regeringsuppdrag Personskerhet i tunnlar (in Swedish).
Karlskrona: The National Board of Housing, Building and
Planning.
Commission on Slope Stability 1995. Anvisningar fr
Slntstabilitetsutredningar (in Swedish). Linkping:
Royal Swedish Academy of Engineering Sciences.
Davidsson, G., Lindgren, M. & Mett, L. 1997. Vrdering
av risk (in Swedish). Report P21-182/97. Karlstad: The
Swedish Rescue Service Agency.
Fldeskommittn 1990. Riktlinjer fr bestmning av dimensionerande flden fr dammanlggningar. Slutrapport
frn Fldeskommittn (in Swedish). Statens Vattenfallsverk, Svenska Kraftverksfreningen and Sveriges
Meteorologiska och Hydrologiska Institut.
Folkhlsoinstitutet 2003. Swedens new public health policy. National public health objectives for Sweden. Report:
2003:58. Stockholm: Swedish National Institute of Public
Health.
Harrami, O., Strmgren, M., Postgrd, U. & All, R.
(in press). Accidents, risk analysis and safety
managementdifferent perspective at a Swedish safety
authority. In Proceedings of ESREL 2008 AND 17th
SRA EUROPE CONFERENCEAnnual Risk, Safety and
Reliability Conference, Valencia Spain, 2225 September
2008. Rotterdam: Balkema.
Hultman, M. 2005. Att frst riskeren kunskapsversikt av
olika kunskapsperspektiv. KBM:s Forskningsserie Report
nr. 8. Stockholm: Swedish Emergency Management
Agency.
Lundgren, L.J. & Sundqvist, G. 1996. Varifrn fr
miljvrds-byrkraterna sin kunskap? In Lars J Lundgren
(ed.), Att veta och att graOm kunskap och handling inom miljvrden (in Swedish): 129171. Lund:
Naturvrdsverket Frlag.
Lupton, D. 1999. Risk. Milton Park: Routledge.
Mattsson, B. 2000. Riskhantering vid skydd mot olyckorproblemlsning och beslutsfattande (in Swedish).
Karlstad: Swedish Rescue Services Agency.
Morgan, M.G. & Henrion, M. 1990. UncertaintyA guide
to dealing with uncertainty in quantitative risk and policy
analysis. Cambridge: Cambridge University Press.
Reid, S.G. 1999. Perception and communication of risk, and
the importance of dependability. Structural Safety 21(4):
373384.
Renn, O. 1992. Social theories of risk. In S. Krimsky & G.
Dominic (eds) 1992. Social theories of risk. Westport:
Praeger Publications.
Summerton, J. & Berner, B. (eds) 2003. Constructing risk
and safety in technological practise. London: Routledge.
Svensk Energi, Svenska Kraftnt & SveMin 2007. Riktlinjer fr bestmning av dimensionerande flden fr
dammanlggningar (in Swedish). Stockholm: Swedenergy, Svenska Kraftnt and SveMin.
Welander, G., Svanstrm, L. & Ekman, R. 2004. Safety
Promotion: An Introduction. Stockholm: Karolinska
Institutet.

406

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Regulation of information security and the impact on top management

commitmentA comparative study of the electric power supply sector
and the finance sector
J.M. Hagen
Gjvik University College

E. Albrechtsen
SINTEF Technology and Society/Norwegian University of Science and Technology, Trondheim, Norway

ABSTRACT: The paper compares how information security in critical infrastructures is regulated in two different sectors and how the regulations can influence organizational awareness. It also compares how organizational
information security measures are applied in the sectors, and discusses how the sectors can learn from each
other. The findings document considerable differences in legal framework and supervision practices, in use
of organizational information security measures and the top management engagement. Enterprises belonging
to the finance sector have more widespread use of organizational security measures, and the respondents are
also more satisfied with the management engagement and the organizations performance according to the legal
requirements. The paper argues that information security audit by authorities can be one important contribution
to information security awareness, and top management commitment to security, and that the sectors can learn
from each other by sharing information on how they deal with information security.

INTRODUCTION

Sundt (2006) and Lobree (2002) claim that laws and

regulations affect the implementation of information
security in organizations. There is, however, limited
information publicly available about how compliance
to information security laws is currently dealt with.
This paper addresses how the regulatory and supervisory activities are carried out and impact information
security practices and organizational awareness in the
finance and the electric power supply sector in Norway. It also discusses how each sector could learn from
each other.
Finance and electric power supply are both critical infrastructures; additionally, electric power supply
is the basis infrastructure for all kind of service
production dependent on computers and electronic
communication services.
Norwegian enterprises have gradually increased
their dependence on IT systems (Hagen 2007). In
addition, studies on critical infrastructures and vulnerability of the society have over the last years
documented that also previously redundant physical
systems, like electricity production and distribution,
now are critically dependent on electronic communication and the Internet, and that connecting the

IT systems to the Internet has increased the risk for

system breakdowns and serious failures (Nystuen &
Hagen 2002, Hagen 2003). Norwegian enterprises
are also exposed to computer crime, and there exist
numerous weaknesses in the applied security systems
(Kredittilsynet 2004, Hagen 2007, Hole et al. 2006).
Norwegian authorities, with support from the legal
framework, supervise and measure the compliance
to information security regulations within the critical
infrastructures. By law, authorities can force the organizations to adopt information security measures. Yet,
certain questions arise:
Do the laws and accompanying supervision process
by the authorities have any effects on the enterprises organizational security awareness and top
management commitment?
Are there any differences in security practices
between the sectors?
If there are differences, how could the enterprises
within the different sectors learn from each other to
strengthen information security?
The paper discusses how audit by authorities can contribute to organizational security awareness and top
management commitment to security. Section 1 introduces the problems to be analyzed. Section 2 gives

407

state of the art regarding the managements role in

information security. Section 3 presents the applied
research method. Section 4 provides an overview of
the Electric Power Supply and the Finance Industry
in Norway and the applied supervision methodologies. In section 5 we compare the two sectors studying
the security attitudes and applied organizational security measures. Discussion of the findings is given in
Section 6. Section 7 provide the answers of the research
questions, and section 8, show the way ahead, and concludes that it will be useful to study security cultures
within organizations.
2

THE MANAGEMENTS ROLE

Information security has developed from a strict technological discipline to become a multidisciplinary
responsibility for top management (Lobree 2002,
Sundt 2006). Williams (2007) claims that even the
board needs to be assured of effective risk management and the sharing of responsibility for information security by a number of individuals within the
organization.
Information security law places responsibility for
information security on the management and the
boards. Dealing with authorities is a typical management task. Thus, supervision from authorities may be
one way to raise top managements awareness. We
expect that an engaged top management is important
for how the organization adopts information security measures and complies with the law. Studies of
safety have documented that management involvement is important for the safety work within companies
(Simonds 1973, Simonds & Shafai-Sharai 1977). It
is reasonable to believe that this experience could be
transferred to the information security field.
3

METHOD

We have carried out two exploratory case studies;

one on financial law and supervision, and one on the
energy law and supervision within the hydroelectric
power supply sector. Data on supervision practices was
collected during the summer 2006 from textual materials and personal interviews with representative from
the Norwegian supervisory authorities and four companies. Authorities in Sweden, Finland, Denmark and
UK were also contacted by mail (Hagen et al. 2006).
In the spring 2007, a web based survey was carried out among IT-officers in different sectors (Hagen
et al. unpubl.). This survey addressed organizational
security measures like user education, procedures and
controls and tools, compliance to law and the effectiveness of supervision. Out of a total of 87 answers,
only 34 were related to the electric power supply and

finance sectors. This is a moderate sample, and

limits our ability to conduct statistical analysis to
reveal associations.
Therefore, we have chosen to present the results as
descriptive statistics. A few hypotheses have also been
formulated and tested by independent samples t-tests.
The independent-samples t-test procedure compares
means for two groups of cases. Ideally, for this test, the
subjects should be randomly assigned to two groups,
but this is not the case when we study enterprises
within the energy and the finance sector. Therefore, we
discuss how differences in other factors, ex. respondents, could be masking or enhancing a significant
difference in means. In addition, the paper has been
through several rounds of validation by the informants.

4
4.1

THE ELECTRIC POWER SUPPLY

AND FINANCE INDUSTRY IN NORWAY
The legal framework

In Norway, the Ministry of Oil and Energy holds

the administrative responsibility for the energy sector. The Norwegian Water Resources and Energy
Directorate (NVE) oversees the administration of the
energy and water resources. The Norwegian electricity
industry finds itself in a unique position internationally, as almost 100% of the countrys electricity
production comes from hydro electric power. Norwegian enterprises need to be licensed in order to
supply and dispose electrical energy. Enterprises act
in accordance with many laws and regulations; however, the most important, dealing with contingency
and information security, are the Energy Act, the
Energy Direction and the Contingency Direction 4
and 6. The regulations deal with the planning, building and conduct of different electrical enterprises,
including emergency preparedness. They place certain
requirements on electrical enterprises protection of
critical information systems and sensitive and classified information, and demand that enterprises provide
Assurance Reports.
The Ministry of Finance holds the superior administrative responsibility for the countrys financial sector.
The Financial Supervisory Authority of Norway oversees Norwayss financial markets and regulates the
financial sector. Like the hydroelectric power enterprises, financial enterprises act in accordance with
many laws and regulations. The most important legal
acts are the Financial Supervisory Act, and the related
Information and Communication Technology (ICT)
Direction. The ICT Direction deals with information security. Thus, one major difference between the
sectors is that the financial sector has within the hydroelectric power supply industry is regulated according
to 4 and 6 in the Contingency Direction.

408

4.2

The supervision methodology

The two Norwegian regulatory authorities chosen for

this study conduct their supervisory controls in different manners, still there are some similarities. In
both sectors, supervision is conducted in the following way: The enterprises go through a pre-evaluation,
which aims to evaluate the risks, importance and status of the enterprises. Then, the authority notifies the
enterprises which are selected to undergo supervision.
The Authority requests information, arranges meetings, and collects more information, analyses it and
documents deviations. In this process the dialog is
important. Then, the enterprises are given a deadline
to close the deviations.
Both the Norwegian Water Resources and Energy
Directorate and the Financial Supervisory Authority
have developed guidelines to assist the enterprises
in their effort to comply with information security
law. In addition, they both offer training and give
advices.
But, there are also certain differences. The supervision methodology of the Norwegian Water Resources
and Energy Directorate is based on NS-EN ISO 19011,
which describes an audit methodology and specifies
requirements for auditors. It has developed a handful of yes/no questions based on the text of the
law. The Financial Supervisory Authority has developed a comprehensive questionnaire based on the
ICT Directive and COBIT, and uses this to audit
the IT processes. This is a far more comprehensive tool than the handful of questions posed by the
Norwegian Water Resources and Energy Directorate.
Within the hydroelectric power industry, information
security supervision is just a part of a more comprehensive security supervision, while the finance
sector supervision is characterized by a sole focus on
information security, and is deeply rooted in the ICT
Direction. In addition, more human resources are dedicated to the information security supervisory process
within the financial sector than in the electric power
industry.
When requirements to close deviations are not
met, both authorities can withdraw the licences. Yet,
punishment practices may vary: The Financial Supervisory Authority of Norway publishes any deviations
that are not closed within the deadline. This may
impact the markets trust to the current financial
enterprise and management. The Norwegian Water
Resources and Energy Directorate require the deviation closed, and any enterprise not closing the deviation risks fines. Also, the addressee of the supervision
report varies: Within the financial sector the report is
addressed to the board, placing responsibility on the
board; within the hydro electrical power supply industry, the supervisory report is addressed to the company
(Hagen, Norden and Halvorsen, 2007).

COMPARATIVE ANALYSES

5.1

The answers

Of the 87 received answers, 21 were from hydro electrical enterprises and 13 from saving banks. Table 1
identifies the distribution of the answers and the
respondents of the questionnaire.
The data shows that within the finance sector more
managers answered the questionnaire compared with
the hydro electric power industry in which responses
from IT personnel dominated.
About half of the respondents report that they outsource some or all IT operation. This corresponds with
the findings in other surveys (Hagen 2007).
5.2

Top management commitment and quality

of security work

The questionnaire included questions about the

respondents judgements of top management engagement and the compliance to information security
law.
Table 2 summarises the results of paired sample
t-tests. A Likert scale was used, ranging from 1 =

409

Table 1.

Respondnets (N = 34).
Electric power

Finance

Manager
IT
Economy
Security
Advisor

1
17
0
2
1

4
1
2
3
2

Total counts

21

12

Table 2. Securty attitudes towards information security

(N = 34). Mean ranging from 1 = strongly disagree to
5 = strongly agree.
Mean

Engaged top
management
Info.sec. is frequently
on the agenda
Legal requirements
are satisfied
Supervision increases
the top managements
engagement

Electric
power

Finance

Sig.

3.33

4.62

0.001

2.29

3.08

0.057

4.14

4.69

0.035

3.95

4.25

0.399

strongly disagree to 5 = strongly agree. The results

show mostly high mean values, which means that the
respondents agree to most statements. Tests of differences in attitudes reveal that the respondents within
the finance sector view top management engagement
to be higher compared with the respondents from the
electric power supply sector. Within the finance sector,
information security is also more often on the managements agenda, and there is also a stronger focus
on meeting legal requirements. There are, however, no
significant differences regarding the respondents view
on the effect of supervision, but a high mean value
indicates that the respondents agree that supervision
has some effect.
Table 3 shows that most organizational security
measures are widely adopted in the finance sector.
Security classification of systems and personnel are,
however, mostly applied by electric power supply
organizations. One possible explanation might be the
security regime that developed after the Second World
War with focus on protection of power plants and
critical infrastructures in case of war.
The survey also inquired about the frequency to
which risk analysis and internal and external audits
were conducted. Independent sample t-tests show significant differences between the sectors; the financial
sector performs more frequently risk analysis and
audits compared with the electric power industry.
Summing up the findings so far, it seems reasonable that the information security awareness is higher
in the finance sector compared with the electric power
industry. This is documented by the respondents
judgement of top management engagement and a more
widespread use of organizational information security
measures.
The data also shows that the overall majority of
both the financial and the electrical power supply
organisations have routines for the reporting of incidents, and that the employees in both sectors often act
according to the routines and report security incidents.

Looking at the share of enterprises experiencing security incidents (Table 5), we see that a larger number
of electric power supply enterprises report incidents
typically caused by insiders, compared with financial enterprises. A subsequent hypothesis may be that
there exists a relationship between high organizational
security awareness and low probability for security
breaches by own employees. The data set is unfortunately too small to conduct a Chi-square test of the
hypothesis.
5.3 How could the sectors learn from each other?
Within the Norwegian electricity industry, focus
placed upon information security has increased during the last years, both through national information
security strategy initiatives and research on protection
of critical infrastructure (Nystuen & Hagen, 2003).
The traditional security focus in the sector has been
on national security, physical security and emergency
in case of natural disasters and war. Information security has become important in particular during the last
1015 years, and is now a critical input in process
operation and trade.
In the finance sector, however, information security has been close to core business ever since money
became electronic signals. If the IT systems are
down or the IT services not secure, this would not

Table 3. Percentage who have applied different organizational security measures in electric power sector (N =
21) and financial sector (N = 13).

Written security policy

Guidelines
Non-disclosure agreement
Employee training
Awareness raising activities
Formal disciplinary process
Classification of systems/personnel

Electric
power

Finance

76%
80%
62%
52%
48%
38%
62%

100%
92%
100%
92%
54%
61%
54%

Table 4. Risk analysis and audits (N = 34). Scale used:

1 = never, 2 < every third year, 3 < yearly, 4 = yearly.
Mean

Risk analysis
Internal audit
External audit

Table 5.

Electric
power

Finance

Sig.
level

3.05
3.10
2.62

3.77
3.85
3.46

0.003
0.002
0.016

Number of enterprises with incidents (N = 34).

N
Electric
power (N = 21)

Abuse of IT systems
Illegal distribution of nondisclosure material
Unauthorized deletion/
alteration of data
Unintentional use violating
security

410

Finance
(N = 13)

10

only prohibit money transfers, but also disrupt the

reputation of the financial enterprise.
Compared with the electrical power supply sector,
the respondents within the finance sector report more
far-reaching security measures, higher management
engagement in information security and more satisfaction with compliance to legal requirements. The
financial sector is also supervised according to a dedicated information security direction and by a far more
comprehensive method.
The electrical power supply sector could learn from
the financial sector how to improve organizational
awareness and adopt organizational information security measures. The financial sector could also learn
from the electric power supply sector how to classify information, personnel and systems into different
security classes, and how emergency exercises can
prepare the organizations for different threat scenarios.
To learn and share knowledge about information
security practices and its effects, mutual trust and a
secure meeting places are needed. The survey data
shows that the organizations are participating in a
wide range of different security forums, but we do
not see many of them participate in the same forum.
The authorities could facilitate meetings to establish connection among enterprises in the different
sectors. Another possibility is that the National Security Agency facilitates such meetings through the
NorCerts forums. This could be a good starting point.
We expect that security systems within both sectors
would benefit from information sharing and mutual
learning. The electric power supply is the most critical
infrastructure in modern society. In case of power outage, most services stop, including financial services.
On the other hand, the electric power supply sector
depends on the finance sector to transfer money when
electricity is traded, and the whole society depends
on both electric power supply and services and infrastructures for money transfers. If any of these critical
service stops, then production and delivery of all kind
of services and goods stop, too. Moreover, in the modern society, there is a new evolving interdependency;
the ICT systems that affect the sectors/domains beyond
their boundaries. Both sectors depend critically on the
ability to communicate, and on ICT providers and
infrastructures that might be located in other countries
(Nystuen and Fridheim, 2007).

We know from this study that both electric power

industry and finance sector authorities have developed
guidelines and in addition provide their supervisory
objects with advice and assistance. The differences in
legal framework, amount of human resources spent
in supervision of information security and applied
supervision methodologies mirror the organizational
security awareness in the two critical infrastructures.
High organizational security awareness corresponds
with less exposure to insider threats. The finance sector has adopted the most comprehensive approach,
which is mirrored in wider use of security measures,
a more engaged top management and less reporting
of insider incidents. Our findings confirm the theories that a legal framework affects how information
security measures are adopted by organizations (Sundt
2006, Lobre 2002).
It is reasonable to believe that the differences in
laws and regulations and how compliance to laws
is supervised can have an impact on organizational
awareness, including top management engagement in
information security. When supervising, the authorities place responsibility on the management. The
quality of the supervision methods and the kind of
applied sanctions might impact the engagement of the
top management. If the management is engaged, it
will be aware of the need for information security
measures to comply with the laws, and assure that
security measures are implemented. In this survey we
found that high management engagement corresponds
with a high degree of adopted security measures and a
lower degree of insider incidents. The findings correspond with the experiences from safety management
(Simonds 1973, Simonds & Shafai-Sharai 1977) and
research on information security law (Sundt 2006,
Lobre 2002).
We must, however, be aware of alternative explanations of the findings. Organizational information
security awareness could be related to maturity in
information security work within the sectors, or the
closeness of information security to core business.
Unfortunately, the data set is neither designed for it,
nor large enough, to conduct statistical analysis to test
the hypothesis.

Supervision from a public authority requires a legal

framework. Without laws and directions, there will
be no public supervisory activities. The supervisory
practices and related legal framework in Norway, Sweden, Denmark, Finland, and the UK is documented in
(Hagen et al, 2007). Norway has close relations to all
these countries, both within the electric power supply
sector and within the finance sector. The Norwegian

6.1

DISCUSSIONS
The differences in legal framework
and supervision practice

The question we raised in this paper asks whether the

laws and accompanying supervision process of the
authorities have any effects on the organizational security awareness of the enterprises under supervision.

6.2

411

Information security supervision

in other countries and the possibility
to generalize the findings

authorities use information security supervision to a

larger extent than the other countries. The Norwegian
financial sector is in an exceptional position because
of the ICT Direction. The usual practice is to include
information security requirements as a minor part of
other laws and directions. In Denmark, however, there
are regulations within the financial industry similar
to the ICT Direction in Norway, and the supervisory process is almost equal to the Norwegian one.
Supervision of information security within the hydroelectric power supply is, however, in a premature phase
compared to Norway, because information security is
not to the same extent regarded as critical. Sweden
does not have information security laws for finance
or hydroelectric power supply, but its Security Act
can be applied on critical infrastructures. Finland has
the Data Security Act, which is relevant for all sectors, yet there are differences regarding the practical
organisation of emergency preparedness. The most
significant differences are, however, related to UK
practices. The UK does not have information security regulation similar to the Nordic countries, but
the Centre for Protection of National Infrastructure
(previous National Infrastructure Security Coordination Centre, NISCC) contributes with knowledge and
advice regarding cyber attacks against critical infrastructures. None of the countries including Norway use
indicators or metrics to measure information security
and compliance to law.
The differences in the national regimes make it difficult to generalize the findings in this study and to
conduct meaningful comparisons. It could, however,
be interesting to study the information security practices within other countries in more detail to learn more
about the effects of different supervision approaches
and also the effect of laws related to more market
driven regimes.
6.3

Validity and reliability of the data

The paper is based on both in depth case studies,

as documented by Hagen et al. (unpubl.) and a survey. The report has been through several rounds of
validation by the informants. These findings, describing the differences in laws and supervision practices,
are therefore considered as both valid and reliable.
Turning over to the survey, the major problem is
the low response rate and the few responses. However, the answers distribute well among the sectors,
enabling some discussion of variations among the sectors. Besides, personal interviews with representatives
from the sectors confirm the findings.
Studying the identity of the respondents to the questionnaires, there were significant differences between
the sectors. In the finance sector, mostly management answered, while in the electric power industry,
mostly IT personnel answered. Both IT personnel and

management should be aware of the use of organizational information security measures examined in the
survey, as the measures would influence in their work
in someway. It is our view that the study produces
valid and reliable results, despite the few answers and
the variations in identity of the respondents to the
questionnaire.

CONCLUSIONS

In this paper we raised three questions:

Do the laws and accompanying supervision process
of the authorities have any effects on the organizational security awareness and top management
engagement of the enterprises under supervision?
Are there any differences in security practices
between the sectors?
If there are differences, how could they learn from
each other to strengthen the information security?
Our findings support the theory that laws have effect
on how organizations adopt information security measures. The findings indicate that both laws and the
quality of the supervisory processes could have an
effect on how the organizations adopt organizational
information security measures.
There are differences between the sectors in legal
framework, supervision methods and type of sanctions
if the requirements are not met. These differences mirror the security practises, attitudes of top management
engagement in information security and the exposure
to insider incidents.
Cross-sector learning and knowledge sharing will
strengthen the overall security culture and can be motivated by the mutual dependency between the sectors.
Existing security forums can be used, or the authorities
could facilitate meetings. In this way, security systems within both sectors can benefit from knowledge
sharing.

THE WAY AHEAD

We will continue to study the effects of organizational

information security measures. Further research will
be conducted on the effect of implemented security
policy and IT-user training on security culture.

ACKNOWLEDGEMENT
The authors would like to acknowledge Jan Hovden,
Pl Spilling, Hvard Fridheim, Kjetil Srlie, shild
Johnsen and Hanne Rogan for their contribution to the
work reported in this paper.

412

REFERENCES
COBIT, 2008: Available at www.isaca.org/cobit
Hagen, J.M. 2003. Securing Energy Supply in NorwayVulnerabilities and Measures, Presented at the conference:
NATO-Membership and the Challenges from Vulnerabilities of Modern Societies, The Norwegian Atlantic Committee, and the Lithuanian Atlantic Treaty Association,
Vilnius, 4th5th December, 2003.
Hagen, J.M., Albrechtsen, E. & Hovden, J. Unpubl. Implementation and effectiveness of organizational information
security measures, Information Management & Computer
Security, accepted with revision.
Hagen, J.M., Norden. L.M. & Halvorsen, E.E. 2007.
Tilsynsmetodikk og mling av informasjonssikkerhet i
finans og kraftsektoren. In Norwegian. [Audit tool and
measurement of information security in the finance and
power sector] FFI/Rapport-2007/00880.
Hagen, J.M. 2007. Evaluating applied information security
measures. An analysis of the data from the Norwegian
Computer Crime survey 2006, FFI-report-2007/02558
Hole, K.J., Moen, V. & Tjstheim, T. 2006. Case study:
Online Banking Security, IEEE Privacy and Security,
March/April 2006.
Kredittilsynet (The Financial Supervisory Authority of Norway), Risko- og srbarhetsanalyse (ROS) 2004, In Norwegian [Risk and vulnerability analysis 2004].

Lobree, B.A. 2002. Impact of legislation on Information Security Management, Security Magazine Practices,
November/December 2002: 4148.
NS-EN ISO 19011, Retningslinjer for revisjon av systemer
for kvalitet og/eller miljstyring, [Guidelines for auditing systems for qualitiy management and environmental
management].
Nystuen, K.O. & Fridheim, H, 2007, Sikkerhet og srbarhet i
elektroniske samfunnsinfrastrukturerrefleksjoner rundt
regulering av tiltak, [Secure and vulnerable electronic
infrastructuresreflections about regulations and security measures], FFI-Report 2007/00941.
Nystuen, K.O & Hagen, J.M. 2003. Critical Information Infrastructure Protection in Norway, CIP Workshop,
Informatik, Frankfurt, a.M, 29.09-02.10.03, 2003.
Simonds, R.H. & Shafai-Sharai, Y. 1977. Factors Apparently
Affecting Injury Frequency in Eleven Matched Pairs of
Companies. Journal of Safety Research 9(3): 120127.
Simonds, R.H. 1973. OSHA Compliance Safety is good
business, Personnel, July-August 1973: 3038.
Sundt, C, 2006. Information Security and the Law. Information Security Technical Report, 11(1): 29.
Williams, P. Executive and board roles in information security, Network Security, August 2007: 1114.

413

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The unintended consequences of risk regulation

B.H. MacGillivray, R.E. Alcock & J.S. Busby
Department of Management Science, Lancaster University, UK

ABSTRACT: Countervailing risks have important implications for risk regulation, as they suggest that gains in
environmental and human health may be coming at a significant cost elsewhere, or even that in some situations,
our cures may be doing more harm than good. This paper expands the prevailing explanation of why risk
reducing measures frequently lead to harmful side-effects which render them uneconomic or even perverse.
Our expansion is three-fold, we: highlight how confirmation bias and various pathologies of complex problem
solving may constrain regulators from foreseeing or even considering the harmful side-effects of proposed risk
reduction measures; argue that limited incentives and capacities for regulatory learning constrain the detection
and correction of harmful side-effects of interventions; and contend that the adversarial nature characterising
many risk conflicts systematically gives rise to perverse trade-offs. We conclude that adaptive, stakeholder-based
forms of regulation are best positioned to reform these pathologies of the regulatory state, and to produce a form
of governance best suited for recognising and addressing the need to make risk trade-offs in what are often highly
charged, contested situations.

A BRIEF HISTORY OF UNINTENDED

CONSEQUENCES

In prehistoric societies, mans tasks were largely

ephemeral problems having no significance beyond
themselves. He collected firewood, hunted prey,
forged tools, and sought mates. The low complexity of both the tasks and the social structures within
which they lay meant that they could be pursued
on an ad hoc basis, and only rarely did the need
to view problems as being embedded within other
problems arise (Drner, 1996). Passing through early
human history, man became more dependent on others and aware of his influence on the environment,
and developed rudimentary economic and political
systems to manage these interdependences. Initially,
small-scale sufficiency economies and councils of
elders were sufficient to coordinate the limited division of labour and facilitate an understanding of causal
relationships between human activities and the natural environment (Hofstetter et al., 2002). However,
as specialisation and differentiation grew, technology
and industrialisation advanced, population burgeoned,
and economic and social systems became increasingly
intertwined, complexity has come to characterise the
modern world. We now must deal with a series of
closely, and often subtly, related problems, with the
consequence that our approach to problem solving
increasingly requires attention to interdependencies
of social and natural systems, and an awareness of

the often latent second and third-order effects of decisions and actions (Drner, 1996). For various reasons,
we seem ill-suited to this task.
This incongruity has led some of our most prominent social theorists and philosophers to focus on the
unforeseen and unintended consequences of human
actions, in contexts ranging from economics, to politics, to social policy (e.g. Popper, 1961; Merton,
1936; Hayek, 1973). The driving impetus was the
idea that unintended consequences should be a central concern of the social sciences, as they constrain
the ability to predict and therefore control the consequences of social interventions. In the past generation,
the baton has been passed to scholars of risk regulation,
a small but influential group of whom have focussed on
the harmful, unintended consequences of risk reducing measures in the public and environmental health
arenas (e.g. Graham and Wiener, 1996; Sunstein,
1990; Viscusi, 1996). For example, airbags may protect adults but kill children; gas mileage standards
may protect the environment at the cost of thousands
of lives annually, as they encourage manufacturers
to sacrifice sturdiness for fuel-efficiency; drug-lags
stemming from stringent testing requirements may
protect the public from potentially adverse effects of
un-tested pharmaceuticals, whilst at the same time
diminishing the health of those who urgently need
them; bans on carcinogens in food-additives may lead
consumers to use non-carcinogenic products which
nevertheless carry even greater health risks, and so on.

415

These countervailing risks have important implications for risk regulation, as they suggest that recent
gains in environmental and human health may be coming at a significant cost elsewhere, or even that in some
situations, our cures may be doing more harm than
good. This paper expands the prevailing explanation of
why risk reducing measures frequently lead to harmful
side-effects (e.g. Graham and Wiener, 1995; Wiener,
1998), and argues that stakeholder-based forms of
adaptive management are the most effective regulatory arrangements for making optimal risk trade-offs
in what are often highly charged, contested situations
requiring tragic choices to be made under conditions
of ignorance.

PERVERSE RISK TRADE-OFFS: THE

PREVAILING EXPLANATION

We shall briefly summarise the dominant explanation

of why externalities arise from risk reducing measures, most prominently associated with Professors
Wiener and Graham (see, e.g. Graham and Wiener,
1995; Wiener, 1998). In short, their argument is that
the fundamental reason that risk reduction measures
sometimes create countervailing risks is the interconnectedness of multiple risks, derived in turn from
the interconnectedness of social and environmental
systems. Of course, countervailing risks are not problematic for risk regulation per se; difficulties arise
when, on balance, the trade-off between target and
countervailing risks serves to render a regulation to
be an inefficient allocation of resources, or, more dramatically, perverse in the sense that we are left with a
net diminution of environmental or public health. And
so, their argument goes, we must think systematically
when considering risk reduction measures. Based on
a series of detailed case studies, these scholars show
that that such systematic thinking is far from the norm
in risk regulation, and argue that perverse or uneconomic trade-offs routinely arise due to the systematic
failure of regulatory bodies to account for the full consequences of their interventions in a rational manner.
This failure is explained in terms of psychology and
with reference to some perverse aspects of regulatory
law and politics.
Graham and Wieners psychological angle posits
that the use of heuristics, chiefly availability, to make
probabilistic judgements leads to systemic error in
risk perception and by extension erroneous regulatory
responses to risk. Similarly, cognitive biases, such as
loss aversion, are held to explain conservative tendencies in both risk assessment (e.g. default conservative
assumptions) and risk management (e.g. the precautionary principle; over-regulation of risks from new
sources vs. old sources), which are argued to further

skew the balancing of target and countervailing risks

away from normative ideals. They also draw attention to risk compensation, wherein the reductions in
risk sought by regulatory interventions are partly offset by behavioural changes which may even shift the
nature of the risk or the group who bears its burden.
For example, safer car designs may encourage faster
or more reckless driving, thus leading to a lower than
expected decrease in risk to the driving public, and a
net increase of risk to pedestrians. The scholars argue
that these compensations are frequently unaccounted
for in regulatory decision making.
From the perspective of regulatory law and politics,
Graham and Wiener invoke public choice theory to
explain why regulatory decision-making often departs
the norm of a balanced, deliberative consideration of
the public good towards satisfying the vested interests of the more vocal interest groups, often with the
effect of offloading countervailing harms on those
whose voices were omitted from the regulatory process. Similarly, the scholars hold that jurisdictional
specialisation and fragmentation of the regulatory
state leave little incentive for regulatory agencies to
monitor or even consider the harmful effects of their
regulations on other domains (e.g. across geographical
boundaries, hazard types, and media, etc.).
To summarise Graham and Wieners view, then,
externalities may arise from risk reduction measures
because environmental and public health problems
are not hermetically sealed, and so unintended, undesirable side-effects of interventions often arise. The
trade-offs between countervailing and target risks are
often uneconomic or perverse because of legal, political and psychological factors which prevent regulators
from thinking rationally and synoptically about their
interventions. Broadly speaking, we find much truth
in this explanation. However, we suggest that the
distorting effect of the availability heuristic is likely
overstated, given that subjective judgements of probability are relatively minor components of overall risk
perceptions (see, e.g. Sjberg, 2000). But this is a
minor quibble. Our broader argument is that there are
additional factors at work in distorting risk trade-offs,
which we expand on in the following section.

EXTENDING THE EXPLANATION

3.1 The psychological perspective

In addition to the errors which arise from heuristic
judgements, and other, unrelated, cognitive biases, we
can finger various pathologies of complex problem
solving as being plausible suspects in the generation of perverse risk trade-offs. In doing so, we
draw upon the dynamic decision making research programme (e.g. Drner, 1996; Brehmer, 1992) which

416

explores individual and group decision making in simulated environments defined by the characteristics of
complexity, opaqueness, and dynamism (e.g. natural
resource management). Given these characteristics,
such simulations are largely analogous to the dilemmas facing regulators tasked with protecting public
and environmental health.
Unsurprisingly, this programme has revealed that
people generally struggle to deal with complex problems; however, the errors committed by the participants were far from random, instead being indicative
of general weaknesses or pathologies in reasoning
and perception when dealing with complex, opaque,
dynamic systems. Those of greatest relevance to us are
(Drner, 1996; Brehmer, 1992): the common failure
to anticipate the side-effects and long-term repercussions of decisions taken; the tendency to assume that
an absence of immediate negative effects following
system interventions serves as validation of the action
taken; and the habit of paying little heed to emerging needs and changes in a situation, arising from
over-involvement in subsets of problems. Rather than
appreciating that they were dealing with systems composed of many interrelated elements, participants all
too often viewed their task as dealing with a sequence
of independent problems (Drner, 1996; Brehmer,
1992).
One conclusion that can been drawn from this is
that peoples mental models often fail to incorporate
all aspects of complex tasks, and so making inferences
about side-effects and latent consequences of actions
is often a bridge too far (Brehmer, 1992). Although
we are to some extent stuck with the limitations of our
cognitive capacities, it is worth bearing in mind that
mental models are not solely internal psychological
constructs, but are to an extent socially constructed.
It thus seems reasonable to assume that, in the context of risk regulation, the incorporation of a broad
range of values, interests, and perspectives within the
decision making process would help to counteract the
tendency to frame the pros and cons of proposed risk
reduction measures in an unduly narrow, isolated manner (i.e.expand regulators mental models). Moreover,
enhancing the mechanisms and incentives for regulatory agencies to monitor and evaluate the impacts
of their interventions could ameliorate the tendency
to neglect latent outcomes of decisions (i.e.lengthen
regulators mental models). We return to this later.
We hypothesise that a further contributory factor
to the phenomenon of perverse risk trade-offs is the
well established phenomenon of confirmation bias.
This refers to the tendency to seek out and interpret
new information in a manner which confirms ones
preconceived views and to avoid information and interpretations which questions ones prior convictions.
The reader may, quite rightly, point out that there is a
veritable litany of psychological biases that have been

identified, and then ask why we consider this one to

be particularly central to the problem of perverse risk
trade-offs. The reason is that disputes over proposed
public and environmental health measures are often
highly charged, emotive issues, given that they concern
the protection or sacrifice of the most fundamental values: human life, inter-generational equity, ecological
health, and so forth. In such situations, where people commonly perceive moral absolutes as being at
stake, said people can raise confirmation bias to an art
form. Thus, those drawing attention to the potentially
harmful consequences of proposed risk reduction measures may be viewed as mere reactionaries or industry
stooges, and the information or data which they use to
support their position treated with scepticism or simply ignored. In essence, we propose that confirmation
bias leads to the systematic discounting or even neglect
of the likelihood and extent of potentially harmful side-effects of proposed risk reduction measures,
a clear formula for the generation of perverse risk
trade-offs.

3.2

The perspective of regulatory law and politics

Much ink has been spent highlighting the legal and

political constraints to optimising those trade-offs
inherent to risk reduction measures (e.g. the use of
absolute standards; the narrow and oft conflicting
mandates and missions of regulatory agencies; laws
forbidding the use of cost-benefit and health-health
analysis in evaluating certain proposed interventions;
the traditional legal framing of environmental conflicts as short-term, zero sum questions of authority,
jurisdiction, prohibition and entitlement, etc. See, e.g.
Viscusi, 1996; Sunstein, 1990; Graham and Wiener,
1996). Suffice it to say that we question the wisdom
of these tendencies towards absolutism and the fragmentation of the regulatory state, problems which we
address in our proposals for reform.
However, the debate on the unintended consequences of risk regulation has focussed on how such
legal and political factors constrain regulatory agencies from accounting for the potentially harmful sideeffects of their interventions before they are enacted;
little thought has been given to the detection and correction of these side-effects post implementation. This
is curious, as the limited incentives and capacities of
regulatory agencies to collect feedback and use this
for error correction have been bemoaned by numerous scholars (e.g. Dryzek, 1987a; Weale, 1992). These
limitations seem particularly problematic when one
considers the radical uncertainty characterising much
of the scientific and technical information underpinning regulatory decisions, as well as the inherent
dynamism and stochastic nature of social and natural
environments. And so proposals for regulatory reform

417

that seek to address the problem of perverse risk tradeoffs should look with one eye towards error prevention,
whilst casting the other towards error detection and
correction.
3.3

The sociological perspective

We now turn to consider what a sociological perspective can tell us about the phenomenon of perverse
risk trade-offs. Our central point is that the polarised
nature of debates over proposed regulations, both in
the public arena and within the lobbying process,
is a key precondition for the generation of perverse
trade-offs. For simplicity, consider the case where a
regulatory measure for restricting the use of brominated flame retardants is proposed. Paradigmatically,
we would see various groups lobbying the architects of
the regulatory process (i.e.administrators, bureaucrats,
legislators): NGOs, industry groups, think tanks, and
so forth. Those favouring the proposed measure would
of course highlight the forecast gains in environmental
and human health arising from a ban, whilst those in
opposition will draw attention to the economic costs
and potential countervailing risks which may it may
give rise to (e.g.through reducing the level of protection from fires, through promoting a shift to potentially
hazardous chemical substitutes about which little are
known, etc.).
At a fundamental level, these contrasting policy
preferences arise from what are often sharply contrasting beliefs, values, norms and interests of the
disputants. Of course, that people disagree is hardly
revelatory. The problem which arises is that the current
regulatory state systematically encourages adversarial
relations between these groups, through, for example: the legal framing of many environmental conflicts as zero-sum questions of prohibition, authority
and jurisdiction (Freeman and Farber, 2005); often
relying on lobbying as a proxy for stakeholder consultations (when the latter occur, they tend to be
infrequent and superficial); and in their occasional
resort to the court for final resolution of disputes.
This promotion of competitive rather than co-operative
behaviour encourages disputants to exaggerate differences between one another, to ascribe malign
intentions to the positions of others, and to simplify
conflicts through the formation of crude stereotypes
(Fine, 2006; Yaffee, 1997). Thus, in many situations,
disputes over risk tradeoffs can resemble prisoners
dilemmas, where co-operation could lead to a mutually acceptable solution which balances harm against
harm, but a fundamental lack of trust leaves the participants caught in a zero-sum struggle as they fear that
any compromise would not be reciprocated.
Moreover, in such adversarial settings the underlying scientific, technical and economic data on which
risk trade-offs are ostensibly based is often (mis)used

by the various parties to entrench existing positions

and to discredit opponents. This is particularly problematic given the objective dearth of our knowledge of
many risks to public and environmental health, and the
value-laden assumptions underlying any comparison
of risk against risk (e.g. what price to put on human life,
how to deal with issues such as equity, how to value
ecological status, etc.). This relative indeterminacy
leaves broad scope for the various disputants to interpret risk analysis outcomes in radically different ways,
and when they have access to power or resources, to
shape the very outcomes themselves through adopting
different assumptions and data gathering procedures,
often within the framework of a defensible scientific
methodology. And so even where regulatory agencies
are able to formulate a policy which ostensibly balances between the diametrically opposed positions of
those disputants (i.e.is able to escape the zero sum
trap), it is often based on highly politicised and valueladen data, meaning that the balancing of harm against
harm is illusory.

A BRIEF PROPOSAL FOR REFORMING

THE REGULATORY STATE

In this final section, we briefly argue that adaptive,

stakeholder-based forms of regulation are best positioned to reform these pathologies of the regulatory
state, and to produce a form of governance best suited
for recognising and addressing the need to make risk
trade-offs in what are often highly charged, contested
situations. Our discussion is at a general, abstract level,
leaving any debate over the finer points of such reforms
to scholars of administrative law and regulatory policy.
Stakeholder-based governance refers to an array of
practices where a broad cross-section of stakeholders,
selected to represent different interests, come together,
in person, for long-term dialogue to address policy
issues of common concern, overseen by a neutral
party which initiates, lubricates and oversees discussions, ensuring that they are governed by rules of
reasoned discourse (e.g.ruling out threat, concealment of information, etc.), and in which decisions
are made by consensus rather than diktat or majority rule (e.g.Dryzek, 1987b; Innes and Booher, 1999;
McDaniels et al., 1999). A growing body of research
suggests that stakeholder-based approaches help build
trust and render participants less hostile to the views of
others, providing the grounds for mutual understandings of stakeholder assumptions, interests, values,
norms and perspectives (e.g. Dryzek, 1987b; Innes and
Booher, 1999; McDaniels et al., 1999). In short, they
create an environment which enables participants to
find solutions which accommodate each others interests without harming their own, or learn to view all

418

interests as interconnected and thus conceive of disputes as joint problems in which each has a stake
(Innes and Booher, 1999). Here, the malign influences of narrow mental models, of confirmation bias,
of absolutism, of adversarial relationships, and of the
omitted voice (all, in some way, related to absolutism),
may be expected to be in large part ameliorated.
Of course, this is not a full-proof solution, and
harmful, unintended consequences will still arise from
regulatory measures derived from even the most holistic process, in part due to the profound uncertainties
characterising many public and environmental health
risk dilemmas, and the temporal nature of social values, interests, and perspectives. It is not uncommon
for public, governmental or scientific perceptions of
the rationality of past risk trade-offs to migrate over
time, leading to calls for corrective measures or even
a sharp reversal of the path already taken, such that
as observers of the decision process we are left with
a sense of dj vu, and a feeling that the governance
of risk resembles a Sisyphean challenge. Thus, it is
crucial that the regulatory state adopt a more adaptive
approach (e.g. McDaniels et al., 1999), in the sense
of viewing decision making iteratively, of placing a
strong emphasis on the role of feedback to verify the
efficacy and efficiency of the policies enacted, and of
appreciating the wisdom of learning from successive
choices.
5

CONCLUSIONS

To conclude, we have argued that stakeholderbased, adaptive approaches to regulatory governance

should reduce the incidence of perverse risk tradeoffs through a) integrating a broader range of values, perspectives, interests, and scientific and technical information into regulatory decision making;
b) transforming what were previously framed as zerosum disputes into co-operative searches for mutually
acceptable solutions; and c) promoting the early detection and correction of undesirable side-effects of regulatory measures through providing mechanisms and
incentives for learning.

REFERENCES
Brehmer, B. 1992. Dynamic decision making: human control
of complex systems. Acta Psychologica. 81(3):211241.
Drner, D. 1996. Recognizing and avoiding error in complex
situations. New York: Metropolitan Books.
Dryzek, J. 1987a. Ecological rationality. Oxford: Blackwell.
Dryzek, J. 1987b. Complexity and rationality in public life.
Political Studies. 35(3):424442.
Fine, G.A. 2006. The chaining of social problems: solutions
and unintended consequences in the age of betrayal. Social
Problems. 53(1):317.
Freeman, J. and Farmer, D.A. 2005. Modular Environmental
Regulation. Duke Law Journal. 54:795.
Graham, J.D. and Wiener, J.B. 1995. Risk versus risk: tradeoffs in protecting health and the environment. Harvard
University Press.
Hayek, F.A. 1973. Law, legislation and liberty. London:
Routledge and Kegan Hall.
Hoffstetter, P., Bare, J.C., Hammitt, J.K., Murphy, P.A. and
Rice, G.E. 2002. Tools for comparative analysis of alternatives: competing or complimentary perspectives? Risk
Analysis. 22(5): 833851.
Innes, J.E. and Booher, D.E. 1999. Consensus building and
complex adaptive systems: a framework for evaluating
collaborative planning. Journal of the American Planning
Association. 65(4):412423.
Merton, R.K. 1936. The unanticipated consequences of
purposive social action. American Sociological Review.
1(6):894904.
McDaniels, T.L., Gregory, R.S. and Fields, D. 1999. Democratizing risk management: successful public involvement
in local water management decisions. Risk Analysis.
19(3): 497510.
Popper, K.R. 1961. The poverty of historicism. London:
Routledge and Kegan Hall.
Sjberg, L. 2000. Factors in risk perception. Risk Analysis.
20(1):112.
Sunstein, C.R. 1990. Paradoxes of the regulatory state.
University of Chicago Law Review. 57(2): 407441.
Viscusi, W.K. 1996. Regulating the regulators. University of
Chicago Law Review. 63(4):14231461.
Weale, A. 1992. The new politics of pollution. Manchester:
University of Manchester Press.
Wiener, J.B. 1998. Managing the iatrogenic risks of risk
management. Risk: health, safety and environment. 9(1):
3982.
Yaffee, S.L. 1997. Why environmental policy nightmares
recur. Conservation biology. 11(2): 328337.

ACKNOWLEDGEMENTS
The authors would like to thank the Leverhulme Trust
for funding this research.

419

Maintenance modelling and optimisation

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A hybrid age-based maintenance policy for heterogeneous items

P.A. Scarf
Centre for OR and Applied Statistics, University of Salford, UK

C.A.V. Cavalcante
Federal University of Pernambuco, Brazil

R.W. Dwight
Faculty of Engineering, University of Wollongong, Australia

P. Gordon
Faculty of Engineering, University of Wollongong, Australia

ABSTRACT: This paper considers a hybrid maintenance policy for items from a heterogeneous population.
This class of items consists of several sub-populations that possess different failure modes. There are a substantial
number of papers that deal with appropriate mixed failure distributions for such a population. However, suitable
maintenance policies for these types of items are limited. By supposing that items may be in a defective but
operating state, we consider a policy that is a hybrid of inspection and replacement policies. There are similarities
in this approach with the concept of burn-in maintenance. The policy is investigated in the context of traction
motor bearing failures.

INTRODUCTION

The occurrence of a mixture in a number of situations

is discussed by Jiang and Jardine (2007). For example, a mixture arises in a population of heterogeneous
items where some items have a short characteristic
life giving rise to early failure and other items have
a long characteristic life and are the subject of agerelated wear-out. When the time to failure of an item
has a mixture distribution, the failure behaviour is not
immediately clear, since a mixture can give rise to
different types of hazard rate functions and so
different failure behaviour (Jiang and Murthy, 1998).
In this sense, it is not straightforward to determine
a suitable maintenance policy to reduce the earlyphase failures and at the same time take account of
failures during a wear-out phase. Thus, standard maintenance policies such as age-based replacement or
regular inspection may not be the most appropriate
policies for such items.
Inspection policies are concerned with determining the state of the item inspected and then maintaining accordingly. Many articles deal with this
problem (Valdez-Florez and Feldman, 1989). The
standard inspection policy checks an operating unit at
successive times Tk (k = 1, 2, . . .). The most common assumption is that any failure is detected at the

next inspection or check and is corrected immediately

(Nakagawa, 2005). A derivative approach is based on
delay time analysis. The delay time is the time lapse
from when a system defect could first be noticed until
the time when its repair can no longer be delayed
because unacceptable consequences such as a serious catastrophe might arise due to failure (Christer,
1999). The delay time is essentially the time between
defect arrival and failure due to this defect. In delaytime inspection models, the focus is both on defects
and on failure. The majority of the inspection policy
approaches only take account of the periodicity of state
verification over the life of the item or some alternative
planning horizon.
In age-based replacement, an item is replaced at
failure and also preventively replaced when the item
reaches age T . A block replacement policy replaces
an item on failure and at times k, k = 1, 2, . . .
regardless of the age of the item (Barlow and Proschan,
1965). The presumption here is that an item is located
in a socket (in a system) and together the item and
socket perform some operational function within the
system (Ascher and Feingold, 1984). These models are
appropriate for items that possess an increasing failure
(hazard) rate function.
Items that are heterogeneous, arising from a
mixed population, require a different approach to

423

maintenance. Finkelstein and Esaulova (2001b) argue

that periodic policies do not take account of systematic changes that occur in the pattern of ageing of items
from a mixed population. In this paper, we propose a
maintenance policy that is a hybrid of inspection maintenance and age-based replacement. The inspection
phase of the policy deals with early failures. The agebased replacement phase deals with normal wear-out.
We will consider items for which failure implies
immediate cost-consequences. In this context, for
inspection to be viable, we suppose that a defect
may arise prior to failure, and that these defects are
detectable at inspection. Before developing this model
in section 4, we discuss mixtures of failure distributions in general. Section 5 present a numerical example
based on traction motor bearing failure.
Note that the model we propose is related to burnin maintenance. In this policy, items are subjected to
a burn-in test. According to Cha et al., (2004), burnin is a widely used method to improve the quality of
products or systems after they have been produced.
The implication is that the population of items produced are heterogeneous and poor quality items (with
short operational lives) will be screened during burnin. This early life screening will be analogous to the
inspection phase of the hybrid maintenance model that
we propose. Burn-in maintenance modelling is concerned with determining, in combination, optima for
the burn-in time and the subsequent preventive maintenance policy for the operational phase (Drapella and
Kosznik, 2002; Jiang and Jardine, 2007). Our model
is different in that both inspection and preventive
replacement will be carried out during the operational
phase of the item.
2

MIXED FAILURE DISTRIBUTIONS

where
w(t) = p(1 F1 (t))/(1 FMixed (t)).
The lifetime distribution for a mixture of items
from two sub-populations is illustrated in Figure 1.
Mixtures of this kind do not necessarily have an
increasing failure (hazard) rate function. Some examples that support the last conclusion have been given
by several authors (Glaser, 1980; Gupta and Gupta,
1996; Finkelstein and Esaulova, 2001a; Block et al.,
2003). Jiang and Murthy (1998) discuss mixtures
involving two Weibull distributions. Eight different
behaviours for the failure (hazard) rate function for
the mixed distribution are evident, depending on the
parameter values of the underlying Weibull distributions. For the general case, there are five parameters;
the shape of the failure (hazard) rate is dependent
on the values of the two shape parameters, 1 and
2 , the ratio 2 /1 of the two scale parameters (and
not their individual values) and, finally, the mixing
parameter p.
Other authors have explored mixed failure distributions through their mean residual life function
(Abraham and Nair, 2000; Finkelstein, 2002).
The fitting of Weibull distributions to failure data
requires care. It is often the case that the data possess
underlying structure that is not immediately apparent
due to, for example, inspections, left and right censoring, or heterogeneity. It would be unfortunate to
fit a two-parameter Weibull distribution to failures
that arise from a mixture, and then adopt an agebased for the items based on the fitted two-parameter
Weibull since the implied critical replacement age
would be inappropriate for both sub-populations of
items (Murthy and Maxwell, 1981). A full discussion

FMixed = pF1 + (1 p)F2 ,

probability density

According to Barlow and Proschan (1965), mixtures of

distributions arise naturally in a number of reliability
situations. For example, suppose two distinct assembly
lines produce units with distributions F1 and F2 , F1 =
F2 . After production, units from both assembly lines
flow into a common shipping room, so that outgoing
lots consist of a random mixture of the output of the
two lines. With this illustrative example, the authors
define the mixed distribution as:
(1)

where p and 1p are the proportions of items from line

1 and 2 respectively. From equation (1), the probability
density function and failure (hazard) rate function are
(Jiang and Murthy, 1998):
fMixed (t) = pf1 (t) + (1 p)f2 (t),
hMixed (t) = w(t)h1 (t) + [1 w(t)]h2 (t),

age/years

Figure 1. Mixed distribution (____ ). Underlying Weibull

distributions, We (1 = 3, 1 = 2.5) (... . ) and We
(2 = 18, 2 = 5) (-----), mixing parameter p = 0.1.

424

of the fitting of Weibull distributions to data is given

in Jiang and Murthy (1995). In particular, they discuss
the appearance of Weibull plots when the underlying
distribution is a three parameter or a mixed Weibull
distribution.
3

MAINTENANCE MODELS WITH MIXED

DISTRIBUTIONS

As indicated above, there have been many contributions on the theme of mixed distributions. However,
there has been less discussion of maintenance policies
for items with mixed failure distributions. Here we
review the main contributions in this area.
Murthy and Maxell (1981) proposed two types
of age-replacement policy for a mixture of items
from two sub-populations, 1 and 2. In policy I, it is
assumed that the lifetime distribution of items for each
sub-population is known, but it is not known if
an operating item is of type 1 or 2. In the policy II, it is assumed that the decision maker can,
by some test that costs per item, determine if
an item is of type 1 or 2, and then subsequently
replace items from sub-population 1 at age T1 or
at failure, and replace items from sub-population
2 at age T2 or at failure, which ever occurs
first. Finkelstein (2004) proposed a minimal repair
model generalized to the case when the lifetime
distribution function is a continuous or a discrete
mixture of distributions. As was stated in the introduction, some adaptations of burn-in models have
being proposed by Drapella and Kosznik (2002) and
Jiang and Jardine (2007). The objective of these
models is to find optimum solutions for a combined
burn-in-replacement policy, in order to take into consideration the change in ageing behaviour of items
from a heterogeneous population, represented by a
mixed distribution. The cost advantage of the combined policy over the separate policies of burn-in and
replacement is quite small. The combined policy is
also much more complex, and therefore it is difficult
to decide if the combined policy is superior (Drapella
and Kosznik, 2002). Jiang and Jardine (2007) argue
that preventive replacement is more effective in the
combined policy.
The burn-in process is not always suitable since
a long-burn in period may be impractical. Then,
the early operational failure of short-lived items
may become a possibility. For this reason, in the
next section, we propose a model that is similar to a combined burn-in-replacement policy, but
instead of a burn-in phase we propose a phase of
inspection to detect the weak items. This can then
accommodate, during operation, the evolution of
the ageing behaviour of items from a heterogeneous
population.

THE MODEL

We use the following notation. For an item,

X is the age at defect arrival with probability density
function fX ;
H is the delay time from defect arrival to failure
with probability density function fH , independent
of X ;
is the interval between inspections;
K is the number of inspections that will take place
during the inspection phase which has length K;
Y is the age at failure so that Y = X + H ;
T is the age at preventive replacement (K).
We assume that:
fX (x) = pf1 (x) + (1 p)f2 (x) where pis the mixing
parameter and f1 (x) and f2 (x) follow Weibull distributions with characteristic lives 1 , 2 and shape
parameters 1 , 2 .
Inspections are perfect in that defects present will
be identified;
Defective items are replaced at inspection instantaneously and the average cost of replacement of a
defective item is CR ;
Failed items are immediately apparent, cause operational failure of the system, and are replaced
instantaneously with average cost CF ;
At the critical replacement age, T , preventive
replacement of an item is instantaneous and again
costs CR < CF ;
The cost of inspection is CI < CR ;
On replacement of items (whatever the state of the
replaced item), the system is restored to the as-new
state.
In this way, we assume that the population of items
consists of two sub-populations, one weak with a
shorter lifetime and one strong with a longer lifetime,
and failures are anticipated by a defective state. In
order to reduce early failures, all items are inspected
with frequency 1/ during the inspection phase up to
age K. An item will be replaced at the time of the ith
inspection if it is defective; it will be replaced on failure; it will be replaced preventively at T if it survives
to this age. The objective of inspection is to prevent
early-life failures of weaker items. Inspections act as
a natural, operational burn-in process, since the weak
items will fail much earlier than strong items. The
objective of preventive replacement, which takes place
over a much longer time-scale, is to reduce wear-out
failures in later life.
The failure model has three states: good, defective,
failed. In the good and defective states, the system is
operational. The notion of a defective state allows us to
model inspections: if the delay-time is zero (two-state
failure model), then, given our assumption that item

425

failures lead to immediate operational failure, inspection is futile. Note that the model might to be extended
to consider a mixed population of delay-times, a
proportion of which are zero (Christer and Wang,
1995). This effectively relaxes the perfect inspection
assumption because it implies that a proportion of failures cannot be prevented by inspection. We do not
consider this model here however.
The decision variables in the model are K, T and
. KandT are age-related, so that on replacement, the
inspection phase begins again. Thus, the maintenance
model is analogous to age-based replacement. The
as-new replacement assumption implies that we can
use the renewal-reward theorem and hence the longrun cost per unit time as objective function.
Within this framework, the length of a renewal
cycle (time between renewals), V , can take different
values, and

T x

+
(x + h)fH (h)fX (x)dhdx
K 0

+ T

(1FH (T x))fX (x)dx+

fX (x)dx .

In the same way, developing equation (3) we have

that when K > 0 the expected cost per cycle is
E(U ) =

K

(iCI + CR )
i=1

(1 FH (i x))fX (x)dx

(i1)

V = i

(2a)
+

if (i 1) < X < i X + H > i, (i =

1, . . . , K). Thus, for example, V = , and the item is
replaced at first inspection, if X < X + H > .
Also,
V = Y,

(i 1) < Y < i,

V =T

(i1)

T
+ (KCI +CF )

(FH (T x))fX (x)dx

(i = 1, . . . , K);

K < Y < T ,

if X > K X + H < T ;

i=1

(2b)

if X > (i 1)X + H < i,

V = Y,

i
K

[(i1)CI +CF ]
FH (ix)fX (x)dx

+ (KCI + CR )

(2c)

T
(1 FH (T x))fX (x)dx
K

and

if X > K X + H > T .

(2d)

fX (x)dx .

The cost incurred per cycle, U , is given by

iCI + CR

(i 1)CI + CF

U=

KCI + CF
KCI + CF

These expressions simplify when K = 0 to

if V = i, i = 1, . . . , K,
if (i 1) < V < i,
i = 1, . . . , K,
if K < V < T ,
if V = T .

E(V ) =

Developing equations (2a-d), we have that when

K > 0 the expected renewal cycle length is

E(V ) =

K

i=1

and

(1 FH (i x))fX (x)dx

+T (1FH (T x))fX (x)dx+ fX (x)dx ,

(3)

T x

(x + h)fH (h)fX (x)dhdx

T
(FH (T x))fX (x)dx

E(U ) = CF

(i1)

i
K


ix

i=1 (i1)

+CR (1FH (T x))fX (x)dx+ fX (x)dx.

(x + h)fH (h)fX (x)dhdx

426

Expressions for E(U ) and E(V ) in the case K =

0 could also be derived by noting that the hybrid
policy reduces to age-based replacement with critical
T
age T . Then we have that E(V ) = 0 (1 FY (y))dy
and E(U ) = CF FY (y) + C
 Ry (1 FY (y)), where
Y = X + H and so FY (y) = 0 FH (y x)fX (x)dx.
Using the above expressions for E(U ) and E(V ),
the optimal hybrid policy of inspection up to age K
and replacement at age T for items from a heterogeneous population can be determined by minimizing
the long-run cost per unit time:
C (T , , K) =

loglog{ 1/(1-i/N))}

Frequency

4
2

Scarf et al. (2005) consider the lifetimes of bearings

in 375V d.c. traction motors used by a commuter railway, and investigated preventive replacement policies
for these bearings. The railway company uses 2296
traction motors (typically 32 per train), and over a
period of study observed 39 bearing failures. These
are shown in a Weibull plot (Fig. 2).

-2

-4

-6

-8
10
time to failure/years

NUMERICAL EXAMPLE

E(U )
E(V )

with respect to K, and T .

It is interesting to note that hybrid inspection and
replacement policies that have been developed to date
(e.g. Wang and Christer, 2003) are based on the notion
of increasing defect arrival rate and minimal repair.
For these policies, inspections will tend to be carried
out with increasing frequency as the item reaches the
critical age for replacement. This is in direct contrast
to the policy developed in this paper.

10

20 30

Figure 2. Weibull plot for traction motor failures. For a

single population of items, this implies 2.5 and 30
years.

Figure 3.
motors.

3
4
5
6
time to failure/ years

Histogram of bearing failure times for 39 traction

A histogram of bearing failure times is shown in

Figure 3. It is plausible that these failures are early life
failures of items (here bearings) that arise from a mixed
population. Furthermore, inspections and replacement
may account for the reduced number of failures at ages
1, 2, 3, and 4, so that the mixing parameter is somewhat more than 39/2296. Plausible values may be in the
range be p = 0.04 to p = 0.10. We will assume a mixture with two sub-populations of bearings, the weak
items with a Weibull distribution of time to defect (time
in good state) with characteristic life 1 = 3 (years)
and shape parameter 1 = 2.5, and the strong items
with a Weibull distribution of time to defect (time in
good state) with characteristic life 2 = 18 and shape
parameter 2 = 5. Note, the parameters of the distribution of defect arrival time for the strong items
is not based on the data here. As motors were preventively replaced at 7 years, we would expect to see
only very few failures of long-lived (strong) items. An
exponential distribution was arbitrarily chosen for the
delay times, with mean =1/2 (year). As the time to
failure is the sum of the time to defective state and
the delay time, the actual failure time distribution will
have a characteristic life of approximately 3.5 years.
The variance of time to failure will of course be larger
than that of the underlying time to defect distribution.
A hybrid maintenance policy for such a heterogeneous population of bearings is now investigated. For
the purpose of this investigation, we take the preventive replacement cost as our unit of cost (CR = 1), and
consider a range of inspection and failure replacement
costs.
The results are shown in Table 1. As can be observed
here, when the sub-populations are well separated
(large 2 or small 1 ), a hybrid policy with K <
T is optimal. Two phases are then apparentthe
early inspection phase, and the later non-inspection

427

Table 1. Optimum hybrid policy for various values of cost parameters and failure model parameters. Long run cost per unit
. Unit cost equal to the cost of preventive replacement (C = 1). Time unit here taken to be one
time of optimum policy is C
R
year although this is arbitrary.
1

CI

CF

2.5
2
3
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5

3
3
3
2
4
3
3
3
3
3
3
3
3
3
3
3
3
3

5
5
5
5
5
4
3.5
3
5
5
5
5
5
5
5
5
5
5

18
18
18
18
18
18
18
18
12
24
18
18
18
18
18
18
18
18

0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.05
0.15
0.1
0.1
0.1
0.1

0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.25
1
0.5
0.5
0.5
0.5
0.5
0.5

0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.025
0.075
0.05
0.05

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
2
20

0.224
0.225
0.222
0.216
0.228
0.253
0.269
0.288
0.313
0.171
0.243
0.198
0.185
0.254
0.202
0.233
0.103
0.310

10.25
10.25
10.25
10.25
10.81
10.45
10.23
10.21
7.51
13.67
10.29
11.44
9.86
10.64
11.25
10.40
14.56
9.33

wear-out phase. As 2 becomes smaller and the subpopulations are less distinct, then K T , and it is
optimum to inspect over the entire life. When the cost
of inspection is varied, the optimum policy behaves as
expectedlower inspection costs lead to more inspections. Also, a longer mean delay time leads to more
inspections and vice versa, implying that inspections
are only effective if there is sufficient delay between
defect arrival and consequent failure. The effect of
varying the mixing parameter can also be observed in
this table.

CONCLUSION

0.85
0.87
0.83
0.60
1.44
0.99
0.88
0.75
0.78
0.84
2.74
1.33
2.96
0.64
0.68
1.75
0
0.48

K
5
5
5
5
7
10
11
13
9
5
1
8
1
7
16
2
0
10

of failure is large or the proportion of weak items in the

mixed population is large, regular inspection in early
life is recommended.
The similarity of the model to combined burn-inreplacement policies is discussed. The hybrid maintenance policy can be generalized and extended. Hybrid
inspection and block replacement policies may be
developed in a similar manner although the calculation of the long-run cost per unit time will be more
difficult. Extensions to repairable systems could also
be considered.

ACKNOWLEDGEMENTS

In this paper, a hybrid maintenance policy for items

from a heterogeneous population is proposed. A limited number of maintenance models for these kinds
of items have been developed to date. In particular
we consider items that arise from a mixture of 2 subpopulations. The first sub-population represents weak
or low quality items (or possibly poor installation of
items), the second stronger, more long-lived items.
The concepts of delay-time in inspection maintenance
and age-based replacement in preventive maintenance
are combined in a hybrid policy that mitigates early
failures of weak items and extends the age at preventive replacement of strong items. The behaviour of the
policy is investigated for various values of the parameters of the underlying mixture and costs of inspection,
preventive replacement, and failure replacement. This
behaviour is as might be anticipatedwhere the cost

The work presented here was carried out during the

visit of the second author to the University of Salford.
This visit was supported by CAPES (the Brazilian National Mobility Programme), grant number:
1045/07-5.

REFERENCES
Abraham, B. & Nair, N.U. 2000. On characterizing mixtures of some life distributions, Statistical Papers 42(3),
387393.
Ascher, H. & Feingold, H. 1984. Repairable Systems Reliability. New York: Marcel Dekker.
Barlow, R.E. & Proschan, F. 1965. Mathematical Theory of
Reliability. NewYork: Wiley.

428

Block, H.W., Savits, T.H. & Wondmagegnehu, E.T. 2003.

Mixtures of distributions with increasing linear failure
rate. Journal of Applied Probability 40(2), 485-504.
Cha, J.H., Lee, S. & Mi, J. 2004. Bounding the optimal
burn-in time for a system with two types of failure, Naval
Research Logistics. 51(8), 10901101.
Christer, A.H. 1999. Developments in delay time analysis for
modeling plant maintenance. Journal of the Operational
Research Society, 50(1), 11201137.
Christer, A.H. & Wang, W. 1995. A delay-time based maintenance model for a multi-component system. IMA Journal
of Management Mathematics 6(2), 205222.
Drapella, A. & Kosznik, S. 2002. A short communication
combining preventive replacement and burn-in procedures. Quality and Reliability Engineering International
18(5), 423427.
Finkelstein, M.S. 2002. On the shape of the mean residual
lifetime function. Applied Stochastic Models in Business
and Industry 18(2), 135146
Finkelstein, M.S. 2004. Minimal repair in heterogeneous populations. Journal of Applied Probability 41(1),
281286.
Finkelstein, M.S. & Esaulova, V. 2001a. On an inverse problem in mixture failure rates modeling Applied Stochastic
Models in Business and Industry 17(2), 221229. yes
Finkelstein, M.S. & Esaulova, V. 2001b. Why the mixture
failure rate decreases. Reliability Engineering and System
Safety 71(2), 173177.
Glaser, R.E. 1980. Bathtub and related failure rate characterizations. Journal of the American Statistical Association,.
75(371), 667672.

Gupta, P.L. & Gupta, R.C. 1996. Ageing characteristics of

the Weibull mixtures. Probability in the Engineering and
Informational Sciences 10(4), 591600.
Jiang, R. & Murthy, D.N.P. 1995. Modeling failure-data by
mixture of 2 Weibull distributions: a graphical approach,
IEEE Transactions on Reliability, 44(3), 477488.
Jiang, R. & Murthy, D.N.P. 1998. Mixtures of Weibull
distributionsparametric characterization of failure rate
function. Applied Stochastic Models and Data Analysis
14(1), 4765
Jiang, R. & Jardine, A.K.S. 2007. An optimal burn-in
preventive-replacement model associated with a mixture distribution. Quality and Reliability Engineering
International 23(1), 8393
Murthy, D.N.P. & Maxwell, M.R. 1981. Optimal age replacement policies for items from a mixture. IEEE Transactions
on Reliability 30(2):169170.
Nakagawa, T. 2005. Maintenance Theory of Reliability.
Springer, London
Scarf, P.A., Dwight, R. & Al-Musrati, A. 2005. On reliability
criteria and the implied cost of failure for a maintained
component. Reliability Engineering and System Safety
89(2), 199207.
Valdez-Florez, C. & Feldman, R.M. 1989. A survey of
preventive maintenance models for stochastically deteriorationg single-unit systems. Naval Logistics Quarterly
36(4), 419446.
Wang, W. & Christer, A.H. 2003. Solution algorithms for
a nonhomogeneous multi-component inspection model.
Computers & Operations Research 30(1), 1934.

429

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A stochastic process model for computing the cost of a condition-based

maintenance plan
J.A.M. van der Weide
Department of Applied Mathematics, Delft University of Technology, Delft, The Netherlands

M.D. Pandey
Department of Civil Engineering, University of Waterloo, Waterloo, Canada

J.M. van Noortwijk

HKV Consultants, Lelystad, The Netherlands
Department of Applied Mathematics, Delft University of Technology, Delft, The Netherlands

ABSTRACT: This paper investigates the reliability of a structure that suffers damage due to shocks arriving
randomly in time. The damage process is cumulative, which is a sum of random damage increments due to
shocks. In structural engineering, the failure is typically defined as an event in which structures resistance
due to deterioration drops below a threshold resistance that is necessary for the functioning of the structure.
The paper models the degradation as a compound point process and formulates a probabilistic approach to
compute the time-dependent reliability of the structure. Analytical expressions are derived for costs associated
with various condition and age based maintenance policies. Examples are presented to illustrate computation of
the discounted life-cycle cost associated with different maintenance policies

INTRODUCTION

In industrialized nations, the infrastructure elements

critical to economy, such as bridges, roads, power
plants and transmission lines, are experiencing aging
related degradation, which makes them more vulnerable to failure. To minimize the risk associated
with failure of an infrastructure system, inspection
and replacements are routinely carried out. Because
of uncertainty associated with degradation mechanisms, operational loads and limited inspection data,
probabilistic methods play a key role in developing
cost effective management models. The theory of
stochastic processes and the renewal theorem have
been fundamental to the development of risk-based
asset management models.
The theory of renewal process has been discussed
in several monographs (Cox 1962, Feller 1957 and
Smith 1958 and Barlow and Proschan 1965). The compound renewal process involves two dimensions of
uncertainties, i.e., the inter-arrival time of the event
(shock or damage) and its magnitude are modeled as
random variables (Smith 1958). The concept of using
compound renewal process for modeling degradation
was summarized in Barlow and Proschan (1965). The
analytical formulation of the compound homogeneous

Poisson process is simpler (Mercer and Smith 1959).

Later this model was generalized to compound nonhomogeneous Poisson process by Nakagawa and his
co-workers (Nakagawa 2005, Nakagawa and Mizutani
2007). These models are fundamental to developing
and optimizing condition-based maintenance strategies (Dekker 1996). The use of compound renewal
process in engineering reliability is very limited in
practical applications due to its mathematical complexity and involved computation associated with the
evaluation of convolution integrals. Our experience
also suggests that clarity of mathematical derivation
is essential to promote practical engineering applications.
The objective of this paper is to present a conceptually simple and intuitive interpretation of renewal
processes with applications to modeling conditionbased maintenance policies. The proposed derivation
is very general and it can be reduced to special cases
of homogeneous or non-homogeneous Poisson shocj
processes. Expressions are derived to compute lifecycle costs and other parameters of interest such as
mean life-cycle length.
The failure of a structure occurs when its resistance
due to deterioration drops below a threshold resistance that is necessary for adequate functioning of the

431

structure. This paper investigates the reliability of a

structure that suffers damage due to shocks arriving
randomly in time. The degradation is assumed to be
cumulative sum of random damage increments due to
shocks. In fact we consider a more general version
of the shock model considered in Ito et al., (2005).
The paper models degradation as a compound point
process and formulates a probabilistic approach to
compute time-dependent reliability of the structure.
Analytical expressions are derived for costs associated with various condition and age based maintenance
policies. Simpler expressions are obtained in case of
the shock process being the Poisson process. Examples are presented to illustrate computation of the
life cycle cost associated with different maintenance
policies.
This paper is organized as follows. In section 2, we
present a general model of the stochastic degradation
process. section 3 describes a mathematical framework
to evaluate the total maintenance cost of a system over
a defined planning period. Three types of maintenance
policies are given in section 4 and specific results are
derived for the non-Homogeneous Poisson process and
illustrative examples are included. Conclusions are
presented in section 5.

function given as
G(x) = P(Y1 x).

(4)

It is also assumed that the damages increment Yj and

the shock process N are independent. Since damage
is additive, the total damage Z(t) at time t is given as
N (t)

Z(t) =

Yj .

(5)

j=1

Using the total probability theorem and independence of the sequence Y1 , Y2 , . . . and N (t), we can
write the distribution of cumulative damage as
P(Z(t) x) =

 j


j=0


Yi x, N (t) = j .

(6)

i=1

The distribution of the sum of j damage increments

is obtained by convolution of G(x) given as
G ( j) (x) = P(Y1 + + Yj x),

(7)

which can also be expressed as

2
2.1

MODEL OF DEGRADATION PROCESS

G ( j+1) (x) =

Model

In this paper the degradation process is modeled by a

cumulative damage model, where the system suffers
damage due to shocks. We model the occurrence times
of shocks as the points in a stochastic point process
on [0, ). For t 0, we denote the total number of
shocks during (0, t] by N (t), N (0) 0. Define the
probability of getting j shocks in (0, t) as
Hj (t) = P(N (t) = j),




G (j) (x y) dG(y)

0
x

G(x y) dG (j) (y).

(8)

Note that G (0) (x) = 1 for x 0.

Now the distribution of the total damage can be
written in a more compact form as
P(Z(t) x) =

(1)

Hj (t)G (j) (x).

(9)

j=0

and the expected number of shocks as

R(t) = E(N (t)).

(2)

The probability that the jth shock occurs at time Sj

during (0, t] is given as
Fj (t) = P(Sj t) = P(N (t) j) =

Substituting H0 (t) = 1 F1 (t) and Hj (t) = Fj (t)

Fj+1 (t), j 1, it can be written as
P(Z(t) x) = 1 +


[G (j) (x) G (j1) (x)]Fj (t).
j=1

(10)
Hi (t).

(3)

i=j

Consider that the jth shock causes damage of a

random amount Yj , and it is assumed that damage
increments, Y1 , Y2 , . . . , are iid with the distribution

Suppose damage exceeding a limit xcr causes a structural failure, this equation can be used to compute
reliability as P(Z(t) xcr ). This is a fundamental
expression in the theory of the compound point process, which can be used to derive probabilities of other
events associated with a maintenance policy.

432

2.2

j = j j

Condition-based maintenance

When the total damage Z(t) exceeds an unacceptable level K, it is referred to as the degradation
failure of structure. The failure prompts a corrective
maintenance action (CM) involving the replacement
or complete overhaul of the component (as good as
new repair). On the other hand, a preventive maintenance action (PM) is triggered when the total damage
exceeds a maintenance threshold level k, k < K. For
sake of generality, it is also assumed that the component will be replaced after certain age T0 , whichever
occurs first. After a maintenance action the structure
is good-as-new. We assume that costs are cK for CM,
c0T for replacement at time T0 and c0k for PM before
time T0 . Let T be the first time at which a maintenance
action has to be done and let C denote the associated
cost. Note that T is a random variable, and T0 is a
decision variable and t is used to indicate an arbitrary
(non-random) point in time. We need to derive the
joint distribution of the pair (T , C) to calculate the life
cycle cost.
Define the event that the total damage exceeds the
PM threshold in shocks j to (j 1) ( j = 1, 2, . . . ) as
Aj =

 j1


Yi k <

i=1

j


Yi .

= Aj

j


For any n 1, the collections {A1 , . . . , An , Bn }

CM
PM
CM
and {APM
1 , A1 , . . . , An , An , Bn } are both finite
partitions of the sample space, so
n


i + n =

i=1

n

i=1

i +

n


i + n = 1.

i=1

The joint distribution of (T , C) can now be described

as follows. If N (T0 ) = n, n 1, then

(Sj , c0k ) on APM

j ,

(T , C) = (Sj , cK ) on ACM
j ,

(T0 , c0T ) on Bn .

j n,
j n,

(12)


= Aj
ACM
j

j


A new structure is installed in service and after first

maintenance action at time T1 with cost C1 , a new
cycle starts. Denote the time to the next maintenance
action by T2 and the cost by C2 . We assume that the
sequence (Tn , Cn ) of random vectors is an iid sequence
with


Yi > K .

i=1

Finally, the event that no PM takes place at at jth shock

or time Sj is defined as

Yi k

MAINTENANCE COST MODEL

3.1 General

Yi K

and

i=1

i=1

i=1

Bj =

i = G (j) (k).

i=1

 j


j


If no shocks occurred up to time T0 , i.e. N (T0 ) = 0,

then (T , C) = (T0 , c0T ). We define for convenience in
notation B0 = , so P(B0 ) = 1.

The other two events that PM or CM have to be done

at the jth shock are defined as
APM
j

j = 1

Ai .

(11)

i=j+1

Denote the probabilities of these events by j = P(Aj ),

CM
j = P(APM
j ), j = P(Aj ) and j = P(Bj ). We
can express these probabilities in the distribution G as
follows
j = G (j1) (k) G (j) (k)
 k
j =
[G(K x) G(k x)] dG (j1) (x)

(Tn , Cn ) = (T , C),
where (T , C) is defined in Section 2. The times at
which the cycles start will be denoted by Sn = T1 +
+ Tn , n 1 and S0 0. Let N (t) denote the total
number of maintenance actions during [0, t]:
N (t) = n Sn t < Sn+1 .

(13)

The total cost K(t) up to time t associated with this

maintenance strategy is

K(t) =

N (t)

j=1

433

Cj .

(14)

Let PK be the probability that we have to perform a

corrective maintenance (CM) at time T .

A well-known result is that the long-term expected

average cost C(T ) per time unit is (Tijms 2003)
C(T ) = lim

E(C)
1
K(t) =
.
t
E(T )

(15)

PK =
=

N (t)


erSj Cj 1{Sj t} .

(17)

n



j Hn (T0 )

n=1 j=1

(16)

where r > 0 is a given discount rate. Let K(t, r) denote

the discounted cost over [0, t] of the given maintenance
policy, which is given as

K(t, r) =

P(C = cK ; N (T0 ) = n)

n=1

Consider exponential discounting with discount

factor
D(t) = ert ,

j Fj (T0 ).

The probability that at time T0 the total damage is

still below the managerial level k and the cost of the
maintenance action is c0T is equal to
P(C = c0T ) = P(T = T0 ) = 1 Pk

j=1

=1

The long-term expected equivalent average cost per

time unit is
rE(CerT )
C(T , r) = lim rE(K(t, r)) =
t
1 E(erT )

(21)

j=1

j Fj (T0 ).

(22)

j=1

It follows that
(18)

P(C = c0k ) = 1 (P(C = cK ) + P(C = c0T ))

=

see Weide et al., (2008).

j Fj (T0 ).

(23)

j=1

3.2

Key results

In order to derive the cost rate using Eq.(15) or (18),

a number of probability terms and expectations have
to be evaluated. This section summarizes such expressions, whereas their derivations are presented in the
Appendix.
There are three possible ways for the renewal of
structure at time T : preventive replacement when
Z(T ) = k or age exceeds T0 or complete failure. The
time interval between the renewals, also referred to as
renewal cycle, is a random variable and its expected
value is given as
E(T ) =


j=0

T0

Hj (x) dx.

(19)

For the expected cost for the first maintenance

we get
E(C) = c0T (1 Pk ) + c0k (Pk PK ) + cK PK
= c0T +


cK j + c0k j c0T j Fj (T0 ).

j=1

(24)
If we assume c0 = c0k = c0T and cK = c0 + K ,
we obtain a simple expression
E(C) = c0 + K PK .

(25)

A general expression for expected discounted cost

in an interval T is derived, as shown in Appendix, as

It follows immediately from Eq.(11) that the probability that a maintenance action is necessary before
time T0 equals

E(CerT ) = erT0

[c0 + K Cn ] Hn (T0 )

n=0

Pk = P(T < T0 ) = 1

n Hn (T0 ).

(20)

n=0


n=1

434


(c0 Bn + cK Cn )
0

T0

Hn (x)rerx dx.

(26)

where Bn =

n
j=1

j and Cn =

n
j=1

j . It follows that

E(CerT ) = erT0 E(C)

+r


(c0 Bn + cK Cn )

T0

Hn (x)erx dx.

where as usual G(K)

= 1 G(K). It follows that the
long-term discounted value of the expected equivalent
average cost rate is given by Eq.(18) as

Taking C 1, the expected (discounted) length of the

renewal cycle can be obtained as

E(e

)=1


n

n=0

T0

Hn (x)re

rx

dx.

(27)

r0

r0

C(T , r) =

All the expressions derived above are completely general, without any specific assumptions about the form
of the point process N that describes the random
arrival of shocks in time, see section 2. Note that
lim C(T , r) = lim

E(CerT ) = c0 (1 I (T0 ))


1 erT0 H0 (T0 ) I (T0 ) ,

+ K G(K)

n=1

rT

Since ACM
= {Yi > K}, APM
= {Yi K} and An =
1
1
for all n 2, we get

rE(CerT )
E(C)
=
= C(T ).
1 E(erT )
E(T )

rc0 (1 I (T0 ))
I (T0 )


1 erT0 H0 (T0 ) I (T0 )

rK G(K)
.
+
I (T0 )

4.2 Model 2
= , j 1, which leads to
In this case k = K and APM
j
E(erT ) = 1

G (j) (K)

MAINTENANCE POLICIES

and

This section reanalyzes the following three maintenance policies discussed by Ito et al., (2005). For
simplicity, we will always assume in this section that
c0T = c0k = c0 and cK = c0 + K .
1. Model 1. The system undergoes PM at time T0 or at
the first shock S1 producing damage Y1 , whichever
occurs first. This means that the PM level k = 0.
2. Model 2. The system undergoes PM at time T0 or
CM if the total damage exceeds the failure level
K, whichever occurs first. This means that the PM
level k = K.
3. Model 3. The system undergoes PM only when the
total damage exceeds the managerial level k. This
means that T0 = .


E(CerT ) = cK 1

K erT0

E(erT ) = 1 I (T0 ),

T0

H0 (x)rerx dx.


Hn (x)rerx dx

n Hn (T0 ).

4.3 Model 3
Since in this case T0 = , it leads to


n=0

Hn (x)rerx dx.

and
E(CerT )

n


[c0 j + cK j ]
=
n=1

T0

Using eq.(18), the long-term expected equivalent

average cost per unit time follows from this.

where
I (T0 ) =


n

n=0

Model 1

In this case P(Bn ) = 0 for all n 1 because

k = 0. Using eqn.(27), the expected cycle length is
obtained as


n=0

E(erT ) = 1
4.1

Hj (x)rerx dx,

j=0

T0

j=1

Hn (x)rerx dx.

Using eq. (18), the long-term expected equivalent average cost per unit time follows from these
expressions.

435

where (a, x) denotes the upper incomplete gamma

function

POISSON SHOCK PROCESS

In this Section we analyze the case when the shock

process N (t) is a non-homogeneous Poisson Process
(NHPP) with a continuous intensity function (t). If
the intensity (t) is constant, then the Poisson
process is called homogeneous. The number of shocks
during (0, t] is now Poisson distributed with mean


R(t) =

(28)

t a1 et dt.

Further


(u) du.

(a, x) =

Fj (T0 ) =

T0

Hj1 (x)(x) dx

The probability of exactly j shocks during (0, t] is

Hj (t) = P(N (t) = j) =

R(t)j R(t)
e
,
j!

j = 0, 1, . . .


 a 
1
(j) j, T0b .
b
(j 1)!

The expected cycle length is obtained as

(29)
The probability density of the time Sj at which the jth
shock occurs is

E(T ) =

u > 0,

j 1.

Fj (t) = P(N (t) j) =

Hi (t),

(31)

Fj (t) =

E(C) = c0 + K PK
= c0 + K


Hj1 (u)(u) du.

Hj (x) dx

and expected cycle cost is derived as

or, using the formula for the probability density of the

jth shock,
t

T0

i=j

 

1 b 1/b  G (j) (k)

=
b a
j!
j=0


a 
(j + 1/b) j + 1/b, T0b ,
b

(30)

The probability Fj (t) that the jth shock occurs before

time t is then

G (j) (k)

j=0

R(u)j1
fj (u) =
(u)eR(u)
(j 1)!
= Hj1 (u)(u),





(j) j, ab T0b
(j 1)!
j=1

[1 G(K x)] dG (j1) (x).

(32)

So, the long-term expected average cost C(T0 , k)

per unit time is can be derived from the above two
equations.

Consider a special case with power law intensity model

(t) = at b1

and

R(x) =

a b
x .
b

5.1

Then
Hj (x) =

In this section, we analyze three maintenance policies

discussed in section 4 under the assumption that the
arrival of shocks follow the Poisson process with a
power law intensity function.
In case of the maintenance policy of Model 1 and
without discounting, the long-term expected average
cost C(T0 ) per unit time is obtained as

 a 
1  a b j
x exp xb ,
b
j! b

and substituting y = (a/b)xb , we get



T0

Example

Hj (x) dx

 

1 b 1/b 1 
a 
=
(j + 1/b) j + 1/b, T0b ,
b a
j!
b

436





c0 + K G(K)
1 1, ab T0b

 .
  
1 b 1/b
(1/b) 1/b, ab T0b
b a

190

220
C (T0)

180

C (T ,r)
0

Asymptotic Value

Asymptotic Value

200

170

180

160

160
150

140
140

120
130

100

120

80

110
100
0

0.5

1.5

2.5

3.5

60
0

Figure 1. Plot of C(T0 ) and the asymptotic value for the

case a = 2 and b = 2, c0 = 20, K = 180, G(K)

= 0.40.

2.5

3.5

[1 G(K x)] dG (j1) (x)

c0 + K G(K)
,

/2

can be simplified to G (j1) (K) G (j) (K). So

Figure 1 contains a plot of the long-term expected

average cost C(T0 ) per unit time as function of T0 .
It is interesting to see what happens if we consider
discounting. Taking again a = b = 2, we get

 k



(j) j, ab T0b
[1 G(K x)] dG (j1) (x)
(j

1)!
0
j=1





(j) j, ab T0b  (j1)
(K) G (j) (K)
G
=
(j 1)!
j=1

r r 2 /4
e
(erf (T0 + r/2) erf (r/2)) ,
2

and the long-term expected equivalent average cost

per unit time is obtained by substitution of I (T0 ) in
the expression derived in Subsection 4.1. The limiting
value as T0 is




a

j=0

T0b

j

 a 
exp T0b G (j) (K).
b

So, the long-term expected average cost C(T0 , K) per

unit time is
 a b  (j)
 a b j
 
c0 + K
G (K)
j=0 1 b T0 exp b T0
 


 .
1 b 1/b G (j) (K)
(j + 1/b) j + 1/b, ab T0b
j=0
b a
j!

r(c0 + K G(K))(1
I ())
,
I ()
where
I () =

1.5

we have k = K and the expression



c0 + K G(K)(1
eT0 )
E(C)


=
E(T )
/2 erf (T0 )

I (T0 ) =

Figure 2. Plot of C(T0 , r) and the asymptotic value for the

= 0.40,
case a = 2 and b = 2, c0 = 20, K = 180, G(K)
r = 0.04.

As an example, take a = b = 2. The intensity (t) =

2t and R(t) = t 2 and we get

0.5

r r 2 /4
e
(1 erf (r/2)) .
2

In case of maintenance policy 3, the long-term

expected average cost C(T0 , k) per unit time is
derived as

Figure 2 contains a plot of the long-term expected

equivalent average discounted cost C(T0 , r).
In the rest of this Section we only consider the case
without discounting. In case of maintenance policy 2,

437

c0 + K
1
b

  k

(j1)
(x)
j=1 0 [1 G(K x)] dG
.
 b 1/b  G(j) (k)
(j
+
1/b)
j=0
a
j!

CONCLUSIONS

we get

The paper derives the expressions for computing

discounted life cycle cost rate associated with maintenance policies involving the age-based and conditionbased criteria for preventive replacement. The paper
generalizes the results presented in the literature (Ito
et al 2005), which is the main contribution of this
paper. The proposed model is currently extended to a
more general case in which shocks follow the renewal
process.

E[CerT ] = c0T erT0

E[CerT ] =

n Hn (T0 )

n=0


(c0k j + cK j )E[er Sj ; Sj T0 ].

(33)

j=1


The term E er Sj ; Sj T0 can be expressed in the
distribution functions Fj of Sj as

APPENDIX

This Appendix contains the derivations of the Key

Results in section 3.

E[er Sj ; Sj T0 ] =
= erT0 Fj (T0 ) +

E[CerT ; N (T0 ) = n]




T0

erx dFj (x)

0
T0

Fj (x)rerx dx,

n=0

or in terms of the probabilities Hn (t) as

For n = 0 we have
E[CerT ; N (T0 ) = 0] = c0T erT0 H0 (T0 ),

E[er Sj ; Sj T0 ] =

and for n 1 we split the expectation over the partiCM

PM
CM
tion elements {APM
1 , A1 , . . . , An , An , Bn }. Using
definition (12) of (T , C), we get

= c0k

j E[e

r Sj

; N (T0 ) = n]

hn ,

n=j

where

E[CerT ; N (T0 ) = n]
n


hn = erT0 Hn (T0 ) +

T0

Hn (x)rerx dx.

j=1

+ cK

n


j E[e

r Sj

; N (T0 ) = n]

If we substitute this last expression in (33) we get, after

interchanging the order of summation

j=1

+ c0T erT0 n Hn (T0 ).

E[CerT ]
= c0T erT0

Since
n



j E[er Sj ; N (T0 ) = n]


j=1

E[er Sj ; N (T0 ) = n]

n=j


(c0k Bn + cK Cn )hn ,
n=1



where Bn = nj=1 j and Cn = nj=1 j .
Re-arranging the terms, we can also write
E[CerT ] = c0T erT0 H0 (T0 )

j E[er Sj ; N (T0 ) j]

+ erT0

j=1

n Hn (T0 ) +

n=0

n=1 j=1


(c0T n + c0k Bn + cK Cn )Hn (T0 )
n=1

j E[er Sj ; Sj T0 ],

j=1


n=1

438

(c0k Bn + cK Cn )
0

T0

Hn (x)rerx dx.

(34)

If c0 = c0T = c0k and cK = c0 + K , this formula

simplifies to

REFERENCES


(c0 + K Cn )Hn (T0 )
E[CerT ] = erT0
n=0


+
(c0k Bn + cK Cn )

T0

Hn (x)rerx dx,

(35)

n=1

which is the same as formula (26). Substituting C 1

in (34) and noting that


n=1

T0

Hn (x)rerx dx =

T0

(1 H0 (x))rerx dx

= 1 erT0

T0

H0 (x)rerx dx,

we get formula (27)

E[erT ] = erT0 +


(1 n )
n=1

=1


n

T0

T0

Hn (x)rerx dx

Hn (x)rerx dx.

(36)

n=0

Differentiating with respect to r followed by substitution of r = 0 yields

E[T ] =


n=0

T0

Hn (x) dx,

which is the same as formula (19).

439

[1] Barlow, R.E., and Proschan, F., 1965. Mathematical

Theory of Reliability. Wiley, New York.
[2] Cox, D.R., 1962. Renewal Theory. Methuen, London.
[3] Dekker, R., 1996. Applications of Maintenance Optimization Models: A Review and Analysis. Reliability
Engineering and System Safety. (51): 229240.
[4] Feller, W., 1957. An Introduction to Probability
Theory and Its Applications, 3 edn. Wiley, New York.
[5] Ito, K., Qian, C.H. and Nakagawa, T., 2005. Optimal preventive maintenance policies for a shock
model with given damage level. Journal of Quality
in Maintenance, 11(3): 216227.
[6] Mercer, A. and Smith, C.S., 1959. A Random
Walk in Which the Steps Occur Randomly in Time.
Biometrika, 46: 3055.
[7] Nakagawa, T., 2005. Maintenance Theory of Reliability. Springer series in reliability engineering. London:
Springer.
[8] Nakagawa, T. and Mizutani, S., 2007. A Summary of
Maintenance Policies for a Finite Interval. Reliability
Engineering and System Safety.
[9] Smith, W.L., 1958. Renewal Theory and Its Ramification. Journal of the Royal Statistical Society, Series
B (Methodological) 20(2)(1): 243302.
[10] Tijms, H.C. 2003. First Course in Stochastic Models
John Wiley and Sons, New York, NY
[11] Weide, J.A.M. van der, Suyono and Noortwijk,
J.M. van, 2008. Renewal theory with exponential and
hyperbolic discounting. Probability in the Engineering and Information Sciences, 22(1):5374.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A study about influence of uncertain distribution inputs in maintenance

optimization
R. Mullor
Dpto. Estadstica e Investigacin Operativa, Universidad Alicante, Spain

S. Martorell
Dpto. Ingeniera Qumica y Nuclear, Universidad Politcnica Valencia, Spain

A. Snchez & N. Martinez-Alzamora

Dpto. Estadstica e Investigacin Operativa Aplicadas y Calidad, Universidad Politcnica Valencia, Spain

ABSTRACT: An important number of studies have been published in the last decade in the field of RAM+C
based optimization considering uncertainties. They have demonstrated that inclusion of uncertainties in the
optimization brings the decision maker insights concerning how uncertain the RAM+C results are and how
this uncertainty does matter as it can result in differences in the outcome of the decision making process. In
the literature several methods of uncertainty propagation have been proposed. In this context, the objective of
this paper focuses on assessing how the choice of input parameters uncertain distribution may affect output
results.

INTRODUCTION

Safety related systems performance optimization is

classically based on quantifying the effects that
testing and maintenance activities have on reliability, availability, maintainability and cost (RAM+C).
However, RAM+C quantification is often incomplete in the sense that important uncertainties may
not be considered. An important number of studies have been published in the last decade in the
field of RAM+C based optimization considering
uncertainties (Bunea & Bedford, 2005, Marseguerra
et al., 2004, Martorell et al., 2007). They have demonstrated that inclusion of uncertainties in the optimization brings the decision maker insights concerning how
uncertain the RAM+C results are and how this uncertainty does matter as it can result in differences in the
outcome of the decision making process.
Several methods of uncertainty propagation have
been proposed in the literature. Some of the existing
methods follow an approach based on estimating tolerance intervals, exploiting the advantages of order
statistics to provide distribution free tolerance intervals for the output. The pioneering work was done
by Wilks (Wilks, 1941). Later on, Wald (Wald, 1942)
generalized the Wilkss method for multivariate and
independent cases. Finally, Guba (Guba et al., 2003)

extended the method considering dependence between

the objective functions.
In this context, the objective of this paper focuses on
assessing how the choice of input parameters uncertain distribution in the input parameters may affect
the output. This work uses the results obtained in
a previous paper (Martorell et al., 2006) where the
estimation of different parameters related to equipment reliability and maintenance effectiveness were
carried out using the Maximum Likelihood (MLE)
Method. As a consequence of the parameter estimation process the results obtained in the RAM+C
based optimization are uncertainty. Herein the optimization problem is formulated as a multi-objective
problem where reliability and cost act as decision criteria and maintenance intervals act as decision variables.
A tolerance interval based approach is used to address
uncertainty, which assures a certain level of tolerance/
confidence is achieved with the minimum number of
samples.
The effect of input parameters distribution is analyzed using a normal distribution. This assumption
is the most realistic one since the parameters have
been estimated by maximum likelihood method which
ensures the normal distribution of the parameters.
A second study is performed considering the dependent or independent input parameters.

441

w = t (1 ) +

RELIABILITY AND COST MODELS

As established in the introduction, the objective of

this paper is focused on the maintenance optimization
of safety related equipment, based on reliability and
cost criteria, and considering uncertainty in the equipment reliability characteristics and in the maintenance
effectiveness.
In this context, the objective of this section is to
present the expressions of the objective functions,
averaged reliability and cost, as functions of the
decision variables (maintenance intervals), reliability
model parameters and maintenance effectiveness.
Two models of failure distribution, Weibull and
linear, and two imperfect maintenance models, PAS
(Proportional Age Set-Back) and PAR (Proportional
Age Reduction, are considered.
2.1


 

M 2
R(t) = exp
t mM +
2



t mM + M

R(t) = exp

w =t

m1


(1 )k mk1

w = t m1

(4)

where t is the chronological time, the maintenance

effectiveness which range in the interval [0,1], the
time between two maintenance activities and m the
number of maintenance tasks executed in a period L.
Assuming that the period between maintenance
activities is constant and equal to M, Eqns (3) and
(4) are simplified to:
w = t mM +


 

M 2
R(t) = exp
t (1 ) +
2
2


t (1 ) + M2

R(t) = exp

(2)

k=0

(5)

(8)

(1)

(3)

(7)

Following the same reasoning, if the PAR approach

is considered the expressions corresponding to
the reliability function are given by Eqns. (9) and (10)
for linear and Weibull distributions, respectively.

The general expression of reliability function associated to linear and Weibull failures distribution
respectively are given by:

where w is the age of the component, which depends

on the imperfect maintenance model selected, is the
linear aging rate and and are the shape factor and
the characteristic time which represents the time scale,
respectively.
To take into consideration imperfect maintenance,
the expressions of PAS and PAR models described in
(Martorell et al., 1999) are used. These expressions
are the following:

(6)

Now, substituting Eqns (5) and (6) into Eqns (1) and
(2) the reliability functions of PAS and PAR models
are obtained as continuous functions of time.
Thus, expressions corresponding to the PAS
approach considering linear and Weibull distributions
are given by Eqns (7) and (8):

Reliability function

 
R(w) = exp w2
2
   
w
R(w) = exp

M
2

(9)

(10)

Using Eqns. (7) to (10) the averaged reliability

function can be obtained as:

1 L
R=
R(t)t
(11)
L 0
Where R(t) is obtained from Eq. (9) and (10).
2.2 Cost function
The relevant costs in analyzing maintenance optimization of safety-related equipment include the cost
of performing preventive and corrective maintenance
and the cost associated with replacing the component
(Martorell et al., 2002). The following expressions are
used to quantify these costs
cma
M
1
cc = 8760
( + h M) cca
M
coa
co = 8760
L
cm = 8760

(12)
(13)
(14)

where the term cm represents a yearly cost contribution

as consequence of performing preventive maintenance

442

on the component over a year period, cc represents a

yearly cost contribution as consequence of performing
corrective maintenance and co is the overhaul cost that
represents the yearly cost associated with replacing the
component with periodicity L. In addition, parameters
cma and cca represent the cost associated with conducting each preventive and corrective maintenance,
respectively, coa represents the total cost of replacing
the component while is the cyclic or per-demand
failure probability. For the PAS model, the average
hazard function h is obtained considering the linear and Weibull distributions, obtaining the following
expressions:
M(2 )
h =
2

functions the MOP can be expressed as:

min

 e(C) + (1 ) e(R)

subject to : R Rr ;

C Cr

(19)

that is, the minimization of convex combinations of

both efficiency functions, being  the weighting coefficient and e(C) and e(R) the efficiency of each feasible
solution which can be evaluated as:
e(C) =

C Cr
Cr Co

(20)

e(R) =

Rr R
Ro Rr

(21)

(15)

h =

( M + L (1 ))
2

(17)

The point (Cr , R0 ) is obtained to maximize the

equipment reliability where the cost function evaluated at its initial value, Ci , acts as restriction, and
(C0 , Rr ) is the result of taking as objective function
the equipment cost which must be minimized keeping
the reliability function greater than its initial value, Ri .
The problem of optimization of the maintenance considering reliability and cost criteria has
been solved using Sequential Quadratic Programming
(SQP) (Biggs, 1975).

h =

(M + 2L (1 )) (M)
L (1 ) (2)

(18)

h =

M1
()

(1 (1 ) )

(16)

If the PAR model is considered the expressions of

the averaged hazard function are given by:

Finally, the global cost of the equipment C can be

derived by summing up the corresponding cost contributions of the relevant components using Eqns. (12)
to (14).

MULTI-OBJECTIVE OPTIMIZATION
PROCEDURE

A Multiobjective Optimization Problem (MOP) considers a set of decision variables x, a set of objective
functions f(x) and a set of constraints g(x) based on
decision criteria. In our problem, the MOP consists
in determining the maintenance intervals, over the
replacement period, on each component of the equipment, which maximize the equipment reliability, R,
while minimize its cost, C, subject to restrictions
generated by an initial solution (Ci , Ri ), usually determinate by the values of current maintenance intervals
implemented in plant.
Applying the so called weighted sum strategy, the
multiobjective problem of minimizing the vector of
objective functions is converted into a scalar problem
by constructing a weighted sum of all the objective
functions. In particular, if we have two objective

NON PARAMETRIC METHOD

OF UNCERTAINTIES PROPAGATION

The problem to tackle in this section is how to bound

the uncertainty present in objective functions optimum
values. In the uncertainty analysis N sets of input
parameters x = (x1 , x2 , . . . , xN ) containing variables
are selected randomly. After N runs with fluctuating inputs, using a crude Monte Carlo sampling, we
obtain N random varying output vectors which carry
information on the fluctuating input.
The statistical evaluation of the output parameters
is based on the distribution-free tolerance limits elaborated by Wilks (Wilks, 1941) and extended by Guba
(Guba et al., 2003).
We assume the output comprise p dependent variables. Carrying out N runs, we get a sample matrix
Y = [yij ] where column j corresponds to the result
obtained in the j-th sampling for the p variables.
Using this matrix can be obtained the appropriate
intervals [Li, Ui]/, that is, the construction of
p pairs of random variables Li (y1 , y2 , . . . , yN ) and
Ui (y1 , y2 , . . . , yN ), i = 1, 2, . . . , p such that



L1

443

U1

Lp

g(y1 , . . . , yp )y1 . . . yp > =

Up

(22)

Since the probability of coverage depends on

the unknown joint density function of the output
g(y1 , . . ., yp ), it is necessary to use a reasonable procedure such that the probability is independent of the
joint distribution function of outputs variables.
The generalized version of this method, known as
truncated sample range, consist of being L as the
greatest of the r smallest values in the sample and
U the smallest of the r largest values, that is, let
y(1), y(2), . . .y(N) be the sample values of y arranged
in order of increasing magnitude, then the tolerance interval [L,U] is taken as [y(r), y(N r + 1)]
such that the couple coverage/confidence is achieved.
Coverage () measures the proportion of the distribution included in the random interval [L,U] while
confidence () is the confidence level.
In this work we use the not truncated version where
r = 1, so the tolerance interval becomes [y(1), y(N)],
the minimum and maximum, respectively, values in
the sample, so coverage/confidence levels achieved
depend only in the sample size N. This version consist
of arranging all rows in the sample matrix in order of
increasing magnitude of first row and selecting y1 (1)
as L1 and y1 (N) as U1 . Then, first and last columns
are removed from the arranged sample matrix which is
arranged again, now, in order of increasing magnitude
of second row. From this updated sample matrix, we
choose y2 (1) as L2 and y2 (N 2) as U2 , then, first
and last columns of this matrix are removed to continue selecting tolerance limits for next variables. We
continue this embedding procedure to the last row of
the sample matrix and define p-dimensional volume
Vp = [L1, U1] [L2, U2] [Lp, Up]

(23)

which achieves Eq. (22) depending only on the sample

size N.
Now, it is necessary to find the sample size N
which achieves the coverage/confidence levels previously selected, since achieving these levels depends
only on the selected sample size.
The relation between coverage/confidence levels
and the sample size can be evaluated as:
Nkp 


j=0


N j
(1 )Nj
j

(24)

where and are couple coverage/confidence levels,

N the searched sample size, k the number of limits for
the tolerance intervals (k = 1, 2 one or two sided confidence levels) and p the number of objective functions
compromised in the output. Then, confidence level is
the value, in N-kp, of the distribution function of a random Binomial variable with sample size N and success
probability , the coverage level.

Table 1. Number of runs to obtain tolerance intervals for

k p at / levels.

0.90

0.95

0.99

kp

0.90

0.95

0.99

1
2
3
4
1
2
3
4
1
2
3
4

22
38
52
65
45
77
105
132
230
388
531
667

29
46
61
76
59
93
124
153
299
473
628
773

44
64
81
97
90
130
165
198
459
662
838
1001

Table 1 shows the number of runs needed to determine tolerance intervals for some usual couples of
coverage/confidence levels.
5

APPLICATION CASE

The application case is focused on the optimization process of preventive maintenance associated to
motor-operated safety valves of a Nuclear Power Plant.
The equipment consists of two main components
(actuator (A) and valve (V)) in serial configuration.
Reliability (, , and ) and maintenance effectiveness
() parameters have been previously estimated using
the Maximum Likelihood Estimation (MLE) method.
So, the problem considers how the uncertainty associated to reliability and to maintenance effectiveness
parameters affect on maintenance optimization process based on system reliability and cost (R+C)
criteria.
Table 2 shows the distribution probability, the
imperfect maintenance model and the reliability data
for the actuator and valve necessary to quantify the
equipment reliability. These values represent mean
values obtained in the estimation process. In addition, single cost data necessary to quantify the yearly
equipment cost are showed in Table 3.
Additionally, maximum likelihood estimators have
the property to be distributed asymptotically. Thus,
it is possible obtain the parameters joint distribution
which is given by:

 C)
(, , A , , V ) N(,

(25)

 the varianceBeing
 the mean vector and C
covariance matrix. By using the MLE method the
following values are obtained:

444

Table 2.

Reliability data.

Distribution
IM model

Table 3.

Actuator

Valve

Weibull
PAS
0.8482

7.4708
15400

Linear
PAR
0.7584
1.54E-9

Single cost data for actuator and valve.

Component

cca
[C
=]

cma
[C
=]

coa
[C
=]

Actuator
Valve

3120
3120

300
800

1900
3600

7.4707
15397

 = 0.8482
1.73e 9
0.7584

N(15397, 40808)
N(1.7343e 9, 2.7619e 20)

(26)

and
0.1572
5.5646
 = 1.6944e 3
C

0
0

N(7.4707, 0.1572)
A N(0.8482, 2.2587e 4)

intervals for two outputs have been performed. After

arranging the rows of the sample matrix of runs, the
first and the last value of each output is selected as its
lower and upper tolerance limit, after explained removing of columns explained, obtaining like this 0.95/0.95
coverage/confidence limits for all solutions obtained
in the optimization process which constitute upper and
lower tolerance limits to the Pareto front. Figures 1 and
2 show the tolerance limits obtained for reliability and
cost values, respectively.
Now, we analyze the influence in the tolerance limits considering that the reliability and effectiveness
maintenance parameters are normally distributed but
assuming they are not dependent, thus is:

5.5646
4.0808e + 4
2.3730
0
0

V N(0.7584, 8.4261e 4)

The values associated to the mean and the standard

deviation of the normal distributions are obtained from

1.6944e 3
0
0

2.3730
0
0

2.2587e 4
0
0
(27)

0
2.7619e 20 2.9647e 12
0
2.9647e 12 8.4261e 4

The optimization process is performed under reliability and cost criteria y = {R,C} and the maintenance
intervals for each component act as decision variables.
The equipment reliability and associated cost have
been quantified using the analytical models previously
introduced. A Sequential Quadratic Programming
(SQP) method is used as algorithm to optimization.
(Biggs, 1975).
Both, equipment reliability and cost functions are
considered to be deterministic in the sense that when
all necessary input data for the model are specified
they provide only one value for every output. However,
as inputs of the equipment reliability and cost models fluctuate according to distribution law reflecting
uncertainty on parameters and equipment reliability
and cost will fluctuate in repeated runs. In this case a
multivariate normal distribution whose parameters are
given by Eqns. (25) and (27) is used to characterize
uncertainty.
Following distribution free tolerance intervals
approach discussed in section 4 to address uncertainty,
Wilks equation results in 153 runs to achieve levels of 0.95/0.95 for coverage/confidence in two sided

the mean vector and variance-covariance matrix given

by Eqs. (25) and (27). Figures 3 and 4 show the results
obtained in the optimization process for reliability and
cost functions, respectively.

0,861

0,860

0,859

0,858
3000

3100

3200

3300

3400

3500

3600

Figure 1. R-C plot of uncertain results considering dependency and parameters normally distributed (Tolerance limits
for cost function).

445

Figure 2. R-C plot of uncertain results considering dependency and parameters normally distributed (Tolerance limits
for reliability function).

Figure 3. R-C plot of uncertain results considering independency and normal distributed parameters (Tolerance limits
for cost function).

Figure 4. R-C plot of uncertain results considering independency and normal distributed parameters (Tolerance limits
for reliability function).

Figure 5. R-C plot of uncertain results considering independency and parameters uniform distributed (Tolerance limits
for cost function).

Figure 6. R-C plot of uncertain results considering independency and parameters uniform distributed (Tolerance limits
for reliability function).

Comparing the tolerance intervals obtained in the

two cases analyzed an important reduction of the
tolerance intervals is observed when the uncertain
parameters are considered normally distributed and
dependent.
Finally, we analyze the influence in the tolerance
limits due to the lack of knowledge of the joint
distribution function. According to the principle of
insufficient reason a uniform distribution is used to
characterize epistemic uncertainty in absence of information. So, we assume that the parameters estimated
(, , and ) follow uniform distributions. Upper and
lower limits of the uniform distribution associated to
each parameter are obtained as 2 . Figures 5 and
6 show the results obtained to optimization process for
reliability and cost functions, respectively.
Comparing the results showed in Figures 46 no significant difference in the tolerance intervals obtained
assuming a normal or uniform distribution and independence among the parameters is observed.

446

CONCLUDING REMARKS

This paper presents an analysis of the results obtained

in the preventive maintenance interval optimization
of a safety-related equipment based on reliability and
cost criteria and considering that reliability and maintenance effectiveness parameters are random variables
which introduce uncertainty in the decision making
process.
So, how the input parameters distribution affect
to the output results is analyzed. Thus, different
input parameters distribution and independency or
dependency among parameters were considered. The
application case shows that the effect on the tolerance
intervals of the type of distribution of the parameters
is not significant. However, significant differences are
observed when dependence or independence among
the parameters is considered.
ACKNOWLEDGMENTS
Authors are grateful to the Spanish Ministry of Education and Science for the financial support of this
work in the framework of the Research Project Ref.
ENE2006-15464-C02-01 which has partial financial
support from the FEDER funds of the European Union.
REFERENCES
Biggs MC. Constrained minimization using recursive
quadratic programming. Towards Global Optimization,
North-Holland 1975: 341349.
Bunea C, Bedford T. The effect of model uncertainty on maintenance optimization. IEEE Transactions on Reliability
2002; 51(4): 486493.

Gill PE, Murray W, Wright MH. Practical Optimization.

London Academic Press 1981.
Guba A, Makai M, Pal L. Statistical aspects of best estimation
method-I. Reliability Engineering & System Safety 2003;
80: 217232.
Malik MAK. Reliable preventive maintenance scheduling.
AIIE Transactions 1979; 11: 221228.
Marseguerra M, Zio E, Podofillini L. Optimal reliability/availability of uncertain systems via multi-objective
genetic algorithms. IEEE Transactions on Reliability
2004; 53(3): 424434.
Martorell S, Sanchez A, Carlos S, Serradell V. A tolerance interval based approach to address uncertainty
for RAMS+C optimization. Reliability Engineering &
System Safety 2007; 92: 408422.
Martorell S, Sanchez A, Carlos S, Serradell V. Age-dependent
reliability model considering effects of maintenance and
working conditions. Reliability Engineering & System
Safety 1999; 64: 1931.
Martorell S, Sanchez A, Carlos S, Serradell V. Simultaneous
and multi-criteria optimization of TS requirements and
maintenance at NPPs. Annals of Nuclear Energy 2002;
29(2): 147168.
Nutt WT, Wallis GB. Evaluation of nuclear safety from the
outputs of computer codes in the presence of uncertainties.
Reliability Engineering & System Safety 2004; 83: 5777.
Rocco CM, Miller AJ, Moreno JA, Carrasquero N, Medina M.
Sensitivity and uncertainty analysis in optimization programs using an evolutionary approach: a maintenance
application. Reliability Engineering & System Safety
2000; 67(3): 249256.
Sanchez A, Martinez-Alzamora N, Mullor R, Martorell S.
Motor-operated valve maintenance optimization considering multiple failure modes and imperfect maintenance
models. Proceedings of ESREL 2007.
Wald A. Setting of tolerance limits when the sample is large.
Annals of Mathematic Statistics 1942; 13: 389399.
Wilks SS. Determination of the sample size for setting tolerance limits. Annals of Mathematic Statistics 1941; 12:
9196.

447

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Aging processes as a primary aspect of predicting reliability and life

of aeronautical hardware

J. Zurek,
M. Zieja & G. Kowalczyk
Air Force Institute of Technology

T. Niezgoda
Military University of Technology

ABSTRACT: The forecasting of reliability and life of aeronautical hardware requires recognition of many
and various destructive processes that deteriorate the health/maintenance status thereof. The aging of technical
components of aircraft as an armament system proves of outstanding significance to reliability and safety of the
whole system. The aging process is usually induced by many and various factors, just to mention mechanical,
biological, climatic, or chemical ones. The aging is an irreversible process and considerably affects (i.e. reduces)
reliability and life of aeronautical equipment.

INTRODUCTION

The aging processes that affect aeronautical equipment

are to a greater or lesser degree correlated with the
items time of operation or the number of cycles of
its operation. In terms of aging processes the items
of aeronautical equipment can be divided into three

groups (Zurek
2005):
those with strongly correlated changes in values
of diagnostic parameters with time or amount of
operation,
those with poorly correlated changes in values
of diagnostic parameters with time or amount of
operation, and
ones showing no correlation changes in values
of diagnostic parameters with time or amount of
operation.
For the items representing the first group one can
predict the instance of time when the diagnostic parameters boundary condition occurs. One can also predict
the time instance of the items safe shut down and then
plan appropriate maintenance actions to be carried out.

1. Health/maintenance status of any item included

in the aeronautical equipment can be described
with diagnostic parameters available throughout
the operational phase, and designated in the following way:
X = (X1 , X2 , X3 , . . . , Xn )

(1)

2. Values of diagnostic parameters change due to

aging processes going on all the time. It is assumed
that these changes are monotonic in nature; they
can be presented in the following way:


Xi = Xi Xinom  ,

i = 1, 2, 3, . . . , n

(2)

where: Xi = absolute value of deviation of the

diagnostic parameter from the nominal value;
Xi = current value of the i-th parameter;
Xinom = nominal value of the i-th parameter.
3. Any item of the aeronautical equipment is serviceable (fit for use) if the following dependence
occurs:

A METHOD TO FORECAST RELIABILITY

AND LIFE OF SOME SELECTED ITEMS
OF AERONAUTICAL EQUIPMENT

Xi Xi

(3)

where: Xi = absolute value of boundary deviation

of the diagnostic parameter from the nominal value.

What has been assumed in the already developed

method is as follows (Zurek

2005):

449

To be clearer, the following terms (notations)

have been introduced:
Xi = zi
g
Xi

g
zi

Equation (9) is now rearranged to take the form of a

partial differential equation of the Fokker-Planck type:
u(zi , t)
u(zi , t) 1 2 2 u(zi , t)
= C
+ C
t
zi
2
zi2

(4)
(5)

where: zi = absolute value of deviation of the

diagnostic parameter from the nominal value;
g
zi = absolute value of boundary deviation of the
diagnostic parameter from the nominal value.
Equation (3) can be, therefore, written down in
the following form:

(10)

Since C is a random variable, an average value of

this variable is introduced. It has the form:
Cg
Cf (c)dc

E[C] =

(11)

Cd
g

zi zi

(6)

4. Values of changes in diagnostic parameters grow

randomly.
5. Changes in diagnostic parameters accepted for the
assessment of health/maintenance status of individual items of aeronautical equipment are independent random variables, i.e. any change of any
of these parameters does not result in any change
in values of other parameters.
6. The method has been dedicated to some selected
items of the aeronautical equipment, namely to
those for which the rate of changes in diagnostic parameters can be described with the following
dependence:
dzi
=C
dt

(7)

where: C = operating-conditions dependant random variable; t = calendar time.

The dynamics of changes in values of deviations
of assumed diagnostic parameters, if approached randomly, is described with a difference equation. One
arbitrarily chosen parameter zi has been accepted for
analysis. The difference equation for the assumptions
made takes the form (Tomaszek 2001):
Uzi ,t+t = PUzi zi ,t

u(zi , t)
u(zi , t) 1 2 u(zi , t)
= b
+ a
t
zi
2
zi2

(9)

where: u(zi , t) = time-dependant density function of

changes in diagnostic parameter.

(12)

where: b = E[C]an average increment of value of

deviation of the diagnostic parameter per time unit;
a = E[C 2 ]a mean square increment of value of
deviation of the diagnostic parameter per time unit.
We need to find a partial solution of equation (12),
one that at t 0 is convergent with the so-called
Dirac function: u(zi , t) 0 for zi  = 0, but in such
a way that the function integral u(zi , t) equals to unity
for all t > 0. This solution takes the form:
u(zi , t) =

(zi B(t))2
1
e 2A(t)
2A(t)

(13)

where:
t

t
adt = at

A(t) =
0

(8)

where: Uzi ,t = probability that at the instance of time

t the deviation of a diagnostic parameter takes value
zi ; P = probability that the value of the deviation
increases by value zi within time interval of t.
Equation (8) takes the following form if function
notation is used:
u(zi , t + t) = u(zi zi , t)

where: f (c) = density function of the random variable

C; Cg , Cd = upper and lower values of the random
variable C.
Taking account of equation (11) while considering
formula (10) the following dependence is arrived at:

B(t) =

bdt = bt
0

Function (13) is a probabilistic characteristic of

changes of the diagnostic parameter due to effects
of aging processes, the rate of which can be determined with equation (7). Density function of changes
in value of the diagnostic parameter can be used
directly to estimate reliability and life of an aeronautical device, the health/maintenance status of which is
estimated with this parameter. Applying the density
function of changes in values of the diagnostic parameter to determine distribution of time of exceeding the
boundary condition is a good example of such a solution. Probability of exceeding the boundary value by

450

the diagnostic parameter can be presented using density functions of changes in the diagnostic parameter
(Tomaszek 2001):

dependence (20):

=

t 2 f (t)zig dt (E [T ])2

Q(t, zi ) =

1
2 at

(zi bt)2
2at

dz

(14)
Hence

Zi

=
2

To determine the density function of time of exceedg

ing the admissible value of deviation zi for the first
time one should use the following dependence:
f (t) =

g
Q(t, zi )
t

(15)

Substitution with equation (14), introduced in equation (15), gives:

f (t) =

1
2 at

(zi bt)2
2at

dz

(16)

azi +

 g 2
 g 2
z
b zi
5a2
+ 4 i2
3
b
4b
2b

(21)

The presented method of determining the distribution of time of exceeding the boundary condition
by the diagnostic parameter allows of finding the
density function) of time of reaching the boundary state. On the basis thereof one can determine
reliability of a given item of aeronautical equipment, the health/maintenance status of which is estimated by means of the diagnostic parameter under
consideration:
t
R(t) = 1

Zi

f (t)zig dt

(22)

Using properties of the differentiation and integration, the following dependence is arrived at:
g

f (t)zig =

(20)

(zi bt)2
zi + bt
1
e 2at

2t
2 at
g

(17)

The probability density function that determines

distribution of time of the diagnostic parameters value
passing through the boundary condition allows also
of calculating the aeronautical items life. Therefore,
the level of risk of exceeding the boundary condition
should be found:
t

Equation (17) determines density function of time

of exceeding the boundary condition by values of the
diagnostic parameter. What is to be found next is the
dependence that determines the expected value of time
of exceeding the boundary condition by the diagnostic
parameter:

E [T ] =

tf (t)zig dt

The value of time, for which the right side of equation (23) equals to the left one, determines life of
an item of aeronautical equipment under conditions
defined with the above-made assumptions.

Hence
g

(23)

(18)

E [T ] =

f (t)zig dt

Q(t)zig =

z
a
a
z
zi
+ i + 2 = i + 2
2b 2b 2b
b
2b

(19)

We also need to find the dependence that determines the variance of distribution of time of exceeding
the boundary condition by the diagnostic parameter. In general, this variance is determined with

ESTIMATES OF LIFE AND RELIABILITY

OF AIRBORNE STORAGE BATTERIES

Airborne storage batteries are those items of aeronautical equipment that show strong correlation between
changes in values of diagnostic parameters and time or
amount of operating time. Capacitance Q is a diagnostic parameter directly correlated with aging processes
that take place while operating airborne storage batteries, one which explicitly determines the expiry date
thereof (Fig. 1). The presented method allows of estimating the reliability and residual life of airborne
storage batteries using diagnostic parameters recorded

451

in the course of operating them. A maximum likelihood method was used in order to estimate parameters
a and b in the equation (17). Gained with the hitherto
made calculations for the airborne storage batteries
12-SAM-28 are the following characteristics of the
density function of time of exceeding the boundary
condition by values of the diagnostic parameter f (t)
and the reliability function R(t). They are shown in
Figs. 2 and 3.
With the above-presented method applied, the following values of life T and residual life Tr have been
gained for particular storage batteries 12-SAM-28 (see
Table 1).

Storage batteries 12-SAM-28

1,2

596
280

330

R(t)

0,8

159
170

0,6

525
112

0,4

180
0,2

574
109

0
0

10

20

30

40

50

60

70

80

90

100

Service life - t[months]

Figure 3. Characteristic curves of the reliability function

R(t) for storage batteries 12-SAM-28.

CONCLUSIONS

The method introduced in this paper allows of analysis

of health/maintenance status of some selected items
of aeronautical equipment because of the nature of
changes in values of diagnostic parameters available

Storage batteries 12-SAM-28

29
28
109

Capacitance - Q[Ah]

27

574

26

180
25

112

24

525

23

170
159

22

330
21

Table 1.

Estimated values of life and residual life.

STORAGE BATTERIES 12-SAM-28

No

Battery No.

T [months]

Tr [months]

1
2
3
4
5
6
7
8
9
10

109
574
180
112
525
170
159
330
280
596

27.7
29.2
34.54
41.19
50.71
47.8
26.13
23.73
29.12
20.85

13.76
11.28
17.5
23.47
32.17
30.38
10.61
6.98
12.32
5.7

280

20

596

19
18
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Service life - t[months]

Figure 1. Characteristic curves of the changes in capacitance values for storage batteries 12-SAM-28.

Storage batteries 12-SAM-28

0,06
596

0,05

280
330

f(t)

0,04

159
170

0,03

525
112

0,02

180
574

0,01

109
0
0

10

20

30

40

50

60

70

80

90

100

Service life - t[months]

Figure 2. Characteristic curves of the density function f (t)

for storage batteries 12-SAM-28.

throughout the operational phase thereof. Determination of how the values of diagnostic parameters and
deviations thereof increase enables determination of
time interval, within which a given item remains fit
for use (serviceable).
Dependence for the rate of changes in value of
the diagnostic parameter, i.e. equation (7), is of primary significance in this method. The method will
not change substantially if other forms of this dependence (i.e. equation (7)) are used. These different
forms may result in changes of coefficients in the
Fokker-Planck equation (10), which in turn will result
in changes of the dependences for both an average
value and variance of the density function of changes
of the diagnostic parameter. The method offers also
a capability of describing aging and wear-and-tear
processes within a multi-dimensional system. The
above-presented method allows of:
assessment of residual life of some selected items of
aeronautical equipment with the required reliability
level maintained,

452

 estimation of reliability and life of some selected

items of aeronautical equipment on the grounds of
diagnostic parameters recorded in the process of
operating them,
verification of the process of operating some
selected items of aeronautical equipment to maintain the required level of reliability between particular checks.
The way of proceeding suggested in the method
under examination can be adopted for specific characteristics of aging and wear-and-tear processes
that affect various items of aeronautical equipment
throughout operational phase thereof.

REFERENCES

Jazwinski, J. & Zurek

J. 2007. Wybrane problemy sterowania
zapasami. Radom.
Tomaszek, H. & Wrblewski, M. 2001. Podstawy oceny
efektywnosci eksploatacji systemw uzbrojenia lotniczego. Warszawa.

Zurek,
J. 2006. Zywotno
sc smiglowcw. Warszawa.

Zurek,
J. & Tomaszek, H. 2005. Zarys metody oceny
niezawodnosci statku powietrznego z uwzglednieniem
uszkodzen sygnalizowanych i katastroficznych (naglych).
Warszawa.

453

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

An alternative imperfect preventive maintenance model

J. Clavareau
Universit Libre de Bruxelles (U.L.B), Service de mtrologie Nuclaire, Belgium
Grant F.R.I.A, Belgium

P.E. Labeau
Universit Libre de Bruxelles (U.L.B), Service de mtrologie Nuclaire, Belgium

ABSTRACT: All systems are subject to aging. When a component is aging, its global performances are decreasing. In order to reduce the effect of aging on a system, preventive maintenance actions can be performed. Yet the
rejuvenation of a unit that can be achieved thanks to these interventions is most of the time not total, and does not
correspond to the classical as-good-as-new assumption. Imperfect maintenance models have been proposed in the
literature in order to embody the partial efficiency of preventive actions. This paper reviews the approaches that
are available in the literature to model imperfect maintenance. It also proposes to extend and modify some of them,
in order to obtain a simple and mathematically exploitable model, associated with a more realistic time-dependent
behavior of the component than that corresponding to the application of previous imperfect maintenance models.

INTRODUCTION

All systems are subject to aging. When a component is aging, its global performances are decreasing.
As a consequence of aging, the failure probability is
increasing and the productivity of the operated system
is lower. In order to reduce the aging effect on a system,
preventive maintenance actions can be performed. The
most easily modeled preventive maintenance is the
replacement of the system: after maintenance, the system is As Good As New (AGAN). The maintenance
policies based on preventive replacement of the system have been studied from a long time, starting from
Barlow & Proschan 1965.
The reality is often quite different: preventive maintenance interventions are made to reduce the aging
effect or to delay the apparition of these effects but
these actions are not perfect. The goal of imperfect preventive maintenance is to maintain the performances
of the system within an acceptable range. Also, the
efficiency and the cost of each maintenance intervention will be dependent on which type of actions is
undertaken. For example, in a car, oil, brake pads or
tires have to be periodically changed, but none of these
actions will make a new car. The way they affect the car
state, the cost of each action and also the optimal periods to carry them out will be different from one action
to the other. We also know that, despite all possible
maintenance actions, the car ages. When the performances of the system tend to go outside the acceptable
range mentioned above, it becomes soon or later more

interesting to replace it than to keep maintaining it.

This trade-off between maintenance vs. failure costs
and new investments defines the end of the working
life of the system.
In practice, all the variables affecting the aging
process are actually not known. The issue of reducing the evolution of all the different degradation
factors to a probability law for the component lifetime,
which is mathematically exploitable and depending on
few parameters, remains open. Most authors consider
strictly increasing failure rates, usually in the form of
a single-mode, two-parameter Weibull law (see eq.(2)
below).This form is actually quite different from the
commonly accepted bathtub curve, as it constraints the
failure rate to start from a zero value and often leads
to estimations of the shape parameters lying between
1 and 2, hence to an exponent lower than 1 in the
time-dependent expression of the failure rate.
In this paper, we propose, in section 2, a quick
overview of maintenance efficiency models available in the literature and in section 3 we propose a
model extending previous models to a more realistic
description of the aging process, including the maintenance interventions. section 4 briefly summarizes the
different points treated in section 3.
2

MODELS REVIEW

As previously said, in Barlow & Proschan 1965,

models with AGAN policies are studied.

455

Brown & Proschan 1983 envisaged a repair model

in which a repair is minimal (As Bad As Old) with a
probability p and degraded with a probability 1 p.
Minimal repair means that the failure rate after repair
is the same as before the failure. The degradation
of the component considered is an increase of the
failure rate. With a constant failure rate between failures, this model leads to a piecewise increasing failure
rate in time. A variant of this model for preventive maintenance can be deduced from there. In this
preventive maintenance model, the preventive intervention is AGAN with a probability p and ABAO with
a probability 1 p.
Generally, imperfect preventive maintenance
models can be divided in two classes (Doyen & Gaudoin 2004) depending on whether the maintenance
actions affect the failure rate (Nakagawa 1986, Nakagawa 1988, Zequeria & Brenguer 2006) or the effective age (Canfield 1986, Kijima 1988, Malik 1979,
Martorell et al., 1999) of the component. We can also
cite a hybrid model (Lin et al., 2000) for which the
effect of the maintenance actions reduces the value of
both the failure rate and the effective age. Obviously
the value of the failure rate is determined by the value
of the effective age, so there is a direct link between
the two classes. Yet we will see below that the maintenance effect modeled with one of the two classes is
not straightforwardly obtained with the other one.
These works discuss ways of modeling the maintenance effects either in case of preventive maintenance
(PM) or in case of corrective maintenance (CM). Here
we will summarize the main formulations in the case
of preventive maintenance with minimal repair (the
components state just after repair is the same as just
before failure) which is equivalent to compute the distribution of the first failure. If repair are not minimal
the intervention effect can be modeled exactly in the
same way that the intervention effect of a preventive
imperfect maintenance. In this work we will thus focus
on the formulation where CM is minimal.
The distribution of the failure time is completely
characterized by the conditional failure intensity
defined by:
t 0, t = lim

dt0

1
P(Nt+dt Nt = 1| Ht )
dt

behavior:
(t) =

t 1
( ) .

(2)

where is the scale parameter (in time unit) and is

the shape parameter of the distribution.
2.1 Review of failure rate impact models
2.1.1 Arithmetic Reduction of Intensity (ARI) model
Following Doyen & Gaudoin 2004, we will define the
Arithmetic Reduction of Intensity (ARI) as follows:
the failure rate after maintenance is taken equal to
the failure rate before maintenance, minus a given
quantity.
We can distinguish three particular cases:
The reduction is a fraction of the augmentation of
the failure rate from the last maintenance. We have,
immediately after the M th PM intervention:

+
+
TM = TM (TM TM 1 ).

(3)

where is the efficiency of the maintenance

between 0 and 1, TM is the time of the M th PM
action,
TM the value of the failure rate just before
the maintenance and +
TM the value of the failure rate
just after the maintenance.
This model is called ARI1 and gives the following
failure rate at time t:
t = (t) (TMt )

(4)

where Mt is the is the number of maintenance

actions up to time t.
We can see that the failure rate at time t is thus given
by only one subtraction.
The reduction is proportional to the global wear out
before the maintenances. This model is called ARI
and we have:

(1)

where Ht is the past history of the component, i.e. the

set of all events having occurred before t and Nt the
number of failures observed up to time t. For the sake
of simplicity, we will use the expression failure rate
instead of conditional failure rate in the sequel of
the text.
All the above cited references assume that the initial
failure (i.e, the failure rate before any maintenance
action) rate is strictly increasing. The initial failure
rate is then most of the time given by a Weibull-like

456

+
Tm = Tm Tm .

(5)

Considering a constant time interval TPM between

two consecutive preventive maintenance actions and
minimal repair in case of failure, the ARI model
gives the following expression of the failure rate at
time t:
t = (t)

M
t 1


(1 )j ((Mt j)TPM )

(6)

j=0

where Mt is the number of maintenance actions

performed up to time t. The reduction is thus proportional to a potentially infinite sum, hence the
name of the model.

 An intermediate case between the two previous

ones can be considered. If we keep only the first
m terms in the sum in (6), we have the ARIm
model:
t = (t)

min(m1,M
 t 1)

(1 )j ((Mt j)TPM )

j=0

(7)

The ARI1 and ARI models are then particular cases of the ARIm model. In the ARI family of
models, the value = 0 means that the intervention is ABAO but, when = 1, the intervention is
not AGAN because the failure rate evolution with
time is different from the evolution of the initial
failure rate of the component. This behavior is at
the same time an advantage and a drawback of the
model: there is a part of the aging, related to the
working time, that is unavoidable, but the replacement of a component is not included in the model
of impact. Figure 1 below illustrates the behavior
of different models including ARI1 and ARI for a
given initial failure rate. We can see that the ARI
models result in a piecewise vertical translation of the
failure rate value and hence keep the wear out trend.
Figure 1, below, gives an example of the resulting failure rate for different maintenance efficiency
models, including ARI1 and ARI models.
2.1.2 Nagakawa model
Nakagawa 1986, 1988 give a model where each PM
resets the value of the failure rate to 0. But after a PM,
the failure rate evolution is increased by an adjustment
factor bigger than 1. We have:
Mt
+
Tm = (t Mt TPM )

(8)

This Nagakawa model assumes that the failure rate

of a component after a PM action is a product of the
adjustment factor and the failure rate before the action.
The value of the adjustment factor is a measure of the
PM quality. Indeed, if is equal to 1, the maintenance
is AGAN. This model seems non-intuitive because in
a certain way, the PM action increases the failure rate.
If the factor intends to emphasize the fact that there
is an irreversible part in the aging, it seems that must
then depend on the working time. Indeed, in this situation, the higher the frequency of the PM, the higher
the increase of the failure rate evolution, what is unexpected. Also, there is no ABAO possibility because
the model always resets the failure rate to 0.

failure mode, with the failure rate (t), and the nonmaintainable failure mode, with the failure rate h(t).
They propose to take into account dependence between
these two competing failure modes. The failure rate
of the maintainable failure mode is the sum of the
initial failure rate (t) plus a positive value p(t)h(t).
They suppose that without imperfect PM the maintainable failure mode is AGAN and the non-maintainable
failure mode remains unchanged.
Under this model the failure rate of the component
at time t is given by:
t = h(t) + (t Mt TPM ) + p(t Mt TPM )h(t)

(9)

With this model, the efficiency of maintenance is

decreasing with time.
In this model, there is no flexibility in the maintenance impact to adapt it to a specific type of
maintenance action. The maintenance efficiency does
not depend on the undertaken actions on the component, but only on the internal parameters which are the
maintainable and the non-maintainable failure rates,
as well as the model of dependence between these two
failure modes. It also seems difficult to determine the
dependence function p(t) for practical cases.
2.2 Review of effective age impact models
In these models, the actual working time is replaced
by an effective age, denoted t in the estimation of the
failure rate at time t. The value of this effective age
should represent the state of wear out of the component. The concept of effective age is sometimes also
referred to as virtual age in the literature.
As for the failure rate model, following Doyen &
Gaudoin 2004, we will define the Arithmetic Reduction of Age (ARA). In the ARA model, the effective
age after maintenance results from a decrease of the
effective age before the maintenance. We can also distinguish between an ARA1 model, an ARA and the
general ARAm model:
ARA1 : the reduction of the effective age due to the
intervention is proportional to the increase of age
between two maintenance actions:
T+M = TM (TM T+M 1 ),

(10)

with the efficiency such that 0 1 and

t = (t ).
ARA : the reduction of the effective age due to
the intervention is proportional to the age before
maintenance:

2.1.3 Maintainable and non-maintainable failure

mode model
In Zequeira and Brenguer 2006, the failure process
is separated in two failure modes: the maintainable

457

T+M = TM TM ,
with 0 1 and t = (t ).

(11)

 ARAm : by analogy with the ARIm , with a constant

maintenance time interval and a constant efficiency,
we can define:
t = t

min(m1,M
 t 1)

(1 )j (Mt j)TPM (12)

j=0

with 0 1 and t = (t ).
If = 1, we have an AGAN intervention and, if
= 0, we have an ABAO intervention.
The model from Canfield 1986, the Kijima1 model
(Kijima 1988) and the Proportional Age Setback
(PAS)(Martorell et al., 1999) are all equivalent to the
ARA1 model.
The model from Malik 1979, the Kijima2 model
(Kijima 1988) and the Proportional Age Reduction
(PAR)(Martorell et al., 1999) are all equivalent to the
ARA model.
It is obvious that with a failure rate of the form (2)
and = 2, the ARA models are equivalent to the ARI
models.
Note also that, with a constant efficiency and a
constant maintenance time interval, the ARA1 model
gives:
t = t Mt TPM

(13)

and the ARA model gives:

t = (t Mt TPM ) +

Mt


(1 )i TPM

(14)

i=1

For the ARA model, the constant efficiency leads

to an upper bounded effective age:
lim T+M =

1
TPM .

(15)

Indeed, using (14), we have:

TPM lim

M



(1 )i = TPM

i=1

1
1
1 (1 )

1
TPM

(16)

considering that is between 0 and 1.

This upper bound of the effective age implies an
upper bound in the failure rate of the component.
We can see this behavior in Figure 1, where the
ARA curve is the lowest one.
We can see, in Figure 1, that, obviously, the ARA
model results in a piecewise translation in time of the
initial failure rate.

Figure 1. Conditional failure rate for models ARI1 , ARI ,

ARA1 and ARA .

2.3 Extended models

In Wu and Clements-Croome 2005, the authors extend
the Nakagawa and the Canfield model. They assume
that the maintenance efficiency is a random variable.
This can be suitable when the maintenance efficiency
is estimated between two boundaries or to emphasize
the fact that each maintenance operation may have a
different impact, thus, embodying the variability in the
resulting state of the component.
To avoid the problem of the upper bound in age
Fouathia and al. 2005 propose not only to use, for
the ARA model, a random variable in the maintenance efficiency but also to assume that the expected
value of this efficiency is decreasing with the number of PM interventions undertaken on the component. In Clavareau and Labeau 2006, the same
idea is used but with an expected value of the
efficiency decreasing with the effective age of the
component.
Martorell et al., 1999 proposes in addition to the
PAS and PAR models to use the Accelerated Life
Model (ALM) to describe the influence of the working conditions on the effective age. Between the
maintenance actions, the augmentation of the effective age is no longer linear but it is proportional to
T
a factor (z) = e z , where z is the vector of q
working state variables and T the transposed vector of the q corresponding regression coefficients.
In their applications, they assume a constant value
for the maintenance efficiency and constant working
conditions. If the working conditions are constant,
this model results in a change in the value of the
scale parameter depending on the specific use of the
component.
Related to this problem of working conditions we
can cite the proportional intensities and proportional
hazard model. Kumar & Klfesjo 1994 review the
existing literature on the proportional hazard model.

458

In Samrout et al., 2008, the authors propose that a PM

action reduces the effective age of the component by a
reduction factor (such as in an ARA model) depending on the costs of the intervention. Percy and Kobbacy
2000 propose that the intensity function is given by:
T
(t) = 0 (t)e x for hazards following a PM and by:
T
(t) = 0 (t)e y for hazards following a CM, where t
measures time since the most recent event and x is the
vector of working state variables, which can possibly
be modified by the maintenance actions.

MODEL EXTENSION

3.1

Bi-Weibull failure rate

The previously cited papers always consider strictly

increasing failure rates in their applications. The latter follow a power law according to equation (2). This
form is quite different from the bathtub curve usually
agreed upon as the most realistic time-dependent evolution of the failure rates. After the infant mortality
region, there is a possibly long period of life where
the components aging is limited, or even negligible,
as long as the physical characteristics of the system
under study are or are maintained within an acceptable range. Nonetheless the component is aging and
at some point, the aging process accelerates, and it is
preferable to replace the component than to maintain
it. To approach this behavior, we can use an initial failure rate constant in a first period of time, before being
increased by an additional contribution following the
power law:
(t) = 0

if t

t 1
)
(t) = 0 + (

if t >

(17)

The corresponding cumulative distribution is given

by the following bi-Weibull law:

F(t) =

1 e0 t
t
1 e0 t( )

if t
if t >

(18)

Later, we will call 0 the constant failure rate and

the location parameter.

Barlow and Proschan model of imperfect repair, there

is no difference between the ABAO and the AGAN
interventions before aging has started.
3.2.1 Location parameter impact model
Because the aim of the PM is to maintain the unit
performance in an acceptable range, one can imagine that, before aging started, the effect of a PM is
to delay the wear-out onset of the component. We
can model this delay by using, instead of an effective age, an effective location parameter in equations
(17) and (18). The effect of a PM will then be to
increase this effective location parameter. Two models
can be proposed. In the first one the location parameter is incremented by an amount proportional to the
maintenance period TPM :
T+M = TM + TPM

where the maintenance efficiency is between 0 and

1. When is equal to 0, the intervention is ABAO and
when is equal to 1, the intervention is AGAN.
Equation (17) gives:
T+M = 0 +

M


TPM = 0 + M TPM

With this assumption of a constant failure rate on an

initial period, the previous models of imperfect intervention are partly irrelevant as they assume the aging
has already started from the beginning of the component operation. Indeed, it would be meaningless to
reduce this constant failure rate if no major change
is made on the component. In the same way, in the

(20)

i=0

where 0 denotes the initial value of the location

parameter.
In the second model, the location parameter is equal
to the initial location parameter incremented by an
amount proportional to the age of the component:
T + = 0 + T .
M

(21)

If we suppose that repairs are minimal, we have

Ti = M TPM and thus equation (21) is equivalent to
equation (20).
We can also see that the model of equation (21) is
equivalent in terms of failure probability to the ARA
model. Indeed, if we compute the value T+M T+M ,
because, in the model from equation (21), the age
after maintenance, T+M , is the same as the age before
maintenance, TM , we have:
T+M T+M = TM TM 0

3.2 Maintenance model

(19)

(22)

And the quantity TM TM is equal to the effective

age after intervention in the model ARA (equation
(11)). Thus, when the effective age at time t, t , is lower
than the location parameter, the delay till the aging
onset due to the imperfect PM can be symmetrically
obtained by reducing the value of the effective age or
by increasing the value of the location parameter.

459

3.2.2 Extended effective age model

We propose thus to use the effective age as a degradation criterion. Depending on the previous interventions, the effective age will give the global performance state of the component. With this assumption
we only have one parameter describing the entire aging
process of the component. Before any intervention on
the component, its effective age is equal to its working
time. The interventions can modify the value of the
effective age and between two consecutive interventions, in normal working conditions, it is assumed that
the increase of the effective age is equal to the time
between the two interventions.
In most of the works cited in the review of section 2,
excepted Zequeira & Brenguer 2006, Fouathia et al.,
2005 and Clavareau and Labeau 2006, the maintenance action is characterized by a constant efficiency
and a constant cost at each intervention, no matter what
maintenance period is considered. With an effective
age model, it leads to considering that the absolute
effect of the maintenance is always higher when the
maintenance time interval is increasing, but with no
impact on the intervention cost. When we are trying to
optimize the interval of maintenance, it seems not logical, that the same intervention, with the same costs,
can better restore the component, only because the
component is more aged.
The amount of effective age decrease after a PM
intervention will obviously depend on the action
undertaken during this intervention. We propose to
not fix a priori the efficiency of the maintenance but
to relax this assumption and let the maintenance effect
vary as a function of the undertaken action in the
maintenance interventions. We will have:
T+M = TM M

(23)

The amount of the decrease, M , will characterize

the effect of the M th maintenance. The value of M
can be interpreted directly as the augmentation of the
expected residual lifetime of the component due to the
intervention. With this formalism, if M is equal to
0, the intervention is ABAO, if M is equal to TM the
maintenance is AGAN, if M is lower than TM it corresponds to an imperfect intervention and possibly if M
is greater than T the intervention is more than perfect.
M
Note that the ARA1 and ARA models are particular cases of this more general model. In ARA1 ,
the maintenance effect, M , is equal to TPM , and, in
ARA , we have M = TM .
In the case of the initial failure rate (15) we have a
resulting conditional failure rate given by:
t = 0

if t

t 1
)
t = 0 + (

if t >

(24)

When the effective age at time t, t , is lower than the

location parameter, the failure rate of the component
is constant and the maintenance can delay the start of
the aging process. When the age is greater than the
location parameter, if, mathematically, relation (23)
is still applicable, the meaning is different. We have
two different cases: If M is greater than TM , the
aging of the component is stopped and the failure rate
is again constant, decreased down to the value 0 . If
M is lower than TM , the aging of the component
is not stopped and the failure rate is still increasing but
with a translation in time corresponding to M .
With this model, the corresponding cumulative
distribution at time t, after an intervention at time ts is:

F(t ) =

1 e0 (tts )

1 e0 (tts )( )

t
e0 (tts )( )
s+
( )

if t
if t >
and s+

(25)

if s+ >

where t = s+ + (t ts ) is the effective age of the

component at time t, s+ is the effective age just after
intervention.
3.2.3 Intervention Impact model
When a component ages, part of this aging cannot
be rejuvenated, unless a bigger cost is required. In
practice, this model allows us to envisage that each
maintenance operation may have a different impact.
This impact should not be taken constant. Moreover,
we can assume, as Wu and Clements-Croome 2005,
that M is a random variable embodying the variability
in the resulting state of the component. The value or
the distribution of M must also depend on the age of
the component if the expected effect of a preventive
maintenance is less and less efficient as the number of
maintenance actions undergone by a component gets
higher.
The model also allows considering a limited PM
in order to maintain the initial working conditions
and performances, and a more important PM with a
better efficiency to recover the loss of performances
due to degradation. The latter type of PM should be
performed with a lower periodicity and it entails higher
costs than the limited, usual one. With this point of
view, the decision to replace a component is in competition with the decision to make or not a more important
and costly maintenance with a large effect on the
component state but with a higher risk of failure than
the replacement.
Figures 2 and 3 illustrate the impact of this kind
of mixed maintenance, for a component with an initial failure rate following (24) with 0 = 1/10 u.t.1 ,
= 10 u.t, = 2.5, = 5 u.t.

460

the action undertaken from one intervention to another

one. This variation seems natural when the working
time grows in order to keep the performances of the
system in an acceptable range.
The model is also suitable in the case where the
repairs are not minimal. In this case, each intervention,
including repairs, will affect the effective age.

Figure 2. Comparison of the effective age evolution for two

maintenance policies.

3.2.4 Global performances

When a component is aging, not only its failure rate is
increasing but its global performances are decreasing.
We consider that the component is described by its failure rate, by its consumption rate (working costs by unit
of time) and by its purchase cost. If we consider that
the effective age is a global indicator of the component
state, we can propose to characterize the performances
of the working component by a consumption rate
varying as a function of the value of the effective age.
In the same way as for the failure rate, this consumption rate will increase with the effective age of the
component, embodying the degradation of the performances with aging. The behavior of this aging process
will be dependent on the component. But, as in the case
of the failure rate, it is reasonable to suppose a constant part when the age is not too important and an
increasing part when the age has reached a threshold
value, and a law of the form (24), with appropriate
parameters, can model this behavior.

Figure 3. Comparison of the resulting failure rates for two

different maintenance policies.

In Figure 2 we can compare the effective age

evolution resulting from the ARA1 model on one hand,
and that resulting from the same model with a bigger
impact taking place in the third maintenance intervention on the other hand. Figure 3 compares the resulting
failure rates. We can see that the first intervention
delays the start of the aging, while the second interventions stop the aging process. The bigger impact on
the third maintenance intervention permits to have an
effective age after intervention lower than the location
parameter when the effective age before intervention
was higher than the parameter location. It means that
the actions undertaken during this third intervention
are sufficient to stop the wear out process and rejuvenate the component state to its initial constant failure
rate. The ARA1 (and other models with constant
efficiency) does not consider this possible variation in

SUMMARY AND PERSPECTIVES

After a quick overview of different maintenance

models available in the literature, we propose, in this
paper, to extend some of the considerations of these
models. The first point is to use the effective age of
a component as a global indicator of the components
state. The failure rate at time t will depend directly
on this effective age at time t, but this is also true
for other component performances such as for example the energy consumption of the component. If the
effective age is a global state indicator, we propose that
all the events that act on the component state, as maintenance interventions or working conditions, modify
the value of the effective age.
We think that it is not realistic to consider a strictly
increasing failure rate from the very start of the components lifetime. Indeed, such an assumption leads
to considering that the likelihood to fail of a component starts increasing from the very moment it
is operated. Yet the daily experience of maintaining
equipment leads rather to interpreting maintenance as
a way of preserving as long as possible the present
performances, including the failure rate, of a unit,
until an accumulation of wear-out triggers a rapid
performance degradation, requiring different, more

461

costly maintenance resources. In order to model such a

maintenance impact, we have to consider a bi-Weibull
distribution for the failure times with a constant failure rate and an increasing part, starting after a certain
useful life, which is lengthened via maintenance.
We think also that the maintenance intervention
should not have a constant efficiency but that these
interventions act differently depending on which type
of intervention is actually undertaken on the component. Consequently we propose to relax the common
hypothesis of a maintenance effect proportional to
a given quantity (the maintenance interval or the
effective age before intervention in the most common cases). We think that this formulation allows
us to investigate more realistic issues, such as the
compromise between different interventions (possibly,
including a replacement) leading to different rejuvenation levels of the components effective age but with
different costs and constraints.
In conclusion, the proposed model is thought to
be closer to the reality of the maintenance field. It
allows more flexibility in the modeling but it always
keeps the assumption that the law parameters describing the failure rate are constant in time. Further work
will investigate the possibility that the aging process is
stopped but that it degraded the component, resulting
now in a higher constant failure rate M .
Also events in the environment of the component
can affect its performances, hence the aging of the
component. For example, the failure of another component in the production process can increase the
failure probability of the component without causing
an instantaneous failure. Other events, as for example,
an overvoltage or a difference in the working temperature can cause such a behavior. As we use the effective
age as a global state indicator of the components state,
one could think of modeling the effect of these events
by an increase in the components effective age.

Canfield RV., 1986, Cost optimization of periodic preventive maintenance, IEEE Trans Reliab; 35:7881.
Clavareau J. and Labeau P.E., 2006 Maintenance and
replacement policies under technological obsolescence
Proc. of ESREL06. Estoril (Portugal), 499506.
Doyen L., Gaudoin O., 2004, Classes of imperfect repair
models based on reduction of failure intensity or effective
age, Rel. Engng. Syst. Safety; 84:4556.
Fouathia O., Maun J.C., Labeau P.E., and Wiot D., 2005,
Cost-optimization model for the planning of the renewal,
inspection, and maintenance of substation facilities in
Belgian power transmission system, Proc. of ESREL
2005. Gdansk (Poland); 631637.
Kijima M., Morimura H., Suzuki Y., 1988, Periodical
replacement problem without assuming minimal repair,
Eur. J. Oper. Res; 37:194203.
Kumar D. and Klfesjo B., 1994, Proportional hazard model:
a review, Rel. Engng. Syst. Safety; 44:17788.
Lin D., Zuo MJ, Yam RCM., 2000, General sequential
imperfect preventive maintenance models, Int J Reliab
Qual Saf Eng; 7:25366.
Malik MAK., 1979 Reliable preventive maintenance
scheduling, AIEE Trans; 11:2218.
Martorell S., Sanchez A., Serradell V., 1999, Agedependent reliability model considering effects of maintenance and working conditions , Rel. Engng. Syst. Safety;
64:1931.
Nakagawa T., 1986 Periodic and sequential preventive
maintenance policies, J. Appl. Probab; 23:53642.
Nakagawa T., 1988 Sequential imperfect preventive maintenance policies, IEEE Trans Reliab; 37:2958.
Percy D.F. and Kobbacy K.A.H, 2000, Determining economical maintenance intervals, Int. J. Production Economics; 67:8794.
Samrout M., Chtelet, E., Kouta R., and N. Chebbo,
2008, Optimization of maintenance policy using the
proportional hazard model, Rel. Engng. Syst. Safety;
doi:10.1016/j;ress.2007/12.006.
Wu S., Clements-Croome D., 2005, Preventive maintenance
models with random maintenance quality, Rel. Engng.
Syst. Safety; 90:99105.
Zequeira R.I., Brenguer C., 2006, Periodic imperfect preventive maintenance with two categories of competing
failure modes, Rel. Engng. Syst. Safety; 91:460468.

REFERENCES
Barlow R., Proschan F., 1996, Mathematical theory of
reliability, 1965, SIAM, Philadelphia.
Brown M., Proschan F., 1983, Imperfect repair, J. Appl.
Prob.; 20:851860.

462

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

An imperfect preventive maintenance model with dependent failure modes

I.T. Castro
University of Extremadura, Badajoz, Spain

ABSTRACT: Consider a system subject to two modes of failures: maintainable and non-maintainable. Whenever the system fails, a minimal repair is performed. Preventive maintenances are performed at integer multiples
of a fixed period. The system is replaced when a fixed number of preventive maintenances have been completed.
The preventive maintenance is imperfect and the two failure modes are dependent. The problem is to determine
an optimal length between successive preventive maintenances and the optimal number of preventive maintenances before the system replacement that minimize the expected cost rate. Optimal preventive maintenance
schedules are obtained for non-decreasing failure rates and numerical examples for power law models are given.

INTRODUCTION

The classic preventive maintenance models assume

that the system becomes anew after each preventive maintenance. But actually, the improvement after
each preventive maintenance depends on the system
age as well as the cost and the cumulative number
of preventive maintenance tasks done since the last
replacement. This action is called imperfect preventive maintenance. There is an extensive literature about
imperfect preventive maintenance models, see (BenDaya et al. 2000), (Nakagawa 2005), (Osaki 2002) and
(Pham 2003) for more details.
In this work, the system failures can be grouped
into two modes: maintainable and non-maintainable
analogously to (Lin et al. 2001) and (Zequeira and
Brenguer 2006). A failure rate function is related
to each failure mode. The preventive maintenance is
imperfect because it can only reduce the failure rate of
the maintainable failures but can not change the failure rate of the non-maintainable failures. For example,
the preventive maintenance may include tasks such as
oiling, cleaning, partial system replacement, . . . that
only can improve some failures but they can not eliminate other failures related to the inherent design of the
system.
(Lin et al. 2001) modeled the effects of the preventive maintenance tasks by the reduction of the failure
rate of the maintainable failures and the effective age of
the system using adjustment factors. In a more realistic
approach, (Zequeira and Brenguer 2006) considered
that the failure modes are dependent. This dependency
is expressed in terms of the failure rate: the failure
rate of the maintainable failures depends on the failure rate of the non-maintainable failures. Assuming
that the preventive maintenance is performed at times

kT , where k = 0, 1, 2, . . . and T > 0, the failure

rate of the maintainable failures after each preventive
maintenance is given by
rk,T (t) = r(t kT ) + pk (t)h(t),
kT t < (k + 1)T ,

(1)

where r(t) denotes the failure rate before the first

preventive maintenance, h(t) is the non-maintainable
failure rate and pk (t) is interpreted as a probability
in the following sense: if the system has a nonmaintainable failure at time t then it will undergo also
a maintainable failure with probability pk (t).
Analogously to (Zequeira and Brenguer 2006),
we assume that the two modes of failure are dependent. Preventive maintenance actions are performed
at kT , k = 1, 2, . . . , and they affect the failure
rate of the maintainable failures. We assume that the
occurrence of maintainable failures depends on the
total of non-maintainable failures from the installation of the system. Assuming that the failure rate of
the maintainable failures is zero for a new system,
the maintainable failure rate after the k-th preventive
maintenance action is
r1,k (t) = r1,0 (t kT )aN2 (kT ) ,
kT t < (k + 1)T ,

(2)

where a > 1 and N2 (kT ) represents the number of

non-maintainable failures in [0, kT ] and r1,0 (t) the
maintainable failure rate for a new system. After each
preventive maintenance, the maintainable failure rate
resets to zero. The adjustment factor aN2 (kT ) represents

463

the effect of the wear-out of the system (due to the nonmaintainable failures) in the occurrence of the maintainable failures. Practical applications of this model
are showed in (Zequeira and Brenguer 2006) where
some examples of dependence between maintainable
and non-maintainable failures are explained.
In this work, the preventive maintenance actions
are performed at times kT , k = 1, 2, . . . and the system is replaced whenever it reaches an age of NT after
last renewal. When the system fails, a minimal repair
is performed. Costs are associated with the preventive maintenance actions, with the repairs and with the
replacements. The objective is to determine an optimal length between preventive maintenances and the
optimal number of preventive maintenances between
replacements of the system.
2

FORMULATION

We consider a maintenance model under where corrective and preventive maintenances take place according
to the following scheme.
1. Before the first preventive maintenance, the
maintainable failures arrive according to a nonhomogeneous Poisson process (NHPP) {N1 (t), t
0} with intensity function r1,0 (t) and cumulative
failure intensity function

H1 (t) =

r1,0 (u)du,

t 0.

A maintainable failure is corrected by a minimal

repair with negligible repair time. We assume that
r1,0 (t) is continuous, non-decreasing in t and zero
for a new system, that is, r1,0 (0) = 0.
2. The non-maintainable failures arrive according to
a NHPP {N2 (t), t 0} with intensity function r2 (t)
and cumulative failure intensity function

H2 (t) =

r2 (u)du,

of the maintainable failures in the interval [kT ,

(k + 1)T ), r1,k is given by
r1,k (t) = r1,0 (t kT )aN2 (kT ) ,
where a > 1.
5. The system is replaced at the N -th preventive maintenance after its installation. After a replacement,
the system is as good as new and the replacement
time is negligible.
6. The costs associated with the minimal repairs for
maintainable and non-maintainable failures are C1
and C2 respectively. The cost of each preventive
maintenance is Cm and the replacement cost is
Cr (Cm < Cr ). All costs are positive numbers.
The problem is to determine the optimal length
between preventive maintenances and the total number
of preventive maintenances before the system replacement. The optimization problem is formulated in terms
of the expected cost rate.
From Assumptions 3 and 4, the preventive maintenance tasks are imperfect in a double sense. First, the
preventive maintenance does not affect the failure rate
of the non-maintainable failures. Second, the accumulated wear-out due to the non-maintainable failures
affect the failure rate of the maintainable failures.
This accumulated wearing is not eliminated under preventive maintenance actions and it is showed in the
constant a. If a = 1, non-maintainable failures and
maintainable failures are independent.
From (2), after the successive preventive maintenances, the failure rate of the maintainable failures is
stochastic and we shall use some results of the theory
of the doubly stochastic Poisson processes (DSPP). In
a DSPP, the intensity of the occurrence of the events
is influenced by an external process, called information process, such that the intensity becomes a random
process. An important property of the DSPP is the following. If {N (t), t 0} is a DSPP controlled by the
process (t) one obtains that

t 0.

P[N (t) = n] = E

A non-maintainable failure is corrected by a minimal repair and the repair time is negligible. We
assume that r2 (t) is continuous and non-decreasing
in t.
3. The system is preventively maintained at times kT ,
where k = 1, 2, . . . and T > 0. The preventive
maintenance actions only reduce the failure rate of
the maintainable failures and the failure rate of the
non-maintainable failures remains undisturbed by
the successive preventive maintenances.
4. The non-maintainable failures affect the failure rate
of the maintainable failures in the following way.
Denoting by r1,k , k = 0, 1, 2, . . . the failure rate


1
(t)n e(t) ,
n!

n = 0, 1, 2, . . . .
(3)

From (2), the random measure in t where kT t <

(k + 1)T and k = 0, 1, 2, . . . is given by

k (t) =

r1,0 (u kT )aN2 (kT ) du

kT

= aN2 (kT ) H1 (t kT ),

(4)

where H1 (t) denotes the cumulative failure intensity

function of r1,0 (t).

464

Denoting by N1 (t) the number of maintainable

failures in [0, t] and using (3) and (4), one obtains
that
P[N1 ((k + 1)T ) N1 (kT ) = n]


 N2 (kT )

1 N2 (kT )
n
(a
H1 (T )) exp a
H1 (T ) ,
=E
n!

T
= min {A(T , N ) > Cr Cm } ,
Nopt

where A(T , N ) is given by

A(T , N ) = C1 H1 (T ) NgN (T )

(9)

for z = 0, 1, 2, . . . , the expected number of maintainable failures between the k-th preventive maintenance and the (k + 1)-th preventive maintenance for
k = 0, 1, 2, . . . is given by
E [N1 (kT , (k + 1)T )]
= H1 (T ) exp((a 1)H2 (kT )).

(5)

We denote by C(T , N ) the expected cost rate. From

Assumption 6 and using Eq. (5), one obtains that
C1
C(T , N ) =

H1 (T ) exp((a 1)H2 (kT ))

T
If N exists such that A(T , N ) = Cr Cm , then Nopt
is not unique. Furthermore, if T1 T2 then
T1
T2
Nopt
Nopt
.

Theorem 1.2 Let C(T , N ) be the function given by

(6). When N is fixed, the value of T that minimizes
N
N
C(T , N ) is obtained for T = Topt
where Topt
is the
value that verifies
N
, N ) = Cr + (N 1)Cm ,
B(Topt

(6)

B(T , N ) = C1

N 1


+ C1

gk (T ) r1,0 (T )T H1 (T )

k=0

where the numerator represents the expected cost

between replacements of the system and the denominator the time between successive replacements of the
system.
3

(10)

where B(T , N ) is the function given by

k=0

NT
C2 H2 (NT ) + Cr + (N 1)Cm
,
+
NT


gk (T )

+ C2 [NH2 ((N +1)T )(N +1)H2 (NT )] .

H2 (kT )z
exp{H2 (kT )},
z!

N 1


N
1

k=0

for n = 0, 1, 2, . . . . Furthermore, using

P[N2 (kT ) = z] =

(8)

N 0

N 1



gk (T ) {H1 (T )(a1)kTr2 (kT )}

k=0

+ C2 (NTr2 (NT ) H2 (NT )) .

(11)

Furthermore, if either r1,0 or r2 is unbounded

OPTIMIZATION

The problem is to find the values T and N that minimize the function C(T , N ) given in (6). In other words,
to find the values Topt and Nopt such that
C(Topt , Nopt ) = inf {C(T , N ), T > 0, N = 1, 2, 3, . . . }.
(7)
Theorem 1 and 2 show the optimization problem in
each variable. The proof of these results can be found
in (Castro 2008).
Theorem 1.1 Let C(T , N ) be the function given by
(6). For fixed T > 0, the finite value of N that
T
minimizes C(T , N ) is obtained for N = Nopt
given by

lim r1,0 (t) = ,

lim r2 (t) = ,

N
< .
then Topt

In the problem of optimization given in (7), the

values Topt and Nopt must satisfy the expressions (8)
and (10). In general, one cannot obtain an explicit analytical solution for these equations and they have to be
computed numerically. But, one can reduce the search
of the values Nopt and Topt for a finite set values of the
variable N . For that, we use a similar result to Theorem 11 given in (Zhang and Jardine 1998) p.1118. The
result is the following

465

Lemma 1.1 We denote by T the following

expression
T =

the first preventive maintenance are given by

r1,0 (t) = 6.33t 0.2 ,

Cm
,
1
C(Topt
, 1)

r2 (t) = 0.1739t 0.2 ,

t 0,

(12)
and,

T
be the value of N that optimizes C(N , T ).
and let Nopt
Assuming that either r1,0 or r2 is unbounded, then the
problem of optimization of C(T , N ) given in (7) has
finite optimal solutions and


C(Topt , Nopt ) = min min C(T , N ) .
(13)
T
1N Nopt

T >0

Proof.
The proof of this result can be found in
(Zhang and Jardine 1998). Note that for the model
presented in this paper, from Theorem 1 if T1 T2
T1
T2
then Nopt
Nopt
. This condition is necessary for the
proof of the result.
Remark 1.1 Analogously to (Zhang and Jardine
1
1998), C(Topt
, 1) in (12) may be replaced by any
i
, i) for i {2, 3, . . . }.
C(Topt
4

NUMERICAL EXAMPLES

We assume that the intensity functions of the processes

{N1 (t), t 0} and {N2 (t), t 0} follow a power law
model with non-decreasing failure rates, that is,
ri (t) = i i (i t)i 1 ,

t 0,

i = 1, 2,

where i > 1. Let 1 = 4, 1 = 1.2, 2 = 0.2,

2 = 1.2 be the parameters of the failure rates. Consequently the failure intensities of the processes before

H1 (T ) = (4T )1.2 H2 (T ) = (0.2T )1.2 ,

T 0,

that is, before the first preventive maintenance, the

expected number of maintainable failures is less than
the expected number of non-maintainable failures.
Furthermore, the following costs are associated to the
minimal repairs, to the inspection tasks and to the
system replacement.
C1 = 5,

C2 = 7.5,

Cr = 200,

Cm = 75.

Finally, let a = 1.02 be the adjustment factor.

Figure 1 shows the graphics of C(T , N ) given in (6)
for different values of N . These values of N are
N = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16, 20, 30].
To analyze the values Topt and Nopt that verify (7),
Figure 2 shows an enlargement of Figure 1 to clarify
the procedure of finding the values Topt and Nopt .
By inspection over these values, one obtains that
the optimal values for this range of values are Topt =
3.575 and Nopt = 10 with an expected cost rate of
C(3.575, 10) = 51.281.
Using Lemma 1, one can reduce the search of the
optimal values T and N verifying (7) in a limited range
of values of N . We shall follow the following steps

54

200

N=4

180
N=30

N=20

N=16 N=14

N=11N=10 N=9

N=8

53.5

N=12

160

N=7

140

53

N=6

N=1

C(T,N)

C(T,N)

120
N=5

100

N=4

52.5
N=6

N=3

80

N=2

N=7

52

60
N=8

40

N=9

51.5

20
N=10

51
0

10

15

20

25

30

35

40

2.5

3.5

4.5

5.5

Figure 1. Expected cost C(T , N ) versus T for different

values of N .

Figure 2. Expected cost C(T , N ) versus T for different

values of N .

466

12
1. To find the value Topt
that verifies

50

12
min C(T , 12)) = C(Topt
, 12).
T >0

0
A(0.4869,N) Cr+Cm

Using Theorem 2 and the function B(T , 12) given

12
in (11) Topt
verifies the equation
12
B(Topt
, 12) Cr 11Cm = 0.

Using a root search algorithm one obtains that

12
Topt
= 3.2256 and C(3.2256, 12) = 51.3431.
2. Calculate
T =

Cm
25
= 0.4869.
=
12
51.3431
C(Topt
, 12)

0.4869
that verifies
3. To find the value Nopt

Figure 4.

N 1

For that, from Theorem 1 and Figure 3, we obtain

0.4869
= 86.
by inspection that Nopt
4. Finally, we have to find Topt and Nopt that verify

30

40

50

60

70

80

90

min {min C(T , N )} = C(Topt , Nopt ).

N
, N ) for difFigure 4 shows the values of C(Topt
N
ferent values of N . The values Topt are obtained
using a root search algorithm for the function

68
66
64
62
60
58
56
54

Function C(TN , N ) versus N .

B(T , N ) given by (11) and for the different values of N . By inspection, one obtains that the
optimal values for the optimization problem are
Topt = 3.575 and Nopt = 10 with an expected cost
rate of C(3.575, 10) = 51.281.

1N 86 T >0

20

0.4869
min{C(0.4869, N )} = C(0.4869, Nopt
).

C(Topt , N)

10

CONCLUSIONS

In a system with two modes of failures and successive preventive maintenance actions, we have studied the problem of finding the optimal length T
between successive preventive maintenances and the
optimal number of preventive maintenances N 1
before the total replacement of the system. The two
modes of failures are dependent and their classification depends on the reduction in the failure rate
after the preventive maintenance actions. For fixed
T , an optimal finite number of preventive maintenances before the total replacement of the system
is obtained. In the same way, when N is fixed, the
optimal length between successive preventive maintenances is obtained. If the failure rates are unbounded,
this value is finite. To analyze the optimization
problem in two variables, except some special cases,
we have to use numerical results but one can reduce
the search of the optimal values to a finite case.

52
50

ACKNOWLEDGEMENTS
10

20

30

40

50

60

70

80

90

N
Figure 3.

Function A(0.4869, N ) Cr + Cm versus N .

100

This research was supported by the Ministerio de

Educacin y Ciencia, Spain, under grant MTM200601973.

467

REFERENCES
Ben-Daya, M., S. Duffuaa, and A. Raouf (2000). Maintenance, Modeling and Optimization. Kluwer Academic
Publisher.
Castro, I.T. (2008). A model of imperfect preventive maintenance with dependent failure modes. European Journal
of Operational Research to Appear.
Lin, D., J. Ming, and R. Yam (2001). Sequential imperfect
preventive maintenance models with two categories of
failure modes. Naval Research Logistics 48, 173178.
Nakagawa, T. (2005). Maintenance Theory of Reliability.
Springer-Verlag, London.

Osaki, S. (2002). Stochastic Models in Reliability and

Maintenance. Springer-Verlag, Berlin.
Pham, H. (2003). Handbook of Reliability Engineering.
Springer-Verlag, London.
Zequeira, R. and C. Brenguer (2006). Periodic imperfect
preventive maintenance with two categories of competing
failure modes. Reliability Engineering and System Safety
91, 460468.
Zhang, F. and A. Jardine (1998). Optimal maintenance models with minimal repair, periodic overhaul and complete
renewal. IIE Transactions 30, 11091119.

468

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Condition-based maintenance approaches for deteriorating system

influenced by environmental conditions
E. Deloux & B. Castanier
IRCCyN/Ecole des Mines de Nantes, Nantes, France

C. Brenguer
Universit de Technologie de Troyes/CNRS, Troyes, France

ABSTRACT: The paper deals with the maintenance optimization of a system subject to a stressful environment.
The behavior of system deterioration can be modified by the environment. Maintenance strategies, based not only
on the stationary deterioration mode but also on the stress state, are proposed to inspect and replace the system
in order to minimize the long-run maintenance cost per unit of time. Numerical experiments are conducted
to compare their performance with classical approaches and thus highlight the economical benefits of our
strategies.

INTRODUCTION

Maintenance is an essential activity in any production

facility. In this context, numerous research works have
provided multiple relevant approaches to optimize
maintenance decision based on different characteristics of the system in order to reduce the associated costs and in the same time to maximize the
availability and the safety of the considered system
(Rausand and Hoyland 2004). Condition-Based Maintenance (CBM) approaches (Wang 2002), where the
decision is directly driven by a measurable condition variable which reflects the deterioration level of
the system, have proved their efficiency compared
to classical time- or age-based policies (Gertsbakh
2000) in terms of economical benefits and also in
terms of system safety performance. Most of the
maintenance decision framework does not take into
account some potential variations in the deterioration mode. The classical approaches in maintenance
optimization do not explicit the relationship between
the system performance and the associated operating
environment.
However, in most practical applications items or
systems operate in heterogeneous environments and
loads. The stressful environment and other dynamically changing environmental factors may influence
the system failure rate or at least the degradation
modes. Different deterioration models taking into
account the stressful environment are provided in the
reliability field (Singpurwalla 1995; Lehmann 2006)
but none of them are developed in maintenance area.

One way of capturing the effect of a random environment on an items lifelength is to make its failure rate
function (be it predictive or model) a stochastic process. Ones of the most important approaches are the
well-known proportional hazard rate (Cox 1972) or the
cumulative hazard rate (Singpurwalla 2006; Singpurwalla 1995) which consist to model the effect of the
environment with the introduction of covariates in the
hazard function. However, the estimation of the different parameters are quite complex for systems even
if efficient methodologies exists for the component
level, see Accelerated Life Testing (Bagdonavicius
and Nikulin 2000; Lehmann 2006). Other reliability works on non-parametric methods (Singpurwalla
2006) allow to specify the likelihood by taking into
account all the information available. Nevertheless,
the direct application of these models for maintenance
optimization leads to very complex and inextricable
mathematical equations.
In this context, the objective of this paper is to
develop a maintenance decision framework for a continuous deteriorating system which is influenced by
environmental condition. We consider a system subject to a continuous random deterioration such as
corrosion, erosion, cracks . . . (van Noortwijk 2007).
The deterioration process is influenced by environmental conditions such as, e.g., humidity rate, temperature or vibration levels. First, a specific deterioration
model based on the nominal deterioration characteristics (without any stress), the random evolution of the
environment and the impact of the stress on the system
are developed. Then, different maintenance strategies

469

which integrate the deterioration level but also the

stress information are presented. Finally, numerical
results based on generic data are conducted to compare
these different approaches with a classical approach
and thus will highlight the economical and safety
benefits of our approach.
2

DESCRIPTION OF THE FAILURE PROCESS

We consider a single-unit system subject to one failure

mechanism evolving in a stressful environment. The
failure process results from an excessive deterioration
level. This section is devoted to describe the system
failure degradation, the evolution of the stress and the
relationship between these two processes.
2.1

developments. The gamma process is parameterized

by two parameters and which can be estimated
from the deterioration data. Gamma processes have
received recently a tremendous amount of attention
in the reliability and maintainability literature as
means to model degradation of many civil engineering
structures under uncertainty (vanNoortwijk 2007).

Stochastic deterioration model

The condition of the system at time t can be summarized by a scalar aging variable Xt (Deloux et al.,
2008; Grall et al., 2006; Grall et al., 2002) which
varies increasingly as the system deteriorates. Xt can
be e.g. the measure of a physical parameter linked to
the resistance of a structure (length of a crack, . . . ).
The initial state corresponds to a perfect working state,
i.e. X0 = 0. The system fails when the aging variable is
greater than a predetermined threshold L. The threshold L can be seen as a deterioration level which must
not be exceeded for economical or security reasons.
The deterioration process after a time t is independent
of the deterioration before this time. In this paper, it
is assumed that (Xt )(t0) is a gamma process and the
increment of (Xt )(t0) on a unit of length t , Xt , follows a gamma probability distribution function with
shape parameter t and scale parameter :
ft, (x) =

1
t xt1 ex Ix0
(t)

2.2 Stress process

It is assumed that the system is subject to an environmental stress (e.g. temperature, vibrations, . . . ). It
is considered that the environment condition at time
t can be summarized with a single binary covariate
(Yt )(t0) . It is assumed that Yt is an indicator of the
environment evolution, i.e. it does not model the environment but only indicate if the system is stress or not
(Yt = 1 if the system is stressed and 0 otherwise). The
time intervals between successive state changes are
exponentially distributed with parameter 0 (respectively 1 for the transit to the non-stressed state at
the stress state (respectively stressed to non-stressed
state). At t = 0 the system is in the stressed state with
0
a probability r = 1+
.
0
2.3 Impact of the stress process on the system
deterioration
The system evolves in a stressful environment, it
is considered that the deterioration behavior can be

X(t)

(1)

Thus, the non-decreasing mean degradation can be

determined as:
E(Xt ) = m(t) = t

(2)

Y(t)

and the variance is proportional to the mean degradation:

Var(Xt ) =

m(t)
2

(3)

Note that the gamma process is a positive process

with independent increment, hence it is sensible to
use this process to describe the deterioration (vanNoortwijk 2007). Another interest of the gamma
process is the existence of an explicit probability distribution function which permits feasible mathematical

0
Non-stressed
system

Stressed
system

Figure 1. Evolution of the deterioration process impacted

by the stress process.

470

impacted by this environment. The covariate influence can be reduced to a log-linear regression model
if Yt = y; Xy (t) (0 e y t, ) (Bagdonavicius and
Nikulin 2000; Lehmann 2006) where measure the
influence of the covariate on the degradation process.
Thus, it is assumed that the system is subject to an
increase in the deterioration speed while it is under
stress (i.e. while Yt = 1), the system deteriorates
according to its nominal mode while it is not stressed
(while Yt = 0). The parameters of the degradation process when the system is non-stressed are 0 t (0 = )
and and when the system is under stress 1 t and
with 1 = 0 e . In average the shape parameter
is 0 (1 + r (e 1)). 0 and can be obtained by
using the maximum likelihood estimation. can be
assimilated to a classical accelerator factor and can be
obtained with accelerated life testing method.
The Figure 1 sketches the evolution of the stress
process and its influence on the system deterioration.
3

DEFINITION AND EVALUATION OF THE

MAINTENANCE POLICY

This section presents the maintenance decision framework. First, the structure of the maintenance policy is presented to define when an inspection or a
replacement should be implemented. The mathematical expressions of the associated long-run maintenance cost per unit of time are developed to optimize
the maintenance decision regarding the system state
behavior.
3.1

Structure of the maintenance policy

The cumulative deterioration level Xt can be observed

only through costly inspections. Let cix be the unitary inspection cost. Even if non-periodic inspection
strategies are optimal (Castanier et al., 2003), a periodic strategy is first proposed. The benefit of such a
choice is a reduced number of the decision parameters,
only the inspection period , and an easier implementation of the approach in an industrial context. This
inspection is assumed to be perfect in the sense that it
reveals the exact deterioration level Xt .
A replacement can take place to renew the system when it is failed (corrective replacement) or to
prevent the failure (preventive replacement). A corrective replacement is performed when the system is
observed in the failed state during an inspection on
Xt . We assume the unitary cost of a corrective replacement cc is composed by all the direct and indirect costs
incurred by this maintenance action. Only the unavailability cost cu per unit of time the system is failed
has to be added to cc . The decision rule for a preventive replacement is the classical control limit rule: if
is the preventive replacement threshold, a preventive

replacement is performed during the inspection on Xt

if the deterioration level belongs to the interval ( , L).
Let cp be the preventive replacement cost (cp < cc ).
This maintenance policy is denoted Policy 0 hereafter.
3.2 Cost-based criterion for maintenance
performance evaluation
The maintenance decision parameters which should be
optimized in order to minimize the long-run maintenance cost are:
the inspection period which allows balancing the
cumulative inspection cost, earlier detection and
prevention of a failure;
the preventive maintenance threshold which
reduces cost by the prevention of a failure.
An illustration of the maintenance decision is presented in Figure 2.
The degradation of the unmaintained system state
is described by the stochastic process (Xt )t0 . Let in
t )t0 describes the evoluthe sequel the process (X
tion of the maintained system state. It can be analyzed through its regenerative characteristics: after a
complete replacement of the system (all the system
components are simultaneously replaced), it is in the
as good as new initial state and its future evolution does not depend any more on the past. These
complete system replacement times are regenerations
points for the process describing the evolution of the
global maintained system state. Thanks to the renewal
property, we can limit the study of the process to a
renewal cycle, which significantly reduces the complexity of the analysis. The renewal-reward theorem
(Asmussen 1987) implies that the cost function equals
the expected costs per cycle divided by the expected
length of a cycle:
C ( , ) = lim

E(C(S))
C(t)
=
t
E(S)

(4)

X(t)
corrective
replacement area
preventive
replacement area

Failure due to an
excessive deterioration
level
L

Figure 2. Evolution of the deterioration process and the

stress process when the system is maintained.

471

cu E(Du (S))
E(S)

25

30

(5)

20

Ins pection period

cix E(S)
+ cp + (cc cp )P(Xs > L)
E(S)

15

C ( , ) =

35

where E(W ) is the expected value of the random variable and S is the length of a regenerative cycle. The cost
C(S) is composed of the inspections, replacements and
unavailability costs and can be written:

10

where Du (t) represents the unavailability time before t.

The computation of the long-run maintenance cost
per unit of time is presented in Appendix A, it requires
the evaluation of the reliability function of the maintained deterioration process in the different modes:
stressed or not stressed.

0.0

0.2

0.4

0.6

0.8

1.0

Figure 3. Evolution of the inspection time depending on the

time elapsed in the stress state r .

3.3 Extension of the maintenance policy with the

introduction of the knowledge on the stress
Previously, only the information given by the deterioration level has been used, Policy 0 is based on the
stationary deterioration mode but it can be useful to
adapt the decision for inspection and replacement with
the observed time elapsed in the different operating
conditions. Policy 0 is well-adapted when only an a
priori knowledge on the stress process is available. In
such a way, the knowledge of the exact time elapsed in
the different operating conditions should allow to fit a
better estimation of the present state. We will develop
in the next paragraph two strategies (hereafter denoted
Policy 1 and Policy 2) which offer the opportunity to
adapt the decision (inspection time and replacement
threshold) as a function of the effective time elapsed
in an operating condition (i.e. time elapsed in the stress
condition).
The decision framework for Policy 1 is defined as
follows:
let r(t) be the proportion of time elapsed in the
stressed condition since the last replacement. Let
introduce (r(t) , r(t) ) which are the optimized decision parameters for Policy 0 when r(t) = r .
For example, we can obtain similar curves as
presented in Figure 3 for all values of r(t).
An inspection is performed at time t:
if X (t) r(t) , a replacement is performed and
the system is as good as new. The next inspection is planned at time t + 0 if the system is
non-stressed or at t + 1 if the system is stressed;
if X (t) < r(t) , the system is left as it is and the
next inspection is planned at time t + r(t) .
The environment state is continuously monitored.
r(t) is continuously evaluated and allow modifications in the decision parameters. the next inspection
(r(t) , r(t) ) are determined in function of the new

r(t). If the new inspection period leads to an inspection time lower than the present time, the system is
inspected immediately with the following decision
(t, r(t) ).
Even if Policy 1 allows to minimize the maintenance
cost criterion, it is not easily to implement this policy
in industry, thus we propose another policy, Policy 2.
Policy 2 is the discrete case of Policy 1. In order to
reduce the number of updates of the decision parameters, we propose to introduce thresholds (l1 , . . . , ln ) on
r(t). The decision framework for Policy 2 is defined
as follows:
if r(t) = 0, the decision parameters associated are
(0 , 0 ).
if r(t) = 1, the decision parameters associated are
(1 , 1 ).
if r(t) (lk , lk+1 ), the decision parameters associated are (lk+1 , lk+1 ) if t < k+1 and (t, lk+1 )
otherwise.
Figure 4 illustrates the updating of the decision
parameters depending on r(t) and the number of
thresholds. In this example, only one threshold on
r(t) is considered: l1 . A t = 0, the system is
assumed to be in a new state and the system is stressed
(Y (t) = 1), thus the decision parameters are (1 , 1 )
(see T 1 in the Figure 4). These decision parameters remain unchanged until r(t) l1 . As soon as
r(t) > l1 , see T2 on Figure 4, decision parameters
are updated to (l1 , l1 ). At T 3, r(t) returns above l1
and the new decision parameters (1 , 1 ) leads to a
passed time, thus the deterioration level is inspected
immediately (cf. the two arrows from the curve r(t)).

472

estimate in order to know the time elapsed

in the failed state Du . In this case the cost is
cix + cc + cu .Du
if the cumulative deterioration is lower than
the failure threshold but upper than the corresponding preventive threshold a preventive
replacement is implemented with the following
cost: cix + cp .

Proportion of time
elapsed in the stress
state: r(t)
1

l1

End
Step 2: estimation of the cost criterion: the long
run maintenance cost corresponds to the mean of
the Nh costs.

0
Y(t)

0
0

l1
l1

1
0

T1

T2

T3

Figure 4. Evolution of the inspection time depending on the

state of the stress with Policy 2.

The objective of Policy 2 is to optimize the number

of thresholds in order to balance the number of decision parameters change and the optimal cost given by
policy 1. In this paper, the number of thresholds is
not optimized, we present only the performance of
this policy in function of the number of the number
of thresholds.
The following iteration algorithm is used for optimizing Policy 2:
Step 0: initialization of the maintenance policy:
choice of the number of thresholds: n;
determination of the n policies (n , n ) optimized
with policy 0;
For i = 1:Nh (1000)
Step 1: estimation of the maintenance cost in one
cycle:

This section is devoted to compare the economic

performance of the three proposed policies. The
numerical results are provided here by the use of R
(www.r-project.org) software, specific programs are
developed to numerically evaluate each expectation
in Equation 5 in case of Policy 0 and the classical
numerical gradient procedure provided by the R software is used to determine the optimal cost and thus the
optimized maintenance decision parameters. A Monte
Carlo approach is used in case of the non-periodic
inspection strategies described in section 3.2.
4.1 Economic performance of Policy 1 compared to
Policy 0
The curves in figure 5 are the respective representation
of the optimized cost criterion for the Policy 0 and 1.
They are obtained when the mean proportion of time
elapsed in the non-stressed state, r varies from 0 to 1.

Policy 0
Policy 1

0.55

Cos t

0.65

Step 1.1: initialization of the stress state Y0 and

the deterioration level X0 = 0;
Step 1.2: simulation of all the dates of stress state
change with the exponential law;
Step 1.3: determination of all the inspection periods function of the time elapsed in the stress
state and the n policies optimized according to
the number of thresholds;
Step 1.4: simulation of the successive deterioration increments between two inspections function
of the dates of stress state change and the stress
levels until a repair;
if the cumulative deterioration is lower than the
corresponding preventive threshold, the system is left as it is until the next inspection, this
inspection incurs a cost cix ;
if the cumulative deterioration is greater than
the failure threshold L, a corrective replacement is implemented and the failure time is

NUMERICAL RESULTS

0.60

1
1

0.0

0.2

0.4

0.6

0.8

1.0

Figure 5. Variation of the optimized cost for Policy 0

(dashed curve) and of the estimation of the optimized cost
for Policy 1 when r (0 , 1 ) varies from 0 to 1.

473

0.625

Policy 2 would have better results than Policy 1

if we introduced a cost for each change of decision
parameters. Moreover, Policy 2 is easier to implement
than Policy 1 in an industrial context.

0.620

0.615

Co s t

Policy 0
Policy 1
Policy 2

number of thres holds

Figure 6. Optimized cost variation when the mean proportion of the time elapsed in the stress state varies.

These curves are obtained for arbitrarily fixed maintenance data and operation costs: 0 = 0.1, = 7,
= 0.3, L = 2, cix = 5, cc = 100, cp = 30, cu = 25.
Figure 5 illustrates the benefit of the non-periodic
inspection strategy. In fact, the non-periodic inspection strategy (full line) is always the policy which
minimizes the long-run maintenance cost per unit of
time. When, in average, the proportion of time elapsed
in the non-stressed state, r , tends to 0 (never stressed)
or 1 (always stressed), Policy 1 tends to Policy 0. The
economic benefit of Policy 1 compared to Policy 0
varies from 0% (when r = 0 and r = 1) to 5%.
Finally, the non-periodic scheme takes the advantage here to propose an adaptive inspection interval to
the real proportion of time elapsed in the stress state
which allow significantly better performance than the
periodic scheme.
4.2

CONCLUSION

In this paper different maintenance decision frameworks for a continuous deteriorating system which
evolves in a stressful environment have been proposed.
The relationship between the system performance and
the associated operating environment has been respectively modeled as an accelerator factor for degradation
and as a binary variable. A cost criterion has been
numerically evaluated to highlight the performance
of the different maintenance strategies and the benefits to consider the opportunity to update the decision
according to the history of the system.
Even if the last proposed structure for maintenance
decision framework has shown interesting performance, a lot of research remains to be done. First, the
mathematical model of the cost criterion in this case
has to be evaluated. A sensitivity analysis when maintenance data varies should be performed. Moreover,
during an inspection, the maintenance decisions for
the next inspection are based only on the time elapsed
by the system in the failed state, but it could be interesting to use the information given by the degradation
level as well and so propose a mix of condition - and
stress - based framework for inspections.
APPENDIX A
Let f (x, t) be the probability density function of the
deterioration increment at time t of a system subject
to aging and stress:
f (x, t) =

Economic performance of Policy 2

The curves in figure 6 are the respective representation of the optimized cost criterion for the Policy 0
and 2. They are obtained when the number of thresholds in the case of Policy 2 varies from 1 to 10. The
curve for Policy 2 is the smoothed estimators of the
respective costs obtained by simulations. These curves
are obtained for arbitrarily fixed maintenance data and
operations costs to following values: 0 = 0.1, = 7,
= 0.3, L = 2, cix = 5, cc = 100, cp = 30,
cu = 25, r = 0.4. Figure 6 illustrates the benefit
of the non-periodic strategy compared to the periodic strategy. The economic performance of Policy 2
increases with the number of thresholds and varies
between 0.012% and 2.5% compared to Policy 1.
When the number of thresholds increases Policy 2
tends to Policy 1.

xt1
e
(t)

(6)

with = 0 (e + r (1 e )) The different expectations in Equation 5 can be evaluated by a function of

the reliability of the maintained system at steady-state,
Rm (t). In order to determine Rm (t) for all t (0, S),
two cases are identified:
Case 1: no inspection on the system has been made
before t (t < ). The probability density function
for the system to be in a good state is only a function of the observed deterioration level x (0, L) at
time t. Hence for t < , we have:

474

Rm (t) = P(X (t) < L)

 L
=
f (x, t)dx
0

(7)

 Case 2: at least one inspection has been performed

at t (t ). The probability density function for the
system to have been never replaced is a function of
the observed deterioration level x (0, ) during
the last inspection at time ([t/ ]. ), (where [.] is the
integer part function) and of the deterioration level
reached since the last inspection y (x, L). Hence
for t , we have:
Rm (t) = P(X (t) < L|X ([t/ ]. ) < )
  Lx
=
f (x, [t/ ]. )f (y, t)dydx
0

(8)

E(S), the expected length of a regenerative cycle

can be expressed by the two following scenarios:
the cycle ends by a corrective replacement
a cycle ends by a preventive replacement
and is given by:

E(S) =

0

+

x (Rm ((x 1) ) Rm (x ))dx



Scenario1

0

x (Rm (x1) P(X (x ) < |X (x 1) < ))



Scenario2

(9)
with
P(X (x ) < |X (x 1) < ) =
  y
f (y, (x 1) )f (z, x )dzdy
0

(10)

P(Xs > L), the probability of a corrective replacement on a cycle S is given by:

x (Rm ((x 1) ) Rm (x ))dx (11)
P(Xs > L) =

REFERENCES
Asmussen, S. (1987). Applied Probability and Queues, Wiley
Series in Probability and Mathematical Statistics. Wiley.
Bagdonavicius, V. and M. Nikulin (2000). Estimation in
Degradation Models with Explanatory Variables. Lifetime
Data Analysis 7, 85103.
Castanier, B., C. Brenguer, and A. Grall (2003). A sequential condition-based repair/replacement policy with nonperiodic inspections for a system subject to continuous
wear. Applied Stochastic Models in Business and Industry 19(4), 327347.
Cox, D. (1972). Regression models and life tables. Journal
of the Royal Statistics B(34), 187202.
Deloux, E., B. Castanier, and B. C. (2008). Maintenance policy for a non-stationary deteriorating system. In Annual
Reliability and Maintainability Symposium Proceedings
2008 (RAMS 2008), Las Vegas, USA, January 2831.
Gertsbakh, I. (2000). Reliability Theory With Applications to
Preventive Maintenance. Springer.
Grall, A., L. Dieulle, C. Brenguer, and M. Roussignol (2002). Continuous-Time Predictive-Maintenance
Scheduling for a Deteriorating System. IEEE Transactions on Reliability 51(2), 141150.
Grall, A., L. Dieulle, C. Brenguer, and M. Roussignol (2006). Asymptotic failure rate of a continuously
monitored system. Reliability Engineering and System
Safety 91(2), 126130.
Lehmann, A. (2006). Joint modeling of degradation and
failure time data. In Degradation, Damage, Fatigue and
Accelerated Life Models In Reliability Testing, ALT2006,
Angers, France, pp. 2632.
Rausand, M. and A. Hoyland (2004). System Reliability
Theory-Models, Statistical Methods, and Applications
(Second ed.). Wiley.
Singpurwalla, N. (1995). Survival in Dynamic Environments.
Statisticals Science 10(1), 86103.
Singpurwalla, N. (2006). Reliability and risk A Bayesian
Perspective. Wiley.
van Noortwijk, J. (2007). A survey of the application of
gamma processes in maintenance. Reliability Engineering
and System Safety doi:10.1016/j.ress.2007.03.019.
Wang, H. (2002). A Survey of Maintenance Policies of
Deteriorating Systems. European Journal of Operational
Research 139, 469489.

E(Du (S)), the expected value of the cumulative

unavailability duration before S is given by:

x (Rm ((x 1) ) Rm (x ))dx (12)
E(Du (S)) =
0

475

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Condition-based maintenance by particle filtering

F. Cadini, E. Zio & D. Avram
Politecnico di MilanoDipartimento di energia, Italy

ABSTRACT: This paper presents the application of the particle filtering method, a model-based Monte Carlo
method, for estimating the failure probability of a component subject to degradation based on a set of observations.
The estimation is embedded within a scheme of condition-based maintenance of a component subject to fatigue
crack growth.

INTRODUCTION

When the health state of a component can be monitored, condition-based policies can be devised to
maintain it dynamically on the basis of the observed
conditions (Marseguerra et al., 2002 and Christer et al.,
1997). However, often in practice the degradation
state of a component can not be observed directly,
but rather must be inferred from some observations,
usually affected by noise and disturbances.
The soundest approaches to the estimation of the
state of a dynamic system or component build a posterior distribution of the unknown degradation states
by combining a prior distribution with the likelihood
of the observations actually collected (Doucet, 1998
and Doucet et al., 2001). In this Bayesian setting,
the estimation method most frequently used in practice is the Kalman filter, which is optimal for linear
state space models and independent, additive Gaussian noises (Anderson & Moore, 1979). In this case,
the posterior distributions are also Gaussian and can
be computed exactly, without approximations.
In practice, however, the degradation process of
many systems and components is non-linear and the
associated noises are non-Gaussian. For these cases,
approximate methods, e.g. analytical approximations
of the extended Kalman (EKF) and Gaussian-sum filters and numerical approximations of the grid-based
filters can be used (Anderson & Moore, 1979).
Alternatively, one may resort to particle filtering
methods, which approximate the continuous distributions of interest by a discrete set of weighed particles
representing random trajectories of system evolution
in the state space and whose weights are estimates of
the probabilities of the trajectories (Kitagawa 1987 and
Djuric et al., 2003 and Doucet et al., 2000).
In this paper, the particle filtering method for
state estimation is embedded within a condition-based
maintenance scheme of a component subject to fatigue

crack growth (Rocha & Schueller 1996 and Myotyri

et al., 2006 and Pulkkinen 1991). The crack propagation is modelled as a non-linear Markov process with
non-additive Gaussian noise (Myotyri et al., 2006);
the condition-based maintenance strategy considered
amounts to replacing the component at an optimal time
determined on the basis of a cost model of literature
(Christer et al., 1997); the component remaining lifetime, upon which the optimal replacement decision is
based, is predicted by particle filtering.
The paper is organized as follows. In Section 2, the
Bayesian formulation of model-based state estimation
and the particle filtering method are recalled. Section 3
presents the application of the method to the prediction
of the optimal time of replacement of a component
subject to fatigue. Finally, some conclusions are drawn
in Section 4.

MODEL-BASED STATE ESTIMATION

AND PARTICLE FILTERING

Let us consider a system whose state at the discrete

time step tk = kt is represented by the vector xk .
The model (transition or state equation) describing the
evolution of the system state is an unobserved (hidden)
Markov process of order one:
xk = f (xk1 , k1 )

(1)

where fk : Rnx Rn Rnx is possibly non-linear

and {k , k N } is an independent identically distributed (i.i.d.) state noise vector sequence of known
distribution.
The transition probability distribution p(xk | xk1 ) is
defined by the system equation (1) and the known distribution of the noise vector k . The initial distribution
of the system state p(x0 ) is assumed known.

477

A sequence of measurements {zk , k N } is

assumed to be collected at the successive time steps
tk . The sequence of measurement values is described
by the measurement (observation) equation:
zk = hk (xk , k)

(2)

where hk : Rnx Rn Rnx is possibly non-linear

and {k , k N } is an i.i.d. measurement noise vector
sequence of known distribution. The measurements
{zk , k N } are, thus, assumed to be conditionally
independent given the state process {xk , k N }.
Within a Bayesian framework, the filtered posterior
distribution p(xk | z0:k ) can be recursively computed in
two stages: prediction and update.
Given the probability distribution p(xk1 | z0:k1 )
at time k1, the prediction stage involves using the
system model (1) to obtain the prior probability distribution of the system state xk at time k via the
Chapman-Kolmogorov equation:

p(xk | z0:k1) =

p(xk | xk1)p(xk1 | z0:k1)dxk1

p(0:k | z0:k) (0:k x0:k) d0:k

p (x0:k | z0:k) =

Ns
1 
i
(x0:k x0:k
)
Ns i=1

p (x0:k | z0:k) =

(0:k | z0:k)

p (0:k | z0:k)
(0:k | z0:k)

(0:k x0:k) d0:k

where the Markovian assumption underpinning the
system model (1) has been used.
At time k, a new measurement zk is collected and
used to update the prior distribution via Bayes rule, so
as to obtain the required posterior distribution of the
current state xk (Arulampalam 2002):
p(xk | z0:k1) p(zk | xk)
p(xk | z0:k) =
p(zk | z0:k1)

p (x0:k | z0:k ) =

NS
1 
i
wi (x0:k x0:k
)
Ns i=1 k

(9)

(4)
where:
wki =


p(xk | z0:k1)p(zk | xk)dxk

(8)

and an unbiased estimate is obtained by (Doucet et al.,

2001 and Arulampalam 2002):

where the normalizing constant is

p(zk | z0:k1) =

(7)

i
, i =1, 2, . . . , Ns is a set of independent
where x0:k
random samples drawn from p(x0:k | z0:k ).
Since, in practice, it is usually not possible to
sample efficiently from the true posterior distribution p(x0:k | z0:k ), importance sampling is used, i.e.
i
the state sequences x0:k
are drawn from an arbitrarily chosen distribution (x0:k | z0:k ), called importance
function (Kalos and Whitlock 1986). The probability
p(x0:k | z0:k ) is written as:


(3)

(6)

and assuming that the true posterior probability

p(x0:k | z0:k ) is known and can be sampled, an estimate
of (6) is given by (Kalos and Whitlock 1986):

p(xk | xk1 | z0:k1)p(xk1 | z0:k1)dxk1




p(x0:k | z0:k) =

(5)

The recurrence relations (3) and (4) form the basis

for the exact Bayesian solution. Unfortunately, except
for a few cases, including linear Gaussian state space
models (Kalman filter) and hidden finite-state space
Markov chains (Wohnam filter), it is not possible to
evaluate analytically these distributions, since they
require the evaluation of complex high-dimensional
integrals.
This problem can be circumvented by resorting to
Monte Carlo sampling methods (Doucet et al., 2001
and Doucet et al., 2000 and Pulkkinen 1991 and Seong
2002).
Writing the posterior probability p(x0:k | z0:k ) of
the entire state sequence x0:k given the measurement
vector z0:k as:

i
i
)p(x0:k
)
p(z0:k | x0:k
i
p(z0:k )(x0:k | z0:k )

(10)

is the importance weight associated to the state

i
sequence x0:k
, i = 1, 2, . . . , Ns , sampled from
i
) is the likelihood of the
(x0:k | z0:k ) and p(z0:k | x0:k
observation sequence.
Computing the weight requires knowledge of
the p (z0:k ) = p (z0:k | x0:k ) p (x0:k ) dx0:k , normalizing constant, which cannot typically be expressed in
closed form. It can be shown that, to overcome this
problem, an estimate of the posterior probability distribution p(x0:k | z0:k ) may be computed as (Doucet et al.,
2001 and Arulampalam 2002):

p (x0:k | z0:k) =

478

Ns

i=1

i
w ki (x0:k x0:k
)

(11)

where the Bayesian importance weights w ki are

given by:
w ki =

wki
Ns

j
wk

(12)

j=1

wki =

i
i
)p(x0:k
)
p(z0:k | x0:k
i
(x0:k
| z0:k )

= wki p(z0:k )

(13)

For on-line applications, the estimate of the distribution p(x0:k | z0:k ) at the k-th time step can be obtained
from the distribution p(x0:k1 | z0:k1 ) at the previous
time step by the following recursive formula obtained
by extension of equation (4) for the Bayesian filter p(xk | z0:k ) (Doucet et al., 2001 and Arulampalam
2002):
p (x0:k | z0:k) =

p (x0:k | z0:k1) p (zk | x0:k | z0:k1)

p (zk | z0:k1)

i
i
i.e. (xk | x0:k1
| z0:k ) = p(xk | xk1
) and the nonnormalized weights (16) become (Tanizaki 1997 and
Tanizaki & Mariano 1998):


i
wki = wk1
(16)
p zk | xki

Actually, in many engineering applications, the

measurements are not available at each time step, but
rather at regularly scheduled or even opportunistically
staggered time instants k1 , k2 , . . .. Denoting the observation sequence up to the current time k as the set
{z}k = {zj : j = k1 , . . ., kf k}, the Monte Carlo
estimation formalism described above remains valid,
provided that the weights updating formulas (16) and
(17) are applied only at the observation times k1 , . . . ,
kf , whereas the weights wki remain constant at all other
time instants k  = k1 , . . ., kf .
3

p (xk | x0:k1 | z0:k1) p (x0:k1 | z0:k1) p (zk | x0:k | z0:k1)

=
p (zk | z0:k1)
=

p (xk | x0:k1) p (x0:k1 | z0:k1) p (zk | x0:k)

p (zk | z0:k1)

= p (x0:k1 | z0:k1)

p (zk | x0:k) p (xk | x0:k1)

p (zk | z0:k1)

(14)

The Paris-Erdogan model is adopted for describing the

evolution of the crack depth x in a component subject
to fatigue (Rocha & Schueller 1996 and Myotyri et al.,
2006 and Pulkkinen 1991):
dx
= C(K)n
dt

Furthermore, if the importance function is chosen

such that:
(x0:k | z0:k) = (x0 | z0)

k


(xj | x0:j1 | z0:j )

j=1

= (xk |x0:k1 | z0:k) (x0:k1 | z0:k1)

(15)

the following recursive formulas for the nonnormalized weights wki and wki can be obtained:
wki =

i
| z0:k)
p(x0:k
i
(x0:k
| z0:k)
i )
p(zk | xki )p(xki | xk1
p(zk | zk1)
i
i
(xki | x0:k1
| z0:k)(x0:k1
| z0:k1)

where C and n are constant parameters related to the

component material properties (Kozin & Bogdanoff
1989 and Provan 1987), which can be estimated from
experimental data (Bigerelle & Iost 1989) and K is
the stress intensity amplitude, roughly proportional to
the square root of x (Provan 1987):

K = x
(18)

 n
dx
= e C x
dt

i
p(zk | xki )p(xki | xk1
)

1
i
i
(xk | x0:k1 | z0:k) p(zk | zk1)

i
wki = wki p(z0:k) = wk1

i
p(zk | xki )p(xki | xk1
)
i
(xki | x0:k1
| z0:k)

(17)

with being a constant parameter which may be

determined from experimental data.
The intrinsic randomness of the process is
accounted for in the model by modifying equation (18)
as follows (Provan 1987):

i
p(x0:k1
| z0:k1)

i
= wk1

CONDITION-BASED MAINTENANCE
AGAINST FATIGUE CRACK
DEGRADATION

(15)

The choice of the importance function is obviously crucial for the efficiency of the estimation.
In this work, the prior distribution of the hidden Markov model is taken as importance function,

(19)

where N (0, 2 ) is a white Gaussian noise.

For t sufficiently small, the state-space model
(20) can be discretized as:
xk = xk1 + ek C(Kt)nt

(20)

which represents a non-linear Markov process with

independent, non-stationary degradation increments.

479

For the observation zk , a logit model can be

introduced (Simola & Pulkkinen 1998):
ln

zk
xk
= 0 + 1 ln
+ k
d zk
d xk

(21)

where d is the component material thickness, 0

(, ) and 1 > 0 are parameters to be estimated
from experimental data and is a white Gaussian noise
such that N (0, 2 ).
Introducing the following standard transformations,
yk = ln

expected cost per unit time (k, l)

=

expected cost over the current life cycle (k, l)

expected current life cycle (k, l)
(27)

where:

zk
d zk

k = 0 + 1 ln

by the knowledge of the crack propagation stochastic process and a set of available measurements {z}k
related to it, taken at selected times prior to k. The best
time to replacement lmin is the one which minimizes
the expression (Christer et al., 1997):

(22)
xk
d xk

(23)

then, Yk N (k , 2 ) is a Gaussian random variable with conditional cumulative distribution function (cdf):

FYk (yk | xk) = P(Yk < yk | xk) =

yk k

(24)

where (u) is the cdf of the standard normal distribution N (0,1).

The conditional cdf of the measurement zk related
to the degradation state xk is then:

FZk (zk | xk) = FYk ln

zk
| xk
d zk

zk
1
k
ln
=

d zk

Neglecting the monitoring costs:

expected cost over the current life cycle (k, l) =

2
zk
k
ln
d zk

e
fZk (zk | xk ) =
2

(28)

where:

= cp (1 P (k + l)) + cf P (k + l)

(25)

with corresponding probability density function:

l denotes the remaining life duration until replacement (either preventive or upon failure),
k + l denotes the preventive replacement time instant
scheduled on the basis of the set of observations
{z}k collected up to time k,
d denotes the critical crack threshold above which
failure is assumed to occur (d < d),
cp denotes the cost of preventive replacement,
cf denotes the cost of replacement upon failure,
p (k + i) = P (xk+i > d | x0:k+i1 < d , {z}k ))
denotes the conditional posterior probability of
the crack depth first exceeding d in the interval
(k + i 1, k + i), knowing the component had not
failed up to time k + i 1 and given the sequence of
observations {z}k available up to time k,
P(k + l)denotes the probability of the crack depth
exceeding d in the interval (k, k + l).

d
zk (d zk )
(26)

The maintenance actions considered for the component are replacement upon failure and preventive
replacement. For a wide variety of industrial components, preventive replacement costs are lower than failure replacement ones, since unscheduled shut down
losses must be included in the latter.
At a generic time step k of the components life, a
decision can be made on whether to replace the component or to further extend its life, albeit assuming the
risk of a possible failure. This decision can be informed

P(k + l) = p(k + 1) + (1 p(k + 1))p (k + 2) +

p (k + 1)
l=1

p (k + 1) + 
=
i=2

i1

(1 p (k + j)) p (k + i) l > 1

j=1

(29)
and
expected current life cycle (k, l) =

k +1

(k + l) (1 P (k + l)) + (k + 1) p (k + 1)
=
l
i1



+ (k + i) (1 p (k + j)) p (k + i)

480

i=2

j=1

(30)

Within the Monte Carlo sampling framework of

Section 2, an estimate of the unknown conditional
posterior probability p(k + i) may be obtained as:

p (k + i) =

m:xkm >d

n:xkn >d

wkm
wkn

(31)

where the numerator sums the weights of the sampled

particles which fail for the first time in the time interval (k + i 1, k + i) and the denominator those of the
particles which survive up to time (k + i 1).
The computational scheme proposed has been
applied to a case study of literature (Myotyri 2006).
The parameters of the state equation (21) are C =
0.005, n = 1.3 and = 1, whereas those in the measurement equation (22) are 0 = 0.06, and 1 = 1.25.
The material thickness is d = 100, with a failure
threshold at d = 80 in arbitrary units. The process and measurement noise variances are 2 = 2.89
and 2 = 0.22, respectively. The cycle time step is
t = 1 and the mission time is T = 1000. The observations of the crack size are assumed to be collected
at time steps k1 = 100, k2 = 200, k3 = 300, k4 = 400
and k5 = 500. The parameters of the cost model are
cp = 1.5 104 and cf = 7.5 104 , also in arbitrary
units (Christer et al., 1997). The number of particles
sampled by the Monte Carlo procedure is Ns = 1000.
Figure 1 shows the true crack depth evolution over
the mission time, if no preventive replacement is
undertaken. The component is shown to fail at k =
577(x577 > d = 80).
Figure 2 shows the expected cost per unit time computed in correspondence of the observation times k1 ,
k2 , k3 , k4 and k5 .

Figure 2. Particle filter-estimated cost with critical level

d = 80 at the measurement times k1 = 100, k2 = 200,
k3 = 300, k4 = 400 and k5 = 500.
Table 1.

True crack depth evolution.

Time step
(k)

Minimum E[cost
per unit time]

kmin

lmin

100
200
300
400
500

31.7
31.7
29.5
28.0
28.5

527
527
535
568
535

427
327
235
168
35

Table 1 reports the minimum values of the expected

cost per unit time, the corresponding time instants and
the remaining life durations before replacement.
At the first measurement time k1 = 100, the suggested time of replacement is 527, which is then
successively updated following the additionally collected measurements up to the last one at k5 = 500
when the recommendation becomes to replace the
component at k = 535, or, equivalently, after lmin =
35 time steps. In this updating process, a decrease
of the recommended time of replacement is observed
from 568 at k4 = 400 to 535 at the successive measurement time k5 = 500; this coherently reflects the
true crack depth evolution (Figure 1), which witnesses
a somewhat sudden increase in the period between
the two measurement times k = 400 and k = 500:
the bayesanly updated recommendation based on the
observation at k5 = 500 coherently takes into account
this increase, resulting in a reduced predicted lifetime.

Figure 1.

Expected costs and replacement times.

CONCLUSIONS

In this work, a method for developing a conditionbased maintenance strategy has been propounded. The
method relies on particle filtering for the dynamic

481

estimation of failure probabilities from noisy measurements related to the degradation state.
An example of application of the method has
been illustrated with respect to the crack propagation
dynamics of a component subject to fatigue cycles and
which may be replaced preventively or at failure, with
different costs.
The proposed method is shown to represent a
valuable prognostic tool which can be used to drive
effective condition-based maintenance strategies for
improving the availability, safety and cost effectiveness of complex safety-critical systems, structures and
components, such as those employed in the nuclear
industry.
REFERENCES
Anderson, B.D. & Moore, J.B. 1979. In Englewood Cliffs.
Optimal Filtering. NJ: Prentice Hall.
Arulampalam, M.S. Maskell, S. Gordon, N. & Clapp, T.
2002. A Tutorial on Particle Filters for Online
Nonlinear/Non-Gaussian Bayesian Tracking. IEEE Trans.
On Signal Processing 50(2): 174188.
Bigerelle, M. & Iost, A. 1999. Bootstrap Analysis of FCGR,
Application to the Paris Relationship and to Lifetime
Prediction. International Journal of Fatigue 21: 299307.
Christer, A.H. & Wang, W. & Sharp, J.M. 1997. A state space
condition monitoring model for furnace erosion prediction and replacement. European Journal of Operational
Research 101: 114.
Djuric, P.M. & Kotecha, J.H. & Zhang, J. & Huang, Y. &
Ghirmai, T. & Bugallo, M. F. & Miguez, J. 2003. Particle
Filtering. IEEE Signal Processing Magazine 1937.
Doucet, A. 1998. On Sequential Simulation-Based Methods for Bayesian Filtering. Technical Report. University
of Cambridge, Dept. of Engineering. CUED-F-ENGTR310.
Doucet, A. & Freitas, J.F.G. de & Gordon, N.J. 2001. An
Introduction to Sequential Monte Carlo Methods. In A.
Doucet, J.F.G. de Freitas & N.J. Gordon (eds), Sequential
Monte Carlo in Practice. New York: Springer-Verlag.

Doucet, A. & Godsill, S. & Andreu, C. 2000. On Sequential

Monte Carlo Sampling Methods for Bayesian Filtering.
Statistics and Computing 10: 197208.
Kalos, M.H. & Whitlock, P.A. 1986. Monte Carlo methods.
Volume I: basics, Wiley.
Kitagawa, G. 1987. Non-Gaussian State-Space Modeling
of Nonstationary Time Series. Journal of the American
Statistical Association 82: 10321063.
Kozin, F. & Bogdanoff, J.L. 1989. Probabilistic Models of
Fatigue Crack Growth: Results and Speculations. Nuclear
Engineering and Design 115: 14371.
Marseguerra, M. & Zio, E. & Podofillini, L. 2002. ConditionBased Maintenance Optimization by Means of Genetic
Algorithms and Monte Carlo Simulation. Reliability
Engineering and System Safety 77: 151165.
Myotyri, E. & Pulkkinen, U. & Simola, K. 2006. Application
of stochastic filtering for lifetime prediction. Reliability
Engineering and System Safety 91: 200208.
Provan J.W. (ed) 1987, Probabilistic fracture mechanics and
reliability, Martinus Nijhoff Publishers.
Pulkkinen, U. 1991. A Stochastic Model for Wear Prediction through Condition Monitoring. In K. Holmberg and
A. Folkeson (eds). Operational Reliability and Systematic
Maintenance. London/New York: Elsevier 223243.
Rocha, M. M. & Schueller, G. I. 1996. A Probabilistic Criterion for Evaluating the Goodness of Fatigue Crack Growth
Models. Engineering Fracture Mechanics 53: 707731.
Seong, S.-H. & Park, H.-Y. & Kim, D.-H. & Suh, Y.-S. &
Hur, S. & Koo, I.-S. & Lee, U.-C. & Jang, J.-W. & Shin,
Y.-C. 2002. Development of Fast Running Simulation
Methodology Using Neural Networks for Load Follow
Operation. Nuclear Science and Engineering 141: 6677.
Simola, K. & Pulkkinen, U. 1998. Models for non-destructive
inspection data. Reliability Engineering and System Safety
60: 112.
Tanizaki, H. 1997. Nonlinear and nonnormal filters using
Monte Carlo methods. Computational Statistics & Data
Analysis 25: 417439.
Tanizaki, H. & Mariano, R.S. 1998. Nonlinear and NonGaussian State-Space Modeling with Monte Carlo Simulations. Journal of Econometrics 83: 263290.

482

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Corrective maintenance for aging air conditioning systems

I. Frenkel & L. Khvatskin
Center for Reliability and Risk Management, Industrial Engineering and Management Department,
Sami Shamoon College of Engineering, Beer Sheva, Israel

Anatoly Lisnianski
The Israel Electric Corporation Ltd., Haifa, Israel

ABSTRACT: This paper considers corrective maintenance contracts for aging air conditioning systems, operating under varying weather conditions. Aging is treated as an increasing failure rate. The system can fall into
unacceptable states for two reasons: through performance degradation because of failures or through an increase
in demand of cold. Each residence in acceptable state, each repair and each entrance to an unacceptable state are
associated with a corresponding cost. A procedure for computing this reliability associated cost is suggested,
based on the Markov reward model for a non-homogeneous Poisson process. By using this model an optimal
maintenance contract that minimizes the total expected cost may be found. A numerical example for a real world
air conditioning system is presented to illustrate the approach.

INTRODUCTION

Many technical systems are subjected during their lifetime to aging and degradation. Most of these systems
are repairable. For many kinds of industrial systems,
it is very important to avoid failures or reduce their
occurrence and duration in order to improve system
reliability and reduce corresponding cost.
Maintenance and repair problems have been widely
investigated in the literature. Barlow & Proshan
(1975), Gertsbakh (2000), Wang (2002) survey and
summarize theoretical developments and practical
applications of maintenance models.
With the increasing complexity of systems, only
specially trained staff with specialized equipment can
provide system service. In this case, maintenance service is provided by an external agent and the owner
is considered a customer of the agent for maintenance
service. In the literature, different aspects of maintenance service have been investigated (Almeida (2001),
Murthy & Asgharizadeh(1999)).
Reliability theory provides a general approach for
constructing efficient statistical models to study aging
and degradation problems in different areas. Aging is
considered a process which results in an age-related
increase of the failure rate. The most common shapes
of failure rates have been observed by Meeker and
Escobar (1998), Finkelstein (2002).
This paper presents a case study where an aging
air conditioning system with minimal repair is considered. The Markov reward model is built for computing

the reliability associated cost, accumulated during

the aging systems life span. By using the model a
corrective maintenance contract with minimal reliability associated cost can be defined from the set of
different maintenance contracts available in the market. The approach is based on the Non-Homogeneous
Markov Reward Model. The main advantage of the
suggested approach is that it can be easily implemented
in practice by reliability engineers.

2
2.1

PROBLEM FORMULATION AND MODEL

DESCRIPTION
Reliability Associated Cost and Corrective
Maintenance Contract

We will define the Reliability Associated Cost (RAC)

as the total cost incurred by the user in operation and
maintenance of the system during its lifetime. In that
case RAC will comprise the operating cost, repair cost
and the penalty cost accumulated during the system
lifespan. Therefore,
RAC = OC + RC + PC

(1)

where
OC is the system operating cost accumulated during
the system lifetime;

483

 RC is the repair cost incurred by the user in

operating and maintaining the system during its
lifetime;
PC is a penalty cost, accumulated during system life
time, which was paid when the system failed.
Let T be the system lifetime. During this time the
system may be in acceptable states (system functioning) or in unacceptable ones (system failure). After
any failure, a corresponding repair action is performed
and the system returns to one of the previously acceptable states. Every entrance of the system into a set of
unacceptable states (system failure) incurs a penalty.
A Maintenance Contract is an agreement between
the repair team and the systems owner that guarantees a specific level of services being delivered.
The Maintenance Contract defines important parameters that determine a service level and corresponding
costs. The main parameter is mean repair time Trm ,
where m(m = 1, 2, . . . , M ) is a possible Maintenance
Contract level and M is the number of such levels.
Repair cost crm for individual repair action depends
on repair time and, so, it corresponds to a maintenance
contract level m. It usually ranges between the highest
and lowest repair costs. Thus Trmin Trm Trmax .
The problem is to find the expected reliability
associated cost corresponding to each maintenance
contract and choose the corrective maintenance contract, minimizing this cost. According to the suggested
approach, this cost is represented by the total expected
reward, calculated via a specially developed Markov
reward model.
2.2 Markov reward model for aging system
A Markov reward model was first introduced by
Howard (1960), and applied to multi-state system
(MSS) reliability analysis by Lisnianski & Levitin
(2003) and others (Lisnianski (2007), Lisnianski et al.,
(2008)).
We suppose that the Markov model for the system
has K states that may be represented by a state space
diagram as well as transitions between states. Intensities aij , i, j = 1, . . . , K of transitions from state i to
state j are defined by corresponding failure and repair
rates.
It is assumed that while the system is in any state
i during any time unit, some money rii will be paid.
It is also assumed that if there is a transition from
state i to state j the amount rij will by paid for each
transition. The amounts rii and rij are called rewards.
The objective is to compute the total expected reward
accumulated from t = 0, when the system begins its
evolution in the state space, up to the time t = T under
specified initial conditions.
Let Vj (t) be the total expected reward accumulated
up to time t, if the system begins its evolution at time

t = 0 from state j. According to Howard (1960),

the following system of differential equations must be
solved in order to find this reward:


dVj (t)
= rjj +
aij rij +
aij Vi (t),
dt
i=1
i=1
K

i =j

j = 1, 2, . . . , K

(2)

The system (2) should be solved under specified

initial conditions:
Vj (0) = 0, j = 1, 2, . . . , K.
For an aging system, its failure rate (t) increases
with age. In the case of minimal repair, the intensities aij , i, j = 1, . . . , K of transitions from state i to
state j corresponding to failures are dependent on time.
The total expected reward can be found from differential equations (2), by substitution of formulae for (t)
instead of corresponding aij values.

3
3.1

NUMERICAL EXAMPLE
The system description

Consider an air conditioning system, placed in a computer center and used around the clock in varying
temperature conditions. The system consists of five
identical air conditioners. The work schedule of the
system is as follows. For regular temperature conditions two air-conditioners must be on-line and three
others are in hot reserve. For peak temperature conditions four air-conditioners have to be on-line and one
is in hot reserve. The number of the air conditioners
that have to be on-line define the demand level.
We denote:
c is the system operations cost per time unit.
cr is the repair cost paid for every order of the
maintenance team;
cps is a penalty cost, which is paid, when the
system fails.
An aging process in air-conditioners is described
via the Weibull distribution with parameters =
1.5849 and = 1.5021. Therefore (t) = 3t 0.5021 .
Service agents can suggest 10 different Corrective
Maintenance Contracts, available in the market. Each
contract m is characterized by repair rate and corresponding repair cost (per repair) as presented in the
Table 1.
The operation cost cop , is equal to $72 per year. The
penalty cost cp , which is paid when the system fails,
is equal to $500 per failure.

484

Table 1.

Maintenance contract characteristics.

Maintenance
contract
M

MTTR
(days)

Repair cost
($ per repair)
crm

1
2
3
4
5
6
7
8
9
10

3.36
1.83
1.22
0.91
0.73
0.61
0.52
0.46
0.41
0.37

36
40
46
52
58
66
74
84
94
106

Following Lisnianski (2007), the variable demand,

representing variable weather conditions, may be
described as a continuous time Markov chain with
2 levels. The first level represents a minimal cold
demand during the night and the second level represents peak cold demand during the day. The cycle time
is Tc = 24 hours and the mean duration of the peak is
td = 9 hours. The transition intensities of the model
can be obtained as
d =

1
1
= 0.066 hours1
=
Tc td
24 9

(3)
where

= 584 year 1 ,
N =

1
= 0.111 hours1 = 972 year 1
td

By using the suggested method one will find the

best maintenance contract level m that provides a minimum of Reliability Associated Cost during system
lifetime.
3.2

The transition intensity matrix for the system is as

shown in (3).

 a11 2(t) 0
0
0
0
|


a22
2(t) 0
0
0
|

2
a33
2(t) 0
0
|
 0
 0
0
3
a44
2(t) 0
|

 0
0
0
4
a55
(t) |

 0
0
0
0
5
a66 |

a =

0
0
0
0
0
|
N

 0
N
0
0
0
0
|

 0
0
N
0
0
0
|

0
0
N
0
0
|
 0
 0
0
0
0
N
0
|

 0
0
0
0
0
N
|


0
0
0
0
| d 0


| 0
d
0
0
0
0

| 0
0
d
0
0
0


| 0
0
0
d
0
0


| 0
0
0
0
d
0

| 0
0
0
0
0
d 

| a77 4(t) 0
0
0
0


|
a88
4(t) 0
0
0

| 0
2
a99
3 (t) 0
0


| 0
0
3
a10,10 2 (t) 0

| 0
0
0
4
a11,11 (t) 
| 0
0
0
0
5
a12,12

The Markov Reward model for the system

The state-space diagram for this system is presented

in Figure 1. Rectangles in the left and right sides of the
state space diagram designate states of air conditioners
corresponding to each system state i = 1, . . ., 12. If an
air conditioner is not available, then the corresponding rectangle is dashed. The number of available air
conditioners in any system state i defines the system
performance gi in this state. Also near each state the
corresponding demand wi is presented.
As was written above, technical requirements
demand that two on-line air conditioners are needed
under regular conditions and four in peak condition,
so that there are six acceptable statesstates 14 and
states 78, and 6 unacceptable states: states 56 and
states 912.

a11
a22
a33
a44
a55
a66

= (2(t) + d )
= (2 (t) + + d )
= (2 (t) + 2 + d )
= (2 + 3 + d )
= ( (t) + 4 + d )
= (5 + d )

a77 = (4 (t) + N )
a88 = (4 (t) + + N )
a99 = (3 (t) + 2 + N )
a10,10 = (2 (t) + 3 + N )
a11,11 = ( (t) + 4 + N )
a12,12 = (5 + N )

To calculate the total expected reward, the reward

matrix for the system is built in the following manner
(see Lisnianski & Levitin (2003) and Lisnianski et al.,
(in press)).
If the system is in states 1-5 and 7-11, the operation
cost associated with use of air conditioners, depends
on their number, must be paid during any time unit. In
the states 6 and 12 there are no on-line air-conditioners,
so there are no rewards associated with these states.
The transitions from state 4 to state 5, from state 3
to state 9, from state 8 to state 9 are associated with the
entrance to unacceptable states and rewards associated
with this transitions are the penalty.
The transitions from state 2 to state 1, from state
3 to state 2, from state 4 to state 3, from state 5 to
state 4, from state 6 to state 5, from state 8 to state 7,
from state 9 to state 8, from state 10 to state 9, from
state 11 to state 10 and from state 12 to state 11 are

485

Main

Main

Reserved
Available

Reserved
Available

Reserved
Available

(g=5) > (w=2)

Main

Main

Main

Reserved
Available

Main

(g=5) > (w=4)

N

4
d

Main

Main

Reserved
On-line

Reserved
Available

Reserved
Available

Main

2
(g=4) = (w=2)

Main

Reserved
On-line

Main

Main

Main

Main

Main

Main

Main

Main

Reserved
On-line

Main

Main

Main

Reserved
On-line

Main

Main

Main

(g=4) = (w=4)
N

2
d

Main

Main

Reserved
On-line

Reserved
On-line

Reserved
Available

3
(g=3) > (w=2)

Main

Reserved
On-line

(g=3) < (w=4)

N

3
d

Main

Reserved
On-line

Main

Reserved
On-line

Reserved
On-line

Main

10

(g=2) = (w=2)

(g=2) < (w=4)

N

Main

Main

Reserved
On-line

Reserved
On-line

Reserved
On-line

Main

11

(g=1) < (w=4)

(g=1) < (w=2)

5
d

Main

Main

Reserved
On-line

Reserved
On-line

Reserved
On-line

12

(g=0) < (w=2)

Figure 1.

Reserve
On-line

(g=0) < (w=4)

The state-space diagram for the air-conditioning system.

|
|
|
|
|
|
|
|
|
|
|
|

associated with the repair of the air conditioner, and

the reward associated with this transition is the mean
cost of repair. The reward matrix for the system of air
conditioners is as follows:











r = 










Main

2cop
crm
0
0
0
0
0
0
0
0
0
0

0
2cop
crm
0
0
0
0
0
0
0
0
0

0
0
2cop
crm
0
0
0
0
0
0
0
0

0
0
0
2cop
crm
0
0
0
0
0
0
0

0
0
0
cp
cop
crm
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
0
0

|
|
|
|
|
|
|
|
|
|
|
|

0
0
0
0
0
4cop
crm
0
0
0
0
0

0
0
0
0
0
0
0
4cop
crm
0
0
0

0
0
cp
0
0
0
0
cp
3cop
crm
0
0

0
0
0
0
0
0
0
0
0
2cop
crm
0

0
0
0
0
0
0
0
0
0
0
0
crm

0
0
0
0
0
0
0
0
0
0
0
0
























Taking into consideration the transition intensity

matrix (3), the system of differential equations that
defines the Markov reward model for the air conditioning system for the calculation of the total
expected reward, may be written as shown in (4).

486

Reliability Associated Cost ($)

x 10

Figure 2 shows the expected Reliability Associated

Cost for T = 10 years as a function of the Maintenance Contract Level (m). The seventh level (m = 7),
which provides the minimal expected reliability associated cost ($14574) for the system, corresponds to a
mean repair time of 0.52 days. Choosing a more expensive Maintenance Contract Level, we pay an additional
payment to the repair team. Choosing a less expensive
one, we pay more for penalties because of transitions
to unacceptable states.

3.5
3
2.5
2
1.5
1

CALCULATION RESULTS

4
6
8
Maintenance Contract Level

10

Figure 2. The expected Reliability Associated Cost vs

Maintenance Contract Level.

The system is solved under initial conditions:

Vi (0) = 0, i = 1, 2, . . . , 12 using MATLAB , the
language of technical computing.

CONCLUSIONS

The case study for the estimation of expected reliability associated cost accumulated during system
lifetime is considered for an aging system under minimal repair. The approach is based on application
of a special Markov reward model, well formalized and suitable for practical application in reliability engineering. The optimal corrective maintenance

dV1 (t)
= 2cop (2 (t) + d ) V1 (t) + 2 (t) V2 (t) + d V7 (t)
dt
dV2 (t)
= 2cop + crm + V1 (t) (2 (t) + + d ) V2 (t) + 2 (t) V3 (t) + d V8 (t)
dt
dV3 (t)
= 2cop + 2crm + cp d + 2V2 (t) (2 (t) + 2 + d )V3 (t) + 2 (t) V4 (t) + d V9 (t)
dt
dV4 (t)
= 2cop + 3crm + 2cp (t) + 3V3 (t) (2 (t) + 3 + d ) V4 (t) + 2 (t) V5 (t) + d V10 (t)
dt
dV5 (t)
= cop + 4crm + 4V4 (t) ( (t) + 4 + d ) V5 (t) + (t) V6 (t) + d V11 (t)
dt
dV6 (t)
= 5crm + 5V5 (t) (5 + d ) V6 (t) + d V12 (t)
dt
dV7 (t)
= 4cop + N V1 (t) (4 (t) + N ) V7 (t) + 4 (t) V8 (t)
dt
dV8 (t)
= 4cop + crm + 4cp + N V2 (t) + V7 (t) (4 (t) + + N ) V8 (t) + 4 (t) V9 (t)
dt
dV9 (t)
= 3cop + 2crm + N V3 (t) + 2V8 (t) (3 (t) + 2 + N ) V9 (t) + 3 (t) V10 (t)
dt
dV10 (t)
= 2cop + 3crm + N V4 (t) + 3V9 (t) (2 (t) + 3 + N ) V10 (t) + 2 (t) V11 (t)
dt
dV11 (t)
= 4crm + N V5 (t) + 4V10 (t) ( (t) + 4 + N ) V11 (t) + (t) V12 (t)
dt
dV12 (t)
= 5crm + N V6 (t) + 5V11 (t) (5 + N ) V12 (t)
dt

487

(4)

contract (m = 7), which provides Minimal expected

reliability associated cost ($14574), was found.
REFERENCES
Almeida de, A.T. 2001. Multicriteria Decision Making on
Maintenance: Spares and Contract Planning. European
Journal of Operational Research 129: 235241.
Barlow, R.E. & Proshan, F. 1975. Statistical Theory of Reliability and Life Testing. Holt, Rinehart and Winston: New
York.
Finkelstein, M.S. 2002. On the shape of the mean residual
lifetime function. Applied Stochastic Models in Business
and Industry 18: 135146.
Gertsbakh, I. 2000. Reliability Theory with Applications to
Preventive Maintenance. Springer-Verlag: Berlin.
Howard, R. 1960. Dynamic Programming and Markov
Processes. MIT Press: Cambridge, Massachusetts.
Lisnianski, A. 2007. The Markov Reward Model for a
Multi-State System Reliability Assessment with Variable
Demand. Quality Technology & Quantitative Management 4(2): 265278.

Lisnianski, A., Frenkel, I., Khvatskin, L. & Ding Yi. 2007.

Markov Reward Model for Multi-State System Reliability Assessment. In F. Vonta, M. Nikulin, N. Limnios, C.
Huber-Carol (eds), Statistical Models and Methods for
Biomedical and Technical Systems. Birkhaser: Boston,
153168.
Lisnianski, A., Frenkel, I., Khvatskin, L. & Ding Yi.
2008. Maintenance contract assessment for aging system, Quality and Reliability Engineering International,
in press.
Lisnianski, A. & Levitin, G. 2003. Multi-state System Reliability. Assessment, Optimization and Applications. World
Scientific: NJ, London, Singapore.
Meeker. W. & Escobar, L. 1998. Statistical Methods for
Reliability Data. Wiley: New York.
Murthy, D.N.P. & Asgharizadeh, E. 1999. Optimal decision
Making in a Maintenance Service Operation, European
Journal of Operational Research 116: 259273.
Wang, H. 2002. A survey of Maintenance Policies of
Deteriorating Systems. European Journal of Operational
Research 139: 469489.

488

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Exact reliability quantification of highly reliable systems

with maintenance
R. Bri
Technical University of Ostrava, Czech Republic

ABSTRACT: The paper presents a new analytical algorithm which is able to carry out exact reliability quantification of highly reliable systems with maintenance (both preventive and corrective). A directed acyclic graph
is used as a system representation. The algorithm allows take into account highly reliable and maintained input
components. New model of a repairable component undergoing to hidden failures, i.e. a model when a failure
is identified only at special deterministically assigned times, is analytically deduced within the paper. All considered models are implemented into the new algorithm. The algorithm is based on a special new procedure
which permits only summarization between two or more non-negative numbers that can be very different. If the
summarization of very small positive numbers transformed into the machine code is performed effectively no
error is committed at the operation. Reliability quantification is demonstrated on a real system from practice.
1

INTRODUCTION

It is a Monte Carlo simulation method (Marseguerra &

Zio (2001)) which is used for the quantification of reliability when accurate analytic or numerical procedures
do not lead to satisfactory computations of system
reliability. Since highly reliable systems require many
Monte Carlo trials to obtain reasonably precise estimates of the reliability, various variance-reducing
techniques (Tanaka et al., 1989), eventually techniques
based on reduction of prior information (Baca 1993)
have been developed. A direct simulation technique
has been improved by the application of a parallel algorithm (Bris 2008) to such extent that it can be used for
real complex systems which can be then modelled and
quantitatively estimated from the point of view of the
reliability without unreal simplified conditions which
analytic methods usually expect. If a complex system
undergoes a fault tree (FT) analysis, new and fast FT
quantification method with high degree of accuracy
for large FTs, based also on a Monte Carlo simulation technique and truncation errors was introduced in
Choi & Cho (2007). However, if it is necessary to work
and quantitatively estimate highly reliable systems,
for which unreliability indicators (i.e. system nonfunctions) move to the order 105 and higher (106
etc.), the simulation technique, whatever improved,
can meet the problems of prolonged and inaccurate
computations.
Highly reliable systems appear more often and in
research they are closely connected with a penetrative
increase of progress. We can observe the systems for
example in space applications where complex device

is often designed for an operation lasting very long

time without a possibility of help of human hand.
Safety systems of nuclear power stations represent
other example of highly reliable systems. They have to
be reliable enough to comply with still increasing internationally agreed safety criteria and moreover they are
mostly so called sleeping systems which start and operate only in the case of big accidents. Their hypothetical
failures are then not apparent and thus reparable only
at optimally selected inspective times. The question
is how to model the behaviour of these systems and
how to find their efficient procedure for estimation
and computation of reliability indicators.

2
2.1

A PROBLEM FORMULATION AND INPUT

COMPONENT MODELS
Problem formulation

Let us have a system assigned with a help of a directed

acyclic graph (AG), see (Bris 2008). Terminal nodes of
the AG are established by the definition of deterministic or stochastic process, to which they are subordinate.
From them we can compute a time course of the
availability coefficient, possibly unavailability of individual terminal nodes, using methodology of basic
renewal theory, as for example in (Bris 2007). The
aim is then to find a correspondent time course of
the (un)availability coefficient for the highest SS node
which represents reliability behaviour of the whole
system.

489

2.2

2.3 Derivation of an unavailability coefficient

for a model with repairable elements
and occurrence of hidden failures

Input component models

In the first phase of development we will suppose

exponential distribution for the time to failure, possibly for the time to restoration. Under this condition
we can describe all frequently used models with both
preventive and corrective maintenance by three of the
following models:
Model with elements (terminal nodes in AG) that
can not be repaired
Model with repairable elements (CMCorrective
Maintenance) for apparent failures, i.e. a model
when a possible failure is identified at the occurrence and immediately afterwards it starts a process
leading to its restoration.
Model with repairable elements with hidden failures, i.e. a model when a failure is identified only at
special deterministically assigned times, appearing
with a given period (moments of periodical inspections). In the case of its occurrence at these times
an analogical restoration process starts, as in the
previous case.
An analytical accurate computation of time dependence of the (un)availability coefficient was for the
first two situations explained enough and derived in
Bris & Drabek (2007). Let us remind that in the
first case of the element that can not be repaired a
final course of unavailability coefficient is presented
by a distribution function of time to failure of the
element:
P(t) = 1 et ,
where is the failure rate.
In the second case a relation can be derived on
the basis of Laplaces transformation for a similar
coefficient


+
e(+)t
+ +


1 e(+)t , t > 0
=
+

With the same indication of failure and repair rates

as given above the unavailability coefficient can be
described with the following function:
P( ) = (1 PC ).(1 e )



(e
+ PC 1 +
e ) , > 0,

(2)
where is a time which has passed since the last
planned inspection, Pc is the probability of a nonfunctional state of an element at the moment of
inspection at the beginning of the interval to the next
inspection.
Proof Let Tp is an assigned period of inspections or
examinations of the functional state of an element. Let
us further indicate
PA (t) . . . probability that an element is at time t in
the correct functional state,
PB (t) . . . probability that an element is at the time
t in a failure state,
PC (t) . . . probability that an element is at the time
t in a repair state.
In the first interval, when t < 0, TP ), the situation is qualitatively equivalent like an element that can
not be repaired. In the interval between two inspections
the situation is however different. As long as the given
element is failed, it converts into the state of repair. As
long as we set a time variable as the time which has
passed since the moment of the last inspection, then
this situation can be noted as
PB ( = 0) = 0
The element can be at the moment of inspection
either in the state of repair with a probability Pc , which
can be described as

P(t) = 1

(1)

where is the repair rate.

The third model is rather more general than the
earlier stated in the model with periodical exchanges
Bris & Drabek (2007) where we assumed a deterministic time to end of repair. If we hereby presume
that a time to the end of repair is exponential random
variable, it is necessary to derive an analytical computation of time course of the function of (un)availability
coefficient.

PC ( = 0) = PC
And finally
PA ( = 0) = PA
is a probability that an element at the time of the
inspection is in the correct function. It is apparent that
PA + PC = 1.

490

In the course of time between two inspections followed

one by another it is as follows:

2. In other hypotheses we will need this expression to

be always positive, what is also easy to proof.

PA ( ) PA ( ) d + PC ( ) d = PA ( + d )
PB ( ) + PA ( ) d = PB ( + d )

THE NEW ALGORITHM

PC ( ) PC ( ) d = PC ( + d )

3.1 Probabilities of functional and non-functional

state

After the elementary modification the following set

of differential equations can be obtained, at the known
initial conditions as above:

It is evident that probabilities of a functional p and

non-functional state q comply with a relation
p + q = 1.

PA ( ) + PA ( ) PC ( ) = 0

Taking into consideration the final accuracy of real

numbers in a computer it is important which one from
p or q we count. If we want to fully use the machine
accuracy, we have to compute the smaller one from
both probabilities.

PB ( ) PA ( ) = 0
PC ( ) + PC ( ) = 0
The solution of this set is:


PC
PC
e
e
PA ( ) = PA +



PC
PC
e
e
PB ( ) =
PA +

Example 1.1 If we counted hypothetically on a computer with three-digit decimal numbers, then for the
value of q = 0.00143, we would instead of a correct
value p = 0.99857 have only p = 0.999.
In return for q = 1 p, we would get: q = 1 p =
0.00100.
It is apparent that it gets to a great loss of accuracy
if we counted p instead of q.
Seeing that probabilities of a non-function state of
a highly reliable system is very small, we have to
concentrate on numerical expression of these probabilities. For these purposes it is necessary to undergo
a reorganization in a computer calculation and set certain rules which do not have the influence on accuracy
of the computation at the numeration process.

+ PA + PC
PC ( ) = PC e
Then the probability that an element is not in the
state of correct function inside the interval at the time
will be:
P( ) = PB ( ) + PC ( )
PC
e



PC
e + PC e
PA +

= PA + PC +

3.2 Problems with computer subtraction of two

near real numbers
The probability of non-functioning of the simplest possible element (a model with elements that can not be
repaired) is given by the relation

= (1 PC ) (1 e )



(e e )
+ PC 1 +

Thus the relationship (2) is proved.

Note:
1. For the purposes of an effective computer numerical expression (see further) the expression in the
brackets can be converted into the formation:



1+
(e e )



e
1 e() , > 0
=1

P(t) = 1 et ,
that was supposed as an unavailability coefficient,
is a failure rate. Similarly, for other models of system
elements the computation of an expression
1 ex , for x 0

(3)

is a crucial moment at probability numerical expression of a non-function (unavailability coefficient).

For values x << 1, i.e. near 0, direct numerical
expression written by the formula would lead to great
errors! At subtraction of two near numbers it gets to a
considerable loss of accuracy. On personal computer

491

the smallest number , for which it is numerically

evident that
1 + = 1,
is approximately 1018 . If x 1025 , the real value
of the expression (3) will be near 1025 . A direct
numerical calculation of the expression gives a zero!
As the algorithm was created in a programming
environment Matlab, for the need of this paper was
used the Matlab function exmp1 which enables
exact calculation of the expression (3).
3.3

The numeration substance of probability

of a non-functional state of a node

The probability of a non-functional state of a node of

an AG, for which the individual input edges are independent, is in fact given by going over all possible combinations of probabilities of the input edges. For 20
input edges we have regularly a million combinations.
One partial contribution to the probability of a nonfunctional state of the node in Figure1 has a form:
q1 q2 qi1 pi qi+1 qj1 pj qj+1 q20 ,
where a number of occurring probabilities p (here the
number equals to 2) can not reach m. The probability of a non-functional state of the node is generally
given by a sum of a big quantity of very small numbers.
These numbers are generally very different!
If the sum will be carried out in the way that the
addition runs in the order from the biggest one to the
smallest ones, certainly a lost stems from rounding off,
more than the addition runs in the order from the smallest ones to the biggest values. And even in this second
case there is not possible to determine effectively how
much accuracy has been lost.
Note: In the case of dependence of the input edges
(terminal nodes) we cannot express the behaviour of
an individual node numerically. There is necessary

to work with the whole relevant sub-graph. Combinatorial character for the quantification will stay
nevertheless unchanged.
3.4 The error-free sum of different non-negative
numbers
The first step to the solution of this problem is to find a
method for the accurate sum of many non-negative
numbers.
The arithmetic unit of a computer (PC) works in a
binary scale. A positive real number of todays PC contains 53 valid binary numbers, see Figure 2. A possible
order ranges from approximately 1000 to 1000.
The line indicated as order means an order of a
binary number.
The algorithm for the accurate quantification of
sums of many non-negative numbers consists from a
few steps:
1. The whole possible machine range of binary positions (bites) is partitioned into segments of 32
positions for orders, according to the following
scheme in Figure 3: The number of these segments
will be approx.:
2000
= 63
32
2. Total sum is memorized as one real number, which
is composed from 32 bite segments. Each from
these segments has additional 21 bites used as
transmission.
3. At first a given non-zero number of the sum that
must be added is decomposed according to before
assigned firm borders (step 1) mostly into three
parts containing 32 binary numbers of the number
at most, according to the scheme in Figure 4. The
individual segments are indexed by numbers 163.
4. Then the individual parts of this decomposed number are added to the corresponding members of the
sum number, as in Figure 5.
5. Always after the processing of 220 numbers (the
limit is chosen so that it could not lead to overflowing of the sum number at any circumstances)

order: 1000
Figure 2.

32

One node of an acyclic graph with 20 edges.

Figure 3.

492

-1000

A positive real number in binary scale.

..

Figure 1.

1 ? ?
53 places

31

-1

-2

..
-32 -33 -34

Segments composed from 32 binary positions.

..

end of the 1-st

non zero segment

Figure 4.

3.5 Permissible context of the usage

not leading to the loss of accuracy

32 bits

The probability of a non-functional state of a repairable

element (see section 2.3a model with repairable elements and occurrence of hidden failures) is given by
the formula (2), which can be adapted to

Decomposition of a given non-zero number.

P( ) = (1 PC ) ( ) + PC ( ),

Figure 5.

Adding a number to the sum number.

A+
Figure 6.

q1 q2 (1 qk ) .

Clearance process.

0
0

Figure 7.

where PC is the probability of a non-functional state

of an element at the moment of the inspection at
the beginning of the interval till the next inspection; , are non-negative mechanically accurate and
numerically expressed functions.
One contribution to the computation of a nonfunctional state of a node generally has a form

The corresponding
part of the number

Demonstration of the final summarization.

a modification of the sum number is carried out

which is indicated as the clearance process.
Upwards a transmission is separated (in the following Figure 6 it is identified by a symbol ) which
is added to the upper sum.
6. If a final sum is required, at first the clearance
process has to be carried out. Then the group of
three sums is elaborated, from which the upper is
the highest non-zero one (identified by a symbol
in Figure 7). We make a sum of these three sums
as usual in a binary scale, when p in the following expression is given by an index of the highest
non-zero segment:
sum = 2p + 2p32 + 2p64

In both cases it happens to occur (1 q). It has already

been explained that the recalculation p 1 q, can
come to catastrophic loss of accuracy. A basic question then comes out: Do we have to modify further the
stated patterns for the purpose of removing the subtraction? Fortunately not. In the introduced context
(1 q) , where is expressed numerically in a full
mechanical accuracy there is no loss in mechanical
accuracy! Thanks to rounding off a final product to 53
binary numbers, lower orders of the expression (1q),
i.e. a binary numbers on 54th place and other places
behind the first valid number, can not be practically
applied.
3.6 Determination of system probability
behaviour according to a graph structure
Let all elements appearing in the system are independent. The probability of a non-functional state of a
system, assigned by the help of AG, is thus simply
gained on the basis of estimation of all nodes upwards.
For instance for AG in the Figure 8:
there is carried out numerical expression of the
probability of a non-functional state of terminal
nodes, i.e. elements 8,9,10 and 5,6,7
there is carried out numerical expression of the
probability of a non-functional state of an internal
node 4 which is given by the following sum:

So numbers in their full machine accuracy (53

binary numbers beginning with 1) are the input
for this process of adding. The output is the
only number in the same machine accuracy (53
binary numbers beginning with 1). The number is
mechanically the nearest number to the real accurate error-free sum which contains in principle up
to 2000 binary numbers.

493

q8 q9 q10 + (1 q8 ) q9 q10
+ q8 (1 q9 ) q10 + q8 q9 (1 q10 )
+ q8 (1 q9 ) (1 q10 )
+ (1 q8 ) q9 (1 q10 )
+ (1 q8 ) (1 q9 ) q10

having the value of 275 we obtain the value of 220 . If

we apply the new algorithm, we obtain error-free sum
which equals:

1
2

sum = 220 + 245

2

Error of the first result reaches to about half numbering mantissa of the result number what can be
considered as relevant error. In spite of the fact that the
example is very hypothetical, it clearly demonstrates
importance of the algorithm.

10

RESULTS WITH TESTED SYSTEM

Figure 8.

The system assigned by the help of a structure AG.

there is carried out numerical expression of the

probability of a non-functional state of an internal
node 3 which is given by the only item
q5 q6 q7
there is carried out numerical expression of the
probability of a non-functional state of a terminal
node 2
there is carried out numerical expression of the
probability of a non-functional state of the highest
SS node 1 which is given:
q2 q3 q4 + (1 q2 ) q3 q4 + q2 (1 q3 ) q4
+ q2 q3 (1 q4 )
Note:
Internal numeration of nodes is such that the node with
a less number can not be inferior to the node with
greater number. Nodes are numbered in the decreasing
order of numbers.
3.7

The new algorithm generated in the Matlab was

tested on real highly reliable system from RATP
company (Regie Autonome des Transports Parisiens).
The system was used as a test case activity
(Dutuit, Y. & Chatelet E. 1997) of the Methodological Research working group of the French Institute
for dependability & Safety (ISDF). The system is
demonstrated in Figure 9.
The system is one-out-of-three active parallel system including three non identical components. The two
first components are periodically tested and repaired
with test intervals T1 and T2 (T2 >> T1 ). The third
component is neither tested nor repairable. Each component has a constant failure rate i ; the repair or
replacement duration are negligible; the test intervals
are very different respectively 24 hours and 26280
hours. We can say that the failures of the first two
components are hidden failures, which are revealed
only when the next test occurs. At the first glance we
can say that the system is highly reliable. Our target is
to calculate the dependence of unavailability on time.
As a first step, the dependence of unavailability on
time was calculated for individual components what is

Importance of the new algorithm

The algorithm is based on summation of many nonnegative numbers that can be very different. The output
of the summation is one number in full machine accuracy, which we call error-free sum. Importance of the
error-free algorithm can be explained on the following
very hypothetical example:
Let us have a system given by the acyclic graph
similar to the one in Fig. 1 with 30 input edges. In such
a case the sum for the probability of a non-function
state of the highest SS node can be composed of about
one billion (230 ) summands. If we have to summarize
(by a software on a common PC) one number of a
value 220 with additional 230 numbers all of them

Figure 9.

494

Block diagram of the tested system.

Figure 10. Dependence of unavailability on time (in hours)

for the Component 1.

Figure 13. Dependence of unavailability on time (in hours)

for the whole system.

Figure 11. Dependence of unavailability on time (in hours)

for the Component 2.

Figure 14. Dependence of unavailability on time (in hours)

for the whole system having all components = 1 e 7.

Figure 12. Dependence of unavailability on time (in hours)

for the Component 3.

demonstrated in Figures 1012. Secondly, dependence

of unavailability on time for the whole system in the
course of first 7 years of operation was determined, in
Figure 13.
We see that the maintenance process is very important for the highly reliable system. Thanks maintenance actions the maximal unavailability in the course
of first six years is 2.5e8.
For better demonstration of the new algorithm, let
us take into account that the system is composed from
identical components such as the Component 1, i.e. all
of them having the same failure rate:
=1e7

The only difference between both systems is the

type of maintenance. In Figure 14 we can see dependence of unavailability on time for the system in course
of first 7 years of operation.
We can easily see that the unavailability dropped
by three orders. Maximal unavailability value in the
course of first six years of operation is about 3.3e11.

CONCLUSION

Maintaining the full machine accuracy requires mainly

not to carry out subtraction of near values. All required
outputs are therefore necessary to express in the form
of the sum of numbers with consistent sign (in our case
non-negative).
A problem of a sum of many non-negative numbers can be solved by decomposing a sum into more
partial sums which can be carried out without a loss!

495

The process has been numerically realized within a

programming environment Matlab.
Numerical expression of probabilities of a nonfunctional state of one node of an AG has a combinatorial character. We have to go over all combinations
of input edges behaviour leading to a non-functional
state of the node. The astronomic increase of combinations with the increasing number of elements causes
that the program will be usable only up to a certain size
of a system. Already at moderate exceeding the critical size of the system it comes to enormous increase
of machine time. All computations above run below
1s, on Pentium (R) 4 CPU 3.40GHz, 2.00 GB RAM.
Model with repairable elements with hidden failures, i.e. a model when a failure is identified only
at special deterministically assigned times, has been
analytically elicited within the paper. Final formula
meets the requirement of permissible context which is
required in the presented algorithm.
The algorithm enables to carry out exact unavailability analysis of real maintained systems with both
preventive and corrective maintenance.
6

ACKNOWLEDGEMENT

This work is supported partly by The Ministry of

Education, Youth and Sports of the Czech Republic
project CEZ MSM6198910007 and partly by The
Ministry of Industry and Trade of the Czech
Republicproject FT-TA4/036.

REFERENCES
Marseguerra M. & Zio E. 2001. Principles of Monte Carlo
simulation for application to reliability and availability
analysis. In: Zio E, Demichela M, Piccinini N, editors.
Safety and reliability towards a safer world, Torino, Italy,
September 1620, 2001. Tutorial notes. pp. 3762.
Tanaka T, Kumamoto H. & Inoue K. 1989. Evaluation of a
dynamic reliability problem based on order of component
failure. IEEE Trans Reliab 1989;38:5736.
Baca A. 1993. Examples of Monte Carlo methods in reliability estimation based on reduction of prior information.
IEEE Trans Reliab 1993;42(4):6459.
Bri R. 2008. Parallel simulation algorithm for maintenance
optimization based on directed Acyclic Graph. Reliab Eng
Syst Saf 2008;93:85262.
Choi JS, Cho NZ 2007. A practical method for accurate
quantification of large fault trees. Reliab Eng Syst Saf
2007;92:971-82.
Bri, R. 2007. Stochastic Ageing ModelsExtensions of the
Classic Renewal Theory. In Proc. of First Summer Safety
and Reliability Seminars 2007, 2229 July, Sopot: 2938,
ISBN 978-83-925436-0-2.
Bri, R. & Drbek, V. 2007. Mathematical Modeling of both
Monitored and Dormant Failures. In Lisa Bartlett (ed.),
Proc. of the 17th Advances in Risk and Reliability Technology Symposium AR2 TS, Loughborough University:
376393.
Dutuit, Y. & Chatelet E. 1997. TEST CASE No. 1, Periodically tested paralel system. Test-case activity of European
Safety and Reliability Association. ISdF-ESRA 1997. In:
Workshop within the European conference on safety and
reliability, ESREL 1997, Lisbon, 1997.

496

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Genetic algorithm optimization of preventive maintenance scheduling

for repairable systems modeled by generalized renewal process
P.A.A. Garcia
Universidade Federal FluminenseDepartamento de Administrao, Volta RedondaRJ, Brasil
Fundao GorceixPetrobras (CENPES), Rio de Janeiro, Brasil

M.C. SantAna
Agncia Nacional de Sade, Rio de Janeiro, Brasil

V.C. Damaso
Centro Tecnolgico do Exrcito, Rio de Janeiro, Brasil

P.F. Frutuoso e Melo

COPPE/UFRJNuclear, Rio de Janeiro, Brasil

ABSTRACT: Reliability analyses of repairable systems are currently modelled through punctual stochastic
processes, which intend to establish survivor measures in a failure x repair scenario. However, these approaches
do not always represent the real life-cycle of repairable systems. In order to have a better and most coherent
reality modelling, one has the Generalized Renewal Process (GRP). With this approach, reliability is modelled
considering the effect of a non-perfect maintenance process, which uses a better-than-old-but-worse-than-new
repair assumption. Considering the GRP approach, this paper presents an availability modelling for operational
systems and discusses an optimisation approach based on a simple genetic algorithm (GA). Finally, a case is
presented and the obtained results demonstrate the efficacy of combining GRP and GA in this kind of problems.

INTRODUCTION

The operational experience of Brazilian nuclear power

plants, gained in the last years, led to research on
probabilistic safety assessment (PSA) in the direction
of preventive maintenance scheduling optimization to
enhance operational safety.
Generally, due to the complexity of nuclear systems,
different optimization techniques have been applied.
These are commonly based on evolutionary computing, specifically on genetic algorithms. Nowadays,
one has other heuristics such as particle swarm optimization (Alkhamis & Ahmed, 2005) and ant colony
(Ruiz et al., 2007, and Samarout et al., 2005), among
others. However, genetic algorithms approaches have
been pre-weighed (Taboada et al., 2007, Martorell
et al., 2007, Martorell et al., 2006, and Garcia et al.,
2006).
The use of evolutionary techniques to establish an
optimized preventive maintenance scheduling is justified because these kinds of problems are generally
combinatorial, where one intents to obtain, simultaneously, the best policy for all the components and

maximize the system availability. Indeed, the availability of many components that comprise the entire
system should be considered. In this case, a superimposed stochastic process combining all stochastic
processes concerning each component can be used.
Whatever the optimization method chosen, the
probabilistic modeling is of crucial importance and
should be given appropriately. The present work aims
to base the availability modeling of repairable systems
on a generalized renewal process (Kijima & Sumita,
1986).
2

RANDOM PROCESSES COMMONLY

USED IN PSA

Firstly, defining repairable systems is appropriate.

According to Rigdon & Basu (2000), a system is called
repairable if, when a failure occurs, it can be restored
to an operating condition by some repair process other
than replacement of the entire system.
For situations where downtime associated with
maintenance, repair or replacement actions is

497

2.1 Generalized renewal process

As mentioned, the probabilistic modeling to be considered in this work to approach repair action, especially
imperfect repairs, is the generalized renewal process
(GRP). Nevertheless, for a complete understanding
about GRP, it is necessary to define the concept of
virtual age (Vn ).
The Vn corresponds to the calculated age of particular equipment after the n-th repair action. Kijima
& Sumita (1986) has proposed two ways to modeling
this virtual age. The first one, commonly named type
I, consists basically of the assumption that a repair
action acts just in the step time just before. With this
assumption, the virtual age of a component increases
proportionally to the time between failures:
Vi = Vi1 + qYi

(1)

Virtual Age

negligible, compared with the mean-time-betweenfailures (MTBF), the point processes are used as
probabilistic models of the failure processes. The commonly adopted point processes in PSA are as follows:
(i) homogeneous Poisson process (HPP), (ii) ordinary
renewal processes (ORP) and (iii) non-homogeneous
Poisson process (NHPP). However, these approaches
do not represent the real life-cycle of a repairable
system (Modarres, 2006). Rather, they have some
assumptions that conflict with reality. In HPP and ORP,
the device, after a repair, returns to an as-good-as-new
condition, and in a NHPP the device, after a repair,
returns to an as-bad-as-old condition.
Kijima & Sumita (1986) introduced the concept of
generalized renewal process (GRP) to generalize the
three point processes previously mentioned. With this
approach, reliability is modeled considering the effect
of a non-perfect maintenance process, which uses
a better-than-old-but-worse-than-new repair assumption. Basically, GRP addresses the repair assumption
by introducing the concept of virtual age, which
defines a parameter q that represents the effectiveness
of repair.

V1

t2

tn

Real Age
Figure 1. Visualization of virtual age (Adapted from
Jakopino, 2005).

whereas the assumption of q = 1 leads to an NHPP

(as bad as old). The values of q that fall in the interval
0 < q < 1 represent the after repair state in which the
condition of the system is better than old but worse
than new.
On the basis of this proposition of virtual age,
Kijima et al., (1988) has proposed the following
approach to calculate the conditional probability of
failure:
F(T |Vn = y) =

F(T + y) F( y)
1 F( y)

(3)

where F is the cumulative distribution function.

Considering (without loss of generality), that time
between failures is modeled by a Weibull distribution,
it follows that

i1

q
F(ti , , , q) = 1 exp
tj
j=1

3
(2)

q can be defined as a rejuvenation parameter in both

models.
According to this modeling, the result of assuming
a value of q = 0 leads to an RP (as good as new),

V2

t1

where Vi is the virtual age immediately after the i-th

repair action, and Yi is the time between the (i 1)-th
and i-th failures.
The type II model considers that the repair can
restore the system considering the elapsed time since
the beginning of its life. In this model, the virtual age
increases proportionally to the total time.
Vi = q(Yi + Vi1 )

Vn

ti + q

i1 
j=1 tj

(4)

AVAILABILITY MODELING
OF REPAIRABLE SYSTEMS

A preventive maintenance policy has an important role

to enhance system availability of any power plant. Nevertheless, scheduling planning for preventive maintenance actions must consider aging characteristics.

498

The higher the degradation the lesser the time

between maintenance stops. In turn, to model systems availability under this condition, the following
assumptions were made:
After each maintenance task, the component will
not go back to an as good as new condition;
There is a probability of failure during? time
between maintenance, leading to a corrective action
and influencing the device unavailability;
When a component is selected to be maintained, all
the components associated will be unavailable too;
Time between maintenances is not necessarily
constant.
Based on these assumptions, a modeling approach
to calculate the component availability considering a
maintenance policy is presented.
Considering R(t) as the component reliability, and
as A(t) its availability, and considering, without loss s
of generality, that the failure process can be modeled
by a Power Law process, and time between failures
is modeled by a Weibull distribution. Thus, At (t),
the availability concerning a preventive maintenance
scheduling, is given by

At1 (t) =

t1 

et R1 (t  |V0 )dt  + 1,
R(t|V0 )e

0,

t0

t0 t < t1
t1 t < t1 + man
(5)

whereAt1 (t) is the availability for t0 t t1 , t0 is

the initial time and t1 is the time to the first maintenance stop; V0 is the virtual age of the component
in the beginning of the simulation. This parameter is
quite important, because in aging characterization, the
component is not new, but it has a certain level of aging.
For the other maintenance steps, the modeling is as
follows:

Notice that, if ti + man is higher than the mission

time, this figure is replaced by Tmiss , which is the under
evaluation mission time.
Considering this modeling approach, the mean
availability is

A =

1
Tmiss

ti

n


A(t)dt
i1 +man

MAINTENANCE DISTRIBUTION
AND GENETIC MODELING

The genetic algorithm is used to indicate the preventive

maintenance scheduling for each component, in order
to optimize the system availability along the mission
period. The instants selected by GA for the maintenance accomplishment on a certain component should
follow a distribution pattern. For instance, it is reasonable to suppose that the interval between maintenances
is reduced as the system ages.
This assumption induces choice for a modeling
that looks for solutions to the preventive maintenance
planning with some ordination in that distribution
(Damaso, 2006). The benefit of such approach is to
limit the universe of solutions to be considered, eliminating those that do not have practical sense.Thus,
the search process gains in terms of efficiency and
computational time.
In this work, a proportional distribution was
adopted, where the intervals between preventive maintenances follow a geometric progression (GP).
The first interval, T1 , starts at the operation beginning, in t = 0, and goes until the final instant of the
first intervention, T1 :
T1 = T1 0 = T1 .

(6)
where Ati (t) is the availability between (i 1)-th and
i-th maintenance stops, R(t|Vi1 ) = 1 F(t|Vi1 ),
and F(t|Vi1 ) is done as in equation 4. man is the
maintenance time.

(8)

The subsequent intervals are given as

Ti+1 = Ti ,

At1 (t) =

R(t|Vi1 )e(t(ti1 +rep )

 ti

e(t (ti1 +rep )) R1 (t  |Vi1 )dt  + 1,

ti1

0, ti1 + man t < ti

ti t < ti + man

(7)

i=0 t

0 < 1,

(9)

where i = 1, 2, . . . , n, Ti is the i-th time interval

and is the proportionality factor (common ratio of
the GP). A unitary value of means that the intervention instants are distributed evenly, that is, the
intervals between interventions are all the same. The
last interval,Tn+1 , is set between the final instant of
the last intervention, Tn , and the operation end, Tf , and
it is given as
Tn+1 = Tf Tn .

(10)

Considering that n interventions are foreseen during

the component operation time, the expression for the

499

Table 1. Components parameters modeled by Equations (1)

and (2).

Heat
exchanger 1

Motor 1

Motor 2

V-1

Comp.

1/
(days)

(days)

V0 (days)

V-2

Pump 1
Pump 2
Pump 3
Motor 1
Motor 2
Motor 3
Valve 1
Valve 2
Valve 3
Valve 4
Heat Ex 1
Heat Ex 2

3
3
3
2
2
2
1
1
1
1
5
5

1,12
1,11
1,12
2,026
1,62
2,026
0,73
0,5
0,5
0,56
1,47
1,47

338,5
180
338,5
76
128
76
35,32
105,13
105,13
54,5
321
321

0,51
0,3
0,51
0,146
0,7
0,146
0,54
0
0
0,02
0,35
0,35

75
25
75
165
145
165
150
200
200
45
272
272

Pump 1

Heat
exchanger 2

Pump 2

V-3

V-4
Motor 3

Figure 2.

Pump 3

Schematic diagram of CCWS.

calculation of T1 , as a function of Tf , n and , is

given as
T1 =

Tf (1 )
.
1 n+1

(11)

The other intervals can be evaluated starting from

the known value of T1 , by Equation 9.
For each component, the following values are
determined: number of maintenances, n; proportionality factor, ; and displacement of the maintenance
scheduling, d, the latter allowing to anticipate or delay
the scheduled maintenances to avoid undesirable coincidences between component unavailability. Such a set
of parameters is metaphorically called phenotype, and
it represents a candidate to the solution. The candidate
to the solution is codified in a structure denominated
genotype. One of the coding possibilities is to use
a superstructure, in which 16 bits by parameter was
reserved in the employed GA. In appendix I, a figure
concerning the mentioned structure is shown.
5

Table 2.

APPLICATION EXAMPLE

An important active system of the nuclear power

plant Angra I, located in Angra dos Reis, a Rio de
Janeiro Province, is the Component Coolant Water
System (CCWS). Three motor-driven pumps, two heat
exchangers and a set of valves that make possible,
if necessary, system operation in different configurations, compose this system. The operational and safety
functions of the power plant can be maintained with
only one line in operational condition. This line must
be composed of, at least, one motor driven pump and
a heat exchanger.
In Table 1, o the parameters of each component
necessary for the analysis are listed.

Maintenance days for the components of CCWS.

Components

Maintenance moments (days)

Motor 1
Motor 2
Motor 3
Heat Ex. 1
Heat Ex. 2

54, 108, 162, 216, 270, 324, 378, 432, 486

270
55, 109, 163, 217, 271, 325, 379, 433, 487
270
271

Notice that the repair rate concerns the mean time to

repair considering an unexpected failure. The downtime (one day) associated to a scheduled preventive
maintenance for all components is presented in table 1
In the simulation, mission time is 11/2 years, which
is the time between reactor fuel recharge.
The results obtained are shown in table 2.
Some components, due to the reduced mission time,
i.e., 540 days, will be maintained just at the end of the
mission.
The system availability, associated to the optimized
maintenance schedule, is 99,68%, and it is graphically
displayed as shown in Figures 314.

CONCLUDING REMARKS

The present paper showed a different approach to

model system availability through generalized renewal
processes. By mean of this modeling a simple
GA is adopted to optimize preventive maintenance
scheduling.
Another approach presented in this paper concerns
genetic modeling, which was based on a geometric progression series. In this modeling, the genetic

500

Pump 1

Motor 1
1.000

1.00
0.995

0.80

0.990

Availability

0.90

Availability

0.70
0.60

0.985
0.980

0.50
0.40

0.975

0.30

0.970
0

0.20

100

200

300

400

500

400

500

400

500

Time (days)

0.10
0.00
0

100

200

300

400

500

Figure 6.

Availability of the pump 1.

Time (days)

Figure 3.

Pump 2

Availability of the motor 1.

1.000
0.995
Availability

Motor 2
1.00
0.90
0.80
Availability

0.70

0.990
0.985
0.980

0.60
0.50

0.975

0.40
0.30

0.970

0.20

100

200

300

0.10

Time (days)

0.00
0

100

200

300

400

500

Time (days)

Figure 4.

Figure 7.

Availability of the pump 2.

Availability of the motor 2.

Pump 3
1.000
0.995

Motor 3
1.00

Availability

0.90

Availability

0.80
0.70
0.60
0.50
0.40

0.990
0.985
0.980

0.30

0.975

0.20
0.10

0.970

0.00
0

100

200

300

400

500

100

200

300

Time (days)

Time (days)

Figure 5.

Availability of the motor 3.

Figure 8.

algorithm should find the best proportionality factors which will distributethe times for preventive
maintenance stop along the mission period.
These results are satisfactory and have contributed
to demonstrate the efficacy of the proposed approach
in this particular problem.

Availability of the pump 3.

Further work is underway to consider a higher

mission period and verify, in great detail, the maintenance distribution forall components of the system. We
will also draw upon the test scheduling optimization
problem concerning safety systems in aging condition.

501

Valve 1
1.000

Heat Exchanger 1
1.00

0.995

0.80

0.985

0.70
Availability

Availability

0.90
0.990

0.980
0.975

0.60
0.50
0.40
0.30

0.970
0

100

200

300

400

0.20

500

0.10

Time (days)

0.00

Figure 9.

Availability of the valve 1.

100

200

300

400

500

Time (days)

Figure 13.

Valve 2

Availability of the heat exchanger 1.

1.000

Heat Exchanger 2

0.990

1.00
0.90

0.985

0.80
0.980

Availability

Availability

0.995

0.975
0.970
0

100

200

300

400

500

0.70
0.60
0.50
0.40
0.30

Time (days)

0.20
0.10

Figure 10.

0.00

Availability of the valve 2.

100

200

300

400

500

Time (days)

Valve 3

Figure 14.

1.000

Availability of the heat exchanger 2.

Availability

0.995
0.990

REFERENCES

0.985
0.980
0.975
0.970
0

100

200

300

400

500

400

500

Time (days)

Figure 11.

Availability of the valve 3.

Valve 4
1.000

Availability

0.995
0.990
0.985
0.980
0.975
0.970
0

100

200

300
Time (days)

Figure 12.

Availability of the valve 4.

Alkhamis, Talal M. & Ahmed, Mohamed A. (2005).

Simulation-based optimization for repairable systems
using particle swarm algorithms. In Proceeding of the 37th
Conference on Winter Simulation, pp. 857861.
Damaso, Vincius C. (2006). An integrated optimization
modeling for safety system availability using genetic algorithms. PhD Dissertation, Federal University of Rio de
Janeiro, Nuclear Engineering Program [in Portuguese].
Garcia, P.A.A., Jacinto, C.M.C. and Droguett, E.A.L. (2005).
A multiobjective genetic algorithm for blowout preventer
test scheduling optimization. In Proceeding of Applied
Sumulation and Modeling, Benalmdena, Spain.
Jacopino, A.G. (2005). Generalization and Bayesian solution
of the general renewal process for modeling the reliability
effects of imperfect inspection and maintenance based on
imprecise data. PhD Dissertation, Faculty of the Graduate
School of the University of Maryland, College Park.
Kijima, M. and Sumita, N. (1986). A useful generalization of renewal theory: counting process governed by
non-negative Markovian increments. Journal of Applied
Probability, Vol. 23, 71.
Martorell, S., Carlos, S., Villanueva, J.F., Sanchez, A.I.,
Galvan, B., Salazar, D. and Cepin, M. (2006). Use of
multiple objective evolutionary algorithms in optimizaing

502

surveillance requirements. Rel. Eng. Sys. Saf., Vol. 91,

pp. 10271038.
Martorell, S., Sanchez, A. and Carlos, S. (2007). A tolerance interval based approach to address uncertainty for
RAMS+C optimization. Rel. Eng. Sys. Saf., Vol. 92,
pp. 408422.
Modarres, Mohamed (2006). Risk Analysis in Engineering Techniques, Tools, and Trends. Taylor & Francis.
Boca Raton, FL.
Rigdon, S.E. and Basu, A.P (2000). Statistical Methods for
the Reliability of Repairable Systems. John Wiley and
Sons, New York.
Rubns Ruiz, Garca-Daz, J. Carlos and Maroto, Conceptin
(2007). Considering scheduling and preventive mainte-

nance in flowshop sequencing problems. Computer &

Operational Research, Vol. 34, 11, pp. 33143330.
Samrout, M, Yalaoui, F, Chtelet, E. and Chebbo, N.
(2005). New methods to minimize the preventive maintenance cost of series-parallel systems using ant colony
optimization. Rel. Eng. Sys. Saf, Vol. 89, 9, pp. 346354.
Taboada, Heidi A., Baheranwala, Fatema, Coit, David W. and
Wattanapongsakorn, Naruemon (2007). Practical solutions for multi-objective optimization: An application to
system reliability design problems. Rel. Eng. Sys. Saf.,
Vol. 92, pp. 314322.

APPENDIX I

00101100 01101011 00110010 10000111 00011110 10011001 00010001 00101001 10010001 00111001 00000010 00110011 00000011 10000100 00010011
11101100 00101100 1101011 10011101 01100100 11011101 00111110 11011100 11101101 1101001 01111110 10101100 1101110 11011101 10111110

Genotype:

Phenotype:

Motor 1
Motor 2
n number of foreseen interventions
proportionality factor
d maintenance scheduling displacement
Figure I1.

Motor 3

Pump 1

Schematic representation of the genotype/phenotype structure.

503

Pump 2

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Maintenance modelling integrating human and material resources

S. Martorell
Dpto. Ingeniera Qumica y Nuclear, Universidad Politcnica Valencia, Spain

M. Villamizar, A. Snchez & G. Clemente

Dpto. Estadstica e Investigacin Operativa Aplicadas y Calidad, Universidad Politcnica Valencia, Spain

ABSTRACT: Maintenance planning is a subject of concern to many industrial sectors as plant safety and
business depend on it. Traditionally, the maintenance planning is formulated in terms of a multi-objective
optimization (MOP) problem where reliability, availability, maintainability and cost (RAM+C) act as decision
criteria and maintenance strategies (i.e. maintenance tasks intervals) act as the only decision variables. However
the appropriate development of each maintenance strategy depends not only on the maintenance intervals but also
on the resources (human and material) available to implement such strategies. Thus, the effect of the necessary
resources on RAM+C needs to be modeled and accounted for in formulating the MOP affecting the set of
objectives and constraints. In Martorell et al., (2007), new RAM + C models were proposed to address explicitly
the effect of human resources. This paper proposes the extension of the previous models integrating explicitly
the effect of material resources (spare parts) on RAM+C criteria. This extended model allows accounting
for explicitly how the above decision criteria depends on the basic model parameters representing the type of
strategies, maintenance intervals, durations, human resources and material resources. Finally, an application
case is performed on a motor-driven pump analyzing how the consideration of human and material resources
would affect the decision-making.

INTRODUCTION

Often, RAMS+C (Reliability, Availability, Maintainability, Safety plus Cost) take part of the relevant criteria for the decision-making of concern to maintenance
planning.
The RCM (Reliability Centered Maintenance) is a
good example of a systematic methodology to establish an efficient maintenance plan to cope with all
of the equipment dominant failure causes (Figure 1).
Typically, the main objective in applying the RCM
methodology has been to find out the best set of maintenance strategies that provide appropriate balance of
equipment reliability and availability and associated
costs.
In pursuing the RCM goal the decision-maker must
face at least two main problems: 1) there is not a unique
maintenance plan (i.e. set of maintenance strategies as
technical resources) to cope with all the dominant failure cause as shown in Ref. (Martorell et al., 1995), 2)
frequencies suggested for performing specific maintenance tasks are based on deterministic criteria and they
are normally far away of being optimized. Then, one
challenge for the decision-maker is to select the most
suitable set of maintenance strategies with optimal
frequencies based on reliability, availability and cost

criteria, facing in this way a typical multi-objective

RAM+C-based optimization problem (MOP). Ideally,
the resolution of such a problem should yield to the
selection of the optimal set of maintenance strategies, named herein an optimal maintenance plan,
which would cope with the dominant failure causes
of the equipment and at the same time would maximize its reliability and availability keeping costs at a
minimum.
A weakness of the traditional method above is that
the maintenance planning is formulated in terms of
a MOP where RAM+C act as decision criteria and
maintenance strategies (i.e. maintenance tasks intervals) act as the only decision variables. However, the
appropriate development of each maintenance strategy depends not only on the maintenance intervals
but also on the resources available to implement such
strategies under a given period. Therefore, an extended
maintenance planning (see Figure 1) accounting for
maintenance intervals affecting not only maintenance
strategies but also the necessary resources (human
and material) for their implementation seems to be
more realistic in many situations where the resources
available are limited or they need to be minimized
also. Thus, the effect of the necessary resources on
RAM+C needs to be modeled and accounted for in

505

Failure
Cause #i

Maintenance strategy #j
Task #k
Task #1
Task #2

Dominant
Failure Causes

Working conditions

Maintenance strategy #2

Maintenance Plan

H um a n
Re sou r ces

M a ter ia l
Res ou rc es

fs
w
fN

dN

formulating the MOP affecting the set of objectives

and constraints.
In Martorell et al., (2007), new RAM + C models were proposed to address explicitly the effect of
human resources. A sensitivity study was performed
to analysis the effect observed in the cost and unavailability when human resources are taken into account.
This study concluded that the implementation of a
given maintenance plan supposes to adopt both the
optimized frequencies and resources as only adopting
optimized frequencies without considering the human
resources necessary can suppose large deviations from
the unavailability and cost goals presumed at their
optima.
Based on the model previously presented in (Martorell et al., 2007), the objective of this paper is focused
on the extension of this model integrating explicitly the
effect of material resources (spare parts) on RAM+C
criteria. This extended model allows accounting for
explicitly how the above decision criteria depends on
the basic model parameters representing the type of
strategies, maintenance intervals, durations, human
and material resources.
The paper is organized as follow. First, the
model assumptions are presented. Section 2 presents
RAM+C models that include the effect of not only
human but also material resources on the previous criteria. Section 3 shows an application case which is
performed on a motor-driven pump analyzing how the
consideration of human and material resources would
affect the decision-making. Finally, section 4 presents
the concluding remarks.

RAM+C MODELS

A quantitative model needs to be used to assess how

testing and maintenance choices affect equipment
RAMS+C attributes. Thus, the relevant criteria are
to be formulated in terms of the decision variables

HN

NonSchedu led

ur

Availability

Figure 1. Extended maintenance planning within a RCM

framework.

(degradation mechanisms)

Critical Equipment

Failure
Cause #2

Scheduled

Cost

ds

M aintainability
Strategies
(Technical Resources)

Maintenance strategy #1

Maintainability
Human Resources

Reliability

Failure
Cause #1

Hs

Maintainability
Material Resources

Figure 2. Conceptual framework for the RAM+C based

integration of technical, human and material resources.

using appropriate models. These models have to be

able to show explicitly the relationships among the
RAMS+C criteria (decision criteria) and the variables
(e.g. surveillance and maintenance tasks intervals,
human resources, spare parts, ..) of interest involved
for the decision-making (decision variables).
Figure 2 shows the conceptual framework being
considered in this paper for the RAM+C based integration of technical, human and material resources in
support of the extended maintenance planning. This
figure follows the principles introduced in (Martorell
et al., 2005a) what concerns RAMS+C informed
decision-making and the role of testing and maintenance planning.
Next subsections introduce reliability, availability, maintainability and cost models which depend
explicitly so much on the maintenance intervals as
on the human and material resources. In the models developed, the opportunistic maintenance has not
been considered. This approach extends the models
proposed in Martorell et al., (2007).
2.1 Reliability models
The inherent reliability (R) of an equipment depends
on its dominant failure causes, which in turn depend
on its physical characteristics of design and working
conditions, and on the level of quality testing and maintenance tasks imposed on it to cope with such dominant
failure causes (Martorell et al., 2005a).
Thus, working conditions introduce degradation
() as time passes that affects equipment age (w).
This way, degradation imposes loss of equipment performance, which can be represented in terms of a
deterioration of the equipment failure rate (), which
can result in equipment failure when degradation goes
beyond of the point defined in the functional specification. This situation is specific for each degradation
mechanisms and the corresponding dominant failure
cause, the latter representing a sort of interruption of
functional capability of the equipment.

506

On the other hand, testing and maintenance tasks

affect the component age. In general, one can assume
that each testing and maintenance activity improves
the age of the component by some degree, depending
on its effectiveness, what is often called imperfect
maintenance, which is a natural generalization of two
extreme situations (Bad As Old (BAO) and Good As
New (GAN) models).
There exist several models developed to simulate
imperfect maintenance (Martorell et al., 1999b). In
this paper, two models that introduce the improvement
effect of the maintenance depending on an effectiveness parameter are considered. Both models assume
that each maintenance activity reduces the age of
the component in view of the rate of occurrences
of failures. These models are the Proportional Age
Reduction (PAR) represented by
= 0 +

[1 + (1 ) ( f RP 1)]
2f

(1)

and the Proportional Age Setback (PAS), represented by

= 0 +

(2 )

2f

(2)

where RP is the Replacement Period, f is the frecuency

associated to each testing and maintenance task, and
its effectiveness in preventing the equipment from
developing the particular degradation or failure cause.
2.2

Maintainability models

Maintenance represents all activities performed on

equipment in order to assess, maintain or restore
its operational capabilities. Maintenance introduces
two types of positive aspects: a) corrective maintenance restores the operational capability of the failed
or degraded equipment and b) preventive maintenance increases the intrinsic reliability of non-failed
equipment beyond the natural reliability.
On the contrary, maintenance also introduces
adverse effects, called the downtime effect that represents the time the equipment is out of service to
overcome maintenance (corrective, preventive, repair,
overhaul, etc.). Thus, the adverse effect depends on
the Maintainability (M) characteristics of the equipment. Maintainability represents the capability of the
equipment to be maintained under specified conditions during a given period, which depends not only
on the equipment physical characteristics imposing a
given number of man-hours (H) to perform an individual maintenance task, but also on the human and
material resources (e.g. task force, spare parts) available to perform the task, delay times, etc., which

influence the real downtime (d) for developing this

task (see Figure 3).
Therefore, the downtime depends, between other
factors, on the delay associated with the availability
of human (TDH ) and material resources (TDM ). Both
delay times are assumed equal to 0 (TDH = 0 and
TDM = 0) for a scheduled task.
Based on Figure 3, the downtime, d, can be estimated for a maintenance task using the following
relationship:


d = F(ST ) d  + TDH


+ (1 F(ST )) d  + max (TDH ; TDM )
(3)
where F(ST ) is associated with the cumulative distribution function of a random variable x representing
the probability of demand of spare parts, that is,
F(ST ) = P(x ST )

(4)

being ST the inventory associated to a particular spare.

Thus, eqn. (4) represents the probability that a given
amount of spares x required for performing a given
maintenance task are available, so the delay (if any)
must be due to the human delay only, see eqn. (3).
On the other hand, in case of spares shortage; this
occurs with a probability given by:
1 F(ST ) = P(x > ST )

(5)

the delay is given by the maximum value between

human and material delay, max(TDH , TDM ). In addition, d  is the downtime if it is assumed that delay time
is equal to zero, TD = 0, which can be estimated as:
d =

H
(P NP + E NE ) [NP + NE ]

(6)

being NP and NE the number of own and external

personnel involved in this task, respectively. Both,
own personnel and external workforce have an efficiency associated in performing such task, which
is represented by P and E respectively. Function
Task
Launched

Delay
TD

Task
Starts

Man-Hours
H

Task
Ends

TD = max {TDH, TDM}

TDPH

Human Delay

TDM

Figure 3.

507

Material Delay

Schematic view of equipment maintainability.

[.] represents the law of decreasing effectiveness

as human resources which is formulated as follows
(Ricardo, 1821):
 

1
[N ] = exp K 1 +
N

(7)

where K ranges in the interval [0,1].

2.3

Availability models

Each couple dominant failure cause and maintenance

task is associated at least one contribution to the equipment unavailability, which corresponds to one of the
following equations (Martorell et al., 2002).



1 
1
1 eI + I
ur = 1
I
2

(8)

uS = fS dS

(9)

uN = fN dN G[AOT ]

AOT
dN

Two types of costs are considered: a) costs associated

to performing surveillance and maintenance activities,
and b) costs related to the management of spare parts
inventory. In the following subsections these costs are
developed.
2.4.1 Surveillance and maintenance costs
The relevant costs in analyzing test surveillance and
maintenance (TS&M) optimization of safety-related
equipment include the contributions to the cost model
of standby components, which, in general, undertake
surveillance testing, preventive maintenance and corrective maintenance to restore their operability after a
failure has been discovered during a test (Martorell
et al., 2000b, 1996c, 1999). Each couple dominant failure cause and maintenance task is associated
one cost contribution to the equipment total LCC
(Life-Cycle Cost), which corresponds to one of the
following equations in accordance to the unavailability
contributions above (Martorell et al., 2001):
cS = 8760. fS c1S

(12)

cN = 8760. fN c1N

(13)

cD = 8760. fN (1 G[D]) c1D

(14)

(10)

Eqn. (8) represents the unavailability associated

with an undetected failure corresponding to the particular sort of failure cause being considered. In addition,
I represents interval for performing a scheduled maintenance task that is intended or supposed to detect the
occurrence of such a failure.
Eqn. (9) represents the unavailability contribution
associated with a planned or scheduled testing or maintenance task, where fS and dS represent the frequency
and downtime respectively for the scheduled activity,
given by fS = 1/I and eqn. (3) respectively.
Finally, Eqn. (10) represents the unavailability contribution associated with a non-planned or unscheduled maintenance task, where fN and dN represent the
frequency and downtime respectively for the activity,
given by for example fN = for on failure corrective
maintenance and Eqn. (3) respectively. In addition,
function G[AOT] represents the probability of ending
the non-planned activity before the maximum allowed
outage time given by AOT, which can be formulated by
G[AOT ] = 1 e

2.4 Cost models

(11)

where AOT ranges in the interval [0, +[. Factor AOT

accounts for the fact that maintenance tasks at NPP
operating at full power have limited the allowed outage
time, typically to limit the unavailability contribution
due to unscheduled maintenance activities. It is simple
to realize that when this limitation does not apply, then
AOT and consequently G[AOT ] 1.

Eqn. (12) represents a yearly cost contribution as

a consequence of performing planned or scheduled
testing or maintenance task over a year period. Eqn.
(13) represents a yearly cost contribution as a consequence of performing non-planned or unscheduled
maintenance task over a year period. Eqn. (14) represents the yearly cost contribution associated with
the number of plant outages and the corresponding
loss of production that is estimated to occur over a
year time horizon as a consequence of unscheduled
maintenance activities exceeding the allowed downtime D. In addition, variables c1S and c1N in Eqns.
(12) and (13) respectively represent unitary costs as a
consequence of performing one single task, scheduled
or non-planed respectively, which can be formulated
using the following relationship:
cli = NP cHP TP + NE cHE TE

(15)

where TP = d and TE = max{dmin 0, d} usually,

representing the time spent by the NP own and NE
external personnel respectively. In addition, cHE represents the hourly cost for external personnel supposed
to be a fixed amount, while cHP is the hourly cost for
own personnel, which is not constant as it depends

508

on a number of factors as proposed in the following

equation:
cHP =

Neq

NP S

(fA TPA )

(16)

AP

where the unknown variables S and Neq represent

the annual salary and number of similar components
respectively assigned to everyone of the NP own
personnel. The aggregation extends over all of the
tasks, scheduled or not, affecting the equipment being
performed by Np personnel.
Finally, c1D represents the unitary cost due to a plant
outage.
2.4.2 Costs of spare parts
The cost related to the management of spare parts is
assumed to consist of the following contributions:
Ci = ch + co + cex

inventory, depends on the number of failures and

maintenance tasks in this interval L.
In order to obtain this cost contribution, it it required
first to obtain the average inventory level in the period
L considering the demand of spares associated to
failures. Two possibilities are considered, which are
showed in Figures 4 and 5 respectively.
Figure 4 shows a situation in which the demand of
spare parts, xi , in the period L, is lower than the original
inventory level ST .
Figure 5 represents the case in which the demand,
xi , exceeds the original inventory, ST , in the period L.
When the inventory level is equal to reorder point,
R, new spares are ordered. However, in this case,
the ordered spares arrive after the L1 and the original inventory is not enough for the period L. As it is
observed in Figure 5, L1 is the duration in which the
original spare inventories decrease from ST to 0.

(17)

where the inventory holding costs, ch , includes the

capital cost of money tied up in inventory and physical
cost of having the inventory, co includes the regular and
urgent ordering costs and cex . represents the holding
super plus or excess costs.
The following subsections present a detailed
description of the models adopted to represent the different contributions, which have been developed under
the following assumptions:
1. Demand rate of spares associated to failures is equal
to . The value of the parameter is given by Eqn.
(1) or (2) depending on the imperfect maintenance
model, PAS or PAR, selected.
2. Demand of spares due to failures is modeled using
a Poisson process under constant demand rate .
3. Demand of spares associated to surveillance and
maintenance tasks are planned well in advance, so
there is only cost of regular ordering.
4. Period between two consecutives orders of spares
is equal to L.
5. ST R spare parts are ordered when the ordering
point Ris reached.
2.4.2.1 Inventory holding costs
Inventory holding costs include the capital cost of
money tied up in inventory and the physical cost of
having the inventory. The total average holding cost
is often determined as a percentage of the average
total inventory cost (Crespo, 2007). This percentage
can usually not be the same for all types of items,
for example, it should be high for computers due to
obsolescence.
Demand for spares during the period L, and consequently the variation of the level of spares in the

ST
xi
ST-xi-R

ST-xi

L
Figure 4. Case 1: Original inventory is enough to supply
the demand of spare parts during the period L.

ST

xi
R
xi-ST
L2

L1
L

Figure 5. Case 2: The ordered spares arrive after the L1 and

the original inventory is not enough to supply the demand of
spare parts during the period L.

509

2.5 Aggregation of availability and cost models

for a given strategy

Based on Figures 4 and 5, the average inventory

level, Iav , in the period L can be evaluated as:
Iav =


ST

+ R TDM

i=0

(xi + R) P(x = xi )

(18)

Based on the inventory average level, Iav , given

by eqn. (18), the yearly cost contribution due to the
inventory holding, ch , can be evaluated as:


(19)
ch = p Iav csp + cdp
where p is the percentage of the average total inventory
cost considered, and csp and cdp represent, respectively,
the cost of a spare part and the depreciate cost per spare
part.
2.4.2.2 Inventory ordering costs
The yearly cost contribution due to the inventory ordering, cio , includes: a) costs for a regular order, cro ,
b) cost of expedited order, ceo , and c) urgency or
emergency order cost,cuo ., which can be formulated
as follows:
co =

(ceo + cro + cuo )

ST

(20)

where ceo is a fixed ordering cost for order, cro is the

cost for a regular order, which can be calculated as:
cro = csp ST

(21)

and cuo is the emergency order cost, which is calculated

(see Figure 5) as:
cuo = cu

(xi R) P(xTDM = xi )

(22)

i=R

being cu the emergency cost per spare part and xTDM

the demand of spare part during the delay associated
to material resources.
2.4.2.3 Holding excess costs
Based on Figure 4, the average surplus stock cost can
be evaluated it xi < (ST R) as:
cex = (cop cdp )

ST


((ST R) xi ) P(x = xi )

i=0

(23)
where cop is the opportunity cost per spare part, as a
consequence of that the capital invested in its purchase
is not available for other uses, and cdp represents the
component depreciate cost.

As proposed in Martorell et al., (2007), under the

RCM approach, each couple dominant failure cause
and maintenance strategy is associated a global efficiency affecting partially equipment availability and
associated costs, which are associated with both the
probability that this failure occurs and the development of the strategy itself, which, in turn, depends on
the frequency of developing scheduled and unscheduled tasks belonging to the maintenance strategy and
their corresponding durations and costs. Thus, the
unavailability and cost models for a given strategy can
be formulated accordingly by simply aggregating the
previous single task models for the k tasks involved
in j-nth strategy used to cope with i-nth failure cause
(i j).
Uij =

Uij, k

(24)

Cij, k

(25)

kj

Cij =


kj

Eqns. (24) and (25) are similar to their equivalents

in Martorell et al., (2007). However, notice the sort
contributions and their formulation vary as follows.
What concern unavailability contributions, the
main difference comes from the novel formulation of
the duration of the a maintenance task to cope with
a given dominant failure cause, see eqns. (3) to (7),
which now addresses also the possibility of not having
spare parts available at the time of a demand, in particular for performing an unscheduled task (e.g. corrective
maintenance), which introduces a delay time in starting the task while one is waiting for a spare part that
has been ordered urgently.
What concern cost contributions, one important difference comes also from the novel formulation of the
duration of the maintenance task to cope with a given
dominant failure cause. Moreover, the novel formulation of the cost for a given couple (i j) addresses
now additional cost contributions, i.e. those contributions used to address the cost of managing spare parts
as described in section 2.4.2, which are also formulated for a given couple (i j). The remaining cost
contributions, i.e. those associated with surveillance
and maintenance costs, as described in section 2.4.1,
keep almost the same as formulated in Martorell et al.,
(2007), except what concerns eqn. (15) in this paper,
where an additional third constant term appeared in
the right hand side of this equation, which was used to
account for the cost of the spares used in a given tasks.
This additional term does not make sense now, as this
contribution was included to address in a simplified

510

way the cost of managing spare parts, which however

is well represented now with the extended modeling
of section 2.4.2.
2.6

Equipment based aggregation of availability

and cost models

Following the reasoning introduced above, one can

realize there is a need to find out a set of maintenance
strategies to prevent the component from developing
all of its dominant failure causes since more than one
normally applies.
According to the study in Martorell et al., (2007)
there is not a unique combination of maintenance
strategies to cope with all the dominant failure causes.
Each combination is associated a given equipment
availability and corresponding cost given by
U=

study after performing the RCM process described in

Martorell et al., (1995). In addition, Table 2 shows the
maintenance plan selected in Martorell et al., (2005b),
which allows covering all the dominant failure causes
of the equipment. This table shows the associated
surveillance and maintenance tasks belonging to a type
of strategy identified to be appropriate (Y) or not (N)
to control every cause and the maintenance intervals
(I ), originally optimized.
Tables 3 and 4 show the additional data necessary
for using the models proposed in section 2. Table 3
shows the data related to human resources. Table 4
shows data related to material resources. For sake of
Table 2. Maintenance plan selected to cover dominant
failures causes.
Task

uij

I (hrs)

ij

C=

cij

(27)

ij

APPLICATION EXAMPLE

Lub oil change (t1)

Operational test (t2)
Visual inspection
Motor (t3)
Visual inspection
Pump (t4)

The case of application is performed on the basis of the

results obtained in Martorell et al., (2005b), where the
best maintenance plan to cope with all the dominat failure causes of a motor-driven pump was found. Later,
in Martorell et al., (2007) it was studied how management of human resources affects decision-making
on the optimal frequency for performing the scheduled tasks included in the above maintenance plan.
The objective of the application example in this paper
is on analyzing how management of spare parts as
well as management of human resources affects this
decision-making using the extended models of equipment unavailability and cost contributions presented
in section 2.
For sake of completeness of the presentation,
Table 1 shows again the six dominant failure causes
considered for the motor-driven pump analyzed in this

Table 3.

Motor-driven pump dominant failure causes.

# Cause

Code

Description

c1
c2

IAL
DEM

c3
c4
c5
c6

MBW
PIW
SPD
MIS

Inadequate lubrication
Damaged electric
or electronic module
Motor bearing wear
Pump impeller wear
Set point drift
Misalignment

511

c1

c2

c3

c4

c5

c6

26000
13000

Y
Y

N
N

Y
N

N
N

N
Y

N
N

26000

26000

Data related to human resources.

Parameter
Efficiency
Delay for
unscheduled tasks
Delay for
scheduled tasks
Cost
Neq
Persons (N)
K (law of decreasing
effectiveness)

Table 4.
Table 1.

Failure causes

(26)

Own
personnel

External
personnel

0.9
0h

1
3h

0h

0h

20000 C
= /year
100
[0, 1, 2, 3]

30 C
= /hour

[0, 1, 2, 3]
0.25

Data relating to material resources {c4,t4}.

Parameter

Value

RP
Spare part cost, csp
Emergency cost per order, cu
Fixed ordering cost, cfo
Percentage p
TDM
Reorder point R
Opportunity cost, cop
Depreciate cost, cdp

40 years
1328 C
=
1992 C
=
132 C
=
20%
720
1
66 C
=
33C
= /year

U -C plot for several strategies of human us stocks

0,0515
(1,0,0)

Without stocks
Without Reorder point
Reorder point (R)

0,0510
0,0505

Unavailability

0,0500
0,0495
0,0490

(1,0,1)

(1,0,1)R
(Np, Ne, Stock)

(2,0,0)
(1,0,2)R
(1,2,0) (1,0,3)R
(1,3,0)
(1,4,0)

(1,1,1)
(3,4,0)

(2,4,0)
0,0485

(1,1,2)R
(1,1,3)R

0,0480
0,0475

(4,4,0)

(1,2,1)
(1,1,2)
(1,2,2)
(1,2,3)R
(1,0,3)
(1,2,4)R
(2,2,3)
(2,2,4)R
(3,2,4)R

(3,2,3)

(4,2,3)

(3,2,4)

(4,2,4)

(4,2,4)R

0,0470
0,0465
2000

4000

6000

8000

10000

12000

14000

16000

Cost

Figure 6.

Unavailability and cost effects of using different couples [NP, NE].

simplicity, it has been considered herein that task 4,

which covers dominant failure cause 4 as shown in
Table 3, is the only one that requires spare parts. Data
related to the equipment reliability characteristic and
others not included herein are the same as proposed in
Ref. Muoz et al., (1997).
As said, the results obtained in Martorell et al.,
(2005b) have been adopted as a reference point. Next,
a sensitivity study has been performed to analysis the
effect observed on the unavailability and cost scores
under several when the human and material resources
are taken into account. Several management policies
have been considered, which address own versus external personnel and spare part inventory only for sake of
simplicity.
Figure 6 shows the effect of changing the human
and material resources, i.e. use of different triplets
[NP , NE , ST ] representing own personnel, external
taskforce and spare parts inventory respectively, for
performing the maintenance plan selected with the
periods showed in Table 2. Three alternatives were
considered: a) without stocks, b) with stocks but without reorder point and c) with stocks and reorder point.
Consequently, three curves were found representing
each a non-dominated set of solutions in the space
U-C for the corresponding alternative.
Figure 6 shows the alternative with the lowest
costs corresponds to the case of having no inventory.
Nevertheless, this situation is the one that imposes the

highest unavailability. On the contrary, when a stock of

spare parts is considered it results in a decrease of the
unavailability at expenses of an increase of the costs.
Comparing two alternatives, with or without
reorder point, the former provides better results in
terms of both unavailability and cost, which may
suggest that management of spare parts with reordering point dominates equivalent solutions without
reordering.
4

CONCLUDING REMARKS

This paper proposes the extension of previous models

developed by the authors to integrating explicitly the
effect of material resources (spare parts) on RAM+C.
This novel modeling allows accounting for explicitly
how the above decision criteria depends on the basic
model parameters representing the type of strategies,
maintenance intervals, durations, human resources
and material resources.
An application example is performed on a motordriven pump analyzing how the consideration of
human and material resources affects the decisionmaking. It shows how changes in managing human and
material resources affect both cost and unavailability.
It is observed also, unavailability can be reduced by
introducing a spare inventory, although, logically, this
option supposes a greater cost. Finally, management

512

of spare parts with reordering point provides better

results than without reordering.
ACKNOWLEDGMENTS
Authors are grateful to the Spanish Ministry of Education and Science for the financial support of this work
(Research Project ENE2006-15464-C02-01, which
has partial financial support from the FEDER funds
of the European Union).

REFERENCES
Axster, S. 2006. Inventory Control. Springer, United States
of America.
Crespo, A. 2007. The maintenance management frame work:
models and methods for complex systems maintenance.
Springer series in reliability engineering.
Kaufmann, A. 1981. Mtodos y Modelos Investigacin de
Operaciones. CIA. Editoral Continental, S.A. de C.V,
Mexico.
Leland T. Blank, Anthony Tarquin. 2004. Engineering
Economy. McGraw-Hill Professional. United States.
Martorell, S., Muoz, A., Serradell V. 1995. An approach
to integrating surveillance and maintenance tasks to prevent the dominant failure causes of critical components.
Reliability engineering and systema safety, Vol. 50,
179187.
Martorell S., Sanchez A., Serradell V. 1999. Age-dependent
reliability model considering effects of maintenance and
working conditions. Reliability Engineering & System
Safety, 64(1):1931
Martorell, S., Carlos, S., Sanchez, A., Serradell, V.
2001. Simultaneous and multi-criteria optimization of
TS requirements and maintenance at NPPs. Annals of
Nuclear Energy, Vol. 29, 147168.

Martorell, S., Sanchez, A., Carlos, S., Serradell, V. 2004.

Alternatives and challenges in optimizing industrial safety
using genetic algorithms. Rel. Engng & System Safety,
Vol. 86 (1) 2538.
Martorell, S., Villanueva, J. F., Carlos, S., Nebot, Y.,
Snchez, A., Pitarch J. L. and Serradell, V. 2005a
RAMS+C informed decision-making with application to
multi-objective optimization of technical specifications
and maintenance using genetic algorithms. Reliability
Engineering & System Safety, 87(1):6575.
Martorell, S., Carlos, S., Sanchez, A. 2005b. Use of metrics with multi-objective GA. Application to the selection
of an optimal maintenance strategy in the RCM context.
In Proceedings of European Safety and Reliability
Conference ESREL 2005. Ed. Tailor & Francis Group,
pp. 13571362.
Martorell, S., Carlos, S., Sanchez, A. 2006. Genetic
Algorithm applications in surveillance and maintenance
optimization. In Computational Intelligence in Reliability Engineering. Volume 1: Evolutionary techniques in
reliability analysis and optimization Ed. Springer.
Martorell, S., Villamizar, M., Sanchez, A., Clemente G. 2007.
Maintenance modeling and optimization integrating
strategies and human resources. European Safety
And Reliability Conference (ESREL 2007). Stavanger,
Noruega.
Muoz, A., Martorell, S., Serradell V. 1997. Genetic algorithms in optimizing surveillance and maintenance of of
components. Reliability engineering and system safety,
Vol. 57, 107120.
Ribaya, F. 1999. Costes. Ediciones Encuentro, Madrid.
Ricardo, D. 1821. On the principles of Political Economy and
Taxation. John Murray, Albemale-Street, London.
Sarabiano, A. 1996. La Investigacin Operativa. Edisorfer,
S. L, Madrid.
Sherbrooke, C. 2004. Optimal Inventory Modeling Of Modeling Of Systems, United States of America.
Starr, M. y D. Miller. 1962. Inventory Control: Theory and
Practice,Prentice-H all, Englewood Cliffs, United States
of America.

513

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Modelling competing risks and opportunistic maintenance

with expert judgement
T. Bedford & B.M. Alkali
Department of Management Science, Strathclyde University, Graham Hills Building, Glasgow, UK

ABSTRACT: We look the use of expert judgement to parameterize a model for degradation, maintenance
and repair by providing detailed information which is then calibrated at a higher level through course plant
data. Equipment degradation provides signals by which inferences are made about the system state. These may
be used informally through the use of red/yellow/green judgements, or may be based on clear criteria from
monitoring. Information from these signals informs the choices made about when opportunities for inspection
or repair are taken up. We propose a stochastic decision model that can be used for two purposes: a) to gain
an understanding of the data censoring processes, and b) to provide a tool that could be used to assess whether
maintenance opportunities should be taken or whether they should be put off to a following opportunity or
scheduled maintenance. The paper discusses the competing risk and opportunistic maintenance modeling with
expert judgement and the broad features of the model. Numerical examples are given to illustrate how the process
works. This work is part of a larger study of power plant coal mills reliability.

INTRODUCTION

In this paper we take a highly (subjective) decisionoriented viewpoint in setting up a model to capture
the interaction between system degradation and opportunistic maintenance/inspection. This model is primarily a decision model, rather than a reliability model,
although reliability provides the context. The model
considers the uncertainty about the state of the system
from the maintainer/operator point of view, and models how information gained about the system (which
includes the usage time, but also other pieces of
information) will change that uncertainty.
The model we have is a complicated stochastic
model that we have to simulate because of the lack of
convenient closed form solution, but it is designed to
fit closely to an expert judgement elicitation process,
and hence be relatively easy to quantify through a combination of expert judgement and plant data. Because
of the difficulties of using messy plant data, we give
preference to a combination of expert judgement and
plant data, as the most cost efficient way of quantifying
the model. In other words, the model quantification
process uses expert information to give fine details
of the model, while using plant data to calibrate that
expert data against actual experience. We have used
a strategy of reducing the difficulty of expert judgements by eliciting relative risks as we believe these to
be easier to elicit than absolute risks.

The motivation for looking at this model was the

aim to produce realistic competing risk models, that
capture the way in which failure events are censored by
other events, in particular by maintenance. A number
of competing risk models have been produced with the
reliability area in mind. However they have not adequately captured the fact that much censoring occurs
due to opportunistic maintenance interventions.
Finally, we remark that because the model is supposed to capture the typical behaviour of the system,
we shall not worry too much about describing the tailbehaviour of lifetime distributions which are outside
the regime of practical interest. In practice changes
to system decision making are only ever going to be
incremental and therefore they allow for the revision
of the model further into higher time values if that
becomes necessary.
1.1

Background

The background to this paper is a project studying the

reliability of power-generating plant coal mills. This
equipment is heavily used and must be maintained
regularly. The coal mills accumulate high operating
hours and are necessary for the coal fired boilers in
the power plant generating units. Maintenance activities are planned to a fixed number of operating hours,
but regular unscheduled maintenance is also required
due to rapid degradation. Due to the high costs of lost

515

production, there is a real need to avoid unplanned

shutdowns. However, there are regular opportunities
arising from the failure of systems elsewhere in the
plant. Furthermore, there is a level of redundancy
in the coal mills which may be used when required.
Hence it is the case that equipment in need of repair
and operating in a degraded state may continue to be
used until an opportunity presents itself. The model
described here is designed to capture this type of
situation and to be quantifiable by expert judgements.
1.2

Competing risks

The competing risk problem arises quite naturally

in coal mill reliability. An intuitive way of describing a competing risks situation with k risks, is to
assume that to each risk is associated a failure time
Tj , j = 1, . . ., k. These k times are thought of as the
failure times if the other risks were not present, or
equivalently as the latent failure time arising from each
risk. When all the risks are present, the observed time
to failure of the system is the smallest of these failure times along with the actual cause of failure. For
further discussion of competing risk models in general see ..Tsiatis (1975) and Crowder (2001). Bedford
(2005) also discussed the use of competing risk models in reliability. Specific models are considered in
.Langseth & Lindqvist (2006), .Bunea & Mazzuchi
(2006), .Cooke & Morales-Mapoles (2006). Yann et
al., (2007) introduce a generalized competing risk
model and use it to model a particular case of for which
potential time to next preventive maintenance and corrective maintenance are independent conditionally to
the past of maintenance processes. For more specific
applications in reliability see the surveys of reliability database in perspective see Cooke and Bedford
(2002).
1.3 Opportunistic maintenance
Opportunistic maintenance can be defined as a strategy that preventive maintenance actions on a component can be performed at anytime by other components failure or arrival of preventive replacement
ages of designated components (separate replacement)
or the joint replacement, Kececioglu & Sun (1995).
Opportunistic maintenance can clearly have impacts
on component and hence on system reliability. Opportunities, interrupt replacement options and many undesirable consequences of interruptions are discussed
in Dagpunar (1996). Opportunities occur randomly
and sometime have restricted durations, implying that
only restricted packages can be executed. The main
idea is to set up a model to determine an optimal
package for individual packages and to develop cost
criterion, Dekker & Smeitink (1994). Maintenance

opportunities often depend on the duration and economic dependence for set up costs which may require
a compromise in some circumstances.
The concept of opportunistic maintenance and competing risk modeling is important within an industrial
power plant. The role of expert judgments within
the maintenance area may give an insight and better understanding of the interrelationships between
events which could strongly support our modeling
later. Expert judgment in maintenance optimization
is discussed in Noortwijk, et al. (1992). Their elicited
information is based on discretized lifetime distribution from different experts. That is, they performed a
straightforward elicitation of failure distribution quantiles, rather than the kind of information being sought
for the model we build here.

BASE MODEL

Here we consider the basic model discussed in Alkali

& Bedford (2007). This model explicitly assumes that
some kind of information about the state of the system is known to the decision-maker. The basic idea
behind the model is that the decision makers perceptions about the state of the system is driven by discrete
events whose occurrence changes the underlying failure rate of the system. Such discrete events might be
external shocks, or might be internal shocks as a result
of internal degradation of the system. These shocks
occur randomly but can be observed by the decision
maker. We discuss the single failure rate and multiple
failure rate case.
2.1

Single failure mode model

In this model only one category of failure mode is

considered. The failure time of the model requires
us to define the failure time distribution through a
process rather than through the more conventional failure rate/survival functions approach, although these
quantities could be calculated.
We define an increasing sequence of shock times
S0 = 0, S1 , S2 , . . . and for the period between shocks
(Si1 , Si ], a failure rate i (t), t (0, Si Si1 ]. The
distribution of failure time T is defined conditionally:
Given that T > Si1 the probability that T > Si is
equal to

exp

Si
Si1

i (t)dt .

(1)

and given that t (0, Si Si1 ] it has conditional

failure rate i (t Si1 ).

516

Clearly, given S0 = 0, S1 , S2 , . . . the failure rate for

T is fixed (deterministic), but prior to knowing S0 =
0, S1 , S2 , . . . the failure rate is stochastic.
This definition of a failure time, while relatively
unusual, has the advantage that it links fairly easily
to an elicitation scheme that can be used to assess a
simplified model using expert judgement. The simple parametric version of this model assumes that
there are a finite number of shocks that the times
between shocks are exponential (possibly with different parameters) and the failure rate in between shocks
is constant. Hence the parameters of this model are the
number of shocks n the mean times between shocks
1 , 1 , . . . , n and the failure rate between shocks
1 , 1 , . . . , n (where S1 is the ith shock and i is the
rate in the period (Si1 Si ], and Sn+1 = ). The
expression for the conditional distribution function
given the shock times is given as
F (t|S1 , S2 , . . . . . . , Sn ) = 1 e

j1


2.3 Competing risk data

The model above will generate competing risk data as
follows. Let X be the time of the system failure, and let
Z be the time at which an opportunity would be taken.
The data that would be recorded in the plant records
would be min(X , Z) and the indicator of which event
had occurred. Note that as a consequence of the above
model, and consistent with practice, we can have that
P(X = Z) = 0.
From the plant record data we can estimate only
the subsurvival function for X and Z, but not
the true survival functions. Recall that the survival
functions are
SX (t) = P(X > t), SZ (t) = P(Z > t),

(3)

and the subsurvival functions are

SX (t) = P(X > t, X < Z), SZ (t) = P(Z > t, Z < X )
(4)

(Si+1 Si )i +(tSj )j

i=1

(2)
where i is the failure rate after shocks i, and Sj is the
largest event time less than t.
2.2

A simple quantity that measures the degree of censoring is SX (0) = P(X < Z). This is simply the
probability that the next event will be a failure rather
than a maintenance action.

Opportunistic maintenance time

In order to model the censoring process, we need to

model two things:
how opportunities arise, and
when opportunities are taken.
In the setting we are considering, opportunities arise
typically as a result of faults elsewhere in the plant
typically upstream. As there is a whole mixture of
different fault types occurring, it is natural to assume
that opportunities arise according to a Poisson process.
The modelling of when opportunities are taken is the
area where we can capture the idea that opportunities will be taken with some knowledge of the system
state, and hence that the censoring of failure data will
be correlated to the actual failure times. In this model
we assume that the decision maker will choose to take
the first opportunity after the system has encountered a
critical shock, or been assessed as in a critical state. We
denote this time, the time at which the system would be
maintained as Z. Note that Z can be equivalently written as the first opportunity after the subjective failure
rate has increased above a critical level. This is probably the way the decision makers will be thinking of Z.
However, so as not to confuse the definition of the
model parameters with steps taken in the elicitation
process (where experts are asked for ratios of failure rates, rather than absolute failure rates) we define
Z in terms of a critical shock rather than a critical
failure rate.

2.4 Multiple failure modes

Very often there is more than one failure mode relevant to system performance. In terms of competing
risk, these may be modelled as different variables that
are competing themselves to fail the system. For the
purposes of this paper we assume that there are just
two failure modes. In many cases it seems reasonable
to suggest that different failure modes develop according to statistically independent processes, but as they
develop further there may be some dependence. We
capture these ideas in the following model. Associated to each failure mode there is a set of shock times
S0 = 0, S1 , S2 , . . . . We denote the ith shock associated
to the kth failure mode by Sik and assume that there
are m(k)shocks for the kth mode.
The associated failure rate given that we have just
had the ith shock for Failure Mode 1 and the jth shock
for Failure Mode 2 is denoted byi,j .
Theorem 1.1 Suppose for a given system there are
two modes in which the system can fail. If the two
shock time processes are independent, and i,j can be
written in the form i,j = i + j then the failure times
X1 and X2 from failure modes 1 and 2 respectively, are
independent.
Sketch Proof : Consider the conditional survival joint
probability given the shock times,

517



P X1 > t, X2 > t|S01,... Sm(1)1 , S02,... Sm(2)2

(5)

and by splitting the terms for the shocks the resulting

distribution is given as
P(X1 > t|S01,... Sm(1)1 ).P(X2 > t|S02,... Sm(2)2 )

(6)

Hence the conditional survival probability factorizes because of the additivity assumption on i,j .
Under the assumption that the two shock processes are
independent, we can then say that the unconditional
survival probabilities also factorize.
Hence when the two failure mode processes are
considered to be truly independent, then they can be
modelled as two different cases of the base model, and
the failure rates added together. However, when they
are not considered independent then can capture this
in one of two ways:
1. The shock time processes are dependent
2. The failure intensities are not additive.
The simplest way in which the shock processes
could be dependent is for there to be common shocks
for both failure modes. For the purposes of this paper
we shall not consider more complex forms of dependency between the shock processes. Regarding failure
intensities, we would typically expect that the failure intensities are additive for early shocks, and then
may become superadditive for late shocks, that is
i,j > i + j
Remark 1.1 When there is more than one failure
mode, there is a modelling issue relating to the meaning of the failure intensities. While in the single failure
mode model we can simply consider the failure intensity to cover failures arising from any reason, when
there are two or more failure modes we have to distinguish failure intensities arising from the different
failure modes. This will avoid double counting of any
residual failure intensity not ascribable to those two
failure modes, but also ends up not counting it at all.
Therefore in this case, if there are significant failure
intensities from residual failure causes, then it is best
to explicitly assess these alongside the main failure
modes, so that the failure intensities can be properly
added.

ELICITATION

The purpose of using this modelling approach is that it

makes discussion with experts easier. However, there
is always a problem with obtaining good expert data,
and some parts of the model will be easier to assess
than others. In particular, one would expect

Straightforward: identification of the failure modes,

identification of signals (note that is often done
informally through use of green amber red codings,
but may not be thought of as a signal in the sense
that we are using it).
Moderate difficulty: quantification of distributions
for times to shocks, specification of relative risk
(i.e. ratios of the i,j ).
Very difficult: specification of absolute values for
the i,j .
The elicitation procedures and model quantification/validation steps have to be designed to use information at the least difficulty level possible. Clearly,
the model does require information about the absolute values of the i,j , but we can also use empirical
data to calibrate the model. This represents something
of a departure from previous competing risk models
that we have considered, where the aim was usually to
check identifiability of the modeli.e. whether or not
the parameters could be estimated from the data.
3.1 Single model
Here we assess the set of signals, the mean times to
signals, and the relative increase in risk after a signal.
3.2 Two failure modes
The model assumes that, by default, the failure intensities arising from different failure modes are additive.
Hence they can be elicited in a first round according the procedure used for the single mode situation.
However, to provide a check on interactions, we check
whether there is a critical signal level after which we
could expect the FM of the other to be affected, or
if there is a common signal. Then we explicitly elicit
relative risk values above the critical signal levels.
4

MODEL CALIBRATION

For simplicity, from now on we just consider the single

failure mode case, but the other case works similarly.
Since the elicitation steps above only give the failure
rates up to an unknown constant, we have to calibrate
the overall model. Suppose that the relative failure
rates elicited from the experts are 1 , . . . , n , so that
the actual failure rates are of the form 1 , . . . , n .
The following result allows us to consider the impact
of the calibration variable :
Theorem 1.2 The marginal distribution for X , is
stochastically decreasing as a function of .
Proof.
It is enough to show that the probability of
failure after time t is increasing as a function of . In
turn, for this, it is enough to show that the conditional

518

probability of failure after time t given the shock times,

is increasing. However, that is clear since this is

exp .

j1



(Si+1 Si )i + (t Sj )j

i=1

where j is the largest integer such that Sj < t.

In practice we can use this result to scale model outputs to the observed failure data. However, to do that
we also need to take into account the censoring taking
place in the observed data. What we actually see in
the observed data is the effect of current opportunistic
maintenance decisions.
Theorem 1.3 If the opportunistic maintenance intervention shock level is held constant, then the subdistribution function evaluated at any point tis increasing as
a function of . In particular, the probability of observing a failure, P(X < Z), is increasing as a function of .

first is 0.2 and to the second is 0.8. The distribution

of time between shocks is modelled as exponential.
There are three failure rates associated to the periods
demarked by the two shocks. We suppose that the failure rate ratios are estimated as increasing by a factor
of 10 each time, that is, the failure rate in the second period is 10 times the initial rate, and that in the
third period is 10 times that in the second. Finally, we
assume that the mean time between opportunties is 1.2,
and that there is a major overhaul after time 10 (so that
no times observed will ever be larger than 10 in any
case). Finally, we assume that the current opportunistic maintenance strategy is to take the first opportunity
that arises after the second shock.
Because we do not have a closed form solution to
the model, we simulate it as follows:
1. Simulate the shock times
2. Calculate the conditional distribution function for
the lifetime given the shock times as in Equation
2, and simulate a lifetime.
3. Simulate a set of opportunity times and then choose
the first one beyond the critical shock time, as the
time at which opportunistic maintenance would be
carried out.

P(X = t, t < Z|S1 , S2 , . . . , Sn )

= P(X = t|S1 , S2 , . . . , Sn , t < Z)
P(t < Z|S1 , S2 , . . . , Sn )
= P(X = t|S1 , S2 , . . . , Sn )P(t < Z|S1 , S2 , . . . , Sn ).
The proof is similar to that given above, using the
fact that the opportunistic maintenance time is independent of the failure time, given the shock times, for
The first term is increasing in while the second is
constant in . The above result allows us to use the
empirically observable quantity P(X < Z) as a way
of calibrating the model by finding the appropriate .
Although we have concentrated here on trying to
establish a value for the scaling parameter , it is
worth also looking at other parameters. In particular,
although we assumed that the experts were able to give
mean times until shocks, and the rate of opportunities,
it is possible to see whether a scaling adjustment of
these values would improve the overall fit of the model
to observed data. These are, however, secondary tuning parameters that should only be considered after
the has been fitted. This fine tuning is then carried
out using a different distance quantityfor example
using Kolmogorov-Smirnov distance on the observed
and model-based subdistribution functions.
5

We sampled 1000 cases (in excel) and used these

to numerically estimate the quantities of interest such
as the failure distributions and the probability of
observing a failure.
Assuming that the scaling variable = 1, we get the
following distributions for the underlying failure time,
maintenance time, and their minimum, see Figure 1.
When we increase the scaling parameter to 1.5, this
increases the overall failure rate, thus making it more
likely that a failure will be observed. This is illustrated
in Figure 2.
When we reduce the scaling parameter to 0.5, this
reduces the overall failure rate, making it less likely
that a failure will be observed, as illustrated in Figure 3
below.
1.2
1
Taken
opportunity

0.8

Failure time

0.6

Minimum

0.4

RESULTS

0.2

We illustrate the results of the method proposed here

by using simulated data.
The example supposes that there are two shocks
that occur in the system, and that the mean time to the

0
0

Figure 1.

519

10

Failure time distribution with = 1.

difficult to quantify from plant data both because of

the cost of cleansing the data sufficiently to make it
amenable to statistical analysis, and also because
due to identifiability problemsit may not be possible
to characterize a unique model in any case.
The model discussed here is designed to bring the
mathematical modelling closer in to line with the way
plant operators and maintainers adjust their beliefs
about the reliability of the equipment. Although we
have described the model in terms of shocks occurring
to the system, in practice these may not be hard discrete
events, but the progression of degradation past a standardized level (for example, where a vibration monitor
consistently measures high vibration, or where the criteria used by the operator to move the equipment state
from green to amber state have been met). Such transitions are the ways in which staff assess a change in the
risk of system failure, and therefore it is reasonable
to build the subjective component reliability around
them.
The model described here is a dependent competing
risk model where maintenance times are statistically
dependent on the component lifetime, and where
different failure modes can also be modelled in a
dependent way. In the former case the dependence
arises through the use of signals to the decision maker
about the component state, which both change the
rate of failure and which also lead to opportunities for
maintenance being taken. In the latter case the dependency arises through the same signals marking failure
rate changes for distinct failure modes, and through
other potential interactions.

1.2
1
Taken
opportunity

0.8

Failure time

0.6

Minimum

0.4
0.2
0
0

Figure 2.

10

Failure time distribution with = 1.5.

1.2
1
Taken
opportunity

0.8

Failure time

0.6

Minimum

0.4
0.2
0
0

Figure 3.

10

15

20

Failure time distribution with = 0.5.

Table 1.
values.

Probability of failure with varying

ACKNOWLEDGEMENTS

alpha

P(X < Z)

0.4
0.6
0.8
1
1.2
1.4
1.6

0.230
0.256
0.313
0.346
0.402
0.470
0.599

This research is part of an EPSRC funded project on

dependent competing risks. We would like to thank
Scottish Power for its support in this project.

REFERENCES

Finally we give a table showing how the probability of observing failure depends on the scaling
parameter .
This confirms empirically the theoretical result
given above, and shows that the scaling parameter is a
first order model parameter.
6

CONCLUSIONS

Expert judgement has an important role to play in the

development of fine detail models. Such models are

Alkali, B.M. and T. Bedford. 2007. Competing Risks and

Reliability Assessment of Power plant Equipment. 17th
Advances in Risk and Reliability Symposium (AR2 TS)
17th19th April. Loughborough University.
Bedford, T. 2005. Competing risk modelling in relability, in Modern Mathematical and Statistical Methods
in Reliability, Series on Quality Reliability and Engineering Statistics, Vol 10. Eds A. Wilson, N. Limnios,
S. Keller-McNulty, Y. Armijo, CRC Press.
Bunea, C. and T. Mazzuchi. 2006. Competing Failure Modes
in Accelerated life Testing. Journal of Statistical Planning
and Inference 136(5): 16081620.
Cooke, R. and O. Morales-Mapoles. 2006. Competing Risk
and the Cox Proportional Hazards Model. Journal of
Statistical Planning and Inference 136(5): 16211637.

520

Crowder, N. 2001. Classical Competing Risks. Chapman &

Hall: Boca Raton
Dagpunar, J.S. 1996. A maintenance model with opportunities and interrupt replacement options. Journal of the
Operational Research Society 47(11): 14061409.
Dekker, R. and E. Smeitink. 1994. Preventive Maintenance
at Opportunities of Restricted Durations. Naval Research
Logistics 41: 335353.
French, S. 1986. Calibration and the Expert Problem.
Management Science 32(3): 315321.
Kececioglu, D. and F.B. Sun. 1995. A General Discrete-Time
Dynamic-Programming Model For The Opportunistic
Replacement Policy And Its Application To Ball-Bearing
Systems. Reliability Engineering & System Safety 47(3):
175185.

Langseth, H. and B. Lindqvist. 2006. Competing Risks for

Repairable Systems: A Data Study. Journal of Statistical
Planning and Inference 136(5): 16871700.
Meyer, M.A. and J.M. Booker. 2001. Eliciting and Analyzing
Expert Judgment: A practical guide, SIAM.
Noortwijk, V., J. Dekker, R. Cook and T. Mazzuchi. 1992.
Expert judgement in maintenance optimization. IEEE
Transaction on Reliability 41(3): 427432.
Tsiatis, A., 1975. A nonidentifiabilty aspect of the problem
of competing risks. Proceedings of the National Academy
of Sciences of the USA 72(1): 2022.
Yann Dijoux, L. Doyen and O. Gaudoin., 2007. Conditionally
Independent Generalized Competing Risks for Maintenance Analysis. 5th Mathematical Methods in Reliability
Conference, Strathclyde University, Glasgow Scotland.

521

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Modelling different types of failure and residual life estimation

for condition-based maintenance
Matthew J. Carr & Wenbin Wang
CORAS, Salford Business School, University of Salford, UK

ABSTRACT: Probabilistic stochastic filtering is an established method for condition-based maintenance

activities. A recursive filter can be used to estimate the time that remains before a component, or part, fails
using information obtained from condition monitoring processes. Many monitored processes exhibit specific
behavioural patterns that typically result in specific types of failure. These behavioural patterns are often identifiable using historical data and are defined here, for future monitoring purposes, as potential failure patterns
that result in a specific type of failure.
This paper presents a case application of a model designed to predict the forthcoming failure type using stochastic filtering theory. Firstly, a stochastic filtering model is developed for each potential failure type. Secondly,
a Bayesian model is used to recursively assess the probability that an observed behavioural pattern for a given
component, or part, will result in a specific type of failure. At monitoring points during the operational life of
a component, estimation of the residual life is undertaken using a weighted combination of the output from the
individual stochastic filters. The model is applied to an oil-based condition monitoring data set obtained from
aircraft engines and the benefits of modelling the individual failure types are assessed.
1

INTRODUCTION

Condition-based maintenance (CBM) involves the

scheduling of maintenance and replacement activities using information obtained from one or more
condition monitoring (CM) processes over time. As
with many reliability applications, we are interested
in modelling the time remaining before failure, or
residual life (RL) of a component, part or piece of
machinery. Relevant prognostic techniques include;
proportional hazards models (see Makis & Jardine
(1991) and Banjevic & Jardine (2006)), proportional
intensity models (see Vlok et al., (2004)), neural
networks (see Zhang & Ganesan (1997)) and stochastic filters (see Wang & Christer (2000) and Wang
(2002)).
This paper documents research into the categorisation and modelling of different CM behavioural
patterns that typically result in specific types of failure.
The investigation is an applied version of a methodology originally presented in Carr & Wang (2008b) for
a single vibration CM parameter that is extended and
modified here to cater for multivariate oil-based CM
information.
At monitoring points throughout the life of a component and under the influence of a given failure
type, probabilistic stochastic filtering (see Carr &
Wang (2008a)) is used to model the conditional RL
using the multivariate CM information. A Bayesian

model is used to calculate the probability associated

with each individual failure type occurring and weight
the RL output from the individual stochastic filters
accordingly.

NOTATION

To describe the RL estimation process for a given

component, the following notation is introduced;
i = 1, 2, . . . is an index of the CM points,
ti is the time of the ith CM point, where t0 = 0,
M {1, 2, . . ., r} is a discrete random variable (r.v.)
representing the forthcoming failure type for the
component,
T > 0 is a continuous r.v. representing the failure
time of the component,
At the ith CM point;
Xi is a continuous r.v. representing the RL, where
Xi = T ti > 0, and xi is the unknown realisation,
Yi = {Yi1 , Yi2 , . . ., Yiq } is a random vector of q CM
parameters with realisation yi = {yi1 , yi2 , . . ., yiq } ,
Vi = {y1 , y2 , . . ., yi } is a q i matrix of historical
CM observations,
i = {y1 Y1 y1 + dy, . . ., yi Yi yi + dy}
where dy0.

523

For a data set consisting of N component CM

histories;
nj is the number of component histories for which
a posterior analysis has revealed 
that the jth failure
type occurred and we have N = rj=1 nj .

type is recursively established as

pi (j|Vi )


= g(yi |u, j)fi1 (u + ti ti1 |j, Vi1 )du
0

For the component index, k = 1, 2, . . ., nj ;

pi1 ( j|Vi1 )dyi

bk is the number of monitoring points,

T k is the failure time,
tkbk is the time of the last CM point; T k tkbk > 0.


r
d=1


g(yi |u, d)
0

fi1 (u + ti ti1 |d, Vi1 )du pi1 (d|Vi1 )dyi

(3)

DEFINITIONS

Before monitoring begins, it is necessary to specify

and parameterise the component elements. Under the
assumptions of failure type j;

A point estimate of the RL is evaluated as the

conditional mean

f0 (x|j)dx = Pr(x X0 x + dx | X0 > 0, M = j)

is a prior pdf representing the initial residual life,
g(y|x, j)dy = Pr(y Yi y+dy| x Xi
x + dx, M = j) is a conditional pdf representing a
stochastic relationship between the CM parameters
and the RL,
p0 ( j) = Pr(M = j | X0 > 0) is a prior pmf for the
unknown failure type.

THE FAILURE TYPE RL ESTIMATION

MODEL


E[Xi |Xi > 0, i ] =

However, when utilising the model to evaluate

maintenance and replacement decisions, the description of the posterior conditional pdf for the RL of
the component is usually more useful than a point
estimate. Particularly useful is the cdf
Pr(Xi x|Xi > 0, i ) = Fi(1) (x|Vi )

Pr(x Xi x + dx | Xi > 0, i )
r


fi (x|j, Vi )pi (j|Vi )dx

(4)

x

Under the assumption of an unknown forthcoming

failure type, a posterior pdf for the RL of an individual
component is established at the ith CM point as

= fi(1) (x|Vi )dx =

xfi(1) (x|Vi )dx

(1)

j=1

The posterior conditional pdf for the RL under the

jth failure type is obtained recursively, using stochastic
filtering theory, as

fi(1) (u|Vi )du

(5)

We parameterise f0 (x | j) and g(y|x, j) using maximum likelihood estimation and the nj CM histories
from components that failed according to the jth failure type. The likelihood is the product of the relevant
conditional probabilities for all the histories;

L( j ) =

nj  bk


(dyki )1 Pr(yki Yki yki + dyki |
k=1

i=1

(2)

Xki > 0, M = j, k,i1 ) Pr(Xki > 0|Xk,i1 > 0,

M = j, k,i1 ) (dx)1 Pr(T k tkbk Xkbk

T k tkbk + dx | Xkbk > 0, . . . M = j, kbk )
(6)

where f0 (x|j, V0 ) f0 (x|j). At the ith CM point, a

posterior conditional pmf for the forthcoming failure

where, j is the unknown parameter set. After inserting the relevant pdfs and re-arranging, the likelihood

fi (x|j, Vi ) =

g(yi |x, j)fi1 (x + ti ti1 |j, Vi1 )

g(yi |u, j)fi1 (u + ti ti1 |j, Vi1 )du

524

function can be written as

nj
bk 



L( j ) =
g(yki |u, j)fk,i1 (u + tki tk,i1 |
i=1 0

k=1

j, Vk,i1 )du fkbk (T k tkbk |j, Vkbk )

(7)

See Carr & Wang (2008) for details. Maximisation

of equation (7), with respect to j , is undertaken using
an optimisation algorithm such as the BFGS quasiNewton method on the log of the likelihood function.
To select between different forms for f0 and g, we use
the Akaike information criterion (AIC).
The prior probabilities that each failure type will
occur for a given component are estimated from the
data set and described using the pmf p0 (j) nj /N for
j = 1, 2, . . . , r.
5

ASSESSING THE MODEL

To assess the applicability of the methodology for a

given scenario, we compare the failure type RL estimation model against a single general stochastic filter
for the RL with no failure type assumptions. For the
general model, the posterior conditional RL pdf is
fi(2) (x|Vi ) =

(2)
(x + ti ti1 |Vi1 )
g (2) (yi |x)fi1
(2)
g (2) (yi |u)fi1
(u + ti ti1 |Vi1 )du

INITIAL CASE ANALYSIS

In preparation for a larger project, we applied the

failure type RL estimation model to a test data set
of component monitoring histories (with associated
failure times) obtained from a certain model of aircraft engine. The objective of the analysis is to assess
the applicability and performance of the failure type
model when applied to multivariate oil-based CM
information.
We select forms and parameterise using historical CM data sets and the model is applied to new
component monitoring information for demonstration purposes. The model is also compared with a
general model with no failure type assumptions to
illustrate the benefits of the failure type model.
More information on the analysis will be given at
the conference and in an extended version of this paper.
6.1 The data
The CM data we are considering consists of the parts
per million (ppm) of contaminating metallic particles
in oil lubrication samples that are obtained from a
type of component used in aircraft engines. At each
monitoring point, we observe the ppm of iron (Fe),
copper(Cu), aluminium (Al), magnesium (Mg) and
chromium(Cr). Initially, 10 component CM histories
are used for the purpose of fitting the models. The
CM information from 2 new components is then
used to demonstrate the application of the models and
compare the performance of the failure type RL estimation model and the general model with no failure
type assumptions.

(8)
The comparison is achieved using an average mean
square error (AMSE) criterion. At the ith CM point
for the kth component, the MSE is

MSE ki =

(x + tki T k )2 fi(a) (x|Vi )dx

(9)

6.2 Model fitting

To reduce the dimensionality of the data and remove
any potential collinearity between the individual CM
variables over time, principal components analysis
(PCA) is applied to the covariance matrix of the CM
variables. After PCA, we have a vector of principal
components at time ti ; yi = {yi1 , yi2 , . . ., yik } where,

for models a = 1, 2. For the failure type RL model,

the AMSE is
bk
r 


(10)

represents the cth principal component and Fei , for

example, is the cumulative iron reading at the ith CM
point. Figure 1 illustrates the standardised first principal component over time and the associated failure
times for the 10 components that are to be used to fit
the models.

(11)

6.2.1 Failure type RL estimation model

For the failure type RL estimation model, the individual filters are parameterised using those CM histories
that are relevant to the particular failure type. As

nj

AMSE =

MSE jki /N

j=1 k=1 i=1

and for the general filter, the AMSE is

AMSE =

bk
N 

k=1 i=1

MSE ki /N

yic = c1 Fei + c2 Cui + c3 Ali + c4 Mgi + c5 Cr i

525

influence of the jth failure type) is

Standardised 1st PC

1.5

fi (x|j, Vi ) =

i


(x + ti )j 1 e(j (x+ti ))

0.5

(yz j (x+ti tz ))2

2j2

z=1
0
0

200

400

600

800

1000

1200

1400

1600

1800

0.5

(u + ti

j
)j 1 e(j (u+ti ))

i


(yz j (u+ti tz ))2

2j2

du

z=1

u=0
-1

(15)

1.5

Time

Figure 1. Illustrating the first principal component over

time and the associated failure times for the model fit data.

and using equation (3), the posterior conditional pmf

for the forthcoming failure type is
 
pi (j|Vi ) =

indicated in figure 1, it is possible to group the CM

histories according to the observed failure times and
the behaviour of the first principal component over
time. The first failure type is defined as underlying
behaviour that results in failure within the range 0
T < 1200 and for the second failure type, we have
T 1200 as the specified range.
Under the assumptions of the jth failure type, the
prior pdf for the RL is
f0 (x|j) =

j j j xj 1 e(j x)

(j2 2)1/2 (u + ti )j 1 e(j (u+ti ))

u=0

(yz j (u+ti tz ))2

2j2

pi1 (j|Vi1 )du

z=1

 

(u + ti1 )j 1 e(j (u+ti1 ))

u=0

(12)

i1


(yz j (u+ti1 tz ))2

2j2

du

 
r

z=1

where, x > 0 and j , j > 0 for j = 1, 2. For linearly independent principal components, we have the
combined conditional pdf
g(yi |x, j) =

g(yic |x, j)



 

d=1

(d2 2)1/2 (u + ti )d 1 e(d (u+ti )) d

u=0

(13)

(yz d (u+ti tz ))2

2 2
d

pi1 (d|Vi1 )du



z=1

c=1

For the particular data set under consideration, only

the first principal component is found to be significant, in that, it describes a satisfactory amount of the
variance within the data set. As such, the first principal component is used as the sole input to the models.
The stochastic relationship between the standardised
first principal component and the underlying RL is
described using a normal distribution as

 

(u + ti1 )d 1 e(d (u+ti1 ))

u=0

i1


(yz d (u+ti1 tz ))2

2 2
d


du

(16)

z=1

(14)

Under the assumptions of the failure type model,

the conditional pdf for the RL at the ith CM point,
fi(1) (x|Vi ), is given by equation (1), using equations
(15) and (16), for i = 0,1,2,. . . and r = 2.

where, j (x) is a function of the RL. Note that, when

applying the models to new component information,
the same principle component and standardisation
transformations must be applied to the CM data before
insertion into the models.
Using equations (2), (12) and (13), the posterior
conditional RL pdf at the ith CM point (under the

6.2.2 Failure type 1

In equation (12), a Weibull prior distribution is used
to model the initial residual life under the influence
of a given forthcoming failure type. Using the likelihood function in equation (7) and the 7 component
histories deemed to have failed according to failure
type 1, the shape and scale parameters are estimated as

g(y|x, j) = (j2 2 )1/2 e

(yj (x))2
2j2

526

Table 1. The estimated parameters and selection results

under the influence of failure type 1.

Table 3. The estimated parameters and selection

results for the general model.

1 (x)

(x)

A1 + B 1 x

A1 + B1 /x

A1 + B1 exp{C1 x}

A1
B1
C1
1

1.628
0.006281

0.393

0.031
18.882

0.866

2.326
4.386
0.002746
0.334

ln L( 1 )
AIC

63.036
132.072

165.753
337.506

41.852
91.704

2 (x)

ln L( 2 )
AIC

A + B exp{Cx}

A
B
C

0.735
0.002052

0.696

0.168
14.367

0.894

1.308
3.454
0.004233
0.333

ln L( )
AIC

236.721
479.442

292.852
591.704

71.27
150.54

f0(2) (x) = x1 e(x)

A2 + B2 x

A2 + B2 /x

0.606
0.001624

0.618

0.401
11.761

0.851

88.109
182.218

A + B/x

6.2.4 The general RL estimation model

Without any failure type assumptions, the prior pdf for
the RL is taken to be

Table 2. The estimated parameters and selection results

under failure type 2.

A2
B2
C2
2

A + Bx

118.21
242.42

A2 + B2 exp{C2 x}

(17)

where, x > 0 and , > 0. The relationship between

the standardised first principal component and the
underlying RL is described using the conditional pdf

1.228
3.268
0.003995
0.279

g (2) (y|x) = ( 2 2)1/2 e

13.531
35.062

1 = 0.001606 and 1 = 5.877 respectively. A number of forms are considered for the function 1 (x) in
equation (14) that describes the relationship between
the RL and the observed (and transformed) CM information under the influence of failure type 1. The
estimated parameters and the AIC results are given
in table 1 where, the objective is to minimise the AIC.
The selected function is 1 (x) = A1 + B1 exp
(C1 x) where, A1 = 2.326, B1 = 4.386, C1 =
0.002746 and the standard deviation parameter is 1 =
0.334. In addition, the prior probability that failure
type 1 will occur for a given component is estimated
as p0 (1) = 7/10 = 0.7.

(y(x))
2

(18)

where > 0 and a number of different functional

forms are considered for (x).
For the general model with no failure type assumptions, all 10 CM histories are used for parameterisation. The likelihood function of equation (7) is used to
estimate the parameters with the sub-script j removed
from consideration. The estimated parameters for the
prior RL pdf in equation (17) are = 0.0009735 and
= 1.917. The estimated parameters and selection
results for (x) are given in table 3.
The selected function is (x) = A + B exp (Cx)
where, A = 1.308, B = 3.454, C = 0.004233 and
the standard deviation parameter is = 0.333.
From equation (8), the posterior conditional pdf for
the RL under the general model assumptions is
fi(2) (x|Vi )

(x + ti )1 e((x+ti ))
6.2.3 Failure type 2
For failure type 2, the prior Weibull RL pdf is parameterised as 2 = 0.000582 and 2 = 19.267. The
estimated parameters and selection results for the
function 2 (x) from equation (14) are given in table 2.
The selected function is 2 (x) = A2 +
B2 exp(C2 x) where, A2 = 1.228, B2 = 3.268,
C2 = 0.003995 and 2 = 0.279. The prior probability
that failure type 2 will occur is p0 (2) = 0.3.

(u + ti )1 e((u+ti ))

u=0

i


z=1
i


(yz (x+ti tz ))2

2 2

(yz (u+ti tz ))2

2 2

du

z=1

(19)
6.3 Comparing the models
The models are compared using new component data.
The first component is known (in hindsight) to have

527

1
2

0.8
Probability

Probability

0.8
0.6

0.4

0.6
2
0.4

0.2

0.2

100

200

300

400
Time

500

600

700

1
0

500

1000

Figure 2. Tracking the forthcoming failure type over time

for the first new component.

1200

Figure 4. Tracking the underlying failure type over time

for the second new component.

Actual RL

Actual RL

1500

1000
FM Model RL
Estimate

800

Residual Life

Residual Life

1500

Time

General Model RL
Estimate

600
400

1200

FM Model RL
Estimate

900

General Model
RL Estimate

600

200

300
0
0

100

200

300

400
Time

500

600

700

800

0
0

500

1000

1500

Time

Figure 3. Comparing the conditional RL estimates of the

failure type and general models for the first new component.

failed according to failure type 1 and the second

according to type 2. We demonstrate the tracking of the
appropriate forthcoming failure type over time using
equation (16) and compare the mean RL estimates at
each CM point with those obtained using the general
model.
With regard to the modelling of optimal maintenance and replacement decisions, the availability of
a posterior conditional pdf for the residual life is of
greater use than a point estimate. With this in mind,
we use equations (10) and (11) to compare the AMSE
produced by the models for the two new components.
Figures 2 and 3 illustrate the tracking of the failure
type and the conditional RL estimation for the first
new component.
It is evident from figure 2 that the failure type
model correctly tracks the forthcoming failure type
over time. The conditional RL estimation process is
more accurate using the failure type model, as illustrated in figure 3, when compared with the general
model. This is also reflected in the AMSE results with
94884.66 for the failure type model and 105309.92 for
the general model.
Figures 4 and 5 illustrate the tracking of the failure
type and the conditional RL estimation for the second
new component that is known, in hindsight, to have
failed according to failure type 2.

Figure 5. Comparing the conditional RL estimates of

the failure type and general models for the second new
component.

Again, the failure type model correctly tracks the

forthcoming type, as illustrated in figure 4. Figure 5
demonstrates that the failure type model tracks the RL
more accurately and rapidly than the general model.
This is again reflected in the AMSE results with
68172.64 for the failure type model and 84969.59 for
the general model.

DISCUSSION

In this paper, we have presented a brief overview of

a model for failure type analysis and conditional RL
estimation. The modelling concepts have been demonstrated using a trial oil-based data set of component
monitoring observations. Although the data set is relatively small, the results do indicate that the failure
type model could be very useful for CM scenarios
that display different behavioural patterns and have
categorisable failures.
In the case study, the failure types are categorised
according to the time at which the failure occurs and
the behaviour of the various CM processes over time.
However, for operational components, the definition

528

of different failure types can be used to represent

different types of operation, or potentially different
faults in the system that are affecting the future life of
the component.
In the initial analysis, we compared the performance
of the failure type model with a general model with no
failure type assumptions. The models are compared
using a MSE criterion. At each monitoring point, the
MSE criterion compares the fit of the established conditional RL pdf about the actual underlying residual
life. When utilised in maintenance and replacement
models, if the density is tighter about the actual value,
the decisions are improved in the sense that, greater
operational time is available whilst still avoiding the
occurrence of failures. The AMSE is substantially
smaller, particularly in the second case, when using
the failure type model.
We are currently in the process of applying the
model to a much larger project involving multiple
potential failure types that are categorised according
to both the nature of the CM information and the
associated failure times.

ACKNOWLEDGEMENT
The research reported here has been supported by the
Engineering and Physical Sciences Research Council
(EPSRC, UK) under grant EP/C54658X/1.

REFERENCES
Makis, V. and Jardine, A.K.S. (1991) Optimal replacement
in the proportional hazards model, INFOR, 30, 172183.
Zhang, S. and Ganesan, R. (1997) Multivariable trend analysis using neural networks for intelligent diagnostics of
rotating machinery, Transactions of the ASME Journal of
Engineering for Gas Turbines and Power, 119, 378384.
Wang, W. and Christer, A.H. (2000) Towards a general condition based maintenance model for a stochastic dynamic
system, Journal of the Operational Research Society, 51,
145155.
Wang, W. (2002) A model to predict the residual life of rolling
element bearings given monitored condition information
to date, IMA Journal of Management Mathematics, 13,
316.
Vlok, P.J., Wnek, M. and Zygmunt, M. (2004) Utilising statistical residual life estimates of bearings to quantify the
influence of preventive maintenance actions, Mechanical
Systems and Signal Processing, 18, 833847.
Banjevic, D. and Jardine, A.K.S. (2006) Calculation of reliability function and remaining useful life for a Markov
failure time process, IMA Journal of Management Mathematics, 286, 429450.
Carr, M.J. and Wang, W. (2008a) A case comparison of
a proportional hazards model and a stochastic filter for
condition based maintenance applications using oil-based
condition monitoring information, Journal of Risk and
Reliability, 222 (1), 4755.
Carr, M.J. and Wang, W. (2008b) Modelling CBM failure
modes using stochastic filtering theory, (under review).

529

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Multi-component systems modeling for quantifying complex

maintenance strategies
V. Zille, C. Brenguer & A. Grall
Universit de Technologie de Troyes, Troyes, France

A. Despujols & J. Lonchampt

EDF R&D, Chatou, France

ABSTRACT: The present work proposes a two-stages modeling framework which aims at representing both
a complex maintenance policy and the functional and dysfunctional behavior of a complex multi-component
system in order to assess its performances in terms of system availability and maintenance costs. A first stage
consists in a generic model of component, developed to describe the component degradation and maintenance
processes. At the second stage, a system of several components is represented and its behavior is simulated when
given operating profile and maintenance strategy are applied, so as to estimate the maintenance costs and the
system availability. The proposed approach has been validated for a simplified turbo-pump lubricating system.

1
1.1

INTRODUCTION
Industrial context

The large impact maintenance process has on system

performances gives a real importance to its optimization, rendered difficult by the various and sometimes
antagonistic criteria like availability, safety, and costs
that must be simultaneously taken into account.
In order to work out preventive maintenance programs in nuclear and fossil fired plants, EDF applies
on critical systems a dedicated Reliability Centered
Maintenance method (RCM) which makes it possible
to determine the critical failure modes and to help the
experts to propose applicable, effective, and economic
preventive maintenance tasks.
Maintenance strategies are then established considering various possible options among the nature of
maintenance (corrective, preventive), the type of tasks
(overhaul, monitoring, scheduled replacements. . .),
their frequency, the maintenance line (repair on site
or in workshop), etc. and such choices are frequently
supported by experts opinion, good sense and intuition, that is qualitative information, which could
be helpfully complemented by quantitative information resulting from deterministic and/or probabilistic
calculations.
However, due to the probabilistic nature of the
failures, it is not easy to compare different options
on quantified bases. It is difficult to evaluate the
results of a maintenance program application over several years, in terms of availability, safety level, and

costs and decisions are generally based on qualitative

information.
For this reason it appears convenient to develop
methods to assess the effects of the maintenance
actions and to quantify the strategies.

1.2

Scientific context

Numerous maintenance performances and costs models have been developed during the past several
decades, see e.g. (Valdez-Flores & Feldman 1989),
with different interesting objectives reached.
However, they remain difficult to adapt to multicomponent systems and complex maintenance strategies such as those developed and applied in the
RCM context, since most of them are devoted
to simple maintenance strategies (periodic maintenance, condition-based maintenance, age-based maintenance, . . .) with a finite number of actions and
defined effects (perfect inspections, perfect replacements, minimal repair, . . .) or applied on single-unit
systems, see e.g. (Dekker 1996, Moustafa et al., 2004).
Other approaches, based on stochastic simulation,
already allow taking into account more complex maintenance strategies but they generally aim at developing
simulation techniques such as Monte Carlo simulation,
or optimization procedures, see e.g. (Marseguerra and
Zio 2000).
Finally, only a few part of maintenance simulation works give an interest to components degradation

531

and failure phenomena and to the effects of maintenance actions. All these observations have led to the
definition of a general and overall description of all
the aspects related to a system and its components
behavior and the different possible maintenance tasks
applied.

System failure
behaviour model

PERFORMANCES
EVALUATIONS

MODELING APPROACH PRESENTATION

2.1

System operation
model

Generic component
model

System-level
maintenance model

System representation for maintenance

strategies assessment

The assessment of complex maintenance program

performances, resulting for example from implementation of RCM, encounters several methodological
difficulties, whose resolution constitutes the scientific
challenge of this work. These difficulties are due first
to the complexity of the systems on which maintenance programs are implemented (systems constituted
of several dependent components, with several degradation mechanisms and several failure modes possibly
in competition to produce a system failure) and secondly to the complexity of the maintenance programs
(diversity of maintenance tasks).
Since both system behavior and maintenance process are strictly linked to the way system components
evolve and react to maintenance tasks, maintenance
strategies assessment can be done throughout a twolevel system description, representing the system and
its components and the way their relative behaviors
and maintenance processes are interacting together.

Figure 1. Overall structure for multi-component system

main-tenance modeling : white arrows represent interactions
between the various parts of the system level, black arrows
represent the interactions between system and component
levels. Assessment results are obtained in terms of maintenance costs and availa-bility performances evaluations.

these observations have resulted in working out an

overall model, shown on Figure 1, divided into four
parts sharing information.
A generic component model is developed for each
system component and the four modeling levels can be
validated independently and then associated by mean
of interactions in order to simulate complete systems,
as described in the following sections.

MODEL OF COMPONENT

3.1 Generic modeling of a maintained component

2.2

Overall model

Equipment failure modes analysis, carried out for

RCM method application are often based on a FMECA
matrix which contains, if complete, the necessary
information to represent the complexity of the problem, since the data collected characterizes:
components with their failure modes, degradation
mechanisms and applied maintenance tasks;
the system which is composed of these components
and characterized by:
its operation which influences the component
degradation mechanisms;
its unavailability which results from components
failure modes occurrence and/or from the preventive and corrective maintenance tasks carried
out;
maintenance rules depending on the maintenance
policy to gather elementary tasks carried out on the
components.
To model reality as accurately as possible it is convenient to represent the various characteristics and

To model system behavior one must represent how

the components it is made of can evolve during their
mission time through different operational states of
functioning, failure and unavailability for maintenance. These behaviors are consequences of different degradation mechanisms evolution that impact
on components and may lead to some failure mode
occurrences. Thus, it is necessary to describe those
evolutions and the way maintenance tasks can detect
them and repair if necessary, in order to prevent or
correct the effects on the system (Brenguer et al.,
2007).
All the phenomena and aspects that need to be represent and the impact they can have on the others are
described on Figure 2 and are presented more precisely
in paragraphs 3.2 and 3.3.
Here again, a real complexity comes from the
huge amount of possible relations between the various
characteristics to be represented. Indeed, influencing
factors often have an effect on various degradation
mechanisms which can involve various failure modes.
Moreover, symptoma can be produced from various degradations and maintenance tasks can make it
possible to detect or repair several degradations.

532

Operating
profile
System
Operation

Degradation
Mechanism

Environment
Influencing Factors

Degradation
Mechanisms

Failure Modes

System Dysfunction

Level 0

Symptoma

Influencing
Factors

Preventive
Maintenance

Corrective
Maintenance

Maintenance

Effects on system

Figure 2. Generic modeling of a maintained component.

Rec-tangular elements represent the phenomena and aspects
in-volved in a maintained component behavior and arrows
de-scribe how they are interacting together.

Failure Mode
occurence

Level n

Figure 3. Degradation and failure processes modeling.

Rectan-gular elements represent general aspects and phenomena in-volved in the described behavior and oval elements precise their different possible states. Black arrows
represent evolution transitions and grey arrows represent the
impact of one pheno-mena on another.

Degradation
Mechanism

Component unavailability, due either to failure

mode occurrence on an unscheduled way, either to preventive maintenance operation on a scheduled way, is
deterministic in terms of effects on the system operation. Unavailability duration is also linked to resources
sharing problems, since specific materials or maintenance repair teams have to be available at a given time
to perform the needed tasks.

3.2

Level i

Failure
rates

Symptom

Unsignificant
Maintenance:
Repair
Significant
Symptom
observation

Maintenance:
Detection

Component degradation behavior

Component behavior is defined by the way degradation mechanisms can evolve and may lead to some
failure mode occurrence. This evolution is due to the
influence of various factors such as operating conditions, environment, failure of another component,
etc.
As shown on Figure 3, each degradation mechanism
evolution is described through various degradation
levels and at each level an increasing failure probability represents the random apparition of the different
possible failure modes that can appear. The degradation level can be reduced after maintenance actions
performance.
Such a representation is motivated by the fact that
the different degradations that can affect a component
can be detected and observed by specific preventive
maintenance tasks which may lead to preventive repair
on a conditional way. Thus, it seems more appropriate
to describe each mechanism than to describe the global
degradation state of the component.
Due to the large amount of degradation mechanisms and to try to get a generic model which could
be applied in a large part of situations, it appears
convenient to propose various alternatives of degradation mechanism evolution representation in order to

Figure 4. Symptom apparition representation. When a

symp-tom reaches a given threshold it becomes significant
of a de-gradation evolution and its detection may avoid the
degradation observation.

choose among them the most adapted to a given situation, depending on the component, the mechanisms
and obviously the information data available. Different classical approaches can be considered (statistical,
semi-parametric, life-time model,. . .).
Another important aspect related to component
degradation mechanisms is that of symptom, i.e. a
phenomenon that may appear due to one or more mechanism evolutions and whose detection gives information about component degradation without directly see
the degradation itself. If possible, this type of detection can be made by less costly tasks which do not need
the component to be stopped, and so present some real
advantages.
As for degradation evolution, symptom apparition
and evolution can be modeled through some levels and
by means of thresholds that represent its significance,
i.e. the fact that it testifies a degradation evolution.

533

Figure 4 describes the relationships between degradation, symptom and maintenance.

Table 1.

RCM method Maintenance Tasks characteristics.

Task

Activation

Effects

Corrective maintenance

3.3

Component maintenance process

Within RCM method, different types of maintenance

tasks can be performed, devoted to different phenomena to detect and with different effects on the component behavior. The model developed has already been
dedicated to this type of strategies in order to let them
been represented integrally.
Indeed, maintenance activities effectiveness is modeled to represent the ability of preventive actions to
detect components degradations, and the ability of
both preventive and corrective actions to modify and
keep under control the degradation mechanism evolution in order to avoid failure occurrence, as shown on
Figures 3 and 4 with the impact of Maintenance on the
various phenomena described.
Other maintenance policies such as opportunistic
maintenance are defined at the system level model
and their activation are made thanks to information
exchange between the component-level model and the
system maintenance model.
In addition to duration and costs, tasks are differing
in terms of nature, activation condition, and effects on
the component state and on its availability, as shown
in Table 1. Regarding repair tasks effectiveness, repair
action are considered either As Good As New, As
Bad As Old or partial and detection tasks are defined
with some possible non detection and false alarm risk
errors.
A focus can be made on the differences between the
various condition base maintenance detection tasks,
ie external inspections, overhauls and tests, and their
relative vantages and disadvantages. Indeed they are
differing both in terms of unavailability engendered
for the component under maintenance and in terms
of efficiency of the detection with a relative impact
on the task performance cost. In particular, on one
hand, overhaul performance implies both the component scheduled unavailability and a high cost but is
really efficient in terms of detection since it consists
in a long and detailed observation of the component to
evaluate its degradation states and eventually decide
to repair it preventively. On the other hand, external
inspections are less expensive and consist in observe
the component without stopping it. These two vantages imply a larger distance from the degradation
with some associated error risks of non detection or
false alarm. Obviously, this kind of task can easily be
used to observe some eventual symptoma characteristics of one or more degradation mechanisms evolution,
and so some appreciation error can exist when decision of preventive repair are taken (treatment of the
bad degradation mechanisms whereas another one is

Repair

Failure mode
occurrence

Unavailability
failure repair

Predetermined preventive maintenance

Scheduled
Replacement

Time period
elapsed

Unavailability

Condition-based preventive maintenance

External
inspection

Time period
elapsed

Overhaul

Time period
elapsed

Test

Failure observed
during stand-by
period
Symptom >
detection threshold
degradation >
repair threshold

Preventive
Repair

No unavailability
symptom
detection
Unavailability
degradation
detection
Unavailability
failure repair
Unavailability
degradation
repair

still in evolution with an increasing probability of failure mode occurrence). Finally, tests are performed
are expensive but efficient tasks that are performed
on stand-by components to detect an eventual failure before the component activation, but can have bad
effects on it.

SYSTEM MODEL: MAINTENANCE COST

ASSESSEMENT

4.1 Three models to describe the system behavior

The system level consists in representing a system
of several components and simulating its behaviour
when given operating profiles and maintenance strategies are applied, so as to estimate the maintenance
costs and the system availability. This is done through
three different models describing its dysfunction and
functioning behaviour and the maintenance rules.
The system dysfunction model describes all the
degradation/failure scenarii that can affect the system.
It gives out the global performance indicators of the
maintained system. Indeed, it allows the evaluation of
the system unavailability, due either to a failure either
to some maintenance actions, and also the associated
maintenance costs.

534

A global maintenance cost model can be defined by

Equation 1:

The system operation model aims at describing the

nominal behavior and the operating rules of the system. This model interacts with the component models
and evolves according to the operating profile and
to the needs of the system : activating of a required
component, stopping of a superfluous component, . . .
Obviously the operating behavior of the system cannot be described by the simple juxtaposition of the
component-level models and it is necessary to take into
account all the possible interactions and dependences
between components. At this level one can model spare
equipments, activating of defense systems in case of
an equipment failure, stopping of a line in case of
maintenance of one of its component, . . .
In the system maintenance model, one can define
the applied system maintenance strategy. It allows the
description of grouping procedures which are used to
take advantage of economic or technical dependences
between components in the case of opportunistic maintenance. This model also includes resources sharing
and availability problems due to a limited number of
repair teams or specific tools and spare parts stocks.

where T miss = the mission time throughout which the

system operates, ni = number of maintenance task i
performed;ci = cost of maintenance task i; tsu = time
the system is under scheduled unavailability; tuu =
time the system is under unscheduled unavailability;
csu = cost of scheduled unavailability; cuu = cost of
unscheduled unavailability.
According to Equation 1, one can decide to asses
the global costs of strategies that differ in terms of
tasks type and frequency, knowing their relative costs
of operation in order to impact on the associated
durations of scheduled and unscheduled unavailability
engendered.

4.2

Interactions between the different models

The three system-level models and the componentlevel model interact together in order to represent
completely the system behavior, its unavailability and
expenditures, knowing the behavior of its components
and the maintenance tasks that are carried out.
Component-level models give information on components states (failure, unavailability for maintenance)
and on maintenance costs to the three other systemlevel models which evolve according to this input data
and possibly sent feedback data.
As shown on Figure 1, the system operation model
sends information to the component-level models to
active a stand-by component or stop an auxiliary component that has become useless after the repair of the
main component.
The system maintenance model can send data to
the component-level model to force the maintenance
of a component coupled together with a component
already in maintenance.

4.3


Cost(Strategy) =

lim

ni ci + tsu csu + tuu cuu

TMiss

TMiss
(1)

CASE STUDY AND RESULTS

5.1 Model development and validation: Petri Nets

and Monte Carlo simulation
The proposed generic methodology has been implemented using the Stochastic Synchronized Petri nets
(SSPN) and it has been coupled with the Monte Carlo
simulation to compute the performances assessment
of industrial systems, see e.g. (Barata et al., 2002;
Brenguer et al., 2007; Dutuit et al., 1997).
For systems dependability studies, SSPN offer a
powerful and versatile modeling tool that can be used
jointly with Monte Carlo simulation, which is widely
used in this kind of work, see e.g. (Barata et al., 2002;
Simeu-Abazi & Sassine 1999).
In order to validate the modeling approach, simulations of real complex systems behavior have been
made to study the effects of parameters variations,
such as maintenance tasks period, on the system
behavior, with really interesting results obtained that
have encouraging further developments to improve its
specificities.

Maintenance strategy assessment

The overall model presented permit maintenance

strategies assessment by evaluating the performances
obtained from the system on which a given strategy is
applied.
This takes into account both the maintenance costs,
depending on the number of tasks performed and the
relative resources used and the system availability and
unavailability during its mission time.

5.2 Case study: Turbo-pump lubricating system

We provide here results obtained on a part of a simplified turbo-pump lubricating system, described on
figure 5, to underline the main originality of the
proposed approach, that is the capability of multicomponent system modeling and complex maintenance assessment. In particular the objective of the

535

Cl1
Po1

Po2

Ca

Cl2

Failure Mode 1
(Unscheduled
shutdown)

Failure Mode 2
(Impossible starting)

Degradation
Mechanism A
(Bearing Wear)

Degradation
Mechanism B
(Oxydation)

Symptom 1
(Vibrations)

Symptom 2
(Temperature)

Figure 5. Part of a turbo-pump lubricating system made of

two pumps Po1 and Po2 and check valves Cl1 and Cl2 and a
sensor Ca.

case study was to compare different possible strategies in terms of global cost, composed of different
type of tasks, depending on their periodicity.
Expert opinions and information data have been
collected to define components and system characteristics as well as those of the maintenance tasks
possibility performed, to let the modeling approach
been applied and simulated.
Indeed, for each component, degradation mechanism and maintenance tasks basic parameters such as
those in table 2 have to be specified.
Obviously, maintenance tasks are also described
in terms of periodicity and decision rule criteria,
that is which degradation levels can be observed and
when preventive repair are decided to be performed.
These characteristics defined the maintenance strategy
applied and simulated.
To model the particular case presented, for each
component, the main degradation mechanisms have
been characterized in terms of levels of degradation
and relative failure rate for the various possible failure modes, possible symptoma and their probability
or delay of apparition and evolution until some significant thresholds. We also defined the evolution transitions from one degradation level to the successive one
and finally, the influencing factors that have effects
on the mechanisms evolution. In the present case
study, mechanism evolution has been modelled using
a Weibull Life-Time distribution law, whose parameters were depending on the mechanisms described and
the information available from the experts, to compute
the time of the transition from one degradation level
to the successive one.
In particular the following statements, described
on Figures 6 and 7, have been defined regarding
the different relationships between the degradation
mechanisms, associated symptoma and failure modes
considered for each component:
Concerning Sensor Ca, only very rare random
occurrence of an electronic failure has been considered, and no symptom nor degradation mechanism.

Figure 6. Description of the relationships between the

degrada-tion mechanisms, failure modes and symptoma
considered in the modeling of Pumps P01 and P02 behavior.

Failure Mode 3
(No opening)

Failure Mode 4
(External leaks)

Degradation
Mechanism C
(Axis blocking)

Degradation
Mechanism D
(Joint wear)

Symptom 3
(Deposits)

Figure 7. Description of the relationships between the

degrada-tion mechanisms, failure modes and symptoms considered in the modeling of check valves Ca1 and Ca2
behavior.

Then, different tasks were considered to define

the preventive maintenance strategies applied to the
system with condition-base maintenance for pumps
and check valves and systematic maintenance for the
sensor :
Pumps Po1 and Po2 degradation can be notified
directly by practicing overhauls, specifics to each
degradation mechanisms, or indirectly thanks to
external inspections, to detect the symptoms eventually appeared. Then, depending on the detection results, preventive repair can be performed
to avoid failure mode occurrences. As previously
said, overhauls are very effective detection tasks
but engender component scheduled unavailability

536

Table 2.

Basic parameters needed for model simulation.

Degradation mechanism evolution description

Levels

Basic evolution
description

Influencing factors
impact

Minimal and
maximal
degradation
thresholds

Representation chosen
and parameters values
(ex: Weibull law
parameters)

Modification of the
basic parameters
depending on the
influencing factors

Failure
Failure modes that
can occur.
Failure rates.

Symptoma
Symptoma that
appears. Eventual
delay or probability.

Degradation mechanism evolution description

Levels
For each
degradation
level

Maintenance Tasks (External inspections, overhauls, tests)

Observations
Failures, degradations
or symptoma
observed

Effectiveness
Error risk
(non detection
and false alarm)

Parameters
Cost
Duration
Resources

Maintenance tasks (Preventive and corrective repairs)

Impacts
Degradations or Failures
Repaired or reduced

Effectiveness
Repair type
(AGAN, ABAO,
or partial)

for maintenance, whereas external inspections are

made without stopping the component but present
a risk of error in the detection (non detection or
false alarm). Moreover, in the proposed case study,
one of the symptoms can appear due to the two
degradation mechanisms, which implies another
error risk for the repair decision that can be taken
after the component inspection : the possible repair
of the bad mechanism without reducing the failure mode occurrence probability linked to the other
mechanism.
Check-valves Ca1 and Ca2 degradation can also be
notified directly by practicing overhauls specifics
to each degradation mechanisms. Since one of the
degradation mechanisms cannot be detected thanks
to an associated symptom, external inspections can
only be performed to detect the evolution of the
other mechanisms, avoiding the performance of the
relative specific overhaul.
Sensor Ca failure is supposed to be random and
rare so a systematic repair is performed before its
expected occurrence.
Finally, the way the entire system can function
or become failed or unavailable for maintenance has

537

Parameters
Cost
Duration
Resources

been represented by means of classical reliability tools

such as failure and event trees. In particular, the two
branches composed of pump and check valve are operating on a redundancy way and the switch over from
one to the other is made thanks to the sensor which
detect the failure of the activated branch and lead to
the activation of the stand-by one.
Condition-based and systematic maintenance tasks
are performed with a given frequency in order to
detect and repair if necessary component degradation
to avoid their failure. After the activated branch has
failed, the other branch is activated so as to render
possible the corrective repair without creating some
additional unavailability.
The system can be unavailable either on a scheduled
way, for a preventive maintenance task performance,
or on an unscheduled way, due to some component
failure.
Different strategies, differing essentially on the type
of preventive tasks performed on the pumps and check
valves, have been simulated and for each one an evaluation of the relative maintenance costs as defined
by Equation 1 has been computed according to the
variation of the tasks periodicity. The objective was

Minimal strategies global costs

Maintenance global costs

5000

1400

4500

1200

4000
1000

3500
3000

800

2500

600

2000
400

1500
1000

200

500

0
Maintenance strategies

0
Maintenance tasks periodicity increasing

Figure 10. Costs maintenance for maintenance strategies

made of both overhaul and external inspections. Black bar
represent the minimal cost of the only overhauls strategy,
grey bars represent the minimal cost of both overhauls and
external in-spections strategies.

Figure 8. Costs maintenance for a strategy made of

overhauls.

Maintenance global costs

4500
4000
3500
3000
2500
2000
1500
1000
Maintenance tasks periodicity increasing

Figure 9. Costs maintenance for a strategy made of external

in-spections.

to compare the strategies in terms of costs difference

induced by preferring one type of detection task to
the other one, ie overhauls and external inspections,
for the preventive maintenance of pumps and check
valves, given the relative characteristics of the tasks.
Figure 8, presents the results obtained for a strategy only made of overhauls for pumps and check
valves; whereas on Figure 9, only external inspections were performed to detect pumps and check valves
degradation.
As it is easy to note, even if external inspections are
less costly and do not induced some system unavailability, the second strategy evaluation leads to higher
maintenance global costs. That is obviously due to

the fact that all the external inspections tasks represent some non detection and false alarm error risks
which are even more important regarding degradation
mechanism C. It is indeed impossible to detect its evolution through some symptom detection and it has been
assumed that it is done with a very poor efficiency due
to the distance between the degradation evolution and
the task performance.
Thus, it is interesting to compare the advantage
of performing both the different type of tasks. By
so doing, it is indeed possible to control the system
components degradation evolution indirectly when
possible, with external inspection devoted to some
symptom detection, and also on a more direct way
with overhauls. Those are more efficient in terms of
degradation detection and, when performed with a
higher periodicity, the global maintenance costs can
be reduced.
Figure 10 presents the minimal global costs for
different strategies composed as follow:
external inspections supported by overhauls to
detect degradation mechanism A, B and D, ie
those detectable through some symptom observation, with overhauls preformed with a higher
periodicity than external inspections one,
overhauls to detect the evolution of degradation
mechanism C, since it is not convenient in terms
of efficiency to observe it through external inspections.
Again, strategies were differing in terms of tasks
periodicity and are here compared to the minimal cost

538

corresponding of the strategy composed only by overhauls to show that some appropriate combinations
render possible to reduce the global maintenance costs.
6

CONCLUSIONS

The objective of the work presented here is to model

and to simulate maintenance programs in order to provide quantitative results which could support choices
between different maintenance tasks and frequencies. The approach is dedicated to multi-component
systems and RCM type complex maintenance strategies. This is done through a structured and modular
model which allows taking under consideration dependences between system components due either to
failures either to operating and environmental conditions. Maintenance activities effectiveness is modeled
to represent the ability of preventive actions to detect
components degradations, and the ability of both preventive and corrective actions to modify and keep
under control the degradation mechanism evolution
in order to avoid failure occurrence. It also let take
into account resources sharing problems such as repair
teams or spare parts stocks. A case study regarding the
modeling of a turbo-pump lubricating system shows
how the approach can efficiently been used to compare
various maintenance strategies.

tems. In MM2007 proceedings Maintenance Management

conference, Roma, 2728 September 2007.
Barata, J., Guedes Soares, C., Marseguerra, M. &
Zio, E. 2002. Simulation modeling of repairable multicomponent deteriorating systems for on-condition maintenance optimisation. Reliability Engineering and System
Safety 76: 255267.
Dekker, R. 1996. Applications of maintenance optimization
models : a review and analysis. Reliability Engineering
and System Safety 51(3): 229240.
Dutuit,Y., Chatelet, E., Signoret, J.P. & Thomas, P. 1997.
Dependabiliy modeling and evaluation by using stochastic Petri Nets :Application to two test cases. Reliability
Engineering and System Safety 55: 117124.
Marseguerra, M. & Zio, E. 2000 Optimizing Maintenance
and Repair Policies via a combination of Genetic Algorithms and Monte Carlo Simulation. Reliability Engineering & System Safety 68(1): 6983.
Moustafa, M.S., Abdel Maksoud, E.Y. & Sadek, S. 2004.
Optimal major and minimal maintenance policies for deteriorating systems. Reliability Engineering and System
Safety 83(3): 363368.
Simeu-Abazi, Z. & Sassine, C. 1999. Maintenance integration in manufacturing systems by using stochastic
Petri Nets. International Journal of Production Research
37(17): 39273940.
Valdez-Flores, C. & Feldman, R.M. 1989. A survey of
preventive maintenance models for stochastically deteriorating single-unit systems. Naval Research Logistics
Quarterly 36: 419446.

REFERENCES
Brenguer, C., Grall, A., Zille, V., Despujols, A. &
Lonchampt, J. 2007. Modeling and simulation of complex maintenance strategies for multi-component sys-

539

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Multiobjective optimization of redundancy allocation in systems with

imperfect repairs via ant colony and discrete event simulation
I.D. Lins & E. Lpez Droguett
Department of Production Engineering, Universidade Federal de Pernambuco, Recife, Pernambuco, Brazil

ABSTRACT: Redundancy Allocation Problems (RAPs) are among the most relevant topics in reliable system design and have received considerable attention in recent years. However, proposed models are usually
built based on simplifying assumptions about the system reliability behavior that are hardly met in practice.
Moreover, the optimization of more than one objective is often required as, for example, to maximize system reliability/availability and minimize system cost. In this context, a set of nondominated solutionssystem
designs with compromise values for both objectivesare of interest. This paper presents an ACO approach
for multiobjective optimization of availability and cost in RAPs considering repairable systems subjected to
imperfect repairs handled via Generalized Renewal Processes. The dynamic behavior of the system is modeled
through Discrete Event Simulation. The proposed approach is illustrated by means of an application example
involving repairable systems with series-parallel configuration.

INTRODUCTION

Suppose that a system is composed of several subsystems in series and that each one of them may
have a number of redundant components in parallel.
The determination of the quantity of redundant components in order to maximize the system reliability
characterizes a redundancy allocation problem.
The increase of the redundancy level usually
improves not only system reliability, but also increases
the system associated costs. With the purpose of incorporating costs limitations, RAPs are often modeled
as single optimization problems with the objective of
maximizing system reliability subject to costs constraints. However, in real circumstances, one may
desire to consider the associated costs not as a constraint but as an additional objective to be attained.
In these situations, in which multiple objectives are
taken into account, it is necessary a multiobjective
optimization approach in modeling RAP.
RAPs are essentially combinatorial since the aim
is to find optimal combinations of the components
available to construct a system. The complexity of
such problems may considerably increase as the number of components grows leading to situations where
the application of classical methods such as integer programming (Wolsey 1998) is prohibitive in
light of the time required to provide results. Alternatively, heuristic methods such as Genetic Algorithms
(GAsGoldberg 1989, Michalewicz 1996) and Ant
Colony Optimization (ACODorigo & Sttzle 2004)

can be adopted. Both GA and ACO are stochastic

approaches based in nature. GA mimics the natural evolution process and has been widely applied in
solving RAPs (Cantoni et al.,, 2000, Busacca et al.,
2000, Taboada & Coit 2006, Taboada et al., 2007).
However, GA demands a substantial computational
effort when considering large-scale problems. ACO,
in turn, is based on the behavior of real ants and has
been proposed by Dorigo et al., (1996) to solve hard
combinatorial optimization problems.
In a number of practical situations, components
and systems are repairable, that is, after a failure
they undergo maintenance actions that do not consist
in their entire replacement (Rigdon & Basu 2000).
Moreover, such maintenance actions often restore
the system to an intermediate condition between as
good as new and as bad as old conditions, that
is, imperfect repairs are performed. In order to
model failure-repair processes of systems subjected
to imperfect repairs, Generalized Renewal Processes
(GRPKijima & Sumita 1986) can be applied.
The consideration of repairable systems composed
of components under deterioration processes and subjected to imperfect repairs (with significant downtime
time) provides a realistic handling of RAPs. On the
other hand, the introduction of such real characteristics makes the evaluation of system failure-repair
process analytically intractable. Therefore, Discrete
Event Simulation (DESBanks et al., 2001) can be
applied in order to obtain some system performance
measures such as system availability. Basically, DES

541

attempts to imitate system behavior by randomly generating discrete events (e.g. failures) during simulation
time (mission time). In addition, the DES flexibility
permits the introduction of many system real aspects,
such as the taking into account the availability of
maintenance resources during mission time.

1.1

Previous works

Shelokar et al., (2002) exemplify the application of

ACO coupled with a strength Pareto-fitness assignment to handle multiobjective problems in reliability
optimization. In the context of electrical systems,
Ippolito et al., (2004) propose a multiobjective ACO
to encounter the optimal planning strategy for electrical distribution systems. The authors use as many
ant colonies as many handled objectives and the algorithm is divided in two different phases: a forward
phase in which each ant colony attempts to optimize
separate objectives and a backward phase in which
nondominated solutions are taken into account. ACO
was applied by Liang & Smith (2004) to solve a RAP
with the single objective for maximizing system reliability subject to costs and weight constraints. They
assume nonrepairable components with constant reliability values over mission time. Nahas & Nourelfath
(2005) use ACO to find the best set of technologies to
form a series system with the aim of obtaining maximal
reliability given budget constraints. So as in Liang &
Smith (2004), reliability values are fixed during mission time. Zhao et al., (2007) use a multiobjective ant
colony system approach in order to maximize reliability in a RAP with a series-parallel system formed by
k-out-of-n: G subsystems. It is important to emphasize
that the authors do not obtain a set of nondominated
solutions, since they only aim to maximize system
reliability subject to cost and weight constraints. Nevertheless, each component featuresreliability (r),
cost (c) and weight (w)are summarized in the quotient (r/(c + w)), which is used as a problem-specific
heuristic information during the execution of ACO.
The majority of the works in the literature regarding ACO for solving RAPs does not give enough
attention to the system reliability aspect. RAPs
are usually built based on simplifying assumptions
about the reliability behavior of the system and
of its components in order to facilitate the problem modeling and solution. These simplifications
(e.g. considering nonrepairable components with
constant reliability values over mission time) are
often non-realistic and do not permit the evaluation of some important real-world problems. Indeed,
models for repairable systems in RAPs using GA
as optimization algorithm assume that components
have constant failure rates, i.e., that they have an
underlying Exponential distribution to model the

times between failures (see, for example, Busacca

et al., (2001), Chiang & Chen (2007), Juang et
al., (2008)). This hypothesis does not allow for the
incorporation of the effects of component deterioration and may incur in grotesque estimation errors of
some important performance measures such as system availability (Bowles 2002). Although Cantoni
et al., (2000) tackle repairable systems with imperfect repairs, they consider constant failure rates from
the moment that a component returns into operation until the occurrence of the very next failure.
Moreover, they use a Brown-Proschan (B-P) model
of imperfect repairs, which is a specific type of failure
intensity model (Doyen & Gaudoin 2004).
In a recent paper, Kuo & Wan (2007) emphasizes
that other optimization approaches than GA, such as
ACO, may be investigated and more widely applied
in solving RAPs. Besides, they also stress that nonrenewable multi-state systems ought to be taken into
account in RAPs approaches.
In this paper, as an attempt to join ACO and nonrenewable systems in RAPs, a multiobjective ACO
algorithm coupled with DES is provided to solve RAPs
in the context of repairable systems with components
subjected to imperfect repairs. The two selected objectives are the system average availability and the total
system cost. DES is used to obtain the former objective and also some parameters required for the cost
calculation over the established mission time. The system cost is composed of the components acquisition
and operating costs, corrective maintenance costs and
costs incurred due to system unavailability. A limited
amount of maintenance resources is considered to give
maintenance support to the entire system. As soon as
a component fails, a maintenance resource is required
and if it is available, a delay time is generated according to an Exponential distribution. Each component
is supposed to have times to failure (TTF) modeled
by Weibull distributions. The times to repair (TTR), in
turn, are assumed to be exponentially distributed with
different means.
The desired outcome is a set of nondominated system designs. The decision maker can then have an idea
of system average availability during mission time and
of the related costs that might have been spent at the
end of that period. Moreover, note that the system lifetime cycle is thus taken into account and it is not a
merely issue of obtaining compromise designs at the
system acquisition moment.
The paper organization is as follows. Section 2
introduces some concepts related to GRP and to
multiobjective optimization. Section 3 details the
use of multiobjective ACO, whereas Section 4
describes the DES approach considered in this work.
In Section 5 an example application is discussed
and the following section gives some concluding
remarks.

542

2
2.1

PRELIMINARIES

of Equation 1 yields:

Generalized renewal processes

Failure-repair processes of repairable items are often

modeled according to either Nonhomogeneous Poisson Processes (NHPP) or Renewal Processes (RP).
Both of them are counting processes meaning that,
in the context of reliability engineering, the times to
repair are negligible compared to the component operational time. NHPP assume that the repairs restore the
system into an operational state with the same condition it had just before failure, that is, corrective
maintenance corresponds to minimal repairs. On the
other hand, RP assume that repairs bring the system
back to a like new condition (a renewal occurs) and
then corrective maintenance corresponds to perfect
repairs. Moreover, in RP, times between failures (TBF)
are independent and identically distributed (i.i.d.) with
an arbitrary density function, whereas in NHPP they
are neither independent nor identically distributed.
Nevertheless, the hypotheses of minimal or perfect
repairs required by either NHPP or RP, respectively,
are often not realistic. In practical situations, maintenance actions usually restore the system to a condition
between these two extremes, that is, they are imperfect
repairs and the component returns into operation with a
condition better than old and worse than new. To overcome this limitation, Generalized Renewal Processes
(GRP) can be used to model failure-repair processes of
components subjected to imperfect repairs. GRP are
a sort of virtual age models, in which repair actions
result in a reduction in the component real age. A
parameter q (effectiveness or rejuvenation parameter)
is introduced in the model and it assumes values related
to the maintenance action efficacy (see Table 1). The
common values of q are in [0,1].
The rejuvenation parameter is used in the calculation of the component virtual age (Vn ) as defined as
follows:
Vn = Vn1 + qXn

(1)

where Xn is the time between the (n 1)th and the n

th failure. By definition, V0 = 0. Thus, the expansion

Vn = q

n


Xi

(2)

i=1

where i Xi , i = 1, . . ., n, is the real component age.

The definition of virtual age presented in Equation
1 and in Equation 2 is in accordance with Kijima Type
I model (Kijima & Sumita 1986, Yaes et al., 2002),
which assumes that the repair acts just on the very last
failure and compensates only the damage accumulated
in the interval between the (n 1)th and n th failures.
Hence, it reduces only the additional age Xn .
The cumulative distribution of the ith TBF(xi ) can
be calculated by a probability distribution function
conditioned on the (i 1)th component virtual age,
as follows:
F(xi |vi1 ) = P(X xi |V > vi1 )
=

F(vi1 + xi ) F(vi1 )
1 F(vi1 )

(3)

If TBF are supposed to have a Weibull distribution,

Equation 3 becomes


 v   v x 
i1
i1
i
F(xi |vi1 ) = 1 exp

(4)
When TTR are not negligible if compared to the
component operational time, they may be considered in the evaluation of its failure-repair process. In
this situation, there are two stochastic processesone
regarding the failure process and the other related to
the repair process. The superimposing of these two
stochastic processes results in an alternating process
that characterizes the component state (either operational or unavailable due to repair action). However,
when the system is composed of several components
each one having an alternating process, the analytical
handling of the entire system failure-repair process
becomes infeasible. In these cases, DES may be used
as an alternative to overcome such difficulties.
2.2

Multiobjective optimization

Table 1. Repair classification according to

values of parameter q.

The general formulation of a multiobjective optimization problem is as follows:

Value

Repair type

Max z = [ f1 (x), f2 (x), . . ., fk (x)]

q=0
0<q<1
q=1

Perfect
Imperfect
Minimal

(5)

Subject to
gi (x) = 0, i = 1, . . ., p

543

(6)

hi (x) 0, i = p + 1, . . ., m

(7)

where z is the vector formed by k objective functions,

x is the n-dimensional vector of decision variables, p is
the number of equality constraints gi (x) and mp is the
quantity of inequality constraints hi (x). Frequently, in
the multiobjective context, a unique solution that optimizes all objectives is very difficult to be found or it
does not exist. Thus, a set of compromise solutions
among objectives, i.e. nondominated solutions, may
be encountered. A solution is said to be nondominated
if, for all objectives, it has a performance at least as
great as the performances of the other solutions and,
at least for one of the objectives, its performance overcome the performance of the others. In mathematical
terms:
x 1  x 2 fh (x 1 ) fh (x2 ), h
fh (x1 ) > fh (x2 ), for some h

and
(8)

where  denotes that x 1 dominates x 2 (i.e., x 1 is

nondominated) considering a maximization problem.
Otherwise, if a minimization problem is taken into
account the symbols and > must be replaced by
and <, respectively. Once the set of nondominated
solutions (also known as Pareto front) is obtained, the
decision maker can choose any of the solutions to be
implemented according to his preferences.
Optimization methods usually handle multiobjective problems by transforming them into a single objective problem and then applying classical mathematical
programming methods such as linear and nonlinear
programming (Luenberger 1984). The Weighted Sum
Method and the -Perturbation Method are examples
of traditional multiobjective approaches (Coello et al.,
2002, Deb 1999). However, a significant drawback in
applying those methods is the fact that they have to
be executed several times in order to obtain different
nondominated solutions. Moreover, objective functions that do not have some features such as continuity
and differentiability render these classical methods
useless.
Alternatively, stochastic optimization methods
based on nature such as ACO can be applied in optimizing multiobjective problems given their flexibility
in handling each objective separately. Since these
methods consider various potential solutions simultaneously, a number of Pareto solutions can be reached
in a single execution of the algorithm.

MULTIOBJECTIVE ACO

The multiobjective ACO (multiACO) put forward in

this work is a problem specific algorithm which is

based on the single objective Ant System for the Traveling Salesman Problem (TSP) proposed by .Dorigo
et al., (1996). The following subsections are dedicated
to the discussion of the proposed multiACO.
3.1 Input data
Initially it is necessary to identify the number of
subsystems in series (s), the maximum (ni,max ) and
minimum (ni,min ) number of redundant components in
ith subsystem, i = 1, . . ., s. Each subsystem can be
composed by different technologies, which may have
different reliability and cost features. Hence, it is also
required the quantity of available component types (ci ,
i = 1, . . ., s) that can be allocated in each subsystem. With this information, the ants environment is
constructed.
Moreover, information about components TTF and
TTR distributions and the components related costs
need to be specified. The ACO specific parameters
nAnts, nCycles, , , and Q are also required and
they are discussed in Subsection 3.3.
3.2 Environment modeling
The environment to be explored by the ants is modeled as a directed graph D = (V , A), where V is the
vertices set and A is the arcs set. D has an initial vertex
(IV) and a final vertex (FV) and is divided in phases
that are separated by intermediate vertices. An intermediate vertex indicates the ending of a phase and also
the beginning of the subsequent phase and therefore
is common to adjacent phases. In this work, a phase
is defined as the representation of a subsystem and
vertices within a phase represent either its extremities
or the possible components to be allocated in parallel in such subsystem. The quantity of vertices of the
ith phase is equal to ni,max cti plus the two vertices
indicating its beginning and its ending.
Vertices are connected by arcs. Firstly, consider a
problem with a unique subsystem. Hence, there is only
one phase with IV and FV, but intermediate vertices
are not necessary. IV is linked to all vertices within
the existent phase (except FV), which in turn are connected with each other and also with FV. Now suppose
a problem involving two subsystems. Then, an intermediate vertex plays the role of FV for the first phase
and also the role of IV to vertices within the second
phase. All vertices in second phase (except the vertex indicating its beginning) are linked to FV. These
instructions can be followed in the case of s subsystems. Arcs linking vertices within ith phase belong to
such phase.
For the sake of illustration, Figure 1 shows an
example of a directed graph representing an ants environment, where s = 2, n1,min = 1, n1,max = 2, ct1 = 2,
n2,min = 1, n2,max = 1, ct2 = 4.

544

where is the pheromone evaporation between times

t and t + m. The quantity vw is obtained by

vw =

nAnts


vw,k

(11)

k=1

Figure 1. Example of directed graph representing an ants

environment.

3.3

Transition probabilities and pheromone

updating

SC = Cmax Ak /Ck

Suppose that an ant k is in vertex v at time t. Then ant

k must choose a vertex w to visit at time t + 1. Several
limitations are imposed in this selection: (i) vertex w
must not have been visited by ant k previously; (ii) if
ant k is in phase i and it has not visited at least ni,min
vertices yet, w can not be the very next intermediate
vertex (or FV, if phase i is the last one); (iii) if ant k
has already visited ni,max vertices, than w must be the
subsequent intermediate vertex (or FV, if phase i is the
last one). Besides, let Wk be the set of allowed vertices
that ant k can be at t + 1. Hence, the probability of ant k
going from v to w is given by the following expression
(.Dorigo & Sttzle 2004):

pkvw

(9)

uWk

where vw (t) is the pheromone quantity on the arc linking vertices v and w at time t, vw is the visibility
of the same arc, is the relative importance of the
pheromone quantity, and is the relative importance
of vw .
In this work, the amount of pheromone is initially
set to 1 / ni,max ct i for each arc within the ith phase.
A cycle is finished when all ants reach FV. When that
occurs, the pheromone quantity on arcs is updated (i.e.,
multiACO is an ant-cycle algorithmsee Dorigo et
al., (1996), Dorigo & Sttzle (2004)). Let m be the
highest number of visited vertices in a cycle. Then,
at time t + m, the pheromone quantity in each arc is
updated in accordance with the rule:
vw (t + m) = (1 )vw (t) + vw

(10)

(12)

where Cmax is the maximum system cost obtained in

the current cycle and Ak and Ck are, respectively, the
system mean availability and the system total cost met
by ant k. Cmax is used in order to bring the denominator
in SC into (0,1].
Visibility (vw ) is a problem specific heuristic
information that does not change during algorithm
execution. Arcs with ending vertex equal to an intermediate vertex or to FV have their visibilities set to
1. Conversely, if w represents a component, then arcs
toward w have their visibilities defined as follows:
vw =

[vw (t)] [vw ]

=
[vu (t)] [vu ]

in which vw,k is the pheromone laid by ant k on arc

linking vertices v and w between times t and t + m,
and nAntsis the total number of ants. If ant k has not
crossed such arc during the cycle, then vw,k = 0.
Otherwise, vw,k is proportional to the entire solution
contribution (SC). Hence, in the latter case, vw,k =
Q SC, in which Q is a constant. In this work, SC is
defined as a quotient involving the obtained values by
ant k for the handled objectives:

MTTFw
cai,max

MTTFw + MTTRw
caw

(13)

where MTTFw and MTTRw are, respectively, the mean

time to failure and the mean time to repair of the component represented by w. In addition, cai , max is the
maximum acquisition cost observed in available components of the ith subsystem (phase) and caw is the
acquisition cost of the component represented by w.
Similarly to the case of SC, cai,max is used in Equation 10 to bring the cost measure into (0,1]. Note that
the first factor in Equation 11 is the mean availability of the considered component, provided that it was
renewed after corrective maintenance action. In this
paper, however, components are subjected to imperfect repairs and such interpretation only holds up to the
first failure. Besides, instead of involving other type of
costs, the factor regarding cost only uses the acquisition cost. This is due to the fact that the acquisition cost
does not change over the component lifetime. Nevertheless, Equation 11 is still applied as an attempt to
locally guide ants to vertices representing components
with higher mean availabilities and lower costs.

545

3.4

Dominance evaluation

The desired result of multiACO is a set of nondominated solutions (N ), which may contain all compromise system designs found during the algorithm run.
Therefore each ant is evaluated for each objective and
has a number of associated objective values equal to
the quantity of objectives (each objective is treated
separately).
The set N is updated at the end of each cycle.
Firstly, however, a set of candidate solutions (CS) to
be inserted in N is obtained by the assessment of the
dominance relation among ants within the cycle under
consideration. If ant k is dominated by other ants in
the current cycle, then it is not introduced in CS. Otherwise, if ant k is nondomidated in relation to all ants
in the present cycle, then it is inserted in CS. Next, it is
necessary to evaluate the dominance relation of each
element in CS in relation to solutions already stored
in N . Suppose that ant k is in CS, then: (i) if ant k is
dominated by elements in N , then ant k is ignored; (ii)
if ant k dominates solutions in N , then all solutions
dominated by ant k are eliminated from N and a copy
of ant k is inserted in N .
3.5

The algorithm

Figure 2 shows the pseudocode of the proposed multiACO. The required input data was discussed in
Subsection 3.1.

DISCRETE EVENT SIMULATION

DES attempts to mimic the behavior of real systems

by randomly generating discrete events during simulation time (Banks et al., 2001). The inherent flexibility
of DES permits the introduction of several real world
aspects in reliability problem modeling.
In this paper, DES is coupled with the multiobjective ACO described in the previous section at the
objectives evaluation stage. Each system design represented by an ant has some of its dynamic features
(e.g. mean availability, mean number of repairs per
component, among others) assessed via DES.
Suppose that ant k has its associated system design
evaluated over a predefined mission time (tM ). In DES,
firstly tM is divided into n equal steps. For each step,
a number of events (e.g. failures, start/end of repair)
are generated for every component. If a component
fails, it is necessary to assess if there are sufficient
maintenance resources to perform the repair. If so, the
failed component must wait a certain amount of time
until such resources are ready to initiate the repair.
Otherwise, if no resources are available, the failed
component is inserted in a queue in order to wait
until its required resources become available. After

Figure 2.

Pseudocode of the proposed multiobjective ACO.

a repair, the component is immediately brought back

into operation. Moreover, at the end of a step, system state (either available or unavailable) is evaluated.
Since the entire system state depends on the state of
its components, it is necessary to construct the relationship among components in order to obtain system
logic. In the implemented DES algorithm, such construction is done by means of a BDD (Binary Decision
Diagramssee Rauzy (2001)). These stochastic transitions are replicated several times for every step
until tM is reached. Figure 4 presents the coupling of
multiACO + DES.
The system availability is estimated via DES
as follows. Let zik [=1(available); =0(unavailable)],

546

and each component number of repairs and operating

time.

EXAMPLE APPLICATION

This section presents an example application in which

the desired result is a set of nondominated solutions
representing alternative designs for a system subjected
to imperfect repair. The two objectives are the mean
availability, which is obtained by means of DES, and
the system total cost. The latter objective is composed
by the acquisition (CA ), operating (CO ) and corrective
maintenance (CCM ) costs of components and also by
the costs incurred due to system unavailability (CU ).
These costs are defined as follows:
CA =

mi
s 


caij xij

(14)

i=1 j=1

Figure 3. Pseudocode
estimation.

for

the

system

availability

zk [=1(available); =0(unavailable)] and tk be the component state, system state and time at the kth step,
respectively. Moreover, let ck be a counter of the number of times that the system is available by the end
of the kth step, hi () and mi () be the time to failure
and repair time probability density functions of the
component i, and A(tk ) be the system availability at
time tk . If nC is the number of system components, a
DES iteration can be written in pseudocode as shown
in Figure 3.
In a nutshell, the algorithm described above may be
thought in the following way: while the process time
ti is lower than the mission time tk , the following steps
are accomplished: (i) the time to failure i of the component i is sampled from hi (); (ii) ti is increased by i ;
(iii) the condition (ti tk ) means component i ends
kth step on an available condition (zik = 1); otherwise,
component i failed before the kth step; (iv) in the latter, the repair time xi is sampled from mi (t) and ti is
increased by it; if (ti tk ) the component iends kth
step under a repair condition and therefore unavailable
(zik = 0); (v) upon assessing the states of the nC components at the kth step, the system state zk is assessed
via the corresponding system BDD; (vi) finally, the
counter ck is increased by zk . The aforementioned procedure is repeated M times, a sufficiently large number
of iterations, and then the availability measure A(tk )
at kth step is estimated dividing the values ck by M .
A number of random variables are obtained via DES
and fed back to multiACO with the aim of calculating
the objectives (that is, system mean availability and
system total cost): system operating/unavailable time

where s is the number of subsystems, mi is the quantity

of components in subsystem i, caij is the acquisition
cost of the jth component type of the ith subsystem
and xij is the quantity of that component.
CO =

xij
mi 
s 


coij toijk

(15)

i=1 j=1 k=1

where coij is the operating cost per unit time for the jth
component type of the ith subsystem and toijk is the
operating time of the kth copy of that component.
CCM =

xij
mi 
s 


ccmij nijk

(16)

i=1 j=1 k=1

where ccmij is the corrective maintenance cost for the

jth component type of subsystem i and nijk is the quantity of repairs to which the kth component is subjected
over mission time.
CU = cu tu

(17)

where cu is the cost per unit time related to system

unavailability and tu is the system unavailable time
during mission time. Hence, C = CA + CO + CCM +
CU , in which C is the system total cost. The variables
toijk (Equation 13), nijk (Equation 14) and tu (Equation
15) are obtained via DES of the system dynamics.
Moreover, it is considered a repairable system
with 3 subsystems in series, S1 , S2 and S3 . The
minimum and maximum number of components in
parallel and the quantity of different technologies for
each subsystem are listed in Table 2.

547

Figure 4.

Flowchart of the used multiACO + DES.

Table 2.

Subsystems features for example application.

S1
S2
S3

ni,min

ni,max

cti

1
1
1

3
6
4

5
5
3

Components are supposed to have their failure process modeled according to a GRP. More specifically,
it is assumed a Kijima Type I model with TTF given
by Weibull distributions with different scale (, in
time units), shape () and rejuvenation (q) parameters.
On the other hand, TTR are exponentially distributed
with different parameters () per component type.
In addition, components are subjected to imperfect
repairs. As soon as a component fails, the repair does
not start immediately and it is necessary to check
the availability of required maintenance resources. If
resources are available, a random time representing
the logistic time for resource acquisition is generated
according to an Exponential distribution with = 1,
i.e., the failed component must wait up to such time
to go under repair. Otherwise, the failed component
waits in queue for the required maintenance resources.

Components characteristics are presented in Table 3.

The components corrective maintenance and operating costs are taken as 2% and 10% of the acquisition
cost, respectively. Moreover, cu = 500 monetary
units.
The parameters for the multiACO were nAnts =
100, nCycles = 50, = 1, = 1, = 0.5 and Q =
1. System designs were evaluated over a mission time
equals to 365 time units. A set of 47 nondominated
solutions were obtained, as shown in Figure 5. Some
of the alternative system designs indicated in Figure 5
are presented in Figure 6.
5.1

Return of investment analysis

All system designs in the set of nondominated solutions are optimal in accordance with the multiobjective
approach. However, the decision maker may select
only one system design to be implemented. In order to
guide such selection, he can make a return of analysis
investment, that is, observe the gain in system mean
availability in relation to the required investment in
the corresponding system design. Mathematically the
return of investment (ROI ) is:
ROI = (Ak Ak1 )/(Ck Ck1 )

548

(18)

Table 3.

S1

S2

S3

Components features for example application.

fTTF

fTTR

ca

co

ccm

1
2
3
4
5
1
2
3
4
5
1
2
3

Wei(40, 1.8, 0.5)

Wei(38, 1.6, 0.5)
Wei(34, 1.5, 0.5)
Wei(31, 1.4, 0.5)
Wei(30, 1.5, 0.5)
Wei(20, 1.4, 0.4)
Wei(27, 1.5, 0.4)
Wei(22, 1.2, 0.4)
Wei(29, 1.4, 0.4)
Wei(23, 1.4, 0.4)
Wei(19, 1.2, 0.6)
Wei(25, 1.8, 0.6)
Wei(22, 1.4, 0.6)

Exp(1.5)
Exp(1.2)
Exp(0.9)
Exp(1.0)
Exp(1.1)
Exp(0.9)
Exp(1.2)
Exp(0.8)
Exp(1.1)
Exp(1.1)
Exp(0.5)
Exp(0.6)
Exp(0.6)

9900
9400
8500
8800
8200
7000
8700
7800
9100
7500
5500
5800
5200

198
188
170
176
164
140
174
156
182
150
110
116
104

990
940
850
880
820
700
870
780
910
750
550
580
520

Figure 5.

Nondominated set for example application.

where k and k 1 are adjacent solutions in the Pareto

front, and Ak , Ak1 , Ck , Ck1 are their respective
system mean availabilities and system costs.
As an example, Table 4 presents the return of investment of solutions B and C, and solutions D and E.
Note that the investment on system design is of 7,034
monetary units to go from solution B to C for a gain of
about 0.05 in mean availability, whereas such investment is 133,276 monetary units to gain about 0.001
in mean availability when upgrading from configuration D to E.

Figure 6.

System designs for selected solutions in Figure 5.

Table 4.

ROI of some selected system designs.

Solution

Mean availability

Cost

ROI

B
C
D
E

0.763128
0.816275
0.990152
0.991839

348357
355391
773650
906926

7.555.106
1.200.108

paper presented an attempt to tackle RAPs considering the more realistic repairable system behavior
consisting of imperfect repairs. This was achieved by
coupling multiobjective ACO and DES. In this context,
the dynamic behavior of potential system designs was
evaluated by means of DES, providing the decision
maker a better comprehension of the incurred costs
during mission time due to different mean availability
values.
The proposed multiobjective ACO algorithm can
certainly be improved by applying more sophisticated
pheromone updating rules and coupling it with a local
search algorithm with the aim of obtaining more accurate solutions. Moreover, some improvements can be
done in the reliability modeling. For example, the
amount of maintenance resources could be a decision variable itself, and hence the decision maker
could have the compromise system designs taking into
consideration the required quantity of maintenance
resources over mission time.

CONCLUDING REMARKS

Although there are some works in the literature applying ACO in RAPs, they often make simplifications
concerning system reliability behavior that are usually not satisfied in practical situations. Therefore this

REFERENCES
Banks, J., Carson, J.S., Nelson, B.L. & Nicol, D.M. 2001.
Discrete event system simulation. Upper Saddle River:
Prentice Hall.

549

Bowles, J.B. 2002. Commentarycaution: constant failurerate models may be hazardous to your design. IEEE
Transactions on Reliability 51(3): 375377.
Busacca, P.G., Marseguerra, M. & Zio, E. 2001. Multiobjective optimization by genetic algorithms: application to
safety systems. Reliability Engineering & System Safety
72: 5974.
Cantoni, M., Marseguerra, M. & Zio, E. 2000. Genetic
algorithms and Monte Carlo simulation for optimal plant
design. Reliability Engineering & System Safety 68:
2938.
Chiang, C.-H. & Chen, L.-H. 2007. Availability allocation
and multiobjective optimization for parallel-series systems. European Journal of Operational Research 180:
12311244.
Coello, C.A.C., Veldhuizen, D.A.V. & Lamont, G.B. 2002.
Evolutionary algorithms for solving multiobjective problems. New York: Kluwer Academic.
Deb, K. 1999. Evolutionary algorithms for multicriterion
optimization in engineering design. In: Proceedings of
Evolutionary Algorithms in Engineering and Computer
Science (EUROGEN99).
Dorigo, M., Maniezzo, V. & Colorni, A. 1996. Ant system:
optimization by cooperating agents. IEEE Transactions
on Systems, Man and Cybernetics 26(1): 2941.
Dorigo, M. & Sttzle, T. 2004. Ant colony optimization.
Massachusetts: MIT Press.
Doyen, L. & Gaudoin, O. 2004. Classes of imperfect repair
models based on reduction of failure intensity or virtual
age. Reliability Engineering & System Safety 84: 4556.
Goldberg, D.E. 1989. Genetic algorithms in search, optimization, and machine learning. Reading: Addison-Wesley.
Ippolito, M.G., Sanseverino, E.R. & Vuinovich, F. 2004.
Multiobjective ant colony search algorithm for optimal
electrical distribution system strategical planning. In:
Proceedings of 2004 IEEE Congress on Evolutionary
Computation. Piscataway, NJ.
Juang, Y.-S., Lin, S.-S. & Kao, H.-P. 2008. A knowledge
management system for series-parallel availability optimization and design. Expert Systems with Applications
34: 181193.
Kijima, M. & Sumita, N. 1986. A useful generalization of renewal theory: counting process governed by

non-negative Markovian increments. Journal of Applied

Probability 23: 7188.
Kuo, W. & Wan, R. 2007. Recent advances in optimal reliability allocation. IEEE Transactions on Systems, Man and
Cybernetics 37(4): 143156.
Liang, Y.-C. & Smith, A.E. 2004. An ant colony optimization
algorithm for the redundancy allocation problem (RAP).
IEEE Transactions on Reliability 53(3): 417423.
Luenberger, D.G. 1984. Linear and nonlinear programming.
Massachusetts: Addison-Wesley.
Michalewicz, Z. 1996. Genetic algorithms + data structures = evolution programs. Berlin: Springer.
Nahas, N. & Nourelfath, M. 2005. Ant system for reliability optimization of a series system with multiple-choice
and budget constraints. Reliability Engineering & System
Safety 87: 112.
Rauzy, A. 2001. Mathematical Foundations of Minimal
Cutsets. IEEE Transactions on Reliability 50: 389396.
Rigdon, S.E. & Basu, A.P. 2000. Statistical methods for the
reliability of repairable systems. New York: John Wiley &
Sons.
Shelokar, P.S., Jayaraman, V.K. & Kulkarni, B.D. 2002. Ant
algorithm for single and multiobjective reliability optimization problems. Quality and Realiability Engineering
International 18(6): 497514.
Taboada, H. & Coit, D.W. 2006. MOEA-DAP: a new multiple
objective evolutionary algorithm for solving design allocation problems. Under review. Reliability Engineering &
System Safety.
Taboada, H., Espiritu, J. & Coit, D.W. 2007. MOMS-GA:
a multiobjective multi-state genetic algorithm for system
reliability optimization design problems. In print. IEEE
Transactions on Reliability.
Wolsey, L.A. 1998. Integer programming. New York: John
Wiley & Sons.
Yaes, M., Joglar, F. & Modarres, M. 2002. Generalized renewal process for analysis of repairable systems
with limited failure experience. Reliability Engineering
& System Safety 77: 167180.
Zhao, J.-H., Liu, Z. & Dao, M.-T. 2007. Reliability optimization using multiobjective ant colony system approaches.
Reliability Engineering & System Safety 92: 109120.

550

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Non-homogeneous Markov reward model for aging multi-state system

under corrective maintenance
A. Lisnianski
The Israel Electric Corporation Ltd., Haifa, Israel

I. Frenkel
Center for Reliability and Risk Management, Industrial Engineering and Management Department,
Sami Shamoon College of Engineering, Beer Sheva, Israel

ABSTRACT: This paper considers reliability measures for aging multi-state system where the system and
its components can have different performance levels ranging from perfect functioning to complete failure.
Aging is treated as failure rate increasing during system life span. The suggested approach presents the nonhomogeneous Markov reward model for computation of commonly used reliability measures such as mean
accumulated performance deficiency, mean number of failures, average availability, etc., for aging multi-state
system. Corresponding procedures for reward matrix definition are suggested for different reliability measures.
A numerical example is presented in order to illustrate the approach.

INTRODUCTION

Many technical systems are subjected during their

lifetime to aging and degradation. After any failure, maintenance is performed by a repair team. The
paper considers an aging multi-state system, where the
system failure rate increases with time.
Maintenance and repair problems have been widely
investigated in the literature. Barlow & Proshan
(1975), Gertsbakh (2000), Valdez-Flores & Feldman
(1989), Wang (2002) survey and summarize theoretical developments and practical applications of
maintenance models. Aging is usually considered as
a process which results in an age-related increase of
the failure rate. The most common shapes of failure
rates have been observed by Gertsbakh & Kordonsky
(1969), Meeker & Escobar (1998), Bagdonavicius &
Nikulin (2002), Wendt & Kahle (2006). An interesting
approach was introduced by Finkelstein (2002), where
it was shown that aging is not always manifested by
the increasing failure rate. For example, it can be an
upside-down bathtub shape of the failure rate which
corresponds to the decreasing mean remaining lifetime
function.
After each corrective maintenance action or repair,
the aging systems failure rate (t) can be expressed
as (t) = q (0) + (1 q) (t), where q is an
improvement factor that characterizes the quality of
the overhauls (0 q 1), (t) is the aging systems
failure rate before repair (Zhang & Jardine 1998). If

q = 1, it means that the maintenance action is perfect

(system becomes as good as new after repair). If
q = 0, it means that the failed system is returned
back to a working state by minimal repair (system
stays as bad as old after repair), in which failure
rate of the system is nearly the same as before. The
minimal repair is appropriate for large complex systems where the failure occurs due to one (or a few)
component(s) failing. In this paper we are dealing
only with minimal repairs and, therefore, q = 0. In
such situation, the failure pattern can be described
by non-homogeneous Poisson process and reliability
measures can be evaluated by using non-homogeneous
Markov reward model.
In this paper, a general approach is suggested
for computing reliability measures for aging MSS
under corrective maintenance with minimal repair.
The approach is based on non-homogeneous Markov
reward model, where specific reward matrix is determined for finding any different reliability measure.
The main advantage of the suggested approach is that
it can be easily implemented in practice by reliability
engineers.
2

MODEL DESCRIPTION

According to the generic Multi-state System (MSS)

model (Lisnianski & Levitin 2003), MSS output performance G(t) at any instant t 0 is a discrete-state

551

continuous-time stochastic process that takes its values

from the set g = {g1 , g2 , . . ., gk }, G(t) g, where gi is
the MSS output performance in state i, i = 1, 2, . . ., k.
Transition rates (intensities) aij between states i and j
are defined by corresponding system failure and repair
rates. The minimal repair is a corrective maintenance
action that brings the aging equipment to the conditions it was in just before the failure occurrence.
Aging MSS subject to minimal repairs experiences
reliability deterioration with the operating time, i.e.
there is a tendency toward more frequent failures. In
such situations, the failure pattern can be described
by a Poisson process whose intensity function monotonically increases with t. A Poisson process with
a non-constant intensity is called non-homogeneous,
since it does not have stationary increments (Gertsbakh 2000). Therefore, in this case corresponding
transition intensities will be functions of time aij (t).
2.1

Non-homogeneous Markov reward model

A systems state at time t can be described by a

continuous-time Markov chain with a set of states
{1, . . ., K} and a transition intensity matrix a = |aij (t)|,
i, j = 1, . . ., K. For Markov reward model it is assumed
that if the process stays in any state i during the time
unit, a certain cost rii is paid. It is also assumed that
each time the process transits from state i to state j
a cost rij should be paid. These costs rii and rij are
called rewards (Hiller & Lieberman 1995). A reward
may also be negative when it characterizes a loss or
penalty. Such a reward process associated with system states or/and transitions is called a Markov process
with rewards. For such processes, in addition to a transition intensity matrix a = |aij (t)|, i, j = 1, . . ., K,
a reward matrix r = |rij |, i, j = 1, . . ., K should be
determined (Carrasko 2003).
Let Vi (t) be the expected total reward accumulated
up to time t, given the initial state of the process as
time instant t = 0 is in state i. According to Howard
(1960), the following system of differential equations
must be solved under specified initial conditions in
order to find the total expected rewards:

If, for example, the state K with highest performance level is defined as the initial state, the value
VK (t) should be found as a solution of the system (1).
It was shown in Lisnianski (2007) and Lisnianski
et al., (2007) that many important reliability measures
can be found by the determination of rewards in a
corresponding reward matrix.

2.2 Rewards determination for computation

of different reliability measures
For an availability computation, we partition the set of
states g, into g 0 , the set of operational or acceptable
system states, and g f , the set of failed or unacceptable states. The system states acceptability depends
on the relation between the MSS output performance
and the desired level of this performance demand,
which is determined outside the system. In general
case demand W (t) is also a random process that can
take discrete values from the set w = {w1 , . . ., wM }.
The desired relation between the system performance
and the demand at any time instant t can be expressed
by the acceptability function (G(t), W (t)) (Lisnianski & Levitin 2003). The acceptable system states
correspond to (G(t), W (t)) 0 and the unacceptable states correspond to(G(t), W (t)) < 0. The last
inequality defines the MSS failure criterion. In many
practical cases, the MSS performance should be equal
to or exceed the demand. Therefore, in such cases the
acceptability function takes the following form:
(G(t), W (t)) = G(t) W (t)

and the criterion of state acceptability can be expressed as

(G(t), W (t)) = G(t) W (t) 0

j =i

(1)


In the most common case, MSS begins to accumulate rewards after time instant t = 0, therefore, the
initial conditions are:
Vi (0) = 0,

i = 1, 2, . . ., K

(2)

(4)

Here without loss of generality we assume that

required demand level is constant W (t) w and
all system states with performance greater than or
equal to w corresponds to the set of acceptable states
and all system states with performance lower than w
correspond to the set of unacceptable states.
We define the indicator random variable

K
K


dVi (t)
= rii +
aij (t)rij +
aij (t)Vj (t)
dt
j=1
j=1

i = 1, 2, . . ., K

(3)

I (t) =

1, if G(t) g0 ,
0 otherwise.

(5)

The MSS instantaneous (point) availability A(t) is

the probability that the MSS at instant t > 0 is in one

552

in a Markov reward model should be defined as

of acceptable states:
A(t) = Pr{I (t) = 1} =

Pi (t)

(6)

ig0

where Pi (t) is the probability that at instant t the system

is in state i.
For an aging MSS an average availability is often
used. The MSS average availability A (T ) is defined as
a mean fraction of time when the system resides in the
set of acceptable states during the time interval [0, T ],
1
A (T ) =
T


rjj =

w gj ,
0,

(7)

w gj > 0,
w gj 0.

(10)

All transitions rewards rij , i  = j should be zeroed.

Therefore, the mean reward Vi (T ) accumulated during the time interval [0, T ], if state i is in the initial
state, defines the mean accumulated performance
deficiency

T
A(t)dt

if
if

Vi (T ) = E

T


(W (t) G(t))dt

(11)

To assess A(T ) for MSS, the rewards in matrix r

can be determined in the following manner.
The rewards associated with all acceptable states
should be defined as 1.
The rewards associated with all unacceptable states
should be zeroed as well as all the rewards associated
with all transitions.
The mean reward Vi (T ) accumulated during interval [0, T ] defines a time that MSS will be in the set of
acceptable states in the case where state i is the initial
state. This reward should be found as a solution of the
system (1). After solving the system (1) and finding
Vi (t), MSS average availability can be obtained for
every i = 1, . . . , K:

Ai (T ) = Vi (T ) T

(8)

Usually the state K is determined as an initial state

or in another words the MSS begins its evolution
in the state space from the best state with maximal
performance.
Mean number Nfi (T ) of MSS failures during the
time interval [0, T ], if state i is the initial state. This
measure can be treated as a mean number of MSS
entrances into the set of unacceptable states during
the time interval [0, T ]. For its computation rewards
associated with each transition from the set of acceptable states to the set of unacceptable states should
be defined as 1. All other rewards should be zeroed.
In this case the mean accumulated reward Vi (T ),
obtained by solving (1) provides the mean number of
entrances into the unacceptable area during the time
interval [0, T ]:
Nfi (T ) = Vi (T )

(9)

Mean performance deficiency accumulated within

interval [0, T ]. The rewards for any state number j

where E - expectation symbol and G(0) = gi .

Mean Time To Failure (MTTF) is the mean time
up to the instant when the MSS enters the subset of
unacceptable states for the first time. For its computation the combined performance-demand model should
be transformed - all transitions that return MSS from
unacceptable states should be forbidden, as in this case
all unacceptable states should be treated as absorbing
states.
In order to assess MTTF for MSS, the rewards
in matrix r for the transformed performance-demand
model should be determined as follows:
The rewards associated with all acceptable states
should be defined as 1.
The reward associated with unacceptable (absorbing) states should be zeroed as well as all rewards
associated with transitions.
In this case, the mean accumulated reward Vi (t)
defines the mean time accumulated up to the first
entrance into the subset of unacceptable states or
MTTF, if the state i is the initial state.
Reliability function and Probability of MSS failure
during the time interval [0, T ]. The model should be
transformed as in the previous case all unacceptable
states should be treated as absorbing states and, therefore, all transitions that return MSS from unacceptable
states should be forbidden. Rewards associated with
all transitions to the absorbing state should be defined
as 1. All other rewards should be zeroed. The mean
accumulated reward Vi (T ) in this case defines the
probability of MSS failure during the time interval
[0, T ], if the state i is the initial state. Therefore, the
MSS reliability function can be obtained as:
Ri (T ) = 1 Vi (T ), i = 1, . . ., K

553

(12)

NUMERICAL EXAMPLE

According to the state space diagram in Figure 1 the

following transition intensity matrix a can be obtained:

Consider a multi-state power generating unit with

nominal generating capacity 360 KWT. Corresponding multi-state model is presented in fig.1 and has 4
different performance levels: complete failure level
(g1 = 0), two levels with reduced capacity (g2 =
215 KWT, g3 = 325 KWT), and level of perfect
functioning (g4 = 360 KWT).
Aging was indicated as increasing transition failure
rate 42 (t) = 7.01 + 0.2189t 2 . Other failure rates are
constant: 41 = 2.63 year 1 and 43 = 13.14 year 1
Repair rates are the following: 14 = 446.9 year 1 ,
24 = 742.8 year 1 , 34 = 2091.0 year 1 .
The demand is constant w = 300 KWT and power
unit failure is treated as generating capacity decreasing
below demand level w.
The state-space diagram for the system is presented
in Figure 1.
By using the presented method we assess the MSS
average availability, mean total number of system
failures, accumulated mean performance deficiency,
Mean Time To Failure and Reliability function for 5
years time interval.


14

 0
a=
 0

41

0
24
0
42 (t)

0
0
34
43



14

24


34

(41 + 42 (t) + 43 )
(13)

)
In order to find the MSS average availability A(T
according to introduced approach we should present
the reward matrix r in the following form.

0
   0
r = r ij  = 
0
0

0
0
0
0

0
0
1
0


0

0
0 
1

(14)

The system of differential equations (1) will be

presented as the following:

g4 = 360

43

Figure 1.

24

42 (t)

w = 300

g1= 0

34

g3 = 325

g2 = 215

dV1 (t)
dt
dV2 (t)
dt
dV3 (t)
dt
dV4 (t)
dt

41

= 14 V1 (t) + 14 V4 (t)
= 24 V2 (t) + 24 V4 (t)
= 1 34 V3 (t) + 34 V4 (t)
= 1 + 41 V1 (t) + 42 (t) V2 (t) + 43 V3 (t)
(41 + 42 (t) + 43 ) V4 (t)

14

State space diagram of generated system.

(15)

The system of differential equations must be sold

under initial conditions: Vi (0) = 0, i = 1, 2, 3, 4.
The results of calculation one can see in Figure 2.
Calculation results are presented for two cases: for
aging unit with 42 (t) = 7.01 + 0.2189t 2 and for nonaging unit where 42 = 7.01 constant.
In order to find the mean total number of system
failures Nf (t) we should present the reward matrix r
in the following form:


  


r = rij = 



554

0
0
0
1

0
0
0
1

0
0
0
0

0
0
0
0

(16)

The system of differential equations must be sold

under initial conditions: Vi (0) = 0, i = 1, 2, 3, 4.
The results of calculation are presented in Figure 3.
In order to find Accumulated Performance Deficiency we should present the reward matrixes r in the
following form:

Average Availability

0.998
0.996
0.994


 300
   0
r = r ij  = 
 0
 0

0.992
0.99
0.988
0.986
0

2
3
Time (years)

(18)

The system of differential equations (1) will be

presented as follows:

Figure 2. Calculation the MSS average availability. (Dash

dot line: 42 (t) = 7.01 + 0.2189t 2 . Bold line: 42 = 7.01
constant).

Mean Total Number of System Failures


0 0 0 

85 0 0 
0 0 0 
0 0 0 

dV1 (t)
dt
dV2 (t)
dt
dV3 (t)
dt
dV4 (t)
dt

60
50
40

= 300 14 V1 (t) + 14 V4 (t)

= 85 24 V2 (t) + 24 V4 (t)
= 34 V3 (t) + 34 V4 (t)
= 41 + 42 + 41 V1 (t) + 42 (t) V2 (t)
+ 43 V3 (t) (41 + 42 (t) + 43 ) V4 (t)

30

(19)

20
10
0
0

2
3
Time (years)

Accumulated Performance Deficiency (KW.Hours)

Figure 3. Mean total number of system failures. (Dash dot

line: 42 (t) = 7.01 + 0.2189t 2 . Bold line: 42 = 7.01
constant).

The system of differential equations must be sold

under initial conditions: Vi (0) = 0, i = 1, 2, 3, 4.
The results of calculation are presented in Figure 4.
For computation of the Mean Time To Failure and
the Probability of MSS failure during the time interval
the state space diagram of generated system should be
transformedall transitions that return system from

Then the system of differential equations (1) will

be obtained:
dV1 (t)
dt
dV2 (t)
dt
dV3 (t)
dt
dV4 (t)
dt

= 14 V1 (t) + 14 V4 (t)
= 24 V2 (t) + 24 V4 (t)
= 34 V3 (t) + 34 V4 (t)
= 41 + 42 + 41 V1 (t) + 42 (t) V2 (t)
+ 43 V3 (t) (41 + 42 (t) + 43 ) V4 (t)
(17)

61320
52560
43800
35040
26280
17520
8760
0
0

2
3
Time (years)

Figure 4. Accumulated performance deficiency. (Dash dot

line: 42 (t) = 7.01 + 0.2189t 2 . Bold line: 42 = 7.01
constant).

555

unacceptable states should be forbidden and all unacceptable states should be treated as absorbing state.
The state space diagram may be presented as follows.
According to the state space diagram in Figure 5
transition intensity matrix a can be presented as
follows:


0

0
a= 
 41 + 42 (t)

dV0 (t)
=0
dt
dV3 (t)
= 1 34 V3 (t) + 34 V4 (t)
dt
dV4 (t)
= 1 + (41 + 42 (t)) V0 (t) + 43 V3 (t)
dt
(41 + 42 (t) + 43 ) V4 (t)
(22)



0
0

,
34
34

43 (41 + 42 (t) + 43 ) 
(20)

In order to find Mean Time To Failure we should

present the reward matrixes r in the following form


   0
r = r ij  =  0
 0

The system of differential equations (1) will be

presented as follows:

0
1
0

0
0
1

(21)

The system of differential equations must be sold

under initial conditions: Vi (0) = 0, i = 0, 3, 4.
The results of calculation are presented in Figure 6.
In order to find Probability of MSS failure during
the time interval [0, T ] we should present the reward
matrixes r in the following form




   0 0 0 



r = r ij =  0 0 0 
(23)
 1 0 0 
The system of differential equations (1) will be
presented as follows:

g4=360

dV0 (t)
=0
dt
dV3 (t)
= 34 V3 (t) + 34 V4 (t)
dt
dV4 (t)
= 41 + 42 (t) + (41 + 42 (t)) V0 (t)
dt
+ 43 V3 (t) (41 + 42 (t) + 43 ) V4 (t)

34

(24)
The system of differential equations must be sold
under initial conditions: Vi (0) = 0, i = 0, 3, 4.

0.12

Mean Time to Failure (years)

g3=325

w=300

0
Figure 5. State space diagram of generated system with
absorbing state.

0.1
0.08
0.06
0.04
0.02
0
0

Figure 6.

556

0.2

0.4
0.6
Time (Years)

Mean time to failure.

0.8

MSS Reliability Function

1
0.8
0.6
0.4
0.2
0
0

Figure 7.
[0, T ].

0.1

0.2
0.3
Time (years)

0.4

0.5

MSS reliability function during the time interval

The results of calculation the MSS reliability function according the formulae (12) are presented in
Figure 7.
From all graphs one can see age-related unit reliability decreasing compared with non-aging unit. In the
last two figures graphs for mean time to failure and
reliability functions for aging and non-aging unit are
almost the same, because of the fact that first unit failure usually occurs within short time (less than 0.5 year
according to Figure 7) and aging impact is negligibly
small for such short period.
4

CONCLUSION

1. The non-homogeneous Markov reward model was

developed for reliability measures computation for
aging MSS under corrective maintenance with
minimal repair.
2. The method is based on a different reward matrix
determination for the non-homogeneous Markov
reward model.
3. The suggested approach is well formalized and
suitable for practical application in reliability engineering.
4. The numerical example is presented in order to
illustrate the suggested approach.

Barlow, R.E. & Proshan, F. 1975. Statistical Theory of Reliability and Life Testing. Holt, Rinehart and Winston:
New York.
Carrasco, J. 2003. Markovian Dependability/Performability
Modeling of Fault-tolerant Systems. In Hoang Pham (ed),
Handbook of Reliability Engineering: 613642. London,
NJ, Berlin: Springer.
Finkelstein, M.S. 2002. On the shape of the mean residual
lifetime function. Applied Stochastic Models in Business
and Industry 18: 135146.
Gertsbakh, I. 2000. Reliability Theory with Applications to
Preventive Maintenance. Springer-Verlag: Berlin.
Gertsbakh, I. & Kordonsky, Kh. 1969. Models of Failures.
Berlin-Heidelberg-New York: Springer.
Hiller, F. & Lieberman, G. 1995. Introduction to Operation
Research. NY, London, Madrid: McGraw-Hill, Inc.
Howard, R. 1960. Dynamic Programming and Markov
Processes. MIT Press: Cambridge, Massachusetts.
Lisnianski, A. 2007. The Markov Reward Model for a
Multi-state System Reliability Assessment with Variable
Demand. Quality Technology & Quantitative Management 4(2): 265278.
Lisnianski, A., Frenkel, I., Khvatskin, L. & Ding Yi. 2007.
Markov Reward Model for Multi-State System Reliability Assessment. In F. Vonta, M. Nikulin, N. Limnios, C.
Huber-Carol (eds), Statistical Models and Methods for
Biomedical and Technical Systems. Birkhaser: Boston,
153168.
Lisnianski, A. & Levitin, G. 2003. Multi-state System Reliability. Assessment, Optimization and Applications. World
Scientific: NJ, London, Singapore.
Meeker. W. & Escobar, L. 1998. Statistical Methods for
Reliability Data. Wiley: New York.
Trivedi, K. 2002. Probability and Statistics with Reliability,
Queuing and Computer Science Applications. New York:
John Wiley & Sons, Inc.
Valdez-Flores, C & Feldman, R.M. 1989. A survey of
preventive maintenance models for stochastically deteriorating single-unit systems. Naval Research Logistics 36:
419446.
Wang, H. 2002. A survey of Maintenance Policies of
Deteriorating Systems. European Journal of Operational
Research 139: 469489.
Wendt, H. & Kahle, W. 2006. Statistical Analysis of Some
Parametric Degradation Models. In M. Nikulin, D. Commenges & C. Huber (eds), Probability, Statistics and
Modelling in Public Health: 26679. Berlin: Springer
Science + Business Media.
Zhang, F. & Jardine, A.K.S. 1998. Optimal maintenance
models with minimal repair, periodic overhaul and complete renewal. IIE Transactions 30: 11091119.

REFERENCES
Bagdonavicius, V. & Nikulin, M.C. 2002. Accelerated life
models. Boca Raton: Chapman & Hall/CRC.

557

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On the modeling of ageing using Weibull models: Case studies

Pavel Praks, Hugo Fernandez Bacarizo & Pierre-Etienne Labeau
Service de Mtrologie Nuclaire, Facult des Sciences Appliques, Universit Libre de Bruxelles (ULB), Belgium

ABSTRACT: Weibull models appear to be very flexible and widely used in the maintainability field for
ageing models. The aim of this contribution is to systematically study the ability of classically used one-mode
Weibull models to approximate the bathtub reliability model. Therefore, we analyze lifetime data simulated
from different reference cases of the well-known bathtub curve model, described by a bi-Weibull distribution
(the infant mortality is skipped, considering the objective of modeling ageing). The Maximum Likelihood
Estimation (MLE) method is then used to estimate the corresponding parameters of a 2-parameter Weibull
distribution, commonly used in maintenance modeling, and the same operation is performed for a 3-parameter
Weibull distribution, with either a positive or negative shift parameter. Several numerical studies are presented,
based first on large and complete samples of failure data, then on a censored data set, the failure data being
limited to the useful life region and to the start of the ageing part of the bathtub curve. Results, in terms of
quality of parameter estimation and of maintenance policy predictions, are presented and discussed.
1

INTRODUCTION

Our contribution is motivated by a common situation

in industrial maintenance, where the failure dataset is
often limited, not only in size, but also in time to the
useful period and the very beginning of ageing, thanks
to the preventive maintenance operations. Although
lots of Weibull models are available in the literature
(for detailed information, consult e.g. (Murthy, Xie,
and Jiang 2004)), there is a trend in industry to use,
for the maintenance applications, simple one-mode
Weibull models with a limited number of parameters
to cover such cases.
The main aim of the paper is to analyze how acceptable this approximation is, by systematically studying
the ability of classically used one-mode Weibull models to approximate the bathtub reliability model. For
this reason, we would like to illustrate how satisfactorily the ageing model corresponding to the bathtub
curve reliability model - in which the infant mortality
is skipped, considering the objective of modeling ageing - can be estimated by a one-mode Weibull model
with 2 or 3 parameters. Indeed, 2- or 3- parameter
Weibull models though commonly used in maintenance optimization do not map well more than one
part of the bathtub curve reliability model. We will see
that randomly sampling failure times from the bathtub
model and subsequently estimating by the maximum
likelihood method the parameters of one-mode, 2- or
3- parameter Weibull models can lead to unexpected
results in terms of parameter estimation as well as to
erroneous predictions of an optimal maintenance policy. Although properties of Weibull models are widely

studied, a similar systematical benchmark study is

hardly found in the current literature.
The paper is organized as follows. We begin with a
definition of the four-parameter model describing the
assumed reality (i.e. our reference cases) in Section 2.
In Section 3 we present the one-mode Weibull models with whom the reference cases are approximated.
Section 4 deals with the maintenance cost models that
have been chosen to compare the optimal maintenance
period predictions associated to the reference cases
and to the one-mode Weibull approximations. Parameter estimations for one-mode Weibull models inferred
from samplings of the reference bathtub curves are
presented in Section 5. Results are presented and
discussed in Section 6 according to the quality of
the lifetime distribution and to the optimal maintenance period predicted, as mentioned above. Finally,
conclusions are summarized in Section 7.
2

FOUR-PARAMETER MODEL OF THE

BATHTUB CURVE

We will use as reference model a three-parameter

Weibull distribution with a positive location parameter, combined with an exponential model. We therefore
assume the following hazard function:


t 1
(t) = 0 +
H (t )
(1)

Symbol 0 0 denotes the failure rate associated to random failures. Symbols >0, >0 and

559

 0 are called scale, shape and location parameters,

respectively. H (.) is the step function.
The 3-parameter Weibull rate models the failure
mode due to ageing, and is active only for t>. In
other words, a failure of the system in time t>0 can
occur by:
a) a contribution due to the exponential distribution,
which describes a random failure of the system for
any t>0
b) a contribution due to the Weibull distribution for
t> , where is a positive location parameter (a shift).
It means that we assume that this failure mode, which
is based on the effect of ageing of the system, can
occur only if t>.

APPROXIMATE WEIBULL MODELS USED

IN THE COMPARISON

We assume the following probability density functions

(PDF) for our parameter estimation study:
Weibull model with 2 parameters:
fa (t) =

a
a

t
a

a 1

t
a

t
a

a

(2)

a

(3)

The scale parameter symbol a has the following

meaning: Solving equation (3) for t = a we have:


Fa (a )

a
a

a

= 1 e1

1e

1 0.3679 = 0.6321

(4)

So, the scale parameter a is the 63rd percentile

of the two-parameter Weibull distribution. The scale
parameter is sometimes called the characteristic
life. The shape parameter describes the speed of
the ageing process in Weibull models and has no
dimension.
Weibull model with 3 parameters including a positive shift >0:
fb (t) =

b
b

t b
b

b 1

tb
b

fc (t) =

c
c

tc
c

c 1

e
4

c
c

tc
c

c

c

H (t)

(6)

MAINTENANCE COSTS ANALYSIS

We assume the following hypotheses in our analysis

of the maintenance policy:
Costs are kept constant and do not change with time
(interest rate or inflation are not considered).
Repair durations are contemplated as negligible.
Failures are detected instantaneously.
Labour resources are always available to repair.
The following three maintenance strategies (Kececioglu 1995; Vansnick 2006) are considered.
4.1 Periodic replacement: As Bad as Old (ABAO)

When = 0 and 0 = 0, the model is reduced

to the 2-parameter Weibull law. The corresponding
cumulative distribution function (CDF) is:
Fa (t) = 1 e

avoiding to implicitly assume c (0) = 0 The PDF

has the following conditional form:

H (t b ) (5)

In this policy an item is replaced with a new one at

every Tp time units of operation, i.e. periodically at
time Tp , 2Tp , 3Tp , . . . . If the component fails before
Tp time units of operation, it is minimally repaired so
that its instantaneous failure rate (t) remains the same
as it was prior to the failure.
The expected total cost will be presented per unit
time and per preventive cost, and the predetermined
maintenance interval is denoted by Tp . For the analysis,
it is interesting to introduce a new coefficient r, the
ratio of the curative cost over the preventive cost:
r=

Cc
Cp

(7)

The total expected cost per unit time can be

written as
(Tp ) =

Cp + Cc E(N (Tp ))
Tp

(8)

where E(N (Tp )) is the expected number of failures in

time interval [0, Tp ]. If the element is not replaced
but only minimally repaired (ABAO), the expected
number of failures at time Tp is
 Tp
 0
(t) dt.
k dt
(9)
E(N (Tp )) =
0

tp

Finally, the total expected cost per unit time and per
preventive cost is given by

Weibull with 3 parameters, including a negative

shift <0. The Weibull model with a negative shift
can be used to model pre-aged components, hence

560

(Tp )
1
r
=
+ E(N (Tp )).
Cp
Tp
Tp

(10)

4.2

Age replacement: AGAN (As-Good-As-New)

The AGAN maintenance strategy consists of replacing the component with a new one after it has failed
or when it has been in operation for Tp time units,
whichever comes first. The total expected total cost per
unit time per preventive cost of a component depends
on a summation of contributions of preventive and
curative costs:
(Tp )
R(Tp ) + r(1 R(Tp ))
.
=
 Tp
Cp
R(t) dt

(11)

4.3

Block replacement: AGAN

In this maintenance strategy an item is replaced

preventively (maximum repair) at time points
Tp , 2Tp , 3Tp , and so on. In addition to this, the component is maintained correctively (maximum repair)
each time it fails. The total expected total cost per
unit time per preventive cost can be expressed as
(Kece-cioglu1995)
(Tp )
1 R(Tp )
1
.
=
+ r  Tp
Cp
Tp
R(t) dt

(12)

PARAMETER ESTIMATIONS OF ONE-MODE

WEIBULL LAWS BASED ON SAMPLINGS
FROM THE BATHTUB CURVE MODEL

The aim of the section is to show the possible paradox

between the classical reference to the bathtub-shaped
failure rates and the usual way of resorting to basic,
one-mode Weibull laws to model ageing and deduce
maintenance policy optimization. This is first performed by assuming that the real behavior of the
failure rate obeys a bathtub shape, in which the infant
mortality is neglected (so that a bi-Weibull is used),
by sampling failure data from this assumed reality,
then by estimating the parameters of reduced 2- or
3-parameter Weibull models and by comparing the
predictions of optimal maintenance period based on
this estimation with that deduced from the assumed
reality. In a second time, we also analyse failure data
censored at the real MTTF point, which is an approximation of the industrial reality where data are usually
limited.
Parameters (, , ) and 0 have to be chosen to
define what is the initial reality of the failure density
of the component.
Actually, three of these parameters correspond to
characteristic times:

parameter is the characteristic time of the ageing

failure mode, and
parameter is a time shift corresponding to the delay
before the onset of ageing.
5.1 Characterization of useful period
In this section, we specify the value of the parameter
related to the useful period (i.e. period with a constant
failure rate 0 ) of our model of reality.
We set 1/0 = 14000 h, because this value is
close to that estimated for reciprocating compressors
whose modelling were an inspiration of our study
(Vansnick 2006). This arbitrary choice hence defines
the time units of the problem, and only 3 independent
parameters remain.
5.2

Parameters describing ageing

In this section, we specify values of parameters related

to the ageing part of our model of reality. We assume
the following parameters in our tests:
The values (2, 2.5, 3, 3.5, 4) are assumed for the
shape parameter . We have chosen these values
because when = 2, the increasing failure rate part
(due to ageing), is increasing linearly with time and
for values of higher than 2 the curve is increasing
non-linearly.
The value of the scale parameter is selected in the
following way:
Since MTTF = + (1 + (1/)) holds for a
single Weibull law, the straightforward observation
is that is a characteristic time on which the ageing
process spans.
Then we can use the following quantity 0 as a
measure of the relative duration of the ageing failure
mode with respect to the purely random one. Let us
assume that 0 < = 1 holds in our cases because
the ageing period lasts less time than the useful life
period. Then we use the following range for values
(0 is kept constant in all cases):
0 = (1, 0.9, 0.8, 0.7, 0.6)

(13)

Even if this assumption might be valid in many

cases, in some situations however, one could meet
a behavior of the component where ageing starts
quite quickly, even if it is then likely to be a slower
process than in other cases.
The value of the location parameter is selected
from the following equation:

parameter 1/0 represents the Mean Time To Failure (MTTF) of the random failure mode taken
alone,

561

R() = (0.9, 0.8, 0.7, 0.6, 0.5, 0.4)

(14)

Here R() denotes the reliability function of the

reference model at the end of the useful period

(i.e. period with constant failure rate 0 ):

R() = 1 F() = exp(0 )

(15)

R() is chosen between 0.9 and 0.4 because for values over 0.9 the useful life would be too short and
values under 0.4 are not worth to be analysed. This
is because, for a constant failure rate, the residual
reliability at t = MTTF is equal to 0.3679. Due to
this, R() = 0.4 is a good limit as in practice a piece
of equipment will not be operated longer.
In order to express what reality is being considered in each case, it can be specified by the
3-uple (0 , , R()). For example, the reality
(0 = 1, = 3, R() = 0.8) can be expressed as
(0 , , R()) = (1, 3, 0.8).
In total, 5 5 6 = 150 samples of data have been
analysed, in each of the following two situations:
1. In the first situation, we used a complete set of
N = 1000 failure data taken from the Model of
Reality, see eq. 1, with the purpose of estimating
the quality of the approximation obtained with the
single-mode Weibull laws.
2. In the second situation we used a set of N = 100
observation data, however in this case suspended
data are also considered. The MTTF point is chosen as the predetermined maintenance interval Tp .
All values greater than this period will be considered as suspended data. This is an approximation of
the industrial reality where failure data are usually
limited.
6

RESULTS OF THE ESTIMATIONS

In this section the predetermined maintenance interval

(Tp ) obtained with the 2- and 3- parameters Weibull
models will be compared to the Tp given by the
reference model, which was presented in Section 5.
As examples, the next figures show the results of the
analysis done in these two different situations. Figures
1 and 2 show sorted results (i.e. in the 150 reference
cases considered) of the optimal predetermined maintenance intervals for the reference model (denoted as
4P ABAO) with bounds corresponding to the minimum cost multiplied by the factor 1.05 (denoted as
4P Int. Sup and 4P Int. Inf, respectively). We assume
the ratio between the curative cost over the preventive cost r = 10 in all cases. These figures also show
results of the poor ability of the 2- and 3- parameters Weibull models (denoted as 2P ABAO and 3P
ABAO, respectively) to provide fitting of the assumed
reality.
In the first situation, the results estimated from 2and 3- parameters Weibull models are conservative:
the estimated values of the shape parameter were in

Figure 1. Optimal predetermined maintenance intervals for

Situation 1.

Figure 2. Optimal predetermined maintenance intervals for

Situation 2.

the interval [1.17, 1.9] for the 3-parameter Weibull

model with a positive location parameter, and the same
range was obtained for the 2-parameter Weibull model.
On this account, the ability of these approximations
to model the wear-out effects given by our reference
model is limited. It does not cover the reference model,
which is represented by the shape parameters within
the interval [2, 4]. In fact, the estimation using the
Weibull model with a positive location parameter is
not very beneficial: Because of random failures, represented by the second part of the bathtub model,
the effect of a positive shift of the location parameter is negligible in our cases: the estimated values
of the location parameter were very close to zero
( 2.108 h).
It seems that the paradox in the Weibull estimation leads to a conservative maintenance policy in
the first situation. Optimal maintenance intervals estimated by one-mode Weibull models are almost in all
cases smaller compared to the optimal maintenance
period of the assumed reference model: Estimated values of Tp for the 3-parameter Weibull model with a
positive location parameter and also for 2-parameter

562

Figure 3. Parameters of the reality (1, 3, 0.8), Situation 1.

Failure rate.

Figure 4. Parameters of the reality (1, 3, 0.8), Situation 1.

Expected cost per unit time and per preventive cost.

Weibull model correspond to 35 82% of the values of Tp obtained with the reference models. When
the 3-parameter Weibull model with a negative location parameter is assumed, the situation is better: The
estimated values of Tp reached 43-100 percent of the
reference Tp . In 7 cases (i.e. in 4.66% of all assumed
cases), the value of Tp was correctly estimated.
In the second situation (i.e. with censored data), the
results estimated from the 2- and 3- parameter Weibull
models are also not desirable: In 36.66% of the estimations done we obtain a value of the shape parameter
<1 (a rejuvenation!). This means that there is no
minimum of the expected cost per unit time and per
preventive cost function. It is thus not interesting to
schedule a systematic preventive maintenance; this is
completely wrong in comparison the reference model.
In 43.34% of the estimations, the estimated value of the
shape parameter obtained lies in the interval [1, 1.2].
This values also does not correspond to the reference
model and the expected cost per unit time and per preventive cost becomes almost constant. Finally, in 20%
of the analyses, the estimated value of the shape parameter is between 1.2 and 1.62. This fact does not enable
us to model correctly the wear-out effects given our
reference model.
As it was said in Section 5.2, in order to express
what reality is being considered in each case, the
parameters that determine it can be expressed by the
3-uple (0 , , R()). Let us analyse the behavior of
two selected realities more deeply.

In Figure 3, the hazard function of the reality and

its 2- or 3-parameters approximations for the first situation are presented. The symbol Zero represents
the MLE results related to the Weibull model with 2
parameters. The symbol Neg represents the Weibull
model with 3 parameters including a negative shift
and finally the symbol Pos represents the Weibull
with 3 parameters, including a positive shift. We can
see that the wear-out period of the assumed bathtub
reality is hardly modelled by the one-mode distributions. Although the Weibull 2P and 3P approximations
model accurately only the useful period of the reality,
it can be seen that it is not a problem to find the optimal
preventive maintenance period Tp in both situations,
see Figures 4 and 5. This happens because in these situations the optimal preventive maintenance period is
found before the ageing process becomes strong. For
this reason, approximations appear acceptable.
In this example it is also worth noticing the
MLE results obtained with the Weibull 2P and 3P
approximations in the second situation: (, , ) =
(12487, 1.11, 0) and (, , ) = (12319, 1.18, 91).
We observe that these results of parameter estimations
do not imply similar results in estimating of Tp s, see
Figure 5. This little difference in shape parameters
implies that the Weibull 2P displays difficulties to find
a minimum in its corresponding expected cost and per
unit time per preventive cost function.

6.1

6.2 Reality (0 , , R()) = (0.6, 2, 0.6)

Reality (0 , , R()) = (1, 3, 0.8)

In this case, obtaining good policy predictions may not

be determined by the wear-out period approximation
quality in this case.

563

Figure 6 contains failure rates for situation 1. Although

cost predictions in situation 1 remain numerically
non-problematic, see Figure 7, both approximation
functions display difficulties to find a minimum in
their corresponding expected cost per unit time and

Figure 7. Parameters of the reality (0.6, 2, 0.6), Situation 1.

Expected cost per unit time and per preventive cost.

Figure 5. Parameters of the reality (1, 3, 0.8), Situation 2.

Expected cost per unit time and per preventive cost.

Figure 6. Parameters of the reality (0.6, 2, 0.6), Situation 1.

Failure rate.

per preventive cost function in the second situation,

such like 2P approximations from Figure 5: The MLE
results obtained with the Weibull 2P and 3P approximations in the second situation have shape parameters
very close to 1: (, , ) = (10355, 1.09, 0) and
(, , ) = (10353, 1.12, 62).

CONCLUSIONS

The main aim of the paper was to systematically

study the ability of classical one-mode Weibull models to approximate the bathtub reliability model. The
novelty of the work includes a construction of a benchmark study, in which 150 carefully selected test-cases

were analyzed. We presented various numerical experiments showing that a random sampling from the
well-known bathtub reliability model and subsequent
MLE of Weibull models with 2 or 3 parameters lead
to potentially dangerous shortcuts.
For huge, complete sets of failure data, the estimated shape parameters are almost in all cases inside
interval [1.17, 1.9] for the Weibull model with a positive location parameter or close to 6, when the Weibull
model with a negative location parameter is assumed.
These estimations do not correspond at all to wear-out
failures represented in our reference model by shape
parameters within the interval [2, 4]. This paradox in
the Weibull estimation leads to a conservative maintenance policy, which involves more cost and can be
even risky: Too short maintenance periods can reduce
the reliability of the system, because not all of these
interventions finish successfully. Moreover, this estimation drawback is observed in cases with a large set
of failure data (1 000).
For limited sets with censored data, the estimated
values of the shape parameter are often close to
one or even less than one and estimated predetermined maintenance intervals are not connected to
the reference model. It could be very interesting to
repeat tests contained in this contribution and also
to provide sensitivity analysis of estimated parameters with goodness-of-fit tests (Meeker and Escobar
1995).
The message of our contribution should be kept
in mind during the Weibull parameter estimation
process, in which also real failure data are analyzed:
Although the Weibull models are flexible and widely
used, we would like to point limitations of ability
of one-mode Weibull models with 2 or 3 parameters

564

to successfully fit more than exactly one part of the

bathtub curve reliability model.
ACKNOWLEDGMENT
Pavel Praks (from VB-Technical University of
Ostrava, the Czech Republic)s postdoctoral stay at the
Universit Libre de Bruxelles, Belgium is part of the
ARC Project - Advanced supervision and dependability of complex processes: application to power
systems.

REFERENCES
Kececioglu, D. (1995). Maintainability, Availability and
Operational Readiness Engineering Handbook. Prentice
Hall.
Meeker, W. Q. and L. A. Escobar (1995). Statistical Methods
for Reliability Data. John Wiley & Sons.
Murthy, D. N. P., M. Xie, and R. Jiang (2004). Weibull Models.
John Wiley & Sons.
Vansnick, M. (2006). Optimization of the maintenance of
reciprocating compress or based on the study of their performance deterioration. Ph. D. thesis, UniversitLibre de
Bruxelles.

565

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On-line condition-based maintenance for systems with several modes

of degradation
Amlie Ponchet, Mitra Fouladirad & Antoine Grall
Universit de Technologie de Troyes, Institut Charles Delaunay, France

ABSTRACT: In this paper, a maintenance model for a deteriorating system with several modes of degradation
is proposed. The time of change of mode and parameters after the change are unknown. A detection procedure
based on an on-line change detection/isolation algorithm is used to deal with unknown change time and unknown
parameters. The aim of this paper is to propose an optimal maintenance versus detection policy in order to
minimize the global maintenance cost.

INTRODUCTION

In this paper, we consider a system which nominally

deteriorates according to a given mode. The mean
deteriorating rate increases suddenly at an unknown
time. It is supposed that at this time a change of mode
is occurred. On-line information is available from online monitoring about the change of mode. To deal with
the unknown change time an adequate on-line change
detection algorithm is used. In this framework, we propose to develop a maintenance model and optimize the
decision rule by taking into account an on-line change
detection algorithm.
The particularity of this model is the presence of
unknown deterioration parameters in the second mode.
These parameters are unknown but the set to which
they belong is known. We deal with these parameters
by using an adequate detection/isolation algorithm in
the framework of the conditional maintenance.
Throughout this paper we deal with a system which
nominally deteriorates according to a given known
mode and its mean deteriorating rate increases suddenly at an unknown time. The time of change of
the degradation mode is called the change time. An
adaptive maintenance policy based on on-line change
detection procedures is already proposed in (Fouladirad, Dieulle, and Grall 2006; Fouladirad, Dieulle,
and Grall 2007) when the parameters of the accelerated mode are known. In this paper we suppose that
the parameters of the accelerated mode are unknown.
They take values belonging to a known set. As theses
parameters are no longer known in advance (even if
they take values belonging to a known set), the methods proposed in (Fouladirad, Dieulle, and Grall 2006;

Fouladirad, Dieulle, and Grall 2007) are no-more

appropriate. Hence, in this paper, to take into account
the presence of unknown parameters in the accelerated
mode, condition-based maintenance policy embedded
with a detection/isolation procedure is proposed. The
on-line change detection/isolation algorithm should be
adequate in the sense that it should minimize the detection delay and have a low false alarm rate and a low
false isolation rate. The proposed parametric maintenance decision rule which uses an online detection
procedure of the change of mode is optimized in order
to minimize an average maintenance cost criterion.
The considered system is subject to continuous
random deterioration and a scalar aging variable can
summarize the system condition (Wang 2002). When
the system deteriorates the ageing variable increases
and a failure occurs when the ageing variable (the system state) exceeds a known fixed threshold L called
failure threshold. The failure threshold corresponds to
the limit value of wear beyond which the mission of the
system is no longer fulfilled. The information about
the system state is only collected through inspections,
it means that such systems are usually subject to nonobvious failures. Therefore, the system is known to
be failed only if the failure is revealed by inspections.
The system can be declared as failed as soon as a
defect or an important deterioration is present, even
if the system is still functioning. In this situation, its
high level of deterioration is unacceptable either for
economic reasons (poor products quality, high consumption of raw material, etc. . . ) or for safety reasons
(high risk of hazardous breakdowns). An example is
given in (van Noortwijk and van Gelder 1996), where
the problem of maintaining coastal flood barriers is

567

modeled. The sea gradually erodes the flood barrier

and the barrier is deemed to have failed when it is no
longer able to withstand the pressure of the sea.
If through inspections it is discovered that the system has failed, a corrective maintenance operation
immediately replaces the failed system by a new one.
Preventive maintenance actions are performed in order
to avoid a failure occurrence and the resulting period of
inactivity of the system. The preventive maintenance
operation is less costly than the corrective maintenance
operation. The preventive maintenance action takes
place when the system state exceeds a predetermined
threshold known as the preventive threshold.
The inter-inspection interval times and the value of
the preventive threshold are two factors which influence the global maintenance cost. For example, in
the case of costly inspections it is not worthwhile to
inspect often the system. But if the system is scarcely
inspected, the risk of missing a failure occurrence
increases. In (Grall, Dieulle, and Berenguer 2002)
or (Dieulle, Brenguer, Grall, and Roussignol 2003)
and (Brenguer, Grall, Dieulle, and Roussignol
2003) authors propose condition-based inspection/replacement and continuous monitoring replacement policies for a single-mode deteriorating system.
In those previous works, a maintenance cost model
is proposed which quantifies the costs of the maintenance strategy and propose a method to find the
optimal strategy leading to a balance between monitoring and maintenance efficiency. When the system
undergoes a change of mode it seems reasonable to
incorporate the on-line information available about the
system in the maintenance decision rule. In (Fouladirad, Dieulle, and Grall 2006; Fouladirad, Dieulle,
and Grall 2007) authors studied the on-line change
detection in the framework of the condition based
maintenance in the case of known parameters after
the change.
In this paper an adaptive maintenance policy based
on an embedded optimal on-line change detection
algorithm is proposed. The originality is due to the fact
that the parameters after the change (i.e. the parameters of the second mode) can take unknown values, but
these values belong to a known and finite set.
In section 2, the deteriorating system is described.
In section 3, an adequate maintenance decision rule is
proposed. Section 4 is devoted to the presentation of
the on-line change detection algorithm. A method for
the evaluation of the maintenance cost is proposed in
section 5. In section 6 numerical implementations of
our results are presented and analyzed.
2

SYSTEM DESCRIPTION

The system to study in this paper is an observable system subject to accumulation of damage. The system

state at time t can be summarized by a scalar random

ageing variable Xt . When no repair or replacement
action has taken place, (Xt )t0 is an increasing stochastic process, with initial state X0 = 0. If the state of
the process reaches a pre-determined threshold, say
L, the system is said to be failed. The system can be
declared as failed as soon as a defect or an important
deterioration is present, even if the system is still functioning. This means that it is no longer able to fulfill
its mission in acceptable conditions. The threshold L
is chosen in respect with the properties of the considered system. It can be seen as a safety level which has
not to be exceeded. The behavior of the deterioration
process after a time t depends only on the amount of
deterioration at this time.
The parameters of the deterioration process (Xt )t0
can suddenly change at time T0 . After T0 the mean
deterioration rate suddenly increases from a nominal
value to an accelerated rate at time T0 . The first mode
corresponds to a nominal mode denoted by M1 and the
accelerated mode is denoted by M2 . In this paper, the
case of two degradation modes is treated but this choice
is not restrictive. The results exposed in this paper
can be generalized to the case of multiple degradation
modes which can be subject to further works.
In this paper, it is assumed that the deterioration
process in mode Mi (i = 1, 2), denoted by (Xti )t0 , is
a gamma process i.e. for all 0 s t, the increment
of (Xti )t0 between s and t, Yts = Xti Xsi , follows a gamma probability distribution function with
shape-parameter i .(t s) and scale parameter i .
This probability distribution function can be written
as follows:
y

yi (ts)1 e
1
fi (ts),i (y) =

(i (t s))
ii (ts)

1{y0} .
(1)

The average deterioration speed rate in mode Mi is

i (t s).i and its variance is i (t s).i2 .
It should be recalled that gamma process is a positive process with independent increments. It implies
frequent occurrences of tiny increments which make
it relevant to describe gradual deterioration due to
continuous use such as erosion, corrosion, concrete creep, crack growth, wear of structural components (van Noortwijk 2007; Cooke, Mendel, and
Vrijling 1997; inlar, Baant, and Osman 1977;
Blain, Barros, Grall, and Lefebvre 2007; Frangopol, Kallen, and van Noortwijk 2004). Furthermore,
the gamma process is the existence of an explicit
probability distribution function which permits feasible mathematical developments. It has been widely
applied to model condition-based maintenance, see
(van Noortwijk 2007).

568

The originality of this paper is due to the existence

of unknown parameters after the change. In (Fouladirad, Dieulle, and Grall 2006; Fouladirad, Dieulle, and
Grall 2007) the second mode parameters are known
in advance but in this paper the second mode parameters are no longer known in advance. They Take their
values in the following set:
S = {(21 , 21 ), . . . (2K , 2K )}

(2)

As card(S) = K, there are K different possibilities

of second degradation mode.
It is supposed that the mode M2 corresponds to an
accelerated mode so in the second mode the parameters are such that 2 .2 > 1 .1 . The case of a slower
second mode is discarded because the systems we consider can not be stabilized nor deteriorate slower than
before.
3

MAINTENANCE POLICY

In the framework of this paper, the considered deteriorating systems cannot be continuously monitored.
Hence, the deterioration level can only be known
at inspection times. We shall denote by (tk )kN the
sequence of the inspection times defined by tk+1 tk =
t (for k N) where t is a fixed parameter. If the
deterioration level exceeds the threshold L between
two inspections, the system continues to deteriorate
until the next inspection. If the deterioration exceeds
the threshold L between two inspections, then the system continues to deteriorate until the next inspection
and can undergo a breakdown. The failure of such a
system being able to have disastrous consequences, a
preventive maintenance action has to take place before
and the inter-inspection time t has to be carefully
chosen in order to be able to replace the system before
the failure. The preventive replacement leads to restore
the system in a good as new state. The maintenance decision scheme provides possible preventive
replacements according to the measured amount of
deterioration.
There are two possible maintenance actions at each
inspection time tk :
the system is preventively replaced if the deterioration level is close to L,
if a failure has occurred (Xtk > L) then the system
is correctively replaced.
A detection method can be used to identify the
unknown time of change of degradation mode T0 .
The on-line change detection algorithm presented in
section 4 is used to detect the change mode time
T0 when the parameters after the change belong to
the set presented in (2). The maintenance decision is
based on a parametric decision rule according to the

change detection result. As in (Saassouh, Dieulle, and

Grall 2005),(Fouladirad, Dieulle, and Grall 2006) and
(Fouladirad, Dieulle, and Grall 2007), the preventive
maintenance decision is based on different preventive
thresholds corresponding to each of the possible deterioration modes of the system. Such maintenance policies are extensions of inspection/replacement structures for single mode deteriorating systems. Let Anom
and Aac be the decision thresholds associated to
the limit cases corresponding to the single-mode
deterioration system. The decision threshold Anom
(respectively Aac ) is chosen in order to minimize the
cost criterion of the nominal single-mode deteriorating system. In the nominal mode the threshold Anom
is effective and as soon as the system is supposed to
have switched in the second mode (accelerated mode)
then threshold is adapted from Anom to Aac .
The possible decisions which can arise at each
inspection time tk are as follows:
If Xtk L the system has failed then it is correctively
replaced.
If no change of mode has been detected in the current cycle and if Xtk Anom then the system is
preventively replaced.
If a change of mode is detected at time tk or has been
detected earlier in the current cycle and if Xtk Aac
then the system is preventively replaced.
In all other cases, the decision is reported to time
tk+1 .
As a consequence of the previous decision rule, if
the system is deteriorating according to the mode i,
i = 1, 2, and if a change of mode is detected at time
tdetect = kt, k N, where tdetect T0 , the two
following scenarios can arise:
If Xtdetect < Aac then the system is left unchanged
and a replacement is performed at time tn > tdetect
such that Xtn1 < Aac Xtn .
If Xtdetect Aac then the system is immediately
replaced.
The parameters of the maintenance decision rule are
respectively the thresholds Aac , Anom , the interinspection interval t and the parameters of the
detection algorithm.

ON-LINE CHANGE DETECTION/

ISOLATION

The on-line abrupt change detection, with low false

alarm rate and a short detection delay, has many important applications, including industrial quality control,
automated fault detection in controlled dynamical systems. Authors in (Basseville and Nikiforov 1993)
present a large literature on the detection algorithms

569

in complex systems. A first attempt to use an optimal on-line abrupt change detection in the framework
of maintenance policy is presented in (Fouladirad,
Dieulle, and Grall 2006), (Fouladirad, Dieulle, and
Grall 2007). The on-line change detection algorithms
permit to use online available information on the deterioration rate to detect the occurred abrupt change
time. These algorithms take into account the information collected through inspections, so they treat with
on-line discrete observations (i.e. system state at times
(tk )kN ). In (Fouladirad, Dieulle, and Grall 2007) is is
supposed that a prior information on the change time
is available. In this case an adequate on-line detection
algorithm which takes into account the available prior
information on the change time is proposed.
The authors in (Fouladirad, Dieulle, and Grall
2006), (Fouladirad, Dieulle, and Grall 2007) considered the case of two deteriorating modes (one
change time) and known parameters after the change.
In this paper, the aim is to propose an adequate
detection/isolation method when the accelerated mode
parameters can take unknown values. These values
belong to a known set defined in (2).
We collect observations (Xk )kN at inspection times
k N. Let be Yk for k N the increments of the
degradation process. Therefore Yk for k N follows a
gamma law with density fi = fi t,i according to the
degradation mode Mi , i = 1, 2. We shall denote fl =
f2l t,2l , l = 1, . . . , K, the density function associated
to the accelerated mode when (2 , 2 ) = (2l , 2l ).
We shall denote by N the alarm time at which
a -type change is detected/isolated and , =
1, . . . , K, is the final decision. A change detection/isolation algorithm should compute the couple
(N , ) based on Y1 , Y2 , . . . . We shall denote by Pr 0
the probability knowing that no change of mode has
occurred, Pr lT0 the probability knowing that the change
of mode has occurred at T0 . Under Pr lT0 the increments Y1 , Y2 , . . . , YT0 1 have each the density function
f1 and a change at T0 has occurred and YT0 is the
first observation with distribution fl , l = 1, . . . , K.E0
(resp. ElT0 ) is the expectation corresponding to the
probability Pr 0 (resp. Pr lT0 ).
The mean time before the first false alarm of a j
type is defined as follow:

= max l ,

E0 (N =j ) = E0 inf {N (k) : (k) = j}

(3)

k1

l = sup esssup
T0 1

= ElT0 (N T0 + 1|N T0 , X1 , . . . , XT0 1 ).

for a given minimum false alarm rate or false isolation
rate:

The mean time before the first false isolation of a j

type is defined as follow:


ElT0 (N =j ) = ElT0 inf {N (k) : (k) = j}

k1

min EiT0 (N =j ) = a

min

(4)

(6)

0iK 1j =iK

where E0T0 = E0 .
Let us recall the detection /isolation algorithm initially proposed by (Nikiforov 1995). We define the
stopping time N l in the following manner:
N l = inf N l (k),
k1

(7)

N l (k) = inf t k, min Skt (l, j) h

0j =lK

(8)

where Skt (l, j) is the the Likelihood ratio between fl and


(Yi )
fj : Skt (l, j) = ti=k log fflj (Y
, and f0 () = f1 ().
i)
The stopping time and the final decision of
the detection/isolation algorithm are presented as it
follows:
N = min{N 1 , . . . , N K }

= argmin{N , . . . , N
1

(9)
}

(10)

In (Nikiforov 1995) author proved that the mean

time to the detection in the worst case defined by
(5) satisfies the following relations:
max ElT0 (N )
= min


fl ln

ln(a)
as a

(11)

min lj

(12)

 
fl
d 0 j  = l K
fj

(13)

1lK 0j =lK

lj =

(5)

1lK

1lK

As it is initially proposed by (Lorden 1971), usually

in on-line detection algorithms the aim is to minimize the mean delay for detection/isolation in the
worst case:

where lj is the Kullback-Leibler distance. The

detection/isolation algorithm presented in this section
reaches asymptotically the lower asymptotic bound
ln(a)
initially proposed by (Lorden 1971).

570

EVALUATION OF THE MAINTENANCE

POLICY

Each time that a maintenance action is performed on

the system, a maintenance cost is incurred. Each corrective (respectively preventive) replacement entails a
cost Cc (respectively Cp ). Since a corrective maintenance operation is performed on a more deteriorated
system, it is generally more complex and consequently
more expensive than a preventive one. Hence it is
supposed that Cp < Cc . The cost incurred by any
inspection is Ci . In the period of unavailability of the
system (i.e the time spent by the system in a failed
state) an additional cost per unit of time Cu is incurred.
All direct and indirect costs are already included in
the unit costs Ci , Cc , Cp , Cu . The maintenance policy is evaluated using an average long run cost rate
taking into account the cost of each type of maintenance actions. Let us denote by Np (t) the number of
preventive replacements before t, Nc (t) the number
of corrective replacements before t, du (t) the cumulative unavailability duration of the system before t and
Ni (t) the number of inspections before t. We know that
Ni (t) = [t/t] where [x] denotes the integer part of
the real number x. Let us denote by T the length of a
life-time cycle and TL the random time at which the
system state exceeds threshold L. The property of the
regeneration process (Xt )t0 allows us to write:
C = lim

E(C(T ))
E(C(t))
=
t
E(T )

(14)

In this work, the cost criterion is optimized as a

function of the parameter of the considered maintenance policy: the detection threshold h defined in
(7,8).

NUMERICAL IMPLEMENTATION

In this section we apply the maintenance policy presented in this paper to the case of a system with
two degradation modes and four possible accelerated
modes (K = 4).
The proposed maintenance policies are analyzed by
numerical implementations. Throughout this section,
the values of the maintenance costs are respectively
Ci = 5, Cp = 50, Cc = 100 and Cu = 250. For
the numerical calculations it is supposed that in the
nominal mode M1 , 1 = 1 and 1 = 1. Hence,
the maintenance threshold Anom is equal to 90.2.
The previous value is the optimal value which minimizes the long run maintenance cost for a single mode
deteriorating system in mode M1 . For this optimization (from Monte Carlo simulations), we use a single
degradation mode results with t = 4. The couple
(Anom , t) = (90.2, 4) is the optimal couple which
minimizes the long run maintenance cost for a single
mode deteriorating system in mode M1 . T0 is simulated by a uniform law from Monte Carlo method. To
evaluate each maintenance policy, four different accelerated modes are considered. So the parameters of the
accelerated mode belong to the following set:
(2 , 2 ) {(2, 1), (1, 3), (2, 2), (1, 7)}

where

The properties of the different second modes are

presented table 1. The threshold Aac corresponds to
the optimal value which minimizes the long run maintenance cost for a single mode deteriorating system in
mode M2 .

C(t) = Ci Ni (t) + Cp Np (t) + Cc Nc (t) + Cu du (t).

We know that
E(Np (T )) = Pr(cycle ends by a preventive repl.),

6.1

E(Nc (T )) = Pr(cycle ends by a corrective repl.)

 
T
,
E(Ni (T )) = E
t

Parameter optimization

In this section, the problem of parameter optimization

for the considered maintenance policies is investigated. The parameter of interest is the detection
threshold h defined in (7) and (8).

E(du (T )) = E(T TL ){1{TL <T },

Let us set the inspection scheduling function
introduced in (Grall, Dieulle, and Berenguer 2002)
be constant then the threshold Anom and the interinspection time t can be obtained by numerical
minimization of the cost criterion of the single-mode
deteriorating system in mode M1 . The threshold Aac
corresponds to the optimal threshold for the singlemode deteriorating system in mode M2 .

Table 1.
mode.

Characteristic data of the second degradation

Real second mode cases

2
2
Aac

2
1
85.6

1
3
74.6

2
2
73.7

1
7
51.6

571

The optimal value of h which leads to a minimal

maintenance cost is numerically calculated. To define
the optimal value of h, the maintenance cost, the false
alarm rate and the isolation rate are obtained for different values of h in the interval [0, 15]. This interval
is chosen because the cost remains stable around the
same values after h = 15. The values of h corresponding to the lowest maintenance cost, lowest false alarm
rate and highest correct isolation are defined. To study
the impact of the variation of the threshold h on the
properties of the maintenance policy, in addition to
the maintenance cost, the probability of preventive and
corrective maintenance for different values of h in the
interval [0, 15] are calculated. The choice of h is not
always based on the value which minimizes the maintenance cost or which can optimizes the properties of
the detection algorithm. For example, it is not sensible
to take a value of h leading to the lowest maintenance
cost if it corresponds to a false isolation rate close to 1.
We present in table 2 the properties of the maintenance versus detection/isolation algorithm corresponding to the value of h which leads to the lowest
maintenance cost. It can be noted that except the case
1 (2 = 2 and 2 = 1) the maintenance costs are very
low in comparison with the conservative case when
only one threshold is used presented in table 5.
Table 2. Optimal maintenance policy corresponding to the
lowest maintenance cost.
Real second mode cases

Maintenance Cost
Detection threshold
False alarm rate
Correct isolation rate

1.98
1
0.9
0.87

1.99
1
0.89
0.03

1.99
1
0.89
0.04

2.00
1
0.88
0.05

Table 3. Optimal maintenance policy corresponding to the

lowest false alarm rate.
Real second mode cases

Maintenance Cost
Detection threshold
False alarm rate
Correct isolation rate

1.99
5
0.016
1

2.22
7
0.02
0

2.37
6
0.014
0

2.67
15
0.24
0

Table 4. Optimal maintenance policy corresponding to the

highest correct isolation rate.
Real second mode cases

Maintenance Cost
Detection threshold
False alarm rate
Correct isolation rate

1.98
12
0.018
1

2.09
2
0.3
0.19

2.17
2
0.31
0.2

2.34
2
0.27
0.3

Table 5. Optimal costs corresponding to the maintenance

with one preventive threshold.
Case

Costs

1.97

2.21

2.36

2.66

In table 3 the properties of the maintenance versus detection/isolation algorithm corresponding to the
value of h which leads to the lowest false alarm rate
are exposed. It can be noticed that maintenance costs
are very close to costs when only one threshold is used
(without detection procedure). The use of the detection
algorithms when a low false alarm is requested doesnt
improve the quality of the maintenance policy. In this
configuration, a maintenance policy without detection
procedure seems to be adequate. The results in table
4corresponds to the properties of the maintenance versus detection/isolation algorithm corresponding to the
value of h which leads to the highest correct isolation.
In this table except the case 1 (2 = 2 and 2 = 1)
the highest correct isolation is not very high but the
corresponding false alarm rate and maintenance costs
are acceptable. The maintenance cost is still lower than
the maintenance policy without detection procedure.
In the first case (2 = 2 and 2 = 1) the correct isolation rate is always very high. This should be due to the
global optimization of the detection threshold h. This
optimization is more sensitive to the properties of the
first case where the two modes are very close. It is
possible that if in the optimization procedure, for each
second mode l = 1, . . . , K, a detection threshold hl in
equation (7) is used, the result of the correct isolation
could be different. But this method requests a complex
optimization procedure and the feasibility is arguable.
If the only criteria is the result of the maintenance
(low maintenance cost) we can neglect the value of
false alarm rate and false isolation. But if in the
maintenance procedure the properties of the detection algorithms are of great importance we can not
base our choice only on the maintenance cost and we
should take into account the properties of the detection
algorithm.
In figure 1 the maintenance properties corresponding to the accelerated mode (2 = 1, 2 = 3) are
depicted. To illustrate the results the threshold h varies
in [0, 15]. The maintenance cost is stable around 2.2
and reaches its minimum value 1.99 for h = 0. The
probability of corrective maintenance is very low and
the probability of preventive maintenance is very high.
We can say that there is mostly a preventive policy.
In figure 2 the detection algorithm properties corresponding to the accelerated mode (2 = 2, 2 = 2)
are depicted. To illustrate the results the threshold h
varies in [0, 15]. The false alarm rate is very high for
the small values of h and it decreases as h grows. For

572

h > 4, the false alarm rate is close to 0. In the contrary

as h grows the detection delay also grows. It is sensible that for a low detection threshold the false alarm
is very significant and as the detection improves and
the false alarm decreases the detection delay appears.
It is natural that with 0.9 false alarm rate the detection
delay is inexistent.
In figure 3 the correct isolation rate corresponding to the accelerated mode (2 = 2, 2 = 1) is
depicted. To illustrate the results the threshold h varies
in [0, 15]. These results show the good quality of the
detection/isolation algorithm.

Cost versus h
3.0
2.8
2.6
2.4
2.2
2.0
1.8
1.6
1.4
1.2
1.0

10

15

Proba of corrective maintenance vs h

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

10

15

Proba of preventive maintenance vs h

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

7
0

Figure 1.

10

In this paper we have proposed a maintenance policy combined with an on-line detection method. This
policy leads to a low maintenance cost. We took into
account the possibility that the system could switch
on to different accelerated modes. By considering this
possibility the proposed detection algorithm is also
an isolation algorithm. We have noticed that the lowest false alarm rate and false isolation rate does not
always necessarily corresponds to the lowest maintenance cost. The proposed algorithm has generally low
correct isolation rate and some times high false alarm
rate. The aim in the future works is to improve these
properties in order to obtain a low cost maintenance
versus detection policy which can easily isolate the
real accelerated mode.

Maintenance properties when 2 = 1, 2 = 3.

False alarm rate vs h

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

10

CONCLUSION

15

15

mean delay vs h
20
18
16
14
12
10
8

REFERENCES

6
4
2
0

Figure 2.
2 = 2.

10

15

Detection algorithm properties when 2 = 2 and

correct isolation rate vs h

1.5

1.0

0.5

0.0
0

Figure 3.

10

15

Correct isolation rate when 2 = 2 and 2 = 1.

Basseville, M. and I. Nikiforov (1993). Detection of

Abrupt Changes: Theory and Application. Prentice. Hall,
www.irisa.fr/sigma2/kniga.
Brenguer, C., A. Grall, L. Dieulle, and M. Roussignol
(2003). Maintenance policy for a continuously monitored
deteriorating system. Probability in the Engineering and
Informational Science 17(2), 235250.
Blain, C., A. Barros, A. Grall, and Y. Lefebvre (2007, Jun.).
Modelling of stress corrosion cracking with stochastic
processes - application to steam generators. In T. Avenand
J. Vinnem (Eds.), Risk, Reliability and Societal Safety,
Proceedings of the European Safety and Reliability Conference 2007 (ES-REL 2007), Stavanger, Norway, 2527
June 2007, Volume 3, Stavanger, Norway, pp. 23952400.
Taylor and Francis Group, London.
inlar, E., Z. Baant, and E. Osman (1977). Stochastic
process for extrapolating concrete creep. Journal of the
Engineering Mechanics Division 103(EM6), 10691088.
Cooke, R., M. Mendel, and H. Vrijling (Eds.) (1997). Engineering Probabilistic Design and Maintenance for Flood
Protection. Kluwer Academic Publishers.
Dieulle, L., C. Brenguer, A. Grall, and M. Roussignol
(2003). Sequential condition-based maintenance scheduling for a deteriorating system. EJOR 150(2), 451461.

573

Fouladirad, M., L. Dieulle, and A. Grall (2006). The use

of online change detection for maintenance gradually
deteriorating system. In ES-REL 2006 Congress, Estoril,
Portugal, 1822 september, 2006.
Fouladirad, M., L. Dieulle, and A. Grall (2007). On the
use of bayesian on-line change detection for maintenance of gradually deteriorating systems. In ESREL 2007
Congress, Stavanger, Norway 2527 June.
Frangopol, D., M. Kallen, and J. van Noortwijk (2004).
Probabilistic models for life-cy cle performance of
deteriorating structures: review and future directions.
Progress in Structural Engineering and Materials 6(4),
197212.
Grall, A., L. Dieulle, and C. Berenguer (2002). Continuoustime predective maintenance scheduling for a detetiorating system. IEEE transactions on reliability 51(2),
141150.

Lorden, G. (1971). Procedure for reacting to a change in

distribution. The Annals of Matematical Statistics 42,
18971908.
Nikiforov, I. (1995). A generalized change detection problem.
IEEE Transactions on Information Theory 41, 171187.
Saassouh, B., L. Dieulle, and A. Grall (2005, 30 June). Adaptive maintenance policy for a deteriorating system with
random change of mode. pp. 5. ESREL 2005. Avec acte.
van Noortwijk, J. (2007). A survey of the application of
gamma processes in maintenance. In Reliability Engineering and System Safety.
van Noortwijk, J. M. and P. van Gelder (1996). Optimal
Maintenance Decisions for Berm Breakwaters. Structural
Safety 18(4), 293309.
Wang, H. (2002). A survey of maintenance policies of
deteriorating systems. European Journal of Operational
research 12, 469489.

574

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Opportunity-based age replacement for a system under two types

of failures
F.G. Bada & M.D. Berrade
Department of Statistics, Centro Politcnico Superior, University of Zaragoza, Zaragoza, Spain

ABSTRACT: This paper is concerned with an opportunity-based age replacement policy for a system under two
types of failures, minor and catastrophic. We consider a general distribution for the time to the first opportunity,
dropping the usual assumption of exponentially distributed times between opportunities. Under this model
the system undergoes a minimal repair whenever a minor failure occurs whereas a perfect restoration follows
any catastrophic failure and after the N minor failure. The system is preventively replaced at maintenance
opportunities arising after instant S and also at the moment its age reaches T . We take into account the costs
due to minimal repairs, perfect repairs, opportunity based replacements and preventive maintenances. We focus
on the optimum policy (T , N ) that minimizes the long-term cost per unit of time, providing conditions under
which such optimum policy exists.

INTRODUCTION

Preventing systems from failing constitutes a significant task for manufacturers because of the costs and
dangers incurred due to failures. Such importance
has make the design of maintenance policies to be an
essential research area in reliability theory. The preventive activities carried out to reduce the breakdown
risk are combined in practice with corrective actions
that restore the failed systems to the operating condition as soon as possible. Depending on the quality
of the corrective actions, the repair can be classified
into perfect or imperfect. The former brings the system
back to an as-good-as-new condition whereas the latter
restores the system somewhere between as-good-asnew and as-bad-as-old. The as-bad-as-old restoration
is known as minimal repair.
Many maintenance policies consider that preventive replacements can be carried out at any moment.
However the work interruption time is likely to produce high costs. If so, taking advantage of inactivity
periods is recommended, delaying the maintenance
until the moment when the mechanism is not required
for service. In general such opportunities are a type
of event occurring at random, e.g., when another
unit in the same system undergoes a failure. Repairs
and replacements carried out outside these so-called
maintenance opportunities can cause cost ineffective maintenances and may be even responsible for
high economic loses.
The following references are based on the principle
that an event type can be considered as an opportunity.

(Dekker and Smeitink 1991) consider a block replacement model in which a component can be replaced
preventively at maintenance opportunities only. Maintenance opportunities occur randomly and are modeled
through a renewal process. (Dekker and Dijkstra 1992)
establish conditions for the existence of a unique
average control limit policy when the opportunities
arise according to a Poisson process. (Jhang and Sheu
1999) propose an opportunity-based age replacement
with minimal repair for a system under two types of
failure. (Iskandar and Sandoh 2000) present a maintenance model taking advantage or not of opportunities
in a random way. (Satow and Osaki 2003) consider
that intensity rate of opportunities change at specific age. (Coolen-Schrijner et al., 2006) describe an
opportunity-based age replacement by using nonparametric predictive inference for the time to failure of a
future unit. They provide an alternative to the classical
approach where the probability distribution of a units
time to failure is assumed to be known. (Dohi et al.,
2007) analyze this type of replacement policies which
result to be less expensive than those carried out at any
time without taking opportunities into account.
In this work we present an opportunity-based maintenance model for a system that may undergo two types
of failures, minor and catastrophic. Minor failures are
followed by a minimal repair to bring the system back
into use whereas catastrophic failures are removed
with a perfect repair. A perfect repair is also carried out
after the N th minor failure. From time S on, opportunity maintenances occur which are independent from
the system time to failure. In addition the maintenance

575

policy is completed with a preventive replacement at

age T . We assume costs derived from minimal and
perfect repairs as well as both opportunity-based maintenance and preventive maintenance costs, aiming at
minimizing the expected cost per unit of time. This
work is concerned with the conditions under which
there exists an optimum policy (T , N ). To this end the
model is described in Section 2 where the cost function is also derived. Sections 3 and 4 contain the results
related to the existence of the optimal policy.
2

THE MODEL

Consider a system that may experience two types of

failures. Provided that a failure occurs at time x, then
it belongs to the minor failures class with probability p(x) or it is a catastrophic failure with probability
q(x) = 1 p(x). The former are removed by means
of a minimal repair whereas the system is restored
with a perfect repair after catastrophic failures. In
addition, the system is preventively maintained at
age T .
Let r(x) be the failure rate of the time to the first
failure. Hence, the corresponding reliability function
is given as
F(x) = e

x
0

x
0

1. the renewal that follows a catastrophic failure or the

N th minor failure, whichever comes first, occurring
before S.
2. the renewal that follows a catastrophic failure or the
N th minor failure, whichever comes first, occurring
after S.
3. a maintenance opportunity.
4. a preventive maintenance.
Let p1 denote the probability of a cycle ending
after the corrective maintenance that follows events 1)
and 2), whereas p2 and p3 represent the corresponding probabilities of events 3) and 4). The foregoing
probabilities are calculated next.


p1 =
0

= H (S, N ) +

N
1


Dk (x)e

x
0

the previous integrals represent the probability of a

corrective maintenance in [0, S] and [S, T ] respectively.
H (y + S, N )dG(y)

H (x, N )dG(x S)

Integrating by parts with u = H (x, N ) and dv =

dG(x S) we arrive at

k
p(u)r(u)du
,
k!

N
2


T S

p2 =

p2 = H (S, N ) H (T , N )G(T S)
k = 0, 1, . . .

+ r(x)DN 1 (x)e

G(x S)dH (x, N )

q(x)r(x)Dk (x)e

k=0

and its density function

h(x, N ) =

G(x S)dH (x, N )

r(u)du

where
 x

k=0

Dk (x) =

G(x S)dH (x, N )

The distributions corresponding to the first catastrophic failure and the N th minor failure are independent. The reliability function corresponding to that
occurring first is given by

T
S

q(u)r(u)du

H (x, N ) =


dH (x, N ) +

r(u)du

A failure happening at instant x belongs to catastrophic class with probability q(x) = 1 p(x). The
reliability function of the time to the first catastrophic
failure is
e

Renewal opportunities independent from the system arise from time S on with 0 < S < T . Let G(x)
denote the reliability function of the time elapsed from
S to the first maintenance opportunity. We assume
that whenever the system is renewed, G(x) remains
the same.
A cycle, that is, the total renewal of the system is
completed after one of the four events described next

x
0

x
0

r(u)du

r(u)du

p3 = H (T , N )G(T S)
The following formula provides the mean length of
a cycle

576

E( ) =

xdH (x, N )

+


T S

(y + S)H (y + S, N )dG(y)

where the four foregoing terms correspond respectively to the mean length of a cycle ending after events
previously numbered as 1) to 4). The third term in the
formula can be rewritten as

(N 1)G(x S)m(x)p(x)r(x)

DN 1 (x)s(x)dx



H (x, N )G(x S)dx

xG(x S)dH (x, N )

y+S

y+S



H (x, N )G(x S)dx

H (x, N )dx +

m(x) = e

T S

H (x, N )G(x S)dx

x
0


dx




h(x, N 1)dx

y+S


p(u)r(u)du dG(y)

x

r(u)du

y+S

Let us denote MR1 , MR2 , MR3 and MR4 the mean

number of minimal repairs in cycle ending after one
of the four events described above.
In what follows we denote by s(x) and m(x) the
following functions:
s(x) = e

p(u)r(u)du dG(y)



x

y+S

p(x)r(x)DN 2(x)e

xdH (x, N ) + SH (S, N )

T


p(u)r(u)du dG(y)

T S 

Hence,


H (x, N 1)q(x)r(x)

y+S




p(u)r(u)du h(x, N 1)dx

0
T S

MR3 =

G(x S)

+ SH (S, N )




kDk (x)s(x)q(x)r(x)m(x)dx

k=0

xH (x, N )dG(x S) = T H (T , N )G(T S)

N
1


xH (x, N )dG(x S)

E( ) =


p(u)r(u)du h(x, N 1)dx

G(x S)

Integrating by parts with u = xH (x, N ) and dv =

dG(x S) leads to
T

x
0

MR2 =

(N 1)m(x)p(x)r(x)DN 1 (x)s(x)dx



+ T H (T , N )G(T S)

kDk (x)s(x)q(x)r(x)m(x)dx

k=0

xG(x S)dH (x, N )

1
SN


MR1 =


=

T S

H (y + s, N 1)



y+S


p(u)r(u)du dG(y)

0
p(u)r(u)du


=

q(u)r(u)du

577


H (x, N 1)
0


p(u)r(u)du dG(x S)

Integratingby parts with dv = dG(x S) and u =

x
H (x, N 1) 0 p(u)r(u)du , we get



p(u)r(u)du H (S, N 1)

MR3 =

opportunity arises with cost c2 and also when it is

time for preventive maintenance, being c3 the corresponding cost. Minimal repairs are worth c4 each.
Hence the mean cost of a cycle is expressed by the
following formula




p(u)r(u)du H (T , N 1)G(T S)

E[C( )] = c1 p1 + c2 p2 + c3 p3 + c4 MR
= c1 H (S, N ) + c2 H (S, N )
 T
+ (c1 c2 )
G(x S)dH (x, N )

p(x)r(x)H (x, N 1)G(x S)dx



+ (c3 c2 )H (T , N )G(T S)
 S
+ c4
p(x)r(x)H (x, N 1)dx

p(u)r(u)du
S

G(x S)h(x, N 1)dx


MR4 = G(T S)


p(u)r(u)du

2
N


Dk (x)q(x)r(x)e

The key theorem of the renewal reward processes

ensures that the expected cost per unit of time can be
expressed as the ratio of the expected cost of a cycle
to its expected length, that is

x
0

r(u)du

dx

k=0



+ G(T S)

Q(T , N ) =

p(u)r(u)du
0

p(x)r(x)DN 2 (x)e


x
0

r(u)du

dx

= G(T S)

h(x, N 1)dx
T

= G(T S)

OPTIMUM T WHEN N IS SET

M (T , N ) = R(T , N )L(T , N ) C(T , N ) = 0



E[C( )]
E[ ]

The value of T minimizing the cost function is

obtained equating to zero the corresponding deriva,N )
= 0 which is equivalent to the
tive that is, dQ(T
dT
equation below

p(u)r(u)du
0

p(x)r(x)G(x S)H (x, N 1)dx

+ c4


p(u)r(u)du H (T , N 1)

where

L(T , N ) = E[ ]

Therefore the mean number of minimal repairs in a

cycle is given by

R(T , N ) = c4 r(T ) + (c2 c3 )l(T S)

MR = MR1 + MR2 + MR3 + MR4

 S
=
p(x)r(x)H (x, N 1)dx

+ (c1 c3 c4 )(q(T ) + p(T )Z(T ,N ))r(T )

with l(x) being the failure rate corresponding to the
time to the first maintenance opportunity distribution,
G, and

C(T , N ) = E[C( )]

p(x)r(x)G(x S)H (x, N 1)dx

Upon catastrophic failures the system is perfectly

restored at cost c1 and so is when a maintenance

DN 1 (T )
Z(T , N ) = N 1
k=0 Dk (T )

578

Z(T , N ) verifies that 0 Z(T , N ) 1 and is increasing with T provided that its corresponding derivative is

N
2  T


k=0

dZ(T , N )
DN 2 (T )
= p(T )r(T )
+ p(T )r(T )
dT
Dk (T ))2


N 1 (N 2+j)!
1
1
j=1 (j1)!(N 2)! DN 2+j (T ) j N 1

1
Dk (T ))2
( Nk=0

(c4 p(x)r(x)Dk (x)b(x)dx

(c1 c3 )
0

DN 1 (x)(r(x)

+ l(x S))b(x)dx (c1 c3 )

N
1


B(T , k)e(T , N )

k=0

Given that

where b(x), B(T , k) and d(T , k) are given by

d(q(T ) + p(T )Z(T , N ))

dT
dZ(T , N )
dq(T )
(1 Z(T , N )) + p(T )
=
dT
dT

b(x) = e

x
0

r(u)+l(uS)du

B(T , k) =


whenever q(T ) is an increasing function with T so is

q(T ) + p(T )Z(T , N ). Assuming that r(t) is increasing and l(x) is decreasing, along with c2 < c3 and
c1 c3 c4 > 0 then R(T , N ) is also increasing. Moreover the derivative of Q(T , N ) keeps the same sign of
M (T , N ) which exhibits the same monotonic behavior than R(T , N ). Therefore as M (0, N ) = c3 < 0,
if lim M (T , N ) > 0 and under the foregoing
assumptions, then Q(T , N ) has a finite minimum TN .
Such optimum policy is the unique zero of the equation
M (T , N ) = 0. In addition the corresponding optimum
cost is Q(TN , N ) = R(TN , N ).

Dk (x)b(x)dx,

k = 0, 1, . . .

0
T

d(T , k) =

l(x S)Dk (x)b(x)dx,

k = 0, 1, . . .

Dk (T )F(T )G(T S)
,
B(T , k)

e(T , k) =

k = 0, 1, . . .

By means of the next identity


0

p(x)r(x)DN 1 (x)b(x)dx

= DN (T )F(T )G(T S)
 T
+
DN (x)(r(x) + l(x S))b(x)dx
0

OPTIMUM N WHEN T IS SET

When the time of the preventive maintenance T

is fixed in advance, the optimum N should verify
both Q(T , N ) Q(T , N + 1) and Q(T , N 1) >
Q(T , N ). The foregoing conditions are equivalent to
W (T , N ) c3 and W (T , N 1) < c3 respectively,
being W (T , N ) defined as follows
W (T , N ) = c4

+ (c1 c3 c4 )(e(T , N ) e(T , N + 1))



d(T , N + 1)
d(T , N )

(1)
+ (c1 c2 c4 )
B(T , N ) B(T , N + 1)

B(T , k)
B(T , N )

p(x)r(x)DN 1 (x)b(x)dx + (c1 c2 )

N 1


d(T , k) d(T , N )

k=0
N
2  T

k=0

k=0

k=0

W (T , N + 1) W (T , N )

 N

B(T , k) (c4 (F(T , N + 1) F(T , N ))
=

N 1

which is obtained integrating by parts in the previous

expression, we get

k=0 B(T , k)
B(T , N )

(c1 c3 )q(x)r(x)Dk (x)b(x)dx

N
2  T

k=0

N 1

(c1 c3 )l(x S)Dk (x)b(x)dx

where
T
F(T , k) =

r(x)Dk (x)b(x)dx
,
B(T , k)

k = 0, 1, . . .

Following similar strategies to those provided in

(Nakagawa 2005) (chapter 4, Theorems 4.3 and 4.4)
we get that F(T , N ) is increasing with N if r(x) is
also increasing whereas if l(x) is decreasing with N
d(T ,N )
so is B(T
,N ) . Moreover it can be easily proved that

579

e(T , N ) increases with N . Then, whenever r(x) and

l(x) are increasing and decreasing respectively along
with c1 c3 c4 < 0 and c1 c2 c4 > 0, it
follows from (1) that W (T , N ) is increasing. Hence,
under the foregoing assumptions the optimum N when
T is previously set, NT , is the minimum N satisfying
W (T , N ) c3 . In case that W (T , N ) < c3 for all N
then NT = .
5

CONCLUSIONS

The high cost incurred due to some preventive maintenances motivate carrying out opportunity-based policies. This paper provides conditions under which
an optimum opportunity-based policy exists in two
cases, the optimum T for a given N and the optimum N when N is fixed. Such conditions involve
an increasing failure rate of the time to failure and
a decreasing failure rate of the time to the first
opportunity apart from cost-related conditions. Concerning the simultaneous optimization of both T and
N we consider the use of the following algorithm
proposed by (Zequeira and Berenguer 2006) and
(Nakagawa1986):
1. Set N = 1
2. If Q(TN +1 , N + 1)) < Q(TN , N ), then go to step 3
or to step 4 otherwise
3. N = N + 1
4. Set N  = N
The optimal policy turns out to be (T , N ). Note
that the foregoing algorithm doesnt ensure a global
optimum but just a local one. Moreover obtaining

conditions that guarantee the existence of an optimum

(T , N ) seems to be a difficult task.

REFERENCES
Coolen-Schrijner, P., F. Coolen, and S. Shaw (2006). Nonparametric adaptive opportunitybased age replacement
strategies. Journal of the Operational Research Society.
Dekker, R. and M. Dijkstra (1992). Opportunitybased age
replacement: exponentially distributed times between
opportunities. Naval Research Logistics (39), 175190.
Dekker, R. and E. Smeitink (1991). Opportunitybased block
replacement. European Journal of Operational Research
(53), 4663.
Dohi, T., N. Kaio, and S. Osaki (2007). Discrete time
opportunistic replacement policies and their application.
Recent advances in stochastic operations research. World
Scientific.
Iskandar, B. and H. Sandoh (2000). An extended opportunitybased age replacement policy. RAIRO Operations
Research (34), 145154.
Jhang, J. and S. Sheu (1999). Opportunity-based age replacement policy with minimal repair. Reliability Engineering
and System Safety (64), 339344.
Nakagawa, T. (1986). Periodic and sequential preventive
maintenance policies. Journal of Applied Probability (23),
536542.
Nakagawa, T. (2005). Maintenance Theory of Reliability.
Springer.
Satow, T. and S. Osaki (2003). Opportunity-based age
replacement with different intensity rates. Mathematical
and Computer Modelling (38), 14191426.
Zequeira, R. and C. Berenguer (2006). Optimal scheduling
of non-perfect inspections. IMA Journal of Management
Mathematics (2), 187207.

580

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Optimal inspection intervals for maintainable equipment

O. Hryniewicz
Systems Research Institute, Warsaw, Poland

ABSTRACT: Maintainable technical systems are considered whose failures can be revealed only by special
inspections. It is assumed that these inspections may be imperfect, i.e. they result may suggest wrong decisions.
Two optimization models are considered: one in which the coefficient of availability is maximized and second in
which related costs are minimized. For both models approximately optimal solutions have been found. Presented
examples show that these solutions are very close to the exact solutions when the time to failure is exponentially
distributed. The paper is illustrated with two numerical examples.

INTRODUCTION

For some technical systems their reliability state (operation or failure) cannot be directly observable. For
example, when a failure of a system is not of catastrophic nature it can be recognized either by doing
some measurements or by observing its performance.
Technical equipment of a production line serves as
a good example in such a case, when failures result
rather in worse performance (measured by observed
percentage of produced nonconforming items) than
in complete stoppage of a production process. Oneshot systems such as back-up batteries are examples
of other systems of that kind. Their reliability state is
not directly observable until the moment when they
are needed. In all such cases there is a need to perform
inspections in order to reveal the actual reliability state
of a considered system. Inspections policies may vary
from very simple, when times between consecutive
inspections (inspection intervals) are constant, to very
complicated, computed with the usage of additional
information about the system.
The problem of finding optimal inspection intervals has attracted many researchers since the works of
Savage (1956), Barlow et al. (1960) and some other
authors published in late 1950s and early 1960s. The
first general model was presented by Barlow, Hunter
and Proschan (1963) who assumed perfect inspections.
Later generalizations of this model were proposed, i.a.
by Menipaz (1978) and von Collani (1981). Among
relatively recent papers devoted to this problem, and
presenting simple and applicable in practice models,
we may list the papers by Baker (1990), Chung (1993),
Vaurio (1994), and Hariga (1996). More complex
inspection models are presented in recent papers by
Fung and Makis (1997), Chelbi and Ait-Kaidi (1999),
Wang and Christer (2003), and Berger et al. (2007).

There also exist papers, such as papers by Badia et al.

(2001, 2002), Vaurio (1999), and Khan et al. (2008)
presenting very complex models which e.g. include
the possibility of preventive maintenance actions, but
in our opinions these models are too complicated for
using them by the majority of practitioners.
In the paper we consider the problem of the optimal choice of inspection intervals for maintainable
(renewable) equipment. We assume that the state of
a system is evaluated periodically, and when its failure
is revealed the system is renewed in the sense of reliability theory. We consider two types of optimization
tasks.
1. Minimization of costs incurred by inspections and
failures of the system; applicable for the case of
e.g. production equipment.
2. Maximization of availability, applicable e.g. for
one-shot equipment.
The structure of the paper is the following. In
the second section we present mathematical models
describing periodically inspected maintainable equipment. Optimization tasks are formulated and solved
in the third section. Simple practical applications are
presented in the last section of the paper.
2

MATHEMATICAL MODEL

Consider renewable equipment whose state can be

revealed only as a result of special inspection. The
considered equipment can be in two exclusive reliability states: the up state, denoted as state 1, and the
down state, denoted as state 2. Let us assume that
a random time to failure T of this equipment, i.e. the
time of transition from state 1 to state 2, is distributed
according to a known cumulative distribution function

581

F(t). The actual reliability state of this equipment is

revealed by periodical inspection, performed every h
time units. When the failure is revealed the system is
renewed or replaced by a new one.
Let us assume that inspections may not be perfect.
In such a case there exists the probability of a false
alarm , which is the probability of revealing nonexistent failure of the inspected equipment, and the
probability of not revealing the actual failure. Such
probabilities are always greater than zero when the
actual reliability state of the considered equipment is
evaluated as the result of a statistical analysis of results
of measurements.
It is easy to show that the expected number of
of inspections performed during the time when the
system is up can be expressed as follows

A1 (h) =

i [P (T < (i + 1) h) P (T < ih)]

i=0

(1)
R (ih) ,

i=1

where R(ih) = 1 F(ih). Thus, the expected number

of false alarms is equal to
EI (h) = A1 (h) .

(2)

The expected number of inspections during the time

when the system is down entirely depends upon the
probability , and is equal to
EII (h) = A2 =

1
1

(3)

There are numerous examples of such systems. For

instance, the state of certain production process which
produces items whose quality can be evaluated by
destructive tests may be revealed by periodical sampling of its output. Another simple example of such
a system is a battery which back-ups power supply
of other equipments. Such a system is a typical oneshot system whose state is usually monitored during
periodical inspections.
In our model we assume that all actions last some
time during which the system is out of service. Let us
assume that the expected time of inspection is equal to
0 , and does not depend upon the actual state of the
inspected system. When the inspection reveals a false
alarm the expected additional out of service time is
equal to a , and when it reveals an actual failure the
expected repair (renewal) time is equal to r . Therefore the expected time between consecutive renewals

is given by
Tr = [A1 (h) + A2 ] h + A1 (h) (0 + a )
+ A2 0 + r

(4)

Hence, the stationary availability coefficient for

this equipment is given by
K (h) =

,
Tr (h)

(5)

where = E(T ) is the expected time to failure.

From the analysis of (4) and (5) it is obvious then
in case of h tending 0 the expected number of inspection, and thus the expected time devoted to inspections
tends to infinity. Hence, the availability in such a case
tends to zero. Similarly, when h tends to infinity the
expected duration of the state II is infinite, and the
availability also tends to zero. Hence, there exists an
optimal inspection interval hk , for which the availability K(h) attains its maximum. We will consider
the problem of finding this optimal value in the next
section of this paper.
The optimization of the inspection interval h can be
also considered as the problem of the minimizations
of some costs. Let us assume that the average cost of
each inspection is equal to c0 . Moreover, we assume
that the average additional cost of a false alarm is equal
to cr , and the average cost of systems renewal is equal
to cr . We can also distinguish the average cost related
to a failure cf , but this cost as unavoidable can be also
included in the cost of renewal cr which has the same
nature. There exist also losses which depend upon the
duration of state 2. We assume that this cost is proportional to the duration of state 2, i.e. the time between
the failure and the moment of its detection, and this
cost calculated per time unit is equal to cl .
Now we can calculate the expected costs of inspections C1 , the expected costs related to false alarms C2 ,
and the expected costs related to the failure C3 . They
can be calculated from the following expressions:
C1 (h) = [A1 (h) + A2 ] c0

(6)

C2 (h) = A1 (h) ca

(7)

C3 (h) = cf + cr + {[A1 (h) + A2 ] h } cl

(8)

The total expected cost related to one renewal period

is now given by
C (h) = C1 (h) + C2 (h) + C3 (h)

(9)

Similarly to the previous case its easy to show that

for h tending to zero and for h tending to infinity the

582

expected cost C(h) is infinite. Thus, there exists an

optimal inspection interval hc , for which the expected
cost C(h) attains its minimum. We will consider the
problem of finding this optimal value in the next
section of this paper.

OPTIMAL INSPECTION INTERVALS

Optimal values of the inspection interval h can be

found using a standard optimization procedure. First,
we find the derivative with respect to h of the goal
function K(h) or C(h), respectively. Then we have to
equate these derivatives to zero arriving at equations
whose solutions are equal to the respective optimal
values of h. In the case of the maximization of K(h)
given by (5) this equation is expressed as follows
hA1 (h) + [A1 (h) + A2 ] + A1 (h) (0 + 1 ) = 0
(10)

the objective function is in this case expressed as

K (h)
= 

1 + A2

eh



eh 1


h
1 h + e 1 (r + A2 0 ) + 0 + a
(14)


In order to find the optimal value of h we have to

find the derivative of (14) with respect to h, and then
to equate this derivative to zero. As the result of these
calculations we arrive at the following equation:
A2 e2h + heh + eh [0 + a + (2A2 1)]
(A2 1) = 0

This nonlinear equation can be solved numerically

using, e.g., the Raphson-Newton method. Application
of this method requires the computation of the derivative of the left-hand side of (15). This derivative is
expressed as
D = 2A2 e2h + heh

where
A1 (h) =

+ eh [1 + (0 + a ) + (2A2 1)] .
if (ih) ,

(11)

i=1

(15)

(16)

When our objective is to minimize the cost given

by (9) in case of the exponentially distributed lifetime
the goal function is expressed as follows:

and f (.) is the probability density function of the time

to failure T .
In case of the minimization of C(h) given by (9) the
respective equation is given by


1
[hcl + (c0 + ca )] + cl A2 h
C(h) = eh 1

A1 (h) (hcl + c0 + ca ) + A1 (h) cl + cl A2 = 0. (12)

When we equate to zero the derivative of (17) we

arrive at the following nonlinear equation:

The main computational problem related either to

the maximization of K(h) given by (5) or the minimization of C(h) given by (9) is caused by difficulties with
the calculation of A1 (h) according to (1), and its derivative given by (11). This function can be expressed in
a closed form only for specially defined probability
distributions F(t). The only popular probability distribution for which A1 (h) is given by a close formula
is the exponential distribution with the hazard rate
equal to the reciprocal of the expected time to failure,
i.e. = 1/ . In this case by simple calculations we
can show that

cl A2 e2h cl heh eh [(c0 + ca )

A1 (h) =

1
.
eh 1

(13)

Let us consider the problem of the minimization

of (5) when the time to failure is exponentially distributed. By elementary calculations we can find that

+ cr + cf cl .

+ cl (2A2 1)] + cl (A2 1) = 0.

(17)

(18)

This equation can be solved only numerically. When

for this purpose we use the Raphson-Newton method
we need to find the derivative of the left-hand side
of (18). This derivative is given by the following
expression:
DC = 2cl A2 e2h cl eh (1 + h)
eh [c0 + ca + cl (2A2 1)] .

(19)

Even in the simplest case of the exponentially

distributed time to failure the solutions of the optimization problems require numerical computations. In
both cases it is necessary to find an appropriate initial
value of h. This value can be found using the following
approximate optimization procedure which is based on

583

the following approximation proposed by Hryniewicz

(1992)
A1 (h)

0.5
h

Table 1. Comparison of approximately optimal and optimal inspection intervals.

(20)

This approximation is valid for when h is significantly smaller than , and what is more important
is valid for any probability distribution of the time to
failure.
When we apply this approximation to the objective
function given by (5) we obtain
K (h) =

h
,
W1 h2 + W2 h + W3

(21)

ZK

hK

hK,opt

1
2
3
5
10

31.622
63,245
94,868
158,114
316,228

31.485
62,701
93,645
154,728
302,804

time to failure as 1000 time units, and will vary the

value of

where

a + 0
A2 0.5

W1 = A2 0.5,

(22)

ZK =

W2 = 0.5a + r + (A2 0.5) 0 ,

(23)

W3 = (a + 0 ) .

(24)

which determines the relation between hK and . The

results of this comparison are presented in Table 1.
The results presented in Table 1 show that even
when inspection interval is comparable with the
expected time to failure the difference between the
approximately optimal solution and the exact optimal solution is negligible from a practical point of
view. Moreover, even in the worse case, i.e. when
ZK = 10, the values of the availability coefficients are
nearly equal for both optimal solutions. It is interesting, however, how the difference between the length of
the approximately optimal inspection interval and the
length of the exactly optimal inspection interval may
influence the value of the availability coefficient. Let
us consider an example where = 1000, = 0.01,
= 0.1, a = 100, 0 = 60, and r = 200. The
value of ZK is equal to 9.991, and the length of the
approximately optimal inspection interval is equal to
hK = 315.94. Under the assumption of the exponential distribution of the time to failure by minimization
of (14) we find that the length of the exactly optimal
inspection interval is in this case equal to 302.54, i.e.
it is about 5% shorter. However, when we compare
the availability coefficients calculated for these two
values we obtain 0.61266 and 0.61281, respectively.
Thus, even in the case when the length of the optimal inspection interval is comparable to the expected
time to failure, and the compared optimal values are
visibly different, the difference between availability
coefficients may be neglected. In more realistic examples, when ZK is much smaller, this difference is even
smaller.
Now, let us find the similar approximate solution of
the optimization problem defined as the minimization
of costs given by (9). The objective function is in this

The derivative of (21) with respect to h gives a

simple equation for the approximately optimal value
of h
W1 h2 + W3 = 0.

(25)

Hence, the approximately optimal inspection interval is given by the following simple expression:

a + 0
hK =
.
A2 0.5

(26)

It is interesting that the optimal solution does not

depend on the expected time of renewal actions. However, this is hardly surprising as the idle time of the
system related to these actions is unavoidable, and is
not influenced by the inspection policy. It is also interesting that in the case of perfect error-free inspections
the optimal inspection interval can be found from a
very simple formula
hK,p =


2 0 .

(27)

Let us evaluate the accuracy of the approximate

solution of the considered optimization problem. We
compare the approximately optimal inspection intervals calculated from (26) with the optimal values
calculated from (15) for the case of the exponential
distribution of the time to failure. We fix the expected

584

(28)

case given by

C(h) =
0.5 [hcl + (c0 + ca )] + cl A2 h
h
+ cr + cf cl .

Table 2. Comparison of approximately optimal and optimal inspection intervals.

(29)

The derivative of (28) with respect to h also gives

a simple equation for the approximately optimal value
of h
cl (A2 0.5) h2 (c0 + ca ) = 0.

hC =

c0 + ca
1
.
A2 0.5
cl

(31)

Also in this case the optimal solution does not

depend upon costs of renewal, cr , and failure, cf , as
these costs are in this model unavoidable, and are not
influenced by the inspection policy. In case of perfect
inspections, the respective optimal inspection interval
is now given by
hC,p =


c0
2 .
cl

hc

hC,opt

1
2
3
5
10

31.622
63,245
94,868
158.114
316.228

31.620
62,868
93,803
154,861
302,879

(30)

Hence, the approximately optimal inspection interval is given by the following simple expression:


ZC

(32)

We may also note the similarity between approximately optimal solutions in both considered models, suggesting the existence of certain equivalence
between both approaches.
In order to evaluate the accuracy of the approximate
solution of this optimization problem we compare the
approximately optimal inspection intervals calculated
from (30) with the optimal values calculated from (18)
for the case of the exponential distribution of the time
to failure. As in the previous case we fix the expected
time to failure as 1000 time units, and will vary the
value of

c0 + ca
1
ZC =
(33)
cl
A2 0.5
which determines the relation between hC and . The
results of this comparison are presented in Table 2.
The results presented in Table 2 show exactly the
same properties of the optimal inspection intervals as
it was presented in Table 1 for the previously considered model. What is more interesting, the accuracy
of approximate solutions, measured in terms of differences between optimal and approximately optimal
results, is very similar for both models. This observation confirms our suggestion that the equality of
ZK and ZC means that the consequences of making

inspections, measured either in terms of availability

or in terms of related costs, are equivalent.

NUMERICAL EXAMPLES

Let us consider two applications of the proposed

solutions to the optimization problems. As the first
example let us consider a production machine which
produces metal cans. The performance of this machine
is described by its ability to produce tight non-leaky
cans, and is measured by the probability of the production of a good can. There are two types of failures
of this machine. The first one is a catastrophic, and
may be noticed immediately after its occurrence. We
are interested, however, in failures of the second type
which are manifested by the increased probability of
the production of potentially leaky cans.
Let us assume that the acceptable quality level,
expressed in terms of the probability of the production of a leaky can, is equal to p1 = 0.001, i.e.
there is on average not more than one nonconforming
can for one thousand of produced cans. The nonacceptable level, whose occurrence is treated as the
occurrence of machines failure is equal to p2 = 0.01,
i.e. we consider the process faulty if there is on average
one or more nonconforming can for each one hundred of produced cans. The current performance of
the machine is measured using a sampling inspection plan taken from the international standard ISO
2859-1 (1999) for which the lot sample size is equal
to n = 125 cans. The alarm is raised when at least
one nonconforming can is found among the sampled
cans. Thus, the probability of a false alarm is equal
to = 1 (0.001)125 = 0.1176 and the probability of not revealing the actual failure is equal to
= (1 0.01)125 = 0.7153. Let us also assume that
the cost of this sampling action is equal to c0 = 1
unit (i.e. we relate all the considered costs to the cost
of inspection). When the alarm is raised additional
500 cans are tested, and when all of them are free
from defects the alarm is considered as false. Otherwise, the machine is considered as failed, and renewal
actions have to be undertaken. Note, that in this case

585

the probability of not revealing a failure is equal to

0.0019, so this additional procedure may be considered as practically error free. The additional cost of a
false alarm is in the considered case equals ca = 4. If
we assume that the expected time between consecutive
failures is equal to = 1000 time units, and the loss
per time unit caused by the work of a failed machine
equals cl = 10, then we have ZC = 0.221, and the
approximately optimal inspection interval, calculated
according to (32) is equal to 6.986 time units. When
the time to failure is exponentially distributed the optimal inspection interval, calculated as the solution of
(18) with the precision to 3 decimal places, is exactly
the same.
Let us now consider another practical example
where the evaluation of costs is difficult or even hardly
possible. In this example we consider a UPS battery
backup system which backups the power supply of a
continuously working computer system. The failure
of the UPS system occurs when its batteries are discharged or/and its switching system is out of order.
The state of the batteries can be evaluated immediately. However, the state of the switching system can
be evaluated by tests which last on average 1 time unit.
Let us assume now that the probability of a false alarm
is equal to = 0, 05 (i.e. on average one of every
20 routine tests triggers an unnecessary alarm), and
the average time necessary to reveal that this alarm
is actually false is equal to 5 time units. Moreover,
let us assume that the probability of not detecting an
existing failure is equal to = 0.1, and - as in the previous example - let us assume that the expected time
between failures is equal to = 1000 time units. The
approximately optimal inspection interval, calculated
according to (26) is equal to 45.227 time units. When
the time to failure is exponentially distributed the optimal inspection interval, calculated as the solution of
(15) equals 44.948 time units, but the availability coefficient (calculated for different values of the renewal
time) are nearly the same.
Both examples show that the proposed approximate
method for the calculation of the optimal inspection interval allows finding of the optimal solution
with the help of a simple pocket calculator. What is
more important, and perhaps even surprising, that the
approximate solutions are very close to the exact ones.
REFERENCES
Badia, F.G., Berrade, M.D., Campos, C.A. 2001. Optimization of inspection intervals based on costs. Journal of
Applied Probability, 38: 872881.

Badia, F.G., Berrade, M.D., Campos, C.A. 2002. Optimal inspection and preventive maintenance of units with
revealed and unrevealed failures. Reliability Engineering
& System Safety, 78: 157163.
Baker, M.J.C. 1990. How often should a machine be
inspected? International Journal of Quality and Reliability Management, 4 (4): 1418.
Barlow, R.E., Hunter L.C., Proschan F. 1960. Optimum checking procedures. In: Proc. of the Seventh
National Symposium on Reliability and Quality Control,
9: 485495
Barlow R.E., Hunter L.C. Proschan F. 1963. Optimum
checking procedures. Journal of SIAM, 11: 10781095.
Berger, K., Bar-Gera, K.,Rabinowitz, G. 2007. Analytical
model for optimal inspection frequency with consideration of setup inspections. Proc. of IEEE Conference on
Automation Science and Engineering: 10811086.
Chelbi, A., Ait-Kadi, D. 1999. Reliability Engineering &
System Safety, 63: 127131.
Chung, K.-J. 1993. A note on the inspection interval of a
machine. International Journal of Quality and Reliability
Management, 10(3): 7173.
Collani von, E. 1981. On the choice of optimal sampling
intervals to maintain current control of a process. In:
Lenz, H.-J., et al. (Eds.) Frontiers in Statistical Quality
Control: 3844, Wuerzburg, Physica Verlag.
Fung, J., Makis, V. 1997. An inspection model with generally
distributed restoration and repair times. Microelectronics
and Reliability, 37: 381389.
Hariga, M.A. 1996. A maintenance inspection model for a
single machine with general failure distribution. Microelectronics and Reliabiliy, 36: 353358.
Hryniewicz, O. 1992. Approximately optimal economic process control for a general class of control procedures.
In: H.J. Lenz et al. (Eds.) Frontiers in Statistical Quality
Control IV: 201215, Heidelberg, Physica Verlag.
ISO 2859-1: 1989(E): Sampling procedures for inspection
by attributes. Part 1. Sampling schemes indexed by
acceptable quality level (AQL) for lot-by-lot inspection.
Khan, F.I., Haddara, M., Krishnasamy, L. 2008. A new
methodology for Risk-Based Availability Analysis. IEEE
Transactions on Reliability, 57: 103112.
Menipaz, E 1978. On economically based quality control
decisions. European Journal of Operational Research, 2:
246256.
Savage, I.R. 1956. Cycling. Naval Research Logistic Quarterly, 3: 163175.
Vaurio, J.K. 1994. A note on optimal inspection intervals. International Journal of Quality and Reliability
Management, 11(6): 6568.
Vaurio, J.K. 1999. Availability and cost functions for periodically inspected preventively maintained units. Reliability
Engineering & System Safety, 63: 133140.
Wang, W., Christer, A.H. 2003. Solution algorithms for
a nonhomogeneous multi-component inspection model.
Computers & Operations Research, 30: 1934.

586

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Optimal periodic inspection of series systems with revealed

and unrevealed failures
M. Carvalho
University of Minho, Braga, Portugal

E. Nunes & J. Telhada

University of Minho, Braga, Portugal

ABSTRACT: Maintenance is one of the main used tools to assure the satisfactory functioning of components
and equipment and the reliability of technological systems. Literature on policies and maintenance models is
enormous, and there are a great number of described contexts where each politics of requisite maintenance is
selected to satisfy technical and financial restrictions. However, by assuming very simplified conditions, many
studies have a limited applicability in reality. Considering a maintenance policy based on periodic inspections,
a model is presented in this article that determines the time interval between inspections that minimizes the
global cost of maintenance per unit of time. It is assumed that the system consists on n series components. It
is recognized the occurrence of failures that are immediately revealed and failures that are only revealed at the
first inspection after their occurrence. The model also incorporates repairing times of components, but both
duration of inspections and duration of preventive maintenance are neglected. The analytical development of
the model allows us to obtain a closed-form function to determine the optimal time period between inspections.
This function will be discussed and a numerical example will be presented.

INTRODUCTION

From an operational perspective, the most critical

phase in the life cycle of a system of technological
nature is the phase of operation and maintenance. This
is also the phase that contributes with the biggest parcel
for the Life Cycle Cost (LCC) of the system. This is the
reason because the problem of estimating the instants
of time for inspections is considered as of great interest or (many times) of primordial importance. These
instants are scheduled to carry out the necessary preventive maintenance of the system in order to maintain
it running normally at a pre-specified level of service.
In the last few decades, the problems of maintenance and substitution of systems have been extensively studied by many researchers. Optimizing the
maintenance costs has been clearly the most common
formulated objective function and many related models had been proposed for those problems. Kallen &
Noortwijk (2006) and Bada et al., (2002) develop
models of maintenance policies with imperfect inspections for single component systems, whose failures are
(or are not) revealed randomly at the instant of failure.
Their models neglected the duration of maintenances
and consider that those lead the component to a state
of as good as new.

Chiang & Yuan (2001) and Wang & Zhang (2006)

also propose similar maintenance models, aiming the
optimization of the global cost, but consider that both
the preventive and corrective maintenances are not perfect. Yet, under the same optimizing objective, other
authors propose models to multi-component systems.
For example, Bris et al., (2003) present a plan of
periodic maintenance for series and parallel systems,
where the failures are only detected at inspections.
Zequeira & Brenguer (2005) develop an analogous
maintenance policy for a parallel system with two
components. Barros et al., (2006) add to the previous
model, the possibility of the failures to be detected at
the instant of its occurrence. In all of these works, the
repairing times are neglected.
2

THE MODEL

This paper develops a policy of periodic inspections

comprising preventive maintenance actions for n independent components in series. It is considered that the
failure of a component is immediately revealed with
probability p and it is not revealed with probability
1 p. The term failure does not necessarily imply that
the system stops working at such an occurrence, but it

587

NOTATION
C
Total cost of maintenance per cycle
C1
Cost of maintenance of revealed failures per cycle
C2
Cost of maintenance of unrevealed failures per cycle
I1
Number of inspections until the occurrence of a revealed failure per cycle
I2
Number of inspections until a detention of an unrevealed failure per cycle
D
Time of not detention of bad functioning for cycle
U
Down-time per cycle
X
Lifetime of the system - E[X ] - (MTBF)
n
Number of components in the system
T
Period of inspection
T
Optimal period of inspection
Probability of performing corrective maintenance to k components (k = 1, . . . n) after inspection
Pr (CMk )
CI
Cost of each inspection plus preventive maintenance
CD
Cost of not detecting a bad functioning (per unit of time)
CU
Cost per down-time unit
CR
Cost of each corrective maintenance to a component
RS (t)
Reliability of the system for a time of mission t
Rk (t)
Reliability of the component k for a time of mission t
1
Cycle of functioning with revealed failure
2
Cycle of functioning with failure not revealed
R
Repairing time of a component (MTTR)

should be understood as the imperfect functioning of

the system.
The basic considerations of the model are:
Whenever an failure is revealed, a corrective maintenance is immediately carried out;
The duration of corrective maintenances are taken
into account, but assumed as constants;
Unrevealed failures are detected at inspections only;
The inspections are periodic, perfect and do not have
any effect on the reliability of the system;
In a given inspection, if it is found that no failure has
occurred in the system (since the previous inspection), only a preventive maintenance is taken; if it
is found that one or more failures have occurred,
then a identical number of corrective actions must
be performed as well as the regular preventive maintenance. Both types of maintenance, corrective and
preventive, restitutes the system (i.e., each of its
components) to the condition of as good as new;
Duration of inspections and preventive maintenances are neglected (null values).
The structure of the series system determines that
the failure of a component implies the failure of the
system. If the failure of a component is revealed at
the instant of its occurrence, it will be immediately
repaired. In this case, it is considered that only one
repair is conducted; unrevealed failures that may have
occurred before will be repaired at the next scheduled
inspection.
On the other hand, if a given failure (in a given component) is not detected immediately at its occurrence

but only at the following inspection, failures in other

components that can effectively occur in the midtime are considered as being of the unrevealed type,
therefore supposing that the system continues to work
uninterruptedly (but imperfectly) until the inspection.
Finally, it is admitted only one repair facility, so whenever two or more failures are detected at a given
inspection, the down time, U , will be sum of the
repairing times of the damaged components.
Considering then a maintenance policy based on
periodic inspections and maintenances that are scheduled for instants iT , (i = 1, 2, . . .), it is intended
to determine the optimum time interval or period
between two consecutive inspections, T , that minimizes the average total cost of maintenance for unit of
time,O[T ]. This cost can be express as:

O [T ] =

E [C ( )]
E [ ]

(1)

In the previous expression, represents the functioning cycle, which is defined as the time interval
between two consecutive renewals of the system. The
length of the cycle depends on the type of the failures occurred. The occurrence of a revealed failure
determines the end of a cycle of functioning and the
commencement of a new (after repairing). In this case,
the ending cycle, 1 , is estimated from the life time
of the system and the down-time associated with the
repairing of the failure. Thus, the average cycle of

588

The average time of unavailability, E[U ], is:

functioning is given by:

E [1 ] = E [X ] + R

(2)

On the other hand, when an unrevealed failure

occurs, it will be detected and repaired at the first
inspection after its occurrence. In this case, supposing that no revealed failure occurs in the mid-time, the
cycle, 2 , extends to a larger time interval that comprises the expected time of life of the component, the
remaining time until next inspection and the sum of all
repairing times needed to be covered by the inspection.
That is, the average cycle of functioning becomes:
E [2 ] = E [X ] + E [D] + E [U ]

(3)

Since revealed (unrevealed) failures occur with

probability p (1 p), the average time of a functioning
cycle is then given by:
E [ ] = E [1 ] p + E [2 ] (1 p)
2.1

(4)

Average total cost of maintenance

per functioning cycle

Concerning to failures, it is assumed that two possible scenarios can occur in the system: (1) the failure
is immediately detected at its occurrence (revealed
failure) which implies the execution of an immediate
corrective maintenance, and (2) the failure occurs but
is only detected in the first inspection after the occurrence (unrevealed failure). In the first case, the maintenance costs per functioning cycle must include the
cost of the expected number of inspections until the
failure occurrence plus the cost of their implicit preventive maintenances, the down-time cost associated
to the (unproductive) time that is necessary to complete the maintenance, and the cost of the corrective
maintenance itself that is necessary to repair the damaged component. In this scenario, the expected cost
per functioning cycle is then given by:
E [C (1 )] = CI E [I1 ] + CU R + CR

E [U ] = R E [CM ]

(8)

where E[CM ] represents the average number of corrective maintenances per functioning cycle.
2.2 Probabilities of inspections (and preventive
maintenances) and corrective maintenances
per functioning cycle
We consider that down-times are considerably inferior
to the interval of time between two inspections. In the
scenario 1, where an occurred failure is immediately
detected and repaired, if the life time of the system is
lesser than the interval of time until the first inspection,
T , then no inspection per cycle is performed. If the life
time extends behind the instant of the first inspection,
T , but not the instant of the second inspection, 2T , then
an inspection per cycle is counted. In general then, if
the life time of the system lies between the ith and the
(i + 1)th inspection, thus i inspections per cycle must
be counted. Thus, the expected number of inspections
per cycle, E[I1 ], for i = 1, 2, . . ., comes as:
E[I1 ] = 0 Pr(0 X < T ) + 1 Pr(T X < 2T ) +
+ i Pr[iT X < (i + 1)T ] +

(9)

That is,
E [I1 ] =

+


RS (iT )

(10)

i=1

The same reasoning applied to scenario 2 of unrevealed failures, implying that the average number of
inspections per cycle, E[I2 ], with i = 1, 2, . . . , is given
by:
E [I2 ] = 1 Pr (0 < X T ) + 2 Pr (T < X 2T )
+ + i Pr [(i 1) T < X iT ] +

(5)

(11)
In scenario 2, the global cost must incorporate an
additional parcel representing the cost associated with
running the system under bad functioning conditions
due to one or more unrevealed failures In this case, the
expected cost per cycle is given by:

(6)

where the average time of not detecting any failure

(implying bad functioning) is given by:
E [D] = T E [I2 ] E [X ]

E [I2 ] =

+


RS ((i 1) T )

(12)

i=1

E [C (2 )] = CI E [I2 ] + CD E [D] + CU E [U ]
+ CR E [CM ]

That is,

(7)

Considering again the scenario 1, the average number of corrective maintenances per functioning cycle,
E[CM ], is given by:
E [CM ] =

589

n

k=1

k Pr (CMk )

(13)

where Pr(CMk ) represents the probability of occurring k corrective maintenances in a cycle, and is
given by:
Pr (CMk ) =

n
n



a (T ) = CI

n


...

i1 =1 i2 =i1 +1

1 Rj (T )


t=i1 ,i2 ,... ,ik

j=i1 ,i2 ,... ,ik

n


Rt (T )
(14)

k=1

Rk (T )

and
b (T ) = p (E [X ] + R )

(15)

k=1

2.3

+ (1p) T

Average total maintenance cost per unit of time

E [C ( )] = CI


+


+


RS (iT )


n

Rk (T )
+ (CU R + CR ) n
i=0

k=1

CD E [X ] (1 p)

RS (iT )+R n

n



Rk (T )

k=1

Our aim is to determine the optimal time interval

between inspections, T , that minimizes the average
total maintenance cost per unit of time, O[T ]. As we
will show later, there is no such a finite optimum under
certain conditions. Basically, the existence of an absolute minimum for the function O[T ] depends on the
relative amplitude among the various costs integrated
into the model. And, independently of such relationship among costs, it can be shown that T +
whenever O[T ] CD .
In the remaining part of this sub-section, we
prove the sufficiency of a relationship among costs
for the existence of a optimal period T . Bada
et al., (2002) verified the validity of this relationship for the case of a single component system following a maintenance policy with imperfect
inspections.

RS (iT ) + CU R + CR p

+ (CI + CD T )

(20)

i=1

+

i=0

From the equations above, the average total cost per

functioning cycle can be formulated by the following
condensed equation:


RS (iT ) + p (CU R + CR CI

CD R ) + (1 p) (CU R + CR CD R )


n

Rk (T ) CD E [X ]
(19)
n

The average number of corrective maintenances per

functioning cycle can still be defined by:
E [CM ] = n

+

i=0

ik =ik1 +1

where

(16)

Now, applying equation (1) and taking in account

the relationships expressed in (2), (3), (4) and (16),
the average total cost per unit of time comes as:

 +

RS (iT ) + CU R + CR p
CI
i=1
 +


O [T ] =
n


p (E [X ] + R ) + (1 p) T
RS (iT ) + R n
Rk (T )
i=0

RS (iT ) + (CU R + CR ) n
Rk (T ) CD E [X ] (1 p)
i=0
k=1
 +


n


p (E [X ] + R ) + (1 p) T
RS (iT ) + R n
Rk (T )

(CI + CD T )
+

k=1

i=0

k=1

The function a(T ), as defined by equation (19), is

a continuous function in R+ , where:

It can be expressed as:

O [T ] = CD +

a (T )
b (T )

(17)

(18)

lim a (T ) > 0

T 0

590

and

lim a (T ) = CI + p (CU R + CR CI CD R )

Table 1 analyzes three numerical cases differentiated

by the probability (or percentage) of occurring unrevealed failures: p = 0.1, p = 0.5 e p = 0.9. In all
cases, we consider the existence of 10 components in
series, whose lifetimes are represented by independent
Weibull distributions.
The reliability function of component k is then
given by:

T +

+ (1 p) n (CU R + CR CD R )
CD E [X ]
The expression defined by equation (20) shows that
b(T ) > 0, T R+ . Taking in account equation (19),
if lim a (T ) < 0, that is, if
T +

E [X ]
>

(21)
then there exists T0 R+ such that a(T0 ) < 0. Applying equation (18), we have that:
a (T0 )
< CD = lim O [T ]
T +
b (T0 )

This result proves that the condition defined in (21)

is sufficient for the existence of an absolute minimum
for the function O[T ].
Table 1.

 T 

Rk (T ) = e

> 0, > 0, T 0

The reliability function of the system with 10

components in series is:

CI +p (CU R +CR CI CD R )+(1p) n (CU R +CR CD R )

CD

O [T0 ] = CD +

NUMERICAL EXAMPLES

RS (T ) = e10

 T 

> 0, > 0, T 0

The expected lifetime of the system is:

 1 + 1 ,
E[X ] =
1

10 /

> 0, > 0

Assuming = 1, the values of the optimal inspection periods and corresponding minimal cost are calculated for each cases. The unit costs considered in
the examples were CI = 10, CD = 100, CR = 25 and

Optimum inspection time and optimum cost when the time to failure is a Weibull distribution.
p = 0.1

p = 0.5

p = 0.9

E [X ]

O [T ]

O [T ]

O [T ]

0.2
0.5
0.8
1
1.1
1.2
1.3
1.4
1.5
2
2.1
2.2
2.3
2.4
2.5
3
4
5
6
10
20

0.0012
0.02
0.0637134
0.1
0.118959
0.138069
0.157124
0.175968
0.194491
0.28025
0.295865
0.31096
0.325544
0.339628
0.353226
0.414484
0.509708
0.579325
0.632048
0.755685
0.867637

0.196586
0.205087
0.213363
0.221413
0.229242
0.265415
0.327678
0.369116
0.394981
0.444197
0.477394

100
100
100
100
100
100
100
100
100
100
97.7365
93.1971
89.2353
85.7534
82.6741
71.4919
59.7013
53.2903
49.0747
39.9996
31.8144

0.250725
0.259736
0.268439
0.276828
0.313843
0.362972
0.390392
0.408787
0.449002
0.478998

100
100
100
100
100
100
100
100
100
100
100
98.4482
94.3313
90.6934
87.4593
75.5591
62.7978
56.017
51.6696
42.7585
35.5176

0.414277
0.428473
0.442472
0.456246
0.469762
0.532297
0.628862
0.695681
0.743316
0.843921
0.921921

100
100
100
100
100
100
100
100
100
100
98.0627
93.6608
89.7803
86.3365
83.2616
71.8094
59.224
52.4872
48.299
40.5767
35.3

591

equal (p = 0.5); this feature is illustrated by Figure 2

for the case of = 2.5.
The condition defined in equation (21) indeed
reveals that it is a sufficient condition for the existence
of a (finite) optimal value T . However, it is clearly
not necessary, as can be demonstrated by the values of
Table 1.

O[T]
400
350
300
250
200
150

100
50
1

This paper developed a model for a maintenance policy

supported by periodic inspections, suitable for applying to technological series systems of n independent
components. The model tolerates the occurrence of
both revealed and unrevealed failures, and the inspections are considered perfect and instantaneous. The
repairing times are admitted as constant values.
A relationship among the involved costs of maintenance, sufficient for the attainment of an optimal
inspection period, was developed and analyzed. Under
way, we are studying the possibility of developing a
sufficient and necessary condition for the existence of
optimal values inspections periods. We are extending
the herein proposed model to integrate other types of
uncertainty inherent to real systems by making use of
fuzzy set theory. Moreover, the work presented in this
paper can be extended in several directions that would
enhance its applicability too, such as k-out-of-n and
parallel systems.

Figure 1. Optimum inspection time and optimum cost when

the time to failure is a Weibull distribution.

O[T]
300

p=0.1

p=0.5

250
200
150
100
50

p=0.9
0.5

1.5

CONCLUSIONS

Figure 2. Optimum inspection time and optimum cost when

the time to failure is a Weibull distribution with = 2.5.

REFERENCES

CU = 50. It is still considered that R = 0.005. The

results were obtained by using Mathematica.
Table 1 shows that an optimal value for the inspection time T does not always exist. Whenever it does
not exist, that is, when T +, it is verified that
O[T ] CD . This means that, under such circumstance, it is not economically compensatory to perform
inspections to the system, and the cost of maintenance
of the system per unit of time corresponds solely to
the cost of the non-detention of damages per unit of
time. In the case of the previous examples this value
has been fixed at CD = 100.
For the costs under consideration, whenever O[T ]
admits a minimum, it can be observed that as bigger
is parameter , the bigger is the value of the optimum time (T ) and the lesser is the maintenance cost.
Figure 1 illustrates this effect for p = 0.1. Also, it
becomes clear that as the fraction of revealed failures decreases, the system must be inspected with an
increasingly frequency. In turn, the optimal maintenance cost per unit of time seems to be maximal when
the fraction of revealed and unrevealed failures are

Bada, F.G., Berrade, M.D. & Campos, C.A. 2002. Optimal inspection and preventive maintenance of units with
revealed and unrevealed failures. Reliability Engineering & Systems Safety 78: 157163.
Barros, A., Berenguer, C. & Grall, A. 2006. A maintenance
policy for two-unit parallel systems based on imperfect monitoring information. Reliability Engineering &
Systems Safety 91: 131136.
Bris, R., Chtelet, E. & Yalaoui, F. 2003. New method to minimize the preventive maintenance cost of series-parallel
systems. Reliability Engineering & Systems Safety82:
247255.
Chiang, J.H. & Yuan, J. 2001. Optimal maintenance policy for
a Markovian system under periodic inspection. Reliability
Engineering & Systems Safety71: 165172.
Kallen, M.J. & van Noortwijk, J.M. 2006. Optimal periodic
of a deterioration process with sequential condition states.
International Journal of Pressure Vessels and Piping 83:
249255.
Wang, G.J., & Zang, Y.L. 2006. Optimal Periodic Preventive Repair and Replacement Policy Assuming Geometric
Process repair. IEEE Transactions on Reliability55, 1:
118122.
Zequeira, R.I. & Brenguer, C. 2005. On the inspection
policy of a two-component parallel system with failure
interaction. Reliability Engineering & Systems Safety71:
165172.

592

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Optimal periodic inspection/replacement policy for deteriorating systems

with explanatory variables
Xuejing Zhao
Universit de Technologie de Troyes, TROYES Cedex, France
Lanzhou University, Lanzhou, P. R. China

Mitra Fouladirad & Christophe Brenguer

Universit de Technologie de Troyes, TROYES Cedex, France

Laurent Bordes
Laboratoire de Mathmatiques Appliques
Universit de Pau et des Pays de lAdour, PAU Cedex, France

ABSTRACT: This paper discusses the problem of the optimization of condition based maintenance policy for
a stochastic deteriorating system in presence of covariates. The deterioration is modelled by a non-monotone
stochastic process. The process of covariates is assumed to be a finite-state Markov chain. A model similar
to the proportional hazards model is used to represent the influence of the covariates. In the framework of a
non-monotone system, we derive the optimal maintenance threshold, optimal inspection period and the optimal
delay ratio to minimize the expected average maintenance cost. Comparison of the expected average costs under
different conditions of covariates and different maintenance policies is given by numerical results of Monte
Carlo simulation.
Keywords: condition based maintenance, covariates, Markov chain, proportional hazards model, nonmonotone system, maintenance, expected average cost

INTRODUCTION

Optimal replacement problems for deteriorating

systems have been intensively studied in the past
decades (see Wang (2002), Brenguer et al. (2003)).
Most of the attention has been focused on statical environment and increasing degradation systems
(Lawless and Crowder (2004), Kharoufeh and Cox
(2005), van Noortwijk (2008), Dieulle et al., (2006)).
Recently more interest and attention are given to two
approaches. In the first approach the problem is to
deal with degradation models including explanatory
variables (covariates), these variables describe the
dynamical environment in the experiments of life science and engineering and they are often expressed by
the proportional hazards model (Singpurwalla (1995),
Newby (1994)). Bagdonavicius and Nikulin (2000)
gave a method to model the increasing degradation
by a gamma process which includes time-dependent
covariates. Makis and Jardine (1992) considered
an optimal replacement problem for a system with
stochastic deterioration which depends on its age and

also on the value of covariates. The second approach

consists in considering a non-monotone deteriorating
system with increasing tendency. Authors in Newby
and Barker (2006) (also in Barker and Newby (2008))
studied the optimal inspection and maintenance policy for a non-monotone system, where the decisions
are based on a non-monotone scale variable, they used
the last exit time from a critical set instead of the first
hitting time to determine the optimal policy.
Generally, the cost-based optimal policy depends
on the inspection interval, as well as the preventive
maintenance level. Some of the researches focus on
the periodic inspection scheme, e.g. Kong and Park
(1997) and Park (1988) used a numerical examples
in which cost is considered as a function of preventive maintenance level while the inspection interval is
given. Newby and Dagg (2003) considered the cost as
a function of inspection interval while the preventive
maintenance level is fixed. Jia and Christer (2002)
considered the maintenance cost as the combination
of the preventive maintenance level and inspection
interval.

593

It is also possible to consider the non-periodic

inspection scheme. In Grall et al., (2002) authors considered the sequence inspection/remplacement policies where the inspection intervals depend on the information of the system deterioration (a function of the
deterioration state). Failure is detected only by inspection and a cost of inactivity of the system per unit time
is calculated as soon as failure occurs. In order to deal
with aperiodic inspections, they used the long-term
expected average cost per unit time,which is based
on of the semi-regenerative properties of the maintained system condition with respect to the steady-state
stationary probability distribution of the maintained
system state. Grall et al., (2002) considered a maintenance policy using a multi-level control-limit rule,
where failures are detected immediately.
In this paper we study the optimal policy of periodic
inspection/replacement for a non-monotone deteriorating system with covariates Zt . The stochastic
deteriorating process Dt = D(t|Ft ) represents the
degradation level of the system given the history of the
covariates {Zt } : Ft = {Zs : s t}. Dt is modelled by
the difference of two conditional independent stochastic processes (Conditionally on the covariates Zt ). We
suppose that the covariates Zt {1, . . ., k} form a
Markov chain with finite state space which describes a
dynamical environment. Following an approach similar to the proportional hazards model proposed by
Cox (See Cox (1972), also in Gouno et al., (2004)),
the covariates effect is modelled by a multiplicative
exponential function.
A method similar to the one proposed by Barker and
Newby (2008) is used to give the optimal maintenance
policy. Suppose that the system is inspected perfectly
at the periodic times = { , 2 , . . .}, the system
states are only known at inspection times, and maintenance actions are instantaneous. We shall denote
by Dk the process D(t) at t = k . We define a
failure threshold L and a maintenance threshold Lp
(Lp < L). Suppose that t = (m + 1) is the first
inspection time where D(m+1) Lp , the system is
maintained at the time t = (m + R + 1) if (1)
Dm < Lp D(m+r+1) < L for r = 0, 1, 2, . . ., R,
where R, defined as the delay ratio, is a decision
variable to be determined, and (2) Dt < L for t
[(m + 1) , . . . , (m + R + 1) ). The system is considered to be failed at time GL = inf {t R+ : Dt L}
and to be replaced at the first inspection time after its
failure. The purpose is to propose an optimal maintenance policy for the considered system in order
to minimize the global long-run expected average
maintenance cost per time unit. In the framework of
the non-monotone system presented previously, we
derive the optimal preventive maintenance threshold
Lp , the optimal inspection interval and the optimal delay ratio R which lead to a minimal expected
average cost.

The other particularity of this paper is that we

compare three different cases for the global expected
average maintenance cost: (1) The optimization when
the covariates are defined by a Markov chain; (2) The
optimization when the covariates Zn = i (i = 1, 2, 3)
are fixed; and (3) The weighted mean of the optimal
results for each Z = i (i = 1, 2, 3) weighted by the stationary probabilities of the Markov chain. All results
are illustrated by a Monte Carlo study.
The structure of the paper is as follows. We model
the degradation processes by a stochastic univariate process in Section 2, where the influence of the
covariates is modelled by a multiplicative exponential function. In Section 3 we study the optimization
maintenance problem when there are two thresholds, corrective replacement threshold and preventive
replacement threshold. For different maintenance cost
units, we find the optimal preventive threshold, the
optimal inspection period and the optimal delay ratio.
Finally, we compare the expected average maintenance
cost per unit time for the three cases mentioned above.

STOCHASTIC DETERIORATION PROCESS

In this Section, we consider a single-unit replaceable system in which an item is replaced with a
new one, either at failure or at planned replacement.
The degradation of the system is represented by a
continuous-state univariate stochastic process D(t)
with initial degradation level D(0) = 0. In this paper,
without loss of generality, we suppose that the deterioration system has an upward trend degradation, though
not necessarily monotonically increasing.

2.1 Deterioration model without covariates

Suppose that the system is subject to stochastic deterioration. The level of deterioration is represented by
a continuous-state univariate stochastic process D(t)
with initial degradation level D(0) = 0.
To describe the non-monotonicity of a system, we
suppose that the variation of the degradation at t,
D(t), is represented by a stochastic process A(t) =
X + (t) X (t), the difference of two independent
stochastic processes, where X + (t) and X (t) denote
respectively the degradation and improvement of the
system. Suppose that the system can be observed at
each time unit tk (k = 1, 2, . . .), so only the discrete
stochastic processes D(tk ), X + (tk ) and X (tk ) can be
observed, denoted respectively by Dk , Xk+ and Xk .
The process Dn (n 1) is defined as:
+

Dn = max(Dn1 + Xn1
Xn1
, 0),

594

(1)

where Xn+ , Xn are independent random variables, of

exponential distribution with mean +

n and n resp.
(without loss of generality, we assume that +

n
n ).
The distribution function and the density function
of the variable Xn+ Xn are given by

Fn (x) =

n
+
n +
n

exp(x/
n )1(x0)


+
n
+
exp(x/
)
1(x0) ,
n

+
n + n

1

fn (x) = +
exp(x/n )1(x0)
n + n

+ exp(x/+
n )1(x0) .

+ 1

pij (k) = P(Zk+1 = j|Zk = i)

P(Dk+1 x|Dk = dk ) = P(Ak + Dk x|Dk = dk )

= P(Ak x dk |Dk = dk )
(2)

(Zn )
exp(x/
n (Zn ))1(x<0)

n (Zn )


+ (Zn ) exp(x/+
n (Zn )
+ 1 n +
1(x0) ,
n (Zn ) +
n (Zn )

exp(x/
n (Zn ))
fn (x, Zn ) =
1(x<0)
+
n (Zn ) +
n (Zn )

exp(x/+
n (Zn ))
+ +
.
1
(x0)
n (Zn ) +
n (Zn )

and for x = 0,
P(Dk+1 = 0|Dk = dk ) = P(Ak + Dk 0|Dk = dk )
= P(Ak dk |Dk = dk )
= P(Ak dk )

k exp(dk /k )
.
+

k + k

(3)

So conditionally on Dn = dn , the r.v Dn+1 has a

mixture distribution, with a density of fn (x dn ) in

n
(0, +) and a mass distribution ++
exp(dn /n )
n
n
at x = 0.

(5)

for n 1, where Xn+ (Zn ), Xn (Zn ) are conditionally independent random variables (given Zn ), with
exponential distribution of mean parameters +
n (Zn )
and
n (Zn ) (without loss generality, we assume that

+
n (Zn ) n (Zn )).
The distribution function and the density function
of the variation An+1 (Zn ) = Xn+ (Zn ) Xn (Zn ) are
given by
Fn (x, Zn ) =

(4)

be the transition probabilities of the process Z. Filtration Ft = {Zs : s t} denotes the history of the
covariates.
We assume that the variation of the degradation at
time tn only depends on the covariates at time tn . Let
stochastic process D(t|Ft ) be the degradation level of
system given Ft . This process is observed at discrete
times t = tn (n N ). We shall denote by Dn the
observed process at time t = tn , defined as:
+

(Zn1 )Xn1
(Zn1 ), 0),
Dn = max(Dn1 +Xn1

Conditionally on Dk = dk (k = 1, 2, . . . ), for
x > 0, the r.v Dk+1 has the distribution

= P(Ak x dk ),

Let

+
n (Zn )+

n
So the distribution of Dn =
k=1 Dk can be
derived using the method of convolution and the total
probability formula.
To describe precisely the influence of the covariates
Zn = zn on An , similar to the proportional hazards
model proposed by Cox (1972), we suppose that the

parameters +
n and n depend on zn as follows:
+
+
n (Zn ) = n (Zn )

2.2 Modelling influence of covariates

on degradation

+
+
= +
n exp (1 1(Zn =1) + + K 1(Zn =K) )

We are interested in the influence of covariates on

degradation. The covariate process Z = {Z(t), t 0}
is assumed to be a finite state Markov chain with states
{1, 2, . . ., K} which describe the states of the environment, such as normal, warning, dangerous, etc.
The covariates are available only at the time points tk ,
(k = 0, 1, 2, . . .), when we observe the degradation
process.

+
= +
n exp (Zn ),

(6)

n (Zn ) = n (Zn )

595

=
n exp (1 1(Zn =1) + + K 1(Zn =K) )

=
n exp (Zn ),

(7)


where +
n (n ) denote the degradation rate (improvement rate) of the system when there is no covariates
considered, + = (1+ , 2+ , . . . , K+ ) and =
(1 , 2 , . . . , K ), from (6) and (7), these parameters allow to account the influence of covariates on the
degradation rate.
Considering the symmetrical property of i , without loss of generality, in what following, we assume
that 1+ 2+ . . . K+ and 1 2 . . . K .

Example 2.1: An example of the degradation for 100

days (inspected per 5 days) is given in Figure 1, where
Zn is a 3-state Markov chain with transition matrix

0.95
P = 0.02
0.00

0.05
0.95
0.05

0.00
0.03 ,
0.95

and initial state Z0 = 1, + = (0.2, 0.5, 1) and

= (0.1, 0.1, 0.1), the baseline mean parameters

+
n = 0.5 and n = 0.3. Notice that the stationary
distribution = (1 , 2 , 3 ) = (0.3, 0.5, 0.2).
For simplification, in what follows, we assume that
0 does not depend on n, and Z {1, 2, 3} is a 3-state
Markov chain as described in the above Example 2.1.
For covariates with initial state Z0 = 1, denote that
n = (1n , 2n , 3n ) with in = P(Zn = i|Z0 = 1), (i =
1, 2, 3), is the distribution of Zn then:
(1n , 2n , 3n )
= (P(Zn = 1|Z0 = 1), P(Zn = 2|Z0 = 1),
P(Zn = 3|Z0 = 1))
= (1, 0, 0)P n ,

Degradation

25
20
15
10
5

10

20

30

40

50

60

70

80

90

100

(a)

Covariables

lim in = i .

(8)

where i is the stationary distribution of the Markov

chain.
Furthermore, we shall denote by Dn (Z) and
Dn ()) respectively the increments of the degradation with a covariate process Z a general Markov chain
starting at Z1 = 1 and a steady-state Markov chain with
stationary distribution .
Let us recall that the covariates form a steady-state
Markov chain. Each replacement makes the system
restart from its initial state (D0 = 0) and the covariates Zn follow their trajectory. Let us denote by n the
instant of replacement. Hence (Dt )(t0) et (Dt+n )(t0)
have the same distribution. So the trajectory of the
degradation does not depend on the history before the
replacement, henceforth the deterioration process is a
renewal process.

CONDITION-BASED PERIODIC
MAINTENANCE MODEL

In this section, we study the optimal periodic maintenance policy for the deteriorating system presented
in Section 2. Suppose that the system is assumed
to be a non-monotone stochastic deteriorating system with initial state D0 = 0, and the state can
exclusively be monitored by inspection at periodic
times tk = k (k = 1, 2, . . . ). We now give some
assumptions under which the model is studied.
1. Inspections are perfect in the sense that they reveal
the true state of the system and the explanatory
variables.
2. The system states are only known at inspection
times and all the maintenance actions take place
only at inspection times and they are instantaneous.
3. Two maintenance operations are available only at
the inspection time: preventive replacement and
corrective replacement.
4. The maintenance actions have no influence on the
covariates process.

30

and

3.1 Maintenance decision

0.0

10

20

30

40

50

60

70

80

90

100

(b)

Figure 1. An example of non-maintained deterioration

system (a) and covariate process (b).

Suppose that the system starts at D0 = 0 and is

inspected perfectly at periodic times = { , 2 , . . . }.
The states are only known at inspection times and
maintenance actions are instantaneous. We define
a failure threshold L and a preventive maintenance
threshold Lp (Lp < L).

596

 The system is considered to be failed and correctively replaced at the first hitting time
GL = inf {t R+ : Dt L | D0 = 0}.

(9)

The degradation of the system maybe recover back

to the critical level or below the critical level after
the exceeding (See Fig. 1). So the system can be
still used for a period of time, especially when the
critical level is the preventive threshold. In practice,
it is useful to consider this information because it
can reduce the maintenance cost. So the preventive
maintenance decision is based on the a last hitting
time
HLp = sup{t R+ : Dt < Lp | D0 = 0}.

(10)

Since the last hitting time is not a stopping time, so it

is seemed to have to search alternative way to deal with
this problem. In Barker and Newby (2008) for multicomponent described by a Bessel process with drift,
the maintenance decision is based on the probability
that the system never recovers, in other words, they use
the probability that the last up-crossing occurs between
the current time and the next scheduled inspection.
We deal with the problem as follows: considering the
increasing tendency of the system degradation and
the probability of the recover, at the time t when the
degradation is exceed the preventive threshold, we take
no maintenance action and we continue inspect for a
period of time, say R units time, where R is defined
as the delay ratio, a decision variable to be determined.
At each inspection time tk = k when Dk < Lp ,
three exclusive events occur:

[0, Lp ) after an interval of time less than R time units.

We define I = max{s {1, . . . , R + 1} : D(k+s) <
Lp , Dt < L, t [k , (k+s) ]}, then event E3 is equal to
R
R


{I = s} E3s . In this case, no maintenance action

s=1

s=1

takes place and the decision is reported to (k + s) .

The only cost incurred in interval [k , (k + s) ] is the
inspection cost s Ci .
An example of the maintained system is given in
Figure 2, where the preventive threshold is Lp = 20,
the corrective threshold is L = 30, the system is
inspected for each 5 days with delay ratio R = 3, the
other parameters are same as in Example 2.1.
The total maintenance cost in [0, t] is:
C(t) = Ci Ni (t) + Cp Np (t) + CF NF (t) + Cd d(t),
where Ni (t) (resp. Np (t), NF (t)) is the number of
inspections (of preventive replacements, of corrective
replacements) till time t.
The expected average cost is:
EC = lim

EC (Z) =

E(v(Z))
,
E(l(Z))

(12)

35
30

Degradation

s=1

25

E1s , where E1s means that the system fails at time

t ((k + s 1) , (k + s) ]. In this case, a corrective

replacement should take place at time T = (k + s) ,
and the system returns to the initial state 0. The maintenance cost includes the corrective replacement cost
CF and cumulated unavailability cost Cd d, where
d = T t is the cumulated unavailability time.
(E2 ) Lp < D(k+r) < L for r = 1, 2, . . ., R and Dt <
L (t [k , (k + R + 1) ]) The system crosses the
threshold Lp (but it is still lower than L) and there is no
observation of recovering during a time period R . In
this case, the system is replaced preventively at time
(k + R + 1) with the preventive replacement cost Cp ,
and the system returns to the initial state 0.
(E3 ) D(k+r) Lp for some r = 1, 2, . . ., R and Dt < L
(t [k , (k + R + 1) ]) It means that the system does
not cross the preventive threshold Lp , or it crosses
the threshold Lp (whereas less than L) but returns to

20
15
10
5
0

50

100

150

200

250

300

350

200

250

300

350

(a)

Covariates

s=1

(11)

When the stochastic process (D, Z) forms a regenerative process, we can calculate the expected cost per
time unit as (Rausand and Hyland (2004))

(E1 ) Dt L (for some t [k , (k + R + 1) ]). Which

R+1

is equivalent to
{(k + s 1) < GL (k + s) }
R+1


E(C(t))
.
t

0.0

50

100

150

(b)
+: inspection; : preventive replacement;
: corrective replacement.

Figure 2. (a) Maintained deteriorating system with two

thresholds: preventive and corrective; (b) The corresponding
covariates process.

597

where E(v(Z)) and E(l(Z)) are respectively the

expected cost and expected length of a renewal cycle.
Considering the three above exclusive events E1 to
E3 , and denoting by Vk (resp. Lk ) the total cost (resp.
the total length) from inspection time Tk to the time
when the system is replaced, since the total cost Vk
(the total length Lk ) is the combination of the cost
(the length) in a time interval [Tk , Tk+s ) and the total
cost (the length) after Tk+s (s = 1, 2, . . . , R + 1), we
calculate the total maintenance cost and the length of
a renewal cycle by the following iterative method:

s=1
R+1

(Ci s + Vk+s )1(E3s ) ,
s=1

Lk = (R )1(E2 ) +

R+1


Case I (Normal cost): Ci = 10, Cp = 60, CF

100, Cd = 250;
Case II (Expensive PR): Ci = 10, Cp = 100, CF
100, Cd = 250;
Case III (Expensive inspection): Ci = 100, Cp
60, CF = 100, Cd = 250;
Case IV (Inexpensive unavailability): Ci
10, Cp = 60, CF = 100, Cd = 100.

s 1(E1s )
EC =

(s + Lk+s ) 1(E3s ) ,

R+1

(CF P(E1s ) + E(Ci s + Cd
s=1

((k + s) GL ))1(E1s ) ) + (Cp + RCi )

R+1


E((Ci s + Vk+s ) 1(E3s ) ),

s=1

E(Lk ) = (R )P(E2 ) +

R+1


EC
(k)k .

(14)

1. The optimal maintenance for three parameters:

preventive threshold Lp , the inspection periods
and the delay ratio R
A numerical optimization was used to give the optimal value of the decision variables Lp = 16, =
12, R = 0 for a deteriorating system with the same
system parameters depicted in Example 2.1 when
Ci = 10, Cp = 60, CF = 100, Cd = 250. These optimal values lead to the optimal expected average cost

EC
= 2.8650. Figure 3 gives the iso-level curves of
EC in the function of (Lp , ) and R takes its optimal
value R = 0.

and the expectation will be

P(E2 ) +

k=1

s=1

E(Vk ) =

Optimal Maintenance cost when Zn form a general

Markov chain;
Optimal Maintenance cost when Zn are fixed to Z =
i, (i = 1, 2, 3) respectively;
Weighted mean of the optimum cost for Z = i (i =
1, 2, 3) with weight the steady-state probability:

s=1
R+1


For each case of maintenance cost, we compare the

following three quantities:

R+1

(CF + Ci s + Cd ((k + s) GL ))1(E1s )
Vk =

+ (Cp + RCi ) 1((E2 )) +

Example 2.1. We consider four different cases of unit

maintenance cost:

E(s 1(E1s ) )

40

s=1
35

R+1


E((s + Lk+s ) 1(E3s ) ).

30

s=1
25

The optimization problem is therefore to chose the

value of Lp , and R that minimize the expected
long-run average maintenance cost:
(Lp , , R ) = arg

(Lp

min

, ,R)(0,L)R+ N

EC (Z).

20

15

10

(13)
5

3.2

10

15

20

25

30

Numerical simulation

In this section we give some numerical simulation of

our model, the deteriorating system is the same as

Figure 3. Iso-level curves of expected average maintenance

cost as a function of (Lp , ).

598

Table 1. The optimal preventive threshold, the optimal

inspection period and the expected average maintenance cost
with periodical inspection.

35

Covariates

(10, 60, 100, 250)

(10, 100, 100, 250)

25

Z general
Z =1
Z =2
Z =3
C

(16, 12, 0, 2.8650)

(14, 33, 0, 1.2658)
(12, 21, 0, 2.1614)
(14, 9, 0, 4.3149)
2.6283

(20, 10, 0, 3.9037)

(16, 27, 0, 1.8097)
(16, 18, 0, 3.0943)
(16, 9, 0, 6.1939)
3.7673

20

Covariates

(100, 60, 100, 250)

(10, 60, 100, 100)

Z general
Z =1
Z =2
Z =3
C

(8, 20, 0, 7.6926)

(2, 60, 0, 2.7868)
(4, 36, 0, 4.7785)
(2, 18, 0, 9.536)
5.8076

(16, 12, 0, 2.7734)

(14, 33, 0, 1.2194)
(6, 36, 0, 2.078)
(14, 12, 0, 4.0442)
2.4963

Case1
Case2
Case3
Case4

EC

30

15

10

10

15

20

25

30

(a)
60
Case1
Case2
Case3
Case4
50

Excepted average cost

18

16

Expected average cost

14

12

10

40

30

20

8
10

6
0

10

15

20

25

30

35

40

(b)

2
35

0
0.5

1.5

2.5

3.5

4.5

Case1
Case2
Case3
Case4

5
30

Figure 4. The optimal cost as a function of 3 for 1 =

0.2, 2 = 0.5.

25

EC

20

Results in Table 1 summarize the results of an

optimization for a deteriorating system with different
maintenance costs.
In all cases of different unit maintenance cost
(expensive or inexpensive), the optimal expected average cost under the condition Z = 1( = 1 ) are
the smallest, because the degradation increments are
smaller in comparison with other cases. The maintenance for Z = 2 ( = 2 ) is more costly than that of
Z = 1, and the maintenance for Z = 3 ( = 3 ) is the
most expensive. As a consequence, the parameter
can be used to express the influence of the dynamical
environment on the deteriorating system.
In order to reveal the way that the maintenance cost
is influenced by the system parameters , using the
symmetrical property of , the optimal expected average cost is computed for various values of 3 with fixed
1 and 2 . The result appears in Figure 4, it can be seen
that the optimal expected average maintenance cost is

15

10

10

(c)

Figure 5. The expected average cost as a function of

(a) Lp for optimal R, ; (b) for optimal R, Lp ; (c) R for
optimal Lp , .

an increasing function of the system parameter 3 . In

fact, since expresses the influence of the dynamic
environment, the expected average maintenance cost
under the worst environment has higher cost than that
of better environment.

599

The expected average maintenance cost for system

with a Markov chain is always more than the weighted
mean (C in Table 1) of the optimal cost for the three
statical cases, since in Markov chain we have less
information on the degradation system. As a consequence, the weighted mean of the optimal cost also
gives a lower bound of the maintenance cost for the
deteriorating system.
2. Comparison of the optimal expected average
maintenance cost for different unit maintenance
costs
The influence of each decision variable on the
expected average maintenance cost function is given
in Figure 5. The influence is given by the curves of
the expected average cost as functions of the R,
and Lp respectively. All results show that no matter
what choice of Lp , R and , the expected average
maintenance cost for an expensive inspection (case
III) is always the most expensive, and the inexpensive
unavailability (case IV) is always the cheapest one.
For fixed = 12, R = 0, Figure 5(a) shows
that for a relative smaller preventive threshold, the
preventive replacement cost Cp determine the optimal maintenance cost (only weak dependence to the
cost of unavailability, as indicated by EC (Case I)
EC (Case IV)), so it can be seen that the suitable maintenance policy is mainly the preventive
replacement whereas we take a corrective replacement for a bigger preventive threshold (this will results
in more failure with the cost of Cd , indicated by
EC (Case I) EC (Case III)) in Figure 5(a)).
For fixed Lp = 16, R = 0, Figure 5(b) shows
that for a smaller inter-inspection such that there is
no maintenance cost paid for unavailability, the maintenance policy is mainly the preventive replacement
(EC (Case I) EC (Case IV)), however a corrective replacement action will take place for a bigger
inter-inspection time.
The optimal maintenance cost increases as R
increases when Lp = 16 and = 12 are fixed, as
Figure 5(c) shows.
4

CONCLUSION

In this paper we deal with a non-monotone deteriorating system with covariates, we use a method similar to
the proportional hazards model to account the influence of dynamical covariates, defined by a 3-state
Markov chain.
Expected average cost is calculated, optimum periodic inspection/replacement policies are derived for
different maintenance costs per unit, as a function of
the preventive level Lp , inspection interval and delay
ratio R. The results show that:
1. The optimal average cost is an increasing function
of the parameters .

2. The optimal inspection interval includes the information on R, and in order to optimize the average cost, we can only consider the parameters
(Lp , ).
3. The expected average maintenance cost for system
with a covariate Markov chain is always greater
than the weighted mean of the optimal cost for the
three statical cases.

REFERENCES
Bagdonavicius, V. and Nikulin, M. 2000. Estimation in degradation models with explanatory variables. Lifetime Data
Analysis 7(1): 85103.
Barker, C.T. and Newby, M. 2008. Optimal non-periodic
inspection for Multivariate degradation model. Reliability
Engineering and System Safety (In press).
Brenguer, C., Grall, A. Dieulle, L. and Roussignol, M.
2003. Maintenance policy for a continuously monitored
deteriorating system. Probability in the Engineering and
Informational Sciences 17(2): 235250.
Cox, D.R. 1972. Regression models and life-tables. Journal
of the Royal Statistical Society. Series B 34(2): 187220.
Dieulle, L., Brenguer, C. Grall, A. and Roussignol, M.
2006. Asymptotic failure rate of a continuously monitored
system. Reliability Engineering and System Safety 91(2):
126130.
Gouno, E., Sen, A. and Balakrishnan, N. 2004. Optimal
step-stress test under progressive type-I censoring. IEEE
Transactions on Reliability 53(3): 388393.
Grall, A., Brenguer, C. and Dieulle, L. 2002. A conditionbased maintenance policy for stochastically deteriorating
systems. Reliability Engineering and System Safety 76(2):
167180.
Grall, A., Dieulle, L. Brenguer, C. and Roussignol, M.
2002. Continuous-time predictive-maintenance scheduling for a deteriorating system. IEEE Transactions on
Reliability 51(2): 141150.
Jia, X. and Christer, A.H. 2002. A prototype cost model of
functional check decisions in reliability-centred maintenance. Journal of Operational Research Society 53(12):
13801384.
Kharoufeh, J.P. and Cox, S.M. 2005. Stochastic models
for degradation-based reliability. IIE Transactions 37(6):
533542.
Kong, M.B. and Park, K.S. 1997. Optimal replacement of
an item subject to cumulative damage under periodic
inspections. Microelectronics Reliability 37(3): 467472.
Lawless, J. and Crowder, M. 2004. Covariates and random
effects in a gamma process model with application to
degradation and failure. Lifetime Data Analysis 10(3):
213227.
Makis, V. and Jardine A. 1992. Optimal replacement in the
proportional hazards model. INFOR 30, 172183.
Newby, M. 1994. Perspective on weibull proportional hazards model. IEEE Transactions on Reliability 43(2):
217223.
Newby, M. and Dagg, R. 2003. Optimal inspection and
maintenance for stochastically deteriorating systems II:
discounted cost criterion. Journal of Indian Statistical
Association 41(1): 927.

600

Newby, M.J. and Barker, C.T. 2006. A bivariate process

model for maintenance and inspection planning. International Journal of Pressure Vessels and Piping 83(4):
270275.
Park, K.S. 1988. Optimal continuous-wear limit replacement under periodic inspections. IEEE Transactions on
Reliability 37(1): 97102.
Rausand, M. and Hyland, A. 2004. System Reliability
Theory: Models, Statistical Methods, and Applications
(2 ed.). New Jersey: John Wiley & Sons Inc.

Singpurwalla, N.D. 1995. Survival in dynamic environnements. Statistical Science 1(10): 86103.
van Noortwijk, J.M. 2008. A survey of the application of
gamma processes in maintenance. Reliability Engineering
and System Safety (In press).
Wang, H. 2002. A survey of maintenance policies of
deteriorating systems. European Journal of Operational
Research 139(3): 469489.

601

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Optimal replacement policy for components with general failure rates

submitted to obsolescence
Sophie Mercier
Universit Paris-Est, Laboratoire dAnalyse et de Mathmatiques Appliques, (CNRS UMR 5050), Champs
sur Marne, Marne-la-Valle, France

ABSTRACT: Identical components are considered, which become obsolete once new-type ones are available,
more reliable and less energy consuming. We envision different possible replacement strategies for the old-type
components by the new-type ones: purely preventive, purely corrective and different mixtures of both types of
strategies. To evaluate the respective value of each possible strategy, a cost function is considered, which takes
into account replacement costs, with economical dependence between simultaneous replacements, and energy
consumption (and/or production) cost, with a constant rate per unit time. A full analytical expression is provided
for the cost function induced by each possible replacement strategy. The optimal strategy is derived in long-time
run. Numerical experiments close the paper.

INTRODUCTION

Identical and independent components are considered,

which may be part of a single industrial equipment or
dispatched in different locations, indifferently. Those
components are degrading with time and their random
life-times follow some common general distribution.
At some fixed time, say time 0, new components
appear in the market, issued from a new technology,
which makes them more reliable, less energy consuming and more performing. Such new-type components
may be substituted to the older ones with no problem of
compatibility. There is no stocking of old-type components and after time 0, no old-type component is
available any more (or the industrialist is not allowed
to use old-type components any more, e.g. for safety
reasons). After time 0, any failed component, either
old-type or new-type, is instantaneously replaced by
a new-type one. At time 0, each old-type component
is in use since some random time, with some random
remaining life-time. If the new-type components are
much less energy consuming than the older ones and
if the period of interest is very long, it may then be
expedient to remove all old-type components immediately at time 0 and replace them by new-type ones,
leading to some so-called purely preventive replacement strategy. On the contrary, in case there is not
much improvement between both technologies and
if the period of interest is short, it may be better to
wait until the successive failures of the old-type components and replace them by new-type ones only at
failure, leading to some purely corrective replacement

strategy. More generally, some mixture of both strategies, preventive and corrective, may also be envisioned
(details below) and may lead to lower costs, as will be
seen later. The point of the present paper is to look
for the optimal replacement strategy with respect of a
cost function, which represents the mean total cost on
some finite time interval [0, t]. This function takes into
account replacement costs, with economical dependence between simultaneous replacements (Dekker,
Wildeman, and van der Duyn Schouten 1997), and also
energy consumption (and/or production) cost, with a
constant rate per unit time.
A similar model as here has already been studied in
(Elmakis, Levitin, and Lisnianski 2002) and (Mercier
and Labeau 2004) in case of constant failures rates
for both old-type and new-type components. In those
papers, all costs were addingly discounted at time 0,
contrary to the present paper. In such a context, it
had been proved in (Mercier and Labeau 2004) that in
case of constant failure rates, the only possible optimal
strategies were either purely corrective or nearly pure
preventive (details further), leading to some simple
dichotomous decision rule.
A first attempt to see whether such a dichotomy is
still valid in case of general failure rates was done in
(Michel, Labeau, and Mercier 2004) by Monte-Carlo
(MC) simulations. However, the length of the MC
simulations did not allow to cover a sufficient range
for the different parameters, making the answer difficult. Similarly, recent works (Clavareau and Labeau
2006a) or (Clavareau and Labeau 2006b) e.g. proposed
complex models including the present one, which are

603

evaluated by MC simulations. Here again, the length

of the MC simulations added to the complexity of the
model, do not allow to provide the optimal replacement
strategy according to the data of the model.
The point of the present paper hence is to answer
to the following questions: is the dichotomy proved
in case of constant failure rates still valid in case of
general failure rates? If not (and it will not), what are
the possible optimal strategies? Finally, how can we
find the optimal strategy?
This paper is organized as follows: the model is
specified in Section 2. Section 3 presents the theoretical results both for a finite and an infinite time horizon.
Numerical experiments are lead on in Section 4.
Concluding remarks end the paper in Section 5.
This paper presents the results from (Mercier 2008),
with different numerical experiments however. Due to
the reduced size of the present paper, no proofs are
provided here, which may be found in the quoted paper.

THE MODEL

We consider n identical and independent components

(n 2), called old-type components in the following.
At time 0, such old-type components are up, in activity. For each i = 1, . . . , n, the residual life-time for the
ith component is assumed to be some absolutely continuous random variable (r.v.) Ui , where Ui s are not
necessarily all identically distributed. The ith (oldtype) component is assumed to fail at time Ui . The
successive times to failure of the n old-type components are the order statistics of (U1 , . . . , Un ). They are
denoted by (U1:n , . . . , Un:n ), where U1:n < < Un:n
almost everywhere (a.e.).
All preventive and corrective replacements (by
new-type components) are instantaneous. The following replacement strategies are envisioned:
strategy 0: the n old-type components are immediately replaced by n new-type ones at time 0. This
is a purely preventive strategy. After time 0, there
are exactly n new-type components and no old-type
component any more,
strategy 1: no replacement is performed before the
first failure, which occurs at time U1:n . At time
U1:n , the failed component is correctively replaced
and the n 1 non-failed old-type components are
simultaneously preventively replaced. This hence is
a nearly pure preventive strategy. Before time U1:n ,
there are exactly n old-type components. After time
U1:n , there are exactly n new-type components,
strategy K (1 K n): no preventive replacement is performed before the Kth failure, which
occurs at time UK:n . This means that only corrective replacements are performed up to time UK:n (at
times U1:n , . . . , UK1:n ). At time UK:n , the failed

component is correctively replaced and the n K

non-failed old-type components are simultaneously
preventively replaced. Before time U1:n , there are
exactly n old-type components. After time UK:n ,
there are exactly n new-type components. For K

2, between times Ui:n and Ui+1:n

(1 i K 1),
there are i new-type components and n i old-type
ones (see Figure 1).
strategy n: no preventive replacement is performed
at all. Before time U1:n , there are exactly n old
type components. Between times Ui:n and Ui+1:n
(1 i n 1), there are i new-type components
and n i old-type ones. After time Un:n , there are
exactly n new-type components.
Once a new-type component is put into activity at
time 0 or at time say Ui:n , it is next instantaneously
replaced at failure by another new-type component.
The successive life-times of such components are
assumed to form a renewal process with eventual delay
Ui:n ; the i.i.d. inter-arrival times are distributed as
some non-negative r.v. V with P (0 V < ) = 1
and P (V > 0) > 0. The renewal function associated
to the non-delayed process is then finite on R+ . Let
E stand for the expectation with respect of the probability measure P on (, A) and for A A, let 1A be
the indicator function with 1A () = 1 if A and
1A () = 1 if \A. The renewal function is then
denoted by V with:



V (t) = E
1{V (1) ++V (k) t }
kN

for t 0, where V (1) , . . . , V (k) , . . . are the successive

inter-arrival times. We recall that V (t) corresponds
to the mean number of renewals on [0, t] of the nondelayed process.
The envisionned cost function represents the mean
total cost on some time interval [0, t]. It is denoted by
CK ([0, t]) when strategy K is used. Two type of costs
are considered:
replacement costs, with economic dependence in
case of simultaneous replacements: each solicitation of the repair team is assumed to entail a fixed

Failure of old components:

corrective replacements by new ones

0 U1:n

U2:n

Ui:n Ui+1:n

UK-1:n UK:n

Corrective replacements
of new components

Figure 1.

604

Corrective and preventive replacements.

cost r (r 0). Each corrective and preventive

replacement involves a supplementary cost, respectively cf and cp , to be added to r (0 < cp cf ).
For instance, the cost for preventive replacement of
i units (0 i n 1) which comes along with the
corrective replacement of one unit is r + cf + icp .
energy and/or production cost, with a constant rate
per unit time (eventually negative, in case of a production rate higher than the energy cost rate). The
rates for an old-type and a new type unit respectively are + and , with 0, R.
(The cost rate is higher for an older unit). The
energy/production cost for j new-type units and
k old-type units on some time intervall [t1 , t2 ] is
( j + k ( + )) (t2 t1 ), where 0 t1 t2 and
j + k = n.
All components both new-type and old-type are
assumed to be independent one with each other.
In all the paper, if X is a non-negative random
variable (r.v.), its cumulative density function (c.d.f.)
is denoted by FX , its survival function by F X with
F X = 1 FX and its eventual probability density
function (p.d.f.) by fX . For t R+ , we also set
X t = min (X , t) and x+ = max (x, 0) for any real x.
Finally, we shall use the following notations:

gK (t) :=

1
(CK+1 ([0, t]) CK ([0, t]))
cp

(1)

for all 0 K n 1, we easily derive the following

corollary.
Corollary 2 Let t 0. For K = 0, we have:
g0 (t) = (a 1) FU1:n (t)

r
cp

  t 
F U1:n (t)
+ n bE U1:n



aE V (t) V (t U1:n )+
and, for 1 K n 1, we have:
gK (t) = (a 1) FUK+1:n (t)
  t

t
+ (n K) bE UK+1:n
UK:n


FUK:n (t) FUK+1:n (t)


 
aE V (t UK:n )+ V (t UK+1:n )+
In order to find the optimal strategy according to the
mission time t and to the data of the model as in the
case of constant failure rates (see (Mercier and Labeau
2004)), the point should now be to find out the sign
of gK (t) for 0 k n 1. This actually seems to
be impossible in the most general case. However, we
are able to give some results in long-time run, which
is done in next subsection.

r + cf
1
a=
cp

0
b=
cp

Setting

THEORETICAL RESULTS
3.2

3.1 Cost functions on [0, t]

Comparison between strategies 0, 1, . . . , n

in long-time run

We first give our results for a finite mission time t.

We first compute the limit of gK (t) when t +.

Theorem 1 Let t 0. For K = 0, we have:

Proposition 3 Assume the distribution of V to be

nonarithmetic and E (Ui ) < + for all 1 i n.
Setting gK () := limt+ gK (t) for all 0 K
n 1, we then have:

C0 ([0, t]) = nt + r + ncp (1 + aV (t))

and, for 1 K n, we have:
CK([0, t]) =

gK () = a 1 + b

K



(r + cf )(FUi:n (t)

a
E (V )

(n K)E (UK+1:n UK:n )

i=1

t
+ E(V ((t Ui:n )+ ))) + E(Ui:n
)

for all 1 K n 1 and

cf
a
1+ b
nE (U1:n U0:n )
cp
E (V )


+ (nK) (r+cf )E(V ((t UK:n )+ ))

g0 () =


t
+ cp FUK:n (t) + E(UK:n
) + nt

where we set

605

U0:n := 0.

A first consequence is that, if b

r+cf
,
E(V )

E(V )

0 or

we then have gK () 0
alternatively
for all 0 K n 1 (we recall that a 1 and
r+c
cf cp ). Consequently, if E(V f) , the best strategy
among 0, . . . , n in long-time run is strategy 0. Such
a result is conform to intuition: indeed, let us recall
that stands for the additional energy consumption
rate for the old-type units compared to the new-type
ones; also, observe that r+cf
is the cost rate per unit
E(V )
time for replacements due to failures among new-type
components in long-time run. Then, the result means
that if replacements of new-type components due to
failures are less costly per unit time than the benefit due
to a lower consumption rate, it is better to replace oldtype components by new-type ones as soon as possible.
a
Now, we have to look at the case b E(V
) < 0
and for that, we have to know something about the
monotony of
DK := (n K) (UK+1:n UK:n ),
with respect of K, where DK is the K-th normalized spacing of the order statistics (U1:n , . . . , Un:n ),
see (Barlow and Proschan 1966) or (Ebrahimi and
Spizzichino 1997) e.g.. With that aim, we have to
put some assumption on the distributions of the residual life times of the old-type components at time
t = 0 (Ui for 1 i n): following (Barlow and
Proschan 1966), we assume that U1 , . . . , Un are i.i.d.
IFR (Increasing Failure Rate), which implies that
(DK )0Kn1 is stochastically decreasing. A first way
to meet with this assumption is to assume that all
old-type components have been put into activity simultaneously (before time 0) so that the residual life times
are i.i.d. (moreover assumed IFR). Another possibility
is to assume that all units have already been replaced
a large number of times. Assuming such replacement
times for the i-th unit to make a renewal process with
inter-arrival times distributed as some U (0) (independent of i), the residual life at time 0 for the i-th unit may
then be considered as the waiting time until next arrival
for a stationary renewal process with inter-arrivals distributed as U (0) . Such a waiting time is known to admit
as p.d.f. the function fU (t) such that:
F U (0) (t)
 1R+ (t),

(2)
E U (0)


assuming 0 < E U (0) < +. Also, it is proved in
(Mercier 2008) that if U (0) is IFR, then U is IFR too.
The r.v. U1 , . . . , Un then are i.i.d. IFR, consequently
meeting with the required assumptions from (Barlow
and Proschan 1966).
We are now ready to state our main result:
fU (t) =

a
Theorem 4 If b E(V
) 0, the optimal strategy
among 0, . . . , n in long time-run is strategy 0.
a
In case b E(V
) < 0, assume that U1 , . . . , Un are
i.i.d. IFR r.v. (which may be realized through assuming that Ui stands for the waiting time till next arrival
for a stationary renewal process with inter-arrival time
(0)
distributed as U (0) , where
 (0)U is a non-negative
IFR r.v. with 0 < E U
< +). Assume
too that Ui s are not exponentially distributed. The
sequence (E (DK ))0Kn1 is then strictly decreasing,
and, setting

a1
c:= a
(V
)b
E

and d:=

cf
cp 1
a
b
E(V )

c,

one of the following cases occurs:

if c E (Dn1 ) : the optimal strategy among
0, . . . , n in long time-run is strategy n,
if c > E (D1 ) :
if d > E (D0 ) : the optimal strategy among
0, . . . , n in long time-run is strategy 0,
if d E (D0 ) : the optimal strategy among
0, . . . , n in long time-run is strategy 1,




if E DK0 < c E DK0 1 for some 2 K0
n 1 : the optimal strategy among 0, . . . , n in long
time-run is strategy K0 .
In (Mercier and Labeau 2004), we had proved the
following dichotomy property: in case of constant
failure rates, only purely preventive (0), nearly pure
preventive (1) or purely corrective (n) strategies can
be optimal for finite horizon. We now know from last
point of Theorem 4 that such a property is not valid any
more in case of general failure rates, at least for infinite
horizon and consequently for large t. We now look at
some numerical experiments to check the validity of
the dichotomy property in case of small t.
4

NUMERICAL EXPERIMENTS

We here assume that Ui s are i.i.d IFR random variables with known distribution. Examples are provided
in (Mercier 2008) for the case where the data is the distribution of some U (0) and the common p.d.f. fU of Ui
is given by (2) (see Theorem 4). All the computations
are made with Matlab.
All Ui s and Vi s are Weibull distributed according to W (U , U ) and W (V , V ), respectively, (all
independent) with survival functions:
U
F U (x) = eU x

for all x 0.

606

V
and F V (x) = eV x

 

We finally compute E V (t UK:n )+ with:

We take:


V = 1/ 2.25 103

U = 1/103 ;

(3)

U = V = 2.8 > 1

E(V ((t UK:n )+ ))

 t
=
V (t u)dfUK:n (t)

(4)

n1
K 1
 t
V (t u)FUK1 (u)F UnK (u)fU (u)du

(Ui s are IFR), which leads to

=n

E (U )
10.5, (U )
4.1,
E (V )
14, (V )
5.4.

We also take:

where the renewal function V is computed via the

algorithm from (Mercier 2007).
For finite horizon, the optimization on K is simply made by computing all CK ([0, t]) for K =
0, . . . , n and taking the smallest. For infinite horizon,
Theorem 4 is used.
The optimal strategy is given in Table 1 for different
values of V and t, as well as the asymptotic results (all
other parameters fixed according to (35). We can see
in such a table that the optimal strategy is quickly stable
with increasing t. More precisely, the optimal strategy
for a finite horizon t is the same as the optimal strategy
in long-time run as soon as t is greater than about
3.5 mean lengths of life of a new-type component.
For t about twice the mean life length, the finite time
optimal strategy is already very near from the longtime run one. Also, any strategy may be optimal, even
for small t.
We now plot in Figure 2 the optimal strategy with
respect of t, for V fixed according to (3). We can
see in such a figure that the behavior of Kopt (optimal K) with increasing t is not regular at all. There
is consequently no hope to get any clear characterization of Kopt with respect of the different parameters
in finite horizon as we had in the exponential case in

n = 10; = 0; = 0.06; cp = 1; cf = 1.1; r = 0 (5)

We compute FUK:n using:

FUK:n (x) =
0

FU (x)

n!
t K1 (1t)nK dt
(K 1)!(nK)!

= IFU (x) (K, n K + 1)

for 1 K n, where Ix (n1 , n2 ) is the incomplete
Beta function (implemented in Matlab), see (Arnold,
Balakrishnan, and Nagaraja 1992) e.g. for the results
about order statistics used in this section.
We also use:

FUK+1:n (t) FUK:n (t) =

FUK (t) F UnK (t)
K
 t

t
from where we derive E UK+1:n
due to:
UK:n

 t
t
=
UK:n
E UK+1:n


F UK+1:n (u) F UK:n (u) du

for 0 K n 1 (we recall U0:n := 0).

Table 1.


Optimal strategy according to t and V .

t \ 1V 103
5
10
15
20
25
30
35
40
45
50
75
100

1.2

1.5

1.75

2.25

2.5

3.5

10
10
10
7
9
10
9
10
10
10
10
10
10

10
10
10
7
9
10
9
9
9
9
9
9
9

10
10
10
6
8
9
7
8
8
8
8
8
8

10
10
2
5
8
9
6
7
7
7
7
7
7

10
10
2
5
7
8
5
6
6
6
6
6
6

10
10
1
4
7
7
4
5
6
5
5
5
5

10
10
1
4
6
6
4
4
5
4
4
4
4

10
10
1
3
5
4
2
3
3
3
3
3
3

10
10
0
2
4
3
1
2
2
2
2
2
2

10
10
0
1
3
1
0
1
1
1
1
1
1

10
10
0
1
1
0
0
0
0
0
0
0
0

607

(Mercier and Labeau 2004) and as we have here in

infinite horizon (Theorem 4).
We next plot Kopt in Figures 36 for t fixed (t = 25)
with respect of parameters cf , , r and cp (all other
parameters fixed according to (35), which shows
that Kopt may vary a lot changing one single parameter. Also, one may note that Kopt decreases with cf
(Fig. 3), (Fig. 4) and r (Fig. 5). Such observations

10

Kopt

10

0
0

0.5

Kopt

Figure 5.

1.5

Optimal strategy w. r. of r for t = 25.

10

2
0

10

20

30

40

50

Optimal strategy with respect of t.

Kopt

Figure 2.
10

6
Kopt

0
0

0.2

0.4

Figure 6.

2
0
1

Figure 3.

1.5

2
cf

2.5

0.6

cp

0.8

1.2

1.4

Optimal strategy w. r. of cp for t = 25.

are coherent with intuition which says that preventive

maintenance should be performed all the earlier (or
equivalently new-type components should be introduced all the earlier) as failures are more costly, as
the difference of costs is higher between both generations of components, or as economical dependance
between replacements is higher. Similarly, Figure 6
shows that Kopt increases with cp , which means that
the higher the cost of a preventive replacement is, the
later the preventive maintenance must be performed.
This is coherent with intuition, too.

Optimal strategy w. r. of cf for t = 25.

10
8

Kopt

6
4

2
0
0

Figure 4.

0.02

0.04

0.06

0.08

Optimal strategy w. r. of for t = 25.

0.1

CONCLUSIONS

In conclusion, we have considered here different

replacement strategies for obsolescent components by
others issued from a newer technology. A cost function on a finite and on an infinite time horizon has been
considered, in order to sort the different strategies one

608

with each other. We have seen that the variations of

the optimal strategy with respect of a finite horizon
t is much less regular in the present case of general failure rates than in the case of constant failure
rates as in (Elmakis, Levitin, and Lisnianski 2002) or
(Mercier and Labeau 2004) (see Figure 2). Also, the
main result from (Mercier and Labeau 2004), which
told that the optimal strategy could only be strategy
0, 1 or n, namely (nearly) purely preventive or purely
corrective, is here false: any strategy among 0, 1, . . . , n
may be optimal.
It does not seem possible here to give clear conditions on the data to foretell which strategy is optimal
in finite horizon as in case of constant failure rates. We
however obtained such conditions in long-time run. 1
A few numerical experiments (see others in (Mercier
2008)) seem to indicate that the optimal strategy in
long-time run actually is quickly optimal, namely for
t not that large. The results for long-time run then
seem to give a good indicator for the choice of the best
strategy, even for t not very large.
REFERENCES
Arnold, B. C., N. Balakrishnan, and H. N. Nagaraja (1992).
A first course in order statistics. Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics. New York: John Wiley & Sons Inc.
A Wiley-Interscience Publication.
Barlow, R. E. and F. Proschan (1966). Inequalities for linear
combinations of order statistics from restricted families.
Ann. Math. Statist. 37, 15741592.

Clavareau, J. and P.-E. Labeau (2006a). Maintenance

and replacement policies under technological obsolescence. In Proceedings of ESREL 06, Estoril (Portugal),
pp. 499506.
Clavareau, J. and P.-E. Labeau (2006b). Maintenance et
stratgies de remplacement de composants soumis obsolescence technologique. In In Proc. Lambda-Mu 15, Lille
(France).
Dekker, R., R. E. Wildeman, and F. A. van der Duyn Schouten
(1997). A review of multicomponent maintenance models
with economic dependence. Math. Methods Oper. Res.
45(3), 411.435. Stochastic models of reliability.
Ebrahimi, N. and F. Spizzichino (1997). Some results on
normalized total time on test and spacings. Statist. Probab.
Lett. 36(3), 231243.
Elmakis, D., G. Levitin, and A. Lisnianski (2002). Optimal
scheduling for replacement of power system equipment
with new-type one. In Proc. Of MMR2002 (Mathemathical Methods in Reliability 2002), Trondheim (Norway),
pp. 227230.
Mercier, S. (2007). Discrete random bounds for general random variables and applications to reliability. European J.
Oper. Res. 177(1), 378405.
Mercier, S. (2008). Optimal replacement policy for obsolete components with general failure rates.Appl. Stoch.
Models Bus. Ind. 24(3), 221235.
Mercier, S. and P.-E. Labeau (2004). Optimal replacement
policy for a series system with obsolescence. Appl. Stoch.
Models Bus. Ind. 20(1), 7391.
Michel, O., P.-E. Labeau, and S. Mercier (2004). Monte Carlo
optimization of the replacement strategy of components
subject to technological obsolescence. In Proc. Of PSAM
7- ESREL 04, Berlin (Germany), pp. 30983103.

609

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Optimization of the maintenance function at a company

Smail Adjabi, Karima Adel-Aissanou & Mourad Azi
Laboratory LAMOS, University of Bejaia, Bejaia, Algeria

ABSTRACT: The main objective of this work is optimizing the function of maintenance at the foundry of the
company BCR in Algeria. For this, we use a comprehensive approach involving two global areas: organizational
aspect and technical aspect. As a first step, we analyse the reliability of a certain equipment through a Pareto
Analysis. After that, we present the influence of repair times on the unavailability of this equipment. In order
to calculate the optimal renewal times of some spare parts of an equipment (for they present an important time
of unavailability), we first make an economic evaluation, which leads to express the direct and indirect costs
of maintenance. Finally, in order not to charge an available item (good condition), we give an overview of
implementing a non destructive control.

INTRODUCTION

The essential problem for a maintenance service is

the optimization of its maintenance plan and the
costs involved for the accidental failures. They may
have serious consequences on the technical, economic
and organizational parameters. The industrial maintenance, which aims to ensure the good functioning of
the production tools, is a strategic function in a company. This is bound to the apparition of the new kind
of management, to the technological development and
to the necessity of reduction of the production costs.
Today it is no longer its sole objective to repair work
tools, but also to anticipate and prevent shortcomings,
in order to maximize the overall performance of production, and meet the demands of their customers in
term of quality and quantity while respecting delivery
times.
The objective of this work is to minimize the costs
involved for the accidental failures at this company.
For this, we propose to put in place the tools to organize and optimize the maintenance function through
reliability, availability and techno-economic study.
The main issues which we will try, in this work, to
provide some answers are:
What are the equipments to Study?
What are the performances of these equipments ?
What is the most efficient method of maintenance
according to the performance of these equipments?
How to calculate the optimal time for preventive
renewal?
To answer all these questions, we will determine the
equipments to be studied. It will prepare an analysis
of the reliability and availability of these equipments,

and then we will make a techno-economic study which

will determine the optimal time for renewal. Finally,
we propose methods of the implementation of the nondestructive control.

SELECTION OF EQUIPMENT

This study is directed to the foundry section of the

subsidiary SANIAK (BCR), which presents one of
the major concerns of its managers, because of its
importance in factories and significant downtime.
Given the important number of equipments that
make up the section, it is primordial to target only the
critical equipments. For this purpose, we conducted
an ABC (Pareto) analysis to determine the class of
equipments which, for a year, has spawned the highest downtime. We proposed that Class A is the one
corresponding to the equipments with a percentage
of 38.24% breakdowns which cause over 60% of the
downtime, as illustrated in the Figure 1.
The equipments chosen are presented in the table 1.

3
3.1

ANALYSIS OF EQUIPMENT RELIABILITY

Parametric modeling equipment reliability

For our study, this approach begins with an assumption

that the random variable X lifetime has a common
usage model; the Weibull or an exponential distribution. The estimation of the parameters of each model
was produced by the method of maximum likelihood
using the statistical software R. that can validate the

611

models obtained through the classic test of adequacy

Kolmogorov-Smirnov. The results are in Table 2.
n: Size of sample.
: Parameter form of the Weibull law .
: Parameter scale of the law Weibull.
Dks : Statistics empirical test Kolmogorov-Smirnov.
d(n,0.05) : Quantile tabulated test Kolmogorov-Smirnov
with a level of significance equal to 0.05.

Figure 1.

The results show that the model of the two parameters Weibull is accepted for a level of significance =
0.05 for the facilities: DECRT, ROVT021, ROVT041,
ROVT042, GREN011, GREN021, NOYT020, BASP,
but For the NOYT018 machine, the model of Weibull
is rejected.
In addition to a tendency of these facilities(equipment) to the Weibull law : DECRT, ROVT041, ROVT042, GREN011, GREN021, NOYT020,
their exponentiality in lifetime is validated as well. For
equipment BASP, ROVT021, the exponential model is
rejected.

Diagram ABC (Pareto).

Table 1.

Equipment chosen.

Code

Designation

ROVT021

Forging Machine mechanical ROVETTA

F180 (4602)
Casting Machine to 2 heads BP 225.S A.O
Machine for cutting bars
Machine to produce kernels (H 2,5 Co)
Installation sandblasting 200 Kgs/
Grenailleuse GF
Grenailleuse TAMBRO 1000 EXK
Automatic Machine to produce kernels
Forging Machine mechanics
S/PRESSE FO300 (4604)
Forging Machine mechanics
S/PRESSE FO300 (4603)

BASP
DECRT
NOYT018
GREN021
GREN011
NOYT020
ROVT042
ROVT041

Table 2.

3.2 Non parametric modeling

In order to have another support for the reliability of
the equipments, one has opted, to the non parametric
modeling, in particular the graphic test based on the
TTT-Statistical. This consists of rejecting the Exponential model if one notices a meaningful gap between
the curve and the first bisector in favor of the IFR law
if the curve is concave or for a DFR law if the curve is
convex.
The diagrams Figure 2 and Figure 3, illustrate the
S(ti )
points below ( ri , S(t
) exit of the TTT-Statistical.
r)

Results of the modeling parametric reliability.

Equip

Model adjusted

Parameters

Dks

d(n,5%)

GREN011

197
244

NOYT018

203

NOYT020

195

ROVT021

110

ROVT041

165

ROVT042

112

DECRT

268

BASP

388

= 1.08, = 79.12
= 0.013
= 1.07, = 66.68
= 0.087
= 1.05, = 127.28
= 0.0080
= 1.09, = 138
= 0.0075
= 1.25, = 182.04
= 0.0059
= 1.15, = 176.77
= 0.0059
= 1.14, = 196.45
= 0.0053
= 0.97, = 150.98
= 0.0065
= 1.32, = 55.66
= 0.0195

0.068
0.057
0.043
0.053
0.097
0.088
0.070
0.057
0.055
0.135
0.092
0.075
0.059
0.063
0.050
0.005
0.034
0.139

0.097

GREN021

Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential
Weibull
Exponential

612

0.087
0.095
0.097
0.13
0.108
0.128
0.083
0.069

Table 3.
Equip

Graphic test.
n

Curve of
tendency

GREN011 197 Close to the

first bisector
GREN021 244 Close to the
first bisector
NOYT018 203 Close to the
first bisector
NOYT020 195 Close to the
first bisector
ROVT021 110 Concave
ROVT041 165 Close to the
first bisector
ROVT042 112 Close to the
first bisector
DECRT
268 Close to the
first bisector
BASP
388 Concave

Figure 2. Graphic test for: DECRET, ROVT041, ROVT042

and ROVT021.

Model

Rate

Exponential Constant
Exponential Constant
Exponential Constant
Exponential Constant
IFR
Growing
Exponential Constant
Exponential Constant
Exponential Constant
IFR

Growing

exponential model. That is to say that the failure rate of such equipment is common, which
is confirmed by the test chart that validated the
exponential model of its lifetime.

AVAILABILITY OF EQUIPMENTS

The calculation of the operational availability

involves:
Figure 3. Graphic test for: GREN011,
NOYT018, NOYT020 and BASP.

GREN021,

Reliability by MUT (Mean UP Time);

The maintainability by MTTR (Mean Time To
Failure);
The repair time annexes.
The operational availability is given by:

From the preceding graphs: Figure 2 and Figure 3,

we will pronounce the results described in the table 3.

Dop =
3.3

MUT
MUT
=
,
MUT + MDT
MTBF

(1)

Comparison and interpretation of the results

With:
MDT: (Mean Down Time), which includes the
average repair time and repair time annexes.
When these annexes are negligible before the repair
time, the operational availability will be:

According to the preceding results, one notices homogeneity between the results gotten by the parametric
modeling and those non parametric because they join
themselves in the description of the type of failure.
 For the BASP and ROVT021, the parameter of
shape of the model of Weibull is superior to 1. Their
failure rates are therefore increasing with age, what
corresponds, in the IFR law found by the graphic
test. It means that their breakdowns are due to the
ageing.
 For the rest of the equipment, modeling parametrical reliability of the their lifetimes gave the model
Weibull with parameter shape close to 1 and the

Dop =

MUT

MUT + MTTR

(2)

With this approximation, one can illustrate the

fragility of the management system of maintenance
at this company.
In what follows we will model the maintainability and the down time to estimate respectively MTTR

613

Table 4.

Results of the modeling of the repair times.

Equip

Law adjusted

Parametres

ROVT021
ROVT042
DECRT
ROVT041
BASP
GREN021
GREN011
NOYT020
NOYT018

111
113
269
166
389
245
198
197
204

Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal

m = 0.93,
m = 1.06,
m = 0.74,
m = 0.75,
m = 0.71,
m = 0.78,
m = 0.66,
m = 0.59,
m = 0.55,

Table 5.

= 1.27
= 1.13
= 1.18
= 1.14
= 1.07
= 0.96
= 0.96
= 0.88
= 0.83

Dks

d(n,0.05)

MTTR

0.11
0.08
0.05
0.07
0.06
0.07
0.03
0.06
0.04

0.13
0.12
0.08
0.10
0.06
0.08
0.09
0.09
0.09

5.67
5.48
4.23
4.12
3.62
3.49
3.07
2.68
2.49

Results of the modeling of the immobilization times.

Equip

Law adjusted

Parametres

Dks

d(n,0.05)

MDT

GREN011
GREN021
NOYT018
NOYT020
ROVT021
ROVT041
ROVT042
DECRT
BASP

198
245
204
197
111
166
113
269
389

Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal
Log-normal

m = 1.17 , = 0.93
m = 1.46, = 1.04
m = 1.21, = 0.92
m = 1.20, = 1.01
m = 1.61, = 1.54
m = 1.51, = 1.33
m = 1.86, = 1.36
m = 1.31, = 1.49
m = 1.47, = 1.10

0.05
0.07
0.08
0.05
0.12
0.08
0.13
0.08
0.06

0.09
0.08
0.09
0.09
0.13
0.11
0.13
0.08
0.07

5.00
7.38
5.14
5.49
16.37
11.05
16.35
11.24
7.96

Table 6.

Results of the modeling of the availability times, where Ddif = Dopr Dops and Dres =

Dopr Dops
1Dopr .

Equip

MUT(h)

MTTR(h)

MDT(h)

Dopr

Dops

Ddif

Dres

GREN011
GREN021
NOYT018
NOYT020
ROVT021
ROVT041
ROVT042
DECRT
BASP

76.75
64.68
125.17
133.17
169.2
167.93
186.62
153
51.20

3.07
3.49
2.49
2.68
5.67
4.12
5.48
4.23
3.62

5.00
7.38
5.14
5.49
16.37
11.05
16.35
11.24
7.96

0.93
0.89
0.96
0.96
0.91
0.94
0.92
0.93
0.86

0.96
0.95
0.98
0.98
0.97
0.98
0.97
0.97
0.93

0.03
0.06
0.02
0.02
0.06
0.04
0.05
0.04
0.07

0.43
0.54
0.50
0.50
0.66
0.66
0.62
0.57
0.50

(Mean Time To Failure), and TDM (Mean Down

Time), in order to evaluate Dopr et Dops .
The results are reported in table 4, table 5 and
table 6. In an effort to highlight the potential impact
of the annexes repair time on the unavailability of
equipments, we have adopted the following steps:
 calculate, as a first step, the actual availability Dopr
and Dops ;
 calculate, in a second stage, the unavailability
(Dopr Dops ) trained by the waiting time of repair;

 finally we estimate the rate of unavailability caused

by the repair time annexes.

Calculate of the reports between the operational

unavailability to the supplementary time of left and
the operational unavailability watch that the times
annex of repair represents more the half of the times
unavailability. As analysis of the situation is therefore
indispensable, in order to reduce these times of immobilization. Him also invites to review the politics of the
management of the stocks of the pieces of spare and

614

Table 7.

5.3

The chosen components.

Code

Designation

POT300
THER
DETP
AXCL
VISB

POTENTIOMETER LWG300 N 014311

THERMOCOUPLE NI CR AL 14X2
END OF COURSE 516326E
AXIS FLUTED IMR
SCREW OF BLOCKAGE

Replacement depending on the age

The replacement policy under age is to replace by a

new component with a Cp cost another component as
it reaches the age T . A failure before the age T causes
a breakdown cost Cd + Cp .
Under this policy, the average unit cost is estimated by:
(T ) =

Table 8.

Costs of maintenance.

(Cd + Cp )[1 R(T )] + Cp .R(T )

T
0 R(t)dt

(3)

Code

CP (DA)

Cd (DA)

In cases where T tends to infinity we find the

average cost per usage unit of a corrective maintenance c .

POT300
THER
DETP
AXCL
VISB

47036.96
3136.46
3772.05
6813.46
291.99

24000
28800
19200
48000
24000

C p + Cd
(Cd ) + Cp
=
= c ,
() = 
MUT
R(t)dt
0

(4)

The economic justification of the preventive maintenance comes from the calculation of the gain by using
a preventive maintenance policy.
to put the adequate means in place for a better hold in
charge of the repairs.

Gain = (T0 ) ().

Where T0 the optimum time is a solution of the
derive of (T ), it is a solution of the equation:

OPTIMIZATION OF THE RENEWAL

The importance of the BASP machine possesses in

the chain of production and the rate of unavailability
importing of this one.
It is on this account that we chose to study optimal
replacement for some of its components, in order to
optimize its overall performance.

5.1

Selection of components

we have chosen components which have a fairly large

frequency of use, in order to have enough data. These
components are five in number, presented in the
table 7.

5.2

Evaluation of maintenance costs

The costs involved in a maintenance policy can be

separated into: compressible Cp (cost of preventive
maintenance) and Cd (costs of defiance in service). In
the first category are involved the prices of spare parts
and Workers. Whatever the planned maintenance is, no
gain is possible. For the second category, one variable
is taken into account: the cost of the unavailability
related to maintenance actions to be carried out. The
results are given in the table 8.

(T )

R(t)dt + R(t) =

Cd
.
(Cd Cp )

(5)

if Cp > Cd the equation has no solution, the most

economical renewal is curative.
5.3.1 Research the optimum replacement:
Kelly model
It allows you to determine the optimum time of preventive change, based on Weibull law parameters and
Kelly abacuses. To use these abacuses it is necessary
to determine the parameters of Weibull law form and
the ratio r = CCdp .
for this we have modeled the lifetime of components
with the Weibull law. The results are presented in the
table 9.
The results of calculating the optimum time T0 of
Preventive change are given in the table 10.
5.3.2 Kay model
From the equations (3) and (4) the preferably relationship (T ) < () ) is expressed by the formula:

615

T
 0
0

R(t)dt
R(t)dt

k + (1 k) F(T ), k =

Cp
Cp + Cd
(6)

Table 9.

Modeled the lifetime of components.

Equip

Law

Parametres

D(ks)

d(n,5%)

POT300

16

Weibull

0.143

0.328

THER

26

Weibull

0.267

0.267

DETP

55

Weibull

0.114

0.183

AXCL

17

Weibull

0.168

0.318

VISB

28

Weibull

= 3.86,
= 3284.45
= 2.21,
= 1002.25
= 4.23,
= 2460.00
= 3.39,
= 5156.40
= 2.33,
= 1968.67

0.075

0.257

Table 10.

Optimal times of the renewals abacus of Kelly.

Compo

Parameters of
Weibull law

POT300
THER
DETP
AXCL
VISB

= 3.86,
= 3284.45
= 2.21,
= 1002.25
= 4.23,
= 2460
= 3.39,
= 5156.40
= 2.33,
= 1968.67

Cd
Cp

T0 = X .

0.51

9.18

0.36

5.18

0.52

1279.2

7.05

0.45

2320.38

82.19

0.17

334.67

r=

Figure 4. Graphic procedure of calculation of the optimal

age for AXCL.
Table 11. Calculation of twisted them of the graph exit of
the TTT-Transformation for AXCL.

360.81

This relationship has an interesting graphic interpretation operated by Kay (1976) in the case of several
criteria. Considering the first member of this relationship, it is a function strictly increasing  of t = F(T )
in the interval [0,1]to value in [0,1]. It is written in the
form:
 T =F 1 (t)
(t = F(T )) =


0

R(t)dt

R(t)dt

Ti

Ti Ti1

ti = i/r

S(Ti )

Si
S(T16 )

0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

1000
2400
2500
3200
3200
4500
4500
4588
4800
4840
4900
5240
5290
5900
6400
6912

0
1000
1400
100
700
0
1300
0
88
212
40
60
340
50
610
500
512

0,062
0.125
0.187
0.25
0.312
0.375
0.437
0.5
0.562
0.625
0.687
0.75
0.812
0.875
0.937
1

24000
45000
46400
55500
55500
69800
69800
70592
72288
72568
72928
74628
74828
76658
77658
78170

0
0.307
0.575
0.593
0.709
0.709
0.892
0.892
0.903
0.924
0.928
0.932
0.954
0.957
0.98
0.993
1

Table 12.

(7)

The time T0 minimizes (T ) defined by the relation

(3) exactly when t = F(T0 ) maximizes the expression
(t)/[t + k/(1 k)]. While t abscissa is the point
of intersection of the curve y = (t) with his tangent
equation y =  (t ).[t + k/(1 k)] Following the
point abscissa t = tk = k/(1k) and orderly 0. The
age of optimal replacement T0 = F 1 (t ) thus, comes
from the graphic procedure TTT-transformation, for
component AXCL, as illustrated in the figure 4.
From the graph we have t = i/r = 0.125, and
if we go back to the table 11, it corresponds to T2 =
2400. The optimal time of renewal for the component
is then AXCL T 0 = 2400 (h). Results for the other
components are on the table 12.

Rangi

Optimal time of preventive renewal.

Code

tk

T0

POT300
THER
DETP
AXCL
VISB

1.96
0.11
0.19
0.14
0.012

2.04
0.12
0.23
0.16
0.012

0.038
0.163
0.125
0.036

350
1282
2400
345

5.3.3 Interpretation and comparing results

Based on the results, there is a similarity between the
results found with the two model. They show that:
if Cd > Cp , there is no optimal time for renewal,
and that the most advantageous maintenance is the
curative one.

616

The optimal times of preventive renewal obtained

by the two methods are similar. The difference is due to
the results found by the model Kay, which are based on
the values of the whole sample, or on discrete values,
whereas those found by the model Kelly are obtained
from an adjustment of a continuous law.
6

NON DESTRUCTIVE CONTROL

After the results of the optimal time for renewal are

obtained, and in order not change a component in good
condition, it is proposed to implement the procedures
of organization of a non destructive control (NCD)
which consists of:
Define the objectives and the equipments to be
followed, and evaluate the causes of failures;
Study the feasibility of NCD;
choose the methods and techniques of the Nondestructive inspections to be used;
Study and establish guidelines alarm;
Establish economic assessments;
train all staff concerned.
7

CONCLUSION

In this study we used a global approach that uses

two main areas: The organizational aspect and the
techno-economic aspect. In a first step, thanks to the
ABC analysis, it was possible to identify the equipments that cause more than 60% of immobilization
of the foundry section. After that, we have modeled equipments reliability, by using parametric and
non-parametric approaches, which helped to highlight
the types of failures of these equipments. We have
seen that the majority of these are subject to random

failures. Then we introduced a phenomenon found in

the collection of data, which is the unavailability linked
to repair time annexes, that represents more than half
of the operational unavailability of equipments.
The evaluation of the economic consequences has
led us to express the direct and indirect costs of maintenance in order to calculate the time of the optimal
renewal of certain pieces of equipment that presents
important time (cost) unavailability. This shows that
the decisions of Preventive renewal are not only the
results of a technical study.
For a better system of managing the maintenance,
it is proposed to revise the current system of management, by investing in the implementation.

REFERENCES
Bunea, C. Bedfford, T. (2002). The effect of model uncertainty on maintenance optipization. IEEE, 486493.
Canfield, R. (1983). Cost optimisation of periodic preventive
maintenance. IEEE, 7881.
Cocozza-Thivent, C. (1997). Processus stochastique et fiabilt des systmes. Springer.
Gasmi, S. love, C. and W. Kahle (2003). A general
repair, proportional-hasard, framework to model complex
repairable systems. IEEE, 2632.
Lyonnet, P. (2000). La maintenance mathmatique et mthodes. Tec and Doc.
Pellegrin, C. (1997). Fondements de la Dcision Maintenance. Economica.
Pongpech, J. Murth, D. Optimal periodic preventive maintenance policy for leased equipment. Science Direct.
Priel, V. (1976). La maintenance: Techniques Modernes de
Gestion. EME.
Yves, G. Richet, D. and A. Gabriel (1999). Pratique de la
Maintenance Industrielle. Dunod.

617

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Planning and scheduling maintenance resources in a complex system

Martin Newby & Colin Barker
Centre for Risk Management, Reliability and Maintenance
City University, London

ABSTRACT: The inspection and maintenance policy is determined by the crossing of a critical threshold
by an aggregate performance measure. Rather than examining the first hitting time of the level, we base our
decisions on the probability that the system will never return to the critical level. The inspection policy is state
dependent and we use a "scheduling function" to determine the time to the next inspection given the system
state. Inspection reveals the true state of the system and allows the determination of the appropriate action, do
nothing or repair, and the time of the next inspection. The approach is illustrated using a multivariate system
model whose aggregate measure of performance is a Bessel process.

INTRODUCTION

The models derived in this paper are a natural extension of models which use the first hitting time of a
critical level as a definition of failure (Barker and
Newby 2006). Here we develop a model in which
the system is repaired if the probability of returning to the critical level is small (a last exit time) and
it has not crossed a second level which corresponds
to catastrophic failure. The intention is to maintain
a minimum level of performance. The approach is
appropriate when the system is subject to relatively
minor repairs until it begins to degrade faster and
requires major repair. This is typically the behaviour of
infrastructure and large capital items. It also captures
the behaviour of systems which eventually become
economically obsolete and not worth repairing. The
system is complex in the sense that it consists of a
number of components whose states evolve in time.
The system state is summarized using a Bessel process Rt [0, ) which is transient and thus tends
to increase. Transience ensures that the process will
eventually escape to . There are two critical levels,
and F > . The system is repaired if on inspection it has a small probability of returning to , and
suffers a catastrophic failure if it reaches F. The time
to inspection and repair is determined by a scheduling
function (Grall et al. 2002) which gives the time until
the next action as a function of the current state.
The threshold defines the repair actions and is
incorporated in a maintenance function r. The actions
are determined by the probability that the process
has escaped from [0, ) and F defines the failure of
the system and hence its replacement. The sequence

0
of failure and replacement times GF
constitutes a
renewal process. This embedded renewal process is
used to derive the expected cost per unit time over
an infinite time horizon (for the periodic inspection
policy) and the total expected cost (for the nonperiodic inspection policy). The costs are optimized
with respect to the system parameters.

1.1 Modelling degradation

The system is complex consisting of N components
and its state is an N -dimensional Wiener process
Wt = t + Bt ,

W0 = 0


T
with = [1 , . . . , N ]T , Bt = Bt(1) , . . . , Bt(N )

where Bt(i) is a standard Brownian motion.

The individual processes are not observed and
decisions are based on a performance measure
Rt = Wt 2 , the L2 -norm of Wt . Without loss of generality we assume that = 1. Rt is the radial norm
of a drifting Brownian motion starting at the origin, a
Bessel process Bes0 (, ) starting at the origin with
parameter and drift (Rogers and Pitman 1980)
where
=

1
N 1,
2

= 2

The properties of the Bessel process entail some

changes in the way the model is developed. Because
the radial part of a Brownian motion with drift starting
at x > 0 is not a Bessel process with drift (Rogers and

619

Pitman 1980) we handle repair by adjusting the thresholds. The difficulty is resolved by calculating the
distance remaining between the observed state at repair
and the threshold and represent repair by restarting
the process from the origin (in RN ) and lowering the
threshold to that remaining distance. Extensive treatments of the Bessel process with drift and the radial
Brownian motion are given in (Revuz and Yor 1991;
Pitman and Yor 1981; Rogers and Pitman 1980).

2
2.1

PERIODIC INSPECTIONS
Features of the model

2.1.1 Model assumptions

The model assumes: a) the inspection policy is to
inspect at fixed intervals and are perfect and instantaneous; moreover, the system state is known only at
inspection or failure; b) the system starts from new,
at t = 0 R0 = 0; c) the thresholds are F and < F ;
d) each inspection incurs a fixed cost ci ; e) catastrophic failure are instantaneously revealed as the first
hitting time of the threshold F; f ) the system is instantaneously replaced by a new one at cost Cf ; g) each
maintenance action incurs a cost determined by a cost
function Cr ; h) the transition density for Rt starting
from x is fx ( y) f ( y|x, ).
2.1.2 Settings for the model
The state space in which the process evolves is partitioned into a normal range [0, ), a deteriorated range
[ , { ) and failed [F , ),
R+ = [0, ) [ , F ) [F , +) ,
The decisions are based on the last exit time from
the critical threshold
H0 = sup {Rt | R0 = 0}
tR+

which is not a stopping time. Because H0 is not a

stopping time, we work with the probability of not
returning to the level before the next inspection,
P[H0x ]. The catastrophic failure time is
0
= inf {Rt = F | R0 = 0}
GF
tR+

x
which is a stopping time. The density of GF
is g xF .
Inspection at time t = (immediately before any
maintenance) reveals the systems performance measure R . The level of maintenance (replacement
or imperfect maintenance) is decided according to
0
whether the system has failed GF
or is still work0
. Replacement is determined by the first
ing GF
hitting time of threshold F.

Maintenance actions are modelled using a function

r to specify the amount by which both of the threshold values are decreased. The maintenance function
depends on the probability of return to :

x, P[H0x ] 1 
r(x) =
kx, P[H0x ] > 1 
where 0 <  < 1 and k [0, 1].
Standard models can be recovered:  = 0 corresponds to no maintenance; k = 1 corresponds to
minimal repair (as bad as old); and k = 0 corresponds
to perfect repair (good as new).
The cost function depends on the amount, r, by
which the threshold values are decreased

0,
P[H0r(x) ] 1 
Cr (x) =
Crep , P[H0r(x) ] > 1 
The transience of the Bessel process implies that Cr
is well defined,  (0, 1), R+ such that
,
> ,

P[H0x ] 1 
P[H0x ] > 1 

2.1.3 The framework

0
or
There are two possibilities, the system fails GF
0
is still working GF after crossing . At inspection
time t1 and before any maintenance action, the performance measure is Rt1 = x. Maintenance lowers the
threshold values  r(x) and F  F r(x),
so considering the next interval
0
1. GF
r(x) > : the system survives until the next
planned inspection in units of time. The next
inspection is at t1 + with cost Ci . The cost of repair
at this next inspection is Cr (r(x)). The performance measure at time t1 + is R0 and determines
the reduction in the thresholds.
0
2. GF
r(x) : the performance measure hits the
threshold F r(x) before the inspection at t1 + .
The system fails and is instantaneously replaced
with cost of failure Cf . These failure times form a
renewal process.

Each cycle consists of a sequence of occurrences of

case 1 and ends with case 2 as the system fails and is
replaced.
2.2 Optimal periodic inspection policy
2.2.1 Expected cost per cycle
If R = x at time an inspection prior to any maintenance, we set R + = x and the threshold values
adjusted to F r(x) and r(x) at + just after
the action. A recursive argument yields an analytical

620

expression for the cost of inspection and maintenance

per cycle. The cost per cycle is Vx given that at R = x.
Vx = Cf 1{G0

F (x)

Cx =


R0 
+ Ci + Cr (x) + V 1{G0

F (x)

> }

R0

A = E Cf 1{G0

F r(x) }

(1)

= Cf
0

0
gF
r(x) ( y)dy



R0 
B = E Ci + Cr (x) + V 1{G0
>
}
F r(x)



0
y)
dy
gF
= {Ci + Cr (x)} 1
r(x) (

0
gF
r(x) ( y) dy

F r(x)

vy f0 ( y) dy


vx = Q(x) + (x)

We restructure the expected cost as


vx

F r(x)

= Q(x) + (x)
0

2.2.4 Obtaining solutions

0
The density, gF
, for the first hitting time of a Bessel
process with drift is known only through its Laplace
transform (Pitman & Yor 1981; Yin 1999). The
transform is for > 0



1 2 0
2 + 2
I (F)
E[e 2 GF ] =


I (F 2 + 2 )
Solutions to (2) were obtained by performing numerical inversions of the Laplace transform using the
EULER method (Abate & Whitt 1995).
The Volterra equations (2) and (4) are reformulated
as Fredholm equations



+ 1

vx
lx

with expressions for vx , lx given in (1) and (4)

respectively.

where V is the future cost restarting from the

renewed state 0.
Taking the expectation:
vx = E[Vx ] = A + B

2.2.3 Expected cost per unit time

A standard renewal reward argument gives the cost per
unit time.


vy f0

( y) dy

lx

(2)

= P(x) + (x)
0

K {x, y} vy dy
K {x, y} ly dy

with Q, as in (3), P as in (5) and

with

(x) = 1
0

K {x, y} = 1{yF r(x)} f0 ( y)

0
gF
r(x) ( y) dy

Q(x) = (1 (x)) Cf + (x) {Ci + Cr (r(x))}

(3)

2.2.2 Expected length of a cycle

The expected length of a cycle, lx , is obtained similarly.
The length of a cycle Lx is
0
Lx = GF
r(x) 1{G 0

F r(x) }


R0 
+ + L 1{G0

F r(x) > }

R0

where L is the length of a cycle restarting in state 0.

The expected value is

lx = P(x) + (x)

F r(x)

ly f0 ( y) dy

with defined in (3) and


0
P(x) =
ygF
r(x) ( y) dy + (x)
0

They are solved numerically using the Nystrom

routine with an N point Gauss-Legendre rule. For
xi ( , F], r (xi ) is not defined since xi < 0.
For such values we take r (xi ) = kxi , i.e. repair is
undertaken on the system. If r (xi ) < 0, a cost of
repair is automatically included at the next inspection
time. This seems to be a reasonable assumption since
Rt is positive and will therefore always stay above such
a threshold with negative value, meaning that the last
exit time has already happened and hence that repair
must be considered.
The optimal period of inspection and repair threshold can then be determined as:
( , ) = argmin( )R+ [0,F ] {C0 }

(4)
3

NON-PERIODIC INSPECTIONS

3.1 Features of the model

(5)

The extension to non-periodic inspection policies

shares many features with the periodic policy

621

described in 2.1. The complexities of a dynamic

programming formulation are avoided by introducing
a scheduling function = m(x) which determines the
time to the next inspection based on the observed
system state x. The scheduling function develops the
sequence of inspections in the following way: an
inspection at i reveals Ri = x, the repair is r(x) and
the next inspection is scheduled at m(r(x)).
Different inspection policies are obtained through
the use of three scheduling functions m1 , m2 and m3
modelled on the approach in (Grall et al. 2002).


a1
m1 [x | a, b] = max 1, a
x
b

2
(x b)
(a 1) + 1, 0  x  b
2
m2 [x | a, b] =
b
1,
x > b.

a1

x + a, 0  x  b
m3 [x | a, b] =
b

1,
x>b
All the functions decrease from a to 1 on the interval
[0, b] and then remain constant at 1. The scheduling
function m1 decays linearly; the function m2 is concave, initially steeply declining and then slowing; the
function m3 is convex, initially slowly declining and
then more rapidly. The different shapes reflect different attitudes to repair, m3 tends to give longer intervals
initially and m2 gives shorter intervals more rapidly.
To avoid a circular definition maintenance function
r employs the performance measure just before repair

x, P[H0x m(x)] 1 
r(x) =
kx, P[H0x m(x)] > 1 
with 0 <  < 1, k (0, 1].
The next inspection is scheduled at m (r(x)) units
of time with cost Cr (r(x)), where

0,
P[H0r(x) m (r(x))] 1 
Cr (r(x)) =
Crep , P[H0r(x) m (r(x))] > 1 
3.2

Expected total cost

V x = [Cf + V 0 ]1fails given r(x)} +

0

[Ci + Cr (x) + V Rm(r(x)) ]1survives given r(x)}

r(x))} +

F r(x) m(

[Ci + Cr (x) + V

vx = E[V x ] = A + B + C


 m(r(x)) 0
A = Cf + v0
gF r(x) ( y) dy
0



B = {Ci + Cr (x)} 1



R0m(r(x))

]1{G0

r(x))}

F r(x) >m(

m(r(x))
0

m(r(x))

C = 1
0
F

0
gF
r(x) ( y) dy

0
y)
dy

gF
(
r(x)

0
1{yF r(x)} vy fm(r(x))
( y) dy

which may be re-arranged as

 F
vx = Q(x) + (1 (x)) v0 + (x) K {x, y} vy dy
0

with

(x) = 1
0

m(r(x))

0
gF
r(x) ( y) dy

Q(x) = (1 (x)) Cf + (x) {Ci + Cr (r(x))}

0
K{x, y} = 1{yF r(x)} fm(r(x))
( y)

3.3 Obtaining solutions

While (6) contains x and 0 because we need only the
value of 0 we change the equation to

x = Q(x) + (1 (x))vx + (x)

K{x, yt} y dy

(6)
and solve for x as for the periodic model as in (Barker
Newby 2007). The solution is then obtained by setting
x = 0.

The expected total cost in the non-periodic case may

be deduced with the use of the scheduling function.

= [Cf + V 0 ]1{G0

Taking the expectation

NUMERICAL RESULTS

4.1 Periodic inspection policy

The numerical results are based on Bes0 (0.5, 2) as
Rt and a failure threshold F = 12. The choice
for the costs
 and the maintenance functions parameters are Ci , Crep , Cf = (50, 200, 500) and (k, ) =
(0.9, 0.5). They are varied to investigate the behaviour
of the model. The choices were chosen arbitrarily to
show some important features of both the inspection
and maintenance policies.

622

Optimal parameters given k = 0.9,  = 0.5.

Ci , Crep , Cf

(0.5, 200, 500)

(50, 200, 500)
(500, 200, 500)
(50, 2, 500)
(50, 200, 500)
(50, 2000, 500)
(50, 200, 5)
(50, 200, 500)
(50, 200, 5000)

160

( , )

v0

l0

C0

(1.2, 9)
(1.6, 9.5)
(max, any)
(1.8, 7.5)
(1.6, 9.5)
(1.6, 10.5)
(max, any)
(1.6, 9.5)
(1.4, 8.5)

9255.1
6972
500
2640.2
6972
5542.3
5
6972
90295

20781
208.81
5.83
81.77
208.81
163.69
5.83
208.81
2350.5

0.45
33.39
85.89
32.29
33.39
33.86
0.86
33.39
38.42

140
Expected cost per unit time

Table 1.

120
100
80
60

33.39
20

4.1.1 The influence of the costs

The response of the model to costs is examined with
costs Ci {0.5, 50, 500}, Crep {2, 200, 2000} and
Cf {5, 500, 5000}. The optimal period of inspection , repair threshold and expected cost per unit
time C0 are summarized in table 1. The expected cost
and expected length per cycle at the optimum are v0
and l0 .
As Ci increases the optimal expected cost per unit
time and the optimal period of inspection increase.
Increasing Ci makes inspection more expensive resulting in less frequent inspection and reduces l0 because
there will be more failures.
Changing Crep affects the optimal period of inspection and gives higher values for the optimal repair
threshold. The higher threshold reduces the frequency of repairs hence reducing costs. The optimal
strategy is driven by the repair threshold which determines the frequency of maintenance and thus the
optimal expected total cost.
Increasing Cf increases, and . For a low cost
of failure (i.e. Cf << Ci + Crep ) the optimal strategy
is to let the system fail and then replace it resulting in
a lower cost than a repair or a simple inspection.
4.1.2 Investigating the maintenance actions
The maintenance
 9 1  function considered has parameters
(k, ) = 10
,2

x,
P[H0x ] 12
r(x) =
0.9x, P[H0x ] > 12
The effects of parameters , k and  on the model
with (Ci , Crep , Cf ) = (50, 200, 500) are examined.
In Table 1 the optimal parameters are ( , ) =
(1.6, 9.5).
i. The repair threshold: The optimal solution
depends strongly on when other parameters
remain fixed as is shown in figure 1.
ii. Level of repair: The level of repair increases
from perfect repair with k = 0 to minimal repair

5
9.5
Repair threshold:

15

Figure 1. Effect of parameter on C0 with (a, , ) =

(0.9, 0.5, 1.6).

Table 2. Optimal inspection period &

expected cost per unit time for different values
of k ( = 0.5).
k

C0

0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

3.4
3.2
3.0
2.8
2.4
2.4
2.0
2.0
1.6
1.6
1.4

15.72
17.04
18.80
20.37
23.08
25.18
28.59
30.02
32.78
33.39
36.21

with k = 1. Table 2 shows the uniform effect of the

repair parameter. The cycle length decreases and
the cost increases as k increases.
iii. Repair Parameter: Repair is determined by the
parameter  and the probability
P[H0x ] .
The different values for , 0.1, 0.5 and 0.9 reflect
the decision makers attitude towards repair. Values
close to 1 corresponds almost certain repair and as the
value decreases to 0 repair occurs less frequently, a
riskier position. The results in table 3 show that only
the threshold responds to .
The model adapts itself to the decision makers attitudes to repair (the value of ) by moving the optimal
repair thresholds. As  increases repairs will be considered more often but increases to restrain the
frequency of repairs. The optimal expected cost per

623

Table 3.

Optimal parameters for different .

0.1

C0

8
1.6
33.39

0.5

0.9

9.5
1.6
33.39

11
1.6
33.39

Table 4.
b and .
Repair
threshold

=1
=2
=3
=4
=5

Optimal expected total cost and parameters a,

Scheduling
function

m1
m2
m3
m1
m2
m3
m1
m2
m3
m1
m2
m3
m1
m2
m3

2.2
2.1
2.1
2.2
2.2
2.1
2.4
2.5
2.4
5.2
5.2
5.2
6.5
6.5
6.5

1.5
4.2
0.9
1.7
2.9
1
2.5
2.8
1
3.7
3.8
1.9
0.5
0.7
0.5

1171.7
1169.5
1170.4
1189.1
1194
1189.9
1546.3
1572.1
1547.8
2.3283 105
2.34 105
2.3264 105
3.8437 106
3.8437 106
3.8437 106

a more complex dynamic programming problem to

determine a policy = {1 , 2 , . . . , n , . . . }.
Figure 2. Effect of parameter  on the optimal solution C0
with parameters (k, , ) = (0.9, 9.5, 1.6).

unit time remains constant in the three cases studied. Figure 2 clearly shows that this is not the case
for inspection periods ( , =0.1 ], where =0.1
satisfies


t > =0.1 : P H0 < t > 1 0.1
For most values in this interval, the expected cost per
unit time increases with : the model penalizes a costly
strategy that favors too many repairs. For a period
of inspection greater than =0.1 , the expected costs
per unit time are identical since in all three cases the
approach towards repair is similar: the system will be
repaired with certainty




P H0 < t > 0.9 P H0 < t > 0.5


P H0 < t > 0.1 .
4.2

Non-periodic inspection policy

The results are obtained using Bes0 (0.5, 2) for Rt and

with fF = 5. The different costs and the maintenance functions parameters are (Ci , Crep , Cf ) =
(50, 100, 200), (k, ) = (0.1, 0.5).
The optimal solution is determined by the optimal
parameter values (a , b ) rather than the solution of

4.2.1 The optimal maintenance policy

Table 4 reports the optimal solutions for a range of
thresholds and for the different scheduling functions.
An optimal solution was found in each case. The solutions are not particularly sensitive to (a, b), but range
over several orders of magnitude as the threshold
varies.
4.2.2 The influence of costs
We take an example with scheduling function m1 and
= 3. The optimal parameters (a , b ) and total cost
are summarized in table 4. As Ci increases, the optimal values of a and b increase making inspection less
frequent when the cost of inspection increases.
4.2.3 Investigating the maintenance actions
i. Level of repair: The optimal cost for the
three optimal inspection scheduling functions and
k [0, 1] and repair threshold = 3 are shown in
Figure 3. In all three cases the expected total cost
increases with k implying a reduction in the amount
of maintenance undertaken on the system at each
repair. The system will therefore require more frequent repairs or will fail sooner implying an increase
in the total expected cost value.
ii. Attitude to repair: The attitude of the decision maker towards repair is reflected in the
parameter  [0, 1]. The optimal expected costs
obtained with corresponding optimal parameters
with  = 0.1, 0.5, 0.9 and = 3 are summarized in
table 6. Letting  approach zero means that the

624

decision maker tends to a safer maintenance

approach. Changes in  induce changes in the optimal inspection policy and the resulting optimal
expected total cost (table 6).
Table 5. Optimal expected total cost and parameters (a, b)
for different values of the maintenance costs, = 3.


b
v
Ci , Crep , Cf
a
(5, 100, 200)
(50, 100, 200)
(500, 100, 200)
(50, 1, 200)
(50, 100, 200)
(50, 1000, 200)
(50, 100, 2)
(50, 100, 200)
(50, 100, 2000)

2.4
2.4
2.5
2.4
2.4
2.6
3.2
2.4
2.4

2.4
2.5
2.6
2.3
2.5
2.5
2.9
2.5
2.2

1467.6
1546.3
2296.7
1377.3
1546.3
2971.9
203.67
1546.3
13134

5500
Inspection strategy considered: m1
Inspection strategy considered: m
2
Inspection strategy considered: m

5000

4500

Expected total cost

4000

SUMMARY

The aim of the models derived and investigated in

the present paper extend an earlier paper (Barker &
Newby 2006) by incorporating catastrophic failure
of the system. Catastrophic failure is represented by
introducing a second threshold F to account for catastrophic failure. The threshold is incorporated in the
repair function r as the last exit time from the interval
[0, ). The repair decision depends on the probability of occurrence of this last exit time before the
next inspection. The models proposed hence include
both a stopping time (the first hitting time) and a
non-stopping time (the last exit time). The probability density function of the first hitting time for a
Bessel process with draft being not known explicitly,
the expression for the expected total cost was solved
numerically (a numerical inversion of the Laplace
transform of the first hitting times density function
was required).
The numerical results revealed a strong influence
of the thresholds value and parameter k on both the
optimal period of inspection and the optimal expected
cost per unit time. Letting parameter  vary produced
changes in the optimal repair threshold only, suggesting that the optimal strategy aims at keeping a
relatively constant frequency of repairs.

3500

3000

REFERENCES
2500

2000

1500

1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Figure 3. Optimal expected total cost as a function of k for

the three inspection strategies.
Table 6. Optimal expected total cost and parameters (a, b)
for different values of , = 3.
Repair
parameter

Inspection
policy

 = 0.1

m1
m2
m3

2.3
2.3
2.3

2.4
3.1
2.1

1535
1622.7
1560.3

 = 0.5

m1
m2
m3

2.4
2.5
2.4

2.5
2.8
1

1546.3
1572.1
1547.8

 = 0.9

m1
m2
m3

2.1
2.1
2.1

2.2
4.1
0.9

1169.1
1169.1
1169.9

Abate, J. and W. Whitt (1995). Numerical inversion of laplace

transforms of probability distribution. ORSA Journal on
Computing 7(1), 3643.
Barker, C.T. and M.J. Newby (2006). Optimal inspection
and maintenance to meet prescribed performance standards for a multivariate degradation model. In C. G. S. .
E. Zio (Ed.), Safety And Reliability For Managing Risk,
Volume 1, pp. 475487. European Safety and Reliability
Association: Taylor & Francis.
Barker, C.T. and M.J. Newby (2007). Optimal nonperiodic inspections for a multivariate degradation model.
to appear Reliability Engineering and System Safety,
DOI:10.1016/j.ress.2007.03.015.
Grall, A., L. Dieulle, C. Berenguer, and M. Roussignol (2002,
June). Continuous-time predictive-maintenance scheduling for a deteriorating system. IEEE Transactions on
Reliability 51, 141150.
Pitman, J. W. and M. Yor (1981). Bessel process and infinitely
divisible laws, stochastic integrals. Lecture Notes in
Mathematics, Springer Berlin.
Revuz, D. and M. Yor (1991). Continuous Martingale and
Brownian Motion. Springer Verlag.
Rogers, L. C. G. and J. W. Pitman (1980). Markov functions.
Annals of Probability 9, 573582.
Yin, C. (1999). The joint distribution of the hitting time and
place to a sphere or spherical shell for brownian motion
with drift. Statistics and Probability Letters 42, 367373.

625

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Preventive maintenance planning using prior expert knowledge

and multicriteria method PROMETHEE III
F.A. Figueiredo, C.A.V. Cavalcante & A.T. de Almeida
Federal University of Pernambuco, Brazil

ABSTRACT: Preventive maintenance planning is one of the most common and significant problems faced by
the industry. It consists of a set of technical, administrative and management actions to decrease the component
ages in order to improve the systems availability. There has been a great amount of research in this area
using different methods for finding an optimal maintenance schedule. This paper proposes a decision model,
integrating Bayesian approach with a multicriteria decision method (MCDM) based on PROMETHEE. This
model determines the best solution for the preventive maintenance policy taking both costs and availability as
objectives and considering prior expert knowledge regarding the reliability of the systems. Several building
considerations regarding the model are presented in order to justify the procedure and model proposed. Finally,
a numerical application to illustrate the use of the model was carried out.
1

INTRODUCTION

Maintenance seeks to optimize the use of assets during their life cycle, which means either preserving
them or preserving their ability to produce something safely and economically. In recent decades,
maintenance management has evolved and, nowadays,
it is not just a simple activity, but a complex and
important function. Its importance is due to the great
number of operational and support costs incurred on
systems and equipment. The United States, for example, used to spend approximately 300 billion dollars
every year on the maintenance of plant and operations
(Levitt, 1997).
Since maintenance planning and development are
crucial, their best practice is important for any industry. Maintenance planning can be defined according
to three types of maintenance. One type is corrective maintenance, which is performed after a failure occurs. Other types are preventive maintenance
and predictive maintenance which are considered to
be proactive strategies, because they are performed
before any failure has occurred and act in order to
reduce the probability of a failure.
Preventive maintenance (PM) is deemed as actions
that occur before the quality or quantity of equipment deteriorates. These actions include repair or
replacement of items and are performed to extend
the equipments life, thus maximizing its availability.
Predictive maintenance, on the other hand, concerns
planned actions that verify the current state of pieces of
equipment in order to detect failures and to avoid system breakdown, thus also maximizing the availability

of equipment. However, in order to create a PM program, it is necessary to determine how often and when
to perform PM activities (Tsai et al., 2001).
In this paper, we apply a multicriteria method
to solve the problem of preventive maintenance and
take into consideration costs, downtime and reliability. The following section, Section 2, contains an
overview of the PM problem and discusses previous
PM scheduling work. Section 3 presents explanations
of the PROMETHEE method and describes its procedures. In Section 4, the model applied is presented
along with the results and discussions. Finally, the conclusions are summarized and possible future studies
are suggested in Section 5.

2
2.1

THE PROBLEM
Preventive maintenance problem

This kind of maintenance is not known to be the dominant practice, because people resist spending money
on efforts to forestall failures that might not happen. There are two situations where PM is important:
when it reduces the probability or the risk of death,
injuries and environmental damage or when the task
costs are smaller than the cost of the consequences
(Levitt, 2003).
Of the many reasons to implement a PM policy in an
industry, the greatest is the improvement of a systems
availability. But that is not the only one. There are
many others such as: increasing automation, minimizing administrative losses due to delays in production,

627

reducing stocks of spares and equipment redundancy

reduction, minimizing energy consumption etc (Quan
et al., 2007). Many advantages can be gained from a
well developed PM program, although they depend on
the size and type of plant. The greater ones asset values per square meter are, the greater will be the return
on the PM program (Worsham, 2000).
The PM program will determine what maintenance
activities should be performed on which items of each
piece of equipment, what resources are needed and
how often the activities should be scheduled. It will
only be effective if the tasks demanded for the system
are scheduled at regular intervals (Quan et al., 2007).
Production must be considered while planning
maintenance, because production schedules may be
interrupted by failures or maintenance tasks. Thus
it is necessary to balance how many interventions
will be made to production in order to make as few
interruptions as possible to perform PM and still maintain a healthy system with few repairs or corrective
maintenance needs (Sortrakul et al., 2005).
Numerous studies have been conducted concerning
the problem of preventive maintenance. Bevilacqua
and Braglia (2000) used AHP for the selection of the
best maintenance strategies for an Italian oil refinery.
Wang et al., (2007) extended Bevilacqua and Braglias
studies by applying a fuzzy analytic hierarchy process to deal with decision makers uncertainties as
to their judgment. Samrout et al., (2005) applied a
heuristic method, called ant colony optimization to
a system of series-parallel aiming to minimize the
cost of preventive maintenance. A few studies tried
to solve the preventive maintenance problem by combining genetic algorithms and multicriteria, such as
Lapa et al., (2006), who used reliability and costs
as objectives. Chareonsuk et al., (1997) presented a
PROMETHEE II model to determine the optimal interval. However, in their model, the criteria were limited
to an interval of values for both criteria, thus excluding
the conflict regions. Cavalcante and Almeida (2005,
2007) also applied the PROMETHE method in different papers. In one of them, a PROMETHEE II
methodology was applied, but in addition to not considering any limitation for the criteria values, the
lack or unreliability of data was also considered. In
the other paper, PROMETHEE III was used, which
permitted more rational planning for the problem of
preventive maintenance and also took into account
external uncertainties.
In this paper, we propose an application based on the
early model (Cavalcante & Almeida, 2007), integrating a Bayesian approach with a multi-criteria decision
method (MCDM) based on PROMETHEE, but, unlike
the former, we consider some different aspects regarding uncertainties about life distribution parameters. In
addition, we higtilight some important particularities
of service systems, which require some changes on

the decision-making processes related to preventive

maintenance.
In summary, this paper proposes a multi-criteria
model in order to establish the age t of replacement,
taking into account not only the cost, as in classical
models, but also reliability, as well as how the problem
of data is overcome by means of a Bayesian integrated
approach.
2.2 The service context
Mont (2001) has observed that the service sector
has undergone significant growth with strong repercussions for the economy. This has caused specific
changes in production systems, which increasingly
depend on service for their processes, and which have
resulted in the increasing flexibility of the production
systems themselves.
According to Hirschl et al., (2003), there has to
be a societal change in consumer behavior. Thus, the
idea of keeping a particular piece of equipment for
the longest time possible is now an idea that may be
imputed to consumer mentality. On the producer side,
this means the longer a particular product is used, the
more service work will be needed over the life-cycle
of the product.
The sustainable potential of a system of goods and
services does not only depend on this holistic vision
which embraces diverse criteria to define operational
procedures, but is also based on the structure of new
kinds of relationships or partnerships between the
actors of the process, leading to new convergences of
economic interests and a potential optimization of the
use of resources. Therefore the notion of a product is
no longer just the physical result of a productive process. Rather, it is an integrated concept where goods
and services are mutually dependent; focusing on consumer needs and demands in an integrated way and at
the same time increasing profit and diminishing the
environmental impact by reducing the volume of goods
manufactured (Manzini et al., 2003).
Given the importance of service production systems
and their growth potential for participating in the economy as well as the ever-stronger tendency towards the
formation of chains, the distinctive features of service
production systems are more than enough to justify the
need for a serious study related to planning for their
maintenance.

THE DECISION MODEL

3.1 Multicriteria decision aid

According to Gomes et al., (2002), several methods for
approaching complex processes of decision making
have been developed over many years. Abstractions,

628

heuristics and deductive reasoning were developed,

but only using multi-criteria or multi-objective methods in which the decision makers preference structure
was most loyally represented.
Multi-criteria methods have been developed to support and lead the decision maker in evaluating and
choosing the best solution. The problems analyzed
by the method can be discrete, when it has a finite
number of alternatives, or continuous, when it has
an infinite set of alternatives. It is very important
always to remember that these methods do not indicate the solution of the problem, but, instead they
support the decision process through alternatives or
recommendations for actions (Gomes et al., 2002).
There are two kinds of multi-criteria methods:
compensatory and non-compensatory methods. Compensatory methods represent the American School and
aggregate all the criteria in a unique utility function,
which brings the notion of compensation between the
criteria. Non-compensatory methods do not aggregate
the criteria and for each criterion a weight is defined
that represents its relative importance in the problem.
The best known methods of this type are the families ELECTRE and PROMETHEE that represent the
French School (Almeida et al., 2002).

3.2

The PROMETHEE method

The PROMETHEE (Preference Ranking Organization Method for Enrichment Evaluation) conceived
by Brans consists of two phases: the construction of
an outranking relation and then the exploitation of
outranking values relations (Almeida et al., 2003).
These French school methods, which include the
PROMETHEE family, like the American school, are
used in multi-criteria problems of the type:
max {f1 (x), f2 (x), . . . , fk (x)|x A} ,

Depending on the pre-defined preference function

chosen, it may be necessary to define certain parameters such as q, p and s. According to Brans (1985) the q
indifference threshold is the largest deviation which is
considered as negligible by the decision maker, while
the p preference threshold is the smallest deviation
which is considered as sufficient to generate a full
preference. The parameter s is used in the case of the
Gaussian criterion and it defines the inflection point
of the preference function. It is recommended that a p
and q be defined first, in order to choose the s value
between the interval.
The decision maker, then, decides a weight wj for
each criterion that increases with the importance of the
criterion, and then the weights are normalized to sum
up to unity (Brans et al., 2002).
To define the preference function Pj (a, b) = dj (a,
b) = fj (a)fj (b), that assumes values between 0 and 1,
what is evaluated is how the preference of the decision
maker changes with the difference between the performance of alternatives for each criterion j(Almeida
et al., 2002). The behavior can be represented below:
Pj (a, b) = 0: Preference for a or b
is indifferent,
Pj (a, b) 0: Mild preference for a
in relation to b,
Pj (a, b) 1: Strong preference for a
or b is indifferent,
Pj (a, b) = 1: Strict preference for a
in relation to b .

Once the generalized criteria are defined for each

criterion of the problem, then the calculation of the
multi-criteria preference index or aggregate ranking
(4) of every pair of alternatives must be established
(Almeida et al., 2002).

(1)

where A is a set of decision alternatives and fi (x),

i = 1, . . . , k is a set of criteria of which each alternative is to be evaluated. Each criterion has its own unit
(Chareonsuk et al., 1997).
The application of the PROMETHEE method consists of three steps. First of all, there is the definition
of the generalized criteria, then the calculation of the
multi-criteria index, followed by the determination and
evaluation of an outranking relation (Dulmin et al.,
2003).
The preference behavior of the decision maker
will determine a function Pj (a, b) that assumes values
between zero and one, where a and b are alternatives.
There are six generalized criteria to choose from, when
it comes to defining the preference function (Brans,
1985).

(2)

(a, b) =

where:

n
1 
wj Pj (a, b)
W j=1

W =

n


(3)
wj

j=1

And finally, the next step is carried out, which

involves the determination and evaluation of the outranking relations.
For each alternative a A, two outranking flows
are determined with respect to all the other alternatives
x A.
The leaving flow or positive outranking flow (4)
expresses how an alternative a is outranking all the

629

others, so the higher + (a), the better the alternative

(Brans, 1985).
+ (a) =

 (a, b)
bA

n1

(4)

The entering flow or negative outranking flow (4)

expresses how an alternative a is outranked by all the
others, so the lower  (a), the better the alternative
(Brans, 1985).
 (a) =

 (b, a)
bA

n1

(5)

The complete ranking, which is often requested by the

decision maker , considers the net ranking flow (6).
This particular flow is the balance between the positive
and negative outranking flows. Thus the higher the
(a), the better the alternative (Brans, 1985).
(a) = + (a)  (a)

(6)

After finding the net flow (6), it is possible to apply

the PROMETHEE II method, which results in a total
pre-order, as follows (Brans, 1985):
aPb(a outranks b) iff(a) > (b),
aIb(a is indifferent to b) iff(a) = (b).

(7)

PROMETHEE III associates an interval [xa , ya ] with

each action a, as follows (Brans and Mareschal, 2002):

a,
xa = (a)

ya = (a)
+ a,
where

n is the number

 of actions,

1 

(a)
=
(a, b) (b, a) = 1n (a)

n


2

bA

a2 = 1n
(a, b) (b, a) (a)
,

bA

> 0.

aI III b(a outranks b)iff xa yb and xb ya

3.3 Decision model structure

3.3.1 Preventive maintenance policy in the service
context
Concerning the question of failure control, in particular the breakdown of equipment parts, while the
paradigm of optimization is well applied to the context of the production of goods, it is not efficient in the
context of services since the quality of service is intrinsically related to the perception of the consumer, for
whom the translation into monetary value has little or
no meaning. On the other hand, cost, in a competitive
environment, is of great importance when decisions
are to be made, and cannot be ignored.
As a result, the choice of a preventive maintenance
policy, which in practice refers to the intervals or
frequency of component replacement, represents a balance between the risk of having a breakdown, keeping
in mind the negative perception that this can cause,
and the cost related to the policy adopted. Optimization, then, no longer makes sense, as there is more than
one criterion to be observed.
What is needed then is a more appropriate
approach, for example, multi-criteria decision-aid,
capable of dealing with conflicts among the criteria
involved in the problem as well as the inclusion in
the decision-making process of the decision-makers
preferences.

(8)

Then the method defines a complete interval order,

as represented below:
aP III b (a outranks b) iff xa > yb

many advantages, such as the ease and speed at which

decision makers understand the method itself and the
concepts and parameters involved, since they represent
a more meaningful significance for them to relate to,
such as physical and economical values (Cavalcante
et al., 2005).

(9)

For the problem presented, the use of a noncompensatory method was indicated as the most
suitable solution. Therefore this method favors the
alternatives with the best average performance. The
PROMETHEE method was chosen because of its

3.3.2 Bayesian analyses

Bayesian analyses are used in cases where there
are insufficient data to directly obtain distribution
parameters of the model (Cavalcante et al., 2007).
Accordingly to Procaccia et al., (1997), the feedback from experience can yield objective factual data.
These data allow the estimation of failure rates in operation or probabilities of starting on demand, when
the equipment is initially sound. So the combination of subjective and objective data, prior probability
function and likelihood provided by feedback from
experience, respectively, helps to obtain a posterior
probability density that is richer.
It is extremely important to know that questioning experts can produce much information since they
have different experiences and are even from different departments or industries, so their sensitivity to
a specific problem will not necessarily be the same
(Procaccia et al., 1997).

630

Earlier literature often takes the classical approach

or the combined classical Bayesian approach that presumes true parameters such as average costs per unit
time and failure rates (Apeland et al., 2003).
In the classical approach it is assumed that true
parameters exist, like failure rates related to component types, or means and standard deviations of
failure costs. The presumed true values of the parameters are unknown, and estimating them is only based
on historical hard data. The focus of a combined
classical Bayesian approach is also on the presumed
true probabilities and probability distributions; however, the meaning of uncertainty is different. In this
case, the analyst group is uncertain of the true values and based on available hard data, suitable models
and engineering judgment, they elaborate a distribution that expresses the uncertainty (Apeland & Scarf,
2003).
In the fully subjective approach, the focus, unlike
the other approaches, is on observable quantities, like
the number of failures or production loss over a period
of time, and probabilities and probability distributions are used to describe uncertainty related to these
quantities (Apeland & Scarf, 2003).
In this paper, the model presented uses the combined classical Bayesian approach, because the data
assumed a distribution, the parameters of which are
unknown, and for that matter the uncertainty is related
to the distributions parameters.
Therefore, these parameters will be determined on
evidence of prior expert knowledge based on previous experiences. When a model is based on a priori
knowledge of specialists, it means that it is based on
the hypothesis that a specialist has a reasonable idea
about the subjective probability distribution of a variable , which represents the state of nature and can
have a random or non-random nature (Cavalcante &
Almeida, 2005).
In the Weibull distribution, the lack of data will
not allow the knowledge of both parameters and
instead of that we can have uncertainty about one or
both of them. In this way this(these) unknown parameter(s) is (are) considered random variable(s) and after
elicitation from a specialist it(they) will be considered
as priori distribution(s), () and/or () (Cavalcante
& Almeida, 2007).

3.3.3 Criteria
A Weibull distribution was assumed for equipment
which deteriorates with time. Besides being highly
used, it is also flexible and therefore is proven to be a
good fit in many cases.
The Weibull distribution has two parameters in this
model: , shape parameter and scale parameter
respectively. The probability density function (10) and

the reliability function (11) obtained are:

f (t) =

R(t) = e

 1  t  

t
e

(10)

(11)

For the preventive maintenance problem tackled, a

replacement policy was considered, because it guarantees lower costs in replacements. There are a great
number of models proposed varying in cost, functional
performance and importance in the process. For every
one of those models several criteria can be identified
to describe the preventive maintenance activities.
The replacement policy used is replacement with
age that replaces an item at the moment it fails or periodically, whichever comes first. This policy is only
effective when the replacement costs before the failure are smaller than after the failure. Therefore, the
objective is to find that interval, t, that represents the
smallest costs per unit time.
The costs involved are ca, replacement cost after
failure, and cb, replacement costs before the failure.
To express the expected cost of using the replacement
policy (12), in a mathematical expression is necessary
to associate costs and time, as shown below:
Cm(t) =

ca(1 R(t)) + cbR(t)

t
xf (x)dx + tR(t)

(12)

Since in real situations this variable is not continuous,

a set of possible solutions or alternatives is given in the
numerical application to find the best solution between
the set.
Where
R(t) is the reliability for the time t;
f (x) is the probability density function for time to
failure.

NUMERICAL APPLICATION

The criteria have already been described and the uncertainties have been discussed. The simulation presented
in this section was based on data taken from the context of electrical energy for devices and was found to
be highly adequate to preventive replacement.
For the replacement policy it was necessary to determine some parameters, as shown in Table 1. After the
determination of the parameters or equipment characteristics, the alternatives for the simulation are defined
(see Table 2). The alternatives determined are not chosen randomly, but according to opportune times that
allow preventive maintenance

631

Table 1.

Input data.

Table 4.

Decision evaluation matrix.

Distribution

Parameters

Value

Alternatives

R(t)

Cm(t)

Weibull

ca
cb

2,8
Unknown
10
2

A1
A2
A3
A4
A5
A6
A7
A8
A9
A10

0.986
0.941
0.877
0.8
0.717
0.631
0.548
0.469
0.398
0.334

2.117
1.244
1.025
0.958
0.945
0.954
0.971
0.992
1.013
1.033

Table 2.

Set of actions.

Alternatives

Time
years

A1
A2
A3
A4
A5
A6
A7
A8
A9
A10

1
2
3
4
5
6
7
8
9
10

Table 5. Preference function and criteria

characteristics.
Characteristics

Cm

Max/min
Weights
Preference function
Indifference limit
Preference limit

Max
0.55
V
0.02
0.05
0.23

Min
0.45
V
0.08
0.5

Table 3. Weibull distribution parameters

for unknown .
Table 6.

Weibull distribution parameters

Neta2
Beta2

11
2.34

Since the multi-criteria model proposed assumes

uncertainty about the parameter of the Weibull distribution of life time, the next step must be the elicitation
for the parameters unknown by applying the Bayesian
analyses, in order to obtain the a priori distributions
of the parameter unknown. The parameters of the distribution that deal with the uncertainty of , , are
presented in Table 3.
After that it is possible to calculate the performance
of the alternatives for each criterion according to their
expressions presented in section 3.3.3. The results
obtained are shown in the decision evaluation matrix
(Table 4).
The generalized criteria for the pairwise comparison must be determined, thus come the weights
for each criterion and their parameters depending
on the choice of generalized criteria. Since the
PROMETHEE III method is used, parameter must
also be defined. The values are presented in Table 5.
Once all the parameters are defined and the decision
evaluation matrix obtained, the aggregation process

Ranking of alternatives.

Alternatives

Time
years

+

A1
A2
A3
A4
A5
A6
A7
A8
A9
A10

1
2
3
4
5
6
7
8
9
10

0.54
0.54
0.49
0.44
0.38
0.32
0.26
0.19
0.13
0.07

0.45
0.22
0.12
0.18
0.24
0.31
0.37
0.43
0.49
0.55

0.09
0.31
0.37
0.26
0.14
0.01
0.11
0.24
0.36
0.48

starts. In this process the alternatives are evaluated in

respect of all criteria and the decision-makers preference. As explained in section 3.2, the PROMETHEE
method will determine the multi-criteria preference
index, the input and output flow and finally the net
flow (see Table 6).
Based on these results we can obtain the partial
ranking by PROMETHEE III. The ranking can be
visualized in the illustration in Figure 1.
Thus, the multi-criteria model applied resulted in
the ranking of the alternatives as follows, ordered from

632

0,4

A3 > A2 > A4 > A5

0,2
Net flow

the best to worst:

= A1 > A6 > A7 > A8 > A9 > A10

Therefore, the (A3) is the best alternative since
it is first in the ranking and it represents that the
replacement of the equipment should take place every
(A3)Years.
Besides, we can highlight that some alternatives
were considered indifferent, since their general performance were judged almost the same. This results from
enlarging the indifference notion, which mirrors more
suitably the behavior of the decision maker when analyzing similar alternatives, especially if it was admitted
that there are uncertainties about some criteria of the
problem.
In order to verify the robustness of some parameters of the model, a sensitivity analysis was carried out. Through this analysis, it is possible to
obtain more interpretative results which enhance the
decision-makers understanding of the maintenance
problem.
Therefore, some variations were made up to +/
15% on the values of the parameters of weights. As a
result, in spite of the fact that there was change in the
ranking, the first three alternatives remain the same.
Regarding the parameters called Preference limits
and indifference limits for both criteria, there were no
changes to the best alternative for variations of up to
+/ 15% on these parameters. The most significant
change was for the variation on 15% on the preference limit for the Reliability criterion. In this case, one
tie was introduced, as we can see:
A3 > A2 = A4 > A5

0,2
0,4
0,6
A1 A2 A3 A4 A5 A6 A7 A8 A9 A10
Alternatives

Figure 1.

Ranking of alternatives based on the net flow.

in some intervals. The multi-criteria method applied,

PROMETHE III, allows not only a better understanding of the parameters, concepts and the method by
the decision makers, but, also an amplification of the
notion of indifference between the alternatives. In this
way, for practical purposes, where not merely one criterion is important for the establishment of preventive
periods, for example, for the service context, the decision maker can make use of this model, even if some
problems with data have been detected.
Future studies suggested include the application of
the model to enrich some other problems in maintenance context, where some decision criteria beyond
the cost are very important for the decision-making
process.

ACKNOWLEDGEMENTS
This work has received financial support from CNPq
(the Brazilian Research Council).

= A1 > A6 > A7 > A8 > A9 > A10

In summary, even with substantial changes to the
parameters that form part of the decision model,
the best alternative remains the same and no critical
change was observed to the rankings of alternatives.
5

CONCLUSIONS

In this study, a MCDM model is put forward in order

to determine the frequency and periods at which the
preventive maintenance for a specific item should be
performed in a service maintenance context in the face
of a lack of failure data or external uncertainties.
Based on the results of the numeral application, we
can conclude that the decision model structure was
used effectively, providing a result that can support the
decision as to replacement periods, considering the criteria involved, which can be conflicting with each other

REFERENCES
Almeida, A.T. de & Costa, A.P.C.S. 2003. Aplicaes com
Mtodos Multicritrio de Apoio a Deciso. Recife: Ed.
Universitria.
Apeland, S. & Scarf, P.A. 2003. A fully subjective approach
to modeling inspection maintenance. European Jounral
of Operational Research: p. 410425.
Bevilacqua, M. & Braglia, M. 2000. The analytical hierarchy process applied to maintenance strategy selection.
Reliability Engineering & System Safety: p. 7183.
Brans, J.P. & Vincke, 1985. A Preference Ranking Organization Method: (The PROMETHEE Method for Multiple
Criteria Decision-Making). Management Science Vol. 31,
No. 6: p.647656.
Brans, J.P. & Mareschal, B. 2002. Promethee-Gaia, une
Methodologie dAide la Dcision em Prsence de
Critres Multiples. Bruxelles: ditions Ellipses.

633

Cavalcante, C.A.V. & Almeida, A.T.de. 2005. Modelo

multicritrio de apoio a deciso para o planejamento
de manuteno preventiva utilizando PROMETHEE
II em situaes de incerteza. Pesquisa Operacional:
p. 279296.
Cavalcante, C.A.V. & Almeida, A.T. de. 2007. A multicriteria decision-aiding model using PROMETHEE III for
preventive maintenance planning under uncertain conditions. Journal of Quality in Maintenance Engineering:
p. 385397.
Chareonsuk, C. & Nagarur, N. & Tabucanon, M.T. 1997.
A multicriteria approach to the selection of preventive
maintenance intervals. International Journal of Production Economics: p. 5564.
Dulmin, R. & Mininno, V. 2004. Standardized project management may increase development project success. International Journal of Project Management.
Gomes, L.F.A.M. & Gomes, C.F.S. & Almeida, A.T.de. 2002.
Tomada de Deciso Gerencial. So Paulo: Editora Atlas.
Hirschl, B. & Konrad, W. & Scholl, G. 2003. New concepts
in product use for sustainable consumption. Journal of
Cleaner Production: p. 873881.
Lapa, C.M.F. & Pereira, C.M.N.A. & Barros, M.P.de.
2006. A model for preventive maintenance planning by
genetic algorithms based in cost and reliability. Reliability
Engineering & System Safety: p. 233240.
Levitt, J. 1997. The Handbook of Maintenance Management.
New York: Industrial Press INC.
Levitt, J. 2003. The complete guide to preventive and
predictive maintenance. New York: Industrial Press INC.
Manzine, E. & Vezzoli, C. 2003. A strategic design approach
to develop sustainable product service systems: exam-

ples taken from the environmentally friendly innovation.

Journalof Cleaner Production: p. 851857.
Mont, O.K. 2002. Clarifying the concept of product-service
system. Journal of Cleaner Production: p. 237245.
Procaccia, H. & Cordier, R. & Muller, S. 1997. Application
of Bayseian statistical decision theory for a maintenance
optimization problem. Reliability Engineering and System
Safety: p. 143149.
Quan, G. & Greenwood, G. & Liu, D. & Hu, S. 2007. Searching for multiobjective preventive maintenance schedules: Combining preferences with evolutionary algorithms. European Journal of Operational Research:
p. 19691984.
Samrout, M. & Yalaoui, F. & Chtelet, E. & Chebbo, N.
2005. New methods to minimize the preventive maintenance cost of series-parallel systems using ant colony
optimizations. Reliability Engineering & System Safety:
p. 346354.
Sortrakul, N. & Nachtmann, H. & Cassady, C.R. 2005.
Genetic algorithms for integrated preventive maintenance
planning and production scheduling for a single machine.
Computers in Industry: p. 161168.
Tsai, Y.-T. & Wang, K.-S. & Teng, H.-Y. 2001. Optimizing
preventive maintenance for mechanical components using
genetic algorithms. Reliability Engineering and System
Safety: p. 8997.
Wang, L. & Chu, J. & Wu, J. 2007. Selection of optimum
maintenance strategies based on a fuzzy analytic hierarchy
process. International Journal of Production Economics:
p. 151163.
Worsham, W.C. 2000. Is preventive maintenance necessary?
Maintenance Resources On-Line Magazine.

634

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Profitability assessment of outsourcing maintenance from the producer

(big rotary machine study)
P. Fuchs & J. Zajicek
Technical University of Liberec, Liberec, Czech Republic

ABSTRACT: Most companies have their maintenance plans and method of conducting them in place now.
The maintenance plan is in agreement with the producers recommendations or is changed by the maintenance
technician on the basis of his experiences. The producers recommendations for the warranty period neednt be
the best of course, its best for the producer and its warranty. These maintenance strategies could be optimal
for commonly available equipment, but big rotary machines are often unique. Expensive maintenance done
by the producer could be in the long term more profitable because of their better experiences. This paper is
about total cost assessment of big rotary machine (i.e. a hydrogen compressor used in the refinery industry)
related to selected maintenance. Reliability data about MTBF and MTTR, economical evaluation of conducting
maintenance and last but not leastproduction lost from scheduled shutdown or equipment failure, are all
considered as all of these affect the cost. The companies that have one or more big rotary machines should do
a study of operational costs and they should do profitability assessment of outsourcing maintenance from the
producer. An economical model can find other problems, for example the wrong list of spare parts. The model
proposed in this paper could help companies to do studies for their big rotary machines.
1

Effective control of production risks consists of understanding the risk character in operating practice and in
finding suitable tools for evaluating and optimizing
the risk. One of the most important risks is associated
with failures of expensive and unique equipment.
The producers experience and construction knowledge can decrease the failure risk due to change of
mean time between failures and mean time to repair
or due to a shorter delivery period of spares. Better
maintenance can affect the useful lifetime also.
The paper is about the profitability evaluation of
the maintenance contract with the producer and finding profitability boundary values of the new MTBF,
MTTR and delivery period. The warranty, warranty period and additional warranty cost are not
included, because the paper is related to important
equipment whose cost/losses come mainly from production losses. In case of failure during warranty
period producers pay the spares and the reparation,
but production losses are on the users side.
This is a practical application of known procedures
and the model can be easy used for other suitable
equipments.
2

DP = delivery period of spare parts [day]

LT = useful lifetime [year]
MTBF_o = MTBF; needed parts that have to be
ordered [year]
MTBF_s = MTBF; needed parts in store [year]
MTTR_o = MTTR; needed parts that have to be
ordered (includes delivery period) [day]
MTTR_s = MTTR; needed parts in store [day]
PL = production losses [EUR/h]
PUR = purchase cost [EUR]

INTRODUCTION

NOTATIONS

CONx = annual maintenance contract cost [EUR/year]

635

APPLICATION FIELD OF THIS STUDY

The equipment researched for this study, which measures the profitability of outsourcing maintenance, is
compressor. Other suitable equipment types would be
big pumps, turbines, etc. For this application it is
necessary to highlight the following presumptions.
Equipment is unique
The user does not have the same or similar equipment, which can perform the same function.
The delivery period of spare parts for repairs and
scheduled maintenance is in weeks or months.
Equipment or spare parts are expensive.
Equipment performs important function
Equipment downtime causes stopping or reduction
of production. It generates large production losses.

Maintenance significantly affects the equipment

availability
Preventive maintenance tasks change the reliability,
availability, etc.
Maintenance tasks are difficult because of the
equipment structure.
It is possible to effectively apply condition
monitoring.
4

Table 1.

REQUIRED INPUTS

Ideally all possible costs should be inputs to the model.

Practically it is more useful to focus on the main parts
of cost only. It may not be possible to exactly evaluate
all values. Insufficient statistical data (about failures,
maintenance tasks and costs) have to be replaced by
estimations.
Inaccuracy of some inputs with large cost could be
so important that these inaccuracies are larger than
some smaller input costs (e.g. administration, rent a
land). These items with small costs are recommended
to be ignored.
5

The compressor in a refinery was chosen as an example of this study. Its function is compressed hydrogen
production. A great deal of company production
depends on this equipment because the compressed
hydrogen is necessary to the technological process.
The user and the producer will not specify the input
data because of private data protection, therefore this
has been changed, but it has not impacted on the
method.
Equipment basic information:
input: 3MW
purchase cost (including installation and liquidation): 3 million EUR
condition monitoring is necessary (vibrodiagnostic,
tribology)
recommended period of inspections and overhaul:
4 years
recommended period of general overhaul: 12 years
supposed useful lifetime: 30 years
Remark: The recommended periods of inspections
and overhaul proceed from scheduled outage of the
refinery.
5.1

Current cost associated with the equipment

Cost associated with the equipment can be split into

purchase cost, installation and liquidation cost, operating cost, purchase of spare parts, storage and

[EUR /
year]

Cost

[EUR]

[year]

Purchase cost
Spare parts cost
bearings
rotary
seals system
seals


Spare parts Storage

and maintenance
Labour cost
Monitoring
vibrodiagnostic

tribology

Operating cost
(energy)

3 000 000

30

100 000

50 000
30 000
30 000
6 000
116 000

4
12
12
4

12 500
2 500
2 500
1 500
19 000
1500

Table 2.

BIG ROTARY MACHINE STUDY COMPRESSOR

Current cost.

10 000
3 000
600
3600
1 000 000

Production losses.

Production losses

PL [EUR / hour]

Tasks during outage

Long-term downtime

0
15 000

maintenance cost of spare parts, production losses,

administration and labour cost.
Estimation of these cost are:
.103 EUR/year - administration, installation, liquidation, storage and maintenance of spare parts
.104 EUR/year - labour cost, purchase cost of
spare parts, condition monitoring
.105 EUR/year - purchase cost of equipment
.105 till .106 EUR/year - production losses, operating cost
Using a more exact approach, the values used in the
study are given in table 1.
Production losses per year are not in the table above.
For evaluation of this cost it is necessary to have some
reliability quantities (MTTR & MTBF) and production losses per hour during the downtime. Production
losses per hour are specified in table 2.
5.2 Reliability quantities
This study is economically oriented therefore only the
most likely events are sufficient. The reliability quantities are expert estimations because of the low amount
of statistical data about failures, repairs and preventive
maintenance tasks. Obtaining the required statistical data is often a problem for reliable and unique

636

Table 3.

service offer of the producer (section 5.3). Other

varieties will be compared with this variety 1.

MTBF, MTTR.

Reliability quantities

MTBF [year]

MTTR [day]

Parts in storage
Parts not in storage

15
50

10
10 + 300

Variety 2
The maintenance contract includes these points:

equipments therefore the estimations are the best way

forward.
Undesirable events, which stop this equipment and
hence stop the production process too, are necessary to
investigate or eradicate. These events are mostly bearing failure or sealing system failure. It is supposed
that the mean time to repair (MTTR) is 10 days, if the
needed parts are in storage. If the needed parts are not
in storage, it is necessary to add the delivery period.
The delivery period is approximately 10 months for
this equipment type (include ordering, the administrative delay, etc.). The mean times between these events
(MTBF) and MTTR are given in table 3.
MTBF, when the parts are not in storage, includes
the possibility that the machine crashes promptly after
repair therefore new parts are not available.
5.3

Producer services

The producer offers a wide range of services and is

able to ensure full outsourcing of maintenance. The
offered services are summarized as:
1. Training course
2. Condition monitoring

online vibrodiagnostic
tribology
technology parameters monitoring
annual reports

Training course
Condition monitoring
Delivery, storage and maintenance of spare parts
Technical support

The annual contract price is 60 000 EUR. The estimated benefits for the user are MTBF and useful
lifetime increase to 105% and repairing times decrease
to 95%. Due to contract existence this will not delay
the ordering and administration, so the delivery period
could be about 8 months.
Variety 3
The maintenance contract includes these points:

Training course
Condition monitoring
Delivery, storage and maintenance of spare parts
Technical support
Inspection and overhaul conducting (every 4 years)
General overhaul conducting (every 12 years)

The annual contract price is 75 000 EUR. The estimated benefits for the user are MTBF and useful
lifetime increase to 110% and repairing times decrease
to 90%. Due to contract existence this will not delay
the ordering and administration, so the delivery period
could be about 8 months.
The benefits of the varieties are summarized in
table 4.
5.5

Varieties evaluation - effectiveness index

The varieties differ in annual purchase cost (including installation and liquidation), monitoring cost and
production lost.
Purchase cost is the same, but annual purchase cost
depends on useful lifetime (equation 2).

3. Delivery, storage and maintenance of spare parts

4. Maintenance tasks scheduling and conducting
according to machine condition
periodic inspections
periodic overhauls and general overhauls
5. 24h Technical support
5.4

Annual_purchase_ cos t =

Purchase_ cos t
Useful_lifetime
[EUR/year]

Outsourcing maintenance varieties

The offered services for outsourcing can be summarized into two packages (variety 2 & variety 3).
The current method of maintenance is described in
variety 1.
Variety 1
The first possibility is to continue with the current maintenance strategy without using the standard

(1)

Condition monitoring cost is only in variety 1, other

varieties include it in contract price.
Production losses are divided into three possibilities
due to maintenance tasks. Tasks conducted during the
scheduled stop period dont generate production losses
because all the plant is in outage. The second (third)
possibility enables maintenance tasks during common

637

Table 4.

Benefits of varieties.

Variety

Lifetime extension [%]

MTBF extension [%]

Time to repair reducing [%]

Delivery Period [month]

1
2
3

0
5
10

0
5
10

0
5
10

10
8
8

1
2
3

Lifetime
[year]

MTBF SP in storage
[year]

MTBF SP not in storage

[year]

Time to repair
[day]

Delivery time
[month]

30
31.5
33

15
15.75
16.5

50
52.5
55

10
9.5
9

10
8
8

operational time when the spares are (not) in storage.

These possibilities have different MTBF and MTTR
for each variety and that is why the losses are different.
Total annual production losses are evaluated according
to formula (2).

Annual_PL =

MTTR_s
MTTR_o
+
MTBF_s MTBF_o

24 PL [EUR/year]

5.6

(2)

Concrete values of direct cost and risks for each

variety are given in table 5.
At first sight it is evident the profitability of contract
making due to savings. Average annual savings are
generated most by a decrease of catastrophic failure
risk. The catastrophic failure risk mainly comes from
the spare parts delivery period. The producer was not
able to provide the spare parts prices dependent on
the delivery period, although the user was agreeable
to pay a multiple of the standard prices. The solution
is hence contract making and thereby administrative
delay elimination.
The next indicator (except savings) could be the
effectiveness index of the contract. The index indicates
valuation of investment. The formula for this index
follows,
IEx =

TC1 TCx
CONx

difficult as variety 2 has better effectiveness index and

variety 3 generates more savings. It is caused by nonlinear dependence between the benefits and contract
costs.

Another approach is finding boundary values of

some parameters, which can be modified by maintenance. The parameters are spare parts delivery
period, repair time, mean time between failures
and useful lifetime. The contract is profitable if at
least one of these watched parameters achieves its
boundary value. Evaluation of these indicators corresponds to formulas (5)(8). All formulas compare
contract price with differences between current cost
and cost after one parameter change, with the other
parameters the same. The general formula for all
parameters is:
CONx = risk(DP) + risk(MTBFs)
+ risk(MTTR)
+ annual_ cos t(LT )[EUR/year]

(4)

C_ DP = boundary value of delivery period

C_ RT = boundary value of repair time

(3)

where IEx = effectiveness index of variety x;

TC1 = total cost of variety 1 (current strategy);
TCx = total cost of variety x without annual contract
price.
Investment is profitable if the index is greater than 1.
The values after evaluation are IE2 = 9.2 &
IE3 = 8.8. Both of the possibilities are much better
than the current state. Determining the best variety is

Varieties evaluationprofitability boundary

C_ BF = boundary value of MTBF

C_ LT = boundary value of useful lifetime
CONx =

638

(DP C_DP) 24 PL
C_DP
MTBF_o

= DP

CONx MTBF_o
[day]
24 PL

(5)

CONx =

CONx =

MTTR_s 24 PL
MTTR_o 24 PL [(MTTR_o(1 DP) C_RT + DP] 24 PL
(1 C_RT ) +
MTBF_s
MTBF_o

CONx MTBF_o MTBF_s

C_RT = 1
(MTBF_o MTTR_s + MTBF_s MTTR_o MTBF_s DP) 24 PL
MTTR_s 24 PL
MTTR_s 24 PL

MTBF_s
MTBF_s C_BF
+

CONx =

MTTR_s
MTTR_o
+ MTBF_o
) 24 PL
( MTBF_s
MTTR_o 24 PL
MTTR_o 24 PL

C_BF
=
MTTR_s
MTTR_o
MTBF_o
MTBF_o C_BF
( MTBF_s + MTBF_o ) 24 PL CONx

PUR
PUR
PUR

C_LT =
LT
LT C_LT
PUR LT CONx

After real values substitution into formulas (5)(8)

the boundary values are according to table 6.
5.7

Optimal list of spare parts

The number of spare parts has an indispensable effect

on the equipment total cost. Spare parts generate
Table 5.

Contract
Savings

C_ DP
C_ RT
C_ BF
C_ LT

(7)

(8)

additional cost from their purchase cost, storage and

maintenance, but they can sharply change the rate
between MTBF when the parts are / are not in storage.
It ultimately affects annual risk.
For every reasonable option of spare parts list
should be done cost analysis including risks. The lowest total cost indicates the best spare parts list and in
case of need with maintenance contract making.

Annual cost of each variety.

Purchase cost,
installation,
liquidation
Spare parts cost
Storage and
maintenance
of spare parts
Labour cost
Monitoring
Operating cost
Production lost
tasks during stop
period
failure; SP in storage
failure; SP not in
storage

Table 6.

(6)

Variety 1
[EUR/y.]

Variety 2
[EUR/y.]

Variety 3
[EUR/y.]

1 00 000
19 000

95 238
19 000

90 909
19 000

1500
10 000
3 600
10 00 000

1500
10 000
0
10 00 000

1500
10 000
0
10 00 000

2 40 000
22 32 000

2 17 143
17 10 857

1 96 364
16 29 818

36 06 100
0
0

30 53 738
60 000
4 92 362

29 47 591
75 000
5 83 509

5.8 Summary of results

The presented economical model compares two varieties of outsourcing maintenance with the current
state. The effectiveness indexes confirm profitability
of both varieties. Since both indexes are very high
(evaluate the investment nearly 10 times) the recommended possibility is variety with more savings variety 3.
In case of input uncertainty it is suitable to evaluate
boundary values of profitability and then consider if it
is possible to achieve at least one of these parameters.
Concretely for contract from variety 3 is sufficient for
example decrease delivery period from 300 days to
290 days or increase MTBF by 3%.

Profitability boundary.
Variety 2

Variety 3

292
81
102.5
250

290
76
103.1
400

days
%
%
%

CONCLUSION

Most companies have their maintenance plans and

method of conducting them in place now. The
maintenance plan is in agreement with the producers recommendations or is changed by the maintenance technician on the basis of his experiences.
Maintenance tasks are conducted by company employees or contractors for special branches (electrician,
machine staff etc.)
These maintenance strategies could be optimal
for commonly available equipment, but big rotary
machines (especially compressors, turbines), which
are made individually based on clients requirements)

639

are often unique. Expensive maintenance done by the

producer could be in the long term more profitable
because of their better experiences.
The presented economical model targets risks
related with operational and maintenance cost of
unique equipments and could be used as an instruction
for risk optimization.
ACKNOWLEDGMENT
This research was supported by the Ministry of Education, Youth and Sports of the Czech Republic, project

No. 1M06047 - Research Center for Quality and

Reliability of Production.
REFERENCES
Moubray J. M.: Reliability-centred Maintenance. Second
edition. Butterworth-Heineman, Oxford, 1997.
Fuchs, P.: Vyuit spolehlivosti v provozn praxi, Liberec,
2002.
ISO 14121 (1999): Safety of machineryPrinciples of risk
assessment.
ISO 14040 (2006): Environmental managementLife cycle
assessment.

640

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Simulated annealing method for the selective maintenance optimization

of multi-mission series-parallel systems
A. Khatab & D. Ait-Kadi
Mechanical Engineering Department, Faculty of Science and Engineering, Interuniversity Research Center
on Entreprise Networks, Logistics and Transportation (CIRRELT), Universit Laval, Quebec, Canada

A. Artiba
Institut Suprieur de Mcanique de Paris (Supmeca), France

ABSTRACT: This paper addresses the selective maintenance optimization problem for a multi-mission seriesparallel system. Such a system experiences several missions with breaks between successive missions. The
reliability of the system is given by the conditional probability that the system survives the next mission given
that it has survived the previous mission. The reliability of each system component is characterized by its hazard
function. To maintain the reliability of the system, preventive maintenance actions are performed during breaks.
Each preventive maintenance action is characterized by its age reduction coefficient. The selective maintenance
problem consists in finding an optimal sequence of maintenance actions, to be performed within breaks, so that
to minimize the total maintenance cost while providing a given required system reliability level for each mission.
To solve such a combinatorial optimization problem, an optimization method is proposed on the basis of the
simulated annealing algorithm. In the literature, this method has been shown to be suitable for solving such a
problem. An application example with numerical results are given for illustration.

INTRODUCTION

For reparable systems, several mathematical models have been developed in the literature for optimal
design of maintenance policies (for a survey, see for
example (Cho and Parlar 1991; Dekker 1996)). Nevertheless, most of these models do not take into account
the limitations on the resources required to perform
maintenance actions. This drawback has motivated the
development of a relatively new concept called the
selective maintenance. The objective of the selective
maintenance consists in finding, among all available
maintenance actions, the set of appropriate actions
to be performed, under some operation constraints,
so that to maximize the system reliability or either
to minimize the total maintenance cost or the total
maintenance time. Selective maintenance, as a maintenance policy, is relevant to systems that are required
to operate a sequence of missions such that, at the
end of a mission, the system is put down for a finite
length of time to provide an opportunity for equipment
maintenance. Such systems may include for example manufacturing systems, computer systems and
transportation systems.

Dealing with selective maintenance, the first work

is reported in (Rice, Cassady, and Nachlas 1998).
In this work, Rice et al. consider a series-parallel
system which operates a series of missions. Replacement of failed components is performed during breaks
between successive missions. Since breaks are of finite
length, it may be difficult to replace all failed components. In (Rice, Cassady, and Nachlas 1998), to
determine the optimal replacement to be performed, a
maintenance optimization problem is derived to maximize the system reliability for the next mission. To
solve this problem, an heuristic-based method is proposed to overcome the limitations of total enumeration
method. The selective maintenance optimization problem has been also addressed in (Cassady, Pohl, and
Murdock 2001) (Cassady, Murdock, and Pohl 2001)
(Chen, Mend, and Zuo 1999) (Khatab, Ait-Kadi,
and Nourelfath 2007b) and (Khatab, Ait-Kadi, and
Nourelfath 2007a).
Selective maintenance optimization problems handled by the above mentioned works are limited to the
treatment of only one next mission. However, since the
system is required to perform a sequence of missions,
while requiring short development schedules and very

641

high reliability, it becomes increasingly important

to develop appropriate approaches to manage selective maintenance decisions when the planning horizon
considers more than a single mission. In (Maillart,
Cassady, Rainwater, and Schneider 2005), the authors
consider a series-parallel system where each subsystem is composed of identical components whose time
to failure is exponentially distributed. The system is
assumed to operate a sequence of identical missions
such that breaks between two successive missions are
of equal durations. At the end of a given mission,
the only available maintenance action is the replacement of failed components. At a given time, the
average number of successful missions remaining in
the planning horizon is defined. To maximize such
a number, a stochastic programming model is then
proposed. Numerical experiments are conducted to
perform and compare the results obtained for three
maintenance optimization problems. Nevertheless, the
approach proposed in (Maillart, Cassady, Rainwater,
and Schneider 2005) merely relies on a series-parallel
system with few subsystems each composed of components of identical constant failure rates. Furthermore, replacement of failed components is the only
available maintenance action, missions are of identical
time interval and breaks are also of identical durations.
This paper solves the selective maintenance optimization problem proposed in (Khatab, Ait-Kadi, and
Artiba 2007). In (Khatab, Ait-Kadi, and Artiba 2007)
the authors consider a system composed of series of
subsystems each composed of parallel, and possibly,
different components the lifetime of which are generally distributed. The system operates a sequence of
missions with possibly different durations such that
nonidentical breaks are allowed between successive
missions. During a given mission, a component that
fail undergoes minimal repair while at the end of a
mission several preventive maintenance actions are
available. Each preventive maintenance action is characterized by its ability to affect the effective age of
system components. In (Khatab, Ait-Kadi, and Artiba
2007), the proposed selective maintenance optimization problem consists in finding an optimal sequence
of preventive maintenance actions the cost of which
minimizes the total maintenance cost while providing
the desired system reliability level for each mission. In
the present work, to solve this problem, we present an
optimization method based on the simulated annealing
algorithm (Kirkpatrick, Gelatt-Jr., and Vecchi 1983).
The advantage of such an algorithm, over other methods, is known for its ability to avoid becoming trapped
at local minima during the search process.
The remainder of this paper is organized as follows.
The next section gives some notations and definitions
related to the studied multi-mission series-parallel
system. Section 3 addresses the proposed selective
maintenance optimization model and the problem

formulation. The optimization method is presented in

Section 4, and an application example with numerical
results are provided in Section 5. Conclusion is drawn
in Section 6.
2

MULTI-MISSION SERIES-PARALLEL
SYSTEM DESCRIPTION

Consider a series-parallel system S composed of n

subsystems Si (i = 1, . . . , n). Each subsystem Si
is composed of Ni independent, and possibly, nonidentical components Cij (j = 1, . . . , Ni ). Components, subsystems and the system are assumed to experience only two possible states, namely functioning
state and failure state.
Assume that the system is initially new and required
to perform a sequence of M missions each with known
duration U (m), m = 1, . . . , M . Between two successive missions there are breaks of known length of time
D(m, m + 1) for m = 1, . . . , M 1. Namely, the system operates according to two successive states: Up
state Down state Up state. . . . In the Up
state the system is operating while in the Down state
the system is not operating, but available for any maintenance actions. Such a scenario may arise for systems
that operate for some time per a day and then put into
the down state for the rest of the day.
Let Aij (m) and Bij (m) be the ages of component Cij ,
respectively, at the beginning and at the end of a given
mission m(m = 1, . . . , M ). Clearly, one may write
Bij (m) as:
Bij (m) = Aij (m) + U (m).

(1)

If Xij denotes the lifetime of component Cij , then the

reliability Rij (m) of component Cij to survive mission
m is given such that:
Rij (m) = Pr(Xij > Bij (m) | Xij > Aij (m))
=

Pr(Xij > Bij (m))

Pr(Xij > Aij (m))

R(Bij (m))
,
R(Aij (m))

(2)

where R is the survival time distribution function of

the random variable Xij .
If component Cij is characterized by its corresponding hazard function h(t), then the conditional
reliability Rij (m) can be written as:

Rij (m) = exp

642

Aij (m)

Bij (m)

hij (t)dt


hij (t)dt

= exp(Hij (Aij (m)) Hij (Bij (m)),

(3)

t
where Hij (t) = 0 hij (x)dx is the cumulated hazard
function of component Cij .
From the above equation, it follows that the reliability of subsystem Si and that of the system S are
respectively denoted by Ri (m) and R(m) and given as:
Ri (m) = 1

Ni

(1 Rij (m)),

and

R(m) =

Ri (m).

(5)

i=1

(m = 1, . . . , M 1)

(4)

j=1
n


performed. To construct such a model, the following

decision variable is introduced:

1 if Cij undergoes PM ap
ap (Cij , m) = at the end of mission m,

0 otherwise.

SELECTIVE MAINTENANCE MODEL

AND PROBLEM FORMULATION

In this paper two types of maintenance are considered, namely corrective maintenance (CM) and
preventive maintenance (PM). CM by means of
minimal repair is carried out upon components
failures during a given mission while PM is a
planned activity conducted at the end of missions
(i.e. within breaks) to improve the overall system mission reliability. It is assumed that component failure is operational dependent and the
time in which a given component undergoes minimal repair is negligible if compared to the mission
duration.
Each component Cij of the system is characterized by its hazard function hij (t) and its minimal
repair cost cmrij . The preventive maintenance model
is given on the basis of the age reduction concept initially introduced by Nakagawa (Nakagawa
1988). According to this concept, the age of a
given component is reduced when PM action is performed on this component. In this paper, the vector
VPM = [a1 , . . . , ap , . . . , aP ] represents the P PM
actions available for a given multi-mission system. For
each PM action ap (p = 1, . . . , P) is assigned the cost
cpm(ap ) and the time duration dpm(ap ) of its implementation, the age reduction coefficient (ap ) [0, 1]
and the set Comp(ap ) of components that may undergoes PM action ap . Regarding the values taken by
a given age reduction coefficient (ap ), two particular cases may be distinguished. The first case
corresponds to (ap ) = 1 which means that the PM
action ap has no effect on the component age (the
component status becomes as bad as old), while the
second case is (ap ) = 0 and corresponds to the
case where the component age is reset to the null
value (i.e. the component status becomes as good
as new).
Selective maintenance model attempts to specify a
PM action that should be performed, on which component and at the end of which mission it has to be

(6)

In this paper, the selective maintenance problem

consists in finding an optimal sequence of maintenance actions the cost of which minimizes the total
maintenance cost while providing the desired system
reliability level for each mission. The total maintenance cost is composed of minimal repair cost
CMRij (m) induced by the repair of each component Cij during each mission m, and the preventive
maintenance cost CPMij (m) of each component Cij
that undergoes preventive maintenance at the end of
mission m.
The cost induced by minimal repairs is function
of components failure rates. Following the work of
Boland (Boland 1982), for a given component Cij , the
expected minimal repair cost in an interval [0, t] is:
 t
cmrij
hij (x)dx.
(7)
0

According to the above equation, the minimal

repair cost CMRij (m) induced by component Cij during
mission m is such that:
 Bij (m)
hij (x)dx,
(8)
CMRij (m) = cmrij
Aij (m)

where Aij (m) and Bij (m) represent the ages of component Cij , respectively, at the beginning and at the end
of a given mission m (m = 1, . . . , M ) and Aij (1) = 0
by definition. If component Cij undergoes preventive
maintenance action ap (p = 1 . . . , P) at the end of
mission m, then the value of the component age Bij (m)
is reduced by the age reduction coefficient (ap ). In
this case, the minimal repair cost CMRij (m) assigned
to Cij becomes:


CMRij (m) = cmrij

g((ap ))Bij (m)

Aij (m)

hij (x)dx ,

(9)

where the function g is related to the value taken by

the decision variable ap (Cij , m) and defined to be such
that:

(ap ) if ap (Cij , m) = 1,
g((ap )) =
(10)
1 otherwise.
According to the above equation, the total minimal repair cost CMRij assigned to Cij which undergoes

643

preventive maintenance actions at the end of missions

1, . . . , M 1 is given such that:

CMRij = cmrij

Bij (M )

Aij (M )

+ cmrij

hij (x)dx


1 
P M


p=1 m=1

g((ap ))Bij (m)

Aij (m)

By using components accumulated hazard rates,

Equation (11) may be written as:

1
P M


CMRij = cmrij Hij (M +
Hij (m, p),
p=1 m=1

(12)
where Hij (M ) = Hij (Bij (M )) Hij (Aij (M )) and
Hij (m, p) = Hij (g((ap )) Bij (m)) Hij (Aij (m)).
From Equation (12), it follows that the total cost
CMR of minimal repair, induced by all components
during missions, is given by:
Ni
n 


CMRij .

(13)

i=1 j=1

The total preventive maintenance cost CPMij

assigned to component Cij , which undergoes preventive maintenance actions at the end of missions, is
given by:
CPMij =

1
P M



cpm(ap ) ap (Cij , m).

(14)

p=1 m=1

It follows, from the above equation, that the total

preventive maintenance cost CPM induced by all
system components is:
CPM =

Ni
n 


DPM (m) =

hij (x)dx .
(11)

CMR =

of a given mission. Therefore, time between missions

should be taken into account as an operation constraint.
The total duration DPM (m) spent by preventive maintenance actions at the end of a given mission m is given
by the following formula:
Ni
n 
P 


dpm(ap ) ap (Cij , m).

(17)

p=1 i=1 j=1

Assume that the system has just achieved the first

mission and will operate the remaining missions, and
let R0 denotes the required reliability level of the
system at the beginning of each mission m (m =
2, . . . , M ). The selective maintenance problem is then
formulated as follows: from the vector VPM find
the optimal sequence of PM actions which minimizes
the total maintenance cost Ctotal while providing the
desired reliability level R0 . To derive the mathematical
programming model corresponding to such a problem,
let the vector S = [s1 , . . . , sK ] be the sequence of PM
actions performed so that to keep the system reliability
at the desired level R0 . Roughly speaking, the vector
S is of dimension K  P and composed of elements
of the vector VPM. At the end of a given mission, if
preventive maintenance is required, then the first PM
action to be performed corresponds to the first element s1 of S. Whenever, action s1 is not sufficient to
guaranty the system reliability level, in this case PM
actions s1 and s2 should be performed simultaneously,
and so on. The mathematical programming model corresponding to the selective maintenance optimization
problem is:
MinimizeCtotal (S) = CMR(S) + CPM (S),

(18)

Subjectto :
R(m + 1)  R0 ,

(19)

DPM (m)  D(m, m + 1),

(20)

sK


ap (Cij , m)  1,

(21)

p=s1

CPMij .

(15)
ap (Cij , m) {0, 1},

i=1 j=1

Finally, the total maintenance cost Ctotal to be minimized is given from Equations (13) and (15) such
that:
Ctotal = CMR + CPM .

(16)

To complete the selective maintenance optimization

problem, let note that, due to the limited time (break)
between missions, it may be not possible that all preventive maintenance actions be performed at the end

(22)

i = 1, . . . , n; j = 1, . . . , Ni ;
m = 1, . . . , M 1;

p = s1 , . . . , sK ,

and K  P.

(23)
(24)

where constraint (20) stands that PM actions undertaken at the end of a given mission should be completed
within the allotted time, constraint (21) imposes the
fact that each component may receive almost one PM
action at the end of each mission, while constraint(22)
is a {0, 1}-integrality constraint.

644

4
4.1

OPTIMIZATION METHOD

4.2 Simulated annealing algorithm for the selective

maintenance optimization problem

The simulated annealing algorithm

Simulated annealing (SA) is one of the most local

search metaheuristics which has been widely studied
and shown to be suitable for a wide range of combinatorial optimization problems, as in the case of
the selective maintenance optimization problem formulated in this paper. The SA principle exploits an
analogy between the way a metal cools and freezes
into a minimum energy crystalline structure and the
search for a minimum in a more general system. The
application of this algorithm to solve combinatorial
optimization problems was initiated by Kirkpatrick
et al. (Kirkpatrick, Gelatt-Jr., and Vecchi 1983). The
major advantage of SA algorithm over other methods is
known for its ability to avoid becoming trapped at local
minima during the serach process. Figure 1 presents
an overview of the SA algorithm.
The algorithm starts by an initial solution s generated either randomly or heuristically, and by an
initial temperature Ti . Then, a solution s is randomly
sampled from the neighborhood V (s) of the current
solution s. The solution s is then accepted or rejected
depending on the current temperature Tc and the values of the objective function f at points s and s (i.e.
f (s) and f (s )). Since the selective maintenance optimization problem consists in maximizing the system
reliability, it follows that s will be accepted with probability 1 as a new solution whenever f (s ) > f (s).
However, in the case where f (s )  f (s), s will be
accepted with probability which is function of Tc and
f = f (s)f (s ). Thisprobability
 follows Boltzmann
f
distribution p = exp
. The temperature is
Tc
decreased following the progression formula Tc = Tc
where represents the cooling schedule. The search
process is continued until the termination condition
T = Tf holds, where Tf is the minimal temperature.

In this paper, the simulated annealing algorithm is

used as an optimization technique to solve the selective maintenance optimization problem. The solution
representation is inspired from that of (Levitin and Lisnianski 2000) (see also (Nahas, Khatab, Ait-Kadi, and
Nourelfath 2007)). The element of the vector VPM
of available PM actions are numbered from 1 to P.
The maintenance plan, as a solution, is represented
by a vector S = [s1 , . . . , sK ] with finite length K P
and where sp {1, 2, . . . , P} , for p = 1, . . . , K. The
length of a given solution depends on its feasibility.
The initial feasible solution is derived on the basis of
the following procedure.
4.2.1 Initial solution construction
1. Set the length of S to a constant number Kmax
2. Generate the elements of S from a random permutation of the set {1, . . . , P}
3. Set K = 1
4. Calculate the objective function and the constraint
values by using the K first elements (i.e. PM
actions) of S
5. if (K = Kmax ) and (S is not feasible) then return to
step 2
6. If (K < Kmax ) and (S is not feasible) then set K =
K + 1 and proceed from step 4
To define the appropriate neighborhood, several
structures were investigated. The following procedur
provides the neighbor solution.
4.2.2 Neighboring solution construction
1. Generate randomly a number x from the interval
[0, 1]
2. If (x 0.5) then choose randomly an element S(i)
with 1 i N , and randomly increase or decrease
by 1 the content of S(i), i.e. S(i) = S(i) + 1 or
S(i) = S(i) 1
3. If (x < 0.5) then choose randomly two elements
S(i) and S(j) with 1 i K and K + 1 j P,
and exchange the contents of S(i) and S(j).
It is worth noticing that in order to ensure the
feasibility of a given solution, one needs to evaluate the objective function and the constraints of the
optimization problem. To this end, a procedure was
developed for the feasibility test of a given solution
vector S = [s1 , . . . , sK ].
5

Figure 1.
algorithm.

An overview of the simulated annealing

APPLICATION EXAMPLE

The test problem used is this paper is based on a seriesparallel system composed of n = 4 subsystems Si
(i = 1, . . . , 4) as shown in Figure 2. Subsystems S1 ,

645

Table 2.

Figure 2.
example.

Table 1.

The series-parallel system for the application

Parameters of system components.

Component Cij

ij

ij

cmrij

C11
C12
C13
C14
C15
C21
C22
C23
C31
C32
C33
C34
C41
C42

0.006
0.01
0.006
0.002
0.008
0.002
0.01
0.011
0.002
0.007
0.009
0.006
0.002
0.002

1.36
2.19
1.66
1.93
1.41
1.58
2.57
2.84
2.01
1.34
1.13
1.7
1.27
2.09

1
2.3
1.5
1
1.9
2.6
1.6
1.3
1.2
2.4
1.9
0.7
0.7
0.4

S2 , S3 and S4 are, respectively, composed of N1 = 5,

N2 = 3, N3 = 4 and N4 = 2 components. The reliability function of each component is given by a Weibull
hazard function h(t) = (t) 1 , where and are,
respectively, the scale and the shape parameters. The
accumulated hazard function is then H (t) = (t) .
For each component Cij (i = 1, . . . , 4; j =
1, . . . , Ni ), the scale ij and shape ij parameters and
minimal repair cost cmrij , are randomly generated and
reported in Table 1. The vector VPM of available
PM actions is also randomly genertaed and given by
Table 2. The system is assumed to experience M = 10
missions. Durations of missions and breaks are given
in Table 3.
The proposed algorithm is implemented by using
MATLAB software tool on a 1.86 GHz Intel Core Duo
processor. For required reliability level R0 = 0.90, the
algorithm was tested for several values of parameters
Ti , Tf and . The most appropriate values are found
to be Ti = 80, Tf = 105 and = 0.99. The length
assigned to the solution vector S is set to Kmax = 30.
The best selective maintenance plan obtained for the
mission reliability level R0 = 0.90 is presented in
Table 4.

Parameters of preventive maintenance actions.

PM action p

Comp(p)

(p)

cpm(p)

dpm(p)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

C11
C11
C12
C12
C13
C13
C14
C14
C15
C15
C21
C21
C22
C22
C22
C23
C23
C23
C31
C31
C32
C32
C32
C33
C33
C33
C34
C34
C34
C41
C41
C41
C42
C42

0.51
0.41
0.00
0.48
0.27
1.00
1.00
0.12
0.15
0.00
0.00
1.00
0.37
0.18
0.00
0.42
0.38
0.68
0.48
0.34
0.42
0.71
0.34
0.13
1.00
0.43
0.63
1.00
0.75
0.78
0.52
0.00
1.00
0.58

10.01
14.33
15.08
3.3
6.28
5.93
7.38
18.19
16.4
20.83
19.6
7.73
6.92
16.67
16.99
10.54
13.22
5.04
11.83
13.96
12.25
4.49
13.33
18.55
9.70
1.68
5.45
1.44
8.71
3.68
6.54
17.66
2.53
9.51

0.66
1.02
1.50
0.22
0.49
0.30
0.37
1.62
1.43
2.08
1.96
0.39
0.51
1.41
1.70
0.74
0.96
0.30
0.80
1.04
0.86
0.26
0.99
1.64
0.49
0.12
0.33
0.07
0.50
0.21
0.43
1.77
0.13
0.6

Table 3. Durations of missions and breaks between

successive missions.
U
D

84
9

57
9

59
9

54
11

78
7

68
9

70
9

76
7

84
11

49

According to the available PM actions given by

Table 2, Table 4 gives components and the time at
which they should receive a specific type of PM action
so that to ensure the system reliability level R0 for each
mission. Figure 3 presents the system reliability versus the number of missions. The best total maintenance

cost Ctotal
, induced by the obtained selective mainte
nance plan and minimal repairs, is Ctotal
= 400.64,
while the execution time is about 76 sec .

646

Table 4. The best selective maintenance plan obtained for

the required reliability level 0.90.
Mission m

PM action p

Cij Comp(p)

2
4

11
15,30,26,
20,8
11,34
15,26,34,
30,20
20,30,
15,8
11,20,8,
26,15

C21
C22 , C41 , C33 ,
C31 , C14
C21 , C42
C22 , C33 , C42 ,
C41 , C31
C31 , C41 ,
C22 , C14
C21 , C31 , C14 ,
C33 , C22

5
6
7
8

Figure 3.
horizon.

this problem, an optimization method is proposed on

the basis of the simulated annealing algorithm.
REFERENCES

Mission reliability of the system in the planing

CONCLUSION

In this paper, we proposed a selective maintenance

optimization model for a multi-mission series-parallel
system. Lifetime of each system component is generally distributed. The system operates on a planning horizon composed of several missions such that
between successive missions a break of finite length is
allotted to perform maintenance actions. Missions as
well as breaks are of possibly different durations, and
during breaks a list of preventive maintenance actions
are available for system components maintenance. A
combinatorial optimization problem is formulated the
objective of which consists in finding, during the planing horizon, an optimal sequence of preventive maintenance actions to be performed so that to minimize
the total maintenance cost while providing, for each
mission, the desired system reliability level. To solve

Boland, P. (1982). Periodic replacement when minimal repair

costs vary with time. Naval Research Logistic Quarterly
29(4), 541546.
Cassady, C. R., W. P. Murdock, and E. A. Pohl (2001).
Selective maintenance for support equipement involving multiple maintenance actions. European Journal of
Operational Research 129, 252258.
Cassady, C. R., E. A. Pohl, and W. P. Murdock (2001). Selective maintenance modeling for industrial systems. Journal
of Quality in Maintenance Engineering 7(2), 104117.
Chen, C., M. Q.-H. Mend, and M. J. Zuo (1999). Selective maintenance optimization for multistate systems. In
Proc. of the IEEE Canadian Conference on Electrical and
Computer Engineering, Edmonton, Canada, 14771482.
Cho, D. I. and M. Parlar (1991). A survey of maintenance models for multi-unit systems. European Journal
of Operational Research 51, 123.
Dekker, R. (1996). Application of maintenance optimization
models: a review and analysis. Reliability Engineering
and System Safety 51(3), 229240.
Khatab, A., D. Ait-Kadi, and A. Artiba (2007). Selective
maintenance optimization for multimission series-parallel
systems. European Journal of Operational Research
(submitted).
Khatab, A., D. Ait-Kadi, and M. Nourelfath (2007a). Algorithme du recuit simul pour la rsolution du problme doptimisation de la maintenance slective des systmes srieparallle. Seventh International Conference on Industrial
Enginnering, Trois-Rivires, QC, Canada.
Khatab, A., D. Ait-Kadi, and M. Nourelfath (2007b).
Heuristic-based methods for solving the selective maintenance problem in seriesprallel systems. International
Conference on Industrial Engineering and Systems Management, Beijing, China.
Kirkpatrick, S., C. D. Gelatt-Jr., and M. P. Vecchi (1983).
Optimization by simulated annealing. Science 220(4598),
671680.
Levitin, G. and A. Lisnianski (2000). Optimization of
imperfect preventive maintenance for multi-state systems.
Reliability Engineering and System Safety 67, 193203.
Maillart, L., C. R. Cassady, C. Rainwater, and K. Schneider (2005). Selective Maintenance Decision-Making Over
Extended Planing Horizons. Technical Memorandum
Number 807, Department of Operations, Weatherhead
School of Management, Case Western Reserve University.
Nahas, N., A. Khatab, D. Ait-Kadi, and M. Nourelfath
(2007). Extended great deluge algorithm for the imperfect preventive maintenance optimization of multi-state
systems. REliability Engineering and System Safety
(submitted).
Nakagawa, T. (1988). Sequential imperfect preventive maintenance policies. IEEE Transactions on Reliability 37(3),
295308.
Rice, W. F., C. R. Cassady, and J. Nachlas (1998). Optimal maintenance plans under limited maintenance time. In
Proceedings of Industrial Engineering Conference, Banff,
BC, Canada.

647

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Study on the availability of a k-out-of-N System given limited spares

under (m, NG ) maintenance policy
T. Zhang, H.T. Lei & B. Guo
National University of Defense Technology, Changsha, Peoples Republic of China

ABSTRACT: This paper considers a k-out-of-N system with identical, repairable components under called
(m, NG ) maintenance policy. Under this policy, maintenance is initiated when the number of failed components
exceeds some critical level identified by m. After a possible set-up time of spares replacement, at least NG
components should be good in the k-out-of-N system when it is going to be sent back to user. A multi-server
repair shop repairs the failed components. The operational availability of this kind depends on not only the spare
part stock level, the repair capacity, but also the two parameters m and NG of maintenance policy. This paper
presents a mathematical model of operational availability for repairable k-out-of-N system given limited spares
under (m, NG ) maintenance policy. We can make trade-off between the spare part stock level, the number of
repairmen and two parameters of maintenance policy using this model. From the analysis of an example, we get
the some valuable conclusions.

INTRODUCTION

For its high reliability, k-out-of-N systems has been

more and more used in the equipments of modern
advanced technology, such as the systems of aircraft,
warship, or missile. An example is the Active Phased
Array Radar (APAR). This radar has a cubical shape.
On each of the four sides, it has a so-called face,
consisting of thousands of transmit and receive elements. A certain percentage of the total number of
elements per face is allowed to fail, without losing the
function of the specific radar face. Say that this percentage is 10% and that the total number of elements
is 3000, then we have a 2700-out-of-3000 system.
And the operational availability of these systems is
affected by the number of spare parts, repair teams
and so on. By analyzing the operational availability
with the different number of spare parts and repair
teams under the different maintenance policy, we can
give support to the decision of maintenance of these
systems. For these systems, Ref.[1] provides a simple policy of condition-based maintenance (we call it
m-maintenance policy here), and gives the model of
operational availability of a k-out-of-N system under
this policy. The m-maintenance policy is the policy
that the system asks for the repair of spare parts when
there are m failed components in the system (the available ones are reduced to N m). But actually, Because
of the limit of spare parts, system can be restored to the
state during which the number of available components

is less than N for the rapid restoration of the repairing

equipments. This can be called (m, NG ) maintenance
policy. And the (m, NG ) maintenance policy means
that replacement of components is only initiated if the
number of available components is less than N m.
And after the repair, the number of available components at least is more than NG (N m < NG N )
before the system will be transited back to the user.
When NG = N , (m, NG ) maintenance policy is just
m-maintenance policy.
Some work is available on the operational availability model of k-out-of-N systems. Ref.[2] analyses
the availability of parallel repairable system; Ref.[3]
analyses the operational availability of a 1-out-of-N
system cold standby redundancy system under the
cycle replacement policy; Ref.[4] analyses the operational availability when the failed components are
repaired immediately and it exists just one repair
team which finish the replacement and repair job
of spare part together in the system; Ref.[5] provides a operational availability model of a k-out-of-N
system which exists cold standby redundancy and
the hot standby redundancy together. It also considers the failed components repaired immediately.
Ref.[1] provides a operational availability model of
the hot standby redundancy k-out-of-N system. This
model considers the m-maintenance policy. And this
paper does some research on the operational availability model of a k-out-of-N system under (m, NG )
maintenance policy.

649

DESCRIPTION OF PROBLEM

In this section, we also describe the k-out-of-N system

with hot standby redundancy. At the start of a system
uptime, all N components are as good as new. The
failure process of each component is characterized by
a negative exponential distribution with rate ; where
we assume that the component failure processes are
mutually independent. The system functions properly
as long as at least k (k > 0) components are available.
It takes (m, NG ) maintenance policy for the system.
There are c repair teams which especially repair the
replaced failed components in the repair shop. Each
team can repair one failed component in the same
time. The repairing time has the negative exponential distribution with rate ; where we assume the
time of replacement is short enough to be neglected
and each repair team can independently work without stopping. The failed component is as-good-as-new
after repair. Also it needs time Td to send the system to the repair shop after it was dismantled, and
time Ts to fix the system and transit back to the user
after replacement of spare parts. If insufficient spares
are available, the maintenance completion is delayed
until the number of available components at least is
NG . Now, we want to get the operational availability
Ao of this system given the initial number X of spare
parts.

RESOLVING PROCESS

The system cycle starts with all N components as good

as new. We define the system operation cycle which
is from the system start time to the next and the spare
replacement cycle which is from the spare replacement start time to the next. So the system operation
cycle includes 4 processes: the period of system up,
the period of system down and in transit to depot, the
period of component replacement, and the period of
system in transit back. See Figure 1.
We know the operational availability equals the
expected uptime during a cycle divided by the expected
cycle length. So, we find

Ao =

E(To )
E(To ) + E(Td ) + E(Tr ) + E(Ts )

(1)

where E(To ) is the expected uptime until system down

and maintenance, E(Td ) is the expected time during
system is down and in transit to depot, E(Tr ) is the
expected time of the period of component replacement, E(Ts ) is the expected time during system is
restored and in transit back. Because we assume the
time of replacement can be neglected, E(Tr ) is the time

waiting for sufficient repaired spare parts. Tr is zero

when there are sufficient spare parts.
Given E(Td ) = Td and E(Ts ) = Ts , to compute
Ao , the key point is to compute the excepted operational time of system E(To ) and the excepted lingering
time E(Tr ) which is produced by the waiting for the
demanding spare parts.
3.1 Definition
The state of system is defined as (n, s). It consists
of the number of operational components n and the
number of available spare parts s. The n and s are
constrained by NG n N , 0 s N + X NG
and NG n + s N + X . Before discussing the
formulas of E(To ) and E(Tr ), we define the following
symbols which will be used in the resolving process.
Pr (a, b, c, t) is the probability when the number of
failed components is reduced from a to b by the repair
of c repair teams in the time t.
Pa (n) is the steady probability when the number
of operational components is n at the starting time of
operational period.
Pb (n, s) is the steady probability when the initial
state of operational period of system is (n, s).
p(ns ,ss ),(ne ,se ) is the state translation probability when
the state of system is translated from (ns , ss ) to (ne , se )
during the system operation cycle.
Ps (s) is the steady probability when there are just s
available spare parts at the starting time of the period
of spare replacement.
NM is the number of operational components when
the system asks for the repair, and NM = N m is the
fixed value.
Z is the maximal number of available spare
parts. When the number of components in the system is NG and there is no components needing
repaired, the number of availability spare parts is
maximal. So Z = S + N NG , and the maximal number of components waiting to be repaired is
also Z.
3.2

The expression of Pr (a, b, c, t)

Because of frequent use of Pr (a, b, c, t), we give its

formula firstly.
If we treat c repair men as c servers, a components
which are repaired or waiting for the repair as aguests
which are served or waiting for the service, and the
repairing time of failed components has exponential
distribution with rate , the resolving of Pr (a, b, c, t)
can be treated as the queue model which has several servers and finite guests. Assume that there are
c independent servers with the serving rate . The
sum of the guests which are served or are waiting
for the service is a at the starting time. There are not
new guests. The waiting guest can enter into any free

650

To

Td

Tr

Ts

the system operation cycle

the spare replacement cycle

Figure 1.

the period of system is operational

the period of system is down and in transit to depot

the period of spare replacement

the period of system is restored and in transit back

The cycles and periods of system.

cec(t ) Pr (a 1, b, c, ) for at [0, t]. And we

can get the recursive Equation (4).

server. And guests leave when they finish their services. So, Pr (a, b, c, t) equals to the probability when
the sum of the guests which are served or are waiting for the service is b, after t time. The formula of
Pr (a, b, c, t) can be treated according to the following
conditions.

Pr (a, b, c, t)
= cect

1. When b > a, as there is not new guest, the sum of

the guests which are served or are waiting for the
service cant be increased. This condition cant be
produced, and Pr (a, b, c, t) = 0.
2. When a = b, it means that there isnt any leaving guest finishing the service in t time. When
a > c, the sum of guests is larger than the number of servers, and all servers are in their job with
c guests served in the same time. Thus, the time in
which one guest finishes the service has exponential distribution with serving rate c. When a c,
the sum of guests is smaller than the number of
servers. And all guests can be served, the serving
rate is a. Synthesizing these two conditions, we
can get the serving rate min {a, c} . Whena = b,
all guests have not finished their service in ttime.
So, we can get the following equation.
Pr (a, b, c, t) = e

min{a,c}t

ec Pr (a 1, b, c, ) d (4)

According to Equation (4), when a > c >

b 0,
 t
Pr (a, b, c, t) = cect ec Pr (a1, b, c, )d
= cect


ec c ec ec


Pr (a 2, b, c, )d d
= (c)2 ect

= =

ec Pr (a2, b, c, )(t )d

(2)

(c)ac ect
(a c 1)

ec
0

Pr (c, b, c, )(t )ac1 d

(5)

where Pr (c, b, c, ) = Ccb eb (1 e )cb .

According to the Equation (4), when c b < a,
as the deducing process of Equation (5), we can get:

(3)

4. When 0 b a and a > c, as a > c, there

are c guests served at most in the same time. In
this condition, the probability is Pr (a 1, b, c, ),
when the number of failed components is reduced
from a 1 to b in time by repairing. And
the probability density function is cec(t ) ,
when one guest finishes his service in the left
t time. So, Pr (a, b, c, t) is the integral of

t
0

3. When b < a and a c, a guests are all being

served. And number of the guests who have finished their service in t time is a b. b guests have
not finished in ttime. So, there is the following
equation.
Pr (a, b, c, t) = Cab ebt (1 et )ab

Pr (a, b, c, t) =


(c)ab ect
(a b 1)!

ec Pr (b, b, c, )(t )ab1 d

(6)

where Pr (b, b, c, ) = ec .
Synthesizing all above conditions, we can get the
following computing equation of Pr (a, b, c, t).

651

e min{a,c}t ,

Cab ebt (1 et )ab

ac


cb1

b i
i

C
(1)
C

a cb

cbi

i=0

ac1

C ac
(b+i)t
ct

(e

e
)

Pr (a, b, c, t) =
j
j=1 (c b i)

(t)acj ct

e
+ Ccb

(a c j)!

(1)cb (ct)ac

ect ,

(a c)!

(ct)ab ct

e
,
(a b)!

3.3

The expression of E(To )

In an operational period of system, we use To (n) to

express the operational time of system when the number of available components is n at the starting time of
operational period. It also means the interval when the
state of system that has n available components translates to the state of system that has n NM failed
components. Consequently, there is the following
equation.

To (n) =

nN
M 1

i=0

1
(n i)

(8)

So, we can also get the following equation of E(To ).

E(To ) =

N


(Pa (n) To (n))

n=NG

N

n=NG

Pa (n)

nN
M 1

i=0

1
(n i)

a < b or a < 0 or b < 0

a=b
0bac

(7)

a>c>b0
a>bc0

the number of available components and consists of

the number of available spare parts. Because the failing time and the repairing time of components both
have the exponential distribution, and the probability of system state at the end of operational period
is just related to the system state at the starting time
of operational period, it can form a Markov chain.
Before resolving Pa (n), we can compute Pb (n, s)
firstly.
To get Pb (n, s), the key point is to get the translating matrix of state Q = [p(ns ,ss ),(ne ,se ) ]. The blow is
discussing p(ns ,ss ),(ne ,se ) .
We divide (ns , ss ) (ne , se ) into two phases. The
first phase is (ns , ss ) (NM , sm ). It begins at the starting time of operational period, and ends at the starting
time of component replacement period. The second
phrase is (NM , sm ) (ne , se ). It begins at the starting
time of component replacement period, and finishes
at the end time of component replacement period. The
time of first phase is t1 . t1 equals to the addition of the
time when the state of system is available and the time
used to ask for repair and send to the repair shop. And
we can get the following.

(9)

The equation of Pa (n) is the next thing we are going

to discuss. We have known that the state of system

t1 = To (ns ) + Td = Td +

ns N
M 1

i=0

652

1
(ns i)

(10)

p2 (NM , sm , ne , se ) = Pr (N + X ne , N + S

sm is the number of available spare parts at the starting

time of component replacement period. sm cant be
less than ss which is the number of available spare
parts at the starting time of operational period. But
it is possible larger than the sum (N + X ns ) of
failed components that is repairing or waiting for the
repair. Thus, we can get the constraint ss sm
N + X ns .
Consequently, The translating probability of system
state (ns , ss ) (ne , se ) equals to the addition of the
possibilities for any sm , when the state of system translates from (ns , ss ) to (NM , sm ), and then from (NM , sm )
to (ne , se ).
p(ns ,ss ),(ne ,se ) =

X +N
ns

p1 (ns , ss , NM , sm )

sm =ss

p2 (NM , sm , ne , se )
where p1 (ns , ss , NM , sm ) is the translating probability when the state of system changes from
(ns , ss ) to (NM , sm ) in the first phase. p2 (NM , sm , ne , se )
is the translating probability when the state of system changes from (NM , sm ) to (ne , se ) in the second
phase.
p1 (ns , ss , NM , sm ) equals to the probability when the
number of the components repaired in t1 time is sm ss .
So, we can get following equation.

p1 (ns , ss , NM , sm ) = Pr (L1 , L2 , c, t1 )

ne se , c, ts )

3. When ne = N , sm N NM and se sm N +NM ,

sm can fill the maximal demand of replacement, and
the number of available spare parts is sm N + NM
at the end time of replacement.

p2 (NM , sm , ne , se ) = Pr (X + NM N sm ,
X se , c, ts )

p2 (NM , sm , ne , se )

Pr (Z, Z se , c, ts ) ,

Pr (N + X ne , N + X ne se , c, ts ) ,

Pr (X + NM N sm , X se , c, ts ) ,

0,
ne = NG and sm NG NM
NG < ne < N and sm = ne NM

where L1 = N + X ns ss , L2 = N + X ns sm .
According to (m, NG ) maintenance policy, it should
go through the process of replacement of failed components and waiting for spare parts. Therefore, we can
discuss it according to following conditions.

p2 (NM , sm , ne , se ) = Pr (Z, Z se , c, ts )

(14)

According to (m, NG ) maintenance policy, The condition that ne and sm dissatisfy above conditions, is
impossible to happen. Thus, p2 (NM , sm , ne , se ) = 0.
Synthesizing the Equation (13),(14) and (15), we
can get the equation of p2 (NM , sm , ne , se )

(11)

1. When ne = NG and sm NG NM , the number

of available components goes back to NG , and the
number of spare parts is reduced to 0 at the end time
of replacement process. And p2 (NM , sm , ne , se )
equals to the probability when the number of failed
components is reduced from Z to Z se by c repair
men.

ne = N and sm N NM and se sm N +NM

others
If we take the Equation (12) and (16) in the Equation (11), we can get p(ns ,ss ),(ne ,se ) . According to the
theory of Markov, there is the following equation.

Pb (n, s) =

N X
+N i



Pb (i, j) p(i,j),(n,s)

i=NG

j=0

(12)

2. When NG < ne < N and sm = ne M , sm can fill

the demand of repair, and the system restores to the
state of ne available components. And the number
of spare parts is also reduced to 0 at the end time
of replacement process.

(13)

Order
 = [Pb (NG , 0) Pb (NG , Z)

653

Pb (NG + i, 0) Pb (NG + i, Z i) ]T ,

(15)

p(NG ,0),(NG ,0)

..
.
p(NG ,Z),(NG,0 )
..
.
p(NG +i,0),(NG,0 )
..
.
p(NG +Zi),(NG,0 )
..
.
p(N0 ),(NG ,0)
..
.
p(N ,X ),(NG ,0)

Q =

|
.
| ..
|
.
| ..
|
.
| ..
|
.
| ..
|
.
| ..
|

..
.

..
.

..
.

..
.

..
.

p(NG ,0),(N ,0)

..
.
p(NG ,Z),(N ,0)
..
.
p(NG +i,0),(N ,0)
..
.
p(NG +i,Zi),(N ,0)
..
.
p(NG ,0),(N ,0)
..
.
p(N ,X ),(N ,0)

p(NG ,0),(NG ,Z)

..
.
p(NG ,Z),(NG,Z )
..
.
p(NG +i,0),(NG,Z )
..
.
p(NG +i,Zt),(NG,Z )
..
.
p(N0 ),(NG ,Z)
..
.
p(N ,X ),(NG ,Z)

..
.

..
.

..
.

..
.

..
.

..
.

..
.

..
.

..
.

..
.

P(NG ,0),(N ,X )
..
.
p(NG ,Z),(N ,X )
..
.
p(NG +i,0),(N ,X )
..
.
p(NG +i,Zi),(N ,X )
..
.
P(N ,0),(N ,X )
..
.
P(N ,X ),(N ,X )

we can get
=Q
Pa (n) =

Z


(17)

Pb (n, s)

(18)

s=0

Then, according to the Equation (19) and (20), we

can get Pb (n, s) and Pa (n).
3.4

The expression of E (Tr )

At the starting time of components replacement, when

the number (sm ) of available spare parts satisfies the
least demand (NG NM ) of repair, it can finish the

E (Tr ) =

NG N
M 1 

 
E Tc NG NM sm , N

sm =0

+X NM sm , c



Ps (sm )

p(NG ,0),(NG +i,0)

..
.
p(NG ,Z),(NG +i,0)
..
.
p(NG +i,0),(NG +i,0)
..
.
p(NG +i,Zt),(NG +i,0)
..
.
p(N0 ),(NG +i,0)
..
.
p(N ,X ),(NG +i,0)
T


(19)

..
.

..
.

..
.

..
.

..
.

p(NG ,0),(NG +i,Zi)

..
.
p(NG ,Z),(NG +i,Zi)
..
.
p(NG +i,0),(NG +i,Zi)
..
.
p(NG +i,Zt),(NG +i,Zi)
..
.
p(N0 ),(NG +i,Zi)
..
.
p(NX ),(NG +i,Zi)

|
|
|
|
|
|
|
|
|
|
|

(16)

process of replacement and repair of system components, and Tr can be treated as 0. When sm dissatisfies
the least demand (NG NM ) of repair, it must wait
for the lacking components which are repaired to finish the process of replacement of system components.
So, Tr is the lingering time which is produced by
waiting for the demanding spare parts,
and it just

relates to sm . Assume that Tc s , s , c is the lingering
time which is produced by waiting for the demanding
spare parts, when the number of demanding replaced
spare parts is s , the number of components that
wait for the repairing is s and the number of repair
teams is c at the starting time of component replacement period. Therefore, we can catch the computing
formula of E(Tr ).
When s c, the number of components that wait
for the repairing is less than the number of repair
teams, and the intervalof the next component that
has been repaired is 1 s . When s > c, the
number of components that wait for the repair is
larger than the number of repair teams, and the interval
 of the next component

 that has been repaired is
1 (c). And T c s , s , c equals to the addition of the

654

time repairing one component

and the repairing time

(Tc s 1, s 1, c ) when the number of demanding
spare parts is s 1and the number of components waiting for repairing is s 1. Thus, there is the following
recursive equation.
 

E Tc s , s , c =
(20)

Table 1. The values of the availability for a combination of

m and NG . m is chosen equal to 3.
Parameter of policy
Initial number of
spare parts
X
X
X
X

 

1
+ E Tc s 1, s 1, c
min {s , c}

=1
=2
=3
=4

(3,5)

(3,6)

(3,7)

0.8797
0.8974
0.9106
0.9106

0.8797
0.8974
0.9106
0.9106

0.7726
0.8359
0.8907
0.9106

On the basis of Equation (22), we can deduce

0. 94



0<s s c

0. 92

s > c and s s c

0. 9
Operationalavailability

s 1

(s h)

h=0


s ,
E[Tc (s , s , c)] = c

 

s c s s +c1 1

(ch) ,
c

h0

0,

s > c and s c < s s

others
(21)

For Ps (sm ), because the number of available components must be NM at the starting time of component
replacement period, Ps (sm ) equals to the steady probability when the state of system is (NM , sm ) at the
starting time of component replacement period. And
we have known 0 ss sm , so we can get
Ps (sm ) =

sm
N 


Pa (ns , ss ) p1 (ns , ss , NM , sm )

ns =NG ss =0

(22)
At the end, by the Equation (21) we can compute
to E(Tr ).
4

ANALYSIS OF AN EXAMPLE

Assume that there is a 2-out-of-7 hot standby redundancy system. The failed time of components has the
exponential distribution with rate = 0.005. There
is only one repair man (c = 1) to repair the failed
components replaced. And the repairing time also has
the exponential distribution with rate = 0.1. The
time of system down and in transit to depot is Td = 5.
And time of system in transit back is Ts = 5. When
we set the initial number (X) of spare parts from 1 to
5, the parameter of (m, NG ) maintenance policy corresponds to (3,6), (4,6), (3,7) and (4,7), we want to
know the result of operational availability. Following
the above method, the Table 1 and Figure 2 show the
date table and curve of operational availability when
we set m = 3 and change NG in the (m, NG ) maintenance policy. And the Table 2 and Figure 3 show

0. 88
0. 86
Initial number of s pare parts

0. 84

X=1
X=2
X=3
X=4

0. 82
0. 8
0. 78
0. 76

Figure 2. The values of the availability for a combination

of m and NG . m is chosen equal to 3.

the date table and curve of operational availability

when we set m = 4 and change NG in the (m, NG )
maintenance policy.
When NG = N = 7 the parameter is (3,7) or
(4,7) in (m, NG ) maintenance policy, from Figure 2
and Figure 3, we can find that in the condition of
fixed number of initial spare parts, no matterm = 4
or m = 3, the operational availability in (4,6) maintenance policy and (3,6) maintenance policy are larger
than the operational availability in m = 4 maintenance
policy and m = 3 maintenance policy.
But when the parameter m is fixed, the parameter NG is not always the smaller is the better. In the
example, when the initial number is X = 2, the
(4,6) maintenance policy is superior to the (4,5) maintenance policy. And when the initial number X < 4,
the (3,6) maintenance policy is superior to the (3,5)
maintenance policy.
The following shows the influence of m in (m, NG )
maintenance policy on the operational availability,
when the initial number changes. Table 3 and Figure 4
show the data table and curve of operational availability, when it fixes NG = 6 and changes m.
From Figure 4, we can find that when the initial
number X = 1, the parameter of best maintenance
policy is m = 3. When X = 2, the parameter of best

655

Table 2. The values of the availability for a combination of

m and NG . m is chosen equal to 4.

0. 96
Initial number of spare parts
0. 94

X=1
X=2
X=3
X=4

Initial number of
spare parts
X
X
X
X

=1
=2
=3
=4

(4,5)

(4,6)

(4,7)

0.9142
0.9249
0.9323
0.9382

0.8605
0.9250
0.9322
0.9382

0.7916
0.8351
0.8837
0.9268

Operational availability

Parameter of policy
0. 92

0. 9

0. 88

0. 86

0. 84

0. 82

0. 96
0. 94

Figure 4. The values of the availability for a combination

of m and NG . NG is chosen equal to 6.

0. 92
0. 9
0. 88

0. 86
Initial number of spare parts
0. 84

0. 95

X=1
X=2
X=3
X=4

0. 82

Operational availability

Oper at i onal avai l abi l i t y

0. 8
0. 78

Figure 3. The values of the availability for a combination

of m and NG . m is chosen equal to 4.

0. 9

0. 85

( 4, 6) maintenance policy
( 3, 6) maintenance policy
m=4 maintenance policy
m=3 maintenance policy

0. 8

0. 75

0. 7

Table 3. The values of the availability for a combination

of m and NG . NG is chosen equal to 6.

0. 65
1

Parameter of policy
Initial number of
spare parts
X
X
X
X

=1
=2
=3
=4

(2,6)

(3,6)

(4,6)

(5,6)

0.8238
0.8605
0.8609
0.8609

0.8797
0.8974
0.9106
0.9106

0.8605
0.9250
0.9322
0.9382

0.8636
0.9048
0.9500
0.9533

3
Initial number of spare parts

Figure 5. The values of the availability for a combination

of m, NG and initiation for different values of the number of
spares.

is the better. And it relates to the initial number of

spare parts.
3. When the parameter NG of (m, NG ) maintenance is
fixed, the parameter m of best maintenance policy
is increasing following the increasing of the initial
number of spare parts. And it is also not always the
bigger is the better.

maintenance policy is m = 4. And when X 3, the

parameter of best maintenance policy is m = 5.
From the above analysis of example, we can get the
following conclusions.
1. Given the initial number of spare parts, choosing the reasonable (m, NG ) maintenance policy can
improve the operational availability of k/N system. When we just consider the availability as the
optimization target, (m, NG ) maintenance policy is
superior to m maintenance policy.
2. When the parameter m of (m, NG ) maintenance is
fixed, the parameter NG is not always the smaller

For getting the best maintenance policy directly, we

can build the relational curve between the initial number of different maintenance policy and the operational
availability. Figure 5 shows the curve. From the figure,
we can see that when the initial number is 1, (3,6) maintenance policy is the best maintenance policy. When
the initial number is larger than 1, (4,6) maintenance
policy is the best.

656

CONCLUSIONS

This paper provides an operational availability model

when thek-out-of-N system takes the (m, NG ) maintenance policy with the given number of spare parts.
Through the analysis of an example, we can find that
the influence on operational availability which is produced by initial number of spare parts, repairing ability
and selection of mand NG is expressed clearly from
the model. And at the same time, the model can provides the decision support for the tradeoff between the
maintenance policy and the number of spare parts in
the different conditions.
REFERENCES
[1] Karin S. de Smidt-Destombes, Matthieu C. van der
Heijden, Aart van Harten. On the availability of a
k-out-of-N system given limited spares and repair
capacity under a condition based maintenance strategy[J]. Reliability Engineering and System Safety,
2004, 83:287300.

657

[2] Deng Zhong-yi. Applicability analysis for N Various

Components Parallel Connection and Maintainability
Systems[J]. Journal of Hunan Institute of Engineering.
2003, 13(2)(in chinese)
[3] Smith MAJ, Dekker R. Preventive maintenance in a
1-out-of-n system: the uptime, downtime and costs[J].
European Journal of Operations Research, 1997, 99:
56583.
[4] Fawzi BB, Hawkes AG. Availability of an R-out-ofN system with spares and repairs[J]. J Application
Probability, 1991, 28: 397408.
[5] Frostig E, Levikson B. On the availability of R out of N
repairable systems[J]. Naval Research Logistics, 2002,
49(5): 483498.
[6] Karin S. de Smidt-Destombes, Matthieu C. van der
Heijden, Aart van Harten. On the interaction between
maintenance, spare part inventories and repair capacity for a k-out-of-N system with wear-out. European
Journal of Operational Research. 2006,174: 182200
[7] Karin S. de Smidt-Destombes, Matthieu C. van der
Heijden, Aart van Harten. Availability of k-out-of-N
systems under block replacement sharing limited spares
and repair capacity, Int. J. Production Economics, 2007,
107: 404421.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

System value trajectories, maintenance, and its present value

K.B. Marais
Department of Industrial Engineering, Stellenbosch University, South Africa

J.H. Saleh
School of Aerospace Engineering, Georgia Institute of Technology, USA

ABSTRACT: Maintenance planning and activities have grown dramatically in importance and are increasingly
recognized as drivers of competitiveness. While maintenance models in the literature all deal with the cost
of maintenance (as an objective function or a constraint), only a handful addresses the notion of value of
maintenance, and seldom in an analytical or quantitative way. We propose that maintenance has intrinsic value
and argue that existing cost-centric models ignore an important dimension of maintenance, its value, and in so
doing, can lead to sub-optimal maintenance strategies. We develop a framework for capturing and quantifying
the value of maintenance activities. The framework presented here offers rich possibilities for future work in
benchmarking existing maintenance strategies based on their value implications, and in deriving new maintenance
strategies that are value-optimized.

INTRODUCTION

Maintenance planning and activities have grown

dramatically in importance across many industries.
This importance is manifested by both the significant material resources allocated to maintenance
departments as well as by the substantial number of personnel involved in maintenance activities in companiesfor example over a quarter of
the total workforce in the process industry is said
to deal with maintenance work [Waeyenbergh and
Pintelon, 2002]. This situation, coupled with an
increasingly competitive environment, creates economic pressures and a heightened need to ensure
that these considerable maintenance resources are
allocated and used appropriately, as they can be significant drivers of competitivenessor lack thereof if
mismanaged.
In response to these pressures, the notion of optimality and the mathematical tools of optimization
and Operations Research (OR) have seeped into maintenance planning, and resulted in the proliferation of
optimal maintenance models (see the reviews by
Pham and Wang (1996) and Wang (2002), for example). In each optimal maintenance model developed,
an objective function is first posited, then analytical
tools are used to derive a maintenance policy that
maximizes or minimizes this objective function subject to some constraints. For example, the objective
function can be the minimization of cost (cost rate,
or life cycle cost) of maintenance given a system

reliability and/or availability constraint; conversely,

the objective function can be the maximization of
reliability or availability, given a cost constraint. In
addition to varying the objective function, different
optimal maintenance models are obtained by: (1)
varying for example the system configuration (e.g.,
single-unit systems versus multi-unit systems, k-outof-n systems); (2) by including several degrees of
maintenance (e.g., minimal, imperfect, perfect); (3)
by varying the planning horizon; (4) by using different analytical tools; or (5) by positing different types of
dependencies between the various units in a multi-unit
system.
Yet, while all these models deal with the cost of
maintenance (as an objective function or a constraint),
only a handful of models touches on the notion of
value of maintenance, and seldom in an analytical
or quantitative way (e.g., Macke and Higuchi, 2007).
Wang (2002) highlights a critical idea for the development of a value-based perspective on maintenance
when he suggests that the cost of maintenance as well
as the resulting system reliability should be considered together when developing optimal maintenance
strategies. Where the benefits of maintenance are considered, it is usually in the sense of avoiding the costs
of failure. Interestingly, it is only within the civil
engineering community that the benefits in the sense
of service delivery are considered and cost-benefit
considerations explicitly taken into account in the
development of maintenance strategies (e.g., Macke
and Higuchi, 2007).

659

The argument for dismissing or not focusing on the

value of maintenance, when it is made, goes along
these lines: while it is easy to quantify the (direct) cost
of maintenance, it is difficult to quantify its benefits.
Dekker (1996) for example notes the main question
faced by maintenance management, whether maintenance output is produced effectively, in terms of
contribution to company profits, [ . . . ] is very difficult to answer. Therefore maintenance planning is
shifted from a value maximization problem formulation to a cost minimization problem (see Saleh (2008)
for a discussion of why these two problems are not the
same and do not lead to similar decisions in system
design and operation). Incidentally, in many organizations, maintenance is seen as a cost function, and
maintenance departments are considered cost centers
whose resources are to be optimized or minimized.
In short, as noted by Rosqvist et al., (2007) a costcentric mindset prevails in the maintenance literature
for which maintenance has no intrinsic value.
In this paper, we propose that maintenance has
intrinsic value and that one aspect of this value, the
net present value, can be captured. We argue that
existing cost-centric optimizations ignore an important dimension of maintenance, namely its value, and
in so doing, they can lead to sub-optimal maintenance
strategies. We therefore develop a framework built on
aspects of existing maintenance models for capturing
and quantifying the value of maintenance activities
by connecting an engineering and operations research
concept, system state, with a financial and managerial
concept, the present value (PV). Note that we consider
value as the net revenue generated by the system
over a given planning horizon. We do not consider
additional dimensions of value such as the potential
positive effects of maintenance on environmental or
health impacts. Such effects can be incorporated in
future work, see, for example, Marais et al., (2008)
for a discussion of the quantification of environmental and health impacts of aviation. The system state
refers to the condition of the system and hence its ability to perform and thereby provide a flow of service
(hence generate revenue, or quasi-rent). In order
to build this connection, we first explore the impact
of a systems state on the flow of service the system can provide over timefor a commercial system,
this translates into the systems revenue-generating
capability. Next we consider the impact of maintenance on system state evolution and hence value
generation capability over time. We then use traditional discounted cash flow techniques to capture the
impact of system state evolution with and without
maintenance on its financial worth, or PV. For simplification, we call the results of our calculations the value
of maintenance. Finally, we discuss the advantages
and limitations of our framework. This work offers
rich possibilities for assessing and benchmarking the

value implications of existing maintenance policies,

and deriving new policies based on maximizing value,
instead of minimizing the cost of maintenance.
2

BACKGROUND

This section provides a brief overview of various

maintenance models. The reader interested in extensive reviews of the subject is referred to the survey
papers by Dekker (1996), Pham and Wang (1996) and
Wang (2002). In the following, we discuss (1) maintenance classification, (2) maintenance models, and (3)
maintenance policies.
2.1 Types and degrees of maintenance
Maintenance refers to the set of all technical and
administrative actions intended to maintain a system
in or restore it to a state in which it can perform at least
part of its intended function(s) [Dekker, 1996].
Maintenance type can be classified into two main
categories: corrective maintenance and preventive
maintenance [Pham and Wang, 1996]. Corrective
maintenance (CM), also referred to as repair or runto-failure (RTF), refers to maintenance activities performed after a system has failed in order to restore its
functionality.
Preventive maintenance (PM) refers to planned
maintenance activities performed while the system
is still operational. Its aim is to retain the system
in some desired operational condition by preventing (or delaying) failures. Preventive maintenance is
further sub-divided into clock-based, age-based, and
condition-based, according to what triggers maintenance activities [Rausand and Hyland, 2004]:
Clock-based maintenance is scheduled at specific
calendar times; its periodicity is preset irrespective
of the systems condition (e.g., every Tuesday).
Age-based maintenance is performed at operating
time intervals or operating cycles of the system (e.g.,
every 500 on/off cycles, or every 4,000 hours of
flight).
Condition-based maintenance is triggered when the
measurement of a condition or state of the system
reaches a threshold that reflects some degradation and loss of performance of a system (but not
yet a failure). Condition-based maintenance is also
referred to as predictive maintenance.
Opportunistic maintenance encompasses both corrective and preventive maintenance and is relevant
for multi-unit systems with economic and functional
dependencies in which the failure of one unit, and
hence its corrective maintenance, offers an opportunity to perform preventive maintenance on other still
functional units.

660

Each type of maintenance can be further classified according to the degree to which it restores the
system [Pham and Wang, 1996]. At one end of the
spectrum, perfect maintenance restores the system to
its initial operating condition or renders it as good
as new. At the other end of the spectrum, minimal
repair returns the system to the condition it was in
immediately prior to failing (in the case of corrective
maintenance), or as bad as old. In between these
extremes lies imperfect maintenance, which returns
the system to a condition somewhere in between as
good as new and as bad as old. Finally, there is also
the possibility that maintenance leaves the system in
a worse condition than before the failure, through, for
example, erroneous actions such as damaging adjacent
parts while replacing a faulty unit.
2.2

Maintenance models

Models used to derive optimal maintenance policies

generally cover four main aspects [Dekker, 1996]: (1)
a description of the system being maintained; (2) a
model of how the system deteriorates and the consequences thereof; (3) a description of the available
information on the system and the available response
options; and (4) an objective function and an analytical
framework (or tools) according to which the optimal maintenance policy is to be derived. This section
reviews the four main classes of maintenance models, following the reviews in Pham and Wang (1996),
Doyen and Gaudoin (2004), Tan and Raghavan (in
press), and Wang (2002).
The first class of models developed considered the
possibility only for perfect or minimal repair [Nakagawa, 1979a, b; Pham and Wang 1996]. Thus,
following maintenance, the system is returned to as
good as new with some repair probability p, or to as bad
as old with probability (1-p). This basic concept is then
expanded to take into account time-dependent repair
probabilities, the possibility that maintenance causes
the system to be scrapped or to transition to some intermediate state, and non-negligible repair times (and
hence non-negligible downtime losses).
The second class of models considers maintenance
as improving the failure rate or intensity, and thus
allows the possibility of imperfect maintenance [Block
et al., 1985; Pham and Wang 1996]. It is assumed that
maintenance provides a fixed or proportional reduction in failure rate, or that it returns the system to the
failure rate curve at some time prior to the maintenance activity. Perfect maintenance returns the failure
rate to that of a new system; minimal maintenance
returns it to that of the system immediately prior to
the failure. The improvement factor is the degree of
improvement of failure rate. The improvement factor
is determined based on historical data, experiment,
expert judgment, or by assuming it correlates with

maintenance cost, system age, or the number of prior

maintenance activities [Malik, 1979; Pham and Wang
1996].
The third class of models views maintenance as
reducing the virtual age of the system [Kijima et al.,
1988]. It is assumed that maintenance reduces the age
of the system by some proportion (assuming increasing failure rate, which implies among other things
that the system exhibits no infant mortality). Perfect
maintenance returns the system virtual age to zero,
while minimal maintenance returns the virtual age to
the age immediately prior to the failure. Kijima et al.,
s (1988) original model allowed only a reduction to
the virtual age of the system following the previous
repair effort, though larger reductions in virtual age
can be seen as resulting from more extensive maintenance efforts. Pham and Wang (1996a) assume that
maintenance time increases with subsequent repairs
and consider the reduction in virtual age as decreasing
over timethat is, repairs become successively less
effective over time.
The fourth class of models considers system failures as manifesting as some level of damage or
degradation in response to a shock. These models are
therefore referred to as shock models. Perfect maintenance then reduces the damage to zero, minimal
maintenance returns the damage level to that immediately prior to the failure, and imperfect damage reduces
the damage by some factor greater than zero and less
than 100%. These models also allow the possibility for
less-effective and more expensive repairs over time by
making the reduction in damage a decreasing function
of time and by successively increasing the duration
of maintenance activities over time [Wang and Pham,
1996a, b].
In each case these models have been used primarily to derive maintenance policies that minimize cost
or downtime, or that maximize system availability,
as we discuss in the next sub-section. In Sections 3
and 4 we show how a simple model based on aspects
of these models can be used to quantify the value of
maintenance.
2.3 Maintenance policies
Maintenance policies describe what types of maintenance (repair, replacement, etc.) are considered in
response to what types of events (failure, calendar
time, machine cycles, etc.). In the following, we
confine our discussion to maintenance policies for
single-unit systems with increasing failure rates.
One popular maintenance policy is age-dependent
preventive maintenance where a system is repaired or
replaced at a pre-determined age [Wang, 2002]. The
triggering of maintenance in this case may be predetermined based on machine time (e.g., every 10,000
cycles) or on time elapsed since the last maintenance

661

activity. Under a random age-dependent maintenance

policy, maintenance is performed based on age and
system availability. This policy takes account of the
fact that systems may not be available for maintenance in the middle of a production run, for example.
A further extension of age-dependent replacement is
failure-dependent replacement where the system is
repaired in response to failures and replaced when a
given number of failures has occurred, or at a given
time, whichever occurs first [Nakagawa, 1984]. Many
other variations on the theme of age-dependent maintenance have been proposed; see Wang (2002) for an
extensive review.
An alternative family of maintenance policies,
referred to as periodic preventive maintenance, is
based on calendar time. Here maintenance occurs on
failure and periodically regardless of the failure or
operating history of the system [Wang, 2002]. Variations on this theme are developed by selecting degrees
of repair from minimal to perfect at specific times or in
response to failures. Further variations are developed
by incorporating the failure or operating history of the
system. For example, the level of maintenance may be
dependent on the number of previous repairs [Wang
and Pham, 1999].
Sequential preventive maintenance can be seen as
a variation of periodic PM where the interval between
PM activities is not constant. For example, the PM
interval may be decreased as the system ages, so
that the system does not exceed a certain operating
time without maintenance [Wang, 2002; Nguyen and
Murthy, 1981].

THE VALUE PERSPECTIVE IN DESIGN,

OPERATIONS AND MAINTENANCE:
A QUALITATIVE DISCUSSION

The present work builds on the premise that engineering systems are value-delivery artifacts that provide a
flow of services (or products) to stakeholders. When
this flow of services is priced in a market, this pricing or rent of these systems services allows the
assessment of the systems value, as will be discussed
shortly. In other words, the value of an engineering
system is determined by the market assessment of the
flow of services the system provides over its lifetime.
We have developed this perspective in a number of previous publications; for further details, the interested
reader is referred to for example Saleh et al., (2003),
Saleh and Marais (2006), or Saleh (2008).
In this paper, we extend our value-centric perspective on design to the case of maintenance. Our
argument is based on four key components:
1. First, we consider systems that deteriorate stochastically and exhibit multi-state failures, and we

model their state evolution using Markov chains

and directed graphs
2. Second, we consider that the system provides a
flow of service per unit time. This flow in turn is
priced and a discounted cash flow is calculated
resulting in a Present Value (PV) for each branch
of the graphor value trajectory of the system.
3. Third, given our previous two points, it is straightforward to conceive of the following: as the system
ages or deteriorates, it migrates towards lower PV
branches of the graph, or lower value trajectories.
4. Finally, we conceptualize maintenance as an operator (in a mathematical sense) that raises the system
to a higher PV branch in the graph, or to higher
value trajectory. We refer to the value of maintenance, or more specifically the Present Value (PV)
of maintenance, as the incremental Present Value
between the pre- and post-maintenance branches of
the graphs.
In the following section, we set up the analytical
framework that corresponds to this qualitative discussion.

MAINTENANCE AND PRESENT VALUE

BRANCHES

In developing our value model of maintenance, we

make a number of simplifying assumptions to keep
the focus on the main argument of this work. These
assumptions affect the particular mechanics of our calculations but bear no impact on the main results, as will
be shown shortly. Our assumptions are the following:
1. We restrict ourselves to the case of perfect maintenance; in addition we assume that maintenance
does not change the systems deterioration mechanism.
2. We restrict ourselves to the case of single-unit
systems.
3. We only consider systems that exhibit an increasing
failure rate. In other words, as our systems age, they
become more likely to deteriorate in the absence of
perfect maintenance.
4. The systems in our model can be in a finite number of discrete states, and the current state depends
only on the prior state, though the state transition
probabilities may be time-dependent. This assumption allows us to model the state evolution of the
system as a Markov process.
5. The systems in our model have no salvage value at
replacement or end of life
6. Finally, for simulation purposes, we consider discrete time steps, and assume that the duration of
maintenance activities is negligible compared with
the size of these time steps.

662

These assumptions will be relaxed in future work.

In the following, we consider first how a system
deteriorates under no-maintenance and introduce the
concept of value trajectories of the system. Next,
we show how maintenance moves the system onto
higher present value trajectories. Finally, we develop
expressions for the present value of maintenance.
We illustrate the framework by means of numerical
examples.

p11(2)

p11(1)
p12(1)

p11(0)

p22(1)

p22(2)

p23(1)

pnn(i)

pnd(i)

Deteriorated

Deteriorated

p23(2)

We consider a k-state discrete-time Markov deteriorating system with time-dependent transition probabilities as shown in Figure 1, for the no-maintenance
case with three states. The states are numbered from
1 through k in ascending order of deterioration where
state 1 is the new state and state k is the failed state.
The time-dependence allows us to take account of the
fact that a new (or deteriorated) system will become
more likely to transition to the deteriorated (or failed)
state as it ages (time-dependence implies dependence
on the virtual age of the system). With no maintenance the failed state is an absorbing state whence
it is not possible to transition to either of the other
states. Further, it is not possible to transition from the
deteriorated state to the new state without performing maintenance. In other words, the system can only
transition in one direction, from new to failed, perhaps via the deteriorated state (but the system has no
self-healing properties).

pnf(i)

pdd(i)

p13(0)

Bworst

Figure 2. Three-state system evolution over time with no

maintenance.

The transition matrix for a system with k states and

no self-healing is given by:
p (i) p (i) p (i)
11
12
1k
p22 (i) p2k (i)
0

P(i) =
..
..
..

.
.
.
0
0
0

pmn (0) pmn (i) pmn (i + j)

Failed

i, j 1
1m<nk

(2)

If we define 0 to be the initial probability distribution of the system, the probability distribution after
j state transitions is:

1
model

(1)

i is the index of the time step considered, and P(i) is

in effect P(ti ) in which ti = iT . For simplification
purposes, we retain only the index i in our notation.
The transition probabilities can be derived from
failure rates as shown by [Macke and Higuchi, 2007].
We represent the evolution of the system over time
using a directed graph, as shown in Figure 2 for a threestate system. This representation expands on Figure 1
and allows in effect an easy read of the time-dependent
transition probabilities, which is difficult to visualize using the traditional Markov chain representation
(Figure 1).
We assume that the probability of transitioning to a
lower state increases over time, and correspondingly
that the probability of remaining in a given state (other
than the failed state) decreases over time:
p11 (0) p11 (i) p11 (i + j)

pdf(i)

Figure 1. Three-state
maintenance.

New

p13(1)

Failed

Deterioration under no-maintenance,

and value trajectories

New

p22(2)
p23(2)

p12(0)

4.1

B1

p12(2)
p13(2)

of

system

with

no

j = Pj P2 P1 0

663

(3)

The likelihood of the system following a particular

branch over N steps is given by the products of the
transition probabilities along that branch:

For convenience we assume that the system is

initially in the new state, that is:
0 =

(4)

Next, we consider that the system can generate

um (t) revenue per unit time when it is in state m;
a degraded system having lower capacity to provide
services (hence generate revenues) than a fully functional system. This um (t) is the expected utility model
of the system or the price of the flow of service it
can provide over time. We discretize time into small
T bins over which um (t) can be considered constant.
Therefore


p Bj =
pi (Bj )

(7)

i=1

The right side of Eq. 7 is shorthand for the product

of the transition probabilities along the branch Bj .
Finally, the expected Present Value of the system
over all the branches is calculated by weighting the
Present Value of each branch by its likelihood:
PV (N ) =

p(Bj ) PV (N , Bj )

(8)

all branches

um (i) = um (iT ) um [(i + 1) T ]

(5)

To simplify the indexing, we consider that the revenues the system can generate between (i 1)T and
iT are equal to um (i) T .
Each branch of the graph represents a particular
value trajectory of the system, as discussed below.
Each time step, the system can remain fully operational, or it can transition to a degraded or failed
state. A branch in the graph is characterized by
the set of states the system can visit for all time
periods considered. For example, the branch B1 =
{1, 1, 1, 1, . . ., 1} represents the system remaining in
State 1 throughout the N periods considered, whereas
Bj = {1, 1, 1, 2, 2, . . ., 2} represents a system starting in State 1 and remaining in this state for the
first two periods, then transitioning to a degraded
state (here State 2) at the third period and remaining in this particular degraded state. Notice that the
branch Bworst = {1, k, k, . . ., k} represents a new
system transitioning to the failed state at the first
transition.
Since each state has an associated utility um (i),a
Present Value can be calculated for each branch over
N periods as follows:
N


PV N , Bj =

Using the Markov chain terminology, this present

value can also be expressed as:
PV (N ) =

(1 + rT )i

(9)

The following simple numerical example will help

clarify this discussion and equations.
Numerical example. Consider a three-state system
with the following transition matrix:

.95
P= 0
0

.04
.9
0

.01
.1
1

The system has a constant revenue model over N

periods, and can generate $100,000/year in State 1,
$60,000/year in the degraded State 2, and $0 in the
failed State 3. Assume a yearly discount rate of 10%.
Next, consider the branches of the system evolution
defined by the transitions in Table 1.
Applying Eq. 5 to this simple case, we obtain the
value trajectories and probabilities shown in Figure 3.
The expected Present Value of this system across
all branches is given by Eq. 7:
PV (N ) = $296, 672

uBj (i)

i=1

k
N 

um (i) i (m)
(1 + rT )i
i=1 m=1

(6)

uB (i) is shorthand for the revenues along the branch B.

rT is the discount rate adjusted for the time interval
T. Notice that Branch 1 has the highest Present Value
since the system remains fully functional throughout
the N periods (in State 1), and branch Bworst has the
lowest Present Value (PVworst = 0) since the system
starts its service life by failing!

664

In the next sub-section, we show how maintenance

moves the system onto higher-value trajectories and
therefore increases the expected present value of the
system.
4.2 Deterioration, Maintenance, and Present Value
Maintenance compensates for the lack of self-healing
in the system, and, from a modeling perspective, given
a three-state system, maintenance allows the system to

Table 1.

Branches for a three-state system evolution over four periods.

Branch

Transitions

B1
B4

{1, 1, 1, 1, 1}
{1, 1, 1, 2, 2}

Comment

The system starts in state 1 and remains in this state throughout the four periods.
The system starts in state 1; it remains in State 1 for two periods, then transitions to
the degraded State 2 in the third period and remains in this State 2.
{1, 1, 2, 2, 3} The system starts in State 1; it remains in State 1 for the first period, then transitions
to the degraded State 2 in the second period; it remains in this degraded state for
the third period, then transitions to the failed State 3 in the fourth period.

B8

PV of Branch 1

PV of Branch 4

PV of Branch 8

$300,000

Present Value

Perfect
maintenance

p(B 1 )=81.4%

$350,000

p(B 4 )=3.2%

$250,000
p(B 8 )=0.3%

$200,000
$150,000
$100,000

New

$50,000
$Period 1

Period 2

Period 3

Period 4

Perfect
maintenance

Figure 3. Illustrative value trajectories for a three-state

system (branches defined in Table 1).

j = Pj P2 P1 0

Deteriorated

Imperfect
maintenance

Failed

Figure 4. Performing perfect maintenance returns the system to the NEW state.
Reliability

move back, (1) from a failed state to a deteriorated state

(imperfect corrective maintenance), (2) from a failed
state to a new state (perfect corrective maintenance),
and (3) from a deteriorated state to a new state (perfect
preventive maintenance). These maintenance-enabled
transitions are shown in Figure 4.
In addition to returning the system to a higherfunctional state, maintenance provides another
advantage: it modifies the time-dependent transition
probabilities. In particular, the transition probabilities
from the new state after perfect maintenance are equal
to those of a new system. That is, performing perfect
maintenance returns the system to the initial transition
probabilities pij (0). Seen from a reliability viewpoint,
perfect maintenance returns the system to the initial
reliability curve, as shown in Figure 5. In the remainder of this work, we focus only on the case of perfect
maintenance.
Figure 6 shows the Markov model for our simple
system under the option of perfect maintenance. The
dotted lines indicate the new transition probabilities
assuming that perfect maintenance occurred at the end
of the previous time step.
Between maintenance activities the system deteriorates according to:

Perfect
maintenance

Perfect maintenance
shifts the reliability
curve to the right

Figure 5.

Impact of maintenance on reliabilit.

(10)

where j is the vector of probabilities of being in states

1 through k.

The effect of perfect maintenance is to return

the system to the initial probability distribution 0
and to the initial transition probability matrix P.

665

and (2) in the case of perfect maintenance, it restores

the initial (more favorable) transition probabilities,
which in effect ensure that system is more likely to
remain in the new state post-maintenance than premaintenance. The two effects of maintenance on the
value of a system are shown in Figure 7 for a threestate system. The following simple examples will help
further clarify this discussion.

p11(i)

p11(0)
New

p12(0)
p12(i)

p13(0) p13(i)

Deteriorated

Example 1: Perfect maintenance and the restoration of

the initial transition probabilities

p22(i)

p23(i)
Failed

Figure 6. Performing maintenance introduces new transition possibilities.

p12(1)
p13(1)

p22(1)
p23(1)

maintenance

p12(0)
p13(0)
p11(0)

p12(0)

p22(1)

New
Deteriorated

pM_2 (B1 ) = p11 (0) p11 (1) ... p11 (n 2)

Failed

and

Maintenance

pM_2 (B1 ) > pnoM_2 (B1 )

p22(2)

p13(0)

(12)

(13)

This result is easily extended for perfect maintenance occurring at the end of any period j:

p23(1)

te
main

nan

ce

p23(2)

pnoM_2 (B1 ) = p11 (2) p11 (3) ... p11 (n)

Whereas if perfect maintenance were performed on

the system at the end of the second period, the initial
transition probabilities are restored, and conditional
probability that the system will remain on B1 increases
and is given by:

p11(1)

p11(0)

Consider Branch B1 in the state evolution of the system

(see Figure 3 and Table 1); B1 as defined previously
corresponds to a system remaining in State 1 (or the
new state) throughout the time periods considered.
Under no maintenance, as the system ages, the likelihood that it will remain on B1 decreases. For example,
a system that remained on B1 until the end of the
second period, as shown in Figure 9, will find its conditional probability of remaining on B1 (given that it
was on B1 at the end of the second period) to be:

pM_j (B1 ) = p11 (0) p11 (1) ... p11 (n j)

and
Figure 7. Maintenance moves the system to a higher PV
branch and, in the case of perfect maintenance, restores the
initial transition probabilities.

(14)

pM_j (B1 ) > pnoM_j (B1 )

Figure 8 illustrates the system evolution over time
under the maintenance and no-maintenance policies.

Thus we can model the transition assuming maintenance at time tm > t0 as:

Example 2: The present value of maintenance

m1 = Pk P2 P1 0
m = 0

(11)

m+k = Pk P2 P1 0
In short, maintenance acts on two levers of value:
(1) it lifts the system to a higher value trajectory,

Consider the three-state system example discussed in

Section 4.1, and assume that the system transitions to a
degraded state at the end of the first period. If no maintenance is performed, the system will remain on the
lower value branches shown in Figure 9 (lower box).
The expected Present Value of these lower branches,
calculated at the end of the first period, is given by

666

Period 1

Period 2

Period 3

Period 4

This incremental Present Value, which in this case

it is calculated at the end of period 1, is what we define
as the Present Value of maintenance:

Period n+1
Time

p11(2)

Branch 1
without
maintenance

p11(n)

p11(3)

p11(1)

PV1 = PV maintenance_1 PV no_maintenance_1

p11(0)

p11(0)

p11(n-2)

p11(1)

Branch 1
with perfect
maintenance
occurring at the
end of the
second period

p11(1)

p11(0)

Perfect
maintenance
occurs here

= $76, 900

Figure 8. Perfect maintenance and the restoration of the

initial transition probabilities.

Period 1

Period 2

Period 3

Period 4
Time

Higher value (and more

likely) branches for the
maintained system

Maintenance lifts a
degraded system to
higher value branches

PV

maintenance_1

PV
provided by
maintenance

PV

Figure 9.
tenance.

Since maintenance is the only cause of the system

accessing these higher value trajectories after it has
degraded, it is appropriate to ascribe this incremental
Present Value to maintenance and term it the Present
Value of maintenance.

no_maintenance_1

Lower value branches of

the degraded system with
no maintenance

The incremental Present Value provided by main-

Eq. 8. Using the numerical values provided in the

previous example, we find:
PV no_maintenance_1 = $165, 600
Next, assume that perfect maintenance is carried out
at the end of period 1 after the system has degraded
to State 2. Perfect maintenance lifts the system up to
higher value trajectories, shown in the upper box in
Figure 10. The expected Present Value of these upper
branches, calculated at the end of the first period,
after the occurrence of perfect maintenance, is given
by Eq. 8. Using the numerical values provided in the
previous example, we find:
PV maintenance_1 = $242, 500
Finally, consider the difference between the
expected Present Value of the branches enabled by
maintenance and the branches without maintenance:
PV = PV maintenance PV no_maintenance

(15)

CONCLUSIONS

In this paper, we proposed that maintenance has

intrinsic value and argued that existing cost-centric
models ignore an important dimension of maintenance, namely its value, and in so doing, they can
lead to sub-optimal maintenance strategies. We argued
that the process for determining a maintenance strategy should involve both an assessment of the value
of maintenancehow much is it worth to the systems stakeholdersand an assessment of the costs
of maintenance.
We developed a framework for capturing and quantifying the value of maintenance activities by connecting an engineering and operations research concept,
system state, with a financial and managerial concept,
the Present Value (PV).
By identifying the impact of system state or condition on PV, the framework and analyses developed in
this paper provide (financial) information for decisionmakers to support in part the maintenance strategy
development. Maintenance, as a consequence, should
not be conceived as just an operational matter and
guided by purely operational matters but by multidisciplinary considerations involving the marketing,
finance, and operations functions within a company.
We are currently expanding this framework to derive
the Net Present Value of maintenance, defined as the
PV of maintenance minus the cost of maintenance.
In this work we made several simplifying assumptions. In particular, we considered value as the net
revenue generated by the system over a given planning
horizon. We did not include additional dimensions of
value such as the potential positive effects of maintenance on environmental or health impacts. Such
effects can be incorporated in future work. Further,
we did not include factors such as safety or regulatory
requirements. Such factors can be easily included as
constraints on the optimisation of value in future work.
One important implication of this work is that the
maintenance strategy should be tied to market conditions and the expected utility profile (ore revenue

667

generating capability of the system). In other words,

unlike traditional maintenance strategies, which are
fixed once devised, a value-optimal maintenance
strategy is dynamic and can change with environmental and market conditions.
Finally, we believe that the framework presented
here offers rich possibilities for future work in benchmarking existing maintenance strategies based on their
value implications, and in deriving new maintenance
strategies that are value-optimized.

REFERENCES
Block, H.W., Borges, W.S. and Savits, T.H. 1985. AgeDependent Minimal Repair, Journal of Applied Probability 22(2): 370385.
Brealy R, Myers C. 2000. Fundamentals of Corporate
Finance. 6th Ed. New York: Irwin/McGraw-Hill.
Dekker, Rommert. 1996. Applications of maintenance optimization models: a review and analysis. Reliability
Engineering and System Safety (51): 229240.
Doyen, Laurent and Gaudoin, Olivier, 2004. Classes of
imperfect repair models based on reduction of failure
intensity or virtual age. Reliability Engineering and
System Safety, 84: 4556.
Hilber, P., Miranda, V., Matos, M.A., Bertling, L. 2007.
Multiobjective Optimization Applied to Maintenance Policy for Electrical Networks. IEEE Transactions on Power
Systems 22(4): 16751682.
Kijima, Masaaki, Morimura, Hidenori and Suzuki, Yasusuke,
1988. Periodical replacement problem without assuming minimal repair. European Journal of Operational
Research 37(2): 194203.
Malik, M.A.K. 1979. Reliable Preventive Maintenance
Scheduling. AIIE Transactions 11(3): 221228.
Marais, Karen, Lukachko, Stephen, Jun, Mina, Mahashabde,
Anuja, and Waitz, Ian A. 2008. Assessing the Impact
of Aviation on Climate. Meteorologische Zeitung, 17(2):
157172.
Nakagawa, T. 1979a. Optimum policies when preventive
maintenance is imperfect. IEEE Transactions on Reliability 28(4): 331332.
Nakagawa, T. 1979b. Imperfect preventive-maintenance.
IEEE Transactions on Reliability 28(5): 402402.
Nakagawa, T. 1984. Optimal policy of continuous and discrete replacement with minimal repair at failure. Naval
Research Logistics Quarterly 31(4): 543550.

Nguyen, D.G., Murthy, D.N.P. 1981. Optimal preventive

maintenance policies for repairable systems. Operations
Research 29(6): 11811194.
Pham, Hoang and Wang, Hongzhou. 1996. Imperfect maintenance. European Journal of Operational Research, 94:
425438.
Rausand M., Hyland A. 2004. System Reliability Theory:
Models, Statistical Methods, and Applications. 2nd Ed.
New Jersey: WileyInterscience.
Rosqvist, T., Laakso, K. and Reunanen, M. 2007. Valuedriven maintenance planning for a production plant. Reliability Engineering & System Safety. In Press, Corrected
Proof.
Saleh, J.H. 2008. Flawed metrics: satellite cost per transponder and cost per day. IEEE Transactions on Aerospace and
Electronic Systems 44(1).
Saleh, J.H. 2008. Durability choice and optimal design
lifetime for complex engineering systems. Journal of
Engineering Design 19(2).
Saleh, J.H., Lamassoure, E., Hastings, D.E., Newman, D.J.
2003. Flexibility and the Value of On-Orbit Servicing: A
New Customer-Centric Perspective. Journal of Spacecraft
and Rockets 40(1): 279291.
Saleh, J.H., K. Marais, K. 2006. Reliability: how much is
it worth? Beyond its estimation or prediction, the (net)
present value of reliability. Reliability Engineering and
System Safety 91(6): 665673.
Sinden, J.A., Worrell, A.C. 1979. Unpriced values: Decisions
without market prices. New York: WileyInterScience.
Tan, Cher Ming and Raghavan, Nagarajan, In Press. A framework to practical predictive maintenance modeling for
multi-state systems. Reliability Engineering & System
Safety. Corrected Proof, Available online 21 September
2007. DOI: 10.1016/j.ress.2007.09.003.
Wang, Hongzhou. 2002. A survey of maintenance policies of
deteriorating systems. European Journal of Operational
Research 139: 469489.
Wang H.Z., Pham H. 1996. Optimal maintenance policies for
several imperfect repair models. International Journal of
Systems Science 27(6): 543549.
Wang, Hongzhou and Pham, Hoang. 1996. Optimal agedependent preventive maintenance policies with imperfect
maintenance. International Journal of Reliability, Quality
and Safety Engineering 3(2): 119135.
Wang H.Z., Pham H. 1999. Some maintenance models and
availability with imperfect maintenance in production
systems. Annals of Operations Research 91: 305318.
Waeyenbergh G, Pintelon, L. 2002. A framework for maintenance concept development. International Journal of
Production Economics 77(3): 299313.

668

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The maintenance management framework: A practical view to

maintenance management
A. Crespo Mrquez, P. Moreu de Len, J.F. Gmez Fernndez, C. Parra Mrquez & V. Gonzlez
Department Industrial Management, School of Engineering, University of Seville, Spain

ABSTRACT: The objective of this paper is to define a process for maintenance management and to classify
maintenance engineering techniques within that process. Regarding the maintenance management process, we
present a generic model proposed for maintenance management which integrates other models found in the
literature for built and in-use assets, and consists of eight sequential management building blocks. The different
maintenance engineering techniques are playing a crucial role within each one of those eight management building blocks. Following this path we characterize the maintenance management framework, i.e. the supporting
structure of the management process.
We offer a practical vision of the set of activities composing each management block, and the result of the
paper is a classification of the different maintenance engineering tools. The discussion of the different tools can
also classify them as qualitative or quantitative. At the same time, some tools will be very analytical tools while
others will be highly empirical. The paper also discusses the proper use of each tool or technique according to
the volume of data/information available.

THE MAINTENANCE MANAGEMENT

PROCESS

The maintenance management process can be divided

into two parts: the definition of the strategy, and the
strategy implementation. The first part, definition of
the maintenance strategy, requires the definition of
the maintenance objectives as an input, which will be
derived directly from the business plan. This initial
part of the maintenance management process conditions the success of maintenance in an organization,
and determines the effectiveness of the subsequent
implementation of the maintenance plans, schedules,
controls and improvements. Effectiveness shows how
well a department or function meets its goals or company needs, and is often discussed in terms of the
quality of the service provided, viewed from the customers perspective. This will allow us to arrive at
a position where we will be able to minimize the
maintenance indirect costs [3], those costs associated
with production losses, and ultimately, with customer
dissatisfaction. In the case of maintenance, effectiveness can represent the overall company satisfaction
with the capacity and condition of its assets [4], or
the reduction of the overall company cost obtained
because production capacity is available when needed
[5]. Effectiveness concentrates then on the correctness

of the process and whether the process produces the

required result.
The second part of the process, the implementation
of the selected strategy has a different significance
level. Our ability to deal with the maintenance management implementation problem (for instance, our
ability to ensure proper skill levels, proper work preparation, suitable tools and schedule fulfilment), will
allow us to minimize the maintenance direct cost
(labour and other maintenance required resources). In
this part of the process we deal with the efficiency
of our management, which should be less important. Efficiency is acting or producing with minimum
waste, expense, or unnecessary effort. Efficiency
is then understood as providing the same or better
maintenance for the same cost.
In this paper we present a generic model proposed
for maintenance management integrates other models found in the literature (see for instance [6,7]) for
built and in-use assets, and consists of eight sequential
management building blocks, as presented in Figure 1.
The first three building blocks condition maintenance
effectiveness, the fourth an fifth ensure maintenance
efficiency, blocks six and seven are devoted to maintenance and assets life cycle cost assessment, finally
block number eight ensures continuous maintenance
management improvement.

669

Effectiveness
Phase 1:
Definition of the
maintenance
objectives and
KPIs

Phase 8:
Continuous
Improvement
and new
techniques
utilization

Phase 7:
Asset life cycle
analysis
and replacement
optimization

Assessment

Figure 1.

Phase 2:
Assets priority
and maintenance
strategy definition

Improvement

Phase 6:
Maintenance
execution
assessment
and control

Effectiveness

Phase 3:
Immediate
intervention
on high impact
weak points

Phase 1:
Balance
Score Card
(BSC)

Phase 4:
Design of
the preventive
maintenance
plans and
resources

Phase 8:
Total Productive
Maintenance
(TPM),
e-maintenance

Phase 5:
Preventive plan,
schedule
and resources
optimization

Phase 7:
Life Cycle
Cost Analysis
(LCCA)

Efficiency

Phase 2:
Criticality
Analysis
(CA)

Improvement
Phase 4:
ReliabilityCentred
Maintenance
(RCM)

Phase 6:
Reliability
Analysis (RA)
& Critical Path
Method
(CPM)

Phase 5:
Optimization
(RCO)

Efficiency

Assessment

Maintenance management model.

Phase 3:
Failure Root
Cause Analysis
(FRCA)

Figure 2. Sample of techniques within the maintenance

management framework.

MAINTENANCE MANAGEMENT
FRAMEWORK

In this section, we will briefly introduce each block

and discuss methods that may be used to improve each
building block decision making process (see Figure 2).
Regarding the Definition of Maintenance Objectives and KPIs (Phase 1), it is common the operational
objectives and strategy, as well as the performance
measures, are inconsistent with the declared overall business strategy [8]. This unsatisfactory situation
can indeed be avoided by introducing the Balanced
Scorecard (BSC) [9]. The BSC is specific for the
organization for which it is developed and allows the
creation of key performance indicators (KPIs) for measuring maintenance management performance which
are aligned to the organizations strategic objectives
(See Figure 3).
Unlike conventional measures which are control
oriented, the Balanced Scorecard puts overall strategy
and vision at the centre and emphasizes on achieving performance targets. The measures are designed
to pull people toward the overall vision. They are
identified and their stretch targets established through
a participative process which involves the consultation of internal and external stakeholders, senior
management, key personnel in the operating units
of the maintenance function, and the users of the
maintenance service. In this manner, the performance
measures for the maintenance operation are linked to
the business success of the whole organization [10].
Once the Maintenance Objectives and Strategy are
defined, there are a large number of quantitative and

Maintenance
Cost Effectiveness

Maintenance cost (%)

per unit produced (7%)

Maintenance
planning
and scheduling

Quality

Learning

PM
Compliance
(98%)

Accomplishment
of criticality analysis
(Every 6 months)

Data integrity
(95%)

Figure 3.

From KPIs to functional indicators [2].

qualitative techniques which attempt to provide a systematic basis for deciding what assets should have
priority within a maintenance management process
(Phase 2), a decision that should be taken in accordance with the existing maintenance strategy. Most of
the quantitative techniques use a variation of a concept
known as the probability/risk number (PRN) [11].
Assets with the higher PRN will be analysed first.
Often, the number of assets potentially at risk outweighs the resources available to manage them. It is
therefore extremely important to know where to apply

670

+
1

Critical

Notice that assessing criticality will be specific to

each individual system, plant or business unit. For
instance, criticality of two similar plants in the same
industry may be different since risk factors for both
plants may vary or have different relative importance.
On some occasions, there is no hard data about historical failure rates, but the maintenance organization
may require a certain gross assessment of assets priority to be carried out. In these cases, qualitative methods
(see example in Figure 5) may be used and an initial

Figure 5.

Maintainability
Working
Time

Safety
Environment

Quality

Delivery

C
C

B.C

B.C

B.C

A.B

1. Define the purpose and scope of the analysis;

2. Establish the risk factors to take into account and
their relative importance;
3. Decide on the number of asset risk criticality levels
to establish;
4. Establish the overall procedure for the identification and priorization of the critical assets.

available resources to mitigate risk in a cost-effective

and efficient manner. Risk assessment is the part of the
ongoing risk management process that assigns relative
priorities for mitigation plans and implementation. In
professional risk assessments, risk combines the probability of an event occurring with the impact that event
would cause. The usual measure of risk for a class of
events is then R = PxC, where P is probability and C
is consequence. The total risk is therefore the sum of
the individual class-risks (see risk/criticality matrix in
Figure 4).
Risk assessment techniques can be used to prioritize assets and to align maintenance actions to
business targets at any time. By doing so we ensure that
maintenance actions are effective, that we reduce the
indirect maintenance cost, the most important maintenance costs, those associated to safety, environmental
risk, production losses, and ultimately, to customer
dissatisfaction.
The procedure to follow in order to carry out an
assets criticality analysis following risk assessment
techniques could be then depicted as follows:

Reliability

A.B

M
A.B

R
B.C

Criticality matrix and assets location.

Figure 4.

Consequence

A.B

10 20 30 40 50

B.C

Non-critical

B.C

Semi-critical

F
r
e
q
u
e
n
c
y

Qualitative criticallity assessment [2].

assets assessment, as a way to start building maintenance operations effectiveness, may be obtained. Once
there is a certain definition of assets priority, we have
to set up the strategy to be followed with each category
of assets. Of course, this strategy will be adjusted over
time, but an initial starting point must be stated.
As mentioned above, once there is a certain ranking of assets priority, we have to set up the strategy
to follow with each category of assets. Of course, this
strategy will be adjusted over time, and will consist
of a course of action to address specific issues for
the emerging critical items under the new business
conditions (see Figure 6).
Once the assets have been prioritized and the maintenance strategy to follow defined, the next step would
be to develop the corresponding maintenance actions
associated with each category of assets. Before doing
so, we may focus on certain repetitiveor chronic
failures that take place in high priority items (Phase 3).

671

Ensure certain
equipment availability
levels

RCM
Implementation phase

RCM
team
conformation

Maintenance strategy

Asset category

Initial
Phase

Reach optimal reliability,

maintainability and
availability levels

Operational
context
definition
and asset
selection

Criticality
Analysis
(level?)

Functional
failures

Function

FMEA
Failure Mode and
Effects Analysis

Failure modes

Effect of
failure modes

Tool to answer the first 5

RCM Questions
Tool to answer
the last 2
RCM Questions

Sustain improve
current situation

Final
Phase

Figure 6. Example of maintenance strategy definition for

different category assets [2].

Figure 7.

Finding and eliminating, if possible, the causes of

those failures could be an immediate intervention providing a fast and important initial payback of our
maintenance management strategy. The entire and
detailed equipment maintenance analysis and design
could be accomplished, reaping the benefits of this
intervention if successful.
There are different methods developed to carry out
this weak point analysis, one of the most well known
being root-cause failure analysis (RCFA). This method
consists of a series of actions taken to find out why a
particular failure or problem exists and to correct those
causes. Causes can be classified as physical, human or
latent. The physical cause is the reason why the asset
failed, the technical explanation on why things broke
or failed. The human cause includes the human errors
(omission or commission) resulting in physical roots.
Finally, the latent cause includes the deficiencies in the
management systems that allow the human errors to
continue unchecked (flaws in the systems and procedures). Latent failure causes will be our main concern
at this point of the process.
Designing the preventive maintenance plan for a
certain system (Phase 4) requires identifying its functions, the way these functions may fail and then
establish a set of applicable and effective preventive
maintenance tasks, based on considerations of system safety and economy. A formal method to do this
is the Reliability Centred Maintenance (RCM), as in
Figure 7.
Optimization of maintenance planning and scheduling (Phase 5) can be carried out to enhance the
effectiveness and efficiency of the maintenance policies resulting from an initial preventive maintenance
plan and program design.
Models to optimize maintenance plan and schedules will vary depending on the time horizon of
the analysis. Long-term models address maintenance capacity planning, spare parts provisioning

Application of
the RCM
logic

Maintenance
plan
documentation

RCM implementation process.

Optimality
Criteria

Equipment
status &
functional
dependencies

Failure
Dynamics

Monte Carlo
Model

PM
Schedule

Preventive
Maintenance
Plan
System
constraints

Figure 8.

Work in process

Obtaining the PM schedule.

and the maintenance/replacement interval determination problems, mid-term models may address, for
instance, the scheduling of the maintenance activities
in a long plant shut down, while short term models
focus on resources allocation and control [13]. Modelling approaches, analytical and empirical, are very
diverse. The complexity of the problem is often very
high and forces the consideration of certain assumptions in order to simplify the analytical resolution of
the models, or sometimes to reduce the computational
needs.
For example, the use of Monte Carlo simulation modelling can improve preventive maintenance
scheduling, allowing the assessment of alternative
scheduling policies that could be implemented dynamically on the plant/shop floor (see Figure 8).
Using a simulation model, we can compare and discuss the benefits of different scheduling policies on the
status of current manufacturing equipment and several operating conditions of the production materials
flow. To do so, we estimate measures of performance
by treating simulation results as a series of realistic

672

CAPEX
Capital Costs
Development
costs

Investment
costs

OPEX
Operational Costs
Operation
costs

Conventional Maintenance

E-maintenance

Top Management

Top Management

Acquisition

Reports

Corrective Maintenance + Security, Environment, Production =

Non Reliability Costs = Risk

Design

Middle Management

Operation + Planned Maintenance Costs.

Middle Management

Investigation

Reports

Login to
iScada

Maintenance Dept
Construction

Precise &
Concise
Information

Remove

Inspections/Complaints

Time (years)

Figure 9.

Maintenance Dept

Assets /
Information Source

Life cycle cost analysis.

experiments and using statistical inference to identify

reasonable confidence intervals.
The execution of the maintenance activitiesonce
designed planned and scheduled using techniques
described for previous building blockshas to be
evaluated and deviations controlled to continuously
pursue business targets and approach stretch values for
key maintenance performance indicators as selected
by the organization (Phase 6). Many of the high
level maintenance KPIs, are built or composed using
other basic level technical and economical indicators. Therefore, it is very important to make sure
that the organization captures suitable data and that
that data is properly aggregated/disaggregated according to the required level of maintenance performance
analysis.
A life cycle cost analysis (Phase 7) calculates the
cost of an asset for its entire life span (see Figure 9).
The analysis of a typical asset could include costs
for planning, research and development, production,
operation, maintenance and disposal. Costs such as
up-front acquisition (research, design, test, production, construction) are usually obvious, but life cycle
cost analysis crucially depends on values calculated
from reliability analyses such us failure rate, cost
of spares, repair times, and component costs. A life
cycle cost analysis is important when making decisions
about capital equipment (replacement or new acquisition) [12], it reinforces the importance of locked
in costs, such as R&D, and it offers three important
benefits:
All costs associated with an asset become visible. Especially: Upstream; R&D, Downstream;
Maintenance;
Allows an analysis of business function interrelationships. Low R&D costs may lead to high
maintenance costs in the future;
Differences in early stage expenditure are highlighted, enabling managers to develop accurate
revenue predictions.
Continuous improvement of maintenance management (Phase 8) will be possible due to the utilization
of emerging techniques and technologies in areas that

Assets /
Information Source

Figure 10. Implementing e-maintenance (http://www.

devicesworld.net).

are considered to be of higher impact as a result of the

previous steps of our management process. Regarding the application of new technologies to maintenance, the e-maintenance concept (Figure 10) is
put forward as a component of the e-manufacturing
concept [14], which profits from the emerging information and communication technologies to implement
a cooperative and distributed multi-user environment.
E-Maintenance can be defined [10] as a maintenance
support which includes the resources, services and
management necessary to enable proactive decision
process execution.
This support not only includes etechnologies (i.e.
ICT, Web-based, tether-free, wireless, infotronic technologies) but also, e-maintenance activities (operations or processes) such as e-monitoring, e-diagnosis,
e-prognosis . . . etc. Besides new technologies for
maintenance, the involvement of maintenance people
within the maintenance improvement process will be
a critical factor for success. Of course, higher levels of knowledge, experience and training will be
required, but at the same time, techniques covering the
involvement of operators in performing simple maintenance tasks will be extremely important to reach higher
levels of maintenance quality and overall equipment
effectiveness.

CONCLUSIONS

This paper summarizes the process (the course of

action and the series of stages or steps to follow)
and the framework (the essential supporting structure
and the basic system) needed to manage maintenance.
A set of models and methods to improve maintenance
management decision making is presented. Models
are then classified according to their more suitable
utilization within the maintenance management process. For further discussion of these topics the reader
is addressed to a recent work of one of the authors [2].

673

REFERENCES
[1] EN 13306:2001, (2001) Maintenance Terminology.
European Standard. CEN (European Committee for
Standardization), Brussels.
[2] Crespo Marquez, A, (2007) The maintenance management Framework. Models and methods for complex
systems maintenance. London: Springer Verlag.
[3] Vagliasindi F, (1989) Gestire la manutenzione. Perche
e come. Milano: Franco Angeli.
[4] Wireman T, (1998) Developing performance indicators for managing maintenance. New York: Industrial
Press.
[5] Palmer RD, (1999) Maintenance Planning and
Scheduling. New York: McGraw-Hill.
[6] Pintelon LM, Gelders LF, (1992) Maintenance management decision making. European Journal of Operational Research, 58: 301317.
[7] Vanneste SG, Van Wassenhove LN, (1995) An integrated and structured approach to improve maintenance. European Journal of Operational Research, 82:
241257.
[8] Gelders L, Mannaerts P, Maes J, (1994) Manufacturing strategy, performance indicators and

674

[9]
[10]

[11]
[12]
[13]

[14]

improvement programmes. International Journal of

production Research, 32(4): 797805.
Kaplan RS, Norton DP, (1992) The Balanced Scorecardmeasures that drive performance. Harvard
Business Review, 70(1): 719.
Tsang A, Jardine A, Kolodny H, (1999) Measuring maintenance performance: a holistic approach.
International Journal of Operations and Production
Management, 19(7): 691715.
Moubray J, (1997) Reliability-Centred Maintenance
(2nd ed.). Oxford: Butterworth-Heinemann.
Campbell JD, Jardine AKS, (2001) Maintenance
excellence. New York: Marcel Dekker.
Duffuaa SO, (2000) Mathematical models in maintenance planning and scheduling. In Maintenance,
Modelling and Optimization. Ben-Daya M, Duffuaa
SO, Raouf A, Editors. Boston: Kluwer Academic
Publishers.
Lee J, (2003) E-manufacturing: fundamental,
tools, and transformation. Robotics and ComputerIntegrated Manufacturing, 19(6): 501507.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Workplace occupation and equipment availability and utilization,

in the context of maintenance float systems
I.S. Lopes
School of Engineering, University of Minho, Braga, Portugal

A.F. Leito
Polytechnic Institute of Bragana, Bragana, Portugal

G.A.B. Pereira
School of Engineering, University of Minho, Guimares, Portugal

ABSTRACT: In industry, spare equipments are often shared by many workplaces with identical equipments
to assure the production rate required to fulfill delivery dates. These types of systems are called Maintenance
Float Systems. The main objective of managers that deal with these types of systems is to assure the required
capacity to deliver orders on time and at minimum cost. Not delivering on time has often important consequences;
it can cause loss of costumer goodwill, loss of sales and can damage organizations image. Maintenance cost is
the indicator more frequently used to configure maintenance float systems and to invest in maintenance workers
or spare equipments. Once the system is configured, other performance indicators must be used to characterize and measure the efficiency of the system. Different improvement initiatives can be performed to enhance
the performance of maintenance float systems: performing preventive maintenance actions, implementation of
autonomous maintenance, improvement of equipments maintainability, increase of maintenance crews efficiency etc. Carrying out improvement based on facts is a principle of Total Quality Management (TQM)
in order to step to business excellence. It requires monitoring processes through performance measures. This
work aims to characterize and highlight the differences and relationships between three types of performance
measuresequipment availability, equipment utilization and workplace occupation, in the context of maintenance float system. Definitions and expressions of these three indicators are developed for maintenance float
systems. The relationship between maintenance float systems efficiency and the referred indicators is shown.
Other indicators are also proposed and compared with the first ones (number of standby equipments, queue
length etc.).

INTRODUCTION

In the past, market demand was very high. It was

required that producers delivered products as quicker
as possible and without defects. Nowadays, within a
competitive market, producers must deliver on time
products that satisfy client requirements. Once client
requirements are always changing, producers have
to constantly redesign and innovate their products.
Small product series are more frequent and require
flexible and reliable technologies. The production
dependence on robots, automatic systems and transport systems makes maintenance an important and
key function in production systems. Some authors
show that maintenance is no longer a cost center, but
should be regarded as a profit generation function
(Alsyouf 2006, Alsyouf 2007) and must be integrated

in the organizations strategic planning (Madu 2000)

Alsyouf (2007) illustrates how an effective maintenance policy could influence the productivity and profitability of a manufacturing process trough its direct
impact on quality, efficiency and effectiveness of
operation. Madu (2000) aimed to demonstrate that reliability and maintainability are crucial to the survival
and competitiveness of an organization, especially
with the rapid proliferation of technologies.
Several methodologies and models concerning
maintenance management can be found in literature,
such as: Reliability Centered Maintenance (RCM),
Total Productive Maintenance (TPM), preventive
maintenance models including predictive maintenance
and inspection models.
Whenever breakdowns are inevitable, the utilization of spare equipments is frequent to minimize

675

units in standby

workstation

in queue

be attended by
maintenance servers

Figure 1.

Maintenance Float System representation.

undesirable effects in production, caused by downtimes. Spares can be efficiently managed when identical equipments are operating in parallel in the
workstation. This type of system is called a Maintenance Float System. Float designates equipments
in standby and equipments waiting for maintenance
actions in the maintenance center. Equipment or
unit involved in a Maintenance Float System (MFS)
switches among different states: operating in workstation, waiting in the queue to be repaired, being repaired
in the maintenance center, waiting until required by the
workstation (see fig. 1).
Some studies present mathematical and simulation
models to configure MFS. One of the first attempts
to determine the number of float units was proposed
by Levine (1965) who uses analytical method based
on traditional reliability theory. The author introduced
a reliability factor based on the ratio MTTR/MTBF.
Gross et al., (1983), Madu (1988), Madu & Kuei
(1996) use Buzens algorithm. Zeng & Zhang (1997)
consider a system where a key unit keeps the workstation functioning and a set of identical units are kept in
a buffer to replace units sent for repairing. The system
is modeled as a closed queue (an M/M/S/F queue),
and the idle probability of the system is obtained. The
optimal values of the capacity of the inventory buffer
(F), the size of repair crew (S) and the mean repair rate
are determined by minimizing the total cost Shankar &
Sahani (2003) consider a float system whose failures
are classified as sudden and wear-out. Units subject
to wear-out failures are replaced and submitted to
preventive maintenance actions after a specific time
period. Based on the reliability function of the system,
authors find the number of floating units needed to
support the active units such that the number of active
units does not change. Most of the studies based on

simulation methods approach the design problem of

MFS through the development of meta-models (models that express the input-output relationship in the
form of a regression equation). Madu (2000) used
Taguchis techniques to construct the meta-model.
Chen & Tseng (2003) used Neural Networks.
Maintenance cost is the indicator more frequently
used to configure MFS and to hire maintenance workers or to invest on spare equipments (Zeng & Zhang
1997; Madu & Kuei 1996; Madu 1999; Chen & Tseng
2003). Lopes et al., (2006) present a cost model to
determine the number of float units, the number of
maintenance crews in the maintenance center and the
time between periodic overhauls. Periodic overhauls
are performed to improve equipments reliability and
to minimize the number of breakdowns.
Once the system is configured, other performance
indicators must be used to characterize and measure
the efficiency of the system and to identify the potential improvement initiatives that can be implemented.
Gupta & Rao (1996) present a recursive method to
obtain the steady-state probability distribution of the
number of down machines of a MFS. A M/G/1 queue
is considered with only one repairman. Gupta & Rao
(1996) use several performance measures to evaluate
the efficiency of MFS: the average number of down
machines, the average number of machines waiting in
queue for repair, the average waiting time in queue, the
average number of operating machines, the machine
availability and the operator utilization. Each state of
the system is characterized by the number of failed
units. Gupta (1997) deals with the same queue model
but considers that the server takes a vacation of random duration every time the repair facility becomes
empty.
Lopes et al. (2007) also determine state probabilities of a float system submitted to preventive
maintenance at periodic overhauls and show the effect
of performing overhauls in equipments involved in a
MFS. Each state of the system is defined by the number of failed units (i) and by the number of equipment
submitted or waiting for an overhaul ( j). Lopes et al.,
(2007) conclude that periodic overhauls optimize the
efficiency of maintenance crews in the maintenance
center. Like Gupta & Rao (1996) several maintenance indicators are defined and determined (average
queue length, probability of waiting queue occurrence,
average number of units not yet replaced etc.).
This work presents and defines three types of
maintenance indicators for MFS: equipment availability, equipment utilization and workplace occupation.
Workplace occupation and other performance measures are determined for a MFS submitted to periodic
overhauls.
This paper is organized as followssection 2
discusses equipment availability. Section 3 studies equipment utilization. In section 4, workplace

676

occupation is defined. Section 5 and 6 present the

indicators for a MFS submitted to periodic overhauls.
Section 7 presents the conclusions.

EQUIPMENT AVAILABILITY

Active
in workstation

Figure 2.

Availability represents the fraction of time that a

repairable equipment is operational. An equipment
can be operational and not be active (operating). The
expression often used is given by equation 1.
A=

Tup
,
Tup + Tdown

MTBF
MTBF + MTTR

be attended in
maintenance center

(2)

Equation 2 assumes that whenever the equipment

is up (operational), it is operating and when it fails the
repair action is immediately started. If equipment does
not start immediately, MTTR must include waiting
time until repair action starts.
Equation 2 is often misused. MTBF and MTTR do
not always use the same measure units. Usually MTTR
is associated with time to repair, but it is very frequent
that MTBF uses measures associated with the phenomena of equipment degradation, such as number
of kilometers or number of utilizations, etc. Availability is a performance indicator which is related to
equipment reliability and maintainability. A higher
availability can be achieved if the equipment reliability could increase or if the repair process could
be faster. It can involve enhancing the repair crew
capacity by adopting new work methods or improving skills through training, or by introducing changes
in equipment that could improve its maintainability or
reliability.
In the context of MFS, the time spent in the
maintenance queue need to be incorporated in the
expression for availability. In this environment, equipment availability is influenced by the efficiency of the
maintenance center which can be characterized by the
respective queue length. The up time (fig. 2) is the sum
of the equipment operating time in the workstation and
the standby time.
Availability = (time in workstation + standby
time)/(time in workstation + time in queue + time
to repair + standby time)

in standby

Cycle time of a unit involved in MFS.

Active
in workstation

Waiting
in the queue

be attended in
maintenance center

(1)

where Tup represents the total time the equipment is

up (operational) and Tdown represents the total time
the equipment is down (not operational), thus meaning
the equipment can not be used. For equipment in the
steady-state, the equation 2 is also frequently used:
A=

Waiting
in the queue

in standby

Figure 3.

Equipment utilization in MFS.

In MFS, if the availability increases, production

capacity increases. For the same level of capacity, spare equipments will not be required so often.
The realization of preventive actions reduces downtime and then increases the availability of equipments
in the workstation. The equipment availability is also
influenced by time to repair. In this respect, different
techniques could be used to reduce time to perform
repairs or to improve maintainability of equipments.
Availability determination requires quantifying
standby time which is complex to determine analytically. Instead of this indicator, the average number
of non-operational equipments could be used to provide similar information. In section 6.1, the average
number of equipments waiting for or under maintenance actions is determined for a MFS, submitted to
periodic overhauls.

EQUIPMENT UTILIZATION

Equipment utilization characterizes the efficiency of

MFS. A high utilization of an equipment means that
spare equipments are often used, thus cost-effective.
It means that equipment spent relatively short time in
standby state, in the maintenance queue and in the
repair station. If the system had been well configured,
equipments involved in a MFS will have its utilization value very close to its availability value. Figure 3
illustrates equipment utilization in a MFS.
Utilization rate = (time in workstation)/(time in
workstation + time in queue + time to repair +
standby time)
The efficiency of MFS could also be assessed
through the average number of equipments in standby
and through the average number of equipments in
queue. These last indicators are easier to determine
analytically and provide similar information. They
also allow the identification of the most critical source

677

for time loss (standby time and time spent in the

queue).

WORKPLACE OCCUPATION

Workplace occupation depends on time between failures and time until replacement of the failed unit
(fig. 4).
Workplace occupation = time in workplace /operating cycle
This operating cycle (Dw ), however, does not correspond to the cycle defined for equipment availability
and utilization calculation purposes. This operating
cycle (fig. 5) begins when an unit starts operating in
the workstation and ends when another unit takes its
place. For MFS, the new unit is different from the
failed one and the replacement is performed before
the conclusion of initial unit repair.
Dw = time in workstation + time until replacement
of failed unit occurs
If the replacement of failed equipment is always
immediate, then the workstation occupation will be 1,
meaning that the workplace is always available when
required. The time to replace equipment is neglected.
The time until a new unit starts operating in the
workplace depends on time to repair and on the number
of spare equipments. Workplace occupation is related
with equipment availability and utilization. Equipment
utilization can be low due to time to repair or due to
standby time. In the first case, workplace occupation is

also low. In the second case, the workplace occupation

is high. In this case, the average number of equipments
in standby can give an idea of the equipment utilization
and also of the workplace occupation. If the average
number of equipments in standby is high, then the
workplace occupation will also be high, once spare
equipments availability is high.
5

WORKPLACE OCUPATION FOR A MFS

SUBMITTED TO PERIODIC OVERHAULS

Based on the model developed in Lopes et al. (2007),

workplace occupation is determined. Lopes et al.
(2007) present a preventive maintenance model with
replacement at constant time intervals or upon failure
for a network with load-dependent maintenance service. Each system state is represented by (i, j), where i
is the number of failed equipments and j is the number
of equipments waiting or under overhaul. Pi,j denotes
the state probability for the MFS. The queue for the
maintenance center follows FIFO (First In, First Out)
discipline. Failed units form a queue and are replaced
while and when spare units are available. Equipment
that needs an overhaul is kept in operation waiting
for its turn to be attended and the replacement will
take place when the overhaul is initiated. If a unit that
needs an overhaul fails before being attended, it will
be replaced if a spare unit is available. In this case, this
unit is treated like a failed unit.
Lopes et al. (2007) address both problems:
number of spare equipments (R) is bigger than the
number of maintenance crews (L); (R L).
number of spare equipments (R) is lower than the
number of maintenance crews (L); (R L).
In this work, the expression for workplace occupation will be determined for the first problem above.

workplace 1

workplace 2

......

workplace M

5.1 Distribution of time to failure

The workstation comprises M identical workplaces
and time to failure for an active equipment in the workstation follows an Exponential distribution f (t) with
its parameter given by 1/f , where f designates the
equipment failure rate.

Workstation

Figure 4.

Replacing units in workstation.

5.2

cycle

Considering that time interval between two successive conclusions of maintenance actions follows the
Exponential distribution:

time until replacement

new unit
starts operating

failure of the
operating unit

new unit
starts operating

Figure 5.

g(t) = et ,
t

Cycle time for a workplace.

Distribution of time to replace an equipment

(3)

where is the service rate of the maintenance crews,

which activity is divided into repair and overhaul

678

actions with a service rate of a and b respectively.

can be determined based on the repair rate (rep ),
on the overhaul rate (rev ) and on the number of
maintenance crews (L), see equation 4.

= Larep + Lbrev

a) Machine ready to be overhauled

ALAR

(4)

t r et
(r + 1)

(ALAR) F

t r+1
0

r+1
t r et
dt =
,
(r + 1)

ALAR

t2

z2

Figure 6.

ALAR

b2

Failure
Failed machine in queue

Cycle time for L < R.

5.3.1 Cycle time

a) Cycle time for the workplace when the operating
equipment is overhauled At the end of the overhaul
interval [0, T ), MFS can be in several different
states. Depending on the system state, different
situations could arise:
There is at least one spare unit available to perform the replacement (AL AR ): the replacement
is made immediately and the overhaul is initiated.
There is at least one spare unit available to
perform the replacement, but there is no maintenance crew available to initiate the overhaul
(equipment will wait active until being replaced)
(AL AR ): if the equipment does not fail while
waiting for an overhaul (NF), the replacement is
made only when a maintenance crew is available,
following the first come first served discipline. If
the equipment fails (F) meanwhile, it is treated as
a failed equipment and will be replaced as soon
as possible.
There is no spare unit and no maintenance crew
available (AL AR ): this situation is addressed
like the previous one.

(7)

AR at least a spare equipment is available;

AR no spare equipment is available;
AL at least one maintenance crew is available;
AL no maintenance crew is available;
F equipment which is waiting for an overhaul fails
before replacement
NF equipment which is waiting for an overhaul
does not fail.

t
X

Active machine

(6)

Based on figure 6, the expression of Tup and D could

be deductedup time is represented by a continuous
line and downtime is represented by a dashed line. At
the left side of figure 6 is represented the cycle time
for the workplace when operating equipment reaches
the end of the overhaul interval and is ready to be
overhauled and at the right side is represented the cycle
time when the operating equipment fails.
Some possible events are considered:

a2

(ALAR) F

ALAR

Average Occupation rate of a workplace

Tup
D

z1

b) Failed machine

The average occupation rate of a workplace (Q) is

determined based on the time interval an equipment
is operating in the workplace (Tup ).
Considering a cycle with a duration D, Q is given
by equation 7.
Q=

(5)

which is the mean of the Gamma distribution. When

r + 1 is an integer, W (t) is designated by Erlang
distribution and (r + 1) = r!
5.3

t1

(ALAR) NF

The average waiting time is given by equation 6.



a1

(ALAR) NF

The time distribution for an equipment that needs

to wait for r + 1 releases of the maintenance center
to be replaced is a Gamma distribution (5). The sum
of independent random variables following Exponential distributions follows a Gamma distribution (see
equation 5).
Wr+1 (t) = r+1

b) Cycle time when the operating equipment fails

When the operating equipment in the workplace
fails, the following situations need to be considered
to identify the time to replace the equipment:

679

T

There is at least one spare unit available (AL AR

AL AR = AR ): the time to replace the equipment
is null (neglecting the time to make the change in
the workplace).
There is no spare unit available (AR AL ): The
average time to replace the equipment could be
determined based on Gamma distribution.
The average duration of a cycle D is given by the
following equation:

T

f (t) P(AL AR ) t + P(AL AR ) t


+P(AL AR ) t dt

Simplifying:

Tup

P(AL AR ) T + P(AL AR ) PNF [T + a1 ]

+P(AL AR ) PF [T + t1 + z1 ]
D = F(T)

P(AL AR ) PNF [T + a2 ]

+P(AL AR ) PF [T + t2 + Zz ]
+

T + P(AL AR ) PNF a1

= F(T) +P(AL AR ) PF t1 + P(AL AR )

PNF a2 + P(AL AR ) PF t2
T
f (t) t dt

+
0

f (t) P(AL AR ) t + P(AL AR ) t


+ P(AL AR ) [t + b2 ] dt

5.4 Average time to replace an equipment

waiting for an overhaul
Before replacing an equipment that is waiting for
an overhaul, all equipments that are already in the
queue have to be replaced first. The number of equipments in the queue is equal to i + j L, i.e., the
number of failed equipments and the number of equipments waiting for an overhaul minus the number of
equipments being attended. Then, the average time to
replace an equipment requiring an overhaul is given
by equation 8.

Simplifying:

D = F(T)

T
+

T + P(AL AR ) PNF a1

+P(AL AR ) PF [t1 + z1 ] +

P(AL AR ) PNF a2

+P(AL AR ) PF [t2 + z2 ]

f (t) t + P(AL AR ) b2 dt

=
a1 = a2

(i + j L) + 1
u

(8)

5.5 Time to replace a failed equipment

For L < i + j < R,

where
P(AL AR ) =

(i + j L) a is the average number of failed

equipments in the queue;
(R L) is the number of available spare equipments
at the instant where maintenance crews become all
occupied.

Pi,j

i+j+1L

P(AL AR ) =

Pi,j

L<i+j+1R

P(AL AR ) =

Pi,j

i+j+1>R

where F(T) is the probability of equipment reaching

the end of the overhaul interval without a failure.
5.3.2 Up time in workstation

Tup = F(T)

P(AL AR ) T + P(AL AR ) PNF [T + a1 ]

+P(AL AR ) PF [T + t1 ]

P(AL AR ) PNF [T + a2 ]

+P(A A ) P [T + t ]
R
F
2
L

If (i + j L) a + 1 R L, the failed equipment

i + j + 1 can be immediately replaced, once there is
spare equipment available.
If (i + j L) a + 1 > R L, the failed equipment i + j + 1 can not be replaced immediately (there
is no spare equipments available). Released equipments will replace failed equipments or equipments
waiting for an overhaul depending of the maintenance
action that is initiated. If the maintenance action is
an overhaul, the released equipment will replace the
equipment submitted to an overhaul. If the maintenance action is a repair, the released equipment will
replace the next non-replaced failed equipment which
is in the queue.

680

Once max[0; (i + j L) a + 1 (R L)] is the

number of non-replaced failed equipments, then the
failed equipment i + j + 1 will be replaced after the
release of max[0; (i + j L) a + 1 (R L)]/a
equipments from the maintenance center (number of
non-replaced failed equipments divided by the fraction
of failed equipments in the queue).
The time elapsed until the replacement is given by
equation 9.
b2 =

5.6

max [0; (i + j L) a + 1 (R L)]

a

(9)

Probability of failure of an equipment

waiting for an overhaul

The probability of failure PF is determined based on

Erlang distributionthe time distribution for replacing an equipment waiting for an overhaul, and on
Exponential distributionthe time distribution for
failures of an active equipment.
  t
PF =

r+1
0

(10)


r+1

=
0

t r et

r!

t

f ef t2 dt2 dt


=

t r et
f ef t2 dt2 dt
r!

r+1


t r et
1 ef t dt
r!



h+ 1 r t
t e
dt t r e(+f )t dt
=
r!
=

r!

r+1

r!
r!

r+1
( + f )r+1

(see equation 12).

 t
tv =

r+1

0 0

f ef d dt
(11)

PF

t1 et2 have the same expressions, once maintenance

rates and expressions for the number of releases of
maintenance center until overhaul be initiated (r =
i + j L) are identical.
Simplifying:

tv =


r+1
r+1 (+

)r+2
f

1
f

1
(+f )r+1



r+1

PF
(12)

5.8 Mean time to replace failed equipment

(initially waiting for an overhaul)
Failed equipments are replaced as soon as possible
and equipments that need an overhaul continue operating until being attended. Equipments waiting for
an overhaul which fail are treated as failed equipments. Therefore, these equipments are replaced
after replacement of lacked equipments in workstation
(failed equipments not replaced).
It is considered that the maintenance system is in the
steady state (state probabilities are time independent)
and it is assumed that system state at the moment of
failure occurrence is independent of the system state
when the overhaul is required.
Then, the average number of lacked equipments
when failure occurs can be calculated using equation 14 below.


rf =

PNF = 1 PF

t r et
r!

Pi,j Max [0; (i + j L) a (R L)]

i+jL

Pi,j

i+jL

(13)

Equations 10 and 11 depend on r, the number of

equipments to be replaced, given by:
r = i + j L.
5.7

Mean time until failure of an active

equipment waiting for an overhaul

The mean time until failure tv of an active equipment waiting for an overhaul is obtained based on
Erlang distribution and on Exponential distribution

where max[0; (i+jL)a(RL)] represents the average number of lacked equipments in the workstation
for the state (i, j).
Using the same logic that is used for determining
b2 (presented above), z1 ez2 are given by equation 15
below.
z1= z2 =

681

(rf + 1)
au

(14)

occupation. It gives information about the workstation

efficiency.
Improvement in the occupation rate can be
achieved by:

OTHER MAINTENANCE INDICATORS

FOR MFS SUBMITTED TO OVERHAULS

6.1

Average number of equipments waiting

for and under maintenance actions

For a MFS submitted to preventive maintenance,

equipment availability determination needs to incorporate time to perform an overhaul and time to
repair. Then, the number of non-operational equipments (the related indicator) includes failed equipments and equipments waiting for and under overhaul,
see equation 16.


Pi,j (i + j)

6.2

(15)

Average number of equipments in the queue

The average number of equipments in the queue (equation 17) allows the identification of the need for
maintenance center improvement.


Pi,j (i + j L)

(16)

i+jL

6.3

Average number of equipments in standby

The average number of lacked equipments in the workstation determined by Lopes et al. (2007) is directly
related to the average number of equipments in standby
(equation 18) and could also be used to assess MFS
efficiency.

i+jL

Pi,j [R (i + j)]


Pi,j max [0; (R L) (i + j L)a]

(17)

Increasing the number of spare equipments

Increasing the number of spare equipments allows
increasing the workplace occupation, but it also
decreases the equipment utilization rate. Utilization
needs to be as close as possible to equipment availability. However, holding cost of spare equipments
and investment made need to be balanced with loss
of production cost.
Increasing the number of maintenance servers
The maintenance center will deliver repaired equipments more frequently, and then equipments will
have a higher availability.
Improving maintenance efficiency
Maintenance efficiency can be enhanced by performing preventive maintenance actions or equipment improvement (enhance reliability or maintainability); changing work procedures and training
operators. Improving maintenance efficiency has
the same effect as increasing maintenance servers,
but it often requires less investment.
Decision makers also frequently use Overall Equipment Effectiveness (OEE), an indicator associated
with TPM methodology, to assess the efficiency of
production lines. OEE quantifies three types of losses:
performance and quality losses and loss associated
with reliability and maintainability. As far as the
loss associated with reliability and maintainability is
concerned, an availability indicator is usually used.
However, as shown in this work, workplace occupation is a better indicator than availability to assess
MFS efficiency. Therefore, we propose a new formula
for OEE to be used in the context of MFS replacing
availability by workplace occupation.

i+j>L

REFERENCES

CONCLUSIONS

The three indicators, equipment availability, equipment utilization and workplace occupation, addressed
in this work are important and need to be used in order
to prioritize improvement initiatives and monitor the
efficiency of MFS. Analytically, it seems complex to
determine equipment availability and equipment utilization. It involves quantifying the time in standby.
However, as shown, some other indicators provide
similar information and are easier to determine.
Decision makers use several kinds of indicators to
define and identify improvement initiatives. However,
one of the most important indicators for production
processes in the context of MFS is the workplace

Alsyouf, I. 2006. Measuring maintenance performance using

a balanced scorecard approach. Journal of Quality in
maintenance Engineering 12(2):133149.
Alsyouf, I. 2007. The role of maintenance in improving
companies productivity and profitability. International
Journal of Production economics 105(1):7078.
Chen, M.-C. & Tseng, H.-Y. 2003. An approach to design
of maintenance float systems. Integrated Manufacturing
Systems 14(3):458467.
Gross, D., Miller, D.R. & Soland, R.M. 1983. A closed
queueing network model for multi-echelon repairable item
provisioning. IIE Transactions 15(4):344352.
Gupta, S.M. 1997. Machine interference problem with
warm spares, server vacations and exhaustive service.
Performance Evaluation 29:195211.

682

Gupta, U.C. & Rao, T.S. 1996. On the M/G/1 machine

interference model with spares. European Journal of
Operational Research 89:164171.
Levine, B. 1965. Estimating maintenance float factors on the
basis of reliability. Theory in Industrial Quality Control
4(2):401405.
Lopes, I., Leito, A. & Pereira, G. A maintenance float
system with periodic overhauls. ESREL 2006. Portugal:
Guedes Soares & Zio (eds); 2006.
Lopes, I., Leito, A. & Pereira, G. 2007. State probabilities of a float system. Journal of Quality in Maintenance
Engineering 13(1):88102.
Madu, C.N. 1988. A closed queuing maintenance network
with two repair centers. Journal of Operational Research
Society 39(10):959967.
Madu, C.N. 2000. Competing through maintenance strategies. International Journal of Quality and Reliability
Management 17(7):937948.

Madu, C.N. & Kuei, C.-H. 1996. Analysis of multiechelon maintenance network characteristic using implicit
enumeration algorithm. Mathematical and Computer
Modelling 24(3):7992.
Madu, I.E. 1999. Robust regression metamodel for a maintenance float policy. International Journal of Quality &
Reliability Management 16(3):433456.
Shankar, G. & Sahani, V. 2003. Reliability analysis of a
maintenance network with repair and preventive maintenance. International Journal of Quality & Reliability
Management 20(2):268280.
Zeng, A.Z. & Zhang, T. 1997. A queuing model for designing
an optimal three-dimensional maintenance float system.
Computers & Operations Research 24(1):8595.

683

Monte Carlo methods in system safety and reliability

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Availability and reliability assessment of industrial complex systems:

A practical view applied on a bioethanol plant simulation
V. Gonzlez, C. Parra & J.F. Gmez
Industrial Management PhD Program at the School of Engineering, University of Seville, Spain

A. Crespo & P. Moreu de Len

Industrial Management School of Engineering, University of Seville, Spain

ABSTRACT: This paper shows a practical view about the behaviour of an industrial assembly in order to assess
its availability and reliability. For that intention it will be used such a complex system like a Bioethanol Plant.
A computerized model will help to create a realistic scenario of the Bioethanol Plant Life Cycle, obtaining an
estimation of the most important performance measures through real data and statistic inference. By this way, it
will be possible to compare and discuss the profit of different plant configurations using the model and following
the initial technical specifications. Basically, the Bioethanol Plant will be divided for that purposes in functional
blocks, defining their tasks and features, as well as their dependencies according to the plant configuration.
Additionally, maintenance information and data bases will be required for the defined functional blocks. Once
these data have been compiled and using any commercial software, it will be possible to carry out a model of the
plant and to simulate scenarios and experiments for each considered configuration. Parameters about availability
and reliability will be obtained for the most important functions, following different plant configurations. From
their interpretation, it will be interesting to consider actions that improve the availability and reliability of the
system under different plant functional requirements. Among other important aspects, it will be researchable
as well a sensitive analysis, i.e., the exploring on how parameters modifications have influence on the result or
final goal.

INTRODUCTION

Nowadays, investors, engineers etc. have to take

into consideration a lot of requirements and conditions in order to avoid risks and hazards on industrial
systems. Components or subsystems have potential
failure modes which have to be in mind from the system initial state and according to its operation modes,
environmental conditions, failure times etc. . .
The failure modelling could be very complicated
because of dependencies or inferences among components and, at the same time, the great amount
of required data. The intention here is to generate
a random number of events under a computerized
model which simulates the plant life scenario. There
are simulation methods that allow us taking into
consideration important aspects on the system operation like redundancies, stand-by nodes, preventive
maintenance, repairing priorities. . .
The use of such methods is rising up when one wants
to predict the general availability or the economical

assessment in a plant. Once simulated the operational

cycles of each component and through their combination and dependencies, we will be able to obtain
the whole system operation cycle. The scope of this
research will be the availability assessment of alternatives configurations in a plant according, not only to
predetermined maintenance strategies, but also to new
maintenance policies.

2
2.1

PROCEDURE
Definition of the system configuration

Description: It is needed to define functional blocks

and their dependencies according to the plant configuration and each block tasks or purposes.
Result: Functional block list/diagram: function,
inputs, outputs, etc. Functional chart of the system containing the relations among blocks and their
reliability features.

687

2.2

Data compiling

Grain (Starch)

Description: Maintenance information and data base

will be required for each considered block.
Result: Schedule with preventive tasks, times,
failure ratios etc.
2.3

Hydrolysis

Saccharification

Evaporation

Centrifuge

Distillation

Fermentation

Drying

Dehydration

DDG

Ethanol

Model construction

Description: Using any commercial software, it will

be possible to make a model of the plant.
Result: Plant computerized model.
2.4

Milling
Water

Syrup

Figure 1.
process.

CO2

General diagram of the Bioethanol obtaining

Simulation

Description: Using the model and with the above mentioned data we will simulate scenarios and experiments
for each considered configuration.
Result: Scenarios list, real and unreal events replication.
2.5

Results and analysis

Description: It will be able to calculate and discuss the

simulation results.
Results: Parameters about availability and reliability will be obtained for the most important functions,
following different plant configurations. From their
interpretation, we will consider actions to improve
the availability and reliability of the system under
plant functional requirements. Among other important
aspects, we will have as well the sensitive analysis, i.e.,
the exploring on how parameters modifications have
influence on the result or final goal.
3

SHORT DESCRIPTION OF THE

BIOETHANOL PROCESS AND ITS
RELATED BLOCK DIAGRAMS

Ethanol is commonly produced by the fermentation

of sugars. The conversion of starch to sugar, called
saccharification, significantly expands the choice of
feedstock to include cereals as corn, wheat, barley
etc. Other alternative route to bioethanol involves the
enzymatic conversion of cellulose and hemi-cellulose.
Completing the picture, ethanol is also commercially
synthesised from petrochemical sources but this does
not fall into the renewable category, which is the frame
of this paper.
Basically, the chemical reactions involved in
bioethanol production can be simplified as follows:

Figure 2. Diagram of the grain reception, cleaning, milling

and storage system.

Conversion of starch into bioethanol involves several more process steps, as. It starts by making a
beer from the milled grain, then distilling off the
alcohol followed by recovery of the residual solids and
recycle of water.
It is necessary to be care with each step of the process to assure an efficient conversion, particularly due
to the fact that it is a biological process where unsuitable reactions can happen causing loss in yield, and
also because different grains have different process
requirements. Some of the most important matters to
have into account in the sequence of process steps are
below mentioned.
Together with this description are also included
Block Diagrams of the main systems in a Bioethanol
plant. These diagrams are only an approaching of an
ideal plant. They try to show briefly the main equipments, devices, as well as material flows inside the
whole plant.
3.1 Slurry preparation
The feed grain is milled to a sufficient fineness in order
to allow water access to all the starch inside each grain.
The meal is then mixed with warm water till a specific
concentration without generating excessive viscosities
downstream.

688

Figure 3. Diagram of the mashing, cooking, liquefaction

and fermentation system.

3.2

Figure 4.
system.

Diagram of the distillation and dehydration

Hydrolysis

The slurry temperature is raised up in order to accelerate the hydrolysis of the grains starch into solution.
Again there is an optimum depending on the grain
typeif the slurry is too hot, the viscosity is excessive and if too cool, the required residence time for
effective hydrolysis is too long.
3.3

Figure 5. Diagram of the centrifugation, evaporation and

drying system.

Saccharification

With enzymes, the dissolved starch is converted to sugars by saccharification, but at a reduced temperature
which again is selected to achieve a balance between a
satisfactory reaction rate, and avoiding the promotion
of unsuitable side reactions and a subsequent loss in
yield.
3.4

Fermentation

The slurry is cooled to a fermentation temperature and

held in large batch fermentation tanks for a specific
time. Fresh yeast is prepared in parallel and added
at the beginning of each batch fermentation cycle.
The fermenting slurry is agitated and also circulated
through external exchangers to remove the heat generated by the fermentation process. On completion of
fermentation the batch is transferred to the Beer Well
and the particular fermentation tank and associated
equipment are cleaned by the plants CIP system.
3.5

Distillation

The beer contains about 812% ethanol. It is continuously pumped to the distillation unit which produces
an overhead stream of about 90% ethanol and water.
Ethanol and water form a 95% azeotrope so it is not
possible to reach 100% by simple distillation.
3.6

Dehydration

The 90% overhead stream is passed through an

absorber containing a molecular sieve which traps the
ethanol while letting the water pass through. Once the
bed in the absorber is full, the feed is switched to a

parallel absorber and the almost pure ethanol is sent

to product storage. The cycle is then repeated.
3.7 Centrifugation
The residual slurry left after distillation is called
Whole Stillage and it contains all the insoluble and
soluble non-starch components from the feed grain,
as well as the yeast which has grown during fermentation. The bulk of the solids, termed Wet Distillers
Grains (WDG) are removed by centrifuge leaving a
Thin Stillage.
3.8 Evaporation
To minimise water consumption, a large portion of the
Thin Stillage is recycled to the front of the process.
Basically, by effect of the evaporation a portion of
water is recycled, and the residual Syrup that contains around 3035% solids is either blended with the
WDG or sold separately to animal feed.
3.9 Drying
WDG is commonly dried as a mixture with the syrup to
about 10% moisture by gas fired rotary dryer. This byproduct, called Dried Distillers Grains and Solubles
(DDGS) is sold to animal feed.
3.10 CO2 Recovery
Fermentation simultaneously generates carbon dioxide which is collected and scrubbed to recover any

689

4.1.3 System availability and unavailability

For the system to be available, each subsystem should
be available. Thus:
Aseries = A1 A2
Figure 6.

Conversely, the unavailability is:

Diagram of the pelletizing system.

UAseries = 1 Aseries = 1 (1 UA1 ) (1 UA2 )

ethanol. It can either be processed into a by-product
or released to atmosphere.

SUMMARY OF FORMULAS
AND CONCEPTS REGARDING
RELIABILITY AND AVAILABILITY

The intention here is to summarise the formulation

regarding the reliability and availability concepts for
two subsystems (non-identical) in series or in parallel.
Additionally, a system with n identical subsystems in
parallel is also here taken into consideration, where
the system will be declared as failed if m or more
subsystems fail (the m_out_n case).
The system characteristics to be formulated here are
such as system failure rate, system Mean Time
Between Failure, system availability and unavailability, and system mean down time. We understand as Failure (f) the termination of the ability of
a component or system to perform a required function.
Therefore, the Failure Rate () will be the arithmetic
average of failures of a component and/or system per
unit exposure time. The most common unit in reliability analyses is hours (h). However, some industries
use failures per year
4.1

= UA1 + UA2 UA1 UA2

4.1.4 System mean down time for repairable
subsystems
If two subsystems are both repairable, one with mean
down time MDT1 and the other MDT2 , then, the mean
down time for the two subsystems in series will be as
follows. At any instance in time, the system is in one
of the 4 states:

The last 3 cases are responsible for the system

being non-functional. It is assumed that the 4th case
has negligible probability. Given the system is down,
the probability that it is because the subsystem #1 is
non-functional is obviously:
1
.
1 + 2
Since subsystem #1 needs MDT1 to repair, the
repair time associated with repairing subsystem #1 is
then:
1
MDT1
1 + 2

Two subsystems in series

4.1.1 System failure rate

Just for one subsystem, the failure rate is 1 . The
probability of failure in dt is 1 dt. For two subsystems in series, the probability of failure in dt
is (1 dt + 2 dt ). The system failure rate is thus
(1 + 2 ) : series = 1 + 2
The reliability function is:
R(t) = exp[(1 + 2 )t].

A similar expression is true for subsystem #2.

Summing them up, one gets:
MDTseries =

MTBF1 MDT2 + MTBF2 MDT1

MTBF1 + MTBF2

4.2 Two subsystems in parallel

Here the two subsystems are repairable. The mean
down times are MDT1 and MDT2 .

4.1.2 System MTBF

From the exponential form of the reliability function,
it is obvious that:
MTBFseries = 1/(1 + 2 ) =

Both subsystems functional.

Only subsystem #1 is non-functional.
Only subsystem #2 is non-functional.
Both subsystems are non-functional.

MTBF1 MTBF2
MTBF1 + MTBF2

4.2.1 System failure rate

If the system just consists of subsystem #1, then the
system failure rate is 1 . The probability of failure
in dt is 1 dt. Adding subsystem #2 in parallel, the
probability for system failure in dt is 1 dt reduced by
the probability that the subsystem #2 is in the failure

690

state. The probability to find the subsystem #2 in the

failure state is given by:

MDT2 << MTBF2

4.3 M out of N parallel subsystems

If a system consists of n parallel, identical subsystems
and the system is down if there are m or more subsystems down, then, the formulas for system failure rate,
system MTBF, system availability, and system mean
down time will be the following ones.

And using:
1
2

Then, the reduced failure rate for subsystem #1 is

then given by: 1 2 MDT2
Likewise, the reduced failure rate for subsystem #2
is: 1 2 MDT1
Consequently: parallel = 1 2 (MDT1 + MDT2 )
4.2.2 System MTBF
Taking the approach that the inverse of the failure rate
is MTBF (true for exponential distribution), one gets:

MTBFparallel = 1/parallel =

4.3.1 System failure rate

If the system just consists of subsystem #1, then the
system failure rate is . The probability of failure in
dt is dt. To have a system failure, we need to have
other (m-1) subsystems in the failure state. The chance
that any one subsystems is in the failure state is given
by MDT/(MTBF+MDT), or (MDT/MTBF), if we
assume MDT<< MTBF. To find (m1) subsystems
in the failure state, the probability is:


MTBF1 MTBF2
MDT1 + MDT2

It is noted that if the two subsystems are not

repairable, then the MTBF for the parallel case is the
sum of the individual MTBFs.
4.2.3 System availability and unavailability
For the system to be available, either subsystem should
be available. Thus:

MDT
MTBF

Conversely, the unavailability is:

UAparallel = 1 Aparallel = 1 (A1 + A2 A1 A2 )
= (1 A1 ) (1 A2 ) = UA1 UA2

(1)

4.2.4 System mean down time for repairable

subsystems
From the definition of:
MDT
MDT

MTBF + MDT
MTBF

It is possible to get the MDT for the parallel case

by using Eq.(1) above.

MDT m1
)
n1 C m1 n
MTBF
n!
=
m MDT m1
(n m)!(m 1)!

m_out_of _n = (

(2)

This is the failure rate for exactly m subsystem

failures. The failure rate for more than m subsystem
failures is going to be smaller by a factor of ( MDT ).
For a consistency check, we consider n = m = 2.
This is a system consisting of two parallel, identical
subsystems. When m = 2 subsystems fail, the system
fails. Eq. (2) for this case is system = 2 (2 MDT )
which agrees with the above mentioned formula.
4.3.2 System MTBF
Taking the approach that the inverse of the failure rate
is MTBF (true for exponential distribution), one gets:
MTBFm_out_of _n = 1/m_out_of _n

MDTparallel
MDTparallel
=
= MTBF1 MTBF2
MTBFparallel
MDT +MDT
1

m1

There are n1 Cm1 ways to group (m-1) subsystems

out of (n-1) subsystems. Also, it is possible to choose
any subsystem to be the #1 subsystem in the analysis.
Putting all together, one has:

Aparallel = A1 + A2 A1 A2

UAparallel

MDT1 MDT2
MDT1 + MDT2

MDTparallel =

Assuming that:

Unavailability =

MDT1
MDT2

MTBF1 MTBF2

Consequently:

MDT2
MTBF2 + MDT2

MTBF2 =

UA1 UA2 =

691

MTBF m
n!
(nm)!(m1)!

MDT m1

5.1 System failure rate

4.3.3 System availability and unavailability

For the system to be available, at least (n-m+1)
subsystems should be available. Thus:
n


Am_out_of _n =

i=nm+1

CO2 = Milling + Fermentation

Ethanol = Milling + Fermentation + Distillation +
Dehydration
DDGS = Milling + Fermentation + Distillation +
Centrifugation + Drying
Syrup = Milling + Fermentation + Distillation +
Centrifugation + Evaporation

n!
Ai (1 A)ni
(n i)! i!

Using the following equality:

1 = [A + (1 A)] =
n

n

i=0

Here it has been included the Hydrolysis and Saccharification in the same process area as Fermentation.

n!
Ai (1 A)ni
(n i)! i!

It is possible to rewrite the availability as:

Am_out_of _n = 1

nm

i=0

5.2 System MTBF system availability (A)

MTBFCO2 =

n!
Ai (1 A)ni
(n i)! i!

MTBFMilling MTBFFermentation
MTBFMilling +MTBFFermentation

MTBFEthanol =

n!
(1 A)m
m! (n m)!

MTBFCO2 MTBFDis+Deh
MTBFCO2 +MTBFDis+Deh ,

Where:
MTBFDis+Deh =

MTBFDistillation MTBFDehydration
MTBFDistillation +MTBFDehydration

MTBFDDGS =

MTBFCO2 MTBFDis+Cen+Dry
MTBFCO2 +MTBFDis+Cen+Dry ,

And the unavailability is given (again, for

MDT << MTBF), by:
UAm_out_of _n

n!
UAm
=
m! (n m)!

Where:
MTBFDis+Cen+Dry =

4.3.4 System mean down time for repairable

subsystems
From the definition of:
UAm_out_of _n =

MDTm_out_of _n
MTBFm_out_of _n

It is possible to get the MDT for the m_out_of_n

case by using the above mentioned formulas for
UAm_out_of _n and MTBFm_out_of _n .

MTBFDis+Cen =

MTBFDis+Cen MTBFDrying
MTBFDis+Cen +MTBFDrying ,
MTBFDistillation MTBFCentrifugation
MTBFDistillation +MTBFCentrifugation

MTBFSyrup =

MTBFCO2 MTBFDis+Cen+Eva
MTBFCO2 +MTBFDis+Cen+Eva ,

Where:
MTBFDis+Cen+Eva =
above mentioned:

Where:

MTBFDis+Cen MTBFEvaporation
MTBFDis+Cen +MTBFEvaporation ,

and as

Consequently:
MDTm_out_of _n =

MDT
m

MTBFDis+Cen =

FORMULATION APPLIED TO THE BLOCK

DIAGRAMS

MTBFDistillation MTBFCentrifugation
MTBFDistillation + MTBFCentrifugation

5.3 System availability (A)

Taking into consideration the General diagram of the

Bioethanol obtaining process included at the beginning of Section 3, it is possible to summarize basically
the Reliability characteristics of the Bioethanol plant
in the following formulas for a configuration in series:

692

ACO2 = AMilling AFermentation

AEthanol = ACO2 ADistillation ADehydration
ADDGS = ACO2 ADistillation ACentrifugation ADrying
ASyrup = ACO2 ADistillation ACentrifugation AEvaporation

5.4

which can be applied to analyze the reliability characteristics of the whole complex system, in this case, a
Bioethanol Plant.

System Mean Down Time (MDT)

MDTCO2
MTBFMilling MDTFermentation +MTBFFermentation MDTMilling
MTBFMilling +MTBFFermentation

MDTEthanol
MTBFCO2 MDTDis+Deh +MTBFDis+Deh MDTCO2
MTBFCO2 +MTBFDis+Deh

Where
MDTDis+Deh
=

MTBFDistillation MDTDehydration +MTBFDehydration MDTDistillation

MTBFDistillation +MTBFDehydration

MDTDDGS
=

MTBFCO2 MDTDis+Cen+Dry +MTBFDis+Cen+Dry MDTCO2

MTBFCO2 +MTBFDis+Cen+Dry

Where
MDTDis+Cen+Dry
=

MTBFDis+Cen MDTDrying +MTBFDrying MDTDis+Cen

,
MTBFDis+Cen +MTBFDrying

And
MDTDis+Cen
=

MTBFDistillation MDTCentrifugation +MTBFCentrifugation MDTDistillation

MTBFDistillation +MTBFCentrifugation

MDTSyrup =
MTBFCO2 MDTDis+Cen+Eva +MTBFDis+Cen+Eva MDTCO2
MTBFCO2 +MTBFDis+Cen+Eva

Where
MDTDis+Cen+Eva
MTBF

MDT

+MTBF

Dis+Cen
Evaporation
Evaporation
=
MTBFDis+Cen +MTBFEvaporation
and as above mentioned,

MDTDis+Cen

CONCLUSION

With this research we pretend to improve the estimations, demonstrating as well how requirements
expressed in initial technical specifications can be
incompatible or even impossible to accomplish for
determined plant configurations. That means, availability expectations on proposed configurations of the
whole plant could be lower, having higher reliability
or mantenability on each functional block, following
the technical requirements in effect.
Additionally, reasonable estimations will be provided for the production availability, which can be
delivered to the final customer in a more realistic
engineering project. These estimations will be based
on validated calculations of functional blocks considered for the model simulation, showing moreover the
importance and opportunity of a sensitive analysis. It
can also be decisive for the final selection the plant
technical configuration.
At the same time, this study can also be used to
adjust some initial requirements in the plant technical
specification. Once the data have been introduced in
the model, they can be adjusted according to the real
equipments included in the offer. By the way, it is possible to study logistical aspects like spare parts amount
in stock.
Finally, not only the availability and reliability are
important, but also the costs estimation is a key factor.
Therefore, an extension of this study could be to transfer the information provided in this research to a life
cycle cost model, with the intention to assess globally
the plant.

MDTDis+Cen
=

MTBFDistillation MDTCentrifugation +MTBFCentrifugation MDTDistillation

MTBFDistillation +MTBFCentrifugation

Once developed these formulas for the main parts of

the process, it is possible to continue breaking down
these as deeper as wanted. It is clear that each area
(milling, fermentation, distillation. . . ) in the plant
has actually its own configuration for its different
devices and equipments, with its specific combination for these subsystems in series or in parallel (see
for instance the different block diagrams also included
in Section 3).
Therefore, and as it has just mentioned, the reliability characteristics can be also broken down till such
a detail level where is possible to apply real values
for such units. There are published many Data Bases
which include real values for process equipments, and

ACKNOWLEDGEMENTS
The author would like to thank the reviewers of the
paper for their contribution to the quality of this work.
REFERENCES
Asociacin Espaola de Mantenimiento. 2005. El Mantenimiento en Espaa: Encuesta sobre su situacin en las
empresas espaolas.
Bangemann T., Rebeuf X., Reboul D., Schulze A., Szymanski
J., Thomesse J.P., Thron M., Zerhouni N.. 2006. ProteusCreating distribuited maintenance systems through an
integration platform. Computers in Industry, Elselvier.
Benot Iung 2006. CRAN Laboratory Research Team
PRODEMAS in Innovative Maintenance and Dependability. Nancy UniversityNancy Research Centre

693

for Automatic Control (CRAN). CNRS UMR 7039

(http://www.cran.uhp-nancy.fr).
Bourne, M. & Neely, A. 2003. Performance measurement
system interventions: the impact of parent company initiatives on success and failure. Journal of Operation and
Management.
Campbell J.D. & Jardine A. 2001. Maintenance excellence.
New York: Marcel Dekker. 2001.
Carter, Russell A. 2001. Shovel maintenance gains from
improved designs, tools and techniques. Elsevier Engineering Information.
Center for Chemical Process Safety (CCPS)Guidelines for
Process Equipment Reliability Data, with Data Tables.
ISBN: 0-8169-0422-7.
Clark J. 1995. Managing Innovation and Change: People,
Technology and Strategy. Business & Economics.
Crespo M.A., Moreu de L.P., Sanchez H.A. 2004. Ingeniera
de Mantenimiento. Tcnicas y Mtodos de Aplicacin a
la Fase Operativa de los Equipos. Aenor, Espaa.
Crespo M.A. 2007. The Maintenance Management Framework. Models and Methods for Complex Systems Maintenance. Londres, Reino Unido. Springer.
Dale & Plunkett 1991. Quality Costing. Chapman Hall.
Dandois, P.A. & Ponte, J. 1999. La administracin del
conocimiento organizacional. El management en el siglo
XXI.
Davenport T. 1993. Process innovation: Reengineering
work through Information Technology. Harvard Business
School Press.
David John Smith. 2005. Reliability, Maintainability and
Risk: Practical Methods for Engineers. ISBN 0-75066694-3 Butterworth-Heinemann.
Deardeen, J. Lilien, G. and Yoon, E. 1999. Marketing
and Production Capacity Strategy For Non-Differentiated
Products: Winning And Losing At The Capacity Cycle
Game. International Journal of Research In Marketing.
Deming W. Edwards 1989. Calidad, productividad y competitividad: la salida de la crisis. Madrid, Ediciones Daz de
Santos.
Dixon J.R. 1966. Design engineering: inventiveness, analysis, and decision making. New York, McGraw-Hill,
Inc.
Dyer R.F. & Forman E.H. 1992. Group decision support
with the Analytic Hierarch Process. Decision Support
Systems.
Earl M.J. 1994. The New and the Old of Business Process Redesign. Journal of Strategic Information Systems,
vol. 3.
EN 13306:2001. Maintenance Terminology. European Standard. CEN (European Committee for Standardization),
Brussels.
European Foundation for Quality Management. 2006. EFQM
Framework for Management of External Resources. By
EIPMEFQM.
Gelders, L., & Pintelon, L. 1988. Reliability and maintenance" in: Doff, R.C. and Nof, S.J. (ed.), International
Encyclopedia of Robotics, Application and Automation,
Wiley, New York.
Goldratt E. 1997. Cadena Crtica. Ediciones Diaz de Santos.
Hammer & Champy 1993. Reengineering the Corporation.
Harper Business.
Hammer M. 1990. Reengineering Work: Dont Automate
Obliterate. Harvard Business Review.

Hansen, M.T.; Noria, N.; Tierney, T. 1999. Whats your

strategy for managing knowledge?. Harvard Business
Review.
Henley, E.J. and Kumamoto, H. 1992 Probabilistic Risk
Assessment: Reliability Engineering, Design & Analysis.
ISBN 0 87942 290 4. IEEE Press, US.
Hyland A & Rausand M, 1994. System Reliability Theory.
Models and Statistical methods. ISBN 0-471-59397-4.
Wiley-Interscience.
Inc, Renewable Fuels Association. 2001. Ethanol Plant
Development Handbook. BBI International. USA
Intelligent Maintenance Centre 2007. www.imscenter.net.
Iserman R. 1984. Process fault detection based on modelling
and estimation methods. Automatica.
ITSMF, IT Service Management Forum 2007. ITIL v3.
Information Technology Infrastructure Library. ITIL v2.
Information Technology Infrastructure.
Jardine A.K.S., Lin D., Banjevic D. 2006. A review on
machinery diagnostics and prognostics implementing
condition based maintenance. Mech, Syst. Signal Process.
Jharkharia S. & Shankarb R. 2005. Selection of logistics
service provider: An analytic network process (ANP)
approach. International Journal of Management Sciente,
Omega 35 (2007) 274289.
Kaplan, Robert S. & David P. Norton 1996. The Balanced
Scorecard: Translating Strategy Into Action . Boston,
MA: Harvard Business School Press.
Kent Allen 1990. Encyclopedia of Computer Science and
Technology. CRC Press.
Klein, M.M. 1994. The most fatal reengineering mistakes.
Information strategy: The Executives J. lO(4) 2128.
Lee J. 1995. Machine performance monitoring and proactive maintenance in computer-integrated manufacturing:
review and perspective. International Journal of Computer
Integrating Manufacturing.
Lee J. 2004. Infotronics-based intelligent maintenance system and its impacts to close-loop product life cycle
systems. Proceedings of de IMS2004 International Conference on Itelligent Maintenance Systems, Arles, France.
Levitt Joel. 2003. Complete Guide to Preventive and Predictive Maintenance. Industrial Press.
Davis, M. 1988. Applied Decision Support. Prentice Hall,
Englewood Cliffs .
Mitchell Ed., Robson Andrew, Prabhu Vas B. 2002. The
Impact of Maintenance Practices on Operational and
Business Performance. Managerial Auditing Journal.
Mobley Keith 2002. An Introduction to Predictive Maintenance. Elsevier.
Moubray J. 1997. Reliability-centered Maintenance. Industrial Press.
Nakajima Seiichi 1992. Introduccin al TPM (Mantenimiento Productivo Total). Productivity Press.
Neely, A.D., Gregory, M. and Platts, K. 1995. Performance
Measurement System DesignA Literature Review and
Research Agenda. International Journal of Operations and
Production Management.
Nonaka & Takeuchi 1995. The Knowledge-Creating Company. USA: Oxford University Press.
Patton, J.D. 1980. Maintainability and Maintenance Management. Instrument Society of America, Research Triangle
Park, NC.
Peters T. & Waterman H.R. Jr. 1982. In Search of Excellence.

694

Pintelon L.M. & Gelders L.F. 1992. Maintenance management decision making. European Journal of Operational
Research.
Porter, M. 1985. Competitive Advantage. Free Press.
Prusak Laurence 1996. The Knowledge Advantage. Strategy
& Leadership.
Ren Yua, Benoit Iung, Herv!e Panetto 2003. A multi-agents
based E-maintenance system with case-based reasoning
decision support. Engineering Applications of Artificial
Intelligence 16 321333.
Saaty, T.L. 1977. A Scaling Method for Priorities in Hierarchical Structures. Journal of Mathematical Psychology,
15: 234281, 1977.
Saaty, T.L. 1980. The Analytic Hierarchy Process. McGraw
Hill.
Saaty, T.L. 1990. How to make a decision: The analytic hierarchy process. European Journal of Operational
Research.
Shu-Hsien Liao 2005. Expert system methodologies and
applications-a decade review from 1995 to 2004.
Elselvier. Expert Systems with Applications 28 93103.

Turban E. 1988. Decision Support and Expert Systems:

Managerial Perspectives. New York: Macmillan.
UNE 66174 2003. Guide for the assessment of quality management system according to UNE-EN ISO 9004:2000
standard. Tools and plans for improvement. UNE.
UNE 66175 2003. Systems of Indicators. UNE.
UNE-EN ISO 9001:2000.
Quality management
systemsRequirements. International Organization for
Standardization.
Wireman, T. 1991. Total Productive Maintenance. Industrial
Press.
Yan S.K. 2003. A condition-based failure prediction and
processing-scheme for preventive maintenance. IEEE
Transaction on Reliability.
Zhu G., Gelders L. and Pintelon L. 2002. Object/objectiveoriented maintenance management. Journal of quality in
maintenance engineering.

695

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Handling dependencies between variables with imprecise probabilistic

models
Sebastien Destercke & Eric Chojnacki
Institut de Radioprotection et de Sret Nuclaire, Cadarache, France

ABSTRACT: Two problems often encountered in uncertainty processing (and especially in safety studies)
are the following: modeling uncertainty when information is scarce or not fully reliable, and taking account
of dependencies between variables when propagating uncertainties. To solve the first problem, one can model
uncertainty by sets of probabilities rather than by single probabilities, resorting to imprecise probabilistic models.
Iman and Conover method is an efficient and practical means to solve the second problem when uncertainty is
modeled by single probabilities and when dependencies are monotonic. In this paper, we propose to combine
these two solutions, by studying how Iman and Conover method can be used with imprecise probabilistic
models.
1

INTRODUCTION

Modeling available information about input variables

and propagating it through a model are two main
steps of uncertainty studies. The former step consists in choosing a representation fitting our current
knowledge or information about input variables or
parameters, while the latter consists in propagating
these information through a model (here functional)
with the aim to estimate the uncertainty on the output(s) of this model. In this paper, we consider that
uncertainty bears on N variables X1 , . . . , XN defined
on the real line. For all i = 1, . . . , N , we note xi a
particular value taken by Xi .
Sampling methods such as Monte-Carlo sampling
or Latin Hypercube sampling (Helton and Davis 2002)
are very convenient tools to simulate and propagate
random variables X1 , . . . , XN . Most of the time, they
j
j
consists in sampling M realizations (x1 , . . . , xN ), j =
1, . . . , M of the N random variables, thus building a
M N sample matrix S. Each line of the matrix S
can then be propagated through a model T : RN
R. When using such sampling technics, it is usual to
assume:
1. That uncertainty on each Xi is representable by a
unique probability density pi associated to a unique
cumulative distribution Fi , with

Fi (x) = Pi ([, x]) =

pi (x)dx.

2. That variables X1 , . . . , XN are independent, that

is that their joint probability distribution is provided by the product of the marginal probability
distributions.
In real applications, both assumptions can be challenged in a number of practical cases: the first when
available information is scarce, imprecise or not fully
reliable, and the second when independence between
variables cannot be proved or is clearly unrealistic. As
shown in (Ferson and Ginzburg 1996), making such
assumptions when they are not justified can lead to
underestimations of the final uncertainty on the output,
possibly leading to bad decisions.
Although there exist some very practical solutions
to overcome either scarceness of the information or
dependencies between variables, there are not a lot
of methods treating both problems at the same time.
In this paper, we propose and discuss such a method,
that combines the use of simple imprecise probabilistic
representations with classical technics used to model
monotonic dependencies between variables (namely,
Iman and Conover method). The paper is divided
in two main sections: section 2 is devoted to basics
needed to understand the paper, and section 3 explains
and discusses the proposed method.

PRELIMINARIES

This section recalls the main principles of Iman and

Conover (Iman and Conover 1982) method to integrate monotonic dependencies in a sampling matrix
and introduces possibility distributions (Baudrit and

697

Dubois 2006) and probability boxes (p-boxes for short)

(Ferson, Ginzburg, Kreinovich, Myers, and Sentz
2003), the two practical probabilistic models we are
going to consider. More details can be found in the
references.
2.1

Integrating monotonic dependencies

in sampling procedures

The first problem we deal with is the integration of

dependencies into sampling schemes. In the sequel,
Si,j denote the matrix element in the ith line and jth
column of S, while Sj and Si respectively denote the
jth column and ith line of S.
Suppose we consider two variables X , Y and a sample (xj , yj ) of size M of these two variables. Then, if we
replace the values xj and yj by their respective ranks
(The lowest value among xj receive rank 1, second lowest rank 2, . . . , and similarly for yj ), their spearman
rank correlation coefficient rs , which is equivalent
to the Pearson correlation computed with ranks, is
given by
 M 2 
6 j=1 dj
rs = 1
M (M 2 1)
with dj the difference of rank between xj and yj .
Spearman correlation rs have various advantages:
i it allows to measure or characterize monotonic (no
necessarily linear) dependencies between variables
ii it depends only on the ranks, not on the particular
values of the variables (i.e. it is distribution-free).
Although Spearman correlations rank are not able to
capture all kinds of dependencies, they remain nowadays one of the best way to elicit dependency structures
(Clemen, Fischer, and Winkler 2000).
Given a sample matrix S and a N N target rank correlation matrix R (e.g. elicited from experts), Iman and
Conover (Iman and Conover 1982) propose a method
to transform the matrix S into a matrix S such that the
rank correlation matrix R of S is close to the target
matrix R. This transformation consists in re-ordering
the elements in each column S,j of S, without changing their values in S, so that the result is the matrix S .
The transformation consists in the following steps:
1. Build a M N matrix W whose N columns are random re-orderings of the vector (a1 , . . . , aM ), where
ai = 1 (i/(M + 1)), 1 being the inverse of a
standard normal cumulative distribution, that is
 2 
 x
1
u
x, (x) =
exp du
2
2
Let C be the N N correlation matrix associated to W .

2. Build a lower triangular N N matrix G such that

G C G  = R with G  the transpose of G. This
can be done in the following way: use Cholesky
factorization procedure to decompose C and R
into C = C C and R = R R , with both
C , R lower triangular matrix (due to the fact that
correlation matrices C, R are, by definition, positive definite and symmetric). Then, G is given by
G = R C 1 and the transpose follows. Note that
G is still a lower triangular matrix.
3. Compute the M N matrix W = W G  .
4. In each column S,j of the original sample matrix,
re-order the sampled values so that they are ranked
as in the column W,j , thus obtaining a matrix S
whose rank correlation matrix R is close to R (but
not forcefully equal, as for a given number M of
samples, rank correlations coefficients can only
assume a finite number of distinct values).
This method allows to take account of monotonic dependencies between the variables in sampling
schemes (and, therefore, in the subsequent propagation), without making any assumptions about the
shape of probability distributions and without changing the sampled value (it just rearranges their pairings
in the sample matrix). It is also mathematically simple and applying it do not require complex tools,
as would other approaches involving, for example,
copulas (Nelsen 2005).

2.2 Modeling uncertainty with sets of probabilities

The second problem concerns situations where available information is scarce, imprecise or not fully
reliable. Such information can come, for instance,
from experts, from few experimental data, from sensors, etc. There are many arguments converging to
the fact that, in such situations, a single probability distribution is unable to account for the scarcity
or imprecision present in the available information,
and that such information would be better modeled by
sets of probabilities (see (Walley 1991), Ch.1) for a
summary and review of such arguments).
Here, we consider two such models: p-boxes and
possibility distributions. They are both popular, simple
and are instrumental to represent or elicit information from experts (for more general models and longer
discussion, see (Destercke, Dubois, and Chojnacki
2007)).
P-boxes (short name for probability boxes) are the
imprecise counterparts of cumulative distributions.
They are defined by an upper (F) and a lower (F)
cumulative distributions forming a pair [F, F] describing the uncertainty: the information only allows us to
state that the true cumulative distribution is between
F and F, and any cumulative distribution F such that

698

0.9

0.5

0.1

0
500 K

Figure 2.
500 K

Figure 1.

600 K

800 K

900 K

600 K

700 K

800 K

900 K

1000 K

Illustration of a possibility distribution.

1000 K

Illustration of a p-box.

probabilities P (Dubois and Prade 1992) such that

P = {P|A RN (A) P(A) (A)}

F F F is coherent with the available information.

A p-box induces a set P[F,F] of probabilities, such that

with P a probability distribution. For a given possibility distribution and for a given value [0, 1], the
(strict) -cut of is defined as the set

P[F,F] = {P|x R, F(x) P([, x]) F}.

= {x R|(x) > }.

P-boxes are appropriate models when experts provide a set of (imprecise) percentiles, when considering
the error associated to sensor data, when we have
only few experimental data or when we have only
information about some characteristics of a distribution (Ferson, Ginzburg, Kreinovich, Myers, and Sentz
2003). Consider the following expert opinion about
the temperature of a fuel rode in a nuclear reactor core
during an accidental scenario:
Temperature is between 500 and 1000 K
The probability to be below 600 K is between 10
and 20%
The probability to be below 800 K is between 40
and 60%
The probability to be below 900 K is between 70
and 100%
Figure 1 illustrates the p-box resulting from this
expert opinion.
Possibility distributions correspond to information given in terms of confidence intervals, and thus
correspond to a very intuitive notion. A possibility distribution is a mapping : R [0, 1] such that there
is at least one value x for which (x) = 1. Given a
possibility distribution , possibility  and necessity
N measures of an event A are respectively defined as:

Note that -cuts are nested (i.e. for two values <
, we have ). An -cut can then be interpreted
as an interval to which we give confidence 1 (The
higher , the lower the confidence). -cuts and the set
of probabilities P are related in the following way
P = {P| [0, 1], P( ) 1 }.
Possibility distributions are appropriate when
experts express their opinion in term of nested confidence intervals or more generally when information
is modeled by nested confidence intervals (Baudrit,
Guyonnet, and Dubois 2006). As an example, consider
an expert opinion, still about the temperature of a fuel
rode in a nuclear reactor core, but this time expressed
by nested confidence intervals:
Probability to be between 750 and 850 K is at
least 10%
Probability to be between 650 and 900 K is at
least 50%
Probability to be between 600 and 950 K is at
least 90%
Temperature is between 500 and 1000 K (100%
confidence)
Figure 2 illustrates the possibility distribution
resulting from this opinion.

(A) = max (x) and N (A) = 1 (Ac )

xA

with Ac the complement of A. For any event A,

N (A) (A), and possibility and necessity measure are respectively interpreted as upper and lower
confidence levels given to an event, defining a set of

PROPAGATING WITH DEPENDENCIES

AND IMPRECISE MODELS

Methods presented in Section 2 constitute very practical solutions to solve two different problems often
encountered in applications. As both problems can be

699

Also note that above sampling procedures have been

considered by Alvarez (Alvarez 2006) in the more
general framework of random sets, of which p-boxes
and possibility distributions constitute two particular
instances. Let us now see how Iman and Conover
method can be extended to such models.

encountered in a same application, it would be interesting to blend these two tools. Such a blending is
proposed in this section.
3.1

Sampling with imprecise probabilistic models

When uncertainty on a (random) variable X is modeled by a precise cumulative distribution FX , then

simulating this variable X by sampling methods usually consists of drawing values coming from a
uniform law on [0, 1], and then to associate the (precise) value F 1 () to each value (see Figure 3.A).
In the case of a N -dimensional problem simulated
by M samples, the j t h sample consists of a vector
j
j
(1 , . . . , N ), to which is associated the realization
j
j
1 j
(F ( )1 , . . . , F 1 ( j )N ) = (x1 , . . . , xN ). Let us
now detail what would be the result of such a sampling
with imprecise models.
P-boxes: since a p-box is described by (lower and
upper ) bounds on cumulative distributions, to each
value do not longer correspond a unique inverse
value, but a set of possible values. This set of possible
values correspond to the interval bounded by the upper
1
(F ) and lower (F 1 ) pseudo inverses, defined, for
all (0, 1] as follows:
F

= sup{x R|F(x) < }

= inf {x R|F(x) > }

See Figure 3.B for an illustration. Thus, given a pbox [F, F], to a sampled value [0, 1] we associate
the interval  such that
 := [F

(), F 1 ()]

Possibility distributions: In the case of a possibility distributions, it is natural to associate to each value
the corresponding -cut (see Figure 3.C for illustration). Anew, this -cut is, in general, not a single
value but an interval.
We can see that, by admitting imprecision in our
uncertainty representation, usual sampling methods do
not longer provide precise values but intervals (which
are effectively the imprecise counterpart of single values). With such models, elements of matrix S can be
intervals and propagating them through a model T will
require to use interval analysis technics (Moore 1979).
Although achieving such a propagation is more difficult than single point propagation when the model
T is complex, it can still remain tractable, even for
high dimensional problems (see (Oberguggenberger,
King, and Schmelzer 2007) for example). Nevertheless, propagation is not our main concern here, and
sampling scheme can be considered independently of
the subsequent problem of propagation.

3.2 Extension of Iman and Conover method

We first recall some notions coming from order theory.
Let P be a set and a relation on the elements of
this set. Then, is a complete partial order if it is
reflexive, antisymmetric and transitive, that is if for all
triplet a, b, c of elements in P
a a (reflexivity)

(1)

if a b andb a, then a = b (antisymmetry)

(2)

if a b and b c, then a c (transitivity)

(3)

and if for two elements a, b, neither a b nor b a,

then a and b are said to be incomparable. A partial
order is total, and is called an order (or a linear
order), if for every pair a, b in P, we have either a b
or b a.
When uncertainty is modeled by precise probabilities, sampled values are precise, and the main reason
for being able to apply Iman and Conover method in
this case is that there is a natural complete ordering
between real numbers, and that to any set of values
corresponds a unique ranking. This is no longer the
case when realizations are intervals, since in most
cases only partial orderings can be defined on sets of
intervals (due to the fact that they can be overlapping,
nested, disjoint,. . . ). Given two intervals [a, b], [c, d],
it is common to consider the partial ordering such that
[a, b] < [c, d] if and only if b < c, and to consider
that two intervals are incomparable as soon as they
overlap. This partial order is commonly called interval order. Adapting Iman and Conover method when
samples are general intervals thus seems difficult and
would result in a not very convenient tool, since one
would have to consider every possible extension of the
partial ordering induced by the interval ordering.
To circumvent this problem and to be able to apply
Iman and Conover method in an easy way on p-boxes
and possibility distributions, we have to define a complete ordering on the elements sampled from these two
representation.
First, note that when uncertainty on a variable X is
modeled by a single (invertible) cumulative distribution FX , there is a one-to-one correspondence between
the ranking of sampled values j [0, 1] and the ranking of corresponding values of X , in the sense that, for

700

1
F

x= F

( )

Fig. 3.A: precise prob.

Figure 3.

( )

1(

Fig. 3.B: p-box

0
Fig. 3.C: possibility dist.

Sampling from precise and imprecise probabilistic models: illustration.

two values i , j , we have

i < j F 1 ( i ) < F 1 ( j ).

(4)

We will use this property to extend Iman and Conover

technics when realizations are either intervals 
coming from a p-box or -cuts of a possibility
distribution.
P-boxes: Consider first a p-box [F, F]. For such a pbox, the ordering similar to Equation (4) means that for
two values , in [0, 1], we have <  <  ,
which is equivalent to impose a complete ordering
between cuts  of the p-box. We note this ordering
[F,F] . Given two intervals [a, b], [c, d], this definition
is equivalent to state that an interval [a, b] [F,F] [c, d]
if and only if a c and b d. Roughly speaking, taking such an ordering means that the rank of an
interval increases as it shifts towards higher values.
Note that the ordering [F,F] is complete only when
intervals are sampled from a p-box, and that incomparability can appear in more general cases (e.g. when
intervals are nested, or when they come from general
random sets).
Possibility distributions: Given a possibility distribution , the ordering similar to Equation (4) is
equivalent to consider a complete ordering on -cuts
induced by inclusion: for two values , in [0, 1], we
have < . We note the ordering
such that
[a, b] [c, d] if and only if [a, b] [c, d]
with [a, b], [c, d] two intervals. Here, the rank of
an interval increases as it gets more precise (narrower). Again, the ordering is complete only when
intervals are sampled from a possibility distribution.
Now that we have defined complete orderings on
intervals sampled from p-boxes and possibility distributions, we can apply Iman and Conover method without difficulty to these models. Nevertheless, one must
pay attention that a same value of rank correlation will

have different meaning and interpretation, depending

on the chosen representation (and, consequently, on
the chosen ordering).
In the case of p-boxes defined on multiple variables,
a positive (negative) rank correlation always means
that to higher values are associated higher (lower)
values. The main difference with single cumulative
distributions is that samples are now intervals instead
of single values. In the case of p-boxes, the application of Iman and Conover method can then be seen as a
simple extension of the usual method, with the benefits that imprecision and scarceness of information
is now acknowledged in the uncertainty model. As a
practical tool, it can also be seen as a means to achieve a
robustness study (concerning either distribution shape
or correlation coefficients). Since correlation coefficients can seldom be exactly known and are often
provided by experts, such a robustness interpretation
appears appealing. Also note that the ordering [F,F]
is a refinement of the classical ordering considered on
intervals, and reduce to the classical ordering between
numbers when samples are single values. All this indicates that using Iman and Conover method on p-boxes
is also equivalent to induce monotonic dependencies
between variables.
Contrary to p-boxes, rank correlations related to
possibility distributions and to the ordering cannot
be considered as an extension of the classical Spearman rank correlations. To see this, simply note that the
ordering , based on inclusion between sets, is not
a refinement of the classical ordering considered on
intervals, and do not reduce to the classical ordering
of numbers when samples are single values. In the case
of possibility distributions, a positive (negative) rank
correlation between two variables X , Y means that to
more precise descriptions of the uncertainty on X will
be associated more (less) precise descriptions of the
uncertainty on Y , i.e. that to narrower intervals will
correspond narrower (broader) intervals. Such dependencies can be used when sensors or experts are likely
to be correlated, or in physical models where knowing

701

a value with more precision means knowing another

one with less precision (of which Heisenberg principle constitutes a famous example). Such kind of
dependencies has poor relation with monotonic dependencies, meaning that using the proposed extension to
possibility distribution is NOT equivalent to assume
monotonic dependencies between variables, but rather
to assume a dependency between the precision of
the knowledge we have on variables. Nevertheless,
if monotonic dependencies have to be integrated and
if information is modeled by possibility distributions,
it is always possible to extract a corresponding p-box
from a possibility distribution, and then to sample from
this corresponding p-box (see (Baudrit and Dubois
2006)).
4

CONCLUSIONS

Integrating known correlation between variables and

dealing with scarce or imprecise information are
two problems that coexist in many real applications.
The use of rank correlation through the means of
Iman and Conover method and the use of simple
imprecise probabilistic models are practical tools to
solve these two problems. In this paper, we have
proposed an approach to blend these two solutions,
thus providing a practical tool to cope (at the same
time) with monotonic dependencies between variables
and with scarceness or imprecision in the information.
Sampling methods and complete orderings related
to possibility distributions and p-boxes have been studied and discussed. They allow to apply Iman and
Conover method to these two models without additional computational difficulties. We have argued that,
in the case of p-boxes, rank correlations can still be
interpreted in terms of monotonic dependencies, thus
providing a direct extension of Iman and Conover
method, with the advantage that it can be interpreted
as an integrated robustness study. The interpretation
concerning possibility distributions is different, as it
is based on set inclusion, and describes some dependencies between the precision of the knowledge we
can acquire on different variables. We suggest that
such correlation can be useful in some physical models, or when sources of information (sensors, experts)
are likely to be correlated.
In our opinion, the prime interest of the suggested
extensions is practical, as they allow to use very popular and efficient numerical technics such as Latin
Hyper Cube Sampling and Iman and Conover method
with imprecise probabilistic models. Moreover, the
proposed extensions can benefits from all the results

concerning these numerical technics (for instance, see

(Sallaberry, Helton, and Hora 2006)).

REFERENCES
Alvarez, D.A. (2006). On the calculation of the bounds of
probability of events using infinite random sets. I. J. of
Approximate Reasoning 43, 241267.
Baudrit, C. and D. Dubois (2006). Practical representations
of incomplete probabilistic knowledge. Computational
Statistics and Data Analysis 51 (1), 86108.
Baudrit, C., D. Guyonnet, and D. Dubois (2006). Joint
propagation and exploitation of probabilistic and possibilistic information in risk assessment. IEEE Trans. Fuzzy
Systems 14, 593608.
Clemen, R., G. Fischer, and R. Winkler (2000, August).
Assessing dependence : some experimental results. Management Science 46 (8), 11001115.
Destercke, S., D. Dubois, and E. Chojnacki (2007). Relating practical representations of imprecise probabilities. In
Proc. 5th Int. Symp. on Imprecise Probabilities: Theories
and Applications.
Dubois, D. and H. Prade (1992). On the relevance of nonstandard theories of uncertainty in modeling amd pooling expert opinions. Reliability Engineering and System
Safety 36, 95107.
Ferson, S., L. Ginzburg, V. Kreinovich, D. Myers, and
K. Sentz (2003). Constructing probability boxes and
dempster-shafer structures. Technical report, Sandia
National Laboratories.
Ferson, S. and L.R. Ginzburg (1996). Different methods are
needed to propagate ignorance and variability. Reliability
Engineering and System Safety 54, 133144.
Helton, J. and F. Davis (2002). Illustration of samplingbased methods for uncertainty and sensitivity analysis.
Risk Analysis 22 (3), 591622.
Iman, R. and W. Conover (1982). A distribution-free
approach to inducing rank correlation among input variables. Communications in Statistics 11 (3), 311334.
Moore, R. (1979). Methods and applications of Interval Analysis. SIAM Studies in Applied Mathematics.
Philadelphia: SIAM.
Nelsen, R. (2005). Copulas and quasi-copulas: An introduction to their properties and applications. In E. Klement
and R. Mesiar (Eds.), Logical, Algebraic, Analytic, and
Probabilistics Aspects of Triangular Norms, Chapter 14.
Elsevier.
Oberguggenberger, M., J. King, and B. Schmelzer (2007).
Imprecise probability methods for sensitivity analysis in
engineering. In Proc. of the 5th Int. Symp. on Imprecise
Probabilities: Theories and Applications, pp. 317326.
Sallaberry, C., J. Helton, and S. Hora (2006). Extension
of latin hypercube samples with correlated variables.
Tech. rep. sand2006- 6135, Sandia National Laboratories, Albuquerque. http://www.prod.sandia.gov/cgibin/
techlib/accesscontrol. pl/2006/066135.pdf.
Walley, P. (1991). Statistical reasoning with Imprecise
Probabilities. New York: Chapman and Hall.

702

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Monte Carlo simulation for investigating the influence of maintenance

strategies on the production availability of offshore installations
K.P. Chang, D. Chang & T.J. Rhee
Hyundai Industrial Research Institute, Hyundai Heavy Industries, Ulsan, Korea

Enrico Zio
Energy Department, Politecnico di Milano, Milan, Italy

ABSTRACT: Monte Carlo simulation is used to investigate the impact of the maintenance strategy on the
production availability of offshore oil and gas plants. Various realistic preventive maintenance strategies and
operational scenarios are considered. The reason for resorting to Monte Carlo simulation is that it provides
the necessary flexibility to describe realistically the system behavior, which is not easily captured by analytical
models. A prototypical offshore production process is taken as the pilot model for the production availability
assessment by Monte Carlo simulation. The system consists of a separator, compressors, power generators,
pumps and dehydration units. A tailor-made computer program has been developed for the study, which enables
to account for the operational transitions of the system components as well as the preventive and corrective
maintenance strategies for both power generators and compressor systems.

INTRODUCTION

By definition, production availability is the ratio

between actual and planned production over a specified period of time [NORSOK Z-016, 1998]. Production availability is considered as an important
indicator of the performance of offshore installations
since it describes how the system is capable of meeting
demand for deliveries.
Recently, the demands of offshore industries for
production availability analysis are not only limited
to knowing the average production evolution but also
present the optimization of the components and systems maintenance strategies, e.g., the maintenance
intervals and spare parts holdings.
Indeed, production availability is affected by the
frequencies of corrective and preventive maintenance
tasks. Furthermore, the spare parts holding requirements must comply with the limits of space and weight.
The items with long lead times for replacement can
also give a serious negative effect on the production
availabilities.
The followings are typical results which are
expected from a production availability analysis;
To identify the critical items which have a dominant
effect on the production shortfall
To verify the intervention and maintenance strategies planned during production
To determine the (minimum) spare part holdings

To satisfy these requirements, stochastic simulation models, such as Monte Carlo simulation, are
increasingly being used to estimate the production
availabilities of offshore installations. They allow
accounting for the realistic maintenance strategies and
operational scenarios [E.Zio et al., 2006].
The purpose of the present study is to develop a
Monte Carlo simulation method for the evaluation of
the production availability of offshore facilities while
accounting for the realistic aspects of system behavior.
A Monte Carlo simulation model has been developed
to demonstrate the effect of maintenance strategies
on the production availability, e.g., by comparing
the system performance without and with preventive
maintenance, and of delays of spare parts for critical
items.
2

SYSTEM DESCRIPTION

A prototypical offshore production process is taken

as the pilot model for the production availability
assessment by Monte Carlo simulation (Figure 1).
2.1

Functional description

The three-phase fluid produced in the production well

enters a main separation system which is a single-train
three-stage separation process. The well fluid is separated into oil, water and gas by the separation process.

703

Gas
Oil
Water
Electricity

Export Gas Compression

Gas Export

Dehydration

Export Gas Compression

Export Gas Compression

Lift Gas Compression

Power Generation

Power Generation
Production
Well

Three-Phase
Separation

Export Oil Pumping

Oil Export

Injection Water Pumping

Figure 1.

Functional block diagram.

The well produces at its maximum 30, 000 m3 /d of oil,

which is the amount of oil which the separator can handle. The separated oil is exported by the export pumping unit, also with capacity of 30, 000 m3 /d of oil.
Off-gas from the separator is routed to the main
compressor unit, with two compressors running and
one standby a 2oo3 voting. Each compressor can
process a maximum of 3.0 MMscm/d. The nominal gas throughput for the system is assumed to be
6.0 MMscm/d, and the system performance will be
evaluated at this rate. Gas dehydration is required
for the lift gas, the export gas and the fuel gas. The
dehydration is performed by a 1 100% glycol contactor on the total gas flowrate, based on gas saturated
with water at conditions downstream of the compressor. The total maximum gas processing throughput
is assumed to be 6.0 MMscm/d, limited by the main
compression and dehydration trains.
To ensure the nominal level of production of the
well, the lift gas is supplied from the discharge of the
compression, after dehydration, and routed to the lift
gas risers under flow control on each riser.
An amount of 1.0 MMscm/d is compressed by the
compressor for lift gas and injected back into the
production well.
Water is injected into the producing reservoirs to
enhance oil production and recovery. The water separated in the separator and treated seawater is injected
in the field. The capacity of water injection system is
assumed to be 5, 000 m3 /d.

Table 1.

Transition rates of the components.

Transition rate (1/hr)

Component

Failure

Repair

Dehydration
Lift gas compressor
Export oil pump
Injection water pump
Three-phase separator
Export gas compressor
Power generation

3.49 104

8.33 102
6.98 102
3.66 102
1.33 102
19.6 102
4.29 102
3.24 102

6.57 104
7.06 104
2.27 104
4.25 104
6.69 104
1.70 103

The 25 MW power requirements on the production

system will be met by 2 17 MW gas turbine-driven
power generation units.
2.2 Component failures and repair rates
For simplicity, the study considers in details stochastic failure and maintenance behaviors of only the 2oo3
compressor system (one in standby) for the gas export
and the 2oo2 power generation system; the other
components have only two states functioning and
failed.
The transition rates of the components with only
two transition states are given in Table 1.

704

Table 2.

Summary of different production levels with the component failures.

Production
level (system
capacity, %)
100%
70%
70%
50%

50%
30%
0%

Example of
failure events

Oil
(km3 /d)

Gas
(MMscm/d)

Water
injection
(km3 /d)

None
Lift gas compressor
Water injection pump
Two export gas compressors
One power generator
Two export gas compressors
and one power generator
together
Two export gas compressors and
injection water pumping
Lift gas compressor and
injection water pump
Dehydration unit
All three export gas compressors
Both power generators

30
20
20
15

6
4
4
3

5
4
0
5

15

10

The compressor and power generation systems are

subjected to stochastic behavior patterns due to their
voting configuration. The failures and repair events
for both compressor and power generation systems are
described in Section 3.1 in detail.

c
i

2 i
0

i
2i

2.3

Production re-configuration

total(3i)

The failure of the components and systems are

assumed to have the following effects on the production level:
Failure of any component immediately causes the
production level to decrease by a step.
Failure of the lift gas compression or water injection
pump reduces the oil production by 10, 000 m3 /day
(30 % of total oil production rate) and the gas
production by 2.0 MMscm/day.
Failure of both the lift gas compression and injection water pumping reduces the oil production
by 20, 000 m3 /day and the gas production by
4.0 MMscm/day.
Failure of two export gas compressors or one generator forces the compression flow rate to decrease
from 6.0 MMscm/day to 3.0 MMscm, facing the
oil production rate to reduce accordingly from
30, 000 m3 /day to 15, 000 m3 /day.
Failure of the dehydration unit, all three export
gas compressors, or both power generator result in
system total shutdown.
The strategy of production reconfiguration against
the failure of the components in the system is illustrated in Table 2.

Figure 2.

State diagram of export compression system.

MAINTENANCE STRATEGIES

3.1 Corrective maintenance

Once failures happen in the system, it is assumed that
a corrective maintenance is immediately implemented
by only a single team apt to repair the failures. In case
that two or more components are failed at the same
time, the maintenance tasks are carried out according
to the sequence of occurrence of the failure events.
The failure and repair events of the export gas
compressor system, and power generation are more
complex than those of the other components. Figure 2
shows the state diagram of the export compression
system. As shown in Figure 2, common cause failures
which would result in the entire system shutdown are
not considered in the study. The compressors in the
export compression system are considered to be identical. The transitions from a state to another one are
assumed to be exponentially distributed.
The export compression system can be in four different states. The state 0 corresponds to two active

705

c
i

2i
0

Table 3. Schedule maintenance interval for compressors

and power generators.

Period,
month (year)

Maintenance
action

2 (1/6)
4 (1/3)
12 (1)

Detergent wash
Service/Cleaning
Baroscopic Inspection/
Generator check
Overhaul or replacement

60 (5)

Downtime
hr (day)
6 (0.25)
24 (1.0)
72 (3.0)
120 (5.0)

2i
Figure 3.

Input the system configuration

and component information

State diagram of power generation system.

Estimate the next transition time

for each component

compressors running at 100 % capacity. The state 1

is corresponds to one of the two active compressors
being failed and the third (standby) compressor being
switched on while the repair task is carried out; the
switch is considered perfect and therefore the state 1
produces the same capacity as the state 0. State 2
represents operation with only one active compressor
or one standby compressor (two failed compressors),
i.e., 50 % capacity; the export compression system
can transfer to state 2 by transition from either the
state 0 directly (due to common cause failure of 2 of
the three compressors) or from state 1 (due to failure
of an additional compressor). The state 3 corresponds
to the total system shutdown, due to failure of all three
compressors.
The same assumptions of the export compression
system apply to the power generation system, although
there are only 3 states given the parallel system logic.
The state diagram is shown in Figure 3.
Repairs allow returning to states of higher capacity
from lower ones.
3.2

Preventive maintenance

The following is assumed for the preventive maintenance tasks considered in the study:
Scheduled preventive maintenance is only implemented to the compressor system for the gas export
and to the power generation system.
Scheduled maintenance tasks of the compressors
and the power generation system are carried out at
the same time, to minimize downtime.
Well should be shutdown during preventive
maintenance.
The schedule maintenance intervals for both systems are given in Table 3.

Determine the shortest transition time

Perform the transition of the component

with the shortest transition time

Evaluate the system capacity

and production availability
Figure 4.

Flow chart for developed simulation program.

MONTE CARLO SIMULATION MODEL

The system stochastic failure/repair/maintenance

behavior described in Sections 2 and 3 has been modeled by Monte Carlo simulation and quantified by a
dedicated computer code.
4.1 Model algorithm
Figure 4 illustrates the flowchart of the Monte Carlo
simulator developed in the study.
First of all, the program imports the system configuration with the detailed information of the components
including the failure rates, repair times, preventive
maintenance intervals and required downtimes. Then,
the simulator estimates the next transition time for
all the components. Note that the next transition
time relies on the current state of each component.
When the component is under corrective or preventive maintenance, the next transition occurs after the
time required for the maintenance action. This maintenance time is predetermined. When the component

706

Table 4.
state.

Estimation of transition time depending on initial

Initial state

Transition time

Corrective
maintenance
Preventive
maintenance
Normal (including
partial load)

Time required for

corrective maintenance
Time required for preventive
maintenance (MTTR)
To be estimated by the direct
Monte Carlo method

Part list
Step 1: Will the stock-out has
direct effect on the offshore
production?

Step 2: Can the part requirement

be anticipated?

Order parts before

a demand occurs

Yes

Hold parts on
offshore platforms

No
Review/revise of maintenance
strategies

Figure 5.

NUMERICAL RESULTS

Delay scenario description

No delay for any preventive and
corrective maintenance
Periodical overhaul or replacement
(preventive maintenance) tasks for
the gas export compressor system
are delayed by 7 days
Corrective maintenance tasks for the
power generation system are delayed
by 7 days
Corrective maintenance tasks for
dehydration system are delayed
by 7 days

Without vs. with preventive maintenance

Firstly, the case in which components can fail and are

repaired following the rules of Sections 2 and 3 has
been considered, but with no preventive maintenance
performed. The average availability of this case turns
out to be 90.03 % .
The value of production availability reduces to 88.6
% with preventive maintenance. This reduction is due
to the unavailabilities of the components during the
maintenance action.
Effects of spare parts delay

Spare parts optimization is one of the most important

maintenance issues in offshore production operation.
Spare parts optimization shall give answers to the
following questions;
Shall spare parts be stored offshore or not?
How many spare parts should be held?
Monte Carlo simulation is an effective tool for
optimization and determination of spare parts requirements. An example of a spare holding decision algorithm generally used to determine spare holdings in
offshore or marine facilities is illustrated in Figure 5.
Monte Carlo simulation can be used to give a quantitative solution to the questions of each step explained

Spare holding decision algorithm [ABS. 2004].

Table 5. Comparison results of the production availability with the delay scenarios.

To investigate the influence of maintenance strategies

on the production availability some cases have been
simulated and compared.
The mission time of 30 years has been considered
as typical for offshore installations.

5.2

No

Yes

is in normal operation condition (not necessarily with

100% capacity), the next transition time is determined
by direct Monte Carlo simulation (Marseguerra & Zio,
2002). Table 4 summarizes how the transition time is
estimated, depending on the initial state.

5.1

No spares holding

Yes

Step3: Can the part be held on

offshore platforms?

No

(Average)
Production
availability
8.86 101

8.36 101
7.78 101
8.37 101

in Figure 6. For example, the effect of the stock--out on

the offshore installations is easily estimated by comparison of production availability results for two cases,
without or with delay of spare parts. Also, it is possible
to determine the required number of parts which must
be held to reduce the consequence of stock-out to an
acceptable level.
The results of the effect of delay of spare parts
on the production availability of the pilot model are
summarized in Table 5. It is assumed that corrective
or preventive maintenance tasks are not immediately
implemented due to stock-out on offshore installations. This situation, although unrealistic, is simulated
to understand the consequences of delay on the production availability. Usually, the rotating equipment

707

installed in offshore platforms is expected to have

a higher frequency of failures than others. And the
failure effect of the compressor and power generation system on production is classified as significant.
With the consideration of frequencies and consequences together, the effect of stock-out for such
systems on production availability should be estimated
in priority during the determination of maintenance
strategies.
6

CONCLUSIONS

A Monte Carlo simulation model for the evaluation

of the production availability of offshore facilities has
been developed.
A prototypical offshore production process is taken
as the pilot model for the production availability
assessment. A tailor-made computer program has been
developed for the study, which enables to account
for the operational transitions of the system components and the preventive and corrective maintenance
strategies.
The feasibility of application of the Monte Carlo
simulation to investigate the effect of maintenance
strategies on the production availability of offshore
installations has been verified.

As future study, it is of interest to formalize the preventive maintenance interval optimization and spare
parts optimization process with the Monte Carlo simulation. To this aim, it will be necessary to combine
the results of the availability assessment based on the
Monte Carlo simulation with the cost information.
The optimization of preventive maintenance intervals
should be determined based on an iterative process
where the overall availability acceptance criteria and
costs fall within the optimal region; the spare parts
optimization will consider the cost of holding different
numbers of spare parts and that of not holding any.
REFERENCES
ABS. 2004. Guidance notes on reliability-centered maintenance. ABS.
E. Zio, P. Baraldi, and E. Patelli. 2006. Assessment of
availability of an offshore installation by Monte Carlo
simulation. International Journal of Pressure Vessel and
Piping, 83: 312320.
Marseguera M. & Zio E. 2002. Basics of the monte carlo
method with application to system reliability. Hagen,
Germany:LiLoLe-Verlag GmbH.
NORSOK Standard (Z-016). 1998. Regularity management & reliability technology. Oslo, Norwat, Norwegian
technology standards institution.

708

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Reliability analysis of discrete multi-state systems by means

of subset simulation
E. Zio & N. Pedroni
Department of Energy, Polytechnic of Milan, Milan, Italy

ABSTRACT: In this paper, the recently developed Subset Simulation method is considered for improving the
efficiency of Monte Carlo simulation. The method, originally developed to solve structural reliability problems,
is founded on the idea that a small failure probability can be expressed as a product of larger conditional failure
probabilities for some intermediate failure events: with a proper choice of the conditional events, the conditional
failure probabilities can be made sufficiently large to allow accurate estimation with a small number of samples.
The method is here applied on a system of discrete multi-state components in a series-parallel configuration.

INTRODUCTION

Let us consider a system made of n components which

are subject to stochastic transitions between different states and to uncertain loading conditions and
mechanisms of deterioration. The resulting stochastic
system life process can be adequately modelled within
a probabilistic approach (Schueller 2007). Within
this framework, the probability of system failure is
expressed as a multi-dimensional integral of the form
P (F) = P(x F) = IF (x)q(x)dx

(1)

where x = {x1 , x2 , . . . , xj , . . . , xn } n is the vector of the random states of the components, i.e. the
random configuration of the system, with multidimensional probability density function (PDF) q :
n [0, ), F n is the failure region and
IF : n {0, 1} is an indicator function such that
IF (x) = 1, if x F and IF (x) = 0, otherwise.
In practical cases, the multi-dimensional integral
(1) can not be easily evaluated by analytical methods
nor by numerical schemes. On the other hand, Monte
Carlo Simulation (MCS) offers an effective means
for estimating the integral, because the method does
not suffer from the complexity and dimension of the
domain of integration, albeit it implies the nontrivial task of sampling from the multidimensional PDF.
Indeed, the MCS solution to (1) entails that a large
number of samples of the values of the component
states vector be drawn from q(); an unbiased and consistent estimate of the failure probability is then simply
computed as the fraction of the number of samples that
lead to failure. However, a large number of samples

(inversely proportional to the failure probability) is

necessary to achieve an acceptable estimation accuracy: in terms of the integral in (1) this can be seen as
due to the high dimensionality n of the problem and the
large dimension of the relative sample space compared
to the failure region of interest (Schueller 2007).
To overcome the rare-event problem, an efficient
approach is offered by Subset Simulation (SS), originally developed to tackle the multidimensional problems of structural reliability (Au & Beck 2001). In
this approach, the failure probability is expressed as
a product of conditional failure probabilities of some
chosen intermediate failure events, whose evaluation
is obtained by simulation of more frequent events.
The problem of evaluating small failure probabilities
in the original probability space is thus replaced by
a sequence of simulations of more frequent events
in the conditional probability spaces. The necessary
conditional samples are generated through successive Markov Chain Monte Carlo (MCMC) simulations
(Metropolis et al., 1953), gradually populating the
intermediate conditional failure regions until the final
target failure region is reached.
In this paper, SS is used for evaluating the reliability of a discrete multi-state system of literature
made of n components in a series-parallel logic (Zio
& Podofillini 2003).
The benefits gained by the use of SS are demonstrated by comparison with respect to a standard
MCS; finally, an analysis of the bias associated to the
estimates provided by SS is also provided.
The remainder of the paper is organized as follows.
In Section 2, a detailed description of the SS procedure is provided. In Section 3, the application to

709

the series-parallel, discrete multi-state system is illustrated. Finally, some conclusions are proposed in the
last Section.

2
2.1

SUBSET SIMULATION
Basics of the method

For a given target failure event F of interest, let F1

F2 . . . Fm be a sequence of intermediate failure events, so that Fk = ki=1 Fi , k = 1, 2, . . ., m. By
sequentially conditioning on the event Fi , the failure
probability P(F) can be written as
P(F) = P(Fm ) = P(F1 )

m1


P(Fi+1 |Fi )

(2)

i=1

Notice that even if P(F) is small, the conditional

probabilities involved in (1) can be made sufficiently
large by appropriately choosing m and the intermediate
failure events {Fi , i = 1, 2, . . ., m 1}.
The original idea of SS is to estimate the
failure probability P(F) by estimating P(F1 ) and
{P(Fi+1 |Fi ) : i = 1, 2, . . . , m 1}. Considering for
example P(F) 105 and choosing m = 5 intermediate failure events such that P(F1 ) and {P(Fi+1 |Fi ) :
i = 1, 2, 3, 4} 0.1, the conditional probabilities can
be evaluated efficiently by simulation of the relatively
frequent failure events (Au & Beck 2001).
Standard MCS can be used to estimate P(F1 ). On
the contrary, computing the conditional failure probabilities in (1) by MCS entails the non-trivial task
of sampling from the conditional distributions of x
given that it lies in Fi , i = 1, 2, . . . , m 1, i.e. from
q(x|Fi ) = q(x)IF (x)/P(F). In this regard, Markov
Chain Monte Carlo (MCMC) simulation provides a
powerful method for generating samples conditional
on the failure region Fi , i = 1, 2, . . . , m 1 (Au &
Beck 2001). The related algorithm is presented in the
next Section 2.2.
2.2

Markov Chain Monte Carlo (MCMC)

simulation

Markov Chain Monte Carlo (MCMC) simulation comprises a number of powerful simulation techniques for
generating samples according to any given probability
distribution (Metropolis et al., 1953).
In the context of the reliability assessment of interest in the present work, MCMC simulation provides an
efficient way for generating samples from the multidimensional conditional PDF q(x|F). The distribution
of the samples thereby generated tends to the multidimensional conditional PDF q(x|F) as the length of the
Markov chain increases. In the particular case of the

initial sample x 1 being distributed exactly as the multidimensional conditional PDF q(x|F), then so are the
subsequent samples and the Markov chain is always
stationary (Au & Beck 2001).
Furthermore, since in practical applications dependent random variables may often be generated by some
transformation of independent random variables, in
the following it is assumed without loss of generality that the
components of x are independent, that is,
q(x) = nj=1 qj (xj ), where qj (xj ) denotes the onedimensional PDF of xj , j = 1, 2, . . . , n (Au & Beck
2001).
To illustrate the MCMC simulation algorithm with
reference to a generic failure region Fi , let x u =
{x1u , x2u , . . ., xju , . . ., xnu } be the uth Markov chain sample drawn and let pj (j |xju ), j = 1, 2, . . ., n, be a
one-dimensional proposal PDF for j , centered at
the value xju and satisfying the symmetry property
pj (j |xju ) = pj (xju |j ). Such distribution, arbitrarily
chosen for each element xj of x, allows generating
a precandidate value j based on the current sample value xju . The following algorithm is then applied
to generate the next Markov chain sample x u+1 =
{x1u+1 , x2u+1 , . . ., xju+1 , . . ., xnu+1 }, u = 1, 2, . . ., Ns 1
(Au and Back 2001):
1. Generate a candidate sample x u+1 = {x1u+1 , x 2u+1 ,
. . ., x ju+1 , . . ., x nu+1 }: for each parameter xj , j =
1, 2, . . ., n, sample a precandidate value ju+1 from
pj (|xju ); compute the acceptance ratio rju+1 =
qj ( j u+1 )/qj (xju ); set x ju+1 = ju+1 with probability min(1, rju+1 ) and x ju+1 = xju with probability
1 min(1, rju+1 ).
2. Accept/reject the candidate sample vector x u+1 :
if x u+1 = x u (i.e., no precandidate values have
been accepted), set x u+1 = x u . Otherwise, check
whether x u+1 is a system failure configuration, i.e.
x u+1 Fi : if it is, then accept the candidate x u+1
as the next state, i.e., set x u+1 = x u+1 ; otherwise, reject the candidate x u+1 and take the current
sample as the next one, i.e., set x u+1 = x u .
The proposal PDFs {pj : j = 1, 2, . . ., n} affect
the deviation of the candidate sample from the current one, thus controlling the efficiency of the Markov
chain samples in populating the failure region. In particular, the spreads of the proposal PDFs affect the
size of the region covered by the Markov chain samples. Small spreads tend to increase the correlation
between successive samples due to their proximity
to the conditioning central value, thus slowing down
the convergence of the failure probability estimators.
Indeed, it can be shown that the coefficient of variation
(c.o.v.) of the failure probability estimates, defined as

710

the ratio of the standard deviation to the mean of the

estimate, increases as the correlation between the successive Markov chain samples increases. On the other
hand, excessively large spreads may reduce the acceptance rate, increasing the number of repeated Markov
chain samples, still slowing down convergence (Au &
Beck 2003).
2.3

The subset simulation algorithm

In the actual SS implementation, with no loss of generality it is assumed that the failure event of interest can
be defined in terms of the value of a critical response
variable Y of the system under analysis (e.g., its output
performance) being lower than a specified threshold
level y, i.e., F = {Y < y}. The sequence of intermediate failure events {Fi : i = 1, 2, . . . , m} can then
be correspondingly defined as Fi = {Y < yi }, i =
1, 2, . . . , m, where y1 > y2 > . . . > yi > . . . > ym =
y > 0 is a decreasing sequence of intermediate threshold values (Au & Beck 2001). Notice that since these
intermediate threshold values (i.e., failure regions) are
introduced purely for computational reasons in SS,
they may not have a strict physical interpretation and
may not be connected to known degradation processes.
The choice of the sequence {yi : i = 1, 2, . . . , m}
affects the values of the conditional probabilities
{P(Fi+1 |Fi ) : i = 1, 2, . . . , m1} in (2) and hence the
efficiency of the SS procedure. In particular, choosing
the sequence {yi : i = 1, 2, . . . , m} arbitrarily a priori
makes it difficult to control the values of the conditional probabilities {P(Fi+1 |Fi ) : i = 1, 2, . . . , m 1}
in the application to real systems. For this reason, in
this work, the intermediate threshold values are chosen
adaptively in such a way that the estimated conditional
failure probabilities are equal to a fixed value p0 (Au
& Beck 2001).
The SS algorithm proceeds as follows (Figure 1).
First, N vectors {x k0 : k = 1, 2, . . ., N } are sampled by
standard MCS, i.e., from the original probability density function q(). The subscript 0 denotes the fact
that these samples correspond to Conditional Level
0. The corresponding values of the response variable {Y (x k0 ) : k = 1, 2, . . ., N } are then computed
(Figure 1a) and the first intermediate threshold value
y1 is chosen as the (1 p0 )N th value in the decreasing list of values {Y (x k0 ) : k = 1, 2, . . ., N }. By so
doing, the sample estimate of P(F1 ) = P(Y < y1 ) is
equal to p0 (note that it has been implicitly assumed
that p0 N is an integer value) (Figure 1b). With this
choice of y1 , there are now p0 N samples among
{x k0 : k = 1, 2, . . ., N } whose response Y lies in
F1 = {Y < y1 }. These samples are at Conditional
level 1 and distributed as q(|F1 ). Starting from each
one of these samples, MCMC simulation is used to
generate (1 p0 )N additional conditional samples
distributed as q(|F1 ), so that there are a total of N

Figure 1. Illustration of the SS procedure: a) Conditional

level 0: Standard Monte Carlo simulation; b) Conditional
level 0: adaptive selection of y1 ; c) Conditional level 1:
Markov Chain Monte Carlo simulation; d) Conditional level
1: adaptive selection of y2 .

conditional samples {x k1 : k = 1, 2, . . ., N } F1 , at
Conditional level 1 (Figure 1c). Then, the intermediate threshold value y2 is chosen as the (1p0 )N th value
in the descending list of {Y (x k1 ) : k = 1, 2, . . ., N }
to define F2 = {Y < y2 } so that, again, the sample estimate of P(F2 |F1 ) = P(Y < y2 |Y < y1 ) is
equal to p0 (Figure 1d). The p0 N samples lying in
F2 are conditional values from q(|F2 ) and function
as seeds for sampling (1 p0 )N additional conditional samples distributed as q(|F2 ), making up a

711

total of N conditional samples {x k2 : k = 1, 2, . . ., N }

at Conditional level 2. This procedure is repeated
for the remaining conditional levels until the samples
at Conditional level (m 1) are generated to yield
ym < y as the (1p0 )N th value in the descending list of
{Y (x km1 ) : k = 1, 2, . . ., N }, so that there are enough
samples for estimating P(Y < y) (Au et al., 2007).
3

APPLICATION TO A SERIES-PARALLEL
DISCRETE MULTI-STATE SYSTEM

not reported for brevity; however, for clarity sake,

the synthetic parameters of the performance distributions (i.e., the mean vj and the standard deviation
vj , j = 1, 2, 3) are summarized in Table 1. Finally,
it is worth noting that the probability of the system
having performance W equal to 0, i.e. being in state
o = {0, 0, 0}, is1.364 103 (this value has been analytically obtained by calculating the exact probabilities
of all the 1331 available system states).
3.2

In this Section, SS is applied for performing the reliability analysis of a series-parallel discrete multi-state
system of literature (Zio & Podofillini 2003).
Let us consider a system made up of a series of =
2 macro-components (nodes), each one performing a
given function, e.g. the transmission of a given amount
of gas, water or oil flow. Node 1 is constituted by
n1 = 2 components in parallel logic, whereas node
2 is constituted by a single component (n2 = 1) so
that the overall number of components in the system

is n = 2b=1 nb = 3.
For each component j = 1, 2, 3 there are zj possible
states, each one corresponding to a different hypothetical level of performance, vj,o , o = 0, 1, . . ., zj 1.
Each component can randomly occupy the discrete
states, according to properly defined probabilities
qj,o , j = 1, 2, 3, o = 0, 1, . . ., zj 1.
In all generality, the output performance Wo associated to the system state 0 = {o1 , o2 , . . ., oj , . . ., on }
is obtained on the basis of the performances vj,o of
the components j = 1, 2, . . ., n constituting the system. More precisely, we assume that the performance
of each node b constituted by nb elements in parallel logic is the sum of the individual performances of
the components and that the performance of the node
series system is that of the node with the lowest performance, which constitutes the bottleneck of the
system (Levitin & Lisnianski 1999).
The system is assumed to fail when its performance
W falls below some specified threshold value w, so
that its probability of failure P(F) can be expressed
as P(W < w). During simulation, the intermediate
failure events {Fi : i = 1, 2, . . . , m} are adaptively
generated as Fi = {W < wi }, where w1 > w2 > . . . >
wi > . . . > wm = w are the intermediate threshold
values (see Section 2.3).

Case 2: 21 discrete states for each component

For each component j = 1, 2, 3 there are now zj =

21 possible states each one corresponding to a different hypothetical level of performance vj,o , o =
0, 1, . . ., 20; thus, the number of available system states
is now 213 = 9261. For clarity sake, the synthetic
parameters of the performance distributions (i.e., the
mean vj and the standard deviation vj , j = 1, 2, 3) are
summarized in Table 2. Finally, in this case, the probability of the system having performance W equal to
0, i.e. being in state o = {0, 0, 0}, is 1.671 104 .
3.3 Subset simulation parameters
In the application of SS to both Case 1 and Case 2,
the conditional failure regions are chosen such that a
conditional failure probability of p0 = 0.1 is attained
at all conditional levels.
In Case 1, the simulations are carried out for m =
3 conditional levels, thus covering the estimation of
failure probabilities as small as 103 .
Table 1. Parameters of the probability distributions of the
components performances for Case 1.
Performance distributions
parameters
Component, j

Mean

Standard deviation

1
2
3

56.48
58.97
92.24

25.17
23.11
11.15

Table 2. Parameters of the probability distributions of the

components performances for Case 2.
Performance distributions
parameters

3.1 Case 1: 11 discrete states for each component

For each component j = 1, 2, 3 there are zj =
11 possible states each one corresponding to a different hypothetical level of performance vj,o , o =
0, 1, . . ., 10; thus, the number of available system states
is 113 = 1331. The probabilities qj,o associated to the
performances vj,o , j = 1, 2, 3, o = 0, 1, . . ., 10, are

712

Component, j

Mean

Standard deviation

1
2
3

58.17
60.66
93.55

24.35
22.32
10.02

3.4.1 Failure probability estimation

3.4.1.1 Comparison with standard Monte Carlo
Simulation (MCS)
Figure 2 shows the failure probability estimates for
different threshold levels w, obtained in a single simulation run, for both Case 1 (top) and Case 2 (bottom).
The results produced by SS with a total of 840 samples (i.e., three simulation levels, each with N = 300
samples) and 1110 samples (i.e., four simulation levels, each with N = 300 samples) are shown in solid
lines. Note that a single SS run yields failure probability estimates for all threshold levels w up to the
smallest one considered (i.e. 103 and 104 for Cases
1 and 2, respectively). For comparison, the analytical failure probabilities (dashed lines) and the results
using standard MCS with 840 and 1110 samples (dotdashed lines) are shown in the same Figures for Cases
1 (top) and 2 (bottom), respectively.

0
10
Analytical
SS (NT = 840)

Failure probability,P(F)

MCS (NT= 840)

-1
10

-2
10

-3
10
0

80

100

80

100

SS (NT = 1110)

-1
10

MCS (NT = 1110)

-2
10

-3
10

-4
10

Discussion of the results

In this Section, the results of the application of SS to

the performance analysis of the system described in
Section 3 are illustrated with reference to both Case 1
and Case 2.

40
60
Failure threshold,w

Analytical

3.4

20

0
10

Failure probability,P(F)

At each conditional level, N = 300 samples are

generated. The total number of samples is thus NT =
300+270+270 = 840, because p0 N = 30 conditional
samples from one conditional level are used to start
the next conditional level and generate the missing
(1 p0 )N = 270 samples at that level.
The failure probability estimates corresponding to
the intermediate thresholds {wi : i = 1, 2, 3}, i.e.
101 , 102 and 103 , are computed using a total
number of samples equal to NT = 300, 570 and
840, respectively. It is worth noting that the number
of samples employed for estimating the probabilities of failure of the system is about 2 times lower
than the total number of available system states, i.e.
1331 (Section 3.1); thus, the computational time
required for estimating the failure probabilities by SS
is substantially lower than that necessary for analytically computing them (i.e., for calculating the exact
probabilities of all the 1331 system states).
Differently, in Case 2, the simulations are carried
out for m = 4 conditional levels, thus covering the
estimation of failure probabilities as small as 104 .
Also in this case, at each conditional level, N = 300
samples are generated such that the total number of
samples is now NT = 1110. The failure probability
estimates corresponding to the intermediate thresholds
{wi : i = 1, 2, 3, 4}, i.e. 101 , 102 , 103 and 104 ,
are computed using a total number of samples equal
to NT = 300, 570, 840 and 1110, respectively. Notice
that the number of SS samples used for estimating
the failure probabilities of the system is about 9 times
lower than the total number of available system states,
i.e. 9261 (Section 3.2).
In both cases, for each components performance
vj,o , j = 1, 2, 3, o = 0, 1, . . ., zj 1, the onedimensional discrete proposal PDF pj,o (j,o |j,o )
adopted to generate by MCMC simulation the random
pre-candidate value j,o based on the current sample
component j,o (Section 2.2) is chosen as a symmetric
uniform distribution, that is, pj,o (j,o |j,o ) = 1/(2lj +
1), if |o o| lj and pj,o (j,o |j,o ) = 0, otherwise,
j = 1, 2, 3, o = o lj , o lj + 1, . . ., o + lj 1, o + lj .
Notice that lj is the maximum allowable number of
discrete steps that the next sample can depart from the
current one. In both cases, the choice l1 = 2, l2 = 2
and l3 = 1 empirically turned out to offer the best
trade-off between estimation accuracy and relatively
low correlation among successive conditional failure
samples.

20

40
60
Failure threshold,w

Figure 2. Analytical failure probabilities (dashed lines) and

their corresponding estimates obtained by SS (solid lines) and
standard MCS (dot-dashed lines) in a single simulation run,
for Case 1 (top) and Case 2 (bottom).

713

Table 3. Mean relative absolute errors [P(F)] made

by both SS and standard MCS with 840 samples in the
estimation of the failure probability P(F) = 1.364
103 (Case 1); these values have been computed for
three batches of S = 200 simulations each.

P(F)
Batch 1
Batch 2
Batch 3

Table 4. Mean relative absolute errors [P(F)] made by

both SS and standard MCS with 1110 samples in the estimation of the failure probabilities P(F) = 1.942 103
and P(F) = 1.671 104 (Case 2); these values have been
computed for three batches of S = 200 simulations each.

Mean relative absolute errors

Mean relative absolute errors

SS

Standard MCS

SS

1.364 103

1.364 103

1.94103

1.67104

0.4327
0.4611
0.4821

0.7265
0.7530
0.6656

0.4181
0.4425
0.4960

0.6409
0.5611
0.6826

P(F)
Batch 1
Batch 2
Batch 3

In order to properly represent the randomness of

the SS and MCS procedures and provide a statistically
meaningful comparison between the performances of
SS and standard MCS in the estimation of a given failure probability P(F) of interest, S = 200 independent
runs of each method have been carried out. In each
simulation s = 1, 2, . . ., S the relative absolute error
s [P(F)] between the exact (i.e., analytically computed) value of the failure probability P(F) and the
corresponding estimate P s (F) (obtained by SS or standard MCS) is computed using |P(F) P s (F)|/P(F).
The performances of SS and standard MCS in the
estimation of P(F) are then compared in terms of the
mean relative absolute error
 [P(F)] over S = 200
runs, computed as 1/S Ss=1 s [P(F)]. This quantity gives an idea of the relative absolute error made
on average by the simulation method in the estimation of a given failure probability P(F) of interest in a
single run.
Table 3 reports the values of the mean relative absolute errors [P(F)] made by both SS and standard MCS
with 840 samples in the estimation of the failure probability P(F) = 1.36413 in Case 1 (Section 3.1): this
value has been chosen as target because it corresponds
to the probability of the system having performance W
equal to 0, i.e. being in state o = (0, 0, 0) which is the
most critical for the system and also the less likely one.
Table 4 presents the values of the mean relative
absolute errors made by both SS and standard MCS
with 1110 samples in the estimation of the failure probabilities P(F) = 1.942103 and P(F) = 1.671104
in Case 2 (Section 3.2). Only for illustration purposes, the results obtained in three batches of S = 200
simulations each are reported for both Cases 1 and 2.
It can be seen from these Tables that in all cases
the mean relative absolute errors made by SS are significantly (23 times) lower than those provided by
standard MCS using the same number of samples. In
particular, it is evident from Table 4 that as the target
probability of failure gets smaller, SS becomes more
and more efficient than standard MCS: for example,

Standard MCS
1.94103
0.6983
0.7793
0.6112

1.67104
1.6670
1.8915
1.6190

in the estimation of P(F) = 1.671 104 SS provides mean relative absolute errors which are even four
times lower than those produced by standard MCS (for
instance, see Batch 2 of Table 4). This result is quite
reasonable: in fact, the estimation of failure probabilities near 104 by means of standard MCS with
1110 samples is not efficient since on average only
1110 104 0.1 failure samples are available in the
failure region of interest. In contrast, due to successive
conditioning, SS guarantees that there are 840, 570,
300 and 30 conditional failure samples at probability
levels P(F) = 101 , 102 , 103 and 104 , thus providing sufficient information for efficiently estimating
the corresponding failure probabilities.
Finally, the computational efficiency of SS can be
compared with that of a standard MCS in terms of the
coefficient of variation (c.o.v.) of the failure probability estimates computed from the same number of
samples.
The sample c.o.v. of the failure probability estimates obtained by SS in S = 200 independent runs
are plotted versus different failure probability levels
P(F) (solid line) in Figure 3, for both Case 1 (top)
and Case 2 (bottom). Recall that the number of samples required by SS at the probability levels P(F) =
101 , 102 , 103 and 104 are NT = 300, 570, 840
and 1110, respectively, as explained in Section 3.3.
The exact c.o.v. of the Monte Carlo estimator using
the same number of samples at probability levels
P(F) 
= 101 , 102 , 103 and 104 are computed

using (1 P (F)) P (F) NT , which holds for NT
independent and identically distributed (i.i.d.) samples: the results are shown as squares in Figure 3, for
Case 1 (top) and Case 2 (bottom). It can be seen that
while the c.o.v. of the standard MCS grows exponentially with decreasing failure probability, the c.o.v. of
the SS estimate approximately grows in a logarithmic
manner: this empirically proves that SS can lead to
a substantial improvement in efficiency over standard
MCS when estimating small failure probabilities.

714

1.4

10
SS (NT = 840) (average 200 runs)

MCS
uncorrelated (lower limit)
fully correlated (upper limit)

NT = 840

Failure probability,P(F)

Coefficien of variation (c.o.v)

1.2

0.8
0.6
NT = 570

0.4

-1

10

-2

10

Analytical
SS (NT = 840) (average 200 runs)

NT = 300

0.2

-3

0 -3
10

-2

-1

10

10
10
Failure probability,P(F)

10

10

20

30

40
50
60
70
Failure threshold,w

80

90

100

100

3.5
SS (NT = 1110) (average 200 runs)
MCS
uncorrelated (lower limit)

-1

Failure probability,P(F)

Coefficientof variation (c.o.v)

3 N = 1110
T

fully correlated (upper limit)

2.5
2
1.5
NT = 840

10

-2

10

Analytical

-3

10

SS (NT = 1110) (average 200 runs)

NT = 570

0.5
NT = 300

0
-4
10

-4

10

-3

-2

10
Failure probability,P(F)

-1

10

10

10

10

20

30

40

50

60

70

80

90

100

Failure threshold,w

Figure 3. Coefficient of variation (c.o.v.) versus different

failure probability levels P(F) for Cases 1 (top) and 2 (bottom). Solid line: sample average over 200 SS runs; dashed
lines: sample average of the lower bound over 200 SS runs;
dot-dashed lines: sample average of the upper bound over
200 SS runs; squares: standard MCS (i.i.d. samples).

3.4.1.2

SS estimates: Bias due to the correlation

among conditional failure samples
To assess quantitatively the statistical properties of the
failure probability estimates produced by SS, the sample mean of the failure probability estimates obtained
in S = 200 independent runs have been computed.
For a given failure probability level P(F) of interest,
the sample mean P(F) of the corresponding estimates

P s (F), s = 1, 2, . . ., S, is P(F) = 1/S Ss=1 P s (F).
Figure 4 shows the sample means of the failure
probability estimates obtained by SS, for both Case 1
(top) and Case 2 (bottom) (solid lines); a comparison with the exact (i.e., analytically computed) failure
probabilities is also given (dashed lines).
The sample means of the failure probability estimates almost coincide with the analytical results,
except at small failure probabilities, near 103 and
104 , where the estimates seem to be quite biased.

Figure 4. Analytical failure probabilities (dashed lines) and

sample averages of the failure probability estimates over 200
SS runs (solid lines) for Case 1 (top) and Case 2 (bottom).

A quantitative indicator of the bias associated to the

estimate of a given failure probability P(F) can be
computed as the relative absolute deviation [P(F)]
between the exact value of the failure probability, i.e.
P(F), and the sample average P(F) of the corresponding estimates, [P(F)] = |P(F) P(F)|/P(F).
Table 5 reports the values of the sample means P(F)
and the corresponding biases [P(F)] produced by
SS in the estimation of P(F) = 1.364 103 in Case 1
(Section 3.1); Table 6 presents the values of the same
indicators referred to the estimation of the failure probabilities P(F) = 1.942103 and P(F) = 1.671104
in Case 2 (Section 3.2). Only for illustration purposes, the results obtained in three batches of S = 200
simulations each are reported for both Cases 1 and 2.
It is evident from Table 6 that the bias of the estimates significantly increases as the target probability
of failure decreases: for instance, in Batch 2 the bias
associated to the estimate of P(F) = 1.942 103
is 0.1865, whereas the one related to the estimate of
P(F) = 1.671 104 is 0.2928. This leads to conclude

715

states, discrete multi-state systems can only occupy a

finite number of states; as a consequence, the generation of repeated (thus, correlated) conditional failure
samples during MCMC simulation may be significant.

Table 5. Sample means P(F) of the failure probability estimates over 200 SS runs and the corresponding
biases [P(F)] produced by SS in the estimation of
P(F) = 1.364 103 (Case 1); these values have been
computed for three batches of S = 200 simulations
each.

CONCLUSIONS

Subset simulation

In this paper, SS has been applied for the reliability

assessment of a system of discrete multi-state components connected into a logic structure. An example of
a simple series-parallel system of literature has been
taken for reference.
The results of SS have been compared to those of
standard Monte Carlo Simulation (MCS) in the estimation of failure probabilities as small as 104 . The
results have demonstrated that as the target probability of failure gets smaller, SS becomes more and more
efficient over standard MCS.
Finally, a word of caution is in order with respect to
the fact that the estimates produced by SS when applied
to discrete multi-state systems may be quite biased
if the number of discrete states is low. This is due
to the correlation between the conditional probability
estimators at different levels: in fact, differently from
continuous-state systems whose stochastic evolution
is modeled in terms of an infinite set of continuous
states, discrete multi-state systems can only occupy a
finite number of states; as a consequence, the number of repeated (thus, correlated) conditional failure
samples generated during MCMC simulation may be
high. Further research is underway on attempting to
estimate the bias.

P(F) = 1.364 103

Batch 1
Batch 2
Batch 3

Sample mean

Bias

1.136103
1.145103
1.065103

0.1672
0.1606
0.2192

Table 6. Sample means P(F) of the failure probability

estimates over 200 SS runs and the corresponding biases
[P(F)] produced by SS in the estimation of P(F) =
1.942 103 and P(F) = 1.674 104 (Case 2); these values
have been computed for three batches of S = 200 simulations
each.
Subset simulation

Batch 1
Batch 2
Batch 3

P(F) = 1.942103

P(F) = 1.671104

Sample mean

Bias

Sample mean

Bias

1.714103
1.579103
1.715103

0.1170
0.1865
0.1164

1.374104
1.181104
1.347104

0.1769
0.2928
0.1934

that the bias due to the correlation between the conditional probability estimators at different levels is not
negligible (Au & Beck 2001).
This finding is also confirmed by the analysis of
the sample c.o.v. of the failure probability estimates
which are plotted versus different failure probability
levels P(F) (solid line) in Figure 3, for both Case 1
(top) and Case 2 (bottom). In these Figures, the dashed
lines show a lower bound on the c.o.v. which would
be obtained if the conditional probability estimates
at different simulation levels were uncorrelated; on
the contrary, the dot-dashed lines provide an upper
bound on the c.o.v. which would be obtained in case
of full correlation among the conditional probability
estimates. From these Figures, it can be seen that the
trend of the actual c.o.v. estimated from 200 runs follows more closely the upper bound, confirming that
the conditional failure probability estimates are almost
completely correlated in both Case 1 and Case 2. The
high correlation between conditional probability estimates may be explained as follows: differently from
continuous-state systems whose stochastic evolution
is modeled in terms of an infinite set of continuous

REFERENCES
Au, S. K. & Beck, J. L. 2001. Estimation of small failure probabilities in high dimensions by subset simulation.
Probabilist. Eng. Mech. 16(4): 263277.
Au, S. K. & Beck, J. L. 2003. Subset Simulation and its
application to seismic risk based on dynamic analysis.
J. Eng. Mech.-ASCE 129(8): 117.
Au, S. K., Wang, Z. & Lo, S. 2007. Compartment fire analysis by advanced Monte Carlo simulation. Eng. Struct.,
in press (doi: 10.1016/j.engstrct.2006.11.024).
Levitin, G. & Lisnianski, A. 1999. Importance and sensitivity analysis of multi-state systems using the universal
generating function method. Reliab. Eng. Syst. Safe. 65:
271282.
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N. &
Taller, A. H. 1953. Equations of state calculations by fast
computing machines. J. Chem. Phys. 21(6): 10871092.
Schueller, G. I. 2007. On the treatment of uncertainties in
structural mechanics and analysis. Comput. Struct. 85:
235243.
Zio E. & Podofillini, L. 2003. Monte Carlo simulation analysis of the effects of different system performance levels
on the importance of multi-state components. Reliab. Eng.
Syst. Safe. 82: 6373.

716

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The application of Bayesian interpolation in Monte Carlo simulations

M. Rajabalinejad, P.H.A.J.M van Gelder & N. van Erp
Department of Hydraulic Engineering, Faculty of Civil Engineering, TUDelft, The Netherlands

ABSTRACT: To reduce the cost of Monte Carlo (MC) simulations for time-consuming processes (like Finite
Elements), a Bayesian interpolation method is coupled with the Monte Carlo technique. It is, therefore, possible
to reduce the number of realizations in MC by interpolation. Besides, there is a possibility of thought about
priors. In other words, this study tries to speed up the Monte Carlo process by taking into the account the prior
knowledge about the problem and reduce the number of simulations. Moreover, the information of previous
simulations aids to judge accuracy of the prediction in every step. As a result, a narrower confidence interval
comes with a higher number of simulations. This paper shows the general methodology, algorithm, and result
of the suggested approach in the form of a numerical example.

of randomly generated data in the unqualified area is

presented by a numerical example.

INTRODUCTION

The so-called Monte Carlo (MC) technique helps engineers to model different phenomena by simulations.
However, these simulations are sometimes expensive
and time-consuming. This is because of the fact that
the more accurate models, usually defined by finite
elements (FE), are time-consuming process themselves. To overcome this problem the cheaper methods
are generally used in the simulation of complicated
problems and, consequently, less accurate results are
obtained. In other words, implementing more accurate models in the Monte Carlo simulation technique
provides more accurate and reliable results; by the
reduction of calculations cost to a reasonable norm,
more accurate plans for risk management are possible.
To reduce the cost of Monte Carlo simulations for a
time-consuming process (like FE), numerous research
projects have been done, primarily in the structural
reliability to get the benefits of not only a probabilistic approach but also to obtain accurate models. For
instance, importance sampling and directional sampling are among those approaches implemented to
reduce the cost of calculations. But still this coupling is
a time-consuming process for practical purposes and it
should be still modified. This research tries to speed up
the Monte Carlo process by considering the assumption that the information of every point (pixel) can
give an estimation of its neighboring pixels. Taking
the advantage of this property, the Bayesian interpolation technique (Bretthorst 1992) is applied to our
requirement of randomness of the generated data. In
this study, we try to present a brief review of the
method and important formulas. The application of
the Bayesian interpolation into the MC for estimation

GENERAL OUTLINES

In the interpolation problem, there is a signal U which

is to be estimated at a number of discrete points.
These discrete points will be called pixels, presented
by ui . These pixels are evenly spaced on a grid of pixels u (u0 , , uv+1 ). Therefore, there are totally
v + 2 pixels. The first and last pixels are called boundary pixels and are treated separately. These boundary
pixels are presented by u0 and uv+1 . As a result, v
presents the number of interior pixels. The total number of observed data points is equal to n which are
distributed in arbitrary (or random) locations among
the pixels. Therefore, the maximum value of n is
equal to v + 2 when there is an observed data point
for each pixel (n v + 2). The locations of the
observed data points are collected in a vector c, so
this vector has n elements which are presented by ci
and i = 1, 2, , n. The vector of observed data points
is called d (d1 , , dn ), and its elements are presented by di . Figure 1 presents an illustration of the
internal and boundary pixels as well as data points.
According to this figure c (1, v 1, v + 2).
3

BAYESIAN INTERPOLATION

The univariate posterior probability density function

(PDF) for an arbitrary pixel uj, given the data d and
the prior information I , will be found by integrating out
all pixels. In this case the sum rule is applied and the

717

following model, f , for an arbitrary pixel ui .

ui = f (ui1 , ui+1 ) =
Figure 1. An illustration of the pixels which data points are
assigned to.

product is integrated all over the multivariate posterior

PDF of all pixels of u except the required pixel uj .
P(u|d, I ) . . . dui . . .
  

ei = ui f (ui1 , ui+1 ) = ui

(2)



1
1
P(ei |) =
exp 2 ei2
2
2

Where P(d|I ) is a normalization constant called

evidence. Therefore, combination of Equations 1 and
2 produces the following equation.

P(uj |d, I )

P(d|u, I )P(u|I ) . . . dui . . .

  

(3)

i=j

This equation presents that to obtain the posterior,

we need to define our likelihood function and the prior.
The likelihood, or in this case more appropriate the
PDF of the data (d) conditional on the pixels (u), is
constructed by making the standard assumptions of
noise. Therefore, according to the Bayesian interpolation technique, there are three main steps should be
taken into account:
1. All the pixels are connected to each other so each
pixel is defined as a function of its neighboring pixels. This is the prior information which is
formulated in Section 4.
2. For the pixels which take the corresponding data
values, the data values are considered the best
estimates. This is described in Section 5.
3. Then the outcome of the previous steps are combined so as to get an estimation of every pixel in
grid, based on the data. In this case, Equation 3 is
used and the result is presented in Section 6.

(5)

The only thing we know about this error is that the

error has a mean of zero (the error is either positive
or negative) with some unknown variance 2 . Using
the principle of Maximum Entropy (Jaynes 2003), we
find the well known Gaussian probability distribution
function of ei presented in Equation 6.

Also, according to the Bayes rule we have:

P(d|u, I )P(u|I )
P(d|I )

ui1 + ui+1
2

(1)

i=j

P(u|d, I ) =

(4)

Having the model defined, the error ei also is

implicitly defined by Equation 5.


P(uj |d, I ) =

ui1 + ui+1
2

THE PRIOR

We expect some logical dependence between neighboring pixels and this expectation is translated in the

(6)

Substituting Equation 5 into Equation 6 and making

the appropriate change of variable from ei to ui , the
PDF of the pixel ui can be obtained by Equation 7.
P(ui |ui1 , ui+1 , )




1
1
ui1 + ui+1 2
exp 2 ui
=
(7)
2
2
2
Assuming that there is no logical dependence
between the errors e1 , . . . , ev , the multivariate PDF
of all the errors is a product of the univariate PDFs.
Then, by making the change of variable from ei to ui
we find the following multivariate PDF for the pixels
u1 , . . . , uv .
P(u1 , . . . , uv |u0 , uv+1 , )



v 
1 
ui1 + ui+1 2
1
exp 2
ui
=
2 i=1
2
(2)v/2 v
(8)
The boundary pixels are treated separately. In fact,
these two pixels are assigned to the first and last position and presented as u0 = v1 and uv+1 = vv+2 . As
a result of using the principle of Maximum Entropy,
the PDF of the boundary pixel u0 is obtained in Equation 9. And a similar equation can be established for

718

the pixel uv+1 .

P(u0 |u1 , )



1
1
exp 2 [u0 u1 ]2
=
2
2

(9)

Combining Equations 8 and 7 using Bayes Theorem, the next equation will be obtained. This equation
is written in a matrix form where u is vector of pixel
positions,
P(u0 , u1 , . . . , uv+1 |)
1

(2 )(v+2)/2 v+2

Q
exp 2
2


(10)

where
Q = uT Ru
and

1.5

0.5

.
R
..

..
.

.
..
0

1.5

0.5

0.5

2
..
.
..
.
..
.

3
..
.

2
..
.

0.5
..
.

..
.
..
.
..
.

0.5
..
.

0.5
0

We have derived the above equation which provides

the PDF for the pixels u0 , . . . , uu+1 using the assumed
prior model presented in Equation 4.
If = 0, we get to the conclusion that our model
(Equation 4) holds exactly. So setting = 0 produces
an extremely informative prior which determines the
values of the pixels. On the other hand, if inf
then the prior relaxes to an extremely uninformative
distribution which lets the values of the pixels totally
free. So in a sense regulates the freedom allowed
to the pixels u0 , . . . , uv+1 .

2
3
0.5 1.5

0
..
.
..
.
..
.

0.5

1.5
1

THE LIKELIHOOD

Apart from our model and prior, we also have n+2 nonoverlapping data points , n v . These data points can
be assigned arbitrarily to any pixel uc where c is an element of the vector c described in Section 2. The value
of c corresponds with the location of the observed data
regarding the pixel numbers (see Figure 2). The error
of the model at the location of any observed data point
is defined as:
ec = uc dc

(11)

Assuming that this error has a mean of zero (the

error is either positive or negative) with some unknown
variance 2 and using the principle of Maximum
Entropy we find that this error has the following
probability distribution function:


1
1
exp 2 ec2
P(ec | ) =
2
2

(12)

Figure 2. An illustration of the pixels which data points are assigned to. The - is a representation of the evaluated values
in the pixels.

719

Substituting 11 into 12 and making a change of variable from the error ec to the data dc , the likelihood
function can be obtained according to Equation 13.


1
P(dc |uc , ) =
exp 2 (dc uc )2
2
2
1


(13)

Again by assuming logical independence between

the errors and making the appropriate substitutions and
changes of variables, the following likelihood function
can be obtained.
P(d1 , . . . , dn |u0 , u1 , . . . , u(v+1) , )

1
1 
2
(dc uc )
=
exp 2
(2 )n/2 n
2 cc

(14)

( Sivia 1996). Equation 19 presents the matrix form of

this function.
P(u|d, , ) P(u|)P(d|u, )


1
(d Su)T (d Su) uT Ru
= v+2 n exp

2 2
2 2
(17)
Equation 19 is conditional on unknown parameters
and , but since we dont know these parameters
we will want to integrate them eventually out as nuissance parameters. We first assign Jefferys prior to
these unknow parameters:
P() =

P( ) =

(18)

Equation 14 can be rewritten in the matrix form as

presented in Equation 15.

Using Bayes theorem we can combine the priors

in Equation 18 with Equation 19 to get the following
equation.

P(d1 , . . . , dn |u1 , . . . , un , )


1
1

T
exp

(d

Su)
(d

Su)
=
(2 )n/2 n
2 2

P(u, , |d) = P( )P()P(u|d, , )



(d Su)T (d Su) uT Ru
1

v+3 n exp

2 2
2 2
(19)

(15)
Where d is a padded vector of length v + 2 where
the data points have been coupled with their corresponding pixels and S is a diagonal matrix with entry
1 for pixels which data points are assigned to them and
0 everywhere else.
For example, the S matrix for the grid in Figure 1,
d = (0, d1 , 0, , 0, d2 , 0, d3 ) becomes:

0 0
0
.. . .
..
.
.
.
.. . .
. 0
.
..
.
. ..
0

0
..
.

..
.
0
..
.
0
..
..
.
.

0
0

0
0

0
0

0
..
.
..
.
..
.

By integrating over all pixel except the target pixel,

uj , the probability distribution function of just one
pixel (uj ) is given as
P(uj |d, , )

= P(u0 , . . . , uv+1 |d, , ) du0 . . . duv+1





0
1


exp

except duj

(d Su)T (d Su) uT Ru

2 2
2 2

1
du0 . . . duv+1


()v+3 n 

(20)

except duj

(16)
For actual evaluation of Equation 20, we refer the
interested reader to ( Bretthorst 1992).
7

ALGORITHM

To couple the Bayesian interpolation approach with

Monte Carlo techniques the following algorithm is
suggested:

THE POSTERIOR

Combining the prior in Equation 10 and with the likelihood presented in Equation 15 we get a function which
is proportional to the posterior PDF of all the pixels

1. Define the interval of variation and the length of

pixels for the variable X . Totally, v + 2 pixels are
to be defined in this interval.

720

2. A random number is generated according to the

PDF of the variable X , and according to its value, its
assigned to a certain location. This location which
is the jth pixel (as presented in Figure 1) is called uj .
3. According to the information of the other pixels and
our assumed model, the PDF of the uj is calculated
by Equation 20.
4. According to the accepted tolerance criteria, it is
decided whether there is a need to calculate the
limit state equation for the jth point or the accuracy
is enough.
5. The calculations are iterated from step 3 and
continues to meet the simulation criteria.
8

NUMERICAL EXAMPLE

One of the important research topics in hydraulic engineering focuses on the impact of water waves on walls
and other coastal structures, which create velocities
and pressure with magnitudes much larger than those
associated with the propagation of ordinary waves
under gravity. The impact of a breaking wave can generate pressures of up to 1000 kN /m2 which is equal
to 100 meters of water head. Although many coastal
structures are damaged by breaking waves, very little
is known about the mechanism of impacts. Insight into
the wave impacts has been gained by investigating the
role of entrained and trapped air in wave impacts. In
this case, a simplified model of maximum pressure of
ocean waves on the coastal structures is presented by
Equation 21.
Pmax = C

p k u2
d

(21)

Where the is density of water, k is the length

of hypothetical piston, d is the thickness of air cushion, u is the horizontal velocity of the advancing wave,
and C is a constant parameter and equal to 2.7 s2 /m.
Having this knowledge, we are willing to find the
probability of the event, when the maximum impact
pressure exceeds 5105 N /m2 for a specific case.
The one dimensional limit state function (LSF) can be
defined by Equation 22, where the velocity parameter
is assumed to be normally distributed as N (1.5, 0.45).
G(u)5 0.98280 u2

(22)

We consider the variation of the variable u in the

interval of [ 3 , + 3 ] where is the mean
value and is the standard deviation of variable u.
This interval is divided to the finite pixels with an
equal distance of 0.01. As a result, there are totally 270
internal pixels defined in this interval. A schematic
view of the all pixels is presented in Figure 2. Pixel 210
is considered as a sample pixel in which we are going

to monitor the change of its PDF during Monte Carlo

process. In this figure, the measured (or observed) data
point is assigned to the first and last internal pixels.
Before we proceed to the simulation process, we
would like to present the probable values of pixel 210
or u210 with the suggested model. Therefore, we need
to use Equation 20 in order to get the required PDF.
Nevertheless, this equation contains and . As a matter of fact, can be integrated out of the equation, but
we need to estimate a value for the . In this case, we
define  = 1 which is called regularizer. Then we can
get the PDF of our regularizer to find its optimal value
which leads to the most narrow PDF. The reader who
is interested in this process is referred to ( Bretthorst
1992). The most probable value of  is estimated to be
2.6 and we use this value during the rest of this work.
As a result, Equation 20 will lead to Equation 23 for
pixel number 210 given two data points: d1 and d2 .
P(uj=210 |d1 , d2 )
=

0.5897 1010

0.3126 109
+ 0.5265 1010 uj + 0.1339 1010 uj 2
(23)

The PDF of u210 given d1 and d2 is depicted in

Figure 3. This figure is a plot of Equation 23. The
mean value of this PDF is -1.97 by assuming a symmetrical PDF. Besides, the 95% accuracy by assuming
a symmetrical distribution leads to the values in the
interval of [-11.28, 7.35]. This interval was obtained
by solving the equation which defines the integration
of a symmetrical area around mean value should be
equal to 0.95. It is a wide PDF and its tails are much
more informative than the Gaussian. In other words,
we expect value of this pixel vary within the interval
having the prior information about the model and just
2 data points.
It is useful to compare this result with the traditional
interpolation problem. In fact, by considering two data
points, there is no other way than we assume a linear
relationship which leads to the value of -1.21 for this
pixel while we do not have any estimation about the
uncertainty. Now, the distinction between two methods is obvious; the applied method enables us to get a
criterion for the uncertainty of the estimated value of
each pixel. This is an huge advantage in the simulation
process. This comparison is depicted in Figure 4. In
this figure there are two data points called A and B.
These two points are the only information which provide point e using a linear interpolation for the pixel
210, where e = 1.21. This is not close to real value
of the limit state function g = 0.0246. Nevertheless,
there is no information over the certainty of the estimated point e from the interpolation. In the other hand,
point f is the mean value of the PDF calculated by the

721

is assigned to a pixel uj , we check if it is necessary to

run the limit state equation, or we can assign its value
regarding our tolerance. To investigate the changes,
we monitor the u210 after 20 realizations of the LSE
(or 20 data points) which are assigned to their location.
As a result, the calculated PDF of u210 given 20 data
points is obtained and depicted in Figure 5. The mean
value of this PDF is 0.013 , and the 95% accuracy
by assuming a symmetrical distribution leads to the
values in the interval of [0.16, 0.19]. This shows
that by implementing more data points, we get a more
precise PDF.
The difference of the results of linear and Bayesian
interpolation at this case is because of the value of
the regularizer(). In this case study its value is set
to be  = 2.60. The effect of epsilon (or which
is inversely related to it) was previously described. In
fact, we can have two extreme situations when we consider two extreme values for . These extreme values
are 0 and infinity. In the first case we just stick to our
data values and in the second case we just consider
our model assumption and leave the other information. Therefore, the difference between e and f should
be related to the value of regularizer.
Since we are not satisfied with the accuracy we continue to generate more data points. Figure 6 presents
the PDF of u210 having 90 data points measured or
calculated. The mean value of this PDF is 0.025 , and
the 95% accuracy by assuming a symmetrical distribution leads to the values in the interval of [0.014,
0.035]. This shows that by implementing more data

Figure 3. This figure presents the probability distribution

function (PDF) of the u210 given 2 measured data points: d1
and d270 .

Figure 4. A comparison between linear interpolation and

the Bayesian technique for the pixel u210 , given 2 measured
data points: d1 and d270 . The exact value of the function
(Equation 22) is depicted by dashed line.

Bayesian technique (f = 1.97). The uncertainty is

shown by its PDF. Having a look at the presented PDF,
a rather wide PDF can be seen; and both of positive
and negative values are expected for this pixel.
From now on we start the Monte Carlo simulation by
generating random numbers. However, before we run
the limit state equation for each random number which

Figure 5. This figure presents the probability distribution

function (PDF) of the u210 given 20 measured data points
given in random locations.

722

points, we get a more precise PDF. Since this interval is small enough, we can assume that we have got
the enough accuracy. Therefore, the simulation effort
has been reduced by 67% for the presented numerical
example.
In fact, the number of simulations in the Monte
Carlo technique depends on several factors. The most
important ones are the tolerance and the distance
between pixels defined for the analysis. In other words,
to get a more precise result we need to implement more
data points. Meanwhile, a higher number of pixels lead
to a higher accuracy.

It is useful to compare the calculated PDFs in

another figure with the same scale. Figure 7 provides
this comparison in which figure (a) presents the PDF
of the pixel at the beginning of simulation where there
are just two data points. Figure 7 (b) presents the PDF
of the same pixel, u210 , when there are 20 data points
randomly generated and assigned to the related pixels. Figure 7(c) again presents the PDF of the same
pixel where the information of ninety pixels are implemented. In this figure, the same scale of axis is selected
to clarify the change of the PDF during the simulation
process.
9

Figure 6. This figure presents the probability distribution

function (PDF) of the u210 given 90 measured data points
given in random locations.

DISCUSSION

The Bayesian interpolation is a technique which can

be nicely coupled with Monte Carlo simulation. In this
study, an attempt is made to have the prior information
of the model incorporated to the current level of the
analysis. This is a step forward in Monte Carlo simulations. In fact, the R in Equation 10 provides a link
between the information of each pixel with its neighborhood. In other words, information of each point
passes through this link and effect the others. Besides,
this approach provides a nice tool to get the other
priors incorporated to the Monte Carlo simulations.
For instance, the limit bounds method ( Rajabalinejad
et al.2007 assumes some other prior information which
can be implemented in this approach.
Nevertheless, the approach presented in this paper
has got two main limitations. The first one is that we
need to use a grid and divide the interval of variation
of the variable to finite number of pixels. The second
limitation is that the pixel are evenly spaced. These
conditions impose large matrices for a large interval,
a small step size, or higher dimensions.

Figure 7. This figure shows the probability distribution function of variable which is assigned to the pixel j = 210. In
Figure (a) Just the information of 2 data points are considered while in Figure (b) and (c), the information of 20 and 90 pixels
are considered, respectively.

723

10

CONCLUSION

The suggested procedure can speed up the Monte Carlo

simulations integrated with finite elements or the other
highly complicated and time consuming processes.
However, in this paper we have limited ourselves into
the finite number of pixels. The proposed method also
provides a tool for implementing informative priors
regarding the considered model. The extension of this
work with an arbitrary length and location of pixels
can provide a more powerful tool and is recommended
for future research projects.
REFERENCES

Jaynes, E. T. (2003). Probability Theory, the Logic of Science.

Cambridge University Press.
Rajabalincjad, M., P. H. A. J. M. van Gelder, and J. Vrijling
(2007). Dynamic limit boundaries coupled with monte
carlo simulations. Submitted to the Journal of Structural
Safety.
Rajabalincjad, M., P. H. A. J. M. van Gelder, J. K. Vrijling, W. Kannng, and S. van Baars (2007). Probabilistic
Assessment of the Flood Wall at 17th Street Canal, New
Orleans. In Risk, Reliability, and Social Safety, Volume
III, pp. 2227.
Sivia, D.S. (1996). Data Analysis: A Bayesian Tutorial.
Clarendon Press.
van Gelder, P. H. A. J. M. (1999). Risks and safety of flood
protection structures in the Netherlands. In Participation
of Young Scientists in the Forum Engelberg, pp. 5560.

Bretthorst, G. L. (1992, july). Bayesian interpolation and

deconvolution. Cr-rd-as-92-4, The advanced sensor diretorate research, Alabama 358985000.

724

Occupational safety

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Application of virtual reality technologies to improve occupational &

industrial safety in industrial processes
Jokin Rubio, Benjamn Rubio, Celina Vaquero, Nekane Galarza, Alberto Pelaz & Jess L. Ipia
Industrial Safety Unit, Fundacin LEIA CDT, Spain

Diego Sagasti
Divisin de Realidad Virtual, EUVE- European Virtual Engineering, Spain

Luca Jord
Unidad de Ingeniera de Producto, AIMME-Instituto Tecnolgico Metalmecnico, Spain

ABSTRACT: Virtual Reality (VR) is emerging as an important tool in the industry sector to simulate humanmachine interaction providing significant findings to improve occupational and industrial safety. In this paper
several VR applications and tools developed for industries in the manufacturing and chemical sector are presented.
These developments combine VR simulations, immersive 3D dynamic simulation and motion capture, addressing
risk assessments in the design process, in normal operation and training.

INTRODUCTION

Virtual Reality is commonly understood as a computer simulation that uses 3D graphics and devices
to provide an interactive experience. It is being used
nowadays in multiple sectors as entertainment (video
games), industry, marketing, military, medicine . . . In
the industry sector, this technology is emerging as an
important tool to simulate and evaluate the humanmachine interaction, especially when this fact may
have important consequences on the safety of people,
processes and facilities or in the environment.
VR training applications are probably the most
broadly used. Military, aeronautic and nuclear industries develop virtual scenarios in order to train operators to avoid risky and costly operations in the real
word. Also, simulators of fork-lift trucks and gantry
cranes have been developed to help worker to reduce
labour risks (Helin K. 2005, Iriarte et al., 2005).
Beyond the training applications, some research has
been done to study the capabilities of VR as a technology to study different safety aspects (risk identification, machinery design, layout and training of
operators) in a manufacturing factory (Mtt T.,
2003). Probably the sector that has shown greater
interest in these technologies in the recent years has
been the automotive industry, where VR is being used
not only as a technology oriented to the final user

(ergonomic and comfortable cars), (Monacelli, G,

2003), but also to reduce costs and time of fabrication and to develop safety and ergonomic assembly
lines for the workers.
Nevertheless, despite of these examples and other
initiatives oriented to improve industrial and occupational safety, VR is still an expensive technology
mainly reduced to research but not accessible to the
majority of the companies.
The Industrial Safety Unit (UDS) in Fundacin
LEIA has an emerging research line focused in VR
technologies, combining virtual reality simulation of
workplaces and processes, immersive 3D dynamic
simulation and operator motion capture. Until now,
several applications and tools, addressed to companies of manufacturing and chemical sectors, have
been developed: 1-risk assessment (safety machinery,
tasks performance and ergonomic assessments), 2-risk
control (implementation & verification of industrial
safety & ergonomic requirements in the design and
redesign of machinery, devices, workplaces and processes) and, 3-operators risk communication, education and training (especially workplace postural
training tools).
In this paper, the work performed by this group and
the results obtained are presented, as well as the future
research in the application of the VR technologies to
other sectors.

727

METHODOLOGY: IMMERSIVE VIRTUAL

REALITY SYSTEM

RESULTS

3.1 Workplace risk assessment

The Industrial Safety Unit (UDS) is provided with a
fully equipped virtual reality laboratory. The capabilities of this laboratory can be summarized as follows
(see Figure 1).
Biomechanical Motion Analysis, provided with IR
video cameras, optical markers, data glove, and
motion capture software tools. This system allows a
movement optical capture and biomechanical analysis of those movements and provides the virtual
reality system with machine vision.
CAD Design and Dynamic Simulation (IBM
Work Station + CATIA/DELMIA software tools).
CATIA/DELMIA work-station is connected to a
Virtual Reality Equipment (next item) and supplies initial virtual environment in order to build
the Virtual Universe.
Immersive Virtual Reality, VirTools v4.0 software
tool (virtual reality engine, the brain of the virtual reality system) and hardware elements, HMD
helmet, data glove, projectors, screens, glasses.
Moreover, as a result of R&D projects (made jointly
with the companies ABGAM and STT), this system
has been optimized with modules to export motioncaptures to DELMIA and to perform automatic risk
ergonomic assessments.
This virtual reality laboratory is being used by this
research group mainly for three safety related applications: workplace risk assessment, design improvement
and to develop education and training applications.

Figure 1. Schematic
capabilities.

illustration

of

the

3D simulation technologies and immersive virtual

reality are being used in UDS to assess processes
and facility risks. In this sense, evaluations of safety
machinery, tasks performance and ergonomics assessments are being developed (Figure 2).
Workers performing their tasks are simulated in
Catia-Delmia software. Mainly safety distances and
worker postures when performing tasks are analysed to
achieve an evaluation of the workplace, using the modules of the software, that include several ergonomic
standards and tools.

3.2 Industrial process design improvement

Immersive VR is being used in the design phase of
the industrial process to simulate labour tasks that
imply safety and ergonomic risks. This is nowadays
being carried out in a collaborative project (TIES) with
another two research centres, AIMME and EUVE.
A system based on CAD design, biomechanical
motion analysis and immersive virtual reality has been
developed (Figure 1). The basic approach being followed in this research is to inmerse final users
in interactive virtual work scenarios developed from
3D machinery designs (Catia, Solid Edge). The uservirtual scenario interaction is analysed in order to
identify the safety/ergonomic risks of those work
environment designs. In this way, designs can be modified prior to the real machinery prototypes/workplace
development, offering a great flexibility in the design
process.
A real example has been developed in order to test
the performance of this system. A specific workplace,
including a hydraulic press, has been modelled. Experimental testing with users has provided preliminary

UDS-Lab
Figure 2.

728

Example of workplace assessment.

results regarding the usefulness of the system to

improve industrial processes design.
The specific scenario and machinery addressed in
this experience was selected attending to its extensive
use in the metallic transformation sector (key sector
in the TIES mentioned project) and the high risks
involved in the manipulation of this specific machinery. Three main steps have been carried out in this
work: virtual scenarios development, ergonomic and
safety modules programming and initial testing.
As a first step, virtual scenarios were developed.
This activity implied mainly the programming of the
avatar and the machinery. The avatar was modelled
to move in the virtual word following the data from
the motion capture system, that is, following the real
movements of the user. User is provided with a Helmet
Mounted Display (HMD), so he can observe the virtual
world while moving within. User is also provided with
a data glove to monitor hand movements, allowing the
user to interact with the virtual world and to unleash
actions like catching and releasing a piece (manipulation of workloads). On the other side, the machinery
(a hydraulic press) was imported to the graphical motor
from its original design in Solid Edge. The mechanical behaviour was modelled according to the press
specifications (the scope of this project only focused
in the mechanical characteristics; other as electrical or
hydraulic where not considered because of the difficult
modelling in our virtual system). Different operating
modes and safety measures were programmed in order
to achieve different work scenarios (some examples
can be seen in Figure 3).
As a second step, software modules incorporating
ergonomic and safety requirements were developed
and integrated in the virtual system. In the interaction
of the user with the virtual word, those developed modules provide with deviation from recommended standard values. Specifically, standard ergonomic methods RULA, REBA, NIOSH and Snook&Ciriello have
been implemented, allowing an automatic ergonomic
evaluation from the motion capture system data (see
Figure 4) . Similarly, safety standards with a focus on
the mechanical risks of the press were reviewed. The
standard UNE-EN294:1993 (Safety distances to prevent danger zones being reached by the upper limbs)
has been programmed so distances among danger
zones, safety measures and user can be checked.
Finally, preliminary testing has been performed
with the developed system (Figures 56).
Users provided with input/output devices have been
immersed in the virtual word, where they have performed their work tasks loading/unloading pieces and
operating the press.
The initial experiences showed that the system
allows the identification of safety and ergonomic risks
of an industrial process in the design phase. Mainly, the
automatic ergonomic evaluation seems to be of special

Figure 3. Operating modes and safety measures of the

modeled press.

Figure 4.

Example of a result of an ergonomic evaluation.

relevance and usefulness; these evaluations have been

validated through comparison with standards methodologies (video recording plus application of ergonomic
standards). On the other side, the programming of
safety standards for mechanical risks into the virtual

729

Figure 7.

Software tool for training.

developed in an industrial process where the training was identified as a key factor for the occupational
health and safety improvement. The tool was prepared
to be used directly by industry trainers and it was
based in a mixture of real video records, 3D simulation (CATIA/DELMIA) and ergonomic evaluation
recognised methodologies. With these software tools,
workers are showed their bad body positions in the
executions of their tasks, the ergonomics risks derived
from them and the correct body positions they should
adopt and the improvement derived. (see Figure 7).

Figure 5 & 6.
workplace.

Immersive virtual reality for the design of a

system offered some assessment results, although the

experience showed that it seems to be more practical
to include this kind of modules in the CAD systems
used in the machinery design process.
Nevertheless, some difficulties have arisen in the
use of immersive VR, mainly derived of visual technical limitation of the HMD used (not stereo), addressing users to have precision problems when virtually
interact with the work environment/machinery. These
problems were partially corrected using mixed virtual reality providing with tactile feedback to users.
It is also planned to make new experiences using an
stereo HMD.
3.3

Workforce education and training

FUTURE PERSPECTIVES

Virtual Reality is currently being used as a promising technology to improve industrial and occupational
health and safety.
In this paper, some applications of this technology
developed by this research group have been presented.
In the near future, our research group plans three
main activities. Firstly, to perform more experiences
with immersive VR, analysing different alternatives to
improve the technical difficulties drafted above, and
developing/testing more work environment scenarios.
Second, to export the gained background to other
industrial sectors, mainly the education and training
applications. Finally, it should be mention that the
experience obtained from these developments is being
used to initiate experiences in order to use VR technology to improve physical and cognitive disabled people
labour integration (Lpez de Ipia, J. 2007).

ACKNOWLEDGMENTS

Tools to train workers in the ergonomic performance

of tasks have been developed and are being used by
the UDS. Specifically a postural training tool was

TIES project was funded by the Spanish Ministry of

Industry, FIT-020600-2006-29.

730

REFERENCES
Helin K. et al., Exploiting Virtual Environment Simulator
for the Mobile Working Machine User Interface Design.
VIRTSAFE Seminar. 04-06.07.2005, CIOP-PIB, Warsaw.
Iriarte Goi, X. et al., Simulador de carretillas elevadoras para
la formacin y disminucin en riesgos laborales: motor
grfico y simulacin dinmica. 7 Congreso iberoamericano de ingenieria mecanica, Mxico D.F., 12 al 14 de
Octubre de 2005.
Monacelli G., Elasis S.C.P, VR Applications for reducing time and cost of Vehicle Development Process. 8th
Conference ATA on Vehicle Architecture: Products, Processes and Future Developments, Florence, 2003.05.16.

Marc J., Belkacem N., MArsot J. Virtual reality: a design

tool for enhanced consideration of usability validation
elements. Safety Science 45 (2007) 589601.
Mtt T. Virtual environments in machinery safety analysis.
VTT Industrial Systems, 2003.
Lpez de Ipia, J.M., Rubio J., Rubio B., Viteri A.,
Vaquero C., Pelaz A., Virtual Reality: A Tool for the
Disabled People Labour Integration. Challenges for Assistive Technology AATE 07, European Conference for the
Advancement of Assistive Technology, San Sebastin,
2007.

731

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Applying the resilience concept in practice: A case study from the oil
and gas industry
Lisbeth Hansson & Ivonne Andrade Herrera
SINTEF, The Foundation for Scientific and Industrial Research at the University of Trondheim

Trond Kongsvik
NTNU Social Studies LTD, Norway

Gaute Solberg
StatoilHydro

ABSTRACT: This paper demonstrates how the resilience concept (Hollnagel et al., 2006) can be used as a
perspective for reducing occupational injuries. The empirical background for the paper is a case study on an
oil and gas installation in the North Sea that had a negative trend in LTI (Lost Time Injury) rates. The HSE
(Health, Safety, Environment) administration initiated a broad process that included the crew on the installation,
the onshore administration and a group of researchers to improve the situation. Instead of focusing the analysis
on incident reports, we applied a proactive view. Thus, we adapted a model for resilience that was used in a
development process. In the context of occupational accidents, we focused on the following factors: sufficient
time, knowledge and competence, resources and including working environment. These factors have been
identified as important for complexity and necessary for the organization to be able to anticipate, perceive and
respond to different constellation of conditions. This paper illustrates to what extent the concept of resilience
was fruitful analytically and as a reflection tool in the development of new HSE measures that are now being
implemented. The links between the resulting HSE measures and the qualities of the resilience concept are
discussed.

1
1.1

INTRODUCTION
Background

The empirical background for this paper is a case study

on an oil and gas installation in the North Sea. The
starting point for the study was historical data that
showed a negative trend in LTI rates. We as an outside
research group were engaged in an effort to improve
their prevention efforts. During our involvement we
tried to shift the focus in the organization from a reactive to a more proactive view on safety. By using an
action research approach (Greenwood & Levin 1998)
we tried to disengage from the focus on the negative
safety results and introduced resilience as an alternative concept for reflecting upon safety. More concrete,
the research question we will try to illuminate in this
paper is: How can resilience be built in practice in
organizations?
Many different strategies are used in safety work
in the oil and gas industry. What many of them have
in common is that history is used to learn about the

future. One example is accident investigations that are

carried out to identify causes and implement measures
to avoid similar incidents to occur. This is in line with
an engineering approach to safety that Hollnagel
(2008) name Theory W. Basic assumptions here are
that systems are tractable, and that systems should be
designed to avoid variability that in it self is regarded
as a threat to safety.
One problem with using historical data as an
approach is that conditions that produce dangerous situations and safety conditions are both complex and
dynamic. The complexity refers to the interdependency of causes and that a specific constellation of
conditions often has to be present for accidents to
occur. Thus, it is often difficult to reduce accidents
to simple cause and effect relationships. The social
and technological context in which accidents occur is
also in constant flux and can be regarded as moving
targets that are difficult to grasp through traditional
accident investigations.
The resilience concept (Hollnagel et al., 2006) represents an approach that implies building mechanisms

733

that make the organization prepared for the unexpected. This can be regarded as a proactive approach
to safety. In a proactive view individuals and organizations must adjust to cope with current conditions.
These adjustments handle different constellations of
conditions that can produce accidents and also successes. Thus a resilient organization (or system) can
adjust its functioning prior to or following changes
and disturbances to continue working in face of continuous stresses or major mishap. Here, variability is
regarded as potentially positive for safety, in line with
what Hollnagel (2008) label Theory Z.
The study is limited to one installation, and can be
regarded as a case study. This implies the exploration
of a bounded system over time involving several
data sources rich in context (Creswel 1998).
1.2

The approach chosen for the case study was action

research combined with use of the resilience concept. In practice, this implied a mutual reflection in
a search conference upon how to strengthen certain
qualities that could make Heidrun TLP more resilient
as an organization. The negative safety results were not
addressed at all in the workshop. Instead we focused
on the positive aspects of the organization and on how
to strengthen these even further.
In the next part of the paper the resilience concept
will be further explored, followed by a description of
how the concept was used in our case. In part four,
the results will be presented, while we in part five will
give an overall discussion of the findings, followed by
the conclusions.
2

The case: Heidrun TLP

The case study was conducted on Heidrun TLP, which

is an oil and gas producing installation in the North
Sea. It is a large tensioned leg platform (TLP) built in
concrete that has operated since 1995. When Heidrun
was designed and built it was regarded as a North Sea
Rolls Royce with sufficient space and high quality equipment. The personnel working on Heidrun are
proud of the installation and feel a strong ownership
towards their working place.
Heidrun TLP is operated by the largest petroleum
company in Norway, StatoilHydro. Approximately
50 % of the workers on the installation are employed by
StatoilHydro while the rest are employed in companies
responsible for drilling and production, companies
responsible for the maintenance and modification
work and another company within StatoilHydro is
responsible for the catering activities onboard. The
personnel on the installation follows the ordinary work
schedule as used in the offshore business, two weeks
on and four weeks off duty.
The offshore industry has a high reputation regarding safety. The safety awareness is high among the
employees and compared to other industries the statistics on occupational accidents are low. Still, the
frequency of occupational accidents has increased during the last years, also on Heidrun TLP. This may be
a consequence of multiple reasons but some of the
causes mentioned are increased personnel turnover,
more inexperienced personnel and a higher activity
level caused by extended modification work. The contract personnel are more exposed to accidents than the
StatoilHydro employees. One explanation for this is
that the contractors are engaged in the most risky and
complex operations. In general the work schedule in
the offshore business is regarded as challenging for the
workers; being away from normal life and the family
for two weeks is one challenge and it is also challenging to change into a work mindset after four week off.

THEORY: THE RESILIENCE CONCEPT

IN SAFETY RESEARCH

The term resilience is not a new concept within

safety. Foster (1993) defined resilience as an ability to accommodate change without a catastrophic
failure, or the ability to absorb shock graceful. Rosness et al., (2004) adapted the resilience definition
to the capacity of an organization to accommodate
failures and disturbances without producing serious
accidents. Resilience has also been defined as the
properties of an organization to make it more resistant
to its operational hazards (Reason and Hobbs 2003).
The resilience engineering term has been discussed
as a transition from traditional view to a systemic
view of safety. (Hollnagel, 2007a)
In the traditional view, accidents are seen as a result
of failure or malfunction of components (humans or
machines). These failures follow a predefined path of
cause and effect. Accident models and their metaphors
provide basis for the prevention alternatives. Accident prevention recommendation based in the Domino
Model (Heinrich, 1930) finds and eliminates causes
of specific accidents, and allows responding to this
specific unwanted event. In the same way, accident prevention recommendations based in the Swiss
cheese model (Reason, 1990) focus on strengthen any
barriers and defences. In the traditional view, risks are
seen as a linear combination of cause and effect and
safety is achieved by constraint variabilities.
Due to technological improvements there has been
a transition from tractable systems to intractable systems. Sociotechnical systems are becoming so complex that work situations are always underspecified.
The designers can not anticipate every contingency.
These systems are no longer bimodal and normal performance is variable. In accordance with Hollnagel
(2007b), performance variability is related to technological systems malfunctions, imperfections and to the
humans that have the tendency to adjust to current

734

In this context, the qualities required for a system

to be resilient are illustrated in Figure 1.
These qualities are related to the ability to:
Respond to regular and irregular threats in a robust
and flexible manner. This is the reactive part of
safety management. The system is designed to a
limited range of responses. There is still a necessity
to adjust responses in a flexible way to unexpected
demands.
Monitor in a flexible way own performance and
external conditions. This monitoring focused on
what it is essential to the operation. In a dynamic
and unpredictable environment, it is required for the
system to be able to have internal monitoring and

Up d a

System

ting

Dynamic developments

Anticipation

Attention

Response

Knowing
what to expect

Knowing
what to look for

Knowing
what to do

Knowledge

Competence
Time

ing

Since it is not possible to describe in detail all

operations and resources are limited, performance
variability is necessary
Many adverse events could contribute to a success
or to a failure. These adverse events are the result
of adaptations to cope with complexity
Safety management must be reactive and proactive. Safety management shall take into account
both hindsight and the ability of the organisation
(system) to make proper adjustments to anticipate
potential threats, monitor risk, revise risk models
and to use resources proactively.

Environment

Learn

conditions. Therefore, the performance variability of

the sociotechnical systems is normal and necessary
resulting in both successes and failures. In the systemic view, accidents and incidents are the result of
unexpected combinations of normal performance variability. Accidents are prevented by monitoring and
damping variability. In this view, risks emerge from
non-linear combination of performance variability.
Hollnagel (2008) defines formally resilience engineering as the intrinsic ability of an organization (or
system) to adjust its functioning prior to or following changes and disturbances to continue working in
face of continuous stresses or major mishaps. It is
not a surprise that there is no unique way to define
resilience. While the majority of definitions focused
on the capability to cope with failures providing a
reactive approach, resilience engineering focused on
the ability to adjust prior to or following a failure.
Resilience engineering explores ways that enhance the
ability of the organizations to be robust and flexible and make the organizations prepared to cope with
the unexpected. This definition focused on variability,
adaptability and unpredictability
We explore resilience engineering and the premises
for resilience engineering will have an influence on
the understanding of the phenomena that we studied
and the solutions that we identified (Hollnagel, 2007).
These premises are:

Resources

Figure 1. Required qualities of a resilient system

(Hollnagel, Woods and Leveson, 2006).

monitor the external conditions that may affect the

operation.
Anticipate risk and opportunities. At this point it is
required to go beyond risk analysis and have imagination to see what may happen and see key aspects
of the future (Westrum, 1993). Is is not only of identify single events but how the may be interact and
affect each other.
Learn from experience implies from actual events
not only collection of data in databases.
In resilience engineering safety is not seen as
the absence of an accident but a dynamic non-event
(Weick and Sutcliffe, 2001) and the capability of the
system to handle unexpected situations. Resilience
acknowledges that individuals and organizations must
adjust to cope with current conditions. These adjustments are always approximate due to current working
conditions where there is a limited amount of information, resources and time. Resilience Engineering
is about to increase the ability of the organisation to
make correct adjustments. The adjustments are influenced by a number of condition, these conditions are
lack of time, lack of knowledge, lack of competence
and lack of resources (Hollnagel and Woods, 2005).
These conditions will facilitate the system to cope with
the unexpected event.
Unexpected events require more time to understand
the situation and decide the proper action. If unexpected events occur in several occasions, they will
affect other activities and there is a possibility of loose
of control. The focus in relation to time should be
when time demands are real and have consequences
for individuals. Knowledge is required to understand
the event and what happened and competence is
related to knowing what to do even if the unexpected
event has gone beyond design limits. An unexpected
event will require the use of resources to regain control.
Finally, the experienced learned from the management
of the unexpected events need to go back to the system
in order to augment response capacity.

735

The resilience engineering concept presented in this

section is adapted to the oil and gas case to analyse
the ability of the organisation to anticipate, monitor,
respond and learn together with the conditions that
influence this ability

METHOD: ADAPTATION AND USE

OF THE RESILIENCE MODEL

Based on the increasing number of occupational accidents, StatoilHydro initiated a project in cooperation
with SINTEF. The scope was to turn the negative
trend identifying safety measures dedicated to the pilot
Heidrun TLP.
The action research approach that was used on Heidrun TLP is a part of an organizational development
(OD) process (Greenwood & Levin 1998). The goal of
this process was to increase the quality of adjustments
and to improve occupational safety in the organization.
In the OD process we focused on working conditions
that influence the ability to make proper adjustments;
sufficient time, knowledge, competence and resources
(Hollnagel and Woods, 2005). In our early information
gathering we saw that the psychosocial work environment on Heidrun TLP also could have a significant
influence on the ability to make proper adjustments,
and therefore added this as a condition.
With this background we developed a working
model that was used to structure the interviews and
as a starting point for the discussions in the search
conference.
Knowledge and Competence was merged in
the model due to pedagogical reasons (although the
difference was explained orally) and the factor psychosocial work environment was added, resulting in
four main factors as indicated in the Figure 2 below.
Based on initial discussions with the safety staff
some external factors where identified, anticipated to
influence the safety level on Heidrun TLP. These external factors are; The safe behaviour programme (a
large safety campaign), open safety talk, cost &

Figure 2.

The adapted resilience model.

contracts, organisational changes, onshore support and last but not least management.
All factors in the resilience model were discussed
through approximately 40 semi-structured interviews,
mostly by offshore workers. All kind of positions
where covered; both internal StatoilHydro and contractor employees. The main findings were extracted
from the interviews and sorted by the factors in the
resilience model.
The findings from the interviews were the main
input to the creative search conference, also gathering around 40 persons. The participants represented
both on and off shore personnel and internal and external StatoilHydro personnel. The two day conference
was arranged with a mix between plenum and group
sessions to discuss and suggest safety measures. The
second day of the conference was dedicated to identification of measures; i.e. how could Heidrun become
a more resilient organization.
4

RESULTS

The safety measures identified in the search conference were sorted and defined as HSE activities. These
were presented and prioritized in a management meeting in the Heidrun organization and the end result from
the project was nine HSE activities;

Safety conversations
Buddy system
Collaboration in practise
The supervisor role
Consistent management
Clarification of the concept visible management
Meeting/session for chiefs of operation
Risk comprehension course
Visualisation of events

These activities are now put into action. Safety

conversations cover both formal and informal conversation were safety is a topic either explicit or
indirectly. The aim of this measure is first of all to
enhance the quality of these conversations by observing good and less good practice and applying training
when necessary. In the Buddy system colleagues
are assigned to take care of new colleagues. This HSE
activity will contribute to an enhanced quality of this
system by observation of practice, exchange of experience and training. In the Collaboration in practice
activity different work groups are brought together to
be more familiar with own work in relation to others
work. The aim is to clarify roles and responsibilities in the common work processes and to increase
knowledge about each others work.
The supervisor role is a role that needs to be
developed and clarified as this role has changed. The
supervisor is the daily manager for the work force on

736

the installation and he has direct contact with the crew

and has a thorough knowledge about the operations.
This activity will aim at clarifying this role and identify
need for enhanced competence. Consistent management will help the managers to agree on common
practice on reactions to insecure behaviour.
The crew onboard the installation request more
Visible management, but at the same time the management claim that they have too little time to be
visible. It is however rather diffuse what is meant by
this expression and the activity will help to clarify this.
Meeting/session for the chiefs of operation shall be
an arena for good and constructive discussions about
safety related topics. This activity will define topics to be addressed and will contribute in design of
fruitful processes for these meetings. Risk comprehensive course shall develop different courses with
the aim to enhance the comprehensive of risk. Finally
the Visualisation of events activity will follow up
and extend visualisation of events through animations
and video and will also encourage the use of drawings
in reporting of events.
5

DISCUSSION

Three main qualities are required for a resilient organization. These are anticipation, attention and response.
These qualities are described in a theoretical way in the
theory section but as an introduction to the discussion
we will give a practical example related to occupation
accidents.
If a group of people onboard an oil installation
shall install a heavy valve together, they need to be
well coordinated. They need knowledge about how
to carry out the operation including who is responsible for what. Competences on the risky situations
they go through in this operation are also essential.
This knowledge represents anticipation, knowing
what to expect. When the operation proceeds they also
need to have competence on how to interpret the situation and what to look for to be aware of the risky
situation, attention is needed. When a risky situation is observed it is crucial that they respond
to it and respond in a correct way. It is not unusual
that an employee do not respond if he sees that an
employee in a higher position do not follow safety procedures. Trust is essential to secure response. Time and
resources is also important to avoid that critical situation are not responded to because they like to get the
job done in due time.
How can the identified HSE activities potentially
influence attention, anticipation and response? The
following Table 1 shows how we interpret this.
The activity Safety conversation covers all conversations where safety is an issue and the purpose is
to enhance the quality of these conversations. When

Table 1. Activities influencing anticipation, attention

and response.
Anticipation
Safety conversation
Buddy system
Collaboration in practice
The supervisor role
Consistent management
Clarification of visible
management
Session for chiefs of
operation
Risk perception course
Visualization

x
x
x
x
x

Attention

Response

x
(x)

(x)
(x)

(x)
(x)

(x)

(x)
x
(x)
(x)

(x)
(x)

safety issues are treated in a proper way they will

increase the knowledge about safety, and clarify anticipations and what to expect. Safety conversations can
also influence attention, e.g. what to look for in terms
of hazards in their daily work. One purpose of safety
conversations between employees and managers is
also to increase the awareness of how to respond to
critical situations.
The Buddy system will in itself contribute to
make newcomers to the installation more familiar and
increase the competence both about the installation
and how work is performed. Increasing the quality of
this system and giving the buddies support so that
they can be more prepared for this role may improve
the newcomers anticipation, attention and response.
Collaboration in practice will especially give a
better clarification of what to expect (anticipation)
regarding how the work that is carried out in a safe
manner.
The supervisors are close to the daily operations.
By increasing their knowledge and skills, this may
have an important effect on anticipation. Indirectly and
dependent on the skills the supervisors acquire, both
attention and the quality of response may increase.
The goal of the activity Consistent management
is to give managers a common understanding of how
to respond to safe and non- safe behavior. Consistent
positive and negative feedback that is regarded fair can
potentially increase both anticipation and attention.
Response regarded as unfair can worsen the psychosocial working environment and thereby decreasing the
two qualities.
A management that is visible to the employees in
terms of safety issues can in itself have a positive effect
on attention. The activity Clarification of visible
management will in the first stage only define the
meaning of the term and will thereby not contribute
to resilience before something is done to make the
managers more visible.

737

Introducing safety issues in meetings for chiefs of

operations in a positive way can increase the managers
knowledge about safety - anticipation.
Both risk comprehension course and Visualization of events can increase knowledge about safety
(anticipation) and also competence on how to be aware
of the risky situations (attention), but this effect is
dependent on a high quality and a proper use.
We see that most of the activities can potentially
improve the anticipation of risk and opportunities.
More uncertain are the influences on appropriate
responses to threats and also on attention; the monitoring of performance and conditions. Attention and
response are the two qualities that are most difficult to
change or improve. Both attention and response can be
regarded as behavior. Thus a change in these two qualities require a behavioral change. Anticipation can be
regarded as a cognitive process, and is as such easier to
change than behavior. Still, behavior change is crucial
in the building of resilience. How the nine activities
actually contribute to behavior change is still an open
question, as the effects have not yet been evaluated.
6

CONCLUSION

The research question for this paper was how resilience

can be built in practice in organizations. We have illustrated that the use of an action research approach, using
search conferences potentially could have a positive
influence on qualities that are required for resilient
organizations; anticipation, attention and response.
Our focus has been occupational injuries, but the
approach could be valid for safety work in general.
The approach and process used in the case study
demonstrates that a proactive approach to safety issues
is motivating for the personnel involved. Statistics and
reports on accidents are widely used to improve safety.
Some fatigue can be observed among the personnel
related to safety work using this approach. The feedback from this project was that the personnel had no
difficulties dealing with the resilience concept as it was
used in the project. Resilience was a construct thatthe

offshore personnel liked to be associated with. One

of the participants in the search conference, a safety
delegate, expressed that this was the most interesting
HSE meeting he had participated in during the last
25 years. The terms from the resilience model have
been adapted and used during daily safety work on the
installation. We may conclude that it is more motivating to use the proactive approach in practical safety
improvement work.

REFERENCES
Creswell, J.W. 1994. Research design: Qualitative & quantitative approaches. Thousand Oaks, California: Sage
Publications.
Greenwood, D.J. & Levin, M. 1998. Introduction to action
research: social research for social change. Thousand
Oaks, California.: Sage Publications.
Heinrich, H.W. 1931. Industrial accident prevention: New
York: McGraw-Hill.
Hollnagel, E., Woods, D. 2005. Joint Cognitive Systems.
Foundations of Cognitive Systems Engineering. Taylor
and Francis, USA.
Hollnagel, E., Leveson, N., Woods, D. 2006. Resilience
Engineering Concepts and Precepts, Aldershoot, Ashgate
Hollnagel, E. 2007a. Resilience Engineering: Why, What
and How. Viewgraphs of presented at Resilient Risk
Management Course, Juan les Pins, France.
Hollnagel, E. 2007b. Principles of Safety Management
Systems: The Nature and Representation of Risk. Viewgraphs of presented at Resilient Risk Management Course,
Juan les Pins, France.
Hollnagel, E. 2008. Why we need Resilience Engineering.
Ecole des Mines de Paris, Sophia Antipolis, France
Reason, J., Hobbs, A. 2003. Managing Maintenance Error,
Ashgate, Aldershot, USA.
Weick, K., Sutcliffe, M. 2001. Managing the unexpected.
Assuring High Performance in the Age of Complexity.
University of Michigan Business School Management
Series John Wiley & Sons, Inc. USA.
Westrum, R. 1993. Cultures with Requisite Imagination. In
Verification and Validation of Complex Systems: Human
Factors Issues, ed. Wise, J, Hopkin, D and Stager, P.
New York: Springer-Verlag, pp 401416.

738

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Development of an assessment tool to facilitate OHS management

based upon the safe place, safe person, safe systems framework
A.-M. Makin & C. Winder
School of Risk and Safety Science, The University of New South Wales, Sydney, Australia

ABSTRACT: A model of OHS management was developed using the safe place, safe person and safe systems
framework. This model concentrates on OHS being a collective responsibility, and incorporates three different
perspectivesan operational level (safe place), an individual level (safe person) and a managerial level (safe
systems). This paper describes the qualitative methodology used in the development of the assessment tool,
including the lessons learnt from the pilot study and preliminary results. This research also promotes the use of a
new style of reporting that identifies areas of strengths as well as vulnerabilities, and uses discreet, non-emotive,
neutral language to encourage an objective, constructive approach to the way forward. The preliminary results
from the pilot study and peer review using the Nominal Group Technique were very encouraging suggesting
that this technique would be useful in directing a targeted approach to systematic OHS management, and that
the safe place, safe person, safe system framework was suitable to be taken to the next stage of wider case study
application.

INTRODUCTION

Hazards in the workplace may arise in many areas,

for example from the infrastructure, hardware and
operating environment; from issues related to the
individual workers skills and strengths and the complexity of human behaviour; or even from the management strategies and methodologies used to direct the
production of goods or services.
A review of the literature suggests that three
main approaches that have emerged to deal with
the corresponding areas where hazards may derive
namelysafe place, safe person and safe systems
(Makin and Winder 2006).
Safe place strategies focus on the existing physical
workplace environment and may include the use of safe
plant and equipment; preventive maintenance, monitoring and inspection; site security; and contingency
plans for emergency situations and ongoing hazard
reviews.
Safe place strategies often involve extensive capital, as it is more difficult to retrofit safety if it has not
been included in the design phase. Changes necessary
to the existing environment are often identified using
the risk assessment and control process. Solutions may
be more reliable when the hazards are predictable and
there is widespread knowledge available. However,
when this is not the case and the hazards are unique
to the particular scenario, or the risks are unforeseen,

it may be necessary to dedicate considerable time and

resources to the development of a customised solution.
Safe person strategies can include the incorporation
of safety into position statements; selection criteria to
protect vulnerable workers; inductions for employees,
visitors and contractors; training and skills development; methods for obtaining feedback; the inclusion
of safe working attitudes into performance appraisals;
and the review of reasons for personnel turnover. The
use of personal protective equipment and provisions
for first aid, injury management, workers compensation and rehabilitation are also included here as
contingency measures for when prevention and control measures have failed. These strategies focus on
the individual worker.
Safe person strategies may be particularly useful
when the workplace is not fixed, for example in call
out work where the physical environment is unknown
and there is heavy reliance on skills and expertise to
manage the existing situation. These strategies may
be limited by the complexities of human behaviour,
such as fatigue and times of job or life stress. These
must be appreciated at the time of application and
allowances or additional controls used to counteract
these parameters.
Safe system strategies are frequently found in templates for occupational health and safety management
systems (OHS MS) and deal with the concept or design
phase, as well as providing guidelines on gathering

739

data to understand the work processes and consider the

bigger picture in review activities. Included here are
provisions for safe design, safe supply and procurement; competent supervision; good communication;
use of consultation; incident management; and means
of self checking via specialist audits or system reviews.
The focus here is on management and systems to
promote a safe working environment.
In order to determine the usefulness of the above
approach, an assessment tool was developed so that
the safe place, safe person, safe system framework
could be applied to determine the hazard profile of an
organisation. The assessment tool comprised of supporting material for each framework element to allow
a risk ranking exercise to be conducted. This included
a definition and scope for each element and risk outcomes if the element was overlooked. The individual
assessments for each element were aided by the development of a series of prompts that considered possible
risk factors and possible prevention and control strategies for each of the elements (see Figure 1 for elements
for Electrical and Stress Awareness).
The risk ranking exercise was conducted as a two
stage process: firstly without taking into account interventions that were already in place so that areas of
vulnerability could be identified should any of the

current control measures fail (the raw hazard profile); secondly, an assessment was to be made on the
risk remaining once existing prevention and control
strategies had been applied (the residual risk profile).
This was to give an indication of the amount of risk
reduction that had been achieved and to help identify
opportunities for improvement. This was performed
by using a risk ranking matrix factoring in a combination of both severity and likelihood and a resulting
allocation of either high, medium-high, medium or
low. It should be noted that the use of non-emotive language was deliberately selected for providing feedback
about hazard profiles as this was considered an important step in breaking down barriers to the improvement
process and avoiding blame. For example words such
as catastrophic or extreme were not used in risk
ranking labels. Where elements were handled with
expertise this was recognised and fed back to organisation by giving it a risk ranking of zeroor well done.
Also, an assessment was made on the level of formality applied to the systems invoked and whether
or not all the elements proposed by the safe place,
safe person, safe system framework had in fact been
addressed by the organisation. The level of formality
was also assessed to recognize where informal systems
were used to manage risks, but did not contain a high

Electrical
All electrical equipment should be handled appropriately by those who are suitably qualified and kept in good working order. Other electrical hazards include
electric shock; static electricity; stored electrical energy, the increased dangers of high voltage equipment and the potential for sparks in flammable/explosive
atmospheres. Where live testing is necessary, only appropriately trained and qualified personnel should do so in compliance with relevant legislation and codes.
The risk is that someone may be injured or fatally electrocuted or cause a fire/explosion by creating sparks in a flammable or explosive atmosphere.
Possible Risk Factors

Possible Prevention and Control Strategies

Electrical Equipment/Goods in Use

Worn Cables

High Voltage Equipment In Use

Messy Power Leads
Use of Unqualified Personnel for Electrical Work
Breakdowns out of Normal Business Hours

Overloaded Power Points

Static Electricity Generated
Live Testing Necessary
Difficulty in Isolating Circuits

Isolation Procedures
Use of Authorised/ Qualified Repairers
Lead/Cable Checks
Circuit Breakers
Use of Residual Current Detectors
High Voltage Procedures
Static Electricity Use of Earthing Devices or Non-Conducting Materials
Lightning Protection

Stress Awareness
Personal skills, personality, family arrangements, coverage of critical absences, resourcing levels and opportunities for employees to have some control over
work load are factored into ongoing work arrangements so as not to induce conditions that may be considered by that particular employee as stressful. Plans are
available for dealing with excessive emails and unwelcome contacts.
The risk is that the employee becomes overwhelmed by the particular work arrangements, and is unable to perform competently or safely due to the particular
circumstances.
Possible Risk Factors
Shift Work
High Level of Responsibility
Frequent Critical Deadlines
No Control Over Work Load
Lack of Necessary Skills

Figure 1.

Possible Prevention and Control Strategies

Night Shifts
Long Shifts
Young Families
No Consultation Prior to Changes
Covering of Absences/ Unforseen Staff Shortages

Training in Necessary Skills Prior to Promotion/ Transfer/ Increased Responsibilities

Consultation Before Changes Made
Plans to Cover Critical Absences
Authority Commensurate with Responsibility
Access to Necessary Resources
Consideration of Individual Family Circumstances
Adequate Staffing Levels
Authority Commensurate with Responsibility
Negotiated Workloads

Examples of supporting material and prompts for elements in the revised assessment tool.

740

level of documentation to record actions. This was

to explore whether there was a connection between
the level of risk reduction achieved and the use of
formal systems, as well as to highlight areas for further
growth.
A pilot study was conducted with a preliminary assessment tool to trial the technique and case
study protocol. The methodology for this qualitative
approach was conducted according to Yin (1989).
A qualitative approach was selected because previous
research investigating means of assessing the effectiveness of OHS MS had shown that one size did
not fit all and there was a need for a more customised approach (Gallagher, 1997). Hence, there
would be difficulty in evaluating the results of the
same tool being applied to different organisations
when each organisation has unique needs and an
unpredictable hazard profile. The qualitative methodology would allow the variables to be better understood and addressed before undertaking a larger scale
quantitative study.
2

METHOD

The initial development of the assessment tool was

based on an extensive review of the literature. Methods to provide internal construct validity such as the
Delphi Method and the Nominal Group Method were
considered as most appropriate for this particular
research as there were no studies known to the authors
where a statistically significant quantitative approach
had been used successfully, and so there were no studies available for comparison of results. Hence the
above mentioned methods were both indicated for the
development of new knowledge where comparison
with others studies was not available and success in
both cases hinged on a balanced and vigorous peer
review process.
The Delphi Method involves the use of an expert
panel, with each member receiving the document for
review. The review is performed independently and
returned. This is to avoid the potential for domination or influence from other members. Multiple rounds
of the review then take place until convergence is
reached. The issue of convergence forms the critical point where the application of the Delphi Method
is challenged. Without skillful determination of the
end point, there is the possibility of creating a very
long, drawn out process (Landetta, 2005; Linstone
and Turoff, 1975).
The Nominal Group Technique is similar to the
Delphi Method, although instead of the information
for review being sent out with a questionnaire, the
panel members are collected together and changes or
improvements are brainstormed in workshop setting
(Delbeqc, 1975b).

Ultimately, the Nominal Group Method was

selected over the Delphi Method for a number of
reasons:
the extensive nature of the literature review which
formed the primary source of input into the process;
the potential for the Delphi Method to become
overly extended if difficulty is encountered reaching
a consensus, or if the return of the reviewed material
is delayed; and
the synergistic effect when working together in a
workshop setting whilst using the Nominal Group
Technique was thought not only to streamline the
process, but also enrich the final outcome.
The Nominal Group Technique is not without limitations, namely the potential for domination of certain
members, and the possibility of groupthink (Petersen,
1975). In order to counteract this potential bias,
the guidelines offered by Delbeqc, Van de Ven and
Gustafson (1975a; b) for conflict resolution and the
constructive use of diverging opinions were observed.
After the Nominal Group Technique was selected
for internal validation by a panel of experts from
academia and industry, a letter of invitation was sent
out. The final panel members included representation by a manual handling expert; a psychologist; an
occupational hygienist; a dangerous goods expert; a
human factors analyst; three members with chemical
engineering experience; an occupational toxicologist;
and industry representatives with experience in manufacturing and design. Three academics were involved
including the chairman of the Standards committee
for the development of AS/NZS 4804/4801: 2001
Occupational Health and Safety Management Systems
(Standards Australia. 2001a, 2001b). All but two of the
invited members were able to attend on the day, but all
provided input to the process.
The actual Nominal Group Technique session was
conducted by assembling the members of the panel
together after the first stage of the pilot study was completed and the preliminary report on the pilot study
produced to demonstrate the application of the assessment tool and findings of the risk ranking exercises
(see Table 1, and Figures 24).
The members of the panel were each given a copy
of the assessment tool and the preliminary report one
month prior to the review date so there was ample
time to read through the information. The Nominal
Group Technique was carried out in two stagesone
for brainstorming of ideas for improvements and the
second for voting on the top five ideas from the
improvements suggested. The brainstorming stages
were split into four sessions so that equal time was
allocated for each of the three sectionssafe place,
safe person and safe systems; as well as time to consider improvements to the format of the final report.

741

Table 1.

Framework model used for OHS MS for pilot study.

Safe place

Safe person

Safe systems

Baseline Risk Assessment

OHS Policy

Ergonomic Assessments
Access/ Egress

Equal Opportunity/
Anti-Harassment
Training Needs Analysis
Inductions -Contractors/ Visitors

Plant/ Equipment

Skill acquisition

Storage/ Handling/ Disposal

Work Organisation

Amenities/ Environment

Accommodating Diversity

Electrical
Noise
Hazardous Substances
Biohazards

Job Descriptions
Training
Behaviour Modification
Health Promotion
Networking, Mentoring,
Further Education
Conflict Resolution
Employee Assistance Programs

Radiation
Installations/ Demolition
Preventive Maintenance
ModificationsPeer
Review/Commissioning
SecuritySite /Personal
Emergency Preparedness
Housekeeping
Plant Inspections/ Monitoring
Risk Review

First Aid/ Reporting

Rehabilitation
Health Surveillance
Performance Appraisals
Feedback Programs
Review of Personnel Turnover

Consultation
Legislative Updates
Procedural Updates
Record Keeping/ Archives
Customer ServiceRecall/
Hotlines
Incident Management
Self Assessment Tool
Audits
System Review

Safe Systems
30%

Safe Systems Safe Place

39%
32%

Safe Person
29%

Figure 2.

Goal setting
Accountability
Due Diligence
Review/Gap Analysis
Resource Allocation/
Administration
Procurement with
OHS Criteria
Supply with OHS
consideration
Competent Supervision
Safe Working Procedures
Communication

Safe Place
37.3%

Safe Person
33.3%

Examples of hazard distributions without interventions (left) and with interventions (right).

The top five ideas were voted upon using a weighting system -five for the most important idea down to
one for the least important of the five ideas selected.
The votes were made in confidence and collected for
counting. The purpose of having all of the members
together was to provide synergy and the opportunity
for explanation by the authors of the assessment tool as
well as to share any lessons learnt from the pilot study.

The external validation of the methodology was

provided by the triangulation of data during the assessment stageseeking to find convergence of observations; questionnaire and interview responses; and
objective documentary evidence. The purpose of this
external validity was not for statistical inference or
to establish causation, but to ensure the coherence of
theoretical reasoning for the development of emergent

742

80
70

Risk Ranking

60
50
40
30
20
10
0

Safe Place

Safe Person

Safe Systems

Figure 3. An example of the risk reduction graph in the revised report (dark columnswithout interventions; light
columnswith interventions).

InductionsContractors/Visitors

All visitors and contractors to the workplace are made aware of any hazards that they are likely to encounter
and understand how to take the necessary precautions to avoid any adverse effects. Information regarding
the times of their presence at the workplace is recorded to allow accounting for all persons should an
emergency situation arise. Entry on site is subject to acceptance of site safety rules where this is applicable.
The risk is that people unfamiliar with the site may be injured because they were unaware of potential hazards.
Procedures for contractors and visitors are working well.

Incident Management

A system is in place to capture information regarding incidents that have occurred to avoid similar incidents from recurring
in the future. Attempts are made to address underlying causes, whilst also putting in place actions to enable a quick
recovery from the situation. Root causes are pursued to the point where they are within the organisations control or
influence. Reporting of incidents is encouraged with a view to improve rather than to blame. Near miss/ hits are also
reported and decisions to investigate based on likelihood or potential for more serious consequences. Investigations are
carried out by persons with the appropriate range of knowledge and skills.
The risk is that information that could prevent incidents from recurring is lost and employees and others at the workplace
continue to contract illnesses or be injured.

This isused more as a database for reporting rather than as a problem solving tool. Selective application of
root cause analysis, corrective action and evaluation may yield significant improvements in this area.

Figure 4.

Excerpts from the preliminary report illustrating the style of reporting used (above).

themes and paradigm shifts. In this sense, the validation process for qualitative research does not try to
achieve the same goals as quantitative research; the
aim is instead to provide multiple perspectives, and
in doing so overcome the potential for bias in each
individual method. A collage is then formed to give
depth and breadth to the understanding of complex
issues (Yin, 1989).

RESULTS

After the peer review process a number of changes

were made to the final framework which represented
the building blocks of systematic occupational health
and safety management, and there was a minor modification to the format of the final report. A letter
was sent out to each of the panel members explaining
the changes made and a copy of the final assessment
tool. Whilst all the votes were carefully considered,
the final decision on the layout was made by the

authors. The final framework comprised of twenty elements for each aspect of the model, making a sixty
element matrix, three more than the original model.
A Risk Reduction graph was added to the final report
to increase clarity and assist in the interpretation of the
final results (see Figure 3).
Three new elements were added: Receipt/Despatch
to cover OHS issues associated with transportation of
materials to and from the workplace; Personal Protection Equipment to address issues related to the safe use
of PPE; and Contractor Management to ensure that all
the lines of responsibility are well understood and that
all the information necessary has been exchanged.
Details of changes within elements in the framework model were:
Training Needs Analysis was incorporated into
Training.
Work OrganisationFatigue and Stress Awareness
was modified to remove stress awareness, which
became its own element.
Noise included more information on vibration.

743

 Access/Egress included a reference to disabled

access/egress.
Risk Review was renamed Operational Risk
Review.
Plant Inspections/Monitoring was renamed Inspections/Monitoring and the explanatory information
modified to reflect the intention that this was not
the walk around inspections associated with housekeeping but more to do with understanding the
process.
Storage/Handling/Disposal had Storage/Handling
removed so the element just refers to Disposal.
Ergonomic Assessment was renamed Ergonomic
Evaluation and had any references to Occupational
Hygiene removed from the explanatory material.
Not many changes were necessary to the format of
the final report, although it was found that the two pie
charts in Figure 2 were difficult to interpret, so the bar
graph in Figure 3 was added to illustrate the relative
risk reduction that taken place.
A number of other changes were incorporated into
the format of the revised OHS assessment tool after
the pilot study was conducted. The most significant
of these was to merge the two assessments (one without interventions and the other with interventions in
place) into the one section. This was to enhance ease
of use and reduce the time taken to conduct the actual
assessments. Also, the tool was colour coded to assist
navigation of the information whilst on site.
Once the changes to the revised OHS assessment
tool were finalised, advertisements were placed in an
Australasian local OHS journal, an OHS email alert as
well as on the home page for UNSW School of Safety
Sciences website, to attract participants into the study.
A brochure was also produced to give details of the
study to those interested in a hard copy and electronic
format. This ensured that all participants received the
same information. A total of eight organisations were
identified for the next phase of project with the revised
assessment tool and improved reporting format.
The size of the organisation to be used for the
case studies was not specified in the advertisements
as the authors were interested to find out what sized
organisations would respond. The pilot study was conducted successfully on a medium sized manufacturer
in the plastics industry. Of the eight organisations that
responded and were eventually selected with the case
studies, one was a small retail business; two were small
to medium sized family businesses and the remainder
were larger organisations. Where larger organisations
were involved, it was considered that as the OHS
assessment tool was originally intended for use in
small to medium sized enterprises the OHS assessment
tool, it could be suitable if the scope of the assessment was limited to a smaller, defined division of the
organisation.

The pilot study was then taken to a second stage that

was completed after the Nominal Group Technique
review had been performed to investigate whether the
risk ranking exercise could be used to make targeted
improvements in the workplace. This was conducted
by asking the organisation to select three elements that
it would like to improve, then choosing three questions
about each element (making a total of nine) that would
be asked each month for a period of four months. The
questions were to be phrased so that they would generate a simple yes or no answer, and one person was asked
to be responsible for the actions nominated. The purpose was to target visible outcomes that would improve
the final risk rating for the particular elements selected.
To improve the objectivity of the controlled self assessment exercise, these answers were to be independently
spot checked each month. Only three elements were
targeted in recognition of the time it takes to implement
changes to a safety program and the desire to keep the
task manageable and hopefully to obtain some early
wins that might encourage management to continue
with the improvement program.
The organisation in which the pilot assessment was
conducted was also asked to identify some means
of measuring whether or not OHS performance was
improving during the implementation of the study.
Guidance material was provided on the strengths and
limitations of various commonly encountered measurement indicators in the project field kit supplied
at the onset of the pilot study (Makin and Winder,
2007). The pilot organisation opted to continue measuring injuries and incidents and the three elements
targeted were: work organisation; access/egress and
incident management. At the end of the four months
the organisation was asked to complete a short evaluation survey. The outcomes of this monitoring process
are shown in Figure 5.
4

DISCUSSION

The use of the Nominal Group Technique was found

to be of great benefit to the development of the OHS
assessment tool by offering an opportunity for vigorous peer review by a group of experts of varying
backgrounds. Not only was this necessary to support
the internal validity of the assessment tool developed,
but it was also found to greatly enrich the final version
of the tool for later use in the multiple case study exercise. Further, each expert was able to bring in detailed
knowledge of their interpretation of the OHS elements,
and the final product could not be said to be a reflection
of any one viewpoint.
The difficulties encountered with the application
of the Nominal Group Technique were mainly related
to the logistical problems of availability of panel
members. As a result two of the members were unable

744

10

LTI's
Medical treatments
First Aid treatments
Reports

Numbers of Reports

0
Jun

Figure 5.

Jul

Aug

Sep

Oct
Month

Nov

Dec

Jan

Injury and incident results after completion of phase 2 of the pilot study.

to attend the workshop on the day, but were able to offer

their comments for review at a later date. Also this task
was performed as an honorary role, so it was necessary
to ensure that the half day was well organised. To assist
this process, agendas were sent out in advance and the
program steered and facilitated by an academic who
was not involved in the actual development of the tool
itself to maintain objectivity. The workshop was able
to adhere to timelines suggested and the process was
considered to run very smoothly. Although there were
clearly differing views from some of the panel members this was not unexpected as each brought their
own perspective and experience and sharing this was
in itself a worthwhile exercise. Where differing opinions remained unresolved, the members were directed
to express their views at the confidential voting stage,
and to cast their votes on the balance of information
available. Once the votes were tallied and the feedback worked into the final version of the assessment
tool, the panel members were given another opportunity to express any concern with the final outcome by
feeding information back to the authors within a reasonable time period. No changes were requested and
most panel members expressed satisfaction with the
final outcome.
Clearly the success of the Nominal Group Technique would be heavily influenced by the range,
breadth and depth of the experience of the panel
members selected and this selection process was
considered to be the most important stage. In this
particular scenario, the process could have been
enhanced by the inclusion of an electrical engineer and
mechanical/structural engineer although a number of
very experienced chemical engineers were present.

Furthermore, the success of the workshop would

not have been possible without the ability to feedback
the experience of the pilot study. This was found to be
crucial in terms of assessing the proposed frameworks
initial viability and many lessons were learnt along
the way so the development of the preliminary assessment tool was a very dynamic process. During the pilot
study methods that did not appear to be workable or
were too cumbersome were quickly modified from the
feedback received at the time. For example, the two
stage assessments were taking too much time when
performed in isolation so they were combined in the
final assessment tool. Furthermore, the triangulation
of data involving observation, interviews and a review
of documentation was found to produce the most fruitful and reliable data when information was sought
from layers of the organisation. It was found that it
was very important to collect the differing perceptions from management, operations personnel and the
individual workers and that these were all slightly different. Managements opinions were found to be more
optimistic, whilst sometimes the individual workers
were more skeptical (and perhaps more realistic) and
operations tended to be somewhere in-between. Where
there was gross misalignment of opinions, these areas
were studied in more depth until a clearer picture of
the situation emerged. Sometimes this would involve
some external research to verify the facts, for example
where this involved the use of hazardous substances
to ensure that the correct requirements were in fact
being used.
The proposed framework itself was found to be very
useful for the OHS management assessment and these
findings are discussed in more detail in a previous

745

article (Makin and Winder, 2008). During the pilot

study the broader context offered by the safe place, safe
person and safe systems model was able to highlight
areas of vulnerability that had perhaps been disguised
by a focus on production and hazards to do with the
physical plant environment such as noise and manual handling. Prior to the study there were significant
issues related to work organisation and the use of inappropriately long shifts that were unresolved. The pilot
study was able to highlight how this was potentially
exposing workers to unnecessarily high levels of risk,
and the longer shift hours were accentuating the problems with noise, fatigue, manual handling as well as
solvent exposure.
The second stage of the pilot study involving a
monthly questionnaire was found to be more difficult
to implement and depended on having someone within
the organisation who was highly motivated to see the
process through. Fortunately, the preliminary results
of the OHS assessment on the pilot study were considered to be very worthwhile by the organisation and
this generated enough enthusiasm to proceed to the
second stage. However, the organisation was undergoing a period of flux and management turnover, so this
second stage was delayed until the situation settled.
Although the OHS assessment was conducted with
the preliminary tool in March and April in 2007, the
second stage wasnt fully underway until the following October even though the agreed follow up actions
were decided in May. Despite these delays, a clear
improvement in OHS performance was observed (see
Figure 4), and it appeared that some time lag was
involved until the effects of the program came into
fruitionsuch as discussing the outcomes of two incident investigations per month at regular, but informal,
operational meetings. Whilst no statistical correlation
was attempted due to the qualitative nature of the study,
a further explanation of the improved trend was the
increased focus and attention on safety and health promoted by the study and the use of a targeted approach
that was realistic. The follow up actions had been set by
the organisation themselves and it was important that
they were in full control of the process. The preliminary assessment was also able to feedback positive
information in areas they had excelled, for example
in inductions for visitors and contractors, and this
was well received and helpful for encouraging their
co-operation with the study.
Finally the reporting style utilised was very well
received and the report was able to be widely distributed. The pictorial representation of key information and colour coding was found to be useful
in the quick dissemination of main points and was
considered to facilitate the interpretation of material to a wider audience from individual workers in a
safety committee setting, to operations personnel and
management.

CONCLUSION

The use of a pilot study and the Nominal Group Technique to trial the application of the safe place, safe
person, safe system model through the development
of an assessment tool was found to be very rewarding and worthwhile, and essential to the integrity
of the research being undertaken. The results have
enabled this research to be taken to the next level
multiple case studies which are currently in progress
and near completion. This qualitative approach is
highly recommended for this particular field of
research and preliminary results from the case studies
suggest that there is much scope for future development and further work, in particular for customising
the current OHS assessment tool for specific industry
fields. Furthermore, the application of this tool was
not limited to small to medium enterprises as originally thought, and may provide a useful benchmarking
exercise across larger organisations where they are
comprised of smaller subsidiary groups.

ACKNOWLEDGEMENTS
The authors gratefully acknowledge Dr. Carlo
Caponecchia, who was facilitator of the nominal group
session, and all contributors to the session.

REFERENCES
Delbecq, A.L., Van de Ven, A., Gustafson, D. H. (1975a)
Profile of Small Group Decision Making. In: Group Techniques for Program Planning. Glenview Illinois: Scott,
Foreman and Company. pp. 1539.
Delbeqc, A., Van de Ven, A., Gustafson, D. H. (1975b) Group
Decision Making in Modern Organisations. In: Group
Techniques for Program Planning. Glenview, Illinois:
Scott, Foresman and Company. pp. 113.
Landeta, J. (2005) Current validity of the Delphi method
in social sciences. Technological Forecasting and Social
Change In press, corrected proof.
Linstone, H.A., Turoff, M. (1975) Introduction. In: The
Delphi Method: Techniques and Applications. Reading,
Massachusetts: Addison-Wesley Publishing Company.
pp. 110.
Gallagher, C. (1997) Health and Safety Management Systems: An Analysis of System Effectiveness. A Report to
the National Occupational Health and Safety Commission: National Key Centre in Industrial Relations.
Makin, A.-M., Winder, C. (2006) A new conceptual framework to improve the application of occupational health
and safety management systems. In: Proceedings of
the European Safety and Reliability Conference 2006
(ESREL 2006), Estoril, Portugal, Taylor and Francis
Group, London.
Makin, A.-M., Winder, C. (2007) Measuring and evaluating safety performance. In: Proceedings of the European

746

Safety and Reliability Conference 2007 (ESREL 2007),

Stavanger, Norway, Taylor and Francis Group, London.
Makin, A.-M., Winder, C. (2008) A new conceptual framework to improve the application of occupational health
and safety management systems. Safety Science. In press,
corrected proof doi:10.1016/j.ssci.2007.11.011
Petersen, D. (1975) Coping with the Group. In: Safety Management: a Human Approach. Deer Park, New York:
Aloray. pp. 205215.
Standards Australia. (2001a). AS/NZS 4801:2001 Occupational Health and Safety Management Systems
Specification with Guidance for Use. Sydney: Standards
Australia International Ltd.

Standards Australia. (2001b). AS/NZS 4804:2001 Occupational Health and Safety Management Systems
General Guidelines on Principles, Systems and
Supporting Techniques. Sydney: Standards Australia
International Ltd.
Yin, R. (1989) Case Study Research: Design and Methods.
Newbury Park, US: Sage Publications.

747

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Exploring knowledge translation in occupational health using the mental

models approach: A case study of machine shops
A.-M. Nicol & A.C. Hurrell
University of British Columbia Centre for Health and Environment Research, Vancouver,
British Columbia, Canada

ABSTRACT: The field of knowledge translation and exchange is growing, particularly in the area of health services. Programs that advance bench-to-bedside approaches have found success in leveraging new research into
a number of medical fields through knowledge translation strategies. However, knowledge translation remains
an understudied area in the realm of occupational health, a factor that is interesting because workplace health
research is often directly applicable to risk reduction activities. This research project investigated knowledge
translation in one occupational setting, small machine shops, where workers are exposed to Metal Working
Fluids (MWF) which are well established dermal and respiratory irritants. Using the mental models approach,
influence diagrams were developed for both scientists and were compared with qualitative interview data from
workers. Initial results indicated that the sphere of influence diagrams would benefit from the inclusion of other
stakeholders, namely policy makers and product representatives. Overall, findings from this research suggest
that there is only minimal transfer of scientific knowledge regarding the health effects of metal working to
those at the machine shop level. A majority of workers did not perceive metal working fluids to be hazardous
to their health. Of note was the finding that MWF product representatives were rated highly as key sources
of risk information. The translation of scientific knowledge to this occupational setting was poor, which may
be due to varying perceptions and prioritizations of risk between stakeholders, lack of avenues through which
communication could occur, an absence of accessible risk information and the small size of the workplaces. The
mental models approach proved successful for eliciting information in this occupational context.

INTRODUCTION

Work is a central feature of life for most adults, providing our livelihood and sense of place, and occupying
at least a third of our waking hours. Not surprisingly,
work can have a profound impact on health. There is
growing recognition that many diseases (e.g., asthma
and chronic obstructive lung disease, joint and tendon disorders, stress-related mental health problems,
and some cancers and communicable diseases) can be
caused or augmented by workplace exposures.
Because of its vital role in adult life, the workplace provides a valuable opportunity for promoting
occupational health. A key feature of occupational
health research is that it is often possible to translate research results into direct prevention activities
in the workplace. Thus, the communication or transfer of occupational health risk information has the
potential to have an immediate and profound effect
on work-related morbidity and mortality.
The process of communicating risk information to
workplaces involves workers, managers, engineers,

researchers, health experts, decision-makers, regulatory bodies and governments. Although most occupational health researchers are aware of the need to
communicate with front-line workers and decisionmakers, there has been little research on knowledge
translation or risk communication processes within
occupational settings. To date, understanding the factors affecting the application of knowledge in this field
remains extremely limited.
The primary objective of this research was to investigate how health information is translated in one
occupational setting, machine shops, where workers
are exposed to complex mixtures of chemical and
biological hazards in the form of metal working fluids (MWFs). These exposures have been linked to
occupational asthma, increased non-allergic airway
responsiveness to irritants, chronic airflow obstruction
and bronchitis symptoms (Cox et al., 2003) and have
been the focus of considerable recent regulatory attention in Canada and the US. This project is linked to an
ongoing study of risk factors for lung disease among
tradespeople in British Columbia, Canada, which has

749

found machinists to be at higher levels of risk for

lung problems than three other trade groups (Kennedy,
Chan-Yeung, Teschke, & Karlen, 1999). Using interview data from experts, workers and managers, this
project aimed to identify knowledge gaps and misperceptions about metal working fluid exposure, to
determine socio-cultural and organization factors that
influence how knowledge is transferred in an occupational context and to determine preferred channels or
pathways for health risk communication.

2
2.1

METHODS

were done over the phone. Each interview took approximately 20-35 minutes. The respondents were asked
open-ended questions that were created using guidance from the experts Mental Model (see 2.2). Workers were queried about their work history and habits,
as well as their knowledge of MWF exposure and
the personal protection strategies they undertook in
the workplace. They were also asked questions about
health effects associated with MWFs, including where
they would look for information on health effects, and
what steps they would take to mitigate these effects.
These open-ended questions were often followed up
by probes designed to elicit more information on a
particular subject matter.

Data collection

Data was collected for this project using the Mental Models methodology developed by Morgan et al
at Carnegie Mellon University (Morgan, 2002). This
method has been previously applied in an occupational
context (Cox et al., 2003; Niewohner, Cox, Gerrard, &
Pidgeon, 2004b). The data was collected in two phases,
beginning with interviews with scientific experts in
the field of MWF exposure and effects, followed by
interviews with workers employed in machine shops.
2.1.1 Expert interviews
A member of the study team who is an academic expert
on the health effects of MWF compiled a list of experts
on MWF and health effects. The list was comprised
primarily of academic researchers from the US and
Europe, but also included US government researchers
and occupational health professionals. Of this list of
experts, the study team was able to contact 16, and 10
of these consented to participate.
The interviews, which were carried out by a single trained research assistant, were conducted over the
phone and lasted for an average of 30 minutes. The
first two interviews were used to pilot test the survey
instrument and were therefore not included in the final
analysis. The respondents were asked open-ended
questions about exposures, health effects, and mitigation strategies relating to MWF in the workplace.
They were also asked about their attitudes and practices relating to the communication of their research
results to decision-makers in industry and regulatory
agencies.
2.1.2 Worker interviews
To recruit machinists, introductory letters were sent
to 130 machine shops in the province of British
Columbia, Canada, and were followed up with at least
one phone call. Twenty-nine workers from 15 different
machine shops agreed to participate in an interview.
The interviews were conducted by four different
trained interviewers. Twenty of the interviews were
done at a private location at the worksite, and nine

2.2

Mental model development

All interviews were audio-taped and transcribed and

entered into NVivo. To construct the expert model,
transcripts from the expert interviews were coded
with a coding schema developed through an iterative process of fitting the codes to the data, based
on grounded theory (Strauss & Corbin, 1998) and
informed by previous mental models work (Cox et al.,
2003; Niewohner, Cox, Gerrard, & Pidgeon, 2004a).
The coding schema was also informed by a literature review of existing chemical information regarding
MWFs, which aided in the development of the model
categories included in the expert mental model (i.e.
exposure processes, machine types, etc.). The initial coding schema was reviewed and validated by an
expert who was part of the research team.
The expert mental model covered five main
domains: shop health and safety factors, MWF
exposure factors, regulatory and economic factors,
health effects, and exposure modes. Within these five
broad categories, related topics, such as information sources, reactive behaviors, and physical safety
barriers emerged as sub-categories.
The transcripts from the worker interviews are
currently being analyzed using a similar grounded
theory-informed method. A worker mental model is
currently under construction.

2.3

Data analysis

For the health effects and information sources analysis, data for each of these constructs was abstracted
from NVivo and reviewed by two members of the
research team who had expertise in the areas of health
effects and risk communication. Data from the workers
were compared and contrasted with the expert model
and areas of both congruence and disconnect were
identified. Results were entered into tables to present
comparisons.

750

RESULTS

3.1

Demographics

Respondents from the 15 participating machine shops

were all male, and represented a range of ages and
levels of experience in the trade. The demographic
details of the machinist respondents can be found in
Table 1.
3.2

Knowledge and description of health effects

Differences were found between the experts and

workers descriptions of the health effects that can
be caused by MWF exposure in the workplace (see
Table 2). In particular, only 28% of the workers were
able to describe symptoms that could occur in the
lungs as a result of MWF exposure (such as cough,
asthma, bronchitis, difficulty breathing). The majority of the experts described respiratory issues in detail,
providing a description of symptoms and specific medical terminology of diseases associated with MWF
exposure such as such as hypersensitivy pneumonitis
(HP), occupational asthma and decreased lung function. Only two of the workers were able to describe
asthma as a potential condition from MWF exposure,
one mentioned decreased lung function, and none
mentioned HP.
While unable to provide any specific terms or symptoms, a further 45% percent of workers were able to

Table 1.

Machinist demographics.

Characteristic

2029
3039
4049
50+
Unknown

3
10
12
2
2

10%
34%
41%
7%
7%

5 to 10
11 to 15
16 to 20
21 plus
Unkown

7
7
6
7
2

24%
24%
21%
25%
7%

<10 people
1150 people
50 plus
Unknown

5
13
10
1

17%
45%
35%
3%

Manual
CNC
Both
Unkown
Other

4
5
15
4
1

14%
17%
52%
14%
3%

Age

identify the lungs as potential site of health problems. Of note, nine of the workers (31%) described
having personal experience with either a lung effect
from MWF exposure or feeling MWF mists in their
lungs.
There was greater concurrence between experts
and workers awareness of specific dermal conditions
that can occur as a result of MWF exposure, including
rash, dry hands, itchy hands, dermatitis and eczema.
Sixty-two percent of workers could identify a specific
dermal health effect such as eczema, although a further
31% were only able to identify the skin in general as
a potential site for health effects. Forty percent of the
workers said that they had experienced adverse effects
on their hands from MWF exposure.
Four of the experts (44%) discussed the association between cancer and MWF exposure, although
proportionally fewer (17%) of the workers described
MWFs as cancer-causing agents. Of the workers who
described cancer, there was a general tendency to mention smoking and its carcinogenic potential in the same
discussion.
There were health effects that workers described
that experts did not, particularly irritation that could
occur in eyes. Two workers also suggested that MWF
could affect blood.
Within the cohort of workers, 21% stated that
MWFs were not harmful to health, even though in
some cases these workers did note that MWF exposure could cause skin problems. Finally, there were
two people in the worker group who stated that they
were unaware of any potential health effects of MWF
exposure.
3.3 Sources of information
Experts and workers were asked slightly different
questions regarding sources of health and safety
Table 2.
workers.

Description of health effects, experts and

# of years in trade

Shop size

Health effects
Described specific health effects
that can occur in the lungs
Described specific health effects
that can occur on the skin
Described a relationship between
MWF exposure and cancer
Central nervous system depression
Eye Irritation
Problems with blood
Poisonous
Stated that MWFs do not
cause health effects

Types of machines

751

Workers Experts
(n = 29) (n = 10)
%
%
28%

70%

62%

70%

17%
3%
17%
7%
3%

40%
10%
0%
0%
0%

21%

0%

information in the workplace. Table 3 presents

responses to the open ended questions How do you
think that workers learn about new scientific advances
in MWFs? and How about safety issues around
MWFs?. While 40% of workers noted that occupational health and safety training was a source of
information, the same number of experts did not think
that workers received such information at all. Material
Safety Data Sheets (MSDS) were ranked fairly low as
information sources for workers amongst the scientific
experts.
Table 4 shows the responses to the following open
ended question that was posed to workers: If you
were going to learn about the health effects of metal
working fluid exposure, or maybe some of the hazards
that you are exposed to in your shop, where would you
go for this sort of information? Suppliers and manufacturers were the most referred to sources, followed
by MSDSs, which is in sharp contrast to the responses
of the experts. Other workers and the internet were
also major sources for workers not described by the
experts.
Workers were also asked what person, agency or
group would you trust the most for information on
MWFs, either about the product or possible associated health effects? The responses to this question,
shown in Table 5, indicate that most workers trust
WorkSafeBC, British Columbias workers compensation board. Various levels and departments within

Table 3. Expert answers to: How do workers learn about

new health and safety issues and advances? (n = 10).
Expert Answer

They dont
Occupational health and safety training
Trade media (magazines, pamphlets)
Union
General news media
MSDS
Govt agencies

40
40
30
30
10
10
10

Table 4. Worker answers to question regarding main

sources of information used.
Worker Answer (n = 29)

Suppliers and manufacturers

MSDSs
Managers or other workers
The internet
Health and safety committee
Government organizations
Container labels

86
69
66
48
41
34
28

Table 5.

Workers trusted sources.

Worker answer (n = 29)

WorkSafeBC
Government
MSDS
Manufacturer/supplier
Other workers
Union
Researchers
Dont know

31
17
14
14
3
3
3
3

Table 6.

Expert communication pathways.

Expert answer (n = 10)

Workplace management
Government health and safety agencies
Workers
Industry/Suppliers
Unions
Government (other than safety agency)
Physician

70
60
50
40
40
10
10

government were the next most trusted, followed by

MSDSs and manufacturers/suppliers.
Table 6 presents results of the questions asked to
experts on how they had attempted to communicate
the results of their MWF research. Most had provided information to either workplace management
or to a government health and safety agency. Half
reported that they had talked to workers directly about
MWF health effects and only one reported talking to
physicians.

DISCUSSION

4.1 Health effects

Good workplace risk communication requires that
workers receive information about the chemicals that
they use and that workers understand the potential
health effects that these chemicals can cause. As
Schulte et al. (Schulte et al., 2003) states effective transfer, receipt and utilization of occupational
health and safety information will only be realized to
the extent to which recipients actually can understand
the information transmitted (p. 522). The results of
this research suggest that while workers are aware that
they are being exposed to MWFs during the course of
their jobs, most have only a generalized understanding of how these compounds may affect the body. Such
results are not unique to this research and have been

752

found in other occupational health research such as

that conducted by Sadhra (Sadhra, Petts, McAlpine,
Pattison & MacRae, 2002).
Of concern were the findings that three quarters of
the workers queried were unable to provide any detail
about the effects that MWF might have on their respiratory tract. In addition, they did not link symptoms
such as cough, difficulties breathing, phlegm, asthma,
and bronchitis to MWF exposure. Researchers such
as Nowak et al (Nowak & Angerer, 2007) indicate
that one of the problems of identifying occupational
disease is that fact that the symptoms arent necessarily correlated directly in time with an exposure and as
such, may happen after a worker has left the workplace.
This mechanism, coupled with a lack of awareness on
the part of the workers about the types of symptoms
that MWFs can cause, makes the correct diagnosis
of occupational respiratory disease very challenging.
Gupta et al (Gupta & Rosenman, 2006) has suggested
that hypersensitivity pneumonitis (HP) rates in the US
are most likely under-reported due to factors such as
inadequate disease recognition. Without information
about workers occupational and environmental exposures, doctors may misdiagnose conditions like HP as
atypical pneumonia. The review by Santos et al (Santos et al., 2007) of diagnostic factors for occupational
asthma found that lack of awareness of the association between symptoms and workplace exposures was
one of the most significant factors contributing to
diagnostic delays.
The workers descriptions of dermal effects were
markedly different from those of respiratory problems,
with a majority of workers being able to describe distinct symptoms of MWF exposure such as rash, itchiness and dry hands. These results may be due to the
fact that many of the workers had actually experienced
these problems personally, or had known others who
had these types of skin problems. Indeed, occupational
skin diseases are the most commonly reported workplace related conditions (Lushniak, 2003). Research
by Sadhra et al (Sadhra et al., 2002) found that workers tended to talk more easily and automatically about
more common health problems rather than those that
were considered more serious. Indeed, many workers in this study noted that they had skin effects, yet
these werent necessarily considered serious, or even
real health effects, even though they were eligible
for compensation. For example, when asked about the
effects of short-term exposure, one worker replied:
As far as sick . . . I wouldnt call what I had being
sick. Its just, you know, you have a rash on your
hand and I did have time off because of that.

were at least aware of the correct areas of the body

that MWFs could affect. Of particular interest from
the worker data was the issue of eye irritation. This
effect was not noted by any of the experts, even though
MSDSs for MWF list eye irritation as a potential health
effect. In fact, a review of MSDS for metal working
fluids found that there was more information about
eye irritation on some sheets than potential respiratory
effects. A review of the published literature revealed
no research focused specifically on eye irritation and
MWF exposure.
4.2

Information sources

The flow or transfer of information between the

expert realm and the workplace realm appeared
to be hampered by a number of barriers in this study.
In particular, the responses from experts and workers on the topic of where to find information on
MWFs showed a significant disconnect between the
groups. None of the workers were ignorant of potential sources of information on the health effects of
MWFs, although 40% of experts believed that workers did not learn about health and safety information.
Workers also identified MSDSs as important information sources, while only 10% of experts believed
that workers learned from MSDSs (this finding is in
keeping with the earlier discussion of effects of MWF
on eyes). Suppliers and manufacturers were the most
commonly mentioned information source by workers,
with 86% of workers stating that they would go to suppliers and manufacturers for information. In contrast,
none of the experts mentioned suppliers and manufacturers. These results are consistent with a study
by Sadhra et al (2002), which found considerable
differences between the information sources mentioned by experts and by workers in the electroplating
industry.
These results suggest that many MWF experts
perceive knowledge translation processes as broken or non-existent, even though experts did report
efforts to communicate their research results to audiences beyond the academic/scientific community. The
majority of experts reported that they communicated
research results to workplace management; however,
most experts were disillusioned about their communication efforts and the potential of these processes to
be translated to those at risk. Experts expressed a variety of opinions as to why they felt that their efforts
to send risk messages to workers were ineffective.
A number of experts directed frustration at workplace
managers and regulatory bodies for seemingly not
heeding scientific advice:

There were relatively few workers who described

health effects that were erroneous, indicating that most

753

. . . the communication [with workplace management] was unsuccessful in that I didnt get any
feedback [ . . . ] on what happened next.

I think we presented [the regulatory body]

with what we thought were positive findings but,
I think, since then weve been a little disappointed that they havent really capitalized on the
research as much as they might have done.
I have to say that we have been very disappointed with the way that the [regulatory agency]
have failed to publish the reports that we did in a
lot of this work.
This perceived lack of responsiveness from
decision-makers who are actually in a position to
effect changes in the workplace was problematic for
experts. This frustrating situation may cause them to
abandon their efforts to communicate with decisionmakers. There is evidence that research dissemination
and uptake is hampered both by researchers imperfect
understanding of decision-making contexts, as well as
by the organizational and/or political pressures facing decision-makers such as regulatory bodies. Lomas
(Lomas, 1996) suggests that structures to support and
improve ongoing knowledge translation and exchange
between researchers and decision-makers are needed
to speed the process of research dissemination and
uptake. He suggests a cultural shift involving new
organizational models for both decision-makers and
researchers, as well as enhanced funding to support ongoing knowledge translation between both
groups.
Other experts, when discussing their communication activities, expressed a level of discomfort with, or
lack of knowledge of, appropriate lay communication
techniques.
I think [communication to workers] is something that scientists overall have to do a lot more
of. They have to interest a lay audience in what
they do, and its an area, I think, we all need to
do a lot more in.
Id like to know a way of getting [health and
safety] information over to people so they actually took heed of advice before they actually got
problems.
Expert: I figure the way that I am doing it, I
would admit, is not the best. I think a program to
directly give your results to the labourers would
be an ideal pathway to go. It is not something that
our department routinely doesif everexcept for
communities. Talking to the workersthats not
something I have ever done and Im not familiar
with that many people who have.
Interviewer: I see. Do you have an idea of how
you would go about developing such a program?

Expert: No. Being an academic, unless there

was funding, I wouldnt know.
These comments demonstrate experts recognition
of their own role in the risk communication process
and their awareness that different communication techniques are necessary to reach worker audiences. This
suggests a need for appropriate training, resources, and
incentives to participate in non-traditional knowledge
translation efforts. Some funding agencies, such as
Canadas Canadian Institutes for Health Research, are
now actively promoting such efforts by requiring academic proposals to have knowledge translation plans,
and by providing funding for research into effective
knowledge translation practices (Canadian Institutes
of Health Research, 2003).
4.2.1 Trust in Information Sources
The role of trust in mediating how risk messages are
perceived, attended to, and acted upon has been widely
acknowledged in risk perception and communication
research. Studies have found that distrust heightens lay
concerns and responses to risk messages, and leads to
questioning of the actions of risk regulators and authorities (Cvetkovich & Lfstedt, 1999). Lay perceptions
of trust and credibility in risk messengers are dependent on three factors: perceptions of knowledge and
expertise; perceptions of openness and honesty; and
perceptions of concern and care (Kasperson, Golding,
& Tuler, 1992). These factors were evident in workers
discussion of trusted information sources.
Worker responses to the question about what person, agency or group they would trust the most
for information on MWF reveal a further disconnect
between the most used sources of information and the
most trusted sources of information. Manufacturers
and suppliers were mentioned most often as a source
of information, yet the provincial workers compensation board was reported to be the most trusted source.
Indeed, many workers specifically noted that they
did not trust manufacturers and suppliers information
source even though they used it. The reason for this
distrust is apparent in the comment of one worker that
theyre just trying to sell you something. Since trust
is seen as a prerequisite to effective risk communication (Kasperson et al., 1992), this demonstration
of distrust in manufacturers is problematic. Although
workers may receive information on MWFs and their
potential impacts on health from these sources, they
may disregard recommendations or ignore precautionary steps due to a perception of dishonesty in the risk
messengers.
Workers identified the provincial workers compensation board as their most trusted source of information. Workers described the board as unbiased,
independent, and non-profit. Many workers

754

pointed out that it was in the boards best (financial)

interest to prevent workers from becoming sick, and
thus it was also in their best interest to provide accurate and balanced information. A number of workers
also mentioned that the board had resources to conduct
research and to make evidence-informed decisions.
Thus, the board fits the factors of expertise, honesty,
and concern put forward by Peters et al (Peters, Covello, & McCallum, 1997). These results suggest that
risk messages delivered by the workers compensation
board may be more likely to be trusted, and thus acted
upon. However, not one worker mentioned the Board
as a source of information that they would use to learn
about hazards associated with MWF. Thus, efforts
would need to be made in order to actively disseminate
information from this source to workers.
5

STRENGTHS AND LIMITATIONS

A strength of this mental models approach rests on the

ability to develop a detailed representations of occupational hazards from different perspectives. However,
these representations rest on interview data that cannot be assumed to be a direct reflection of participants conceptual understandings. Participants may
not mention implicit knowledge of hazards or workplace behaviour, or may edit their comments based on
what they consider appropriate for a conversation with
a researcher.
This study benefited from interviews with a comprehensive range of experts who have studied MWFs
from various disciplinary angles, including occupational hygiene, respiratory medicine, environmental
toxicology, etc. The research community in this area
is small, resulting in a small sample size drawn from
around the world.
In contrast, the sample of workers was relatively
large, but was drawn only from the province of British
Columbia, Canada. Thus, the workers conceptual representations of MWF regulation and use may differ
from experts due to geographic specificities in regulation, economics, and use. In addition, some of the
workers were also participants in an ongoing study of
respiratory health of tradespeople. Thus, these respondents might have been more aware of the respiratory
health effects of MWF (although this hypothesis is not
supported by the results of this study).
6

CONCLUSION

The results of this research have implications not

only for workers but also for the broader fields of
occupational health and safety, occupational medicine
and disease surveillance, and occupational knowledge
translation. Through this Mental Models process we
have determined that there is some overlap between

how workers and experts understand the effects of

MWFs, particularly in the area of dermal exposure, but
that much more attention needs to be paid to providing
workers with a more comprehensive understanding of
the effects of MWF on the respiratory tract.
This study has also illuminated a number of important disconnects between how workers do receive
information as opposed to how they would like to
receive information, an important distinction that may
be impeding the awareness and management of workplace risks. Additionally, this study uncovered a degree
of frustration on the part of experts in their attempts
to communicate their findings and a relatively bleak
view of the current workplace communication milieu
for the worker. Neither of these conditions, as they
stand, will enhance the communication and exchange
of MWF exposure data in the occupational context.
At the outset of this study, manufacturers and suppliers were not expected to play such a key role in
the dissemination of health and safety information on
MWFs. These unexpected findings have led to a third
phase of interviews with a selection of manufacturers and suppliers. The results of these interviews are
expected to shed additional light onto the role that this
group plays in the communication of health and safety
issues relating to MWF.
ACKNOWLEDGEMENTS
The authors would like to thank Dr. Susan Kennedy,
Emily Carpenter, Reid Chambers and Natasha
McCartney for their assistance with this paper. This
project was funded in part by the Canadian Institute
for Health Research.
REFERENCES
Canadian Institutes of Health Research,. (2003). Knowledge translation overview. Retrieved June 29, 2006, from
http://www.cihr-irsc.gc.ca/e/7518.html
Cox, P., Niewohmer, J., Pidgeon, N., Gerrard, S.,
Fischhoff, B. & Riley, D. (2003). The use of mental models in chemical risk protection: Developing a
generic workplace methodology. Risk Analysis : An Official Publication of the Society for Risk Analysis, 23(2),
311324.
Cvetkovich, G.T. & Lfstedt, R. (1999). In Cvetkovich G. T.,
Lfstedt R. (Eds.), Social trust and the management of
risk. London: Earthscan.
Gupta, A. & Rosenman, K. D. (2006). Hypersensitivity pneumonitis due to metal working fluids: Sporadic or under
reported? American Journal of Industrial Medicine, 49(6),
423433.
Kasperson, R.E., Golding, D. & Tuler, S. (1992). Social
distrust as a factor in siting hazardous facilities and
communicating risk. Journal of Social Issues, 48(4),
161187.

755

Kennedy, S.M., Chan-Yeung, M., Teschke, K. & Karlen, B.

(1999). Change in airway responsiveness among apprentices exposed to metalworking fluids. American Journal of Respiratory and Critical Care Medicine, 159(1),
8793.
Lomas, J. (1996). Improving research dissemination and
uptake in the health sector: Beyond the sound of one hand
clappingCentre for Health Economics and Policy Analisys. Department of Clinical Epidemiology and Biostatics.
McMaster University.
Lushniak, B.D. (2003). The importance of occupational skin
diseases in the united states. International Archives of
Occupational and Environmental Health, 76(5), 325330.
Morgan, M.G. (2002). Risk communication: A mental
models approachCambridge University Press.
Niewohner, J., Cox, P., Gerrard, S. & Pidgeon, N. (2004a).
Evaluating the efficacy of a mental models approach for
improving occupational chemical risk protection. Risk
Analysis : An Official Publication of the Society for Risk
Analysis, 24(2), 349361.
Niewohner, J., Cox, P., Gerrard, S. & Pidgeon, N. (2004b).
Evaluating the efficacy of a mental models approach for
improving occupational chemical risk protection. Risk
Analysis : An Official Publication of the Society for Risk
Analysis, 24(2), 349361.

Nowak, D. & Angerer, P. (2007). Work-related chronic

respiratory diseasescurrent diagnosis. [Arbeitsbedingte
chronische Atemwegserkrankungen. Bekommen Sie
am Wochenende besser Luft?] MMW Fortschritte Der
Medizin, 149(4950), 3740.
Peters, R.G., Covello, V.T., & McCallum, D.B. (1997). The
determinants of trust and credibility in environmental risk
communication: An empirical study. Risk Analysis, 17(1),
4354.
Sadhra, S., Petts, J., McAlpine, S., Pattison, H. & MacRae, S.
(2002). Workers understanding of chemical risks: Electroplating case study. Occupational and Environmental
Medicine, 59(10), 689695.
Santos, M.S., Jung, H., Peyrovi, J., Lou, W., Liss, G.M. &
Tarlo, S.M. (2007). Occupational asthma and workexacerbated asthma: Factors associated with time to
diagnostic steps. Chest, 131(6), 17681775.
Schulte, P.A., Okun, A., Stephenson, C.M., Colligan, M.,
Ahlers, H., Gjessing, C., et al. (2003). Information
dissemination and use: Critical components in occupational safety and health. American Journal of Industrial
Medicine, 44(5), 515531.
Strauss, A.L. & Corbin, J.M. (1998). Basics of qualitative research: Techniques and procedures for developing
grounded theorySage Publications.

756

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Mathematical modelling of risk factors concerning work-related

traffic accidents
C. Santamara, G. Rubio, B. Garca & E. Navarro
Instituto de Matemtica Multidisciplinar, Universidad Politcnica de Valencia, Valencia, Spain

ABSTRACT: Work-related traffic accidents suppose an important economic and public health problem. In this
context, it is important to improve our knowledge on the main factors that influence these accidents, in order to
design better preventive activities. A data base coming from an insuring company is analyzed, in which information regarding personal characteristics of the individual is contained, as well as labor situation characteristics.
Cox model is used in order to construct a predictive model of the risk of work- related traffic accident. From the
obtained model we study if personal or labor characteristics act like predicting factors of work-related traffic
accidents.

INTRODUCTION

Work-related traffic accidents suppose an important

economic and public health problem. For instance, it
is reported that it is the largest cause of occupational
fatality in the United Kingdom (Clarke et al., 2005). In
Boufous & Williamson (2006) it is argued that traffic
crashes are one of the leading causes of occupational
fatalities in various parts of the industrialized world.
Clarke et al., (2005) reviews some research carried out
over recent years that suggests that drivers who drive
for business purposes are at an above average risk of
accident involvement relative to the general driving
population.
There exists research on this matter that has led to
certain amount of suggestions in order to manage the
risk to drivers health and safety (DfT 2003), (Bomel
2004), (WRSTG 2001). In this context, it is important to improve our knowledge on the main factors
that influence these accidents, in order to design better
preventive activities.
Different studies give a great weight to the human
factor in the origin of the accidents, (Lawton & Parker
1998), (Lewin 1982). In Cellier et al., (1995) it is
argued that age and professional experience are related
to the frequency and gravity of the accidents, in the
sense that the youngest and the oldest, as well as those
of smaller professional experience have the highest
indices in both aspects. A low educative level is also
related to the most serious accidents, according to
Hjar et al., (2000) (these and other references are
reported in Lpez-Araujo & Osca (2007)).
Some factors can be connected to the professional
situation of the person, like type of company or work,

type of contract, schedule, and others. The aim of

this paper is to construct a predictive model based
in these types of variables, using a data base coming from an insuring company. Further research is
needed in order to extend the used data type, looking
for in other data bases and designing perhaps suitable
questionnaires.
The paper is organized as follows. In section 2 we
present the data. In Section 3 we state the predictive
model by means of proportional hazards Cox model.
In section 4 the model is represented by means of a
useful graphic tool. Finally we do a little discussion in
section 5.
2

DATA

The sample comprises 22249 individuals that suffered

a work accident between january 2006 and february 2008, who joined their companies starting from
1-1-2003. From these accidents, 2190 were traffic
accidents.
Variables considered for this study were age, sex,
professional situation (waged or autonomous, labeled
Situapro2cat), and several company features which we
describe in the following.
Pertemp: relationship between individuals company and the work centre connected with the accident. There are four possibilities: 1. the work centre
belongs to the company; 2. the work centre does not
belong to the company and the relationship is of contract/subcontract; 3. the work centre does not belong
to the company and the relationship is of company of
temporary work; 4. the work centre does not belong

757

Table 1.

Categorical variables.

Variable
Sex
Men
Women
Situapro2cat
Waged
Autonomous
Pertemp
1
2
3
4
Topap
No
Yes
Topni
No
Yes
Topspa
No
Yes
Topspm
No
Yes
Topspp
No
Yes
Toptd
No
Yes

Table 2.

(%)

16954
5295

76.2
23.8

22053
196

99.1
0.9

20905
760
364
220

94.0
3.4
1.6
1.0

21590
659

97.0
3.0

22034
215

99.0
1.0

3802
18447

17.1
82.9

20859
1390

93.8
6.2

19803
2446

89.0
11.0

21916
333

98.5
1.5

Topap: preventive organization personally assumed

by the employer (no/yes).
Topni: there is not any preventive organization
(no/yes).
Topspa: external prevention service (no/yes).
Topspm: common prevention service (no/yes).
Topspp: own prevention service (no/yes).
Toptd: appointed worker or workers (no/yes).
Age values are between 16 and 77 years old,
with mean 33.39 and standard deviation 10.79. The
preceding variables are summarized in Table 1.
3

Variable

exp()

se()

Age
Sex
Pertemp2
Pertemp3
Pertemp4
Topap

0.0537
0.4774
1.1896
1.3117
0.0707
0.4566

0.948
1.612
0.304
3.713
0.932
1.579

0.00252
0.04511
0.23726
0.17971
0.23686
0.10238

21.347
10.584
5.014
7.299
0.299
4.460

Variable

exp()

p-value

lower .95

upper .95

0.0e+00
0.0e+00
5.3e07
2.9e13
7.7e01
8.2e06

0.943
1.475
0.191
2.610
0.586
1.292

Age
Sex
Pertemp2
Pertemp3
Pertemp4
Topap

to the company and the relationship is different to the

previous ones.
The following six variables refer to the preventive
organization of the company.

A PREDICTIVE MODEL FOR

WORK-RELATED TRAFFIC ACCIDENTS

In order to study predictive factors for work related

traffic accidents, a possible approach could be the
timetoevent analysis. We record the follow up of

Cox model.

0.948
1.612
0.304
3.713
0.932
1.579

0.952
1.761
0.485
5.280
1.482
1.929

an individual from the time he/she joins the company,

to the time of traffic accident, which will be a censored
time if the event has not occurred yet when the study is
finished. Then we investigate the influence on the risk
of accident of individuals features, as well as features
related to the company.
This approach would let us to detect the workers
with higher risk of traffic accident at a given time t,
say one or two years after his/her entry in the company.
This would help to adopt suitable preventive measures. Moreover, predictive factors close related to the
company could suggest the advisability of structural
reforms.
A first step could be to use the Cox proportional
hazards model (Cox 1972), and then analyze if it fits
accurately the data. The focus is modelling the risk of
traffic accident at time t, what is obtained from the
hazard function h(t). It is well known that from the
Cox regression model the hazard for individual i at
time t is given by
hi (t) = exp(1 x1i + 2 x2i + + p xpi )h0 (t)

(1)

where h0 (t) is the baseline hazard function, and xki are

the explanatory variables of individual i.
Parameters estimates in the Cox regression model
are presented in Table 2.
We will show in the next section a useful way
of treating the information obtained from the model.
Nevertheless, now we can extract some conclusions
already. For example, keeping the rest of variables
equal, workers with Pertemp = 2 have a risk 0.304
times the risk of workers with Pertemp = 1. In other
words, the risk is minor if the work centre where
the accident took place does not belong to the company and the relationship is of contract/subcontract,

758

Figure 1.

Nomogram.

with respect to the situation of work centre belonging to the company. On the other hand, workers with
Pertemp = 3 have a risk 3.713 times the risk of workers with Pertemp = 1. That is to say, the risk increases
a lot if the work centre does not belong to the company and the relationship is of company of temporary
work. With respect to preventive organization, there is
in this way some conclusion: the risk if the preventive
organization is personally assumed by the employer is
1.579 times the risk if the preventive organization is
not personally assumed by the employer.
Of course, these and other conclusions that may be
extracted from the model are provisional, and further
analysis is necessary, by improving the model, or handling the information with another models, and also
looking for another data base.
4

NOMOGRAM

The model may be represented by means of nomograms. A nomogram is a graphic tool easily interpretable. So, it is an interesting tool to take advantage
of the model. Figure 1 depicts a nomogram for predicting probability of no occurrence of traffic work related
accident at one year and two years after the individual
joins the company. Typographical reasons force us to
rotate the figure, but the natural way of looking at it is
obvious.
To read nomogram draw vertical line from each tick
mark indicating predictor status to top axis (Points).

759

Calculate total points and find corresponding number

on axis (Total Points). Draw vertical line down to axes
to find the workers probabilities of remaining traffic
accident free for one and two years.
For example, a worker with Topap = 1, contributes
approximately 14 points. This is determined by comparing the location of the value 1 on the Topap axis
to the points scale above and drawing a vertical line
between the two axes. In a similar manner, the point
values for the rest of predictor variables are determined
and are summed to arrive at a total points value. For
example, Pertemp = 1 would give 35 points, Sex = 2
gives 14 points, Age = 35 gives about 71 points, what
produces 134 points for a worker with these predictors. This value is marked on the Total Points axis, and
drawing a vertical line down we obtain a probability
of about 0.77 of being free of traffic accident the first
year, and about 0.66 the second year.
We assess the accuracy of the nomogram (and of the
model) using the concordance index (Harrell 2001),
which is similar to an area under the receiver operating
characteristic curve, and applicable to time-until-event
data. On a 0.5 to 1.0 scale c provides the probability that, in a randomly selected pair of individuals in
which one of them suffers an accident before the other,
the individual who had the accident first had the worse
predicted outcome from the nomogram. c = 0.5 represents agreement by chance; c = 1.0 represents perfect
discrimination. A total of 200 bootstrapping resamples
were used to reduce overfit bias and for internal validation (Harrell et al., 1982). We obtained c = 0.68.

These statistical analyses and the nomogram were performed using S-Plus software (PC Version 2000 Professional; Insightful Corp, Redmond, WA) with additional functions (called Design)(Harrell 2001) added.
5

DISCUSSION

The presented approach might be improved if we get

richer data bases. The concordance index could go
up if we take account another features from individuals and companies. It would be very interesting, for
instance, to take account information about recurrent
traffic accidents in an individual. There are several
extensions of Cox model designed to deal with recurrent events that have become popular (Andersen & Gill
1982), (Prentice et al., 1981), (Wei et al., 1989), and
many other useful methods.
The model includes one of the variables referred to
the preventive organization of the company (Topap).
This suggest a connection between organizational
safety culture of companies and work related traffic
accidents. In fact, there are studies on this issue, see
Bomel (2004) and references therein. It is pointed out
in Bomel (2004) that key components of organizational safety culture are training, procedures, planning,
incident feedback, management and communications.
Among reached conclusions it is founded out that the
most critical factors for management of car driver
occupational road risk (ORR) are fatigue, pressure,
training, incident management and communications.
It would be interesting to explore this and related
factors within time-to-event framework.
REFERENCES
Andersen, P. K. & Gill, R. D. 1982. Coxs regression model
for counting processes: a large sample study. Annals of
Statistics 10, 110020.

Bomel 2004. Safety culture and work related road accidents. Road Safety Research Report 51, Department for
Transport, London.
Boufous, S. & Williamson, A. 2006. Work-related traffic
crashes: A record linkage study. Accident Analysis and
Prevention 38 (1), 1421.
Cellier, J., Eyrolle, H. & Bertrand, A. 1995. Effects of age
and level of work experience on occurrence of accidents.
Perceptual and Motor Skills 80 (3, Pt 1), 931940.
Clarke, D., Ward, P., Bartle, C. & Truman, W. 2005. An indepth study of work-related road traffic accidents. Road
Safety Research Report 58, Department for Transport,
London.
Cox, D. R. 1972. Regression models and life tables (with
discussion). Journal of the Royal Statistical Society Series
B 34, 187220.
DfT 2003. Driving at Work: Managing workrelated road
safety. Department for Transport, HSE Books.
Harrell, F. E. 2001. Regression Modeling Strategies. With
Aplications to Linear Models, Logistic Regression, and
Survival Analysis. Springer.
Harrell, F. E., Califf, R. M. & Pryor, D. B. 1982. Evaluating
the yield of medical tests. JAMA 247, 25432546.
Hjar, M., Carrillo, C. & Flores, M. 2000. Risk factors in
highway traffic accidents: a case control study. Accident
Analysis & Prevention 32, 5, 703709.
Lawton, R. & Parker, D. 1998. Individual differences in accident liability: A review and integrative approach. Human
Factors 40, 655671.
Lewin, I. 1982. Driver training: a perceptual motor skill
approach. Ergonomics 25, 917925.
Lpez-Araujo, B. & Osca, A. 2007. Factores explicativos de la accidentalidad en jvenes: un anlisis de la
investigacin. Revista de Estudios de Juventud 79.
Prentice, R. L., Williams, B. J. & Peterson, A. V. 1981. On
the regression analysis of multivariate failure time data.
Biometrika 68, 373389.
Wei, L. J., Lin, D. Y. & Weissfeld, L. 1989. Regression
analysis of multivariate incomplete failure time data by
modeling marginal distributions. Journal of the American
Statistical Association 84, 10651073.
WRSTG 2001. Reducing atwork road traffic incidents.
Work Related Road Safety Task Group, HSE Books.

760

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

New performance indicators for the health and safety

domain: A benchmarking use perspective
H.V. Neto, P.M. Arezes & S.D. Sousa
DPS, School of Engineering of the University of Minho, Guimares, Portugal

ABSTRACT: Given the need of organisations to express their performance in the Health and Safety domain
by positive indicators, such as the gains within that domain and proactive actions carried out to improve work
conditions, a research effort has been carried out in order to respond to this particular need, or at least to have a
valid contribute for that purpose. As a result of this effort, a performance scorecard on occupational Health and
Safety was developed and is briefly presented in this paper.

INTRODUCTION

The contemporary systemic focus adopted in the

domain of Health and Safety (H&S) at work establishes itself as the consequence of an entire evolutional process, either in terms of developing these
issues within the organisations, or at the models of
operational management level. More than a simple
add-value contribute in terms of risk management, a
H&S management system is a philosophical and operational challenge for the organization that attempts to
implement it. The reason for this challenge is because
it requires a structured approach for identification,
evaluation and control of the risk factors of their activity and an ongoing effort to continuously improve its
efficiency and effectiveness.
A management system, despite being structured
according to an international or national standard, it
only will be effective if its implementation results
in a clear improvement in a significant number of
indicators, which are typically designated as performance indicators. While most of the indicators used
to evaluate the organisations management are positive indicators, i.e., they express gains rather than
losses (company profits, no. of customers, market
share, etc.), in the H&S domain the used indicators
are, traditionally, negatives, i.e., indicators related
with issues that companies intend to minimize (accident rates, economical and material losses, costs, etc.).
Despite the significant importance of these issues, it is
clear that they do not fit the need for a proactive profile required by a management system, because they
reflect and are supported predominately by a reactive
behaviour.
The current management models assume a systematic monitoring (Pinto, 2005), privileging proactive

indicators that promote the need for continuous

and updated information, and a preventive approach
regarding the risks factors in the occupational environments. Therefore, it is urgent that the organisations
H&S performance is also characterized by positive
indicators, reflecting gains associated with that particular domain, as well as by a proactive approach
in seeking the most adequate working conditions. In
accordance, the main aim of this study was to identify a possible H&S performance key-indicators
scheme, which might be able to play a very important
role in defining a specific organization position, in
terms of H&S performance, when compared with others companies (Benchmarking). This was carried out
considering also several scenarios (regarding external
competitors, regarding other delegations or units of
the same company or other subsidiary organizations).
2
2.1

CRITICAL FACTORS OF SUCCESS

Structured Matrix of Performance Indicators

Benchmarking is one of the most famous tools used

to support the process of continuous improvement. It
represents a tool that accomplishes with the need for
continuous and systematic analysis and measurement
process (Camp, 1993), using as reference a specific
baseline, or golden standard. This allows, at least
theoretically, thecomparisonof, forexample, practices,
proceduresandperformancestandardsbetweendifferent
organizations and within the same organization. From
a methodological point of view, it can be stated
that Benchmarking is, simultaneously, an approach,
consideringthatitseeksanorganizationalimprovement
and an identification of the best practices at the product
and organizational processes level, and a process of

761

continuous and systematic search, since it implies the

development of an original idea, strongly based on
mechanisms of ownership, adaptation (or adjustment),
optimization and development, but also a technique
for data collection and for knowledge generation.
Benchmarking is, actually, a very successful
methodology in organizational contexts, but it requires
a definition of critical factors of success that emphasise the organisation strategy and mission for improving some key indicators. Based on this assumption and
on the relevance to have H&S performance indicators
that complies with the contemporary organizational
needs, an inventory of the H&S critical factors of success found in the literature (technical, scientific and
normative documents) was developed, after this some
factors identified by the performance indicators listed
used to monitor and/or measure was collected. Finally,
a systematization of the selected elements through a
structured matrix of performance indicators was carried out. This matrix will allow establishing a proposal
for a performance scorecard.
The scorecarding is one of the contemporary
archetypes of management that are clearly framed
by the principles of continuous improvement. Its
aim is the development of systematic organizational
processes of monitoring and measuring that provide
continuous information and allows the implementation of a continuous improvement process (Armitage
& Scholey, 2004). This improvement will be reached
because it favours a reporting based in key indicators,
which will be (or at least, might be) representative of
the performance on critical factors of success for one,
or more, of the organisational domains.
2.2

Performance Scorecard for H&S Management

Systems

The need to establish a matrix of performance results

that considers both proactive and reactive practices,
and that fit within the Portuguese and European standards, led to a proposal for a structured matrix of
performance results to be used in the scope of a H&S
management systems. The international vision, which
was supposed to be the basis of this proposal, was
considered not only because the references arising
from the technical and scientific world framework,
but also due to the requirements imposed at an European and international level regarding H&S regulatory
and legal framework. This international vision is also
stated in the designation selection for the proposal,
SafetyCardPerformance Scorecard for Occupational Safety and Health Management Systems.
However, this was not the only challenge that we
have tried to address, many others were considered in
this the attempt to structure a performance scorecard
of this type. The complexity and multidisciplinarity of
the H&S domain, meant that the developed model had

to reflect the major carriers operating in this area, to

promote a consistent diagnosis of the target fields of
the analysis, in order to provide the identification of
real performance improvements, and, simultaneously,
be applicable to the largest number of organizations.
Some of the aspects that confirm the need for operational ability come from the fact that the structured
matrix contains:
critical success factors and key elements that can,
naturally, evolve into a more minimalist matrix and
with a broader scope;
extremely strategic elements, both for the H&S
domain and for the success of the organization as a
whole, leading to the possibility to be integrated in
an overall performance matrix, such as the Balanced
Scorecard;
a great operational flexibility, leading to a model
that is adaptable to different organizations and/or
temporalities, i.e., that allows to be used entirely, or
segmented according to users needs;
the main principles of scorecarding, as well as the
main requirements of Performance Benchmarking.
Due to some restraints associated with the dimension
of this paper, it will be not possible to explain here the
entire conceptual model and the operational considerations previewed by the SafetyCard. Therefore, we
will try to identify, the main elements of the model,
and to synthesize them in a few points, such as:
Organizational design considers aspects related
to the H&S services organisation model, the coverage assured by technical elements and the systemic
approach of the carried out activities. Accordingly,
the used indicators are related to the type of technical coverage and the systemic focus regarding H&S
operation and management;
Organizational culture considers aspects related
to the beliefs, the rules and standards of behaviour
set by the organization on H&S matter. Therefore,
it considers indicators that refer to organisation and
individual values, rules and codes of conduct and to
the basic assumptions, description and evaluation
of the H&S services performance;
Occupational health services considers aspects
related to the surveillance and health promotion,
thus, it contemplates performance indicators structured on the basis of the two considered segments:
health surveillance and promotion;
Operational service of occupational hygiene &
safety consider indicators related to the work accidents statistical control, the development and design
of training programmes, the planning and implementation of prevention and protection measures.
Therefore, the segments of analysis refer to the field
of the organization and operability of the Hygiene
& Safety services, accidents control and analysis,
training and prevention and protection actions;

762

 Internal emergency plan the definition of a

functional organisation, based on the structural
planning, accountability, selection, preparation and
testing of means and devices that should be able
to respond to emergency situations, assumes here a
role as the main vector of performance. Therefore,
the performance indicators were organized according to three analytical segments: (i) Planning, (ii)
Attributes and Responsibilities and (iii) Devices.
Monitoring and/or measurement services considers analytical segments, such as the (i) control
of workplace environmental conditions, (ii) the
mechanisms of monitoring and/or measurement
and (iii) the implementation of corrective actions.
Therefore, the selection of performance indicators
intends to evaluate the organisation practices in the
field of environmental control, monitoring and/or
measurement;
Work safety equipments considers issues related
to organizational practices in what concerns the
integration of H&S requirements into the process
of selection, acquisition and maintenance of work
safety equipments.
2.3

Performances Weighting and Classification

One of the objectives of the analytic model was to

encourage a global, but also partial, screening of
the H&S performance level, hence that has assumed
both a quantitative and qualitatively character. It was
intended to establish a normalized system of performance indicator, so that all performance measures
could be compared. This process of normalization is
grounded in a numerical basis, where each indicator
always assumes a result between zero and one, and
in some cases they assume a traditional binary score
(zero or one).
Since the entire structure of weights is made on
this basis, from the application of the first coefficients
of importance on the first level of the structure, all
the scores assume a continuous distribution between
the limits associated with each analytic element. The
maximum amount possible in each of the stages will
be equivalent to the value of the multiplier (M), originating that each stage may vary always between zero
and one, even the final classification. The sum of the
scores obtained in each domain represents a total value
of performance, both in quantitative terms and in qualitative terms, since there is the possibility to analyse
the final numerical value through a traditional discrete
scale of classification ( Very good, good, . . .).
3

SAFETYCARD: RESULTS OF A CASE

STUDY

At this point, the aim will be to make a presentation

of the most important obtained results from a pilot

implementation of the SafetyCard (Neto, 2007). The

used organization operate in the Construction branch,
specifically on the Construction of Buildings, and
had 134 employees, 6 of them are H&S practitioners
(4 technicians of level V and 2 of level III).
As some indicators had to be standardised to integrate the matrix of performance results and considering that we had not any reference or Benchmarking
partner, it was used a previous period (past year) to
allow the comparison. The SafetyCard was not applied
entirely, since 6 performance indicators were not applicable to the considered organization and 2 did not have
the data for the needed computation. However, taking
into account the flexibility of the model, the results
matrix of performance has been modelled to that specific company, without losing quality or empirical
relevance. For illustrating this exercise, an xample of
the scorecard for the considere company is presented
in table 1. As it is possible to see in this table, the overall performance result was of 0,740, which according
to the scale considered, reflects a good performance.
Thus, we conclude that this organization had a good
performance in matter of H&S at Work.
The previous results would be the overall assessment of performance, but we can use this analysis and
detail the evaluation. Accordingly, and based in the
analytical domains previously stated, we can mention
the following aspects:
Organizational design consistent organizational
structure, in part this is a result of excellent technique coverage. The point less strong is the reduced
systemic approach, but that is due to the fact of
the H&S management system is being prepared
and implemented. This system is being prepared
under the guidelines of OHSAS 18011:1999/NP
4397:2001, which certainly will bring, in the short
term, benefits to the organization, both in terms of
performance and at a practical level.
Organizational culture characteristic traits of
institutional values, norms and standards of
behaviour and basic assumptions of description and
evaluation in matter of H&S have been identified,
which are transposed into a strong organizational
culture focused to the protection and maintenance
of the acceptable working conditions.
Occupational health services great operational
and structuring sense. The organization assures all
the procedures to promote health in the workplaces
and implement a few mechanisms for monitoring
workers health.
Operational service of occupational health &
safety inadequate monitoring and documentation
at the risk prevention level. The organization had
to pay strong fines, both in monetary terms, and
in terms of severity and absenteeism induced by
accidents.

763

Table 1.

Example of the Performance Benchmaring Scorecard (summary) for the case study.
Weighted segment

Weighted domain

Baseline
weight

M (a)

Partial

M (b)

Partial

Analytic domain

Analytic segment

No. of
indicators

Organizational
design

Technical covering
Systemic focus

4
2

1,00
0,25

0,70
0,30

0,70
0,08

0,05

0,04

Organizational
culture

Values
Behaviour standards
Basics assumptions

3
3
4

1,00
1,00
0,92

0,50
0,20
0,30

0,50
0,20
0,28

0,20

0,20

Occupational health
services

Surveillance
Promotion

6
2

0,85
1,00

0,75
0,25

0,64
0,25

0,10

0,09

Operational service
of occupational
safety & hygiene

Organization
Accidents incidence
Formation
Prevention
Protection

3
10
6
5
3

0,40
0,67
0,76
0,49
1,00

0,05
0,15
0,25
0,40
0,15

0,02
0,10
0,19
0,19
0,15

0,25

0,16

Internal emergency
plan

Planning
Attributes and
responsibilities
Mechanisms

0,90

0,40

0,36

7
10

1,00
0,60

0,25
0,35

0,25
0,21

0,15

0,12

Control of environmental
work conditions
Mechanisms of monitoring
and/or measurement
Corrective action

0,13

0,55

< 0, 07

0,20

0,09

4
2

0,69
1,00

0,30
0,15

0,21
0,15

Maintenance
Safety instructions

4
3

1,00
0,84

0,50
0,50

0,50
0,42

0,05

0,05

Monitoring and/or
measurement
services

Safety equipments

Total

90

0,740

(a) The letter M represents the multiplier associated with the baseline weight and with the maximum score that can be obatined
in a specific segment. (b) The letter M represents the multiplier associated with the segment and with the maximum score that
can be obtained in a specific domain.

Internal emergency plan excellent basis of structuring and planning, with the organization ensuring
the main procedural mechanisms of response to
emergencies (plans, responsibilities and devices).
However, there is an operational weakness, because
the organization does not have evidences of its operational ability (for example, no fire drills were
carried out).
Monitoring and/or measurement service low level
of monitoring of the environmental conditions arising from the adopted processes. The organization
acknowledged the existence of some ergonomic
nature risk factors and occupational exposure to
harmful agents, like noise. However, have not developed specific procedures for evaluating the exposure levels of the workers. This points as become a
critical segment, due to the fact that it penalizes the
organization.
Work safety equipments great strategic importance
is given both to the acquisition and maintenance,

and to the prescription of elements relating to

the work safety equipments. This was the critical
success area where the organization obtained the
best score.

CONCLUSIONS

The best way to conclude one journey is go back to

the starting point. From the literature review, it seems
consensual that organizations need a structured matrix
of positive indicators that go beyond the assessment
of some organizations attributes that do not favour the
idea of a set, and does not fully reflect the overall H&S
performance. So it is important to have a scorecard that
can be able to reflect the structural H&S performance
of an organization, and to allow internal and external
comparisons (Benchmarking in the various possible
scenarios).

764

The performance scorecard that has been developed and implemented reflects applicability and show
technical-scientific relevance, allowing the diagnosis
of a structural H&S Management System. This
diagnosis could be carried out both in terms of work
conditions and organizational values, and in H&S
performance monitoring and/or measurement.
Finally, it is also necessary to highlight that there is
some work to be done, but it is expected that the presented tools could be improved and refined in order
to have a reliable and useful tool for performance
assessment.

Camp, R. 1993. Benchmarking: O Caminho da Qualidade TotalIdentificando, Analisando e Adaptando as

Melhores Prticas da Administrao Que Levam Maximizao da Performance Empresarial. So Paulo: Livraria
Pioneira Editora (in Portuguese).
Neto, H.V. 2007. Novos Indicadores de Desempenho em
Matria de Higiene e Segurana no Trabalho: perspectiva de utilizao em Benchmarking, Dissertation thesis of
the MSc. in Human Engineering, School of Engineering.
Guimaraes: University of Minho (in Portuguese).
Pinto, A. 2005. Sistemas de Gesto da Segurana e Sade
no TrabalhoGuia para a sua implementao, 1a Edio.
Lisbon: Edies Slabo (in Portuguese).

REFERENCES
Armitage, H. & Scholey, C. 2004. Hands-on scorecardingHow strategy mapping has helped one organization
see better its successes and future challenges. CMA
Management 78(6): 3438.

765

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Occupational risk management for fall from height

O.N. Aneziris & M. Konstandinidou
NCSR DEMOKRITOS Greece

I.A. Papazoglou
TU Delft, Safety Science Group, Delft, The Netherlands

M. Mud
RPS Advies BV, Delft, The Netherlands

M. Damen
RIGO, Amsterdam, The Netherlands

J. Kuiper
Consumer Safety Institute, Amsterdam, The Netherlands

H. Baksteen
Rondas Safety Consultancy, The Netherlands

L.J. Bellamy
WhiteQueen, The Netherlands

J.G. Post
NIFV NIBRA, Arnhem, The Netherlands

J. Oh
Ministry Social Affairs & Employment, The Hague, The Netherlands

ABSTRACT: A general logic model for fall from height has been developed under the Workgroup Occupational Model (WORM) project, financed by the Dutch government and is presented. Risk has been quantified
risk for the specific cases of fall from placement ladders, fixed ladders, step ladders, fixed scaffolds mobile
scaffolds, (dis) assembling scaffolds, roofs, floor openings, fixed platforms, holes, moveable platforms and
non moving vehicles. A sensitivity analysis assessing the relative importance of measures affecting risk is presented and risk increase and risk decrease measures are assessed. The most important measures in order to
decrease fatality risk owing to falls from fixed ladders is the way of climbing, for step ladders their location,
for roofs, floors and platforms not to work on them while being demolished, for mobile scaffolds the existence
of safety lines, for fixed scaffolds protection against hanging objects, for work near holes and (de) installing
scaffolds the use of fall arrest and for moveable platforms and non moving vehicles the existence of edge
protection.

INTRODUCTION

Occupational fatalities and injuries caused by falls

from height pose a serious public problem and are the
leaders of occupational accidents in the Netherlands.
Falls from height constitute 27% of the reported 12000
accidents, which have occurred in the Netherlands

between 1998 and 2004, while the number of deaths

is on the average 25 per year.
Several studies have been performed examining the
causes of injuries and deaths from falls, such as those
from the National Institute of Occupational Safety and
Health (NIOSH) Fatal Accident Circumstances and
Epidemiology (FACE) reports (NIOSH, 2000), OSHA

767

report on falls from scaffolds (OSHA 1979), OSHA

report on falls from elevated platforms (OSHA 1991),
the study of McCann (2003) for deaths in construction
related to personnel lifts and the study of HSE (2003)
for falls from height in various industrial sectors.
The Dutch government has chosen the quantitative
risk approach in order to determine the most important paths of occupational accidents and optimize the
risk reduction efforts. It has embarked the Workgroup Occupational Risk Model (WORM) project, as
presented by Ale et al., (2008). Major part of the
WORM project is the quantification of occupational
risk, according to the bowtie methodology developed
within the project and presented by Papazoglou & Ale
(2007).
Of the 9000 analyzed GISAI (2005) occupational
accidents, which have occurred in the Netherlands
between 1998 and 2004, 805 have been classified as
falls from placement ladders, 70 from fixed ladders,
187 from step ladders, 245 from mobile scaffolds, 229
from fixed scaffolds, 78 falls while installing or deinstalling scaffolds, 430 falls from roofs, 415 from
floors, 235 from fixed platforms, 74 falls in holes,
205 from moveable platforms and 206 from non moving vehicles. Logical models for fall from height have
been presented by Aneziris et al (2008).
This paper presents the overall quantified risk, the
specific causes and their prioritization for the following occupational hazards: a) fall from placement
ladders, b) fall from fixed ladders, c) fall from step
ladders d) fall from fixed scaffolds e) fall from mobile
scaffolds e) fall while (dis) assembling scaffolds, f)
fall from roofs g) fall from floor openings h) fall from
fixed platforms i) fall in holes j) fall from moveable
platforms k) fall from non moving vehicles.
The paper is organized as follows. After the introduction of section 1, section 2 presents a general logic
model for fall from height and risk results for all fall
from height cases. Section 3 presents the ranking of the
various working conditions and/or safety measures in
terms of their contribution to the risk. Finally section
4 offers a summary and the conclusions.

LOGICAL MODEL FOR FALL

FROM HEIGHT

In this section a general model for fall from height

is presented, which may be applied in all fall from
height cases, while more detailed models for fall from
ladders, scaffolds, roofs, holes, moveable platforms
and non moving vehicles, are described by Aneziris
et al (2008). Figure 1 presents the fall from height
bowtie. The Center event represents a fall or not from
the structure (ladder, scaffold, roof, hole, moving platform, or non moving vehicle) and it is decomposed

Figure 1.

General bowtie for fall from height.

into the initiating event and the safety measures aiming at preventing a fall. The initiating event represents
working on the high structure, while the primary safety
measures preventing a fall are strength and stability of
structure, user stability and the edge protection.
Strength of structure: The structure should be able
to support the imposed load by the user and the associated loads (persons or equipment). It is applicable
to all fall from height cases, with the exception of
fall in hole. It is defined as a two state event with the
following cases: success or loss of strength
Stability of structure: The structure itself through
its design and material provides the necessary stability
so that it does not tip over. It is applicable to all fall
from height cases, with the exception of fall from roof,
floor, platform and fall in a hole. It is defined as a two
state event with the following cases: success or loss of
structure stability
Use stability: Given a strong and stable structure,
the user should be able to remain on the structure without losing his stability. This measure is applicable to
all fall from height cases. It is defined as a two state
event with the following cases: success or loss of user
stability.

768

Table 1. Support safety barriers affecting primary safety barriers, for all fall from height
accidents.
Type of fall

Support barriers
STRENGTH
Type or condition of ladder
Structural Design Construction
Roof Surface Condition

LADDER
SCAFFOLD
ROOF/FLOOR/FIXED
PLATFORM
MOVEABLE PLATFORM
NON MOVING VEHICLE

STRUCTURE STABILITY
Placement and Protection
Type or condition of ladder
Anchoring
Foundation
Scaffold Protection
Foundation/Anchoring
Position of Machinery/ weight
Foundation
Load Handling

LADDER
SCAFFOLD
MOVEABLE PLATFORM
NON MOVING VEHICLE

USER STABILITY
User Ability
User Ability
User Ability
User Ability
Floor Condition
User Ability
External Conditions
Movement Control
Position of Machinery or
weight on platform
Ability
Load Handling
Working Surface

LADDER
HOLE IN GROUND
ROOF/FLOOR/FIXED PLATFORM
SCAFFOLD
MOVEABLE PLATFORM

NON MOVING VEHICLE

Edge Protection: This measure includes the provision of guardrails that enhance the stability of user. It is
applicable to all fall from height cases with the exception of ladders and non moving vehicles. It can be in
one of the following states: present, failed or absent.
2.1

Condition of Lift/Support
Loading

Support safety barriers

A Support Safety Barrier (SSB) contributes to the

adequate function of the Primary Safety Barriers and
influence the probability with which the primary safety
barrier-states occur. There are three types of support
barriers, those affecting structure strength, structure
stability and user stability.
Measures that affect strength are structure specific
and are presented in Table 1. Condition of the structure surface affects the strength of ladder, roof and
moveable platform. Structural design and construction
affects scaffold strength and loading affects non moving vehicles.

Measures that affect the stability of the structure

are structure specific and are presented in Table 1.
Foundation is a measure affecting all structures with
the exception of ladders. Ladder stability is affected
by its condition, placement and protection.
User Ability is a measure affecting user stability
in all fall accidents. Other measures, such as working
surface, depend on the type of fall accident and are
presented in Table 1.
2.2 Probability influencing entities (PIEs)
In several instances the safety barriers of the model
are simple enough to link directly to easily understood working conditions and measures as in the
barrier Anchoring, which affects the stability of a
fixed scaffold. Assessing the frequency with which
anchoring exists is straightforward.
In other instances, however, this is not possible.
For example, the support barrier Floor surface condition may be analysed into more detailed and more

769

Table 2.

PIES characteristics and frequencies.

PIEs
frequency

Barrier
success
probability
0,265

Barrier

PIEs

ROOF
PROTECTION

Edge Protection
absent
No Edge
Protection Next to
non-supporting parts

0,36

Roof/working
platform/floor (parts)
being built or torn
down
Roof/working
platform/floor (parts)
not intended to
support exerted
weight

0,43

Capacity to keep
balance on roof
Walking
backwards
Unfit: unwell
Hands not free
Overstretching
Substandard
movement (slip, trip)
Outside edge protection
Weather
Slope

0,21

Collective fall
arrest (CFA)
Condition of CFA
Personal fall arrest
Anchorpoints FA
State of Maintenance

0,18

ROOF
SURFACE
CONDITION

ABILITY

PFA

concrete measures that affect its quality. Such specific

measures are: (i) floor which is being demolished; (ii)
floor not able to support weight. Similarly the barrier Fall Arrest may be analysed into the following
measures: i) Use of Collective Fall arrest; ii) Maintenance of Collective Fall arrest; iii) Use of Personal Fall
arrest; iv) Maintenance of Personal Fall arrest. Such
factors have the name of Probability Influencing Entities (PIEs). Each influencing factor (PIE) is assumed
to have two possible levels, Adequate and Inadequate. The quality of an influencing factor is then set
equal to the frequency with which this factor is at the
adequate level in the working places. Then the quality
of the barrier is given by a weighted sum of the influencing factor qualities. The weights reflect the relative
importance of each factor and are assessed by the analyst on the basis of expert judgement. Currently equal
weights have been used. This way the probability of

0,17

0,3

0,17

0,205

0,19
0,09
0,35
0,19
0,2
0,17
0,13
0,32
0,118

0,05
0,16
0,16
0,04

a support barrier to be in one of its possible states is

given by the weighted sum of the frequencies of the
influencing factors (RIVM 2008).
PIEs and their frequencies as well as the failure
probability for the barriers they influence for fall from
roofs are presented in Table 2. All other PIEs for fall
from height models are presented in RIVM (2008).
Frequencies of PIEs have been assessed through surveys of the working condition in the Dutch working
population and reflect the Dutch National Average
RIVM (2008).
2.3

Right hand side (RHS)

The right hand side of the fall bowties in combination

with the outcome of the centre event determine the
consequences of the falls. Four levels of consequences
are used: C1: No consequence; C2: Recoverable

770

1,00E-05

PLACEMENT LADDER

1,00E-06

Substandard movements

1,00E-07

Substandard condition/ fitness of person

Use of hands for support

Substandard position of person on equipment

1,00E-08

Location: signs to prevent impact

Mobile
scaffold

Fixed
scaffold

Installing
scaffold

Roof

Floor

Fixed
platform

10

Nonmoving
vehicle

Step
ladder or
steps

Hole in
ground

Fixed
ladder

Moveable
platform

Placement
ladder

Location: protection against impact

1,00E-09

11

12

Dimension: fixation if long

Angle: prevention against sliding
Angle: secured top
Substandard surface/ support
Placement ladder is in good condition

Recoverable injury /hr

Permanent Injury /hr)

Fatality (/hr)

Surface conditon of steps

Dimension/ type

Figure 2.
hazards.

Condition/ presence of anti-slip

Individual risk per hour for fall from height

0
RISK DECREASE

injury; C3: Permanent injury; C4: Death. Events of

the RHS are the Height of the fall, the type of the surface and the medical attention. More details on these
events are presented by Aneziris et al (2008).
Individual risk of death, permanent injury and
recoverable injury per hour have been assessed according to the methodology presented by Aneziris (2008),
Papazoglou et al (2008) and are presented in Figure 2.
Fall from roof has the higher fatality risk 2.2
107 /hr, followed by fall from moveable platform
8.06 108 /hr.

10

20

RISK INCREASE

30

40

50

60

percentage %

Figure 3. Risk fatality increase and risk decrease for fall

from placement ladder bowtie, for various working conditions (PIEs).

FIXED LADDER
Substandard movements
External force exerted on the
person
Use of both hands for climbing/
descending
Substandard condition/ fitness of
person
Substandard position of person on
equipment

IMPORTANCE ANALYSIS

Safety cage
Fixed ladder is not in good condition
(damaged)
Design of fixed ladder (dimension of
rungs)
Design of arrival area (ergonomics
lay-out, etc.)

To assess the relative importance of each factor influencing the risk from fall, two importance measures
have been calculated.
1. Risk decrease: This measure gives the relative
decrease of risk, with respect to the present state, if
the barrier (or PIE) achieves its perfect state with
probability equal to unity.
2. Risk increase: This measure gives the relative
increase of risk, with respect to the present state,
if the barrier (or PIE) achieves its failed state with
probability equal to unity.
Risk decrease prioritizes the various elements of the
model for the purposes of possible improvements.
It is more riskeffective to try to improve first a
barrier with higher risk decrease effect than another
with lower risk decrease.
Risk increase provides a measure of the importance
of each element in the model to be maintained at its
present level of quality. It is more important to concentrate on the maintenance of a barrier with high risk
increase importance than one with a lesser one. The
effect each PIE has on the overall risk is presented in
Figures 314.

Surface conditon of steps

0

20 40 60 80 100 120 140

percentage (%)

RISK DECREASE

RISK INCREASE

Figure 4. Risk fatality increase and decrease for fall from

fixed ladder bowtie, for various working conditions (PIEs).

3.1 Fall from height-ladders

Placement ladder: The most important measure in
order to decrease fatality risk is the location of signs to
prevent impact and the use of the right type of ladder
which, if used 100% of the time a placement ladder is
used, will decrease risk by 21% and 20% respectively.
The most important measure in order to maintain risk
is to keep the ladder in good condition. If this is not
done risk increases by 54%.

771

Fixed ladder: The most important measure in

order to decrease fatality risk is to use both hands
for climbing, and to be in the right position (not
on top and not overreaching). If these measures are
used 100% a fixed ladder is used, risk decreases
by 34% and 21% respectively. The most important measure to maintain risk is the good physical
condition of the worker and avoidance of external force to be exerted on him. If this is not

FIXED SCAFFOLD
Fall arrestors, safety nets
Protection of scaffold against being struck by a
vehicle
Health checks based on clear criteria for people
working on heights
No ladder placed on top of a scaffold
Safe access of scaffold
Protection against hanging/swinging objects
Footings capable of supporting the loaded scaffold
without displacement.
Scaffold on a level and firm foundation
Scaffold surfaces non-slippery and free from
obstacles

STEP LADDER

Adequate floors (all boards present, stable and not

broken)
Adequate stabilising supports and/or ties to
adjacent structures

Substandard movements

Erection/ modification of scaffold by competent

persons

External force exerted on the person on the

stepladder

Proper design and construction

Use of hands for support

Edge protection absent

Substandard condition/ fitness of person

Regular inspection of scaffolds

Substandard position of person on

equipment

Adequate edge protection

Substandard placement of stepladder

RISK DECREASE

Surface/ support

RISK INCREASE

20

40

60
80
100
percentage (%)

120

140

Location of ladder

Figure 7. Risk fatality increase and risk decrease for fall

from fixed scaffold bowtie, for various working conditions
(PIEs).

Step ladder is in good condition

Surface conditon of steps
Dimension
Condition/ presence of anti-slip shoes

20

40

60

RISK INCREASE
RISK DECREASE

80

100

percentage %
(DE) INSTALLING SCAFFOLD
Fall arrestors, safety nets

Figure 5. Risk fatality increase and risk decrease for fall

from step ladder bowtie, for various working conditions
(PIEs).

Protection of scaffold against being struck by a

vehicle
Health checks based on clear criteria for people
working on heights

MOBILE SCAFFOLD

Safe access of scaffold (from the inside)

Protection against hanging/swinging objects

Fall arrestors, safety nets

Protection of scaffold against being struck by a
vehicle
Health checks based on clear criteria for people
working on heights

Footings capable of supporting the loaded scaffold

without displacement.
Scaffold on a level and firm foundation

No ladder placed on top of a scaffold

Adequate floors (all boards present, stable and not
broken)

Safe access of scaffold (from the inside)

No construction/ adjacent structure to support/

attach scaffolds

No unstable objects used to support scaffold

Scaffold on a level and firm foundation

Adequate stabilising supports and/or ties to

adjacent structures

Scaffold surfaces non-slippery and free from

obstacles
Adequate floors (all boards present, stable and not
broken)

Criteria (eg. Wind-load, height-width ratio, etc.) to

determine anchoring requirements

Adequate breaks (mobile scaffolds only)

Erection/ modification of scaffold by competent

persons

Adequate stabilising supports and/or ties to

adjacent structures
Criteria (Wind-load, height-width ratio) to determine
anchoring requirements
Erection/ modification of scaffold by competent
persons

Proper design and construction

Edge protection absent

Proper design and construction

Adequate edge protection

Edge protection absent

Regular inspection of scaffolds

10

20

30

40

50 60

70

80

90 100

Adequate edge protection

RISK DECREASE

RISK INCREASE

10

20

30

40

50

60

70

80

90

RISK DECREASE

100

RISK INCREASE

percentage (%)

Figure 6. Risk fatality increase and risk decrease for fall

from mobile scaffold, for various working conditions (PIEs).

Figure 8. Risk fatality increase and risk decrease for fall

scaffold while (de) installing it, for various working conditions (PIEs).

772

FALL FROM PLATFORM

FALL FROM ROOF

State of Maintenance

State of Maintenance

Anchorpoints FA
Personal fall arrest

Anchor points FA

Condition of CFA

Personal fall arrest

Collective fall arrest (CFA)

Collective fall arrest

Slope

Weather

Weather

Outside edge protection

Outside edge protection

Substandard movement (slip, trip)

Substandard movement (slip, trip)

Overstretching

Overstretching

Hands not free

Hands not free

Unfit

Unfit

Walking backwards

Walking backwards

Capacity to keep balance on floor

Capacity to keep balance on roof

Roof not intended to support exerted
weight
Roof being built or torn down

Platform overloaded
Platform being built or demolished
EP absent

No EP next to non-supporting parts

EP absent

0
0

10

RISK DECREASE

20

30

40

50

RISK INCREASE

60

70

80

RISK DECREASE

20

40

RISK INCREASE

60

80

100

percentage %

percentage (%)

Figure 11. Risk fatality increase and risk decrease for fall
from platform, for various working conditions (PIEs).

Figure 9. Risk fatality increase and risk decrease for fall

roof, for various working conditions (PIEs).

FALL IN HOLE
FALL FROM FLOOR

State of Maintenance PFA

State of Maintenance FA

Anchor points FA
Anchorpoints FA

Personal fall arrest

State of Maintenance CFA

Personal fall arrest

presence

Collective fall arrest

Weather

Illumination

Outside edge protection

Substandard movement

Substandard movement
(slip, trip)

Overstretching
Hands not free

Walking backwards

Unfit
Walking backwards

External force by
machinery or equipment

Problems to keep balance on floor

Floor not intended to support exerted
weight
Floor being built or demolished

Hole (edge) protection

absent

No EP next to non-supporting parts

Cover absent

Edge protection absent

0
RISK DECREASE

RISK INCREASE

20

40

60

80

100 120 140

RISK DECREASE

10

20

30

40

RISK INCREASE

50

60

70

80

90

100

percentage %

percentage %

Figure 10. Risk fatality increase and risk decrease for fall
from floor, for various working conditions (PIEs).

Figure 12. Risk fatality increase and risk decrease for fall
in hole, for various working conditions (PIEs).

done fatality risk increases by 116% and 108%

respectively.
Step ladder: The most important measure in order
to decrease fatality risk is the location of ladder, so
that it cannot be hit by falling object. This measure
increases risk by 26%, if used 100% of time a step
ladder is used. The most important measure in order
to maintain risk is the placement of the step ladder, so

that it is fully extended. If this is not done risk increases

risk by 85%.
3.2 Fall from height- Scaffolds
Mobile scaffolds: The most important measure in order
to decrease fatality risk is the use of a safety line or
a harness belt, which decrease risk by 60%, if used

773

100% of time a mobile scaffold is used. In case they

do not exist risk will increase by 88%.
Fixed scaffolds: The most important measure in
order to decrease fatality risk is the existence of
protection against hanging/swinging objects, which
decreases risk by 75%, if used 100% of time a fixed
scaffold is used. In case it does not exist fatality risk
increases by 122%.
(De) Installing scaffolds: The most important measure in order to decrease fatality risk is the use of fall
arrestors and safety nets, which decrease risk by 48%,
if used 100% of time a scaffold is installed or deinstalled. In case they do not exist risk will be increased
by 93%.

FALL FROM MOVEABLE PLATFORM

Fixation of temp. Platform
Outside the platform
Substandard movement (slip, trip)
Overstretching
Hands not free
Unfit
Self induced external force by machinery or
equipment
Anchoring
Foundation: stabilisatoren
Foundation: ondergrond
Position of the beams
Position of load
Moveable height hit by vehicle or hanging/swinging
load
User hit by rolling, sliding, swinging or hanging object
Heavy wind
Control failure (operator or machine induced)
Load blocks movement
Unclear control panel
Malfunction (of transmission of movement)
Substandard lifting/hoisting mechanism
Substandard brake
Metal fatigue and corrosion of structural/support parts
Wear and tear of structural/support parts

3.3 Fall from roofs

PFA condition/state
PFA anchoring
Personal Fall Arrest
Collective Fall Arrest lifts
Edge Protection Absent

RISK DECREASE

100 200 300 400 500 600 700 800 900 1000

RISK INCREASE

percentage %

Figure 13. Risk fatality increase and risk decrease for fall
from moveable platform, for various working conditions
(PIEs).

The most important measure in order to decrease

fatality risk is to avoid work on a roof that is being
demolished, which decreases risk by 58% if used 100%
of time while working on a roof. The most important
measures to maintain risk is not to walk on weak spots
of roofs and to maintain roof edge protection. If they
are not followed risk will increase risk by 70%.
3.4 Fall from floors or balconies
The most important measure in order to decrease fatality risk is to avoid work on a floor that is being
demolished, which decreases risk by 21% if used 100%
of time while working on a floor or balcony. The
most important measures in order to maintain risk is
to keep safety nets in a good condition, have sufficient anchoring points for fall arrests and maintain fall
arrest equipment. If each of them is not followed risk
will increase by 116%.

FALL FROM NON MOVING VEHICLE

Hands not free
Unwell
Moving backwards
External force
Substandard movement
Uneven
Objects
Wet
Unknown surface (strength)

3.5 Fall from fixed platforms

Unknown surface
Slope

The most important measure in order to decrease fatality risk is avoid working on a platform that is being
demolished, which decreases risk by 23% if used 100%
of time while working on affixed platform. The most
important measures to maintain risk is to maintain the
edge protection of platforms. If this is not done risk
increases by 70%.

No grip
Round/rolling parts
Unbalanced loading
Unsecured load
Loading and stabilising
Position of the road
Corrosion

3.6 Fall from hole in ground

Edge protection absent

20

40

60

80

100

120

140

160

percentage %
RISK DECREASE

RISK INCREASE

Figure 14. Risk fatality increase and risk decrease for fall
from non moving vehicle, for various working conditions
(PIEs).

The most important measure in order to decrease fatality risk is the existence of fall arrest which decreases
risk by 45% if used 100% of time while working near a
hole. The most important measure in order to maintain
risk is the existence of edge protection. If it is absent
risk will increase risk by 88%.

774

3.7

Fall from moveable platform

The most important measure in order to decrease fatality risk is the existence of edge protection, which
decreases risk by 65% if used 100% of time while
working near on a moveable platform. The most important measure in order to maintain risk is the fixation
of the platform, since its absence will increase risk 9
times.
3.8

Fall from non moving vehicle

The most important measure in order to decrease fatality risk is the existence of edge protection which
decreases risk by 39% if used 100% of time while
climbing on a non moving vehicle. The most important measure in order to maintain risk is securing and
balancing load, since their absence will increase risk
by 139% and 137% respectively.
4

CONCLUSIONS

A general logical model has been presented for quantifying the probability of fall from height and the various
types of consequences following all fall from height
accidents. The model has been used for risk reducing
measures prioritization, through the calculation of two
risk importance measures: the risk decrease and the
risk increase. The calculations were made for fatality
risk.

Aneziris O.N, Papazoglou I.A., Baksteen H., Mud M.L., Ale

B.J.M, Bellamy L.J., Hale A.R., Bloemhoff A., Post J.,
Oh J.I.H., 2008. Quantified risk assessment for fall from
height. Safety Science, Volume 46, Issue 2: 198-220.
GISAI, 2005. Geintegreerd Informatie Systeem Arbeids
Inspectie: Integrated Information System of the Labor
Inspection in the Netherlands.
HSE (2003). Fall from HeightPrevention and risk control
effectiveness, ISBN 07176221 5, http://www.hse.gov.uk/
research/rrpdf/rr116.pdf.
McCAnn M. (2003) Deaths in construction related to personnel lifts, 1992-1999, Journal of Safety Research,34,
507514.
NIOSH, (2000). Worker Death by Falls, US department
of Health and Human Services, www.cdc.gov/elcosh/
docs/d0100/d000057/d000057.html.
OSHA (1979). Occupational fatalities related to scaffolds as found in reports of OSHA fatality/catastrophe
investigations, Washington DC.
OSHA (1991). Selected occupational fatalities related to
vehiclemounted elevating and rotating work platforms
as found in reports of OSHA fatality/ catastrophe investigations , Washington DC.
Papazoglou I.A., Ale B.J.M., 2007. A logical model for quantification of occupational risk, Reliability Engineering &
System Safety 92 (6): 785-803.
Papazoglou I.A, L.J. Bellamy, K.C.M. Leidelmeijer, M.
Damen, A. Bloemhoff, J. Kuiper, BJ.M. Ale, J.I.H.
Oh, 2008, Quantification of Occupational Risk from
Accidents, submitted in PSAM 9.
RIVM 2008 WORM Metamorphosis Consortium. The Quantification of Occupational Risk. The development of
a risk assessment model and software. RIVM Report
620801001/2007 The Hague.

REFERENCES
Ale B.J.M., Baksteen H., Bellamy L.J., Bloemhof A.,
Goossens L., Hale A.R., Mud M.L., Oh J.I.H., Papazoglou
I.A., Post J., and Whiston J.Y.,2008. Quantifying occupational risk: The development of an occupational risk
model. Safety Science, Volume 46, Issue 2: 176-185.

775

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Occupational risk management for vapour/gas explosions

I.A. Papazoglou
TU Delft, Safety Science Group, Delft, The Netherlands

O.N. Aneziris & M. Konstandinidou

NCSR DEMOKRITOS, Aghia Paraskevi, The Greece

M. Mud
RPS Advies BV, Delft, The Netherlands

M. Damen
RIGO, Amsterdam, The Netherlands

J. Kuiper & A. Bloemhoff

Consumer Safety Institute, Amsterdam, The Netherlands

H. Baksteen
Rondas Safety Consultancy, The Netherlands

L.J. Bellamy
White Queen, The Netherlands

J.G. Post
NIFV NIBRA, Arnhem, The Netherlands

J. Oh
Ministry Social Affairs & Employment, The Hague, The Netherlands

ABSTRACT: Chemical explosions pose a serious threat for personnel in sites producing or storing dangerous
substances. The Workgroup Occupational Risk Model (WORM) project financed by the Dutch government
aims at the development and quantification of models for a full range of potential risks from accidents in
the workspace. Sixty-three logical models have been developed each coupling working conditions with the
consequences of accidents owing to sixty-three specific hazards. The logical model for vapour/gas chemical
explosions is presented in this paper. A vapour/gas chemical explosion resulting in a reportable-under the Dutch
law-consequence constitutes the centre event of the model. The left hand side (LHS) of the model comprises
specific safety barriers, that prevent the initiation of an explosion and specific support barriers that influence the
adequate functioning of the primary barriers. The right hand side (RHS) of the model includes the consequences
of the chemical explosion. The model is quantified and the probability of three types of consequences of an
accident (fatality, permanent injury, recoverable injury) is assessed. A sensitivity analysis assessing the relative
importance of each element or working conditions to the risk is also presented.

INTRODUCTION

Chemical explosions pose a very serious threat for

personnel in sites handling hazardous substances.
According to OSHA 201 fire and explosion related
accidents occurred in 2006 in US (OSHA 2008). Only
in the Netherlands 25 explosions occur per year from

which at least 20 can be characterized as chemical

explosions (Baksteen et al., 2008). Chemical explosions may be due to different causes such as creation
of explosive atmospheres and mixtures, sudden ignition of flammable substances, ignition of explosives
or violent exothermic reactions. Their consequences
are usually serious and may involve multiple victims.

777

The Workgroup Occupational Risk Model (WORM)

project has been launched by the Dutch government in
order to manage and reduce occupational risk. The
aim of the project is the quantification of occupational
risk through logical models for a full range of potential risks from accidents in the workspace (Ale et al.,
2008). Data for the development of these models are
derived from the GISAI database (GISAI 2005) of the
Netherlands Ministry of Work which includes approximately 12500 accident cases reported between January
1998 and February 2004.
Of the analysed GISAI accident cases, 126 cases
have been classified as vapour/gas chemical explosions accidents with reportable consequences. The
modelling and the quantification of these cases are
described in this paper. Quantification for other types
of occupational accidents such as falls from ladders
(Aneziris et al., 2008a), or crane activities occupational accidents (Aneziris et al., 2008b) has already
been performed within the WORM project. An overall
assessment of the risk from 63 specific occupational
hazards is given in Papazoglou et al (2008).
From the observed accident cases scenario-models
have been firstly developed to capture the sequence of
events leading to the accident (Bellamy et al., 2007).
The scenario-model is the basis for the logical modelling in the WORM project (Papazoglou 2007). This
logical model consists in successive decomposition of
the overall accident consequence into simpler and simpler events until a final level of event resolution is
achieved. Each level of events is logically interconnected with the more general events of the immediately
upper level. The events of the lower level of decomposition form an influence diagram consisting of two
parts connected by a main event called the Centre
Event (CE) and representing the occurrence of an accident resulting in a reportable consequence (here a
vapour/gas explosion). This is a very important characteristic of the model. Owing to the nature of the
available data that correspond to joint events of explosions resulting in reportable consequences, the Centre
Event refers to events that either result in a reportable
consequence or not (i.e. no explosion or an explosion
without reportable consequences). Usually all events
to the left of this event represent events aiming at preventing the CE from occurring and the corresponding
part of the diagram is called Left Hand Side (LHS).
All events to the right of the CE correspond to events
aiming at mitigating the consequences of the CE and
this part of the model is called Right Hand Side (RHS)
(Papazoglou 2007). In the core of the WORM project,
however, the events to the left are events that influence
the probability of the Centre Event occurring, the latter being an accident with reportable consequence.
The events to the right of the Centre Event, simply
condition the severity of the reportable consequence.
For communication purposes with safety engineers not

familiar with logical models this influence diagram is

called, within the WORM project, bowtie model.
The logical model provides a way for organising
various events from a root cause via the centre event,
ending up with a reportable damage to the health of
the worker. The use of such a model is twofold. On the
one hand it provides the accident sequences, that is, the
sequences of events that lead from a fundamental or
root cause to the final consequence. On the other hand,
it provides a way for quantifying the risk (Papazoglou
2007).
The structure of the paper is as follows: Sections 2
and 3 illustrate the general concept of the logical model
for vapour/gas chemical explosions along with specific details. Section 4 presents the quantification of
the Bowtie. The results and the ranking of the various
working conditions and/or safety measures in terms of
their contribution to the risk are presented in section 5.
Finally section 6 concludes the paper.
2

LOGICAL MODEL FOR VAPOUR/GAS

CHEMICAL EXPLOSIONS

Occupational explosion accidents are accidents where

the injuries are the result of the effects of an explosion. For the purpose of modelling explosions are
characterised by pressure (wave) effects and sometimes the launching of fragments. A distinction has
been made between physical explosions and chemical
explosions. Physical explosions are explosions which
are caused by an over-pressurisation of containment
for any reason other than a chemical explosion inside
the containment. Chemical explosions are explosions
caused by a) vapour or gas mixtures; b) dust; c)
the ignition of (solid) explosives and d) explosive
reactions (explosive run-away reactions, auto-ignition
reactions, combustion reactions).
Four different models have been developed to
include those four types of explosions.
This paper presents the model and its quantification
for vapour/gas chemical explosions.
2.1

Left hand side of the model (LHS)

The left hand side of the model consists of an initiating

event and corresponding safety measures (technical or procedural) aiming at preventing a vapour/gas
explosion with reportable consequences.
2.1.1 Initiating events
The initiating event represents activities where workers are adding substances to a containment (filling,
feeding, pressurising); venting, discharging, releasing, emptying of a containment/substance; opening a
containment (e.g. a valve of an oxygen gas cylinder);

778

closing a containment; working with chemicals

(performing experiments, surface treating, putting
objects in galvanic baths); working with explosives;
using an ignition source (heating, hot work activities,
switching on equipment); manual moving a containment; cleaning a containment; disconnecting a battery
and fighting a fire.
2.1.2 Primary and support safety barriers
A safety barrier is a physical entity, a technical,
hardware, procedural or organisational element in
the working environment that aims either at preventing something from happening (e.g. the CE) or at
mitigating the consequences of something that has
happened. Safety Barriers can be distinguished in
Primary and Support Barriers. A Primary Safety
Barrier (PSB) either alone or in combination with
other PSBs may prevent the initiation of an explosion. A Support Safety Barrier (SSB) sustains the
adequate function of the PSB and influence the probability with which the primary safety barrier-states
occur.
2.2

Right hand side (RHS)

The right hand side of the chemical explosions model

in combination with the outcome of the centre event
determines the consequences of the chemical explosions. Four levels of consequences are used: C1: No
consequence; C2: Recoverable injury; C3: Permanent
injury; C4: Death. The quantification of occupational risk for chemical explosions will be presented
in form of probabilities for the three levels of possible
consequence severity.

SAFETY BARRIERS FOR VAPOUR/GAS

EXPLOSIONS

A vapour/gas explosion occurs when an explosive

mixture is formed and this mixture comes into contact with an ignition source. Consequently the general
objective of the safety functions in relevant situations
is to avoid the simultaneous occurrence of ignition
sources and explosive mixtures. This can be done
by preventing the formation of an explosive mixture
and if this is not possible either by keeping the explosive mixture isolated from the space where ignition
sources exist (or it is likely to exist), or by keeping ignition sources isolated from spaces where an explosive
mixture exists.
Based on the physicochemical characteristics of
explosive mixtures and the process industry experience along with information from the accidents
occurred in the Netherlands, four different situations
where an explosive mixture can come into contact with

an ignition source and pose an explosion hazard for the

operator are distinguished:
1. Working (at locations) with systems with enclosed
flammable substances.
2. Working (at locations) where ventilation is the
suitable measure for preventing the creation of
explosive vapours/gases.
3. Working (at locations) in which explosive atmospheres are normally present.
4. Working (at locations) with flammable substances
which can vaporize resulting in explosive vapour
mixtures.
Chemical explosions from vapour/gas mixtures
were modelled according to the model shown in
Figure 1. First block in the model is the Mission
split. This block splits the initial mission to four
mutually exclusive working environments each with
the potential of one of the four types of explosion. The
mission split values are {48%, 25%, 15%, and 12%}
for the four explosion types respectively. The meaning
of these values is twofold: either they express the percentage of time a worker spends in activities related
with each explosion type (for single worker assessment) or in a multi-worker assessment they express
the percentage of workers working exclusively in the
environment related to each explosion type.
Safety barriers to prevent the four different types of
a vapour/gas explosion are presented in the following
sections.
3.1

Prevention of uncontrolled substance release

This barrier (PSB1 in figure 1) models explosions taking place due to uncontrolled flammable substance
release and the introduction or existence of ignition
sources in the same space. This barrier belongs to
type 1 explosion safety barriers and has one success
state and three failure states:
State 1: Success state corresponding to no explosion
since substance release has been prevented (no release
of flammable substance).
State 2: Failure state that models the release of flammable substance and subsequent explosion given
that an ignition source will be introduced by a human
activity.
State 3: Failure state that models the release of flammable substance and subsequent explosion due to
ignition source introduced by an equipment malfunction. This state models the joint event of flammable
substance release and the introduction of ignition source because of equipment malfunction (e.g. sparks,
shorts).
State 4: Failure state that models the release of flammable substance and subsequent explosion due to
failure to separate the released flammable vapour from

779

Figure 1. Logical model for vapour or gas explosions.

(normally) existing ignition sources. This state models the joint event of flammable substance release and
the failure to isolate this vapour from existing ignition
sources.
When this barrier is in any of the three failure states
an explosion of type 1 may occur.
3.2

Prevention of explosion of flammable

atmosphere in closed space

This barrier (PSB2 in figure 1) models explosions taking place in closed spaces where flammable vapours
are produced and are supposed to be removed by a
ventilation system. Absence or failure of the ventilation system allows the built-up of the explosive vapour
and the explosion takes place either because of the
erroneous introduction of ignition source (by human
activity or equipment malfunction) or failure to detect
the presence of the vapour and the separation from normally present ignition sources. This barrier prevents
type 2 explosions and has six states (1 success state
and 5 failure states) as follows:
State 1: Success state resulting in no explosion since
no ignition sources have been introduced or separation
barriers are in full function.
State 2: Explosion takes place given that flammable
vapours exist and an ignition source is introduced by
human activity.
State 3: Explosion takes place given that flammable
vapours exist and an ignition source is introduced
owing to equipment failure. This involves equipment
which fails and forms an ignition source (e.g. not

possible to turn it off, electrical defect, missing insulation) or which is the wrong type of equipment (use
of non-explosion proof equipment).
State 4: Explosion takes place given normally present ignition sources. Flammable vapours are generated and remained undetected because no provisions or possibilities for the indication/detection
of the presence of explosive mixtures have been
taken.
State 5: Explosion takes place given normally present ignition sources. Indication/detection provisions
are present but failed or diagnosis/response has failed
so flammable vapours have been generated.
State 6: Explosion takes place given normally present ignition sources. Flammable vapours are introduced where the ignition sources are from other areas
through the removal of barriers.
When this barrier is in any of the states 25 above
an explosion of type 2 may occur if flammable vapours
are created.
3.3

Prevention of explosion of flammable

atmosphere in or near a system
that normally produces such atmosphere

This barrier (PSB4 in Figure 1) models explosions

where the explosive atmosphere is always present due
to badly designed process. Explosion occurs where
an ignition source is introduced. No separation of the
flammable atmosphere from the ignition source is possible in this case. This barrier models explosions of
type 3 and has one success and two failure states:

780

State 1: Success state resulting in no explosion because process is designed in a way that no explosive
atmosphere is generated.
State 2: Failure state corresponding to generation of ;explosive atmosphere and subsequent explosion given that an ignition source will be introduced
by a human activity.
State 3: Failure state corresponding to generation of
explosive atmosphere and subsequent explosion due
to an ignition source introduced by an equipment
malfunction.
When this barrier is in any of the two failure states
an explosion of type 3 may occur.

3.4

Prevention of explosion of a flammable

atmosphere created by evaporation
of flammable material

This barrier (PSB5 in Figure 1) models explosions

at locations with flammable substances which can
vaporize suddenly resulting in explosive vapour mixturesexplosion type 4. This barrier has one success
and three failure states.
State 1: Success state resulting in no explosion since
no explosive atmosphere is generated.
State 2: Failure state that models the generation of
explosive atmosphere and a subsequent explosion
given that an ignition source will be introduced by
a human activity.
State 3: Failure state that models the generation of
explosive atmosphere and a subsequent explosion due
to an ignition source introduced by an equipment
malfunction.
State 4: Failure state that models the generation of
explosive atmosphere and a subsequent explosion due
to failure to separate the explosive atmosphere from
existing (normally) ignition sources.
When this barrier is in any of the three failure states
an explosion of type 5 may occur.

3.5

atmosphere barriers in state 2 (e.g. PSB6 with PSB1,

PSB8 with PSB4 and PSB9 with PSB5 in Figure 1).
Introduction of ignition sources due to equipment
malfunction is included in the states of the safety
barriers (see sections 3.13.4 above).
3.6

Adequate ventilation systems ensure that explosive

atmosphere will not be created in confined spaces.
This support barrier has two states: Exists and
Absent. Exists means that ventilation system exists
in the working place but an explosion still may occur
because the existing ventilation system either fails or is
inadequate and an explosive atmosphere has been created. Absent means that no ventilation system exists
so an explosive atmosphere may be created.
This barrier is influencing barrier Prevention of
explosion of flammable atmosphere in closed space
as it is shown in Figure 1. Existence of a ventilation
system implies a lower probability of chemical explosion due to prevention failure in closed spaces than
when a ventilation system is absent.
3.7

These barriers (PSB6, PSB8, PSB9 in Figure 1) represent the introduction of an ignition source through
various human activities (e.g. hot works, human errors,
smoking). Three barriers have been introduced for the
three types of explosions that are due to human ignition
causes (type 1, type 3 and type 4 explosions). The barriers have two states namely success and failure
which correspond to the successful and failed respectively prevention of introduction of ignition sources by
human activities.
Failure of such a barrier can result in an explosion if combined with the corresponding flammable

Personal protective equipment (PPE),

protection other than PPE and emergency
response

Personal Protective equipment (e.g. helmets and safety

glasses to protect from fragments and ear muffs
to avoid drum rupture), Protective barriers (such
as explosion proof areas or doors) and Emergency
response and prompt medical attention in the workplace may mitigate the effects of an explosion. Those
barriers have two states: Exists and Absent meaning
that either the barrier exists or does not exist (absent) or
not used in the working place. PPE, Other protection than PPE and Emergency response influence
all prevention barriers (see also quantification issues
in section 4).
3.8

Barriers preventing the introduction

of an ignition source through human activities

Ventilation systems

Age of operator

Age of the person that was injured by the explosion

may have an influence on the consequences of the
explosion on the victim. This event has two states:
Age of operator <= 50 years old and Age of operator > 50 years old. Age of operator influences all
prevention barriers.
3.9

Probability influencing entities (PIEs)

In several instances the safety barriers of the model

are simple enough to link directly to easily understood
working conditions and measures as in the barrier
Ventilation. Assessing the frequency with which

781

ventilation systems exist in the working environment is

straightforward since either those systems exist or not.
In other instances, however, this is not possible.
For example, the support barrier Other protection
than PPE may be analysed into more detailed and
more concrete measures that affect its quality. Such
specific measures are: (i) explosion suppression systems (explosion proof areas such as control rooms or
explosion resistant doors); (ii) the distance between
the explosion point and the location of the worker.
Similarly the barrier Emergency response may be
analysed into the following measures: i) supervision
or monitoring at the workplace; ii) emergency team
which is always present or standby; iii) first aid and
decompression facilities; iv) multiple rescue possibilities (direct access); and v) professional medical
assistance. Such factors have the name of Probability Influencing Entity (PIEs). Each influencing
factor (PIE) is assumed to have two possible levels, Adequate and Inadequate. The quality of an
influencing factor is then set equal to the frequency
with which this factor is at the adequate level in the
working places. Then the quality of the barrier is given
by a weighted sum of the influencing factor qualities.
The weights reflect the relative importance of each
factor and are assessed by the analyst on the basis of
expert judgement. Currently equal weights have been
used. This way the probability of a barrier to be in one
of its possible states is given by the weighted sum of the
frequencies of the influencing factors (RIVM 2008).

PIES and their values as well as the failure probability for the barrier they influence for the logic model
of vapour/gas chemical explosions are presented in
Table 1.

QUANTIFICATION PROCESS

In general the level of resolution of a logical model

used in ORM was driven by the available data. A
logic model provides a collection of event outcomes
or barrier states which may lead to an accident when
they coexist in particular states. These accidents have
specific consequences. The general form of such a
sequence is:
C = {S1 , S2 , .....Sn , B1 , B2 , . . .Bm }

Analysis of available accident data allowed the

assessment of the number of times such accident
sequences occurred during a given period of time. Surveys of the Dutch working population assessed the
exposure of the workers to the specific hazards over
the same period of time. Consequently it was possible
to assess the probability P(C) of the various accident
sequences. Surveys of the Dutch working places and of
the corresponding conditions allowed the assessment
of the overall probability of some individual barriers
(e.g. see Table 1). If such assessment is made then

Table 1.

PIES characteristics and values.

Barrier

PIE

PIE characteristics

PSB6-

PSB8-

PSB9SSB1-

1
1

SSB2-

SSB3-

SSB4-

Prevention of introduction of an ignition

source (human activity - type 1 explosion)
Prevention of introduction of an ignition
source (human activity - type 3 explosion)
Prevention of introduction of an ignition
source (human activity - type 4 explosion)
Ventilation systems
Personal Protective
Equipment
Explosion suppression systems
control room or doors
Distance to explosion
Supervision/ monitoring at
the location
Emergency team present or
standby
First aid/decompression
facilities
Rescue possibilities (access)
Professional medical
assistance
Population share exposed
>=50 years

SSB5-

(1)

782

PIE
value

Barrier
failure

0.17

0.17

0.08

0.08

0.06

0.06

0.06
0.3

0.06
0.3

0.26

0.25

0.24
0.2

0.23

0.23
0.24
0.24
0.24
0.17

0.17

probabilities of the form P(S1 , S2 , . . ., B1 , . . .Bi , . . .)

can be estimated where (S1 , S2 , . . ., B1 , . . ., Bi , . . .)
are the barriers that can be quantified independently
of the accident data. Then equation (1) can be
written as:
P(C) = P(S1 , . . ., Bi ) P(Bi+1 , . . ., Bj/S1 , . . ., Bi ) (2)
In the above equation P(C), P(S1 , . . . , Bi ) are
known hence P(Bi+1 , . . ., Bj /S1 , . . ., Bi ) can be calculated. The overall objective of ORM was then to be
able to calculate the new value of P(C) given a number of safety measures that have a specific effect and
consequently change P(S1 , . . . , Bi ) to P(S1 , . . . , Bi ).
As a result events that could not be independently
quantified had to be grouped together as (Bi+1 , . . .Bj )
since only the probability of the joint event could be
quantified. In particular within the group of events
(Bi+1 , . . ., Bj ) there were always the events of an accident with specific consequences. For example the
probability of the joint event of P(Failure to detect
explosive atmosphere, explosion, fatality) could be
assessed and hence P(Failure to detect explosives
atmospheres/explosions) could be assessed either but
not the independent probability P(Failure to detect
explosive atmospheres). Based on these constraints the
model presented in section 3 has been developed and
quantified.

5.1 Importance analysis

To assess the relative importance of each factor influencing the risk from vapour/gas chemical explosions
two importance measures have been calculated.
1. Risk decrease: This measure gives the relative
decrease of the risk if the barrier (or PIE) achieves
its perfect state with probability equal to unity.
2. Risk increase: This measure gives the relative
increase of the risk if the barrier (or PIE) achieves
its failed state with probability equal to unity.

Risk decrease prioritizes the various elements of the

model for the purposes of possible improvements. It
is more riskeffective to try to improve first a barrier
with higher risk decrease percentage than another with
lower risk decrease.
Risk increase provides a measure of the importance
of each element in the model to be maintained at its
present level of quality. It is more important to concentrate on the maintenance of a barrier with high risk
increase importance than one with a lesser one.
The effect each barrier has on the risk rate per type
of explosion and on the overall risk is presented in
Tables 2 and 3.
From the risk decrease shown in Table 2 it follows
that Personal Protective Equipment and Other protection than PPE are the barriers with the most
significant effect on the overall risk reduction (higher
values of risk decrease) if made perfect, that is if they
are used 100% of the time. On the other hand, the
results in Table 3 indicate that the barriers with the
highest levels of risk increase are the Ventilation
5 RESULTS
Systems and the Protection other than PPE This
means that these two are the most important barriers
Quantification of the model for vapour gas explosions
to maintain because if failed they cause the highest
given a total exposure of 1.62x109 hours resulted in
risk increase. It should be emphasized that the results
the following probabilities:
shown in Tables 2 and 3 and the corresponding concluProbability of Explosion with reportable
sions are valid for the particular case modelled. This
Consequence (hr 1 ) = 7.78 108
is a worker or a group of workers that are exposed
Probability of Lethal Injury (hr 1 ) = 9.26 109
to the hazards of the four types of vapour/gas exploProbability of Permanent Injury (hr 1 ) = 2.38108
Probability of Recoverable Injury (hr 1 ) = 4.47 108 . sions with the quantitative mission split discussed in
Table 2.

Risk rates per type of explosion and Overall risk decrease for each safety barrier.

Safety barrier
Base case
Prevention of ignition source (type 1 explosion)
Prevention of ignition source (type 3 explosion)
Prevention of ignition source (type 4 explosion)
Ventilation systems
Personal Protective Equipment
Other protection than PPE

Type 1
explosion

Type 2
explosion

Type 3
explosion

Type 4
explosion

Overall
risk rate

Risk
decrease

2.65E-08
2.22E-08

1.79E-08

1.05E-08

2.28E-08

7.78E-08
7.35E-08
7.24E-08
7.65E-08
6.60E-08
4.69E-08
5.02E-08

5.53%
6.94%
1.58%
15.08%
39.66%
35.49%

5.13E-09
2.16E-08
1.04E-08
1.60E-08

6.17E-09
1.28E-08
1.19E-08

783

7.50E-09
7.00E-09

1.63E-08
1.52E-08

Table 3.

Risk rates per type of explosion and Overall risk increase for each safety barrier.

Safety barrier
Base case
Prevention of ignition source (type 1 explosion)
Prevention of ignition source (type 3 explosion)
Prevention of ignition source (type 4 explosion)
Ventilation systems
Personal Protective Equipment
Other protection than PPE

Table 4.

Type 2
explosion

Type 3
explosion

Type 4
explosion

Overall
risk rate

Risk
increase

2.65E-08
4.76E-08

1.79E-08

1.05E-08

2.28E-08

7.78E-08
9.89E-08
1.39E-07
9.72E-08
2.61E-07
1.50E-07
1.60E-07

27.19%
79.35%
25.00%
235.91%
92.52%
106.15%

7.22E-08
4.22E-08
6.43E-08
5.80E-08

2.02E-07
2.98E-08
3.58E-08

1.75E-08
2.10E-08

3.81E-08
4.57E-08

Risk decrease for the three levels of consequences.

Base case
Prevention of ignition source (type 1 explosion)
Prevention of ignition source (type 3 explosion)
Prevention of ignition source (type 4 explosion)
Ventilation systems
Personal protective equipment
Other protection than PPE
Emergency response

Table 5.

Type 1
explosion

Recoverable
injuries

Risk
decrease
(%)

Permanent
injuries

Risk
decrease
(%)

Lethal
injuries

Risk
decrease
(%)

4.47E-08
4.14E-08
4.19E-08
4.41E-08
3.90E-08
2.53E-08
2.85E-08
3.73E-08

7.38
6.26
1.34
12.75
43.40
36.24
16.55

2.38E-08
2.31E-08
2.12E-08
2.33E-08
2.10E-08
1.56E-08
1.56E-08
1.82E-08

2.94
10.92
2.10
11.76
34.45
34.45
23.53

9.26E-09
8.96E-09
9.26E-09
9.13E-09
6.03E-09
6.02E-09
6.06E-09
8.64E-09

3.24
0.001
1.40
34.88
34.99
34.56
6.70

Recoverable
injuries

Risk
increase
(%)

Permanent
injuries

Risk
increase
(%)

Lethal
injuries

Risk
increase
(%)

4.47E-08
6.09E-08
7.61E-08
5.34E-08
1.33E-07
8.99E-08
9.31E-08
6.95E-08

36.24
70.25
19.46
197.54
101.12
108.28
55.48

2.38E-08
2.73E-08
5.41E-08
3.24E-08
6.82E-08
4.30E-08
4.85E-08
4.27E-08

14.71
127.31
36.13
186.55
80.67
103.78
79.41

9.26E-09
1.07E-08
9.26E-09
1.14E-08
6.00E-08
1.68E-08
1.87E-08
1.14E-08

15.55
0.01
23.11
547.95
81.43
101.94
23.11

Risk increase for the three levels of consequences.

Base case
Prevention of ignition source (type 1 explosion)
Prevention of ignition source (type 3 explosion)
Prevention of ignition source (type 4 explosion)
Ventilation systems
Personal protective equipment
Other protection than PPE
Emergency response

section 3. For this specific case the most effective measure for reducing risk is the increase in the percentage
of time that PPEs are used.
This conclusion might change if a different mission
split is considered or if an explosion type is considered
in isolation. For explosion type 2, for example, presence of Ventilation systems at 100% of the cases
decreases the risk rate by 65.5%, whereas the PPE
and Other protection by 28.5% and 33.5% respectively. Thus the risk of a type 2 explosion is reduced
more by the increasefrom the present level- of the
use of ventilation systems rather than the PPE or other
protective measures.

Risk importance measures can be calculated not

only with respect to the probability of an explosion
with reportable consequences but with respect to specific type of consequence, as shown in Tables 4 and 5.
As it can be seen from Table 4 (last column) Ventilation Systems Personal Protective Equipment and
Other protection measures are practically equally
important in terms of their improvement with respect
to fatalities. This is due to the fact that the conditional
probability of a fatal injury given a type 2 explosion
is much higher than for the other types of explosions.
Thus even if type 2 explosions participate only by 25%
in the mission split the increase in the presence of ven-

784

tilation systems is very important for decreasing the

risk of fatality.
As for the risk increase results of Table 5: In all
consequences levels Ventilation systems have the
greater risk increase percentages, while Other protection than PPE is second in ranking for recoverable
and lethal injuries, and Prevention of introduction
of ignition sources for type 3 explosion is second
for permanent injuries. Third barrier in risk increase
ranking is PPE for lethal and recoverable injuries
and Other protection than PPE for the permanent
ones.
In this way all safety barriers can be ranked according to the effect they induce in the overall risk as well as
in the risk of lethal, permanent or recoverable injuries.
6

CONCLUSIONS

A logical model has been presented for quantifying the

probability of vapour/gas chemical explosions and the
various types of consequences following these types
of accidents. The model includes primary and support
safety barriers aiming at preventing chemical explosions. For the quantification of the model the exposure
rates (total time spent in an activity involving each hazard per hour) have been used which was estimated with
user (operators) surveys and real accident data coming from the reported accident database GISAI. The
probability of the consequences of such accidents is
presented in three levels: fatalities, permanent injury
and non-permanent injury. Surveys also provided data
for the working places and the corresponding conditions allowing in this way the assessment of the overall
probability of some individual barriers. The model
has been used for risk reducing measures prioritization through the calculation of two risk importance
measures: the risk decrease and the risk increase. The
calculations were made for the overall risk and the
risk in three levels of consequence severity. Personal
Protective Equipment and Ventilation Systems are
the barriers with the most important risk values in the
overall risk ranking analysis.

REFERENCES
Ale B.J.M., Baksteen H., Bellamy L.J., Bloemhof A.,
Goossens L., Hale A.R., Mud M.L., Oh J.I.H., Papazoglou
I.A., Post J., and Whiston J.Y., 2008. Quantifying occupational risk: The development of an occupational risk
model. Safety Science, Volume 46, Issue 2: 176185.
Aneziris O.N. Papazoglou I.A., Baksteen H., Mud M.L., Ale
B.J.M, Bellamy L.J., Hale A.R., Bloemhoff A., Post J.,
Oh J.I.H., 2008a. Quantified risk assessment for fall from
height. Safety Science, Volume 46, Issue 2: 198220.
Aneziris O.N., Papazoglou I.A., Mud M.L., Damen M.,
Kuiper J., Baksteen H., Ale B.J.M., Bellamy L.J.,
Hale A.R., Bloemhoff A., Post J.G., Oh J.I.H, 2008b.
Towards risk assessment for crane activities Safety Science. doi:10.1016/j.ssci.2007.11.012
Baksteen, H., Samwe, M., Mud, M., Bellamy, L., Papazoglou, I.A., Aneziris, O., Konstandinidou, M., 2008
ScenarioBowtie modeling BT 27 explosions, WORM
Metamorphosis Report.
Bellamy L.J., Ale B.J.M., Geyer T.A.W., Goossens L.H.J.,
Hale A.R., Oh J.I.H., Mud M.L., Bloemhoff A, Papazoglou I.A., Whiston J.Y., 2007. StorybuilderA tool for
the analysis of accident reports, Reliability Engineering
and System Safety 92: 735744.
GISAI, 2005. Geintegreerd Informatie Systeem Arbeids
Inspectie: Integrated Information System of the Labor
Inspection in the Netherlands.
OSHA 2008. Bureau of Labor statistics http://data.bls.gov/
GQT/servlet/InitialPage.
Papazoglou I.A., Ale B.J.M., 2007. A logical model for quantification of occupational risk, Reliability Engineering &
System Safety 92 (6): 785803.
Papazoglou I.A, L.J. Bellamy, K.C.M. Leidelmeijerc, M.
Damenc, A. Bloemhoffd, J. Kuiperd, BJ.M. Alea, J.I.H.
Oh, Quantification of Occupational Risk from Accidents, submitted in PSAM 9.
RIVM 2008 WORM Metamorphosis Consortium. The Quantification of Occupational Risk. The development of
a risk assessment model and software. RIVM Report
620801001/2007 The Hague.

785

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Occupational risk of an aluminium industry

O.N. Aneziris & I.A. Papazoglou
National Center DEMOKRITOS, Athens, Greece

O. Doudakmani
Center for Prevention of Occupational Risk, Hellenic Ministry of Employment and Social Affairs, Thessaloniki, Greece

ABSTRACT: This paper presents the quantification of occupational risk in an aluminium plant producing
profiles, located in Northern Greece. Risk assessment is based on the Workgroup Occupational Risk Model
(WORM) project, developed in the Netherlands. This model can assess occupational risk at hazard level, activity
level, job level and overall company risk. Twenty six job positions have been identified for this plant, such as
operators of press extruders, forklift operators, crane operators, painters, and various other workers across the
process units. All risk profiles of workers have been quantified and jobs have been ranked according to their
risk. Occupational risk has also been assessed for all plant units and the overall company.

INTRODUCTION

Occupational safety and health is a major concern to

many countries and the traditional way to deal with it
is legislation, regulation, standards and safety guidelines. The Ministry of Social Affairs and Employment
in the Netherlands developed Workgroup Occupational Risk Model (WORM) project, a large scale
project during 2003-2007 to improve the level of safety
at workplace, by introducing quantitative occupational
risk. WORM is presented in (RIVM 2008) and its main
achievement is the development of the occupational
risk model, which is built on the detailed analysis of
9000 accident reports in the Netherlands.
Occupational risk is performed for an aluminium
plant, located in Greece. Aluminium industry is the
most dynamic branch of the non ferrous metal working
industry in Europe with significant activity in Greece.
Data regarding workers jobs, activities and hazards
were available, as well as general accident data of
Greek aluminium industries.
This paper is organized as follows. After the introduction of section 1, section 2 presents briefly the
methodology of occupational risk and section 3 the
description of the aluminium plant. Section 4 presents
the job positions of all workers and section 5 occupational risk quantification and results.
2

developed. According to this model occupational risk

in a company is calculated by assessing the hazards
the workers in this company are exposed the duration
of the exposure and the integration of the risk to all
hazards and all workers.
A tree-like structure is used to develop the composite model of ORM as depicted in Figure 1.
The Top Level of the tree corresponds to the entity
under analysis.
The second level provides the type of Companyposition corresponding to a specific type of job along
with the number of people in each position type. There
are n = 1, 2, . . . , N company positions each occupied
by T1 , . . ., Tn employees, respectively.

OCCUPATIONAL RISK

In the framework of the WORM project a model

for the quantifications of occupational risk has been

Figure 1.

787

Composite ORM model structure.

The third level of the tree describes for each

position-type the activities required to perform the corresponding job along with the respective frequencies.
This means that a particular job is described in terms
of a number of activities each one of which is performed a specific number of times over a given period.
Thus the nth job position is characterized by Mi activities A(n, 1), . . ., A(n, m), .., A(n, Mn ) each performed
with annual frequency f(n,m).
Finally, performance of a specific activity is associated with a number of single hazards (out of
the sixty three single hazards) and a corresponding duration of exposure to each and every hazard. Thus activity A(n,m) is associated with hazards
h(n, m, 1), h(n, m, 2), . . ., h(n, m, K nm ). Risk is calculated as a combination of the contributions of Jobs,
activities and Bowties.
2.1

Calculation at the hazard Level

WORM has assessed the risk per hour of exposure for

63 hazards on the basis of the characteristics of the
average Dutch worker. These characteristics express
the working conditions and are quantified as percentage of the time that the worker is working under
specified levels or types of those conditions. Case-specific analyses can be made by adjusting these
characteristics to the specific conditions. These calculations provide the risk as the probability of one of
three possible consequences (recoverable injuries, permanent injuries, and fatalities) for the duration of the
exposure, as presented by Papazoglou et al., (2009)
and Ale et al., (2008).
2.2

Calculation at activity level

Next the risk at the activity level is calculated. A general assumption is that if during one of the actions an
accident occurs resulting in a consequence (recoverable injury, permanent injury or death) the Activity is
interrupted and the overall consequence of the activity
is the same. That is no more exposure to the same or
additional hazards is possible.
Let
n = 1, 2, . . ., Nm is an index over all the dangers of
the mth activity.
p1,k : Probability of No-Consequence in the kth
hazard
p2,k : Probability of recoverable Injury in the kth
hazard
p3,k : Probability of permanent injury in the kth
hazard
p4,k : Probability of death in the kth hazard
p1,m : Probability of No consequence for activity m
p2,m : Probability of recoverable injury for activity m
p3,m : Probability of permanent injury for activity m
p4,m : Probability of death for activity m

Then
P1,m =

Nm


p1,n

n=1

P2,m =

Nm


p2,n

n=1

P3,m =

Nm


Nm

n=1

p1,r

r=1

p3,n

n=1

P4,m =

n1


n1


p1,r

r=1

p4,n

n1


p1,r

(1)

r=1

Where it has been assumed that any consequence

(recoverable injury, permanent injury, death) happening in the kth hazard interrupts the activity and hence
successful completion of the Activity requires successful (no consequences) completion of the preceding
(k-1) dangers. Any other consequence interrupts the
stream of dangers and the activity results in this
consequence.
2.3 Calculation at the job level
A worker job in a given period of time undertakes a
number of activities, where each activity consists of a
number of dangers (Bowties).
There are M activities Am m = 1, 2, 3, . . ., M . Each
activity consists of Nm Dangers, dn n = 1, 2, . . ., Nm.
Each activity is repeated fm times a year m =
1, 2, . . ., M (frequencies). Then the risk for the period
of interest (year) will be calculated as follows:
For each activity (m) we calculate risk per activity
as in section 2.2
Assumption: recoverable injury, permanent injury &
fatality interrupt the activity and no additional exposure is possible.
For each Activity (m) given the annual frequency
fm we calculate the annual risk per activity
Assumption: recoverable injury during the f th
undertaking of the activity does not preclude undertaking of the same activity again for the f + 1, .. up to
the fm time.
aP1,m = (P1,m) )fm
aP2,m = 1 aP1,m aP3,m aP4,m
aP3,m =

1 (1 P3,m P4,m )fm

P3,m
P3,m + P4,m

aP4,m =

1 (1 P3,m P4,m )fm

P4,m
P3,m + P4,m

788

(2)

where: aP1,m, aP2,m, aP3,m, aP4,m annual probability of

no consequence, recoverable injury, permanent injury
and of death of activity m.
For the M activities we calculate the total annual
risk as follows:
R1 =

M


aP1,m

m=1

R2 = 1 R1 R3 R4
R3 =

M


aP3,m

m1


(aP1,m + aP2,m )

m=1

R4 =

M


r=1

aP4,m

m1


m=1

(aP1,m + aP2,m )

(3)

r=1

where: R1, R2, R3, R4 total annual probability of no consequence, recoverable injury, permanent injury and
death.
Again the assumption is made that recoverable
injury during activity (m) does not preclude undertaking of the remaining activities during the year.

the plant consists of six major units, which are the

following: extrusion, surface treatment, die, storage,
packaging and mechanical.
a) Extrusion unit: Aluminium billets are the starting
stock for the extrusion process. They are introduced
into furnaces and heated up to 450 C, cut at the
required length and inserted in a press, where extrusion takes place. Pressure is exerted on the billet
which is crushed against a die. The newly formed
extrusion is supported on a conveyor as it leaves
the press. Depending on the alloy, the extrusion is
cooled after emerging from the die, so as to obtain
sufficient metallurgical properties. The profile is
then stretched, since it is placed in a traction to
obtain the exact geometrical shape. It is then conveyed to a saw, where it is cut in order to obtain the
required commercial length, and transported with
cranes to the ageing treatment. This is performed
in furnaces where aluminium profiles are heated at
185 C for 6-7 hours. Figure 2 presents the block
diagram of the plant.
b) Surface treatment consists of the cleaning, anodizing and coating sections. In the cleaning section
BILLET

2.4 Overall risk

STORAGE

Given a company with N jobs and Tn workers performing the nth job, the overall risk is approximated by the
expected number of workers to suffer each of the three
sequences.
R2,o =

FURNACE 450

BILLET CUT-

R2 T2

R3,o =

R3 T3

STRETCHING

R4,o =

R4 T4

(4)

PROFILE

AGEING FURNACE

where: R2,0 , R3,0 , R4,0 overall risk of recoverable injury,

permanent injury and death.
3

DIES

EXTRUDING

SURFACE TREATMENT

PLANT DESCRIPTION

The aluminium plant produces profiles for various

applications in the industry and building construction. The heart of this industry is the extrusion press
section, where raw material arriving in the form of billets, are transformed to aluminium profiles. Next they
are transferred to the surface treatment section, so as
to acquire the required aesthetical and anti corrosion
properties. Four additional sections are required to
support profile production, which are the die section,
storage, packaging and mechanical support. Therefore

ANODIZING

PACKAGING

PROFILE STORAGE
.

Figure 2.

789

Plant block diagram.

PAINTING

TOOLS

c)

d)

e)

f)

profiles are immersed in a bath of NAOH where

they are cleaned from oils and treated for painting.
In anodizing the metal is covered with a layer of
protective oxide which adds anti corrosion properties, by an electrolytic process. Aluminium profiles
are transported to the painting unit where they are
hanged either horizontally or vertically and painted.
Painting with powder in either horizontal or vertical units provides good aesthetic result of the final
product.
Die section: In this section dies, which are
required in the extrusion press unit for the production of profiles, are treated either mechanically
or chemically, in order to remove traces of metal
which have remained on them. Clean and well
maintained dies influence the quality of the final
product.
Packaging unit: after surface treatment the profiles
are moved into other areas where they are packed
and prepared for transportation to customers. The
plant is equipped with an automatic packaging unit
and profiles are palletized so as to be protected from
surface damage and twisting.
Storage areas for aluminium billets and profiles.
Aluminium billets, which is the starting material of
the plant, are stored in an area close to the extrusion
process and transported by forklifts. Packed aluminium profiles are transported either by forklifts
or by cranes to the storage area.
The machinery section consists of various operations supporting the production of profiles.
COMPANY POSITIONS

This plant has seventy seven workers distributed along

the six units, described in the previous section. There
are twenty six different types of jobs, such as operator

of press extruder, forklift operator, painter etc., which

are described in this section, along with the associated
activities and hazards.

4.1 Extrusion
There are four job positions in this unit: extruder
press operator, extruder worker, stretching and cutting
operators.
a) Press extruder operator: He is responsible for
the cut of the billets to the required length, their loading and unloading on the press extruder, the operation
of the press and completing the required documents.
Therefore his job is decomposed into four activities
(cut of the billets, loading/unloading billets on press,
press operation and completing documents), as presented in Figure 3 and Table 1. While cutting billets,
which occurs every day for two hours, he is exposed
to the following hazards: fall on same level (for 2
hours), contact with falling object from crane (for 0.2
hours), contact with hanging or swinging objects (for
0.2 hours), contact with moving parts of machine (for
0.5 hours), trapped between objects (for 0.6 hours) and
contact with hot surface (for 2 hours). Figure 2 presents
the decomposition of this job into activities and their
associated hazards, while Table 1 presents also the frequency and duration of the activities and the exposure
to hazards. Similar Tables are provided for all jobs
described in this section and presented by (Doudakmani 2007). There are 6 press extruder operators in
this plant working on an eight hour shift basis.
b) extruder worker: His activities are to heat the
dies, install them in the press extruder, and transport
them either by crane or by trolley to the die section.
He is exposed to the same hazards as the extruder
press operator, but with different durations. There are 6
extruder workers working on an eight hour shift basis.

COMPANY

FORKLIFT OPERATOR

LOADING/UNLOADING

PRESS EXTRUDER OPERATOR

CUT OF

PAINTER

PRESS

FALL ON THE SAME LEVEL

CONTACT WITH FALLING

CONTACT WITH HANGING

CONTACT WITH MOVING

TRAPPED BETWEEN OBJECTS

CONTACT WITH HOT SURFACE

Figure 3.

Decomposition of company into job positions, activities and hazards.

790

DOCUMENTS

Table 1.

Activities and associated hazards of press extruder operator.

Exposure to hazard (hrs)

Activities
of press
extruder
operator
and their
frequencies
Cut of the billets
2 hrs/day
Loading/unloading
billets on press
2 hrs/ day
Completing documents 1 hr/ day
Press operation 3 hrs/ day

Trapped
between
objects

Contact
with
moving
parts of a
machine
-operating

Fall on
the
same
level

Contact
with
falling
object
-crane
or load

Contact
with
hanging/
swinging
objects

Contact
with hot
or cold
surfaces
or open
flame

0.6

0.5

.2

0.2

0.3

0.3

.2

0.2

0.1
1

c) stretching operator: He checks the profiles arriving to the stretching press, is responsible for the press
operation and the transportation of the profiles to the
cutting machines. He is exposed to the same hazards
as the extruder press operator but with different duration. There are 6 stretching operators working on an
eight hour shift basis.
d) cutting operator: He operates the saw and cuts
the profiles to the required length. He manually moves
the cut profiles and arranges them on special cases,
which are transported by crane to the ageing furnace.
He is exposed to the following hazards: fall on same
level, contact with falling object from crane, contact
with hanging or swinging objects, contact with moving
parts of machine, trapped between, contact with falling
object during manual operation. There are 6 cutting
operators working on an eight hour shift basis.
4.2

Surface treatment

There are seven different job positions in this unit:

worker at the entrance (or exit) of the anodizing unit,
crane operator, worker at anodizing unit, workers at the
horizontal or vertical painting units, painter, cleaner
and forklift operator.
a) worker at anodizing unit. His activity is to transport profiles by trolleys to this unit. He is exposed
to the following hazards: fall on the same level, hit
by falling object, trapped between object, contact hot
surface and move into object.
b) worker at the entrance (or exit) of the anodizing
unit. The worker in the entrance carries profiles from
trolleys and ties them on vertical poles, where they
will be anodized. At the exit of this unit, when anodizing has been performed, profiles are stored on pallets.
He is exposed to the following hazards: hit by falling
object, trapped between object and move into object.

There are 4 workers with the same job description in

this unit.
c) crane operator. He operates the crane that transports profiles to the anodizing baths. He is exposed
to the following hazards: fall on the same level, hit
by falling object, trapped between object and contact
with chemicals.
d) forklift operator. His job is to transport profiles from the anodizing to the painting section. He is
exposed to the following hazards: hit by falling object,
hit by rolling object and struck by moving vehicle.
e) workers at the horizontal (or vertical) painting
unit. The workers activity is to hang alloys on the
painting unit. He is exposed to the following hazards:
fall on the same level, fall from height, falling object,
flying object, trapped between object, move into object
and contact with hot surface. There are 5 workers with
the same job description in this unit.
f ) painter. He either paints the profiles himself, or
prepares the painting powder and operates the automatic machine. He is exposed to the following hazards:
fall on the same level, falling object from crane,
trapped between object, move into object and contact
with chemicals.
g) cleaner. His main activity is to clean the painting unit. He is exposed to the following hazards:
fall on the same level, hit by falling object, trapped
between objects, move into object and contact with
chemicals.
4.3 Die section
There are five job positions in this section, where dies
are treated either chemically or mechanically. These
are the following: operators installing, sandblasting,
cleaning chemically and hardening dies and also
operators using machine tools.

791

a) operator installing dies on the press excluder. His

main activity is to install the dies on the press extruder,
but he also transports them by a crane or by trolley. He
is exposed to the following hazards: fall on same level,
hit by falling object from- crane, contact by hanging or
swinging objects, contact with falling object manually
transported and trapped between objects.
b) operator sandblasting dies. Apart from his main
activity which is die sandblasting he is involved in
material handling. He is exposed to the following hazards: fall on same level, contact with falling object,
contact with hanging or swinging objects, trapped
between objects, contact with moving parts of machine
and contact with flying objects.
c) operator cleaning chemically dies. Apart from
his main activity, which is cleaning dies, he is involved
in material handling. He is exposed to the following
hazards: fall on same level, contact with falling object,
contact with hanging or swinging objects, trapped
between objects and contact with chemicals.
d) operator hardening dies. Apart from his main
activity, which is hardening dies he is involved in material handling. He is exposed to the same hazards as the
operator who cleans dies.
e) worker using machine tools. His activities ate to
use tools and to handle materials. He is exposed to the
same hazards as the operator sandblasting the die.
4.4

Packaging

There are three job positions for packaging profiles

which are the operator of the automatic packaging
machine, a worker manually packaging and a helper.
a) operator of packaging machine: His activities
are transporting profiles by trolleys, feeding the packaging machine and operating it. He is exposed to the
following hazards: hit by falling objects, contact with
moving parts, trapped between objects, move into
an object, contact handheld tool and stuck by moving vehicle. There are 2 operators with the same job
description in this unit.
b) a workermanually packages alloys and transports
them by trolleys. He is exposed to the following hazards: hit by falling objects, trapped between objects,
move into an object, contact handheld tool and stuck
by moving vehicle. There are 2 workers with the same
job description in this unit.
c) a helper cuts with a saw and packages manually.
He is exposed to the same hazards as the operator of
the packaging machine. There are 2 workers with the
same job description in this unit.
4.5

Storage areas

There are three job positions in the storage area, which

are: forklift operators transporting alloys and profiles
and a worker.

a) forklift operator transporting alloys: his activities

are unloading alloys from trucks, arranging them in the
storage area, transporting them to the process area and
emptying bins with aluminium scrap. The hazards to
which the operator is exposed are: hit by falling object,
hit by rolling object, fall from height and struck by
moving vehicle. There are 2 forklift operators with the
same job description in this unit.
b) forklift operator transporting profiles: his activities are transporting profiles, loading and unloading
trucks, piling profiles and transporting bins of scrap.
The hazards to which the operator is exposed are hit by
falling object, fall from height and struck by moving
vehicle. There are 6 forklift operators with the same
job description in this unit.
c) worker of storage area: his activities are crane
operation, arrangement, transportation and labeling of
profiles and also checking documents. The hazards to
which the operator is exposed are hit by falling object
from crane, hit by hanging or swinging object, fall
from height, fall on the same level, move on object,
trapped between objects and struck by moving vehicle.
There are 10 workers with the same job description in
this unit.
4.6

Machinery section

This section consists of a press operator, an operator

of machine tools, an insulation fitter and a carpenter.
a) press operator: his activities are press operation
and material handling. The hazards he is exposed to are
hit by falling object, move on object, trapped between
objects and contact with moving parts of machine.
b) operator of machine tools: his activities are
operation machine tools and material handling. The
hazards he is exposed to are hit by falling object, hit by
flying object, move on object, trapped between objects
and contact with moving parts of machine.
c) insulation fitter: his activities are fitting insulation and material handling. The hazards to which he
is exposed are the same as the press operator, but with
different duration of exposure.
d) carpenter: his activities are operation with saws
and tools for cutting and also packaging. The hazards to which he is exposed are hit by falling object,
trapped between objects and contact with moving parts
of machine and contact with handheld tools.
5

RESULTS AND CONCLUSIONS

Occupational risk has been calculated for all job position of the plant and is presented in Figures 4 and 5
and Table 2. Figure 4 presents present annual risk
of death and Figure 5 annual permanent and recoverable injury for all job positions in this plant. The
operator at the entrance of the painting unit has the
highest probability of death (3.25105 /yr) followed

792

3,10E-05

2,60E-05

2,10E-05

1,60E-05

1,10E-05

6,00E-06

extruder operator
extruder worker
cutting operator
stretching opertaor
forklift operator -billets
forklift operator -profiles
worker in storage
die installation
die sandblasting
die chemical
die hardening
op. die machine tools
operator entrance painting unit
painter
cleaner
forklift operator- painting
manual handling painting
crane operator-anodizing
worker -entrance anodizing
operator of press for tools
operator of machine tools
insulation fitter
helper packaging
operator of packaging machi
worker packaging
carpenter

1,00E-06

Figure 4.
(/year).

Probability of fatality of worker in the plant

2,50E-04

2,00E-04

1,50E-04

1,00E-04

0,00E+00

extruder operator
extruder worker
cutting operator
stretching opertaor
forklift operator -billets
forklift operator -profiles
worker in storage
die installation
die sandblasting
die chemical
die hardening
op. die machine tools
operator entrance painting unit
painter
cleaner
forklift operator- painting
manual handling painting
crane operator-anodizing
worker -entrance anodizing
operator of press for tools
operator of machine tools
insulation fitter
helper packaging
operator of packaging machi
worker packaging
carpenter

5,00E-05

Permanent Injury

Recoverable Injury

Figure 5. Probability of recoverable and permanent injury

for each worker (/year).

by the worker in the storage area (2.18 105 /yr)

and the worker performing sandblasting of the dies
(1.91105 /yr). The operator of at the entrance of the
painting unit has also the highest probability of recoverable injury (2.33 104 /yr) followed by the worker

in the storage area (1.92 104 /yr) and the worker

manual handling, at the painting unit (1.85104 /yr).
The helper at the packaging unit has the highest probability of permanent injury (2.22 104 /yr) followed
by the operator at the entrance of the painting unit
(2.03 104 /yr) and the worker performing sandblasting of the dies (1.85 104 /yr). The operators at
the entrance of the painting unit and the worker in the
storage area have are the most dangerous jobs owing
to the high probability of death and recoverable injury,
while the helper at the packaging unit is also regarded
as a dangerous job, owing to the high probability of
permanent injury. High fatality risk of workers at the
entrance of the painting unit, at the storage area and
the worker performing sandblasting of the dies can
be further analyzed in order to obtain the most serious hazards, these workers are exposed to. Figure 6
presents these results and it appears that all three workers are exposed to high hazard of contact with falling
object from cranes.
Risk has also been calculated for the six units
of the plant. This quantification is performed as
described in section 2.4 for each unit. The storage
area has the higher expected number of deaths (2.8
104 /year) and recoverable injury (2.76103 /year),
followed by the surface treatment (2.54 104 /year
and 1.97 103 /year respectively). The extruder area
has the higher expected number permanent injury
(2.4 103 /year), followed by the surface treatment
(1.76 103 ). It should be marked that the extruder
unit, the storage and surface treatments areas have
most of workers in the plant, which are 24, 18 and
14 respectively.
The overall annual company risk is equal to
8.44 104 /year for fatality risk, 8.66 103 /year
for permanent injury and 8.18 103 /year for recoverable injury.
Accident data for aluminium plants in Northern
Greece were available, as presented by (Doudakmani
2007). Totally 96 accidents have occurred during the
period 20012005, 2 permanent injuries and 94 recoverable injuries. 29 accidents occurred in the extrusion
unit, 25 in the surface treatment, 23 in the storage
areas, 11 in packaging, 5 in machinery and 3 in
the die section. Figure 8 presents the percentage of
recoverable accidents in each unit together with the
contribution of the unit recoverable risk to the overall plant recoverable risk. In the storage area 24% of
the total recoverable accidents have occurred, while
this area has 27% of the total recoverable risk. The
number of accidents does not give the same ranking
of hazards as risk, since according to risk ranking first
comes the storage area then the surface treatment and
then the extruder unit area, while if the number of accidents was considered for risk ranking the extruder unit
would come first, followed by the storage area and the
surface treatment unit.

793

Table 2.

Occupational risk of the aluminium plant (/year).

Position
in company

Job
positions

Fatality

Permanent
injury

Recoverable
injury

Extruder operator
Extruder worker
Cutting operator
Stretching op.
Forklift op.-billets
Forklift op.-prof.
Worker storage
Die installation
Die sandblasting
Die chemical
Die hardening
Op. die machine tools
Operator entrance painting unit
Painter
Cleaner
Forklift op.paint
Manual painting
Crane op.anodizing
Worker-entrance anodizing
Operator of press for tools
Operator of machine tools
Insulation fitter
Helper packaging
Operator of packaging machi
Worker packaging
Carpenter

6
6
6
6
2
6
10
1
1
1
1
1
5
1
1
1
1
1
4
1
1
1
1
4
4
4

5,44E-06
8,75E-06
5,28E-06
6,53E-06
5,94E-06
8,40E-06
2,18E-05
6,50E-06
1,91E-05
1,23E-05
1,23E-05
9,00E-06
3,25E-05
5,99E-06
3,02E-06
3,60E-06
3,39E-06
1,37E-05
1,54E-05
1,08E-05
2,77E-06
4,90E-06
7,41E-06
7,21E-06
7,10E-06
2,71E-06

8,38E-05
1,11E-04
1,04E-04
1,01E-04
4,28E-05
5,59E-05
1,22E-04
6,33E-05
1,85E-04
1,01E-04
1,01E-04
1,43E-04
2,03E-04
4,98E-05
5,86E-05
2,46E-05
6,78E-05
1,06E-04
1,10E-04
1,34E-04
1,10E-04
1,22E-04
2,22E-04
1,36E-04
9,95E-05
1,74E-04

5,30E-05
7,18E-05
6,15E-05
9,33E-05
7,92E-05
1,14E-04
1,92E-04
5,40E-05
1,17E-04
1,08E-04
1,08E-04
7,53E-05
2,33E-04
1,20E-04
8,70E-05
3,23E-05
1,86E-04
7,01E-05
7,74E-05
6,28E-05
6,47E-05
6,99E-05
1,08E-04
9,11E-05
8,75E-05
7,21E-05

1,56E-04
2,80E-04
5,92E-05
2,54E-04
2,93E-05
6,47E-05
8,44E-04

2,40E-03
1,64E-03
5,93E-04
1,76E-03
1,06E-03
1,16E-03
8,63E-03

1,68E-03
2,76E-03
4,62E-04
1,97E-03
4,86E-04
8,22E-04
8,18E-03

Extrusion unit
Storage area
Die section
Surface treatment
Packaging unit
Machinery
Overall Risk

3,50E-05
3,00E-05
2,50E-05
2,00E-05
1,50E-05
1,00E-05
5,00E-06
0,00E+00

3,00E-03
2,50E-03
2,00E-03

Death
Permanent Injury
Recoverable Injury

1,50E-03

worker in storage

operator entrance
painting unit

1,00E-03

die sandblasting

5,00E-04

Figure 6. Risk of each hazard (/year) for worker in storage, operator in entrance of painting unit and operator at die
sandblasting.

IN
G
C
HI
N
ER
Y
M
A

G
CK
A
PA

D
IE

CE
R
FA

SU

EX
TR
U

DE
R
ST
R
O
R
A
G
E

Struck by moving Vehicle

Contact Hanging or Swinging Objects
Contact flying object

TR
.

0,00E+00

Fall from height

Contact falling object from crane
Trapped between

Figure 7. Annual risk (death, permanent injury, non permanent injury) in each unit.

794

occupational risk in the plant and therefore the most

dangerous job and units can be identified.

3,50E-01
3,00E-01
2,50E-01
2,00E-01
1,50E-01
1,00E-01
5,00E-02
0,00E+00
ER
Y

AC
H

IN

G
IN
G

PA
C

FA
SU
R
Risk

KA

TR

IE

CE

R
O
ST
R

EX
TR

AG

DE
R

REFERENCES

Number of accidents

Figure 8. Risk contribution and percentage of accidents in

each unit.

Occupational risk has been performed for an

aluminium plant and risk of fatality, permanent injury
and recoverable injury have been calculated for all
job positions, plant units and for the whole company.
Risk prioritization can be achieved by quantifying

Ale B.J.M., Baksteen H., Bellamy L.J., Bloemhof A.,

Goossens L., Hale A.R., Mud M.L., Oh J.I.H., Papazoglou
I.A., Post J., and Whiston J.Y., 2008. Quantifying occupational risk: The development of an occupational risk
model. Safety Science, Volume 46, Issue 2: 176185.
Doudakmani O. 2007, Risk assessment of an aluminium
plant, Master Thesis dissertation, Greek Open University.
Papazoglou I.A., Bellamy, L.J., Leidelmeijer, K.C.M.,
Damen, M., Bloemhoff, A., Kuiper, J. Ale, B.J.M. and
Oh, J.I.H., Quantification of Occupational Risk from
Accidents, submitted in PSAM 9.
RIVM Report 620801001/2007, The Quantification of
Occupational Risk. The development of a risk assessment
model and software, WORM Metamorphosis Consortium, 2008.

795

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk regulation bureaucracies in EU accession states: Drinking water safety

in Estonia
K. Kangur
Kings College London, Department of Geography, London, UK
Estonian University for Life Sciences, Tartu, Estonia

ABSTRACT: The objective of the research presented in this paper is to improve understanding of risk regulation
regimes in new European Union member states. In order to do so, the research focuses on drinking water safety
regulation in Estonia. This paper tests the importance of rules, cultures, capacities and design of regulatory
bureaucracies in determining the processes and outcomes of a risk regulation regime. The effect of fragmented
nature of the regime, deep-rooted dominating pressures, institutional capacities, and regulatory actors present
in Estonian drinking water regulation are discussed in this paper.

1.1

RISK REGULATION IN THE EU ACCESSION

COUNTRIES

This article looks at the drivers of a risk regulation

regime functioning in an EU accession country, Estonia. The analysis specifically focuses on the bureaucracy as the determinant of the regulatory regime
effectiveness. The risk regulation regime approach
is used in order to disaggregate the components and
drivers inside the regulatory bureaucracy.
Risk regulation poses a particularly interesting
domain for analysing regulation, as it tends to be constrained by limited resources, competing priorities,
cognitive uncertainties, bounded rationalities, conflicting interests, ungovernable actors, and unintended
consequences (Rothstein et al., 2006). Risk regulation
in new member states offers a particularly compelling
material of analysis as the application of the European Union Directives is more than just adapting the
black-letter of the law.
The member states themselves as well as the
European Commission set high hopes on the implementation of new regulation in accession countries.
However, problems may appear with enforcing the
European Directives as new regulations have to fit with
the deep-rooted dominating pressures, institutional
capacities, and regulatory actors present in the state.
The case of drinking water safety regime is ought to
exemplify some of the problems the regulatory system
of an EU accession state, Estonia is experiencing.

Case of drinking water regulation

In order to understand the functioning of different

components of regulation, we should first have a
look at the institutions in place working for drinking
water safety. The drinking water regulation cuts across
governance regimes related to the Ministry of Social
Affairs and the Ministry of Environmental Affairs in
Estonia.
In the process of EU accession, the European
Drinking Water Directive 98/83/EC (Directive 1998),
(DWD) was adopted to be an Estonian regulation. In
order to protect public health, an Order for drinking
water regulation (Ministry of Social Affairs 2001) sets
risk-based standards for 42 parameters. These include
include maximum allowable concentrations for chemical and microbiological components and organoleptic
characteristics of the water. The Health Protection
Inspectorate under Ministry of Social Affairs supervises quality and safety requirements of drinking
water. The drinking water regime is closely linked
to the water protection programmes of the Ministry
of Environment. The allocation of the EU Structural
support finances for drinking water and sewage system
upgrading is decided in the Ministry of Environment
Water Department.
The municipalities bear responsibility for providing
drinking water supply and treatment services unless
they have given the responsibilities to a private company. After privatisation of the state water supply
system in 1995-1997, most of the municipal water

797

utilities were transformed into public limited companies. Public water supply companies serving larger
settlements belong to the Association of Water Works.
Some smaller municipalities have joined in a municipal syndicate in order to be able to cope with expensive
development of the water supply systems.
This article tries to find explanations for current
functioning of the drinking water safety regulation.
The risk regulation regime approach will be applied in
order to address the mechanics and dynamics as well
as the outcomes of the regulation.

RISK REGULATION REGIME APPROACH

The risk regulation regime (RRR) framework looks at

the governance of risk as a holistic system to analyse why and in which stages new regulation might
fail (Hood et al., 2004). The RRR perspective encompasses the complexity of institutional geography, rules,
practices and animating ideas related to regulation of
a particular hazard.
The RRR approach ambitiously brings out systematic interaction of regime components of informationgathering, standard-setting, and behaviour modification. The nature of the risk, the lay and expert understandings of the risk and its salience, but also interest
groups reactions to the division of costs and benefits of risk regulation may influence its functioning as
forces from the context of regulation (Fig. 1).
Risk regulation as a whole is under-explored in
new member states of the European Union, let alone
though any comparative perspectives in the Baltic
States region. The next section draws on some of
the theoretical knowledge that is available on the

Organised
pressure groupso

Europeanisation processes and the functioning of risk

bureaucracies in new EU accession states.

2.1 Functionality of the risk bureaucracy

The functioning of the core of government is crucial
in order to be able to organise measures for mitigating environmental health risks. The dynamics of
adopting EU regulation into Estonian national legal
system will be analysed. Looking at the informationgathering component of regulation should show the
systems in place for identifying where the weaknesses
might lie in putting the regulations to work. The
transfer of rules into institutions and practices of ruleenforcement needs to be observed as a crucial step
of risk regulation. The influence of regulatory design,
bureaucratic cultures, and the availability of knowledge, finances and administrative capacities will be
assessed as possible determinants of the functioning
of risk bureaucracy (Fig. 2).

2.1.1 The nature of rules

Safety rules apply in order to protect the public from
environmental health hazards. The rules for regulation have to balance the regulatory intervention by the
state (the degree of regulatory bureaucracy) and market forces, but also set the threshold of risk tolerance.
The feasibility analysis behind the rules determines
their workability. The basis of many EU environmental health directives stems from the time before the
1990s. At that time, attention was focused on the
policy formulation, with little regard to the issues
of integration and implementation outcomes (Weale
et al., 2000). Little consideration of the national context when adopting the rules in state regulation may
hinder proper implementation of safety requirements.

Public
salience
Bureaucratic
cultures

RISK REGULATION
BUREAUCRACY

Standard-setting
haviour-modification
ormation gathering

Knowledge
controversies

Figure 1.

External pressures of risk bureaucracy.

Capacities

Figure 2.

798

Nature of
rules

Functionality of
Risk
Bureuaucracy

Regulatory
design

Internal drivers of risk bureaucracy.

2.1.2 Bureaucratic cultures

Regulatory cultures, operating conventions, attitudes
of those involved in regulation, and the formal and
informal processes influence the functioning of risk
bureaucracies. The negotiations associated with the
EU accession demonstrated that in the accession states
a general political passivity did not encourage major
discussions about the suitability of EU regulations in
accessions states (Kramer 2004; Pavlinek and Pickles
2004). In contrast, the national political elites were
proactive in shifting their loyalties, expectations and
political activities toward a new pan-European centre
(Raik 2004).
It is argued that whereas long-established, more
populous EU member states are unlikely to rely so
heavily on compliance for their legitimacy, standing
and reputation, EU newcomers would naturally make
greater efforts to implement faithfully any environmental directives and thus, prove their credentials as
cooperative, reliable and committed member states
.(Perkins and Neumayer 2007). By publicising and
exchanging information about different practices and
reports on progress of enforcing the EU laws, it is
hoped that reputation, mutual learning and competition mechanisms are being set in motion among
member states (Heritier 2001).
One of the big changes in procedural norms of
regulation is the post-soviet countries shift from
command-and-control regulation to more discursive,
non-hierarchical modes of guidance (Swyngedouw
2002; Skjaerseth and Wettestad 2006). An analysis of
Central and Eastern European (CEE) environmental
policies after 1989, however, reveals strong influences of state socialist legacies in determining the
content and practices of regulatory reform (Pavlinek
and Pickles 2004).
2.1.3 Regulatory architecture
The regulatory structure entails the ways in which
institutional arrangements are adopted for comprehensive and robust regulation. Rather than the nation-state
setting their own regulatory agenda as before, in more
recent times regulatory decision-making has shifted
either upwards to the European Commission and
Council (or beyond) or alternatively downwards to the
regional or local level (Swyngedouw 2002; Lfstedt
and Anderson 2003). Introduced extensive complexity, rivalries and overlaps of competencies may inhibit
proper scrutiny of regulatory implementation (Hood
1986), and enforcing regulations can become rigid and
non-reflexive (Baldwin and Cave 1999).
2.1.4 Regulatory capacity
Despite massive EU financial support and long duration of transitional periods, meeting the EU requirements is hindered by the inherent inefficiencies in both

local and regional administrations of accession countries .(Weale et al., 2000; Homeyer 2004; Kramer
2004). The bureaucratic overload and the rushing for
adoption of EU policies has been described as one of
the main reasons for insufficient policy analysis and
the poor quality of legislation in Estonia (Raik 2004).
Due to the disproportionate development of different scientific agendas in the past, (Massa and Tynkkynen 2001), environmental health impact assessments
are often lacking or incomplete in Eastern Europe.
Inadequate scientific capacities on national level
encourage regulators in smaller countries to copy the
research and policy innovations of larger countries
(Hoberg 1999).
Above sections demonstrate the importance of regulatory design, embedded cultures and capacities in
determining the functioning of regulation in EU accession countries. The next section will look at how the
importance of these determinants was assessed in case
of Estonia.
3

ESTONIAN EXPERIENCES WITH

ADOPTING THE EU DRINKING
WATER SAFETY REGULATION

For information gathering on drinking water safety

regulation in Estonia, the study involved an interaction between documentary study and a snowballing
programme of interviews to locate relevant documentary material and elicit attitudes and practice not
contained in documents. Much of the information
required was found in documents detailing doctrine
and practice of regulators, legal and statutory materials from Estonian re-independence time, EU directives
and scientific literature. The open-textured analysis
of these materials was followed by identifying and
filling in the gaps through semi-structured interview
programme with 22 Estonian key actors from regulatory components key institutions (ministerial level,
regional and local administrations and inspectorates)
and representatives of dominant actors in the context
of regulation (scientific experts, water suppliers and
organised interest groups). The fieldwork conducted
from June 2007 to March 2008 revealed complex
insights of the motivations, objectives and expectations of different governance actors with regard to
drinking water safety and its new regulations.
3.1

Bureaucracies and the standard-setting

component of drinking water regulation

Estonian Ministry of Social Affairs Order on drinking

water (2001) applies to all the water supplies providing drinking water for communities larger than 50
people. In the process of adopting the DWD for Estonian regulation, national provisions and even stricter

799

safety requirements could have been added to make the

regulation more protective of the Estonian public. A
derogation period until 2010 was set for smaller operators (serving up to 5000 people) to meet the drinking
water quality requirements, to allow for renovation of
water purification and supply systems.
A group consisting of scientists, representatives of
the larger water companies and the Ministry of Social
Affairs officials was formed as an expert group, in
order to draft the drinking water quality laws. Neither
consumer protection groups nor representatives of the
municipalities were present at the regulation drafting
round-table. The scientists advocated the inclusion of
more restrictions of hazardous substances (e.g., barium) and the setting of minimum values for water
hardness. Water suppliers argued for more risk-based
monitoring schemes. After the group negotiations, the
Ministry of Social Affairs made the decision simply
to adopt the DWD, with no modifications. Three main
factors can be considered as influential for why no
Estonian national modifications were considered.
Firstly, the DWD adoption negotiations in Estonia
were rushed because of the fast pace of EU accession. There was a situation in 2000-2004 where EU
accession countries such as Estonia, Latvia, Lithuania,
Slovakia, Slovenia, Malta and Cyprus were competing with each other for faster integration into the EU
regulatory system. Thus, there was little time for consideration for any alternative policies from that of the
established EU requirements. As a result, the need
for inclusion of additional safety parameters or the
applicability of the EU directives requirements was
not analysed.
Secondly, as the Estonian bureaucratic system has
a long tradition of elite-centred decision-making, its
bureaucrats have had little experience in carrying
out consultations with other relevant parties in the
decision-making process. Skills for interacting with
outside expertise and knowledge about how to make
use of the information gathered from parties that were
present in the national regulation design were missing.
This did not allow for proper addressing of local expertise regarding the drinking water issues. A motive
for not acknowledging national drinking water related
studies may stem from the bureaucrats educational
background, as the interviewed scientists claimed. The
general neglect of education on environmental health
issues has contributed to bureaucrats low awareness
about the drinking water issues in Estonia. Thus, the
Ministry officials might not have been capable of
appreciating the scientific contributions to drinking
water regulation design.
The third aspect that may have driven the policymakers to neglect any national modifications was the
poor extent of communication with other levels of
regulatory regime. The Ministry of Social Affairs
may have overlooked issues such as 23% of mostly

rural population not being under surveillance simply

because there might have not been enough pressure
from the lower levels of regulatory regime (municipalities, local inspectors) to explain the need for
changing the rules. An understanding of the problems
on the ground could easily become obfuscated within
the hierarchical regulatory structure. This was especially the case during the accession period, when the
regulatory cadres were changing rapidly.
3.2

Bureaucracies and information gathering on

the change of states in drinking water safety

Having adopted the new regulations, there was a need

to obtain information about the change of states in
drinking water safety. Regular monitoring of drinking
water quality by operator companies should follow the
Health Inspectorates prescriptions. The frequency of
how often water suppliers are obliged to take full tests
(covering 42 quality requirements) and minimal number of tests (18 quality parameters) depends on the
size of the companys clientele. All economic water
suppliers (providing water from 100 to over 100 000
people) are expected to test four times a year for a
minimum monitoring program. Companies providing
water to less than 10 000 people only have to take
one full test per year. Companies serving 10,000 to
100,000 people take full tests three times a year and
operators with over 100,000 clients as often as over
10 times a year. After every five years, the European
Commission checks on the monitoring results of especially larger water companies (serving more than 5000
people).
In reality, neither the operators nor the Health
Inspectorates test water as often as expected, as the representative of Inspectorate commented in interview.
Tests conducted by smaller companies are sporadic;
and there are very few records on the drinking water
quality provided through private wells. This means that
in addition to the uncontrolled drinking water sources
from private wells (23% of population), the water
delivered by smaller operators (providing for 13% of
population) is also not under systematic surveillance
for any potential health risks.
There are regional differences in water quality. At
the time of adoption, the representatives of water
companies and the scientists pressured for more sitespecific risk-based monitoring requirements. Concentrating more frequent testing on those hazardous
substances found in the area and carrying out fewer
testing of the other substances could have alleviated monitoring pressure on water companies. The
rationale for site-specific monitoring programme was
not considered in designing monitoring guidelines,
though.
There are some reasons that might explain why the
Inspectorates check smaller companies less frequently.

800

Firstly, there is insufficient financial and personnel

capacity for controlling the large number of water suppliers. The representative of the Health Inspectorate
recently interviewed stressed that the inspectors are
under a heavy workload and the finances to conduct
full analyses are scarce.
Secondly, the Inspectorate may not be completely
aware of or simply not give full consideration to the
risks that non-surveillance may pose to the people
obtaining uncontrolled drinking water from smaller
operators. The European Commission demands information only about the larger water companies from the
Estonian Health Inspectorate. Therefore, Inspectors
have no incentives for pushing for wider control.

3.3

The functioning of behaviour-modification

according to the drinking water regulations

The records on behaviour modification according to

the drinking water regulations show that larger companies are now using modern purification techniques
and are for the most part following the set quality
requirements (Sadikova 2005). The scientific studies and Inspectorates data, which is available on the
water supplies in the rural areas, however, shows that
many risk parameters exceed the defined health limits. For example, northern Estonian drinking water
sources are affected by a radioactive bedrock zone
(Lust et al., 2005), the north-eastern part of Estonia has
excessive sulphates, and high fluoride levels are problematic in western areas of Estonia (Karro et al., 2006).
Shallow individual wells are generally subject to nitrogen compounds from fertilizers and microbiological
contamination (Sadikova 2005). These contaminants
have been associated with gastrointestinal, circulatory and/or nervous system diseases. There are still
relatively poor water infrastructure and purification
systems in parts of Estonia. This, together with the
poor information about drinking water quality, puts up
to 36% of the public (customers of smaller companies
and users of private wells) at risk with their health.
The enforcement of the drinking water regulations may be achieved through investments or strict
regulation. Due to the structure of the bureaucratic
enforcement system, the Ministry of Environment has
the power and control over allocation of finances with
respect to water. Thus, environmental concerns have
prevailed above those concerned with public health.
The Ministry of Environment with minimal consultation with the Ministry of Social Affairs designed the
holistic water management plans. Priority given for
larger communities drinking water supplies and sewerage system updating has been cost-beneficial from
the larger environmental and public health points of
view, as the larger communities are safeguarded. However, the safety of inhabitants of large rural areas and

the clients of smaller water suppliers has been compromised due to negligence of health priorities that
should be promoted by the Ministry of Social Affairs.
Another set of enforcement problems is related
to the way regulation works. Investing regulatory
attention into the larger companies may be explained
through the Estonian bureaucratic culture, which prioritizes good performance with respect to the European Commission. As the records that interest the
European Commission most (those on larger company
compliance) show conformity with EU requirements,
the Inspectorates are seen by Brussels to be efficient.
Thus, credibility among the public seem to be a less
important driver for the Inspectorates. This is especially as they are used to top-down decision-making
traditions.
One could presume that if those smallest suppliers
and private wells were not monitored, there would be a
system for informing the individual well users to take
protective measures. Yet, there have not been any information campaigns nor has there been any demand for
more information regarding the drinking water from
the users. A consideration that the private wells are
something that the owners have to manage themselves
is prevailing. The support schemes and information
dissemination for addressing the drinking water quality in individual wells has not been institutionalized
through the Ministry orders either.
4

CONCLUSIONS

Aligning national safety standards with the European Union rules requires from bureaucracies careful
decisions about the organisation of regulatory responsibilities, about the approaches for attaining their
objectives, as well as choices about practical allocation
of ever-scarce capacities to employ these strategies.
This paper focused on the bureaucratic determinants of drinking water safety regulation efficiency
in Estonia. The standard-setting, monitoring and
enforcement activities associated with drinking water
regulation in Estonia may be described as an EU allegiance striving process. Search for power, patronage
and reputation are the main compliance-driving forces
of inspectors on national level, but may also determine
the state bureaucracies behaviour on European level.
Deeply rooted bureaucratic cultures may function
as gatekeepers for the take-up or neglect of more
innovative non-hierarchical modes of enforcement.
National scientific incapacities have carried over to
poor bureaucrats awareness on drinking water safety
issues leading to insufficient local policy analysis and
simply application of preset rules. Available financial
and administrative capacities have led to a reinterpretation of the set standards and some neglect of smaller
operators quality control. Allocating scarce resources

801

and controls to the larger companies has benefited the

viability of larger communities, but smaller and rural
communities appear to have been ignored.
The complexity of the regulatory structure, span
through the EU expert committees, national levels of
government, and their sub-departments, may create
an illusion of regulatory control, yet the real drinking
water safety issues may remain unattended.
ACKNOWLEDGMENTS
This article presents some of the findings of authors
PhD research at Kings College London that is
financed by Estonian Academic Mobility Foundation and Ministry of Science and Education grant SF
0170006s08.
REFERENCES
Baldwin, R. & Cave, M. 1999. Understanding Regulation:
Theory, Strategy and Practice. Oxford: Oxford University
Press.
Directive. 1998. Directive 98/83/EC of the European Parliament and of the council of 3 November 1998 on the
quality of water intended for human consumption. Official
Journal of the European Communities: OJ L 330.
Heritier, A. 2001. New Models of Governance in Europe:
Policy-Making without Legislating. Vienna: Renner
Institute.
Hoberg, G. 1999. Sleeping with an elephant: the American influence on Canadian environmental regulation. In
B. Hutter (ed), A Reader in Environmental Law: 337363.
Oxford: Oxford University Press.
Homeyer, V.I. 2004. Differential effects of enlargement on
EU environmental governance. Environmental Politics
13(1): 5276.
Hood, C. 1986. Administrative Analysis: An introduction to
Rules, Enforcement, and Organizations. Sussex: Wheatsheaf Books.

Hood, C., Rothstein, H. & Baldwin, R. 2004. The Government of Risk. Understanding Risk Regulation Regimes.
Oxford: Oxford University Press.
Karro, E., Indermitte, E., Saava A., Haamer, K. & Marandi,
A. 2006. Fluoride occurrence in publicly supplied drinking water in Estonia. Environmental Geology 50(3):
389396.
Kramer, J.M. 2004. EU enlargement and the environment:
six challenges. Environmental Politics 13(1): 290311.
Lust, M., Pesur E., Lepasson, M., Rajame, R. & Realo, E.
2005. Assessment of Health Risk caused by Radioactivity
in Drinking Water. Tallinn: Radiation Protection Centre.
Lfstedt, R.E. & Anderson, E.L. 2003. European risk policy
issues. Risk Analysis 23(2): 379.
Massa, I. & Tynkkynen, V.P. 2001. The Struggle for Russian
Environmental Policy. Helsinki: Kikimora Publications.
Ministry of Social Affairs. 2001. Requirements for Drinking
Water Quality and Control, and Analysis Methods. In RTL
2001, 100, 1369. Riigi Teataja: Tallinn.
Pavlinek, P. & Pickles, J. 2004. Environmental pasts & environmental futures in post-socialist Europe. Environmental
Politics 13(1): 237265.
Perkins, R. & Neumayer, E. 2007. Implementing multilateral
environmental agreements: an analysis of EU Directives.
Global Environmental Politics 7(3): 1341.
Raik, K. 2004. EU accession of Central and Eastern European
countries: democracy and integration as conflicting logics.
East European Politics & Societies 18(4): 567594.
Rothstein, H., Huber, M. &, Gaskell, G. 2006. A theory
of risk colonization: The spiralling regulatory logics of
societal and institutional risk. Economy and Society 35:
91112.
Sadikova, O. 2005. Overview of Estonian Drinking Water
Safety. Tallinn: Health Inspectorate.
Skjaerseth, J.B. & Wettestad, J. 2006. EU Enlargement
and Environmental Policy: The Bright Side. FNI Report
14/2006. Lysaker: The Fritjof Nansen Institute.
Swyngedouw, E. 2002. Governance, Water, and Globalisation: a Political-Ecological Perspective. Meaningful
Interdisciplinarity: Challenges and Opportunities for
Water Research. Oxford: Oxford University Press.

802

Organization learning

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Can organisational learning improve safety and resilience during changes?

S.O. Johnsen
NTNU, IO Centre, Norway

S. Hbrekke
SINTEF, Norway

ABSTRACT: We have explored accident data from British Rail in the period from 1946 through 2005. Our
hypothesis has been that safety is improved through learning from experience. Based on a quantitative analysis
this hypothesis is tested for the data using a simple regression model. We have discussed the model and its
limitations, benefits and possible improvements. We have also explored our findings based on qualitative theory
from the field of organisational learning, and have suggested key issues to be explored to improve safety and
resilience during changes, such as safety cases, standardisation of training, unambiguous communication and
sharing of incidents and mitigating actions.

1
1.1

INTRODUCTION
Learning from experience

The influence of experience on productivity has been

observed based on data from the Second Word War.
Based on experience, it was observed a systematic
improvement of productivity and reduction of time to
build airplanes and Liberty ships, see Yelle (1979).
Doubling of cumulative production, i.e. cumulative
experience, typically improved productivity by a fixed
percentageand thus reduced the unit costs by a fixed
percentage. This was called the law of experience,
expressed as:
The unit cost of value added to a standard product
declines by a constant percentage (typically between
20 and 30 percent) each time cumulative output doubles. Productivity increases due to experience at a
rate proportional to its value, and hence there is
an exponential relationship between productivity and
experience.
We suggest a similar relationship between safety
and experience. As experience is built our hypothesis
is that accidents and incidents are reduced due to learning among the workforce and management. We are
proposing that accidents due to experience in operations has the same behaviour as productivity due to
experience in operations, our hypothesis is: when
cumulative production is doubled or in general cumulative experience is doubled, accidents are reduced
by a fixed percentage. Based on this hypothesis,
there should be an exponential relationship between
accidents and experience.

Experience can be expressed different, dependent

on the type of industry to be explored. In an industry production environment it could be the number of
units produced. In transportation it could be transport
distance in miles/km or that experience has a simple
linear relationship to time.
According to the hypothesis, when plotting the
number of accidents as a function of experience, we
expect an exponential decrease in the numbers of accidents. Using time (t) as a basis for experience, the
level of accidents should be approximately given by the
exponential expression a ebt . The constant a should
indicate the level of accidents at the starting point, and
the constant b is an indication of the speed of learning
from experience.
1.2 Proposed caseBritish Rail from 1946 to 2005
British Rail (BR) was the main line railway operator
in Great Britain from 1946 to 1994. We have gathered
data on almost all fatal railway accidents from 1946 to
2005, and we have analysed the data with respect to
the hypothesis. We are basing our work on data used
by Evans (2007). In the period from 1946 there have
been major technological and organisational changes.
One major change was the deregulation of British Rail
in 1994, leading to a fragmentation of BR into more
than 100 separate organisations.
The reason to use data from British Rail has been
the number of accidents during the period, making it
possible to discuss and validate a quantitative hypothesis. The large changes introduced by deregulation in

805

1994 and their consequences on safety have also been

explored.
We would expect an increase in accidents after 1994
due the large scale changes, but increased scrutiny of
the effects and an increased focus on accidents could
moderate the negative effect of deregulation.

2
2.1

MODEL DESCRIPTION
Individual and organisational learning

We are using a definition by Weick (1991) to define

learning, .i.e.: ..to become able to respond to taskdemand or an environmental pressure in a different
way as a result of earlier response to the same task
(practice) or as a result of other intervening relevant
experience.
Based on this definition of learning the actors that
learn must sense what is going on, assess the response
based on earlier response or experience and respond
with a different behaviour.
Related to organisational learning we are focusing
on the result of organisational learning, in that we are
observing what the actors in the organisation is actually doing and the results of their actions related to the
level of accidents. Due to learning, we are assuming
that the new and safer practice is carried out regularly,
that it is present in the behaviour of several individuals
engaged in similar tasks, included in procedures and
that new members in the organisation are instructed
in the new practice. This description is based on the
definition of organisational learning as described by
Schn (1983).
2.2

Development of accidents based on experience

Based on the preceding discussion our proposed nullhypothesis is: In the long run accidents follow an
exponential decreasing regression line based on experience, where experience is expressed by time.
In our quantitative analysis we have elaborated the
hypotheses, see section 2.6.
In our proposed quantitative model the level of historical accidents at time t (years after 1946), A(t) is
following a regression line on the form
A(t) = a ebt

(1)

Here a and b are the parameters in the model. Note that

the model considers t as a continuous variable while
we in practice only consider t for the distinct integral
years. This is a weakness of the model.
Examining accident data from year to year may be
challenging due to the stochastic nature of accidents
and especially the variation in number of fatalities
from year to year.

2.3 Short term perturbations

A more complex model could take into account the
short term effects of the dynamics of learning and
focus on experience. Learning could be influenced by
the increased alertness after an accident and the complacency setting in when no accidents are happening.
The increased alertness should lead to fewer accidents,
while the increased complacency could lead to more
accidents. This dynamic behaviour could be modelled
as some sort of a sinus curve interposed on the exponential model. We have not established such a model
yet, but we anticipate that the accident data would
show such a relationship if plotted in a logarithmic
scale.
2.4 Benefits and limitations
The simplicity of the model in (1) is the major argument for the examination of the railway data for
learning by experience.
But such a simple model has its limitation, especially when analysing the results. Accidents are spontaneous and discrete, and should best be modelled by a
Poisson distribution. Also a prediction of the numbers
of accidents in the future must be done in carefulness.
However the regression line approximation illustrates
the accident trend over a rather long time period. Our
purpose is to present the shape of a trend line rather
than specified values (such as number of accident a
certain year). Another benefit of the model in (1), is the
various possibilities of exploration of learning. This
can be quantified in different ways and not only by the
number of accidents alone. This is further discussed
in section 3.1.
Equation (1) can be simplified in a logarithmic
system as a linear model instead of exponential:
ln A(t) = ln a b t

(2)

This representation makes it possible to fit a linear

regression line.
2.5 Model improvements
Further exploration of a model from an accident analyses or statistic analysis point of view can be done by
more detailed regression analysis, life time analysis or
time series modelling.
Evans (2007) uses Generalised Linear Models
(GLMs) to analyse the train data. Here the numbers of accidents during one year are assumed to
be Poisson distributed. In such a model the failure
rate as a function of time can be presented, i.e. a
model with only one parameter. Also different types
of GLMs can be compared in order to achieve a model

806

which gives the best explanation of the variation in

the data.
The accident data can also be treated as life time
data, where a life time is the time until an accident appears. Here NHPP (Non-Homogenous Poisson
Process) or Cox proportional hazard regression are
suitable alternatives.
Experience related to transportation could be
defined not only as time, but as accumulated travel
length as well. Experience in a production environment could be defined as time or accumulated
produced equipment. Both GLMs and Cox regression
can take into account as many explanatory variables
as desired, and tests to check the significance of the
variables in the model can be executed. Thus accumulated travel length (l) could be an explanatory
variable in addition to time, e.g. for the exponential
regression:
A(t) = a e(b1 t+b2 l) .
Another alternative for analysing the train data is
by time series modelling. In the data analysis we will
see that the observations seem to oscillate around the
regression line. Thus e.g. an ARIMA model could be
relevant to estimate and explore.
2.6

Hypothesis testing

The quantitative null-hypothesis proposed in section

2.2, that the accidents follow an exponential decreasing trend, is b is significantly different from 0 in the
regression line described by (1).
To decide whether the values follow an exponential expansion; i.e. if our hypothesis is not rejected,
we evaluate different measures. If the residuals show
independency and seem to be sampled from a normal
distribution, we also check if the P-value, Confidence interval and Standard deviation of b, the
T-statistic and the R 2 -value give reasons to reject the
hypothesis. If one of the following criteria (all output from the analysis) is fulfilled we will reject the
null-hypothesis:
1. The T-statistic for b is in absolute value below 2.0;
i.e. b is not a significant parameter (considered as 0)
and our model reduces to A(t) = a.
2. Standard deviation of b is great compared to the
parameter value; the values scatter.
3. P-values (probability of an accident level being
totally independent of the particular year) is greater
than 0.05.
4. Dependent residuals, i.e. the residuals are not
independent and normally distributed and show
certain trend; i.e. the model does not fit the data
rather well.

5. R 2 (indicates how well the regression line fits the

data) is less than 0.801 .
If our analysis shows none of the above, we cannot
reject the hypothesis that the accident level follows
an exponential decreasing trend, and the hypothesis is
accepted.

DATA ANALYSIS

3.1

Presentation of data

This analysis is based on data of all railway accidents

where the major part is movement and non-movement
accidents. Collisions, derailments, overruns and collisions between train and road vehicles are also classes
of accidents included in the study. Train fires and other
train accidents are not included.
The data represents both fatal accidents and the total
number of fatalities in each accident. Every fatal accident registered has at least one fatality. In addition
the number of million train km per year is part of the
data, and accidents pr million train km is a natural
expression for the accident level.
In our model, the sum of accidents and fatalities pr
million train km is used. As fatal accidents are the
numbers of events in which the fatalities occurred,
this may seem rather odd. However experience and
learning is most probable achieved both through high
frequency of accidents and severe consequences of
accidents from which organisations learn. We have
chosen an equal weighting of both accidents and fatalities pr million train km as the quantitative measure of
accidents. Both frequency and consequences are taken
into account. It is most likely that learning increases
due to the number of accidents and the number of
fatalities in an accident. If we were only considering the number of accidents, years with few accidents
but many fatalities would give a wrong impression of
the learning level. Analysing fatalities only, do not
take into account the learning due to the number of
accidents with few fatalities.
Alternatively a different weighting of the numbers
of accidents and fatalities, respectively, may give a
more appropriate picture of the accident levels. However, an exploration of different weighting shows that
this does not make any significant influence on the
analysis with respect to our hypothesis.

1 The

value of 0.80 should not be considered as a normative value. It is just a value in order to explore the goodness
of fit in addition to what can be seen from the plotted estimated regression line together with the observed data.
In this case we are satisfied with 80 percent or more
explanation of the data variation in the model.

807

It is worth mentioning that the data reported was

in 1991 changed from calendar year to fiscal year
(1 April to 31 March) and in 2005 changed back again.
This means that there may be some data missing/
overwritten during the two transitions for the data used
in this analysis. However, the accidents/fatalities and
the train-km data have been collected over the same
intervals, such that the transitions have very restricted
influence on our analysis.

ten years all values are above the estimated line,

followed by a period in the 90s where the values
are all below the estimated line. The last ten years
with observed data again seems to give independent
residual.
5. R 2 equals 0.91.

3.2

A(t) = 0.27 e0.043t ,

Data analysis

A least square regression is executed to the sum of

the number of accidents and the number of fatalities
pr million train km in the years 1946 thorough 2005.
The best fit is characterized by the sum of squared
residuals with the least value, a residual being the difference between an observed value and the value given
by the model. The regression estimates the parameters
a and b. To evaluate the goodness of fit of the model
to the observed data, we examine all the R 2 value,
the residuals, P-value, T-statistic and confidence interval of the parameters in order to verify the hypothesis
stated.
We see that the observed values seem to nearly
follow an exponential trend.
Results of the ANOVA test:
1. The T-statistic for b is in absolute value a lot greater
than 2.0. The 90% confidence interval of b is
(0.0455, 0.0398).
2. Standard deviation of b is only 4% of the parameter
value.
3. P-values are far less than 0.05.
4. From the plot above the residuals seems to be
independent and their values indicates a normal distribution. The independency becomes less reliable
after the beginning of the 80s. In a period for about

Based on the ANOVA results, the null-hypothesis

cannot be rejected and are thus is accepted. The
estimated regression line from (1) is:

and the corresponding estimated linear model in (2) is

ln A(t) = 1.3 0.043t.
3.3 Short range fluctuations
We observe that the actual accident data seem to oscillate as some sort of a sinus curve around the estimated
curve, see Figure 2. This could be due to natural variation or a result of increased awareness and increased
safety after an accident, followed by complacency and
less safety in a period with no incidents or accidents.
3.4 Discussion of parameters
The parameters a and b estimated are of interest for
comparing different periods or different industries.
The value of a tells where we are at the start of the
learning process, and the value b says something about
how fast we learn. The greater b the faster learning,
but one should note that this may depend on a: If data
considered in the analyses are from a period where the
industry is new we except a greater b than if the data
considered are from a period where the industry is well
established.

1,6

0,5

1,4

1,2

-0,5
-1

-1,5

0,8

-2

0,6
-2,5

0,4
-3

0,2
0
1940

Figure 1.

-3,5
1940

1950

1960

1970

1980

1990

2000

1950

1960

1970

1980

1990

2000

2010

2010

Long range fit of accident data from 1946 to 2005.

Figure 2. Short range fluctuations between actual accident

data and estimated model (2).

808

This can be illustrated using the British Rail data,

e.g. by dividing the period in two, i.e. one period
from 1946 to 1975 and the other from 1976 to 2005.
Analysing these two separately, gives the following
distinctions:
The first period fits the data better than the last.
The first period has a greater b value, and hence
indicates stronger learning than the last.
We have also established a model based on cumulative distances travelled as a basis for experience, and
have also identified a good fit.
The following two remarks must be considered
under examination of data with respect to accidents
and learning:
1. The period (time and length) of analysis: What
period in the history of the industry is considered?
It is probable more likely that learning and development are happening initially. Also the length of
the period plays an important role, due to increased
validity of a large sample size. For a huge dataset
will for instance outliers (observed values rather
distinct from estimated values) have less influence
on the regression line (both establishing and evaluating). A small dataset without any particular trend
will often result in a regression line rather similar
to the horizontal/constant mean value line. Hence,
it is important to consider a data sample of an
appropriate sample size.
2. The number of accidents: If few accidents are
observed, the model does not fit very well. This
could be interpreted as the stochastic character of
accidents.
In the analysis, we have considered the total number
of both accidents and fatalities for all types of accidents. However, dealing with accidents and fatalities
separately do not impact the acceptance of the hypothesis test. As expected the accidents fit the model better
than the fatalities, as fatalities naturally vary more. For
the different types of accidents it is only the class for
movement and non-movement accidents which fits the
model. This accident type also makes the greatest contribution to all accidents. The other two classes, train
collisions and collisions between trains and road vehicles, probably due to few accidents (the model do not
handle 0-values) and large variation in the data, do not
fit the model.

4
4.1

DISCUSSIONS OF SOME RELEVANT

ISSUES
Use of technology to improve safety

Technology could be an important factor in safety,

but often new technology initially is mimicking old

technology. An example is the implementation of electric motors replacing steam engines. At first the large
steam engines were replaced by large electric motors
in factories, using long chain drives to distribute the
power in the factory. Later small electrical motors were
distributed directly in the production process as they
were needed. This removed the danger from the long
chain drives. Thus the efficient and safe use of new
technology was based on exploration and learning in a
period that could take years. This is discussed further
in Utterback (1996).
Improvements in working procedures and safety
thus usually take place after some years of experiences
with new technology.
4.2 Consequences of our model of learning
from accidents
We have proposed a hypothesis, suggesting a model
where the levels of accidents are dependent on experience; more precisely the levels of accidents per year fit
an exponential trend. We have based our model on the
experiences from the Railways in Great Britain. Our
simple model of accidents at time t, is A(t) = a ebt .
The hypothesis indicates that experience and the
level of learning is a key element in reducing accidents.
If we are using the iceberg theory from Heinrich (1959) we are assuming that accidents, near
accidents/incidents and slips are attributed to the same
factors. The distribution between these categories was
suggested by Heinrich to be:

1 Major accident
29 Minor accidents
300 incidents, no-injury accidents
3000(?) slips, unsafe conditions or practices.

The result of more recent studies contradict this

hypothesis, one issue is that severity is dependent on
the energy involved and the distribution of incidents
differ between different types of workplace. However, our point is to focus on learning from minor
accidents or incidents in order to avoid major accidents. Both accidents and deviations are of interest
related to learning. If we manage to focus and document minor accidents, incidents and slips in order
to establish organisational learning; and succeed with
this learning, it should imply that major and minor
accidents should decrease.
A Heinrich pyramid with few major accidents
but documentation, exploration and learning of many
slips could be an indication of a learning and resilient
organisation.
However, if we do not focus, document and explore
minor accidents, incidents and slipsit could lead
to more serious accidents due to less learning from
experience. Thus in en environment of less learning

809

we should find a Heinrich pyramid with more major

accidents but few slips.
At the 2007 HRO conference, in Deauville, May
2830 examples of this were mentioned from aviation
in USA. Airlines that had explored minor accidents,
incidents and slips had none serous accidents. Airlines
that had documented few minor accidents, incidents
and slips had several serous accidents. This could indicate poor learning. Since this is comparison between
companies in the same industry with same energy
level, this could indicate different order of learning.

4.3

Level of analysisIndustry comparisons

We can use the suggested model and perform an analysis on different organisational levels and of different
industries.
By different organisational levels we mean either:
(1) within an organisation, looking on accident data
from different parts of the organisation analysing
which part having the better organisational learning
and thus safer practices, (2) comparing same type
of organisations within a countryidentifying what
organisations having the better organisational learning
system and (3) comparing same type of organisations
between countriestrying to identify what countries
having the better organisational learning systems.
Organisations could be compared to discuss which
kind of organisation has the better accident learning
system. As an example, within aviation you could
identify the airlines having the better accident learning
system.
Different industries (e.g. aviation and shipping)
could be compared to discuss which kind of industries
has the better accident learning system.
In a system with better organisational learning, the
consequences should be that the whole system is learning faster and accidents are decreasing faster, e.g. a
greater value of b. Improved organisational learning
in this context should mean more information sharing
among the involved actors in the whole system and
better understanding of the causes of accidents.
In a system with initially safer technology, the initial
value of a should be lower, meaning that the levels of
accidents are initially lower.
In the airway industry there is international sharing of incidents and accidentsa system that seems
to have improved organisational learning, due to more
incidents and more learning experiences across countries and different organisations. Thus it should be
checked if the airway industry is learning faster and
accidents are decreasing faster, e.g. a greater value of
b in the same period than other industries.
Note that the values of a and b from the regression
results analysing different industries are only comparable when considering the same type of industry

(e.g. transport) as the response measures must be equal

(e.g. accidents per million transport kilometre).
4.4 How to increase safety further?
If improved safety is due to experience and learning
from accidents as suggested by our model, safety may
be improved by focusing more on exploring organisational learning and organisational inquiry as discussed
by Schn (1983). This means that we should explore
what has gone wrong based on using actual data in an
open and testable manner. We should try to explore
double-loop learning, doing reflection in a learning
arena among relevant actors from both inside and
outside the organisation.
Some consequences of that could be an improved
focus on:
Common training to ensure clear communication
and common understanding of what has happened.
Exploring known incidents more thoroughly and
ensure that the employees are aware of what can go
wrong and why. This must involve key actors in the
organisation to ensure that organisational learning is
taking place. A systematic analysis of several incidents together could help identify common errors
and thus increase learning.
Improve and focus on scenario training by discussing and reflections on what can go wrong in
the future, especially with a cross functional team
involving actors from outside the organisational
boundary.
4.5 What should be done in an environment
with no accidents?
In an environment with almost no accidents, it
becomes more important to analyse minor accidents,
incidents and slips in an open manner. Sharing of
incidents must be performed, in order to improve learning. The use of narratives or story-telling could be
used to create understanding and support of a learning
environment.
Scenario training should also be explored to
increase and sustain organisational learning.
4.6 Effect of the deregulation of BR in 1994
When we look at accident data from 1994, we do not
find serious deviation from our model and hypothesis.
There were no large scale increases of accidents, thus
indicating that mitigation of the changes were taking
places such as learning and sharing of experience. One
argument to support this was the increased focus on
safety in the railway system due to the deregulation and
increased public scrutiny of accidents and incidents.

810

The increased focus, information and discussion may

have sustained the low level of accidents after 1994.
Other issues that has been mentioned, from Evans,
has been that the greatest number of fatal accidents are
movement and non-movement accidents. Two particular things that will have contributed to the fall in these
are: (1) A major reduction in the number of trains
without centrally-locked doors (they have now gone
completely). The doors of trains without central locking could be opened while the train was in motion,
and there were fatalities due to passengers falling from
trains, usually because they opened a door. Such events
are now virtually non-existent.
(2) Fewer track workers and better management
of track maintenance work, leading to fewer fatal
accidents to staff working on the track.
4.7

Further exploration of the model

In this paper we discuss neither the use of the model

in prediction nor the uncertainty/confidence bounds.
Data from a restricted period were the latest years are
excluded, are fitted to a model which can be used to
predict the data (most likely in the future). The predicted value can then be compared with the actual
historical value to see whether or not they are inside
the confidence bounds and then verifies the property
of prediction of the model.
4.8

Objectivity, reliability and validity

The proposed hypothesis can be tested on different

industries, and there are clear objective criteria to be
used to check if the hypothesis is correct as described
in 4.1. We do propose that the result is objective.
We are proposing that the result is reliable, due to
the quantitative approach we have used.
Validity is difficult to ascertain, the hypothesis
should be explored in different industries and on different data to be assured of validity. In Duffey (2002)
there are several examples validating these issues.

PRACTICES FROM THE RAILWAY

INDUSTRY

We have discussed some organisational learning issues

with the railway industry related to changes both
locally and changes involving international railway
traffic. Key issues identified during interviews and
workshops have been related to new practices related to
communication, procedures, responsibility, incident
reporting, training and common risk perceptions.
The changes have been related to the changes in
British Rail but also related to railway traffic from
Britain to France.

We have structured the responses based on how

learning is taking place, e.g. issues related to sensing what is going on based on communication and
common training, who has responsibility to act, and
how to assess the response based on earlier response
or experience. The main practices we have identified
are in accordance with our hypothesis related to learning, e.g. a strong focus on practices supporting and
enabling organisational learning that is a focus on:
Sense what is going onbased on unambiguous
communication, and common training;
Unambiguous responsibility:
Respond based on earlier experienceunambiguous
procedures:
Support learning based on experiences and incident
reporting across organisations:
Organisational learningcommon procedures and
perceptions.
This is explored in the following.
1. Sense what is going onunambiguous communication: The use of protocols or formalised communication templates is essential when communicating
cross interfaces. Pre-determined protocols and forms
reduce difficulties in understanding and learning and
should be mandatory.
2. Sense what is going on based on common
trainingStandardised training for operators, focusing on communication and handling of deviations. It
is especially important to establish common models or
perceptions of how incidents happen as described by
Schn (1983). It is also important to share an understanding of culture e.g. perceptions, knowledge
and behaviour between the different companies. Good
experience has been obtained by the use of scenario
training and simulators in addition to normal training.
In a simulatorscenarios including deviations from
normal operations can be tested, and the other side of
the interface can be included.
3. Unambiguous responsibility: Unambiguous or
unclear responsibility (e.g. grey areas of responsibility) should not be tolerated. It is essential to have a
perfect clarity in tasks definition and responsibilities
cross interfaces, especially if learning and exploration
should be done.
4. Respond based on earlier experience
unambiguous procedures: It is essential that different
parties harmonise their procedures so that operators
adopt the same behaviour both during normal operations and exceptions. Based on our hypothesis, learning is based on incidents and accidents. To improve
learning across boundaries it should be important to
decide on one set of rules cross boundaries, and ensure
that both the basic rules are the same and also that the
common understanding of the rules is the same. Translation of a rule from English to French and then back to

811

English could be one way of exploring possible differences in understanding and increase learning in cross
border cooperation between Great Britain and France.
Rules and procedures that are actually used should be
kept as living documents, meaning that the documents should be updated by the working professionals
themselves.
5. Support learning based on experiences and incident reporting across organisations: It should be a
clear obligation to report any condition that could
imply a risk for other companies. All parties must share
their databases regarding events that could improve or
degrade safety, and also share the resulting recommendations. This would ensure a possibility for common
learning and an increased level of safety for all operators. Both interfacing organisations will benefit from
the ability to admit that they are different without inferring value or preference. One partners solution is not
necessary the only right solution, one should share
experiences (both from accidents, fatalities and good
practices) to provide an opportunity to learn from each
other.
6. Organisational learningcommon procedures
and perceptions: Harmonisation of procedures by
project teams cross organisational boundaries should
be done. Experience shows that groups with representatives from each of the companies (or countries)
involved in operations should be established and meet
face to face, to create confidence, common understanding and a good learning environment, and establish harmonised procedures. As part of organisational

learning it should be important to establish proactive

and common risk perception and understanding: It
would be helpful for two different organisations or
actors to agree on a common model (common mental model) for identifying and managing risks and
the resources to control risks. Some of the most difficult issues to resolve are due to differences in the
conceptualisation of risks and risk management.

REFERENCES
Duffey, R. and Saull, J. 2002, Know the Risk: Learning from Errors and Accidents: Safety and Risk in
Todays Technology Butterworth-Heinemann ISBN-13:
978-0750675963.
Evans, A.W. 2007. Rail safety and rail privatization in
Britain, Accident Analysis & Prevention 39: pages
510523.
Heinrich, H.W. 1959. Industrial accident prevention
A scientific approach Mc Graw Hill, New York.
Schn, D.A. 1983 Organisational Learning in Morgan
G (1983) Beyond Method Sage, Beverly Hills, pages
114129.
Utterback, J.M. 1996 Mastering the Dynamics of Innovation HBS, Boston.
Weick, K.E (1991). The non-traditional quality of organizational learning, Organization Science.
Yelle, L.E. 1979. The Learning Curve: Historical Review
and Comprehensive Survey, Decision Sciences 10:
302328.

812

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Consequence analysis as organizational development

Berit Moltu
SINTEF Technology and Society, Trondheim, Norway

Arne Jarl Ringstad

StatoilHydro ASA, Trondheim, Norway

Geir Guttormsen
SINTEF Technology and Society, Trondheim, Norway

ABSTRACT: In this article we argue that consequence analysis is about organisational change and thereby
methodologically should be treated as part of that. Traditionally methods of consequence analysis is not sufficient
or enough. HSE (Health, Safety and Environment) is also about Organisational Development (OD). Both in
information and data gathering, in decision and participation, and in safe and secure implementation of suggested
changes, we consider this argument to be important.
The article is based on R&D projects done in the Norwegian oil company StatoilHydro ASA under the heading
of Integrated Operations (IO). The strategy was to choose several pilot projects in one asset to be analysed by
consequences as far as HSE was concerned. The idea was further to spread the successfully pilots to other assets
after a successful Consequence Analysis (CA).
Our approach to understand organizations is inspired by Science and Technology Studies (STS) and sees
organisations as complex seamless networks of human and nonhuman actants (Actor Network Theory, (ANT)
(Latour 1986)). We understand organisations as the ongoing process created by the interests of different actants
like ICT, rooms, work processes, new ways to work, being organised and managed. This in addition to an
understanding of communities of practice (Levy & Venge 1989) is the point of starting to discuss CA as part of
OD. Another used method is based on the risk analysis tool HAZID (Hazard Identification) witch is used in the
Norwegian offshore industry as a planning tool to identify hazardous factors and to evaluate risk related to future
operations. HAZID were used as a basis for collecting qualitative data in our concept of consequence analysis.
Different method was used to identify positive and negative consequences related to implementation of (IO) in
two cases, the steering of smart wells from onshore, and a new operation model on an offshore installation.
We observed that the methods had qualities beyond just evaluation of consequences. During the interviews
on smart well different groups of actants started to mobilize according to the change process from pilot to
broad implementation, new routines and improvements of the pilot were suggested by the production engineers
even though they have been operating along these lines for years. But now as the pilot might go to broad
implementation, different interests initiated a change of the pilot from the process engineers.
During the interviews and the search conferences on the case of a new operational model, we observed that
the discussions generated a new common understanding among the informants about the pilot, the whole change
process. The method helped to clarify what the changes would mean in day to day operation, how they were
going to work and what the potential consequences could be. It also generated a new understanding of why
changes were proposed.
All these questions are important issues in change management and elements that can be discussed related to
organisational learning. Consequence analysis can be a useful change management and organisational learning
tool, if the traditional design and use of such analysis can be changed.

INTRODUCTION

The oil and gas industry is undergoing a fundamental

change in important business processes. The transition is made possible by new and powerful information

technology. Traditional work processes and organisational structures are challenged by more efficient
and integrated approaches to exploration and production. The new approaches reduce the impact of
traditional obstacleswhether they are geographical,

813

organisational or professionalto efficient use of an

organisations expertise knowledge in decision making
(Kaminski, D. 2004; Lauche, Sawaryn & Thorogood,
2006; Ringstad & Andersen, 2008).
Descriptions of the new approaches exist elsewhere
(e.g. Upstream technology 2007), and will not be
repeated here. The approaches can be subsumed under
the heading Integrated Operations (IO). Numerous
definitions of IO exist in the industry. In StatoilHydro
(2007) IO is defined as:
New work processes which use real time data to
improve the collaboration between disciplines, organisations, companies and locations to achieve safer,
better and faster decisions.
It is generally assumed that improved decision making processes in turn will lead to increased production,
less downtime, fewer irregularities, a reduced number of HSE-related incidents, and in general a more
efficient and streamlined operation.
The fundamental changes in work execution as a
result of IO are illustrated in Figure 1 and are briefly
described below:
The old assembly line work mode is seriously challenged by IO. More tasks can be performed in a
parallel fashion, thereby reducing total time consumption. From a decision making perspective,
parallel work execution means a more iterative and
relational process.
Multidisciplinary teamwork becomes more critical
as the availability of real time data increases, and
work is performed in a parallel fashion more or less
independently of physical location.
Real time data at different locations make it possible for personnel at these locations to cooperate
based on a shared and up-to-date description of the
operational situation.
Videoconferencing and readily access to data and
software tools reduce the need for specialists to be
on location. This increases the availability of expert
knowledge for operational units, and reduces the
time it takes to muster the experts.
The diverse and fundamental changes associated
with IO require a careful and deliberate impleSerial

Parallel

Single discipline

Multidiscipline
teams

Dependent of

Independent of

physicallocation
physical
location

physicallocation
physical
location

Decisionsbased
Decisions
based

Decisionsbased
Decisions
based

onexperience
on
experience

onrealtime
on
realtimedata
data

data

Figure 1.

Changes in work execution as a result of IO.

mentation strategy and adequate tools and methods.

A method that facilitate analysis and prediction across
a broad range of consequences categories are deemed
particularly useful.
However, many traditional consequence analysis
methodologies are concerned with one consequence
category (e.g. safety or cost) and/or are based on one
particular approach to data collection and analysis.
Although it would be possible to utilise a number
of different consequence analyses prior to any IO
implementation, it was decided to develop a new
methodology for consequence analysis particularly
suited for the purpose.
The new method should:
Be suited for analysis of a broad range of consequence categories
Be flexible (i.e. allow the analyst to use different
types of data and data collection methods, and the
method should be flexible across analyses objects
(e.g. a refinery and an offshore installation)
Involve personnel affected by IO in the analysis to
ensure participation in the change process.
The IO-program of StatoilHydro has chosen a strategy from pilot to broad implementation in the efforts
of achieving the visions of IO. A special exemplary
practice due to the IO characteristics in one of the
assets are chosen as a pilot. This practice is first evaluated to be defined a pilot. Then a CA is carried out
with conclusions and recommendations for a broader
implementation or not. The decision is to be taken by
the process owners.
This paper comprises two case studies exemplifying
the new method in use, and a general discussion of pros
and cons of the new method based on several analyses
performed in StatoilHydro in 2007. This discussion
gives an especial emphasis on how CA and OD might
be seen together, as two sides of the same process.
2

METHODS IN CONSEQUENCE ANALYSIS

In the following we present both a theoretically

approach or an attitude underlying the method e.g.
Actor Network Theory, the consequence categories
used as a basis for the CA, the structure of the method
and the practical data techniques used. This method is
developed and used in two pilot cases in the IO program of StatoilHydro, Steering of smart wells from
onshore at Snorre B, and New operation model at
Huldra Veslefrikk).
2.1 Actor network theory-identifying of actants
and controversies
In the IO-case or IO-pilot of Steering of smart wells
from onshore from the field Snorre B in StatoilHydro

814

ASA, SINTEF used a new approach to a CA method,

named Actor Network Theory (Latour, 1986), based
on Science and Technology studies (STS), since this
pilot is very much about the development of a new
complex technology, where as we will see, there might
be a lot of different technologically solutions to this
issue of smart wells.
This pilot was also about the complex interplay
between technology and organization, a seamless
web (Callone, 1986) of how to use and operate this
technology e.g. a network of different actants human
and nonhuman and how they chain in different heterogeneous engineering (Callone, 1986). To study
the local community of practice, (Levy and Venge,
1985), their interactions, negotiations and struggles
more in depth where this technology is in use, gives
an important input to the understanding of the pro et
contras of the pilot, and the potential broader HSE
consequences of such a pilot. The case showed that
technology was not yet frozen as the CA started. In
opposite the work on the CA makes it develops further.
2.2

CA methodvisualization of consequence
categories

A basis was also to identify both positive and negative

consequences related to the categories organization and management, personnel and competence,
operations and regularity, HSE, economy and
company reputation, which are a broader set of
categories than normally in CA.
This figure illustrates a linearity or a cause and
effect chain between the main factors used as an analytically tool. In studying the communities of practices
or real life in these cases, we see this is of course
more a messy matter (Law, 2004). One of the main
activities of researchers is to tidy up the mess, and linearity in-between categories might be one way to tidy
up. The linearity was the basis for the methods further
procedure. But first as a starting point to identify the
different aspects of potentially consequences within
these categories, it is important to identify the most
important groups of actants (Bijker & Pinch, 1985)
participating in the pilot studied. Then the most obvious controversies are important to identify. A quick
visualisation of these is often useful to make as a draft
to be changed as the analysis goes on. The usefulness
of identifying the controversies at an early stage is also
to early be able to investigate whether there is a connection between the controversy and the risk level of
the potentially consequences.
The further methods follows a well known phase
divided, linear, stepwise procedure known in many
analysis, evaluations and change programs.
1. Identification of groups of actants and main controversies.

2. Qualitative data collection, interviews and search

conference.
3. Use of a consequence matrix to help sort raw
data based on the factor categories organization
and management , personnel and competence,
operations and regularities, HSE, economy
and company reputation.
4. Analysis of datausing ANT-analysis, cluster
analysis, chains of argumentations.
5. Evaluation of risk related to the found negative
consequences vs. positive consequences.
6. Conclusions and suggestions.
2.3

Data collectioninterviews and search

conferences

SINTEF further developed the CA method in

order to evaluate a new integrated operation model
(IO-model) to be implemented in StatoilHydro ASA
Huldra Veslefrikk organisation. The method aimed
to identify potential positive and negative consequences related to implementation of the new operation model according to the five consequence categories mentioned above (fig. 2).
In both cases qualitative data were collected through
document studies, thirty individual interviews (Smart
wells, Snorre B), seven individual interviews (New
operation model, Huldra Veslefrikk) and one search
conference (Huldra Veslefrikk), (e.g. Emery & Purser,
1996) with relevant personnel form the Huldra Veslefrikk organization. Search conferences were not effectuated in the Smart Well case due to the initially high
controversy about this pilot.
The interviews could be performed either with individual informants or in groups of people. The choice
depends on how important are the controversies, and
how complex is the understanding of the operational
practice and the communities of practice that follows
that. A combination might also be a good solution.

Figure 2. Visualization of the consequence categories as a

basis for the analysis.

815

IO-project reports and descriptions of the new proposed IO-models in was the basis for the document
studies in both cases. As a basis for the interviews and
the search conferences we used the proposed change
measures needed to implement the new model as an
interview guide. In addition we initially ask for the
history of the initialisation of the pilot to get access to
the main actants and controversies, and to get in touch
with the most important arguments.
The group interviews have the basic aim to gather
information, and might be less time-consuming than
individual interviews, but might get access to more
supervision information. The search conference as
such is more a technique to create an arena for a
common dialogue on especial issues. A combination of these techniques is often been seen to be
fruitful.
In the IO change processes we have seen conflicting interests between management representatives and
trade unions. The search conference can be a useful tool in order to overcome these change process
barriers. The search conference can create openness among the participants (show that things are
what they appear to be); create an understanding of a
shared field (the people present can see they are in the
same world/situation); create psychological similarity
among the representatives; and it can generate a mutual
trust between parties. All these elements are found to
be important in order to achieve effective communication within and between groups (Asch, 1952), and in
this case to bring the planned change process forward
in a constructive direction.
2.4

Use of consequence matrix

In order to sort hypothetical negative and positive consequences after implementation of the suggested pilot,
we used a matrix to help us sort the informants statements within the categories organization and management, personnel and competence, operations
and regularity, HSE, economy and company
reputation. To the positive consequences we tried to
describe which presumptions underlying these statements, and to the challenges found we tried to suggest
compensating actions.
New ICT (Information and Communication Technology), with the use of large screens gives new
possibilities for search conferences. In group interviews and in the search conferences these matrixes
might be collective created, showed on a large screen,
which might give a good enthusiasm, participation
and founding of the results. In a case like this there
will be many arguments and the matrixes gives a nice
way to tidy the messy arguments and easily give an
overview. The concurrent production of this matrix in
a search conference might in addition be timesaving.
A further step might be to use the search conference

and the consequence matrix to also start the analysis.

Here is the point where employees often feel the participation ends, which creates a situation of resistance
at the time of implementation.
2.5

Analysing data

To analyse data, one of the methodologically starting points was to find the controversies and paradoxes
about the smart well technology, and to identify the
different groups of actants that are involved in the
controversies. By identifying the different controversies one also identifies the interests that are connected
to the controversies, and the constellations of interests that the different actants are chained in. Interests
are to be seen as the driving forces for changes.
Interests are what makes things happen both in a positive and a negative way, e.g. interests are also what
make things not happen. If one wants to understand
the OD aspects of a CA, one has to understand the
main interests. And if one want to do Organizational
Change one has to be able to play with the main interests or to be able to play the game, to chain in with the
different interests in different enrolements and translations (Latour, 1986) to make a strong enough chain
to be able to do Change Management, if not it is all in
vane.
Part of the analysis was also to describe which
presumptions underlying the positive consequences
found, and to suggest compensating actions to the
challenges found. The main stakeholder in the analysis is in these cases SINTEF. Consequence analysis
is something in-between an evaluation and scenario
thinking, and trained skilled methodological and analytical skills are of course required. But a higher degree
of participation in the analysis, and to test out the
analysis might be a fruitful idea, and with the search
conference as a tool, a possibility that is not so far away.
But the last responsibility for the analysis should be the
action researchers.
In addition to identify the different aspects of potentially consequences of the pilot mentioned above,
positive as negative, the CA has to do a ranging of
the different arguments by importance. E.g. by risk
level or sometimes by interests (As seen in figure 4).
One way might be to find what argumentations and
chain of argumentations that are used by visualizing
the arguments by cluster analysis. We often end up
with only a few central arguments, as the basis for
conclusion.
The cluster analysis aimed to find clusters in
statements regarding proposed negative consequences
related to one or several IO-model measures. As
a result it was easier to see how several IO-model
measures could cause interaction effects (e.g. severe
negative consequences) within the different categories
shown in figure 2.

816

The clusters of the negative consequences were then

risk evaluated, based on the risk analysis tool HAZID
(Hazard Identification). The HAZID tool defines risk
as probability multiplied with consequence. A risk
level for each consequence cluster was established
by a qualitative evaluation of how probably it was
for each cluster to occur, and how critical it would
be, e.g. how large the interests are concerned to this
consequence.
All steps in this design, except for the risk evaluation of clusters, were carried out in close cooperation with our informants and the StatoilHydro
IO-development team.
The extended focus on employee involvement
through interviews and search conferences must be
characterized as a relatively new approach within
CA-designs. Also the focus on analysing consequences in a broader perspective than just HSE must
be characterized as new, as definitively the approach
of ANT is CA. The effects of this kind of approach
will be discussed further in this paper.
The method as so contains of well known elements,
but the newness in the method is in the combination of
these well known elements as ANT, interviews, matrix,
search conference, cluster analysis, etc.

THE USE OF CONSEQUENCE ANALYSIS

(CA) DATA IN ORGANIZATIONAL
DEVELOPMENT (OD), IN STATOILHYDROS
IO PROGRAM

In the following we describe two different cases based

on the methodology we described in chapter 2.
3.1

Case 1: Implementation of a new integrated

operation model in StatoilHydro ASA Huldra
Veslefrikk organization

The StatoilHydro installations Huldra and Veslefrikk

were classified as tail production installations,
meaning they were in the last phase of production
initiated to prolong the economic lifetime of a field.
This situation can represent rising production costs
and potential lower profitability. In order to obtain the
profitability, the Huldra Veslefrik organization had to
increase the efficiency of the operations and to cut
administrative costs. Implementation of IO was then
seen as a solution, and the organization became a part
of StatoilHydros pilot program for implementation of
IO-models.
An organizational development (OD) process was
started to find a suitable IO-model for the organization. As a basis for the model design, they
emphasized the following design criteria for the new
IO-model:

1. Administrative work tasks offshore to be transferred

to the onshore organization
2. To make the onshore support organization more
operative (e.g. manning the onshore support team
with offshore personnel in rotation)
3. To make the offshore organization more operative (e.g. more time spent on operations, less on
administrative tasks)
4. To enhance handovers between shifts offshore by
improved planning onshore
The OD-process was carried out by using a
method called CORD-MTO (Coordinated Offshore operation and maintenance Research and
DevelopmentMan-Technology-Organization) as a
basis. The process leading up to a proposed IO-model
turned out to be filled with conflicts between the
project management group and labour union representatives. This was mostly due to change management
issues and what the labour unions characterized as
an unfair OD process. We, as an independent part
of the process, also observed a kind of uncertainty
among employees about how the new organization
would look like, and what consequences the structural
changes would have for each individual employee. We
have observed this kind of change anxiety in many
change processes, but in this case we observed that
the process of carrying out a CA, we believe had an
unexpected effect upon this change anxiety.
We observed that the CA method used had qualities beyond just evaluation of consequences. During
the search conferences we observed that the discussions generated a new common understanding among
the informants (employees) about the whole change
process and the new proposed operation model. The
method helped to clarify what the changes would
mean in day to day operation, how they were going
to work and what the potential consequences could
be. It also generated a new understanding of why
changes were proposed. All these issues are important
in change management, and they are elements that can
be discussed related to organisational learning. CA can
therefore be seen as a useful change management and
organizational learning tool if the traditional design
and use of such analysis can be changed.
3.2

Case 2: The pilot Steering of smart wells

from onshore in StatoilHydro ASA

A meeting in the W& D (well and srilling) network

19.11.04 decided that with planning of new fields it
should be prepared for the implementation of down
whole equipment or DIACS valve, e.g. smart wells.
For existing fields the same policy is decided for
planning of new wells. Deviations from this should
be followed by economically calculations. This pilot
is about a potential implementation of smart wells

817

as a part of an IO strategy from pilot to broad

participation, in existing wells in former StatoilHydro. The pilot is named Steering of smart wells from
onshore. The pilot takes place in the field Snorre B,
which came from former Hydro into StatoilHydro.
Originally Snorre B with its technologically inventions came from the former small oil company Saga.
Through performance of a Consequence analysis
SINTEF should help StatoilHydro in take good decisions whether the pilot should be recommended for a
broader implementation or not.
StatoilHydro had about 1015% of the smart wells
worldwide. In December 2006 we found 48 completions of smart wells with altogether 147 DIACS valves.
There is an increasing trend in smart well completion
in the company, and about 100 smart wells with 320
valves over 25 fields were anticipated in 2010.
One of the conclusions of our CA was that there are
no safety consequences because the DIACS vales are
not part of the barrier system. The main consequences
are of potential economically gains in production optimalisation, and potentially changes for the production
engineers if they are changing from todays reactive
operation to a more proactive operation with the use
of realtime data. More personnel resources of the production controllers are needed and it might be a more
24 hrs operation in one or a way.
We mapped the present practice in 6 different assets
(Heidrun, Veslefrikk, Gullfaks hovudfelt, Snorre A,
Gullfaks satelitter and Visund) to see the gap between
the pilot and the present practice in these assets. To
main way of operating smart wells was identified, e.g.
a manual way of operating as we see it in Gullfaks hovudfelt, Snorre A, Veslefrikk, Heidrun. In this solution
the production engineers and the supplier has to travel
offshore to operate the DIACS physically. There is a
common understanding that this is not a good solution. The main controversies are between the pilot of
Snorre B or the solution as they have in Norne, Visund,
Gullfaks.
In the analysis of all the potential consequences
we soon realised that we need to make a distinction
between operations (who is pushing the button) vs.
steering (who is planning and initializing the process)
due to an unclear linking between language or term
and practice, and to thereby avoid misunderstandings.
As a premise for the further analysis it is given as a presumption that onshore is always steering anyhow, the
competence for that is onshore in Drilling & Well and
in Petroleumtechnology. The controversy is weather
SCR (Sentral Controle Room) or the productions engineers/production controllers onshore is the one who
should operate the DIACS valves or if it should done
by a supplier onshore as in the pilot of Snorre B.
Their largest controversy in this pilot is who shall
operate the DIACS valves, who shall push the button, SCR offshore or onshore. And connotated to that

Figure 3. Overview of the main alternatives for smart wells

by existing assets in StatoilHydro.

Figure 4. Main controversies in the analysis of the smart

well pilot.

controversy is the question of choosing and developing

of technology, which solution to choose. Different
technology might support different operational solution e.g. who can push the buttons for operating the
smart wells.
And when it goes from a single autonomous case
to a pilot with possible consequences for other assets
the chain of interests are getting larger. The two main
technologically alternatives are between to digital variants, one with the integration in the OS (operation
station) system and operation from SCR. Many of the
assets want to go in this direction. The alternative of
the pilot is to have a separate operation system which
is used now for the pilot. Operation today is done by
the supplier Well dynamics which has to travel from
their office to StatoilHydro to operate the valves which
influenced the time of reaction from decision to operation if that is important. One of the most important
consequences against an integration in OS is the potential possibilities for external hacking which one avoids

818

with a separate control system as in the pilot. But security is said to be well handled at the statoil@plant.
It also involves larger development cost to integrate
than the separate solution of the pilot. The pilot has a
script that makes an operation from onshore preferred.
An integration in OS opens for a symmetry between
on and offshore operations, and thereby might conserve status quoe, as far as todays situation on who
should operate the valves, whether the pilot might push
a change.
As the CA started a discussion within the pilot
weather the pilot initially was good enough evaluated
to become a pilot or not. While the interviews came
about the production engineers then starts to create
suggestions of what can be changed in the pilot as
they realize that this might be the reality for many colleagues in other assets, and that theirs practice might
be the standardized proactive, even tough they have
not done anything to improve this or to come with the
same suggestions the two three years in-between now
and after the pilot was evaluated.
4

FROM PILOT TO BROAD

IMPLEMENTATION AS A CHANGE
STRATEGY

One of the main strategies to achieve their aims of IO

in StatoilHydro has been to define different locally
existing practices which contains good IO characteristics as a pilot to be considerer a broad implantation
in the other assets after first an evaluation and an then
a broader consequence analysis. The pilot of smart
wells in Snorre B was a locally initiative and a concept that was decided when the field was developed,
that we can see from the choice of concept of platform. Here we see a top down strategy, e.g. the IO
initiative meets a locally bottom up initiative developed in Snorre B. There is a huge variety in practices
among the assets due to local autonomy and different
history, different field caracteristics. When making
such connections with local and more global change
strategies it is important to well inform the pilot about
its chosen status so that everybody knows, to avoid
the killing of locally commitment. This is also important to avoid that the local people dont fell they are
hostages for larger organizational changes in other
assets, into practices that might work well for them,
but they might anyhow create large resistance to these
changes in other fields, they are just not invented here
and does thereby not fit in, and might demand some
trouble to change locally even though they might have
been a successful and smooth practice elsewhere. If
the pilot is not locally enough anchored, the questions
will be posed if it is good enough evaluated locally
and thereby any argumentation to support a broad
implementation might be effect fully be stopped by

potential opponents as a political argument against the

planned changes and not an argument that is based in
professional discussions.
5

CONCLUSIONS

In this paper we argue that consequence analysis is to

be seen as a part of a planned organizational change
process. In fact the organizational change process
starts when the CA starts. Thereby CA and OU should
not be seen as separate parts.
Objective analyses of consequences do not exist. At
the time one starts interviewing about potential consequences of an action, different groups of actants are
starting to chain and to mobilize their common interests, as we see it in the smart well case, and the change
process starts.
The CA might better be seen as a part of a planned
organisational change program, and with trying to
achieve a good dialogue and a collaborate atmosphere
among the parts. It is as we see in the Huldra Veslefrikk
case not easy to achieve a good change process if the
process of analysis in advanced (CORD analysis) has
not followed a good participated process in advance;
it is very hard to achieve that later.
The best advice is to use the energy for change that
is to be found in the mobilizing and chaining of interests. One has to enroll important actants and chains
of important interests if not the OD program will be
in vain.
To succeed one has to understand the concrete operational challenges in the pilot, and the seamless web
of technology and organization and thus these needs
to be described and understood. The CA has one large
advantage dealing with this that OD programme rarely
interfere with. CA might contribute to make OD more
successful.
REFERENCES
Asch, S. (1952). Social psychology. Englewood Cliffs, NJ.:
Prentice-Hall.
Bijker, W.E., Hughes, T.P. Pinch, T. (1987). The Social
Construction of Technological Systems, The MIT Press,
Cambridge, Massachusetts.
Callone, M. (1986). Some elements of a sociology of translation: domestication of the scallops and the fisherman,
John Law (ed.) pp. 196229.
Emery, M. & Purser, R.E. (1996). The search conferencea
method for planning organizational change and community action. San Francisco: Jossey-Bass Publishers.
Latour, B. (1986). Science in Action, Harvard University
Press, Massachusetts.
Lauche, K., Sawaryn, S.J. & Thorogood, J.L. (2006). Capability development with remote drilling operations. Paper
presented at the SPE Intelligent Energy Conference and
Exhibition, Amsterdam, The Netherlands.

819

Lave, J. & Wenger, E. (1991). Situated Learning. Legitimate

peripheral participation. Cambridge University Press.
Law, J. (2004). After Method. Mess in social Science
Research. London, Routledge.
Kaminski, D. (2004). Remote real time operations centre for
geologically optimised productivity. Paper presented at
the AAPG International Conference, Cancun, Mexico.
Moltu, B. (2003). PhD Thesis, BPR in Norwegian! The
mangementconcept og Business Process Reengineering
(BPR) as a culturally praxis, (Nr 2004:81) NTNU, Trondheim, Norway.

Ringstad, A.J. & Andersen, K. (2007). Integrated operations

and the need for a balanced development of people, technology and organization. Paper presented at
the International Petroleum Technology Conference,
Dubai, UA.
StatoilHydro (2007). Integrated operations in StatoilHydro.
Monthly Newsletter, May 2007.
Upstream technology, Feb. 2007, pp. 3637. Interview
with Adolfo Henriquez, ManagerCorporate Initiative
Integrated Operations, StatoilHydro.

820

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Integrated operations and leadershipHow virtual cooperation influences

leadership practice
K. Skarholt & P. Nsje
SINTEF Technology and Society, Trondheim, Norway

V. Heps & A.S. Bye

StatoilHydro, Trondheim, Norway

ABSTRACT: This paper aims to discuss how the use of advanced information and communication technology
impacts leadership practice. The paper is based on a research study accomplished at the Kristin asset on the
Norwegian continental shelf. The technology we explore is Integrated Operations (IO), and how organizations
can benefit from using this kind of technology. We discuss the results of our study, focusing on virtual cooperation
among leadership teams located onshore and offshore in the Kristin organization. To date, some research on how
to succeed in virtual teams exists, but few studies explore leadership in virtual teams. The strength of this study
is the in-depth insight of how and why IO shapes the work practice of leaders and operators/technicians. So far,
few empirical research studies shed light on how IO functions and is experienced by the people involved. The
research has mostly focused on the theoretical models of IO.

INTRODUCTION

In this paper we discuss how virtual cooperation

through the use of Integrated Operations influences
leadership practice. The paper is based on a research
study conducted at the Kristin asset on the Norwegian
continental shelf.

1.1

Integrated operations

Today, several oil companies on the Norwegian continental shelf have implemented IO as a strategic tool to
achieve safe, reliable, and efficient operations. There
are a variety of concepts describing IO, also called
e-Operations and Smart operations. IO allows for a
tighter integration of offshore and onshore personnel,
operator companies, and service companies, by working with real-time data from the offshore installations.
The Norwegian Ministry of Petroleum and Energy
(in white paper no. 38) defines IO as: Use of information technology to change work processes to achieve
improved decisions, remote control of processes and
equipment, and to relocate functions and personnel to
a remote installation or an onshore facility. Thus, IO
is both a technological and an organizational issue,
focusing on the use of new and advanced technology as well as new work practices. The IO technology
implementation is not considered to be a major obstacle in StatoilHydro. The most challenging issue is to

develop new work practices and change management

Henriquez et al (2007).
How technology is able to coordinate and communicate tasks within virtual teams is of great importance.
The IO technology consists of high-quality video
conferencing, shared work spaces and data sharing
facilities. These arenas include so-called collaboration rooms (operation rooms) for rapid responses and
decision-making. The design includes video walls to
share information and involve people in discussions,
having eye contact with each other both onshore and
offshore. IO technology is characterized by vividness
and interactivity. According to Steuer (1992), vividness is the ability of a telecommunications medium
to produce a rich environment for the senses, which
means having a range of sensory input (voice, video,
eye contact, etc.), as well as depth of information bandwidth. Interactivity denotes the degree to which users
can influence form or content in their telecommunications medium. In our study we found that the use of
collaboration rooms creates the sense of being present
in a place different from ones physical location, i.e.,
the sense of being there.
1.2

Integrated operations at Kristin

To explore how IO shapes new leadership practices,

we discuss findings and knowledge acquired from a
research and development project undertaken in 2007.
The purpose of this project has been to describe IO in

821

the Kristin organization, focusing on work practices

concerning operation and maintenance at the platform
with onshore operational support. In addition, the
purpose has been working with organizational development, focusing on which organizational capabilities
Kristin ought to develop.
The Kristin asset, operated by StatoilHydro, is a
condensate and gas field on the Norwegian continental shelf. The platform is located 240 km offshore
from Trondheim, Norway, and production started in
November 2005. The Kristin organization wanted to
develop an IO mindset in order to be able to operate
the platform with a minimum of people on board for
safety reasons and maximize production and operational efficiency, as well as keeping the platform in
optimal technical condition. Compared to large offshore installations having an operation crew of 200
employees, there are only 31 employees working on
the Kristin platform during any shift period. This lean
organization influences the communication on board,
and the communication between offshore and onshore
personnel.
The Kristin organization has two management
teams, one onshore and one offshore, each located in a
collaboration room. There are continuous video links
onshore and offshore, so both management teams can
see each other at all times during the day.
1.3

Virtual leadership teams

How to succeed in virtual teams has been quite well

described in the literature, but there are few studies
exploring leadership in virtual teams. As with traditional team work, research on virtual teams has
demonstrated the importance of effective communication and coordination within virtual teams (Lipnack
and Stamps, 1997). Virtual teams are often characterized by high levels of autonomy rather than direct
control, which will affect leadership practice.
Lipnack & Stamps (1997) define a virtual team as:
A group of people who interact through interdependent tasks guided by a common purpose that works
across space, time and organizational boundaries with
links strengthened by webs of communication technologies. IO is about how members of a geographically distributed organization (offshore and onshore),
participate, communicate and coordinate their work
through information technology.
In this paper, we focus on two types of virtual
management concerning the cooperation between the
onshore and offshore organization at Kristin. First,
this paper explores how the use of IO technology by
virtual management teams influences the cooperation
and communication between the offshore and onshore
management teams. Second, we explore the virtual
cooperation between the onshore technical engineers
and the offshore operators.

Olsen & Olsen (2000) describe which elements

are crucial for success in virtual team work, such
as: the sharing of knowledge, coupling in work, the
need for collaboration to solve tasks, and the need for
technology that effectively supports communication
and decision-making. We explore how these elements
affect cooperation and outcomes in the organization
we study.

METHOD

The empirical data for this study is comprised of observations and interviews. We have interviewed managers
both onshore and offshore, operators offshore within
all functions, and disciplines represented at the platform (electro, mechanic, automation, instrument,
process, among others) and technical lead engineers
onshore within most of the disciplines. The collected
material comprises semi-structured interviews with a
total of 69 informants, as well as extensive participating observations both onshore and offshore. Analyses
of interviews were conducted based on the principles
of grounded methodology (Strauss and Corbin, 1998)
with qualitative coding techniques. Examples of participating observations are being present at formal and
informal meetings in the collaboration rooms both
onshore and offshore, as well as following the work
of the operators when they were out in the process
plant doing maintenance and operation tasks.
The research approach has been co-generative
learning (Elden & Levin, 1991). The basic idea is
that practitioners and researchers create new practices
together and parallel to developing a deeper understanding of the research question in focus. Examples
of this close interaction between practitioners and
researcher in our study are as follows:
During the project period, the researchers and key
practitioners met on a regular basis (every month)
for working sessions. This was central to the development of the analysis of the IO work practice. At
work sessions, difficult issues could be discussed,
misunderstandings were sorted out, and findings
that needed interpretation were discussed.
By holding informal meetings, being embedded
during data collection, etc., the project contributed
to co-learning and reflection on work practices
between researchers and practitioners. A set of
shared notions and concepts was developed, and
thus also a higher awareness of critical organizational aspects.
In addition, the researchers presented and discussed
the project results closely with all people involved in
the projectoffshore operators, onshore and offshore
management, and onshore technical lead discipline

822

engineers. This had the effect of the researcher gaining

deep insight into peoples work practices.
In terms of methodology, project execution followed these steps:
1. Provisional mapping of functions, arenas and
relations. Prioritization of the most important
functions/arenas/relations. (Tools: Workshop with
stakeholders).
2. Collection of data. Information on arenas and relation used as collection guide. Evolving focus with
new information. (Tools: Observations, conversations, interviews).
3. Analysis of data for key findings and observations.
Conducted simultaneously with data collection.
4. Identification of important themes in the material,
using stakeholders to sound the importance and
secure ownership (Tools: Work meetings).
5. Suggesting short-term and long-term actions. Prioritizing actions together with management and
stakeholders. Presenting findings and actions for
management, stakeholders, and employees. (Tools:
Facilitating workshops).

DISCUSSION: INTEGRATED OPERATIONS

AND LEADERSHIP PRACTICE

Certain characteristics of an organizations leadership

practice will have an impact on the outcomes of virtual
cooperation. This will influence the quality of relations between the offshore and onshore organization.
Below we begin by describing the characteristics of the
leadership practice in the organization of study. Next,
we discuss the virtual cooperation and work practice
between the offshore and onshore organization: i) the
virtual management team, ii) the virtual support from
the onshore technical engineers.
3.1

Characteristics of leadership practice

We have explored the leadership at Kristin from a

relational perspective, and find that leadership is
characterized by a transformational and situational
leadership style.
Leadership philosophy and style will impact how
the offshore operators conduct their work, particularly
the way in which they are expected to lead themselves and take responsibility for the operations both
individually and collectively as a team. According to
Wadel (2005), organizational restructuring into flat
organizations and autonomous work teams means that
co-workers to a larger extent have to lead and support
each other. This change in roles and practice among
workers also changes the role of leadership. To explore
this, we have to understand leadership from a relational
perspective as well as from a situational perspective.

According to Bass (1990), transformational leadership means that a leader communicates a vision, which
is a reflection of how he or she defines an organizations goals and the values which will support it.
Transformational leaders know their employees and
inspire and motivate them to view the organizations
vision as their own (Bass and Avioli, 1994). Such
leadership occurs when one or more persons engage
with others in such a way that leaders and followers lift each other to higher levels of motivation. At
Kristin, the concept of integrated operationswhat it
really means for this organizationinvolved defining
a vision and values concerning how the work ought
to be performedon board, and between the offshore
and onshore personnel. Kristin has a lean and competent organization, where the operators/technicians in
the Operation & Maintenance (O&M) team offshore
possess expertise not necessarily found among their
superiors. The motivation at Kristin has been empowerment, which has affected the autonomous work of
the operators and the delegating leadership style.
Another leadership characteristic found at Kristin is
situational leadership, which means that leaders allow
for flexible solutions and actions adapted to the special conditions and situations in the organization. The
lean offshore organization at Kristin with few persons
within each discipline necessitates flexible problemsolving, which includes cooperation across disciplines
to support each others work. Situational leadership is
the opposite of trying to generalize or standardize work
practices and routines. Situational leadership theories
in organization studies presume that different leadership styles are better in different situations, and that
leaders must be flexible enough to adapt their style to
the situation in which they find themselves. A good
situational leader is one who can quickly adapt his or
her leadership style as the situation changes. Hersey
and Blanchard (1977) developed situational leadership
theory. They categorize leadership style according to
the amount of control exerted and support given in
terms of task and relationship behaviours; persuasive,
instructive, participating, and delegating behaviour.
Instructive behaviour means giving precise instructions and controlling execution. Persuasive behaviour
involves defining tasks, but seeking ideas and suggestions from the workers. A participating leadership
style is when the leader facilitates and takes part in
decisions. A delegating behaviour means that leaders delegate the responsibility for decision-making and
execution.
The level of competence among workers will influence whether a supportive or controlling leadership
behaviour is adopted. To lead personnel with low
degree of competence, a manager will define tasks
and supervise the employees closely. On the other
hand, leading highly skilled workers involves delegating tasks and responsibility, and the control lies with

823

the employees. High levels of expertise do not require

as much control from the manager. At the Kristin
platform, the leadership style is adaptive. Depending
on the situation and discipline, it is primarily characterized by a participating, persuasive, or delegating
management style. This is because of the highly skilled
personnel working at the platform. The O&M-crew is
directed primarily by the Operation supervisor, who
is their line manager, but they are also directed by
the technical lead engineers onshore. This is further
discussed in Chapters 3.2 and 3.3.

3.2

Virtual cooperation between the offshore

and onshore management teams

We have examined which kinds of work practices, benefits, and outcomes the Kristin leadership teams, both
offshore and onshore, have achieved by the use of integrated operations. First, we present and discuss how
the management teams actually work in the collaboration rooms. Then we discuss the benefits and how
they are achieved.
At Kristin, the management is organized as follows:
There are two management teams, one onshore and
one offshore, each located in a collaboration room.
The collaboration is supported by the use of video
conferencing and data sharing facilities, where both
management teams can see each other at all times
during the workday. Also, process data is online and
available at both locations and can be shared.
The offshore management team at Kristin is comprised of four managers: a Platform Manager, an Operation Supervisor (O&M), an Operations Engineer, and
a Hotel & Administration Supervisor (H&A). The
management team offshore manages maintenance and
operations in close collaboration with the management onshore. The onshore management is comprised
of a Platform Manager, an Operation Supervisor, an
Operation Engineer, and a Technical Support Supervisor. They share the collaboration room with some
technical engineers, who support operations and modifications offshore. Both the offshore and onshore
management uses the collaboration room on a permanent basis, as their office, and not as a meeting room
like several other assets on the Norwegian continental
shelf do.
The onshore management team is responsible for
giving day-to-day operational support to the offshore
organization, and for the planning of maintenance
programs and tasks on a long-term basis. This takes
place through formal daily meetings and through
informal and ad-hoc dialogue during the day. Each
morning the offshore and onshore management teams
have shared virtual meetings to inform and discuss the last 24 hours of operation and the next 24
hours to come. Here, representatives from different

technical units onshore attend to be informed about

the situation on the platform, and to give advice if
needed.
So, what are the benefits of this close, but still
virtual cooperation? First of all, StatoilHydro has estimated huge savings in operation costs over the first
year from integrated operations. This is a statement
from one of the platform managers at Kristin: Half
of the saving was due to the way we work. The other
half was due to having a quality process plant. Thus,
the reliability of and uptime at Kristin has been very
profitable.
The successful use of collaboration rooms has
affected the economic outcomes. One important
assumption is that everyone using the collaboration
rooms both offshore and onshore know each other well.
They meet in person at informal and formal meetings
onshore quite often, which strengthens the quality of
the virtual work. The random personal contact and the
fact that people know each other makes the distance
leadership more close (Maznevski & Chudoba, 2000).
This is an important criterion for success in virtual
cooperation.
Next, we find that peripheral awareness has developed at Kristin, which means that you develop a deep
understanding of what is going on at the platform.
The condition of peripheral awareness improves the
organizations capability to achieve rapid responses,
which in turn allows for more effective problemsolving and decision-making processes. One example
is the low number of backlog activities concerning
operations, maintenance, and HSE work. Compared
to large installations on the Norwegian continental
shelf, which have a high number of backlog activities,
at Kristin they have managed to handle these issues
effectively as a team.
The contextual differences (different work atmospheres, weather conditions, etc.) offshore and onshore
become less important by the use of collaboration
rooms. In the onshore collaboration room there are
several video walls showing pictures/video of the platform, the technical equipment, and the people working
there.
This daily and close communication creates a situation of shared situational awareness between the
onshore and offshore managers. Rosseau et al. (2004:
1415), Artman (2000), and Patrick and James (2004)
argue that there is an increasing interest in studying
team cognition, based on the fact that teamwork, or
working towards a shared goal, requires information
sharing and coordination. Shared situational awareness represents the overlap between team members, or
the degree to which team members possess the same
situational awareness or shared mental models. Shared
mental models are . . . knowledge structures held by
members of a team that enable them to form accurate explanations and expectations for the task, and

824

Figure 1.

Different levels of integration.

in turn, to coordinate their actions and adapt their

behaviour to demands of the task and other team members (Cannon-Bowers et al 1993: 228 in French et al
2004).
Figure 1 above illustrates different levels of integration in the virtual communication. The highest level
of interaction is the social level.
The challenge is to enable human and technical
elements to work together as integrated units. Communication through the use of technology means
more than the transfer of knowledge and information. Interoperability must be present in each of the
four domains: physical, information, cognitive, and
social (Alberts & Hayes, 2005). Videoconferencing
requires interoperability on many levels; from the
physical (technological) level to the social (organizational) level. At Kristin, the integration between
the onshore and offshore organizations has reached
the social level. This means that the organization has
gained organizational improvements, such as situational awareness or shared understanding and efficient
decision-making processes.
Shared understanding/shared situational awareness
has a significant impact on the ability of teams to
coordinate their work and perform well. Shared understanding affects performance in several ways, such as
predicting the behaviors of team members, increasing satisfaction and motivation, and taking actions
that benefit the team and the outcomes (Hinds &
Weisband, 2003). In the absence of shared understanding, frustrations, conflicts and distrust can develop.
In virtual teams, shared understanding is more difficult to generate. At Kristin, the IO mindset and
technology has improved the ability to obtain shared
understanding.
Below are some statements which illustrate the
benefits of integrated operations:
The collaboration room enables access to important information, where we get to know about each
other tasks and an overall impression of the work
onshore and offshore. Thus, we perform the work as

a management team, and not only as individuals.

(Manager)
One important aspect with integrated operations
at Kristin is the informal communication that happens
16 hours every day. In the operation room I receive a
lot of useful information from the other managers who
are sharing this room with me. (Manager)
The platform management at Kristin expresses that
they aim at behaving as one management team, meaning that they want to co-ordinate problem-solving and
decision-making between shifts. Once a week, even
in their time off, the platform managers arrange a
telephone meeting to discuss and share opinions concerning operation plans and their execution. In this way
they develop hands-on knowledge regarding whats
going on at the Kristin platform, where tasks are being
followed up on and rapidly solved.
We [the management] is quite co-ordinated across
the shifts. We organize a telephone meeting every
Thursday: 1) a meeting among the platform managers,
and 2) a meeting among platform managers and the
O&M-supervisor. This is very special I think, I have
never experienced this leadership practice at other
offshore installations. (Manager)
Performing as an integrated management team, has
influenced the sharing of common values and philosophy concerning how to organize and run the platform. This has been beneficial in terms of operational
efficiency.
I find that values, norms and philosophy at
Kristin are common and shared between the platform
managers. (Manager)
The way we work must not be dependent on whos
at work. It matters how we work as a team. (Manager)
Their goals are consistent leadership behaviors and
to obtain co-ordinated management solutions across
shifts. Nevertheless, this can be challenging to achieve
across different shift periods. For example, there
are situations where the management team forgets to
inform the next shift about all the decisions taken, but
these are not critical decisions.

825

In summary, these are the capabilities or benefits

developed through management use of collaboration
rooms both offshore and onshore:
Efficient problem-solving and decision-making
processes
Common ground: shared situational awareness
Shared leadership practices
Shared values

they manage. One of the technical engineers put it this

way: For some of us, the collaboration room becomes
like a drug. What he means is that you become dependent on being present and available in the collaboration
room. If you are not present, you may miss important issues and discussions of what is going on during
the day.
3.3

In the above we have focused on the benefits of

virtual cooperation among leadership teams onshore
and offshore. We have also examined whether the
collaboration rooms can represent a barrier concerning leadership practice at the platform. The operators
working in the O&M-team offshore are located in an
office landscape next to the collaboration room. We
asked the operators if the co-location of managers
hampers or increases a managers availability.
We found that the managers themselves wish and
try to be available for the operators at any time during
the working day, as this statement illustrates:
My concern is that the management team should
be available at all times during the day, even though we
are located in the collaboration room. I understand if
some of the operators find that it makes us less available, but my wish is to be available as a manager.
(Manager)
Most of the operators are of the opinion that the
offshore management team is available, and feel free
to contact their managers whenever they need to. The
reasons for contacting management during a workday
mostly involve the need for signatures and approvals
of work orders (WO) and working permits (AT). The
operators find that management responds quickly.
Personally, I have no problems with the managers
sitting in the collaboration room. The managers have
told us that if the door is locked they are busy. If not,
we are welcome. (Operator)
Nevertheless, some operators find that the colocalization of managers may be a barrier for calling
upon the managers attention. In some situations, they
are unsure of whether or not they are disturbing them.
This is when the managers are in contact with the management team onshore, or if one of the mangers is
having a telephone conversation. The managers say
that they try to invite people in when they notice them
standing outside the room waiting.
Another challenge concerning virtual management
is that some of the managers spend more and more
time in virtual space (collaboration room). This influences how much time the managers offshore spend
out in the process plant where the workers spend most
of their time. This then influences the amount of time
spent together with the operators. Similarly, the leaders
onshore spend much time in the collaboration room,
and becomes less present and available to the workers

Virtual cooperation between the technical

management onshore and the operators
offshore

At Kristin, the operation and maintenance tasks performed by offshore operators are based on remote support from technical engineers onshore. Their function
is not the management of people, but the management
of technical tasks within operation and maintenance.
For example, there is one domain engineer within
the electro discipline who is planning and supporting the work of the electricians on the platform. This
engineer is a domain expert and system responsible.
A similar situation exists for the other disciplines on
board (mechanics, automation among others). He/she
remotely assists the operations performed on the platform on a daily and long term basis, such as the
planning and prioritizing of operation and maintenance tasks. The crew on the platform is very much
dependent on the skills and knowledge of these system
responsible engineers, and on their availability in the
daily decision-making and task-solving processes.
The work practice and virtual cooperation between
technical engineers onshore and operators offshore is
characterized by telephone meetings, telephone conversations, e-mails, and face-to-face cooperation on
the platform. Meetings rarely take place in the collaboration rooms. For example, the electricians and
mechanics on the platform have weekly telephone
meetings with their technical engineers onshore. In
addition, the technical engineers go offshore to Kristin
23 times a year, on average. This results in personnel onshore and offshore knowing each other well,
and they develop a shared situational awareness of the
operation and maintenance conditions.
We find a lot of shared characteristics between
the different disciplines, such as the close cooperation between operators within different disciplines and
technical engineers. The relation is characterized by
mutual trust, and they refer to each other as good colleagues:
I have a close cooperation with the operators
on Kristin. They are highly skilled, work independently, and know the platform very well. Im in daily
dialogue with them, and we have weekly telephone
meetings. Together we discuss technical challenges
and problems. (Technical engineer)
We are very satisfied with the technical support
from our discipline expert. We appreciate it when he

826

comes offshore. It creates a common understanding of

the technical problems that we are facing. (Operator)
The operators trust the technical engineers ability, experience, and knowledge to support their work
offshore. The engineers have demonstrated competency with the quality of work and behavior necessary
to accomplish production at the platform. According
to Jarvenpaa and Leidner (1999), teams that highly
trust each another tend to perform better. If trust
exists in the relationships it means that much of the
work involved in monitoring and controlling others
becomes less important (McEvily, 2003, pp. 9293),
and this reduces the transaction costs associated with
operations. . . . trust is the willingness to accept
vulnerability based on positive expectations about
anothers intention or behaviors . . . trust represents a
positive assumption about the motives and intentions
of another party, it allows people to economize on
information processing and safeguarding behaviors.
It can be difficult to manage and support people
you do not see. The daily and close relation between
the operators and the technical engineers encourage
positive trust relations. Their relation is characterized by a social level of interaction, (ref. Figure 1).
They know each other quite well (they have met
each other in person), and the technical engineers
are highly committed to the work they are supposed
to manage and support on the platform. These good
relations lead to efficient problem-solving and highquality performance of operations and maintenance on
board. Trusting relations between management and
workers lead to increased responsibility and power
to autonomous workgroups (Skarholt & Finnestrand,
2007). Mac Evily (2003) argues that the tie sustaining
trust becomes stronger because there are additional
dimensions and relational contents. In addition to
exchanging information and advice, friendships are
also developed. Thus, the trust element is the glue, or
the foundation for a flexible structure of communication and enrolment realized through virtual and boundary work (Heps, in Jemielniak & Kociatkiewicz (ed)
(2008).
We find that the technical management onshore
is characterized by a situational leadership style.
Examples of situational leadership are as follows:
Electro: Instructive on strategy, hands-on problemsolving using a participating style
Programmed maintenance: Instructive on strategy,
participating in execution
Well control: Instructive in steering, participating
in POG-activities (production optimizing goals)
The level of complexity concerning the execution of
tasks will influence leadership style; whether the style
is delegating, participating, persuasive, or instructive. The reason behind different management styles is

that some tasks offshore require more instruction than

others, such as well control, which is managed and
decided onshore by a production engineer. On the other
hand, tasks within electro are managed by close cooperation with onshore personnel and are characterized
by a participating behavior.
The overall impression is a qualified support from
the technical engineers onshore, but there are some
challenges. One challenge is that there is a delay in
bringing support to the platform, because of the vast
number of jobs these engineers have to deal with, both
on Kristin and on others platforms they are supporting. In addition, they are involved in many discussions
with contractors and suppliers about how to solve
technical issues. Nevertheless, problems that are not
critical can wait, while some problems need observation and evaluation across shifts. The operators express
an understanding for the discipline experts situation,
and, similarly, the discipline experts express a wish to
be able to respond more rapidly, so that the operators
are able to accomplish their tasks efficiently. This is an
example of how they mutually trust each other in their
efforts to obtain confident results regarding operation
and maintenance at Kristin.
Another challenge is that the technical support from
the onshore organization is very dependent on who
is performing it, because it is based on one persons
expertise. If the technical engineer does not manage to
provide the necessary support to the platform, this role
or function does not work. So, the system is fragile and
is dependent on extrovert, available, co-operative, and
highly skilled engineers.
We have asked the technical engineers onshore how
they experience the availability and involvement of
the management team onshore, located in operation
rooms. We find that the co-localization of managers in
some situations impedes involvement from the technical engineers. The close cooperation between offshore
and onshore management in some situations leads to
quick decisions where the engineers are not includes
as part of the decision-making loop. Thus, the collaboration room can be a barrier for involving the
experts. This is similar to what the operators offshore
experienced. Nevertheless, we find that in critical situations, when their expertise is necessary, the engineers
actively take part in discussions and solutions together
with the management team.
4

CONCLUSIONS

This paper explore on how integrated operations

have an impact on leadership practice, and how
virtual collaboration create integration among management teams and personnel onshore and offshore.
At Kristin, the concept of integrated operations and
the use of collaboration rooms have created shared

827

situational awareness, which is crucial for obtaining efficient problem-solving and decision-making
concerning safety, production and maintenance.
In our study, we find that IO enhances the
experience of integration and common understanding between the onshore and offshore organizations,
where the virtual contact through the use of collaboration is experienced as being in the same room. This
results in better and faster decisions, because both the
onshore and the offshore managements have in-depth
knowledge about the situations/problems.
The challenging aspects with the use of collaboration rooms is that it can impede the managers handson relationships with people outside this room, such
as the relations with the operators/technicians offshore
and the technical engineers onshore. Both groups have
expressed a wish for more involvement from their
management onshore and offshore in problem-solving
tasks.
Our focus has been on how organizations can benefit from the use of new and advanced technology.
The challenge is not the technology itself, but the
organizational aspects, such as developing hands-on
leadership practices, clear roles and tasks, common
goals, trust, and knowledge and skills. These elements
are essential for developing an efficient organization
with motivated and skilled employees and managers.
REFERENCES
Alberts, D.S. and Hayes, R.E. (2005), Power to the Edge,
Command and Control in the Information Age, CCRP
Publication Series.
Artman, H. (2000), Team Situation Assessment and Information Distribution, Ergonomics, 43, 11111128.
Bass, B.M. (1990), From transactional to transformational
leadership: Learning to share the vision. Organizational
Dynamics, (Winter), 1931.
Bass, B.M. and Avolio, B. (1994), Improving Organizational Effectiveness Through Transformational Leadership. Thousand Oaks, Calif.: Sage, 1994.
Eldon, M. and Levin, M. (1991), Co-generative learning: Bringing participation into action research. In
William Foote Whyte (Ed.), Participatory action research
(pp.127142). Newbury Park, CA: Sage.
French, H.T., Matthew, M.D. and Redden, E.S. Infantry
Situation awareness. In Banbury, S. and Temblay (ed)
(2004), A Cognitive Approach to Situation Awareness;
theory and application, Ashgate Publishing Company,
Burlington VT, USA.

Hinds, P.J. and Weisband, S.P. (2003), Knowledge Sharing

and Shared Understanding. In Gibson, C.B., Cohen, S.G.
(ed) (2003): Virtual Teams that Work, Creating Conditions
for Virtual Team Effectiveness, Jossey Bass.
Henriques, A. et al. (2007), Integrated Operations: how Statoil is managing the challenge of change. Management
focus, vol. 25.
Heps, V. (2008), Boundary-spanning practices and paradoxes related to trust among people and machines in a
high-tech oil and gas environment. In: Jemielniak, D. &
Kociatkiewicz (ed), Management Practices in High-Tech
Environments.
Hersey, P. and Blanchard, K.H. (1977), Management of organization behavior: utilizing human resources. (3rd ed.),
Prentice-Hall, Englewood Cliffs, NJ.
Jarvenpaa, S.L. and Leidner, D.E. Communication and trust
in Virtual Teams. Organization Science, 1999, 10(6),
791815.
Maznevski, M.L. and Chudoba, K.M. (2000), Bridging
Space over Time: Global Virtual Team Dynamics and
Effectiveness, Organization Science, 11(5), 473492.
Mayer, R.C., Davis, J.H. and Schoorman, F.D. (1995), An
Integrative Model of Organizational Trust. The Academy
of Management Review; 20,3.
McEvily, B., Perrone, V. and Zaheer, A. (2003), Trust as
an Organizing Principle. Organization Science. 14(1),
99103.
Patrick, J. and James, N., (2004), A Task-Oriented Perspective of Situation Awareness in Banbury, S. and Tremblay
(ed), A Cognitive Approach to Situation Awareness:
theory and application, Ashgate Publishing Company,
Burlington VT, USA.
Rousseau, R., Trembley, S. and Breton, R. (2004), Defining
and Modelling Situation Aeareness: A critical Review in:
Banbury, S. and Tremblay (ed), A Cognitive Approach
to Situation Awareness: theory and application, Ashgate
Publishing Company, Burling.
Skarholt, K. and Finnestrand, H.O. (2007), The influence
of trust in an industrial team organization: How different organizational cultures affect trust building and
maintenance. Paper for the 23rd EGOS Colloquium,
Vienna.
Steuer, J. (1992), Defining Virtual Reality: Dimensions
Determining Telepresence, Journal of Communications,
42(4), 7393.
Strauss, A.L. and Corbin, J. (1998), Basics of qualitative research: Techniques and procedures for developing
grounded theory (2nd ed.). Thousands Oaks, CA: Sage.
Wadel, Carl Cato (2005), Nr medarbeidere m lede hverandreorganisasjonsendring mot mer medarbeiderledelse.
Tidsskrift for Arbejdsliv, 7. rg, nr. 4.

828

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Outsourcing maintenance in services providers

J.F. Gmez, C. Parra & V. Gonzlez
Industrial Management PhD Program at the School of Engineering, University of Seville, Spain

A. Crespo & P. Moreu de Len

Industrial Management School of Engineering, University of Seville, Spain

ABSTRACT: This paper presents a framework about management in maintenance outsourcing in a service
provider company. It proposes key aspects for taking decisions in a well-established and controlled organization.
Cost is not the most important aspect to consider in outsourcing, the decision has to be a global and strategic
idea inside the company. Of course, not only the directors must take part, but also the technical personnel of
maintenance. We are trying to offer a basic guide to establish an outsourcing service, with guidelines and possible
evolution. It is based on a practical view of knowledge management over ten years of professional experience
focused in networks. Below there is a case study which demonstrates a methodology for decision-making and
shows how to optimize the organization without losing differing levels of knowledge. For this, we employ
quantitative and qualitative criteria to obtain a wide consensus and acceptance.

INTRODUCTION

The outsourcing in maintenance is a practice being

increasingly used (Elfing & Duening 1994), especially with services providers. Although the decision
to outsource is not a simple decision, it is a strategic
decision (Click & Duening 2005) for an organization, and as such, it should align itself with the
business to impact positively on the objectives of the
organization.
There are different strategic reasons for which people decide to execute processes of outsourcing. For
example many managers who are making a process
of outsourcing, believe that it is an assignment carried out of responsibility to manage a part of the
business with the supplier. Other motives are primarily economic, issues which endanger the control of
outsourcing.
Faced with this, it is advisable to continue the process guided by decision making-steps, to ensure the
outcome of outsourcing is properly reached in maintenance. In this document we will attempt to provide
a framework for guiding implantations of outsourcing
in service providers companies.
For this, we have structured this document in five
parts. In the first two paragraphs, points 2 and 3, we
begin with a base revision on outsourcing and maintenance in suppliers of services. After, in point 4 we
will develop the reference model that is proposed and
finish with a case study and conclusion.

OUTSOURCING

Outsourcing is defined as the delegation of business

functions totally or partially to another company along
with part of the administrative and operational control.
Therefore, it is established between two companies,
a supplier and a customer, a contractual relationship
governed by service agreements.
Mainly, with a process of outsourcing we are looking for the specialization in activities, not keys for
the organization (Elfing & Duening 1994), such as
systems, accounting, buildings, human resources,
call centres, engineering, logistics, etc. For which,
it can transfer resources which bore those functions
formerly.
The decision of outsourcing is a strategic decision
(Earl 1996), aimed at improving the objectives of the
organization:

Improving Quality
Improving Security
Reducing Cost
Optimizing Resources

Therefore, the organization should focus its efforts

in improving those functions that are a source of competitive advantages and more profitable to the core
business.
Outsourcings have several advantages and disadvantages (Alexander & Young 1996, Halvey & Melby

829

2005, Jharkharia & Shankarb 2005, Tho 2005), and

within the advantages we can list:
Reduction costs, at the same quality to employ a
supplier more specialized
Restructuring costs, changing fixed costs by variable costs in terms of services provided
Stimulates the local employment through contracts
with local firms
Obtaining rapid budget by selling assets
Improvement of quality, for higher specialization
Access to outside expert knowledge
Standardization and access to scale economies
Flushes resources for other purposes
Improves company focus
Improving management of difficult functions to
handle
Optimizing routine tasks
Share the risk by flexibility of demand with the
supplier company
Provides legal guarantee for services
Relationships developed between financial aspects
and levels of service
Starting point for changes in the organization
Speed through reengineering
We also have to consider potential risks and disadvantages which affect any outsourcing plan:
Unfulfilled or questionable expectations, for a
scenario developed to generate the process of
outsourcing.
Changes in the quality for breach of agreements on
services, either by the knowledge or capabilities of
the supplier company, or by errors in the definition
of the same company.
Loss of knowledge or skills through transfer to the
supplier, where it is more difficult to retain and
improve, this happens frequently.
Loss of control over the externalized functions,
source of learning for the internal staff.
Dependence by the supplier could cause adverse
consequences for the client (investments extraordinary).
Loss of security by transferred staff to the supplier,
by hoax and illegal transmission of knowledge and
information to the competence.
Public and internal opinion for outsource jobs to
another company.
Loss of motivation for staff involved in the service,
because it can create a feeling of alienation within
the client company and result in the staff feeling
their jobs are valueless.
Although, the decision about which activities are to
be outsourced, it is often described as the beginning of
the process, however, really the process should begin
much earlier, defining the mechanisms to start from

a stable situation where the organization is controlled,

and avoid a difficult management of change.
The maintenance outsourcing could be an advantage as in other businesses, with the order to devote
most of the internal efforts in the core processes, and
seek the specialization of external agents. Although, it
should be guided by three types of criteria: strategic,
technical and economic.
Organizations often outsource those activities
which have work patterns that are fluctuating in their
burden and performance, and then the maintenance
and especially within distribution networks that meet
this requirement.
Below, we describe the nature of maintenance in
companies of distribution services and consider them
for the decision of outsourcing.
3

MAINTENANCE IN SERVICES PROVIDERS

Maintenance is characterized as a highly complex

field inside business and involves various disciplines:
management, human resources, company economy,
security, and knowledge of the whole production
chain. Another consideration is that maintenance
activities are all times under pressure to reduce costs
more than valuing the benefits or damages that it
avoids (Carter 2001, Mitchell et al. 2002) to the
company. Show of this importance, it is the weight
of O&M activities in GDP, 9.4% in Spain (AEM
2005), and other international studies put it among
the 15%40% depending on sector (Mulcahy 1999,
Mobley 2002).
The concurrence of these disciplines implies that it
can be difficult to determine the appropriate decision
every time.
On the other hand, we can define a service provider
company as those that provide clients certain services that are supported and distributed by a network
infrastructure, such as gas companies, water, electricity, telecommunications, etc. This infrastructure is
often organized and composed of elements prepared in
hierarchical structures and replicated by areas of distribution (Fig. 1). These companies fulfill the following
characteristics:
1. Elements geographically dispersed and in conditions of environment not optimal
2. High number of interconnected elements
3. High number and classes customers
4. Hierarchical structure in networks with levels of
aggregation of customer service
5. The network is dynamic and suffers configurational
and operational changes
6. High needs of human resources and spares
In these types of enterprises, maintenance is a key
department (Earl 1994), by its contribution to look

830

SOURCE
Primary
Connections
Secundary
Connections
Tertiary
Connections
Customer
Link

Figure 1.

Infrastructure of a service provider company.

Figure 2.

after the interests or satisfy the needs of clients and

benefits of enterprises (Zhu et al. 2002).
There are five different types of maintaining to consider for these types of enterprises according to most
of the standards (Crespo 2007, Benoit 2006, Levitt
2003, Wireman 1991):
1. Corrective Maintenance.
2. Preventive Maintenance.
3. Predictive Maintenance.
And the most recently maintenances:
4. Proactive Maintenance, a set of activities designed
to detect and correct an incidence before it occurs
avoiding its effects within the network and in
services (Tanenbaum 1991).
5. Perfective Maintenance. In the spirit of continuous improvement (IEEE 1219 1993, UNE 66174
2003) In the spirit of continuing improvement it
is a set of projects to improve the performance of
the network using the knowledge of maintenance
(Kent 1990), also called Design-out Maintenance
(DOM) (Gelders & Pintelon 1988).
4

OUTSOURCING AND MANAGEMENT

IN SERVICES PROVIDERS

There are many standards models of processes, best

practices, and Information Technologies facilities
(Hammer 1990, Peters 1982), unfortunately no single model works in all conditions. Some of which
referencing will seek to take advantage:
EFQM (2006), TPM (Nakajima 1992) and ISO
9001 (2001), management by process and quality
ITIL (ITSMF 2007), e-business
eTOM (2007), network management
CMMI (2007), system management to evaluate
maturity of companies
The initial situation of the majority of organizations
in distribution services are characterized by:

Objectives of the management model.

Management of a high set of elements from a large

number of suppliers
High dedication to resolution of incidents
Operation not automated, manual, dedication to
repetitive tasks
Reactive management before occurrences
Network documentation in paper format
Absence of unique and updated inventory
The situation tends to become more complex: increasingly difficult to manage systems due to increased
volume and large geographic dispersion. Therefore,
this situation as a starting point for a process of
outsourcing is not the most appropriate.
Recommendation for outsourcing is to establish a
structured management model in maintenance, as a
redesign (Davenport 1993, Hammer 1993) based on
activities, and with objectives of ensuring service quality (Klein 1994), so that it facilitates decision-making
and finds points of improvement more quickly and easily. In sum, a model oriented to customers, processes
and services delivery (Fig. 2). Then, we develop our
model is intended to be a support for outsourcing in
maintenance, searching:
Balance between the fulfillment of internal and
external requirements; strategic, operational and
tactical (Kaplan and Norton 1996)
Transform resources and services in customer satisfaction, fulfilling specifications and, restrictions
on management and cost (Nakajima 1992)
Then the implantation of the reference model is
developed in the following six phases (Fig. 3):
4.1 Mission, objectives and responsibilities
The definition of European standard UNE 13306
(2001) about maintenance management is quite comprehensive, although we have to identify the mission
of maintenance to complete that definition according

831

outsourcing and maintenance, in search of efficiency,

through a balance between quality and cost. The implementation of a complete management system can
reduce between 1030% of the annual budget of maintenance (Crain 2003, PMRC 2004), highlighting the
main improvements upon cost and task control, vital
in the control of outsourcing.
In addition, based on Campbell and Jardine
(2001) and standards, we can consider that the minimum support systems for a Computerized System
Maintenance Management System (CMMS), also
called (MMIS) Maintenance Management Information System (Pintelon & Gelders 1992) are six (Fig. 3):

1 MISSION AND
OBJECTIVES

2 DEPARTMENT
STRATEGY
5 CHANGE
MANAGEMENT
3 PROCESSES
AND ACTIVITIES

4 CONTROL
SYSTEM

5 SELECTION
SUPPLIER

Figure 3.

Phases of the management model.

to the characteristics of the services providers companies: guarantee the service, to ensure the proper
functioning of the services supplied. And starting
from this mission, we define responsibilities to get
the objectives of the department.
4.2

Strategic maintenance outsourcing plan

This phase establishes a strategy to achieve the goal

and maintain a solid and operational network that
ensures services, according to requirements set by
the organization. Strategy is based on three interrelated viewpoints: improvement of services quality,
cost reduction and resources optimization.
4.3

Processes and activities definition

Set all necessary functions to carry out the strategies

and to structure them in processes and activities. To
categorize activities by their strategic value, as their
employment in cost and resources, as for their contribution to quality, taking into account the type of
knowledge in which is based and the control possibilities. In this point, it will determine whether the
implementation of these activities should be carried
out with internal or external resources (Fixler & Siegel
1999, Grossman & Helpman 2002). It is important not
to forget that within outsourcing costs, we can appreciate further preparation, implementation, maintenance
and completion.
4.4

Outsourcing knowledge and documentation

in maintenance control system

This is where it should establish maintenance control system and where it defines the way to assess

4.4.1 Inventory system

A correct management of the configuration within
a network ensures rigorous bulletins and organization (ITSMF 2007), which maintains a history about
the evolution of elements in a planned manner. So
it reduces the risk of loss of control, ageing and
variations in service quality.
4.4.2 Monitoring system
It is key point for proactivity in maintenance. It provides information in real time about status of the
network and services (Lee 1995) with the objective
to ensure maximum network availability (Yan 2003),
with the highest quality, in a rapid and effective
response for incidents, preventing potential problems
before they start (IMSCENTER 2007).
4.4.3 Activities and resources management system
Its mission is managing, planning and documenting
categorizing activities associated with human resources and infrastructure elements. Therefore it is characterized by encompassing knowledge management
of historical data as a source for managing problems
(symptom-cause-solution) and learning. This module is the integration between technology and social
variables: tasks, resources, personnel and organization (Earl 1994). Activities have to be automatic
by work-flow systems to provide increased levels
of availability, reliability, flexibility and speed of
services from the point of view of technical and
economic.
4.4.4 Integration module with the rest
of company systems
This module has an objective to allow impact analysis in maintenance from the point of view of business
operations and clients (Lee 2004). Interconnection
must be conducted and controlled, at least with the
following enterprise systems:
Economics system
Human resources management system
Logistics

832

 CRM, Customer Relationship Management

Documentary management system
Knowledge management system.

4.6

4.4.5 Balance scorecard in maintenance

and with other systems
The balance scorecard is a pillar in evaluation and control of compliance with the department objectives. Its
aims are alignment of the department with company
strategy; relates all activities, processes, systems and
resources all with the operational objectives and strategic (UNE 66174 2003, UNE 66175 2003, Kaplan &
Norton 1996).
To this end, it collects a coherent set of indicators,
financial, about business, customer relationship and
continuous improvement.
4.4.6 Expert support system for taking decisions
This system gives support to take decision with maximum information, to facilitate the achievement of
objectives (Davis 1988). The recommendation is to
be formed as a module that integrates:
1. Decision Support System for decision making
(DSS) (Turban 1988, Bui & Jarke 1984) through
scientific models based on all the information from
the systems.
2. Expert System (ESs) (Shu-Hsien 2005) to emulate
through artificial intelligence, human reasoning
like an expert.
Thismoduleappliesboth, informationmanagement,
as statistical models (Marple 1987) and simulations
to submit patterns and solutions to facilitate decision making in maintenance (Iserman 1984, Jardine &
Banjevic 2006).
4.5

Supplier selection

Once the reach of outsourcing is defined, from a stable

situation, it proceeds to supplier selection and the planning of outsourcing implementation. There are many
considerations to take into account during this negotiation to avoid the risks listed above, but the main point
is that it is a collaborative negotiation or win-to-win
process, with the supplier as a strategic partner. It is
advisable to guide suppliers to offer services based on
levels of their knowledge, and thus avoid the approach
to only reducing cost.
Aspects most favoured to select a supplier are:

Experience in a sector
Flexibility on demand for services
Confidence
Technical and economic solvency
Will of collaboration oriented to services as strategic
support
Transparency of suitable pricing

Management of changes

Planning correct transition is important, it is a learning phase oriented to the supplier for fulfilling agreed
service levels. On the other hand, to ensure business
continuity in outsourcing, it should also be considered
a transitional phase and, a possible reversion distinguishing if it occurs in the transitional phase, at any
time, or at the end of contract.
To work with an outsourcing model of these characteristics implies important changes for everyone,
especially those teams responsible that have to take
a much more participatory role in management.
5

A CASE STUDY IN A
TELECOMMUNICATIONS COMPANY

As an example, to simplify, we will only focus on

the outsourcing decision in a telecommunications
provider to evaluate importance of each activity by
its contribution towards the maintenance goals and to
decide that activities could be outsourced.
From a strategic point of view, (Kaplan & Norton
1996, Campbell & Jardine 2001, EFQM 2006), it
must abide by basic maintenance objectives, which
summarize in the following six categories:
1. Management
2. Economical
3. Production or business
4. Quality or related to customers
5. Security, environmental and ethics
6. Evolution and Improvement
On the other hand, from a tactical point of view
processes of the department should also be taken into
account: corrective, preventive, predictive, proactive
and perfective.
From an operational point of view, it has to consider
maintenance activities. to simplify the study, within
these activities, only the most important activities are
considered.
1. To manage incidents, all kinds of incidences
2. Monitoring alarms and status of network and
services
3. On demand activities, to support other internal
departments in field
4. Preventive activities
5. Predictive activities, analysis to avoid or minimize
future impacts
6. Perfective activities. Improvement plans or tasks
to optimize infrastructure and services
7. Logistics. Stores management and spares
8. Budget and human resources. To control of budget, resources, tools, staff, vehicles, etc. . .
9. Security. Tocontrolsecurity, healthandsafetyrisks
10. Documentation and compilation of processes,
procedures, reports, etc. . .

833

Table 1. Relative Importance between variables. Saaty scales.

1

Same

Weak
1/3
Slightly less

Strong
1/5
Less

Proven
1/7
Much less

Absolute
1/9
Absolute less

(1/2, 1/4, 1/6, 1/8, 2, 4, 6, 8) Intermediate values if it is

necessary

Table 2.

Index of comparisons randomly generated.

ICrandom 0 0,58 0,9 1,12 1,24 1,32 1,41 1,45

n

2 3

For decision-making, we rely on properties of the

AHP method (Saaty 1977, 1980, 1990) for decisions in
group (Dyer & Forman 1992) of selected maintenance
experts from several hierarchical levels. The Analytic
Hierarchy Process AHP is a methodology to synthesize a solution (a matrix) of a complex problem through
a breakdown in parts ordered hierarchically, quantifying and comparing variables in pair with a normalized
and reciprocal scale of relative importance (Tab. 1).
In the use of this method, it can use subjective values, which implies a degree of uncertainty or lack of
reliability. To measure reliability coefficient RC is
used ratio between rate of consistency IC of a comparisons array into pairs and value of the same index
of a comparisons array into pairs randomly generated
(Tab. 2). The reliability is sufficient if RC is smaller
than or equal to 0.10; otherwise, it must be reviewed
to improve its consistency.
RC =
IC =

IC
ICrandom

0.1

max n
n1

geometric mean (3.086) of the six individual values

(3 + 4 + 3 + 2 + 4 + 3) by the sum of the second
column in the figure 4 (14.5), and at the same manner
with the rest of the matrix.
Successive matrixes from activities compared
according to each strategic objective were developed
(with indices RC all less than 0.06, valid), with the
exception of cost, where we employ the activity budget
rate (quantitative criteria). These matrixes are multiplied by their respective eigenvector (W) to obtain in
one column this contribution (Tab. 3).
In short, it obtains the weights for activities in the
Table 4, multiplying each cell of the Table 3 by the
respective cell in the same column of the W vector
of the Figure 5. Then activities are ranked depending
on their importance in relation with the objectives of
maintenance:
1. Budget and Human Resources 17.70%
2. Documentation 13.96%
3. Predictive 13.58%
4. Perfective 12.15%
5. Monitoring 10.87%
6. Preventive 10.31%
Activities less valued are:
7. Security 7.00%
8. Manage Incident 6.84%
9. Logistic 5.11%

Quality

Cost

Production

Man agement

Security

Imp rovement

Quality

3.09

0.50

2.40

0.46

1.12

Cost

0.32

0.30

1.51

0.22

0.55

Production

1.99

3.37

3.36

0.93

1.70

Man agement

0.42

0.66

0.30

0.17

0.46

wij

Security

2.19

4.57

1.07

5.73

1.26

Improvement

0.89

1.82

0.59

2.18

0.79

6.8

14.5

3.8

16.2

3.6

6.1

(1)

(2)

So, the problem is hierarchically structuring with

criteria and alternatives, in three levels:

Figure 4. Matrix completed with the average of six individual comparisons.

Table 3.

1. Goal
2. Maintenance objectives as criteria
3. Activities as alternatives
For valuing objectives, it is used an expert group
poll with qualitative criteria depending on their strategic importance. Each technician of a six group compares them employing table 1 and after, the resulting
matrix (Fig. 5) is built weighing the average of individual values (Fig. 4), e.g. 0.21 (in the second cell
of the first row in the figure 5 is calculated dividing

834

Matrix of activities rates to each objective.

Quality

Cost

Production

Management

Security

Improvement

Manage Incident

0.117

0.058

0.076

0.057

0.056

0.040

Monitoring

0.094

0.104

0.107

0.068

0.109

0.145

On demand
activities

0.026

0.022

0.021

0.028

0.0266

0.027

Preventive

0.074

0.085

0.099

0.074

0.122

0.125

Predictive

0.166

0.109

0.121

0.121

0.115

0.185

Perfective

0.126

0.112

0.122

0.148

0.093

0.162

Logistics

0.039

0.069

0.072

0.031

0.039

0.050

Budget and
Human R.

0.147

0.310

0.27

0.285

0.110

0.073

Security

0.064

0.048

0.037

0.040

0.115

0.069

Documentation

0.148

0.083

0.074

0.149

0.214

0.123

knowledge in maintenance thanks to information

systems facilities, e-maintenance (Yua et al. 2003,
IMSCENTER 2007):

Table 4. Matrix of activities rates according to their importance in relation with strategic objectives.
Qual- Cost
ity

Produc- Manag- Secu- Improvtion

ement
rity
ement

Manage
incident 0.019 0.004 0.019
Monitoring 0.015 0.008 0.027
On demand
activities 0.004 0.002 0.005

0.004
0.004

0.016 0.006
0.032 0.023

0.002

0.008 0.004

Preventive

0.012 0.006 0.025

0.005

0.036 0.019

Predictive

0.026 0.008 0.031

0.008

0.034 0.029

Perfective
0.020 0.008 0.031
Logistics
0.006 0.005 0.019
Budget and
Human R. 0.023 0.023 0.070

0.009
0.002

0.027 0.025
0.012 0.008

0.018

0.032 0.011

Security
Documentation

0.010 0.004 0.009

0.002

0.034 0.011

0.023 0.006 0.019

0.009

0.063 0.019

wij

Quality

Cost

Production

Management

Security

Improvement

Quality

0,15

0,21

0,13

0,15

0,13

0,18

Cost

0,05

0,07

0,08

0,09

0,06

0,09

Production

0,29

0,23

0,27

0,21

0,26

0,28

Management

0,06

0,05

0,08

0,06

0,05

0,08

Security

0,32

0,31

0,28

0,35

0,28

0,21

Improvement

0,13

0,13

0,16

0,13

0,22

0,16

0,159

0,073

0,257

0,062

0,294

0,156

Facilitate manage agreements on service levels

and delivery service reports. Then IT contributes
to effectiveness and efficiency.
Orientation towards services more than elements of
infrastructure, searching for continuous improvement in services and processes to reduce costs and
times, and to improve value and quality.
The aim with this model is to increase decision
reliability with experience, and in accordance with
department strategy:
1. Improved organization and structure
2. Using a rational and logical analysis, it seeks a
solution for a complex problem with various alternatives in conflict and in conditions of uncertainty
[DIXON66]
3. Aligned with company strategy, it considers processes, objectives and activities
4. Employment qualitative criteria, to rationalize
intangible quality and value judgments from
experts to extract specialists knowledge
5. Promote positive attitudes towards improving
maintenance
6. Consensus in groups with different interests
7. Categorize alternatives
8. Improve interactively
9. Report processes for future developments
10. Easy to use and flexible with information available

Figure 5. Weights in strategic criteria (RC = 0.01579,

acceptable).

10. On demand activities 2.49%

This situation conducts processes of externalization
towards these last four routine and repetitive activities with not crucial importance to the core business.
The expert group feels motivated with this decision
for participating, and it is suggested to advance more
in outsourcing after a stable period externalizing:
Monitoring, at least first attention level
Preventive maintenance, guiding planning internally by predictive and perfective maintenance
6

CONCLUSION

This reference model has been implemented in at

least two companies in distribution of telecommunications services, and as such can be developed in a
high depth and customization for certain scenarios.
In addition, it is possible to increase control and

This method reduces time in decision making,

increases quality and security of the final decision,
and produces motivation and satisfaction with goals
and team work.
In conclusion, according to maintenance outsourcing in a service provider, it suggests that to compose
levels of externalisation progressively in time, increasing internal knowledge and control about activities
before to recruit once. That is, to make a partial
outsourcing:
with a flexible contract
guarantee business productivity through service
level agreement
devoting staff to manage contractual relationship
and monitor services
outsourcing should be guided primarily by strategic
criteria
But it should carry out the analysis with caution,
because in the case of outsourcing the level beyond
the norm, there is a point of Irreversibility of decision
where it would be impossible to react, this point
is where the procedure to prevent consequences
expressed would be unacceptable to act upon due to
time and resources.

835

REFERENCES
AEM, Asociacin Espaola de Mantenimiento. 2005. El
Mantenimiento en Espaa: Encuesta sobre su situacin
en las empresas espaolas.
Alexander M. & Young D. 1996. Strategic outsourcing. Long
Range Planning 29 (1): 116119.
Benot Iung 2006. CRAN Laboratory Research Team
PRODEMAS in Innovative Maintenance and Dependability. Nancy UniversityNancy Research Centre
for Automatic Control (CRAN). CNRS UMR 7039
(http://www.cran.uhp-nancy.fr).
Bourne M. & Neely A. 2003. Performance measurement
system interventions: the impact of parent company initiatives on success and failure. Journal of Operation and
Management.
Campbell J.D. & Jardine A. 2001. Maintenance excellence.
New York: Marcel Dekker. 2001.
Carter, Russell A. 2001. Shovel maintenance gains from
improved designs, tools and techniques. Elsevier Engineering Information.
Click R.L. & Duening T.N. 2005. Business Process
Outsourcing: The competitive Adventage. John Wiley &
Sons, Inc.
CMMI Product Team. Software Engineering Institute 2007.
CMMI for Development, Version 1. CMMI-DEV, V1.2,
CMU/SEI-2006-TR-008, ESC-TR-2006-008.
COBIT [Control Objectives for Information and related Technology] 1992. Objetivos de Control para la informacin
y Tecnologas relacionadas. Asociacin para la Auditora
y Control de Sistemas de Informacin, (ISACA, Information Systems Audit and Control Association), y el Instituto
de Administracin de las Tecnologas de la Informacin
(ITGI, IT Governance Institute).
Crespo M.A., Moreu de L.P. & Sanchez H.A. 2004. Ingeniera de Mantenimiento. Tcnicas y Mtodos de Aplicacin a la Fase Operativa de los Equipos. Aenor,
Espaa.
Crespo M.A. 2007. The Maintenance Management Framework. Models and Methods for Complex Systems Maintenance. Londres, Reino Unido. Springer.
Davenport T. 1993. Process innovation: Reengineering
work through Information Technology. Harvard Business
School Press.
Dixon J.R. 1966. Design engineering: inventiveness, analysis, and decision making. New York, McGraw-Hill,
Inc.
Dyer R.F. & Forman E.H. 1992. Group decision support with
the Analytic Hierarch Process. Decision Support Systems.
Earl M.J. 1994. The New and the Old of Business Process Redesign. Journal of Strategic Information Systems,
vol. 3.
Earl M.J. 1996. The Risks of Outsourcing IT. Sloan Management Review. 37. 2632.
Elfing T. & Baven G. 1994. Outsourcing technical services:
stages of development. Long Range Planning 27 (5):
4251.
EN 13306:2001. Maintenance Terminology. European Standard. CEN (European Committee for Standardization),
Brussels.
Fixler D.J. & Siegel D. 1999. Outsourcing and Productivity Growth in Services. Structural Change and Economic
Dynamics.

European Foundation for Quality Management. 2006. EFQM

Framework for Management of External Resources. By
EIPMEFQM.
Gelders L. & Pintelon L. 1988. Reliability and maintenance in: Doff, R.C. and Nof, S.J. (ed.), International
Encyclopedia of Robotics, Application and Automation,
Wiley, New York.
Goldratt E. 1997. Cadena Crtica. Ediciones Diaz de Santos.
Grossman G.M. & Helpman E. 2002. Integration versus Outsourcing in Industry Equilibrium. The Quarterly
Journal.
Halvey J.K. & Melby B.M. 2005. Information Technology Outsourcing Transactions: process, strategies and
contracts. John Wiley & Sons, Inc.
Hammer & Champy 1993. Reengineering the Corporation.
Harper Business.
Hammer M. 1990. Reengineering Work: Dont Automate
Obliterate. Harvard Business Review.
Ian Tho 2005. Managing the Risks of IT Outsourcing.
Elsevier Butterworth-Heinemann.
Intelligent Maintenance Centre 2007. www.imscenter.net.
Iserman R. 1984. Process fault detection based on modelling
and estimation methods. Automatica.
ITSMF, IT Service Management Forum 2007. ITIL v3.
Information Technology Infrastructure Library. ITIL v2.
Information Technology Infrastructure.
Jardine A.K.S., Lin D. & Banjevic D. 2006. A review
on machinery diagnostics and prognostics implementing condition based maintenance. Mech, Syst. Signal
Process.
Jharkharia S. & Shankarb R. 2005. Selection of logistics
service provider: An analytic network process (ANP)
approach. International Journal of Management Sciente,
Omega 35 (2007) 274289.
Kaplan, Robert S. & David P. Norton 1996. The Balanced
Scorecard: Translating Strategy Into Action. Boston, MA:
Harvard Business School Press.
Kent Allen 1990. Encyclopedia of Computer Science and
Technology. CRC Press.
Klein, M.M. 1994. The most fatal reengineering mistakes. Information strategy: The Executives J. 10(4)
2128.
Lee J. 1995. Machine perfomance monitoring and proactive maintenance in computer-integrated manufacturing:
review and perspective. International Journal of Computer
Integrating Manufacturing.
Lee J. 2004. Infotronics-based intelligent maintenance
system and its impacts to close-loop product life cycle systems. Proceedings of de IMS2004 International Conference on Intelligent Maintenance Systems, Arles, France.
Levitt Joel. 2003. Complete Guide to Preventive and Predictive Maintenance. Industrial Press.
M. Davis. 1988. Applied Decision Support. Prentice Hall,
Englewood Cliffs.
Marple S.L. 1987. Digital Spectra Analysis. Prentice
Hall.
Mike Crain 2003. The Role of CMMS. Industrial Technologies Northern Digital, Inc.
Mitchell Ed., Robson Andrew, Prabhu Vas B. 2002. The
Impact of Maintenance Practices on Operational and
Business Performance. Managerial Auditing Journal.
Mobley Keith 2002. An Introduction to Predictive Maintenance. Elsevier.

836

Mulcahy R. 1999. The CMMS technology revolutionwhy

Best-of Breed wil still be best. International Journal of
Maintenance and Asset Management.
Nakajima Seiichi 1992. Introduccion al TPM (Mantenimiento Productivo Total). Productivity Press.
Neely A.D., Gregory M. & Platts, K. 1995. Performance
Measurement System DesignA Literature Review and
Research Agenda. International Journal of Operations and
Production Management.
Peters T. & Waterman H.R. Jr. 1982. In Search of Excellence.
Pintelon L.M. & Gelders L.F. 1992. Maintenance management decision making. European Journal of Operational
Research.
Ren Yua, Benoit Iung, Herv!e Panetto 2003. A multi-agents
based E-maintenance system with case-based reasoning
decision support. Engineering Applications of Artificial
Intelligence 16: 321333.
Saaty T.L. 1977. A Scaling Method for Priorities in Hierarchical Structures. Journal of Mathematical Psychology,
15: 234281, 1977.
Saaty T.L. 1980. The Analytic Hierarchy Process. McGraw
Hill.
Saaty T.L. 1990. How to make a decision: The analytic
hierarchy process. European Journal of Operational
Research.
Shu-Hsien Liao 2005. Expert system methodologies and
applicationsa decade review from 1995 to 2004.
Elselvier. Expert Systems with Applications 28: 93103.

Tanenbaum, Andrew S. 1991. Computer Networks. Ed.

Prentice-Hall.
The Institute of Electrical and Electronics Engineers. Inc.
1993. IEEE 1219. Standard for Software Maintenance.
The Plant Maintenance Resource Center 2004. CMMS Implementation Survey Results2004. The Plant Maintenance
Resource Center.
Tung Bui & Matthias Jarke 1984. A DSS for cooperative
multiple criteria group decision making. STERN School
of Business, Working Paper Series IS-84-45.
Turban E. 1988. Decision Support and Expert Systems:
Managerial Perspectives. New York: Macmillan.
UNE 66174 2003. Guide for the assessment of quality management system according to UNE-EN ISO 9004:2000
standard. Tools and plans for improvement. UNE.
UNE 66175 2003. Systems of Indicators. UNE.
UNE-EN ISO 9001:2000. Quality management systems
Requirements. International Organization for Standardization.
Wireman T. 1991. Total Productive Maintenance. Industrial
Press.
Yan S.K. 2003. A condition-based failure prediction and
processing-scheme for preventive maintenance. IEEE
Transaction on Reliability.
Zhu G., Gelders L. & Pintelon L. 2002. Object/objectiveoriented maintenance management. Journal of quality in
maintenance engineering.

837

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Revising rules and reviving knowledge in the Norwegian railway system

H.C. Blakstad & R. Rosness
SINTEF Technology and Society, Trondheim, Norway

J. Hovden
Department of Industrial Economics and Technology Management,
Norwegian University of Science and Technology (NTNU), Trondheim, Norway

ABSTRACT: This paper presents and discusses four safety rule modification processes in the Norwegian
railway system. It focuses upon the impact from the processes upon railway knowledge and in particular the
ambitions to change from predominantly experience based prescriptive rules towards risk based outcome oriented
rules, i.e. a deductive top-down approach to rule development.
The cases met this challenge with an inductive bottom-up approach to rule development, a strategy given the
name reverse invention. Discussions about the new approach and the processes of reverse invention stimulated
inquiries into railway knowledge that revived this knowledge. It remained uncertain whether the inquires resulted
in actual new knowledge. The new approach also stimulated a reduction of relational and contextual elements
of the railway knowledge. According to theory these elements are important for the ability to decode theoretical
knowledge and to judge its relevance for future use.

1.1

INTRODUCTION

Rules constitute an important record of the organizations learning about its operational dangers (Reason,
1997; Reason et al. 1998). Hale (1990) argues that
the greatest value of safety rules lies in the process of
actually finding out and writing down the rules. The
organization can retain this value by treating the rules
as a living repository of lessons learned in the life of
the system. Accordingly, safety rules are not only a
result of a learning process, they can be seen as a part
of that process.
There is limited scientific knowledge about performing safety rule modifications (Hale & al., 2003).
This also implies that there is limited scientific knowledge of the close relationship between safety rules and
knowledge of activities and related risks in a regulated
area and the possible impact from rule modification
upon this knowledge.
The purpose of this paper is to explore how safety
rule modification can influence organizational knowledge about operational dangers. It presents results
from a case study of four safety rule modification
processes in the Norwegian railway system. Special
attention is given to consequences of the ambition
to change from predominantly experience based prescriptive rules towards risk based outcome oriented
rules.

Safety rules and knowledge

Safety rules have been directed at control of known

risks and can be seen as a result of the knowledge
at the time of the rule making. In Baumards terminology they represent explicit collective knowledge
(Baumard, 1999). However, there is a general concern that it is neither possible nor wise to transfer all
lessons learned into explicit collective knowledge and
rules (Baumard, 1999; Hale, 1990; Rasmussen, 1997;
Reason, 1997).
Furthermore, the knowledge transformed into rules
can differ. Many traditional safety systems have used
safety rules providing detailed prescriptions to the
operative staff of what to do in response to predicted
situations or requirements to the states of the system
(Hale, 1990). The knowledge base for the development of such prescriptive rules were usually extensive
knowledge of the systems functioning combined with
knowledge derived from practical experiences with
accidents and dangerous events (Rasmussen, 1997).
Hence, the prescriptive and experience based safety
rules can represent an explicit collection of information about what to do under certain conditions based
on knowledge derived from lessons learned in the past
(Baumard, 1999; Hale, 1990; Rasmussen, 1997).
There is now an increased use of rules for safety
management and outcomes. When found necessary,
these are supplemented with lower level rules to form

839

a dynamic hierarchy of safety rules, see for instance

Hale & al. (1997), Hovden (1998). Risk assessment is
often required to decide upon the necessity of rules.
Such an approach to the development of safety rules
can be seen as a deductive and risk based top-down
approach to rule development (Hovden, 1998).
The change from experience based prescriptive
rules towards rules for safety management and outcomes based on risk assessments represents a change
in the knowledge base for the rules. This implies a
change in the attention and selection of knowledge
that will be considered relevant and expressed through
the rules, i.e. from knowledge about what to do under
certain conditions towards knowledge about intended
outcomes.
The new approach to rule modification can be
seen as change in the dominating type of rationality
and knowledge. As the tradition has been that rules
have been developed in accordance with rule developers current understanding of the actual situation, the
logic of appropriateness has played an important role
(March, 1994). The knowledge base here is familiarity and experience. This knowledge is rather implicit,
i.e. tacit, and the information treatment is intuitively.
Ellstrm (1996) labels the knowledge perspective of
this tradition intuitively-contextual.
The increased emphasis upon goals in outcome oriented rules requires rule makers to identify possible
alternatives and choose between them in accordance
with their contribution to preset goals of the rules. This
approach is linked to another type of rationality that
March (1994) calls rationalistic choice based decision
making. This form of decision making is preference
based and supposed to apply the logic of consequence.
March argues that the logic of consequence makes
great demands on the abilities of individuals and institutions to anticipate the future and to form useful
preferences.
The introduction of risk analyses represents an additional move in this direction. Perrow (1999) argues
that such a development represents a strengthening
of the tradition of absolute rationality. This is a form
of rationality wherein calculations can be made about
risks and benefits, clearly showing which activities we
should prefer. Risk analyses thus serve as support for
choice based decision-making as described by March
(1994).
The scientific nature of the outcome oriented rules
and risk analyses also resemble the highest level
of technical rationality described by Schn (1991).
Here the strategy is First thinking and then acting.
Ellstrm (1996) labels the knowledge perspective of
this tradition rationalistic. The dominating knowledge base of this knowledge tradition is theoretical
and explicit and the treatment of information is analytical. Lindblom (1959) argues that the rationalistic
strategy implies that decisions have to start from

new fundamentals each time. They only build upon

the past when experiences are embodied in a theory.
The decision-makers are always prepared to start from
the ground up, i.e. from scratch.
Seen together, this theory indicates that to change
an experience based, prescriptive safety rule tradition
towards outcome oriented rules based on risk analyses will require a change in attention and knowledge
tradition of the rules.
1.2

Safety rules and knowledge in the Norwegian

railway system

The Norwegian railway system has a tradition of prescriptive rules directed at the operative staff at the
lower levels of organizational hierarchies. The rules
has been developed with growing knowledge of the
systems technology, activities and interactions and
with experiences of unwanted events or accidents
(Gulowsen & Ryggvik, 2004; Ryggvik, 2004). Much
of the knowledge has been derived from practice and
consisted of collective tacit and explicit knowledge.
This knowledge was shared through an internal educational system, practice oriented trainee programs and
socialization. Here the rules served an important role
for the structure of the education and as knowledge
carriers.
In 1996, steps were taken to open the Norwegian
railway system to new traffic operators. The Norwegian
state owned railway company (NSB) was divided
into the National Railway Administration, which was
responsible for infrastructure management, and NSB
BA, a state owned traffic operator. An independent
regulatory body, the Norwegian Railway Inspectorate,
was established.
The railway sector, and in particular the Norwegian
Railway Inspectorate, has been influenced by the
safety management traditions of the Norwegian oil
industry (Ryggvik 2004). This tradition has emphasized internal control principles with extensive use of
risk analyses and outcome oriented rules. The development has resulted in initiatives, especially from the
Norwegian Railway Inspectorate, to change the tradition of experience based, prescriptive rules towards
outcome oriented rules based on results from risk
analyses.
The intentions of a development towards a deductive and risk based top-down approach to safety-rule
modifications was evident in two different projects.
One project was established for modification of the
traffic safety rules. In general, these rules were
detailed prescriptive action rules that coordinated
the activities of the operative staff involved in traffic
operations. The management of this project encouraged the rule developers to think new and develop
outcome-oriented rules formulated as goals and to
base these upon risk analyses. From the beginning, the

840

rule-imposers were the Norwegian Railway Administration. Later this responsibility was transferred to the
Norwegian Railway Inspectorate.
The other project had as purpose to improve
the management of infrastructure maintenance. One
element in this project was to modify the maintenance rules. These rules were organized in different sets for each subsystem of the infrastructure.
They were mainly detailed prescriptive action or state
rules directed at the operative staff that served both
safety and other purposes. The different subsystems
had varying characteristics regarding time sequencing
of activities, communication and coordination. The
project organized subprojects for the modification of
each rule set.
Also, in this project the management encouraged
the rule developers to think new. This meant to
increase the use of triggering requirements and to
base the rules on risk analyses. The triggering
requirements should define conditions in the
infrastructures that should trigger off maintenance
activities, i.e. define outcomes for maintenance activities. The rule-imposers were the Norwegian Railway
Administration.
On this background, the Norwegian railway system
represented an opportunity to study implementation of
the ongoing changes in rule traditions and its impact
upon knowledge about operations of the system and
associated dangers.
In this paper the term railway knowledge refers
to the individual and collective understanding of functions and interactions of the railway system. This
includes knowledge of the system itself, its activities
and their interactions, the inherent risks and preventive
means.
2

RESEARCH QUESTION AND APPROACH

This study looks at how the intended changes of

regulatory mode and knowledge base were handled
in practical life. The research question is: How did
ambitions to change the safety rule tradition of the
Norwegian railway system (from predominantly experience based prescriptive rules towards risk based
outcome oriented rules) influence railway knowledge?
The question is based on the hypothesis that the
described change in the safety rule tradition will
change railway knowledge. The study will also explore
possible explanations for the identified changes in
railway knowledge and discuss practical implications. This calls for an explorative and qualitative
approach and a case study design was chosen (Miles &
Huberman, 1994; Yin, 1994).
Four cases of safety rule modifications in the Norwegian railway system were chosen for the study. The
project for modification of the traffic safety rules,

hereafter called the Traffic-rule case was chosen

to represent one case in the study. This case was followed until the work was transferred to the Norwegian
Railway Inspectorate.
Among the subprojects of the Maintenance project,
three cases were chosen for the study. These were the
projects modifying rules for the signal-, power supplyand superstructure-infrastructure. These cases were
followed until the rules were approved.
The information for the study was collected by
interviews of 41 people that had been involved in
the modification processes, studies of selected project
documents and participation in 4 different meetings.
The analyses were performed as an iterative process inspired by Grounded theory (Strauss & Corbin,
1998). The analytic tools and results influenced further
data collection and further data-collection developed
the analytic tools. For further presentation of the
method, see Blakstad (2006).
3
3.1

FINDINGS: REVIVAL AND CONSERVATION

OF RAILWAY KNOWLEDGE
A strategy of reverse invention

In all cases the participants of the modification processes tried to think new and to use the intended
deductive top-down approach to rule development, i.e.
to start with the development of higher order outcome
oriented rules based on knowledge from risk analyses. The chosen methods for the risk analyses used
experienced top events as outset for the analyses. The
core attention was directed at the operative level of the
railway system.
However, the cases critically judged outcome oriented rule solutions and risk analytic results through
inquiries into experience based railway knowledge.
This reflected a precautionary concern for safety
(Blakstad, 2006). For example, one of the persons
involved in the Traffic-rule case explained how one
of the railway professionals of the work group always
expressed his worries for safety. He did this even when
he was not able to express why he was worried. These
expressed worries led to inquiries and discussions that
revealed the foundations for these worries. In this way
experience based railway knowledge, even when it was
tacit, served as reference for safe solutions.
The cases soon abandoned the deductive and risk
based top-down strategy to the rule development.
Instead, all cases used a bottom-up approach where
existing low level prescriptive rules and associated
knowledge were used as outset for the development
of outcome-oriented rules. This strategy is given the
name reverse invention in this study.
When the cases changed into processes of reverse
invention the work built upon the railway knowledge of
the existing rules, i.e. knowledge directly expressed in

841

the rules themselves and knowledge about their history

and intended function. It was necessary to inquire
into the intentions and rationale behind the existing
prescriptive rules, a knowledge that was sometimes
difficult to retrieve. In this way existing knowledge
associated with the pre-existing prescriptive rules was
brought forth.
Also, none of the cases found that outcome oriented
rules gave sufficient control of known risks. The Traffic rule case stayed with the prescriptive rule solution
and intended to use and develop their outcome oriented
formulations for educational purposes. However, the
plan for such a textbook was abandoned of economical reasons. The Maintenance-rule cases developed
outcome oriented triggering requirements. These were
supplemented with explanatory texts and subordinated
prescriptive rules. With safety as an important argument, the cases included more prescriptive rules than
intended.
The main explanations the cases gave for the development towards reverse invention and the chosen rule
solutions were that existing lower order rules and associated knowledge was highly trusted and necessary for
safe performance. Therefore, it was experienced as a
waste to start from scratch without taking advantage
of the knowledge associated with these rules. Reference was also made to the important function the rules
had in the educational system for railway personnel;
a change would require a new way of teaching. Outcome oriented formulations were also welcomed as a
means to illustrate the purpose of prescriptive rules in
the educational settings.

3.2

All cases compared the risk analyses with railway

knowledge. The expressed purpose of the comparison was to control that known risk was included in the
analyses. Accordingly, the cases revealed a higher trust
in existing experience-based railway knowledge than
in the results of the risk analyses and railway knowledge served as reference for good quality of the risk
analyses.
Usually the risk analyses and the railway knowledge provided the same conclusions. When this happened, the confidence in both railway knowledge
and risk analyses increased among the participants
of the work and they experienced it as a validation
of both knowledge sources. The interviewees also
gave some examples where the conclusions from risk
analyses and railway knowledge came in conflict. In
such instances the reason for the different results was
questioned and inquiries initiated. The interviewees
expressed less trust in risk analytic results than the
experience based railway knowledge. Therefore, the
most common strategy for the inquiries was to review
the risk analyses. The major concern was whether
the analyses had excluded important railway knowledge. When weaknesses in the analyses were revealed,
the analyses were adjusted to conform to the railway
knowledge. Through this strategy the risk analyses
were brought in accordance with the railway knowledge and agreement was established. When consensus
was reached, the participants experienced this as if the
risk analyses and the railway knowledge validated each
other. Accordingly, also conflicting initial conclusions
resulted in increased confidence in both.
3.3

Risk analyses validated railway knowledge

In spite of the strategy of reverse invention with the use

of existing prescriptive rules and railway knowledge
as the core fundament for the rule development, all
cases incorporated risk analyses in their modification
work. However, the cases demonstrated four different
solutions for the application of risk analyses that gave
them different functions in the work (Blakstad, 2006).
In spite of this there were some common features in
the processes.
One common feature was that the risk analyses initiated inquiries. As one of the interviewees of the
Maintenance-rule cases related:
But I do believe that we have had much to learn
from the RCM. [Reliability Centered Maintenance;
autors comment.] There were systematically asked
questions about consequences of failures etc. And it
is clear that we have this in the back of our mind for
anything we do but we have not always followed the
thinking to the end. And sometimes one experiences
some surprises.

Validation of railway knowledge through

feedback

The work of the cases included different feedback

loops (Blakstad, 2006). First, the cases made inquiries
into the efficiency of the existing rules. Second, the
cases took steps to achieve feedback from practitioners
upon the evolving rule solutions, methods to include
risk analyses in the modification work and risk analysis
results.
The feedback was given in different ways. In the
Traffic-rule case the members used each other and
their network of practitioners holding different functions for feedback. This was mostly done through
informal inquiries and discussions when the work
raised questions. The work group also involved their
resource groups that held different positions in the
organizational hierarchies of the system. In addition,
new statistical material and reports of unwanted events
and accidents that occurred during the process were
used.
The Maintenance-rule cases also took steps to get
feedback on their work. Again, the network of practitioners that the core participants posited became

842

involved to achieve feedback. However, differences

in the organization of the Maintenance-rule cases and
their tasks created different conditions for using those
people directly involved in maintenance performance
for feedback. The cases participated in a joint hearing
process of the modified rules and the risk analyses.
Also these cases looked into new statistics and accident reports. At the end, the Maintenance-rule project
organized a joint two-step hearing process.
3.4

Systematizing and storing knowledge

The descriptions above reveal that the existing prescriptive rules served as an important knowledge base
for the rule development.
The descriptions also illustrates that the processes
of the cases can be seen as a revival of railway
knowledge. Railway knowledge was spread around
the organization. It was sometimes difficult to retrieve
because some of it had a more or less individual and
tacit form. Therefore the inquiries and the following
work with the achieved knowledge implied an articulation of this knowledge and that more people had
access to it. It remained uncertain whether the inquiries
contributed with actually new knowledge.
The knowledge retrieved from the inquiries were
combined and sorted out, discussed, systematized and
to some extent documented. This implied a direction
of attention where some knowledge became more in
focus than others and therefore included in the work.
The processes were governed by a combination of the
incitements to the rule solutions, the frameworks that
the risk analytic methods provided and the risk perception of the participants. For instance, the final report
of the Traffic-rule case comments that the risk analyses did not contribute with any particular unknown
conditions. However, it had a systematizing function, contributed with an overview and drew attention
to conditions known from before. An interviewee of
the Maintenance-rule cases made some of the same
reflections:
At least the systematizing of it [the risk analyses, authors comment] forces one to evaluate and
document what one does. And he continues: . . .
before it was very much based on individualsthe
experience one had within the areas.
The interviewees were asked who the core contributors to the work were. Their answers revealed that
those performing the rule development and the risk
analyses were the main contributors. Their networks
contributed as important supplements. However, there
were differences between the cases. The organizing of
the cases and the tasks differed and influenced how
much the knowledge of these actors became articulated, made collective and combined. There were also

differences between the cases regarding available time

for reflection.
The Traffic-rule case had the best conditions of
the cases for interaction and access to knowledge
resources. The work group was located together and
had dedicated time for the work. It worked as an
interactive team where also the different tasks interacted in iterative processes. Furthermore, this case
had a formalized organization that included many
actors, required communication and written reports
and agenda papers. Thus knowledge from different
sources became articulated and combined and to a
great extent transferred into a written form.
Among the Maintenance-rule cases, only the work
group of the Superstructure case was located together
and had continuity in their work. This case also had
a formalized Board of the branch that it included.
The Signal case that had the role as a pilot, had
more available economic resources and dedicated time
to organize meetings about the risk analyses. The
Maintenance-rule cases were also less formalized than
the Traffic-rule case. Therefore they did not communicate with others and did not produce written
documentation to the same extent.
However, when it came to the rule solutions,
the Traffic-rule case only expressed knowledge in
prescriptive rules while the Maintenance-rule cases
expressed it in three ways: In triggering requirements,
their explanatory texts and in prescriptive rules.
4

DISCUSSION

The results above reveal that the cases used railway knowledge as the core knowledge base for the
rule modification process. The rule modification process revived railway knowledge by making formerly
tacit knowledge explicit. The processes increased the
confidence in this knowledge.
4.1

The strong position of railway knowledge

The cases did not adopt the rationalistic, deductive topdown strategy that was intended for the modification
work. The main explanation was that existing experience based railway knowledge, and in particular the
knowledge associated with existing prescriptive rules,
was seen as too valuable for safety to be abandoned.
Hence, they are on line with Lindbloms critic of the
rationalistic strategy (Lindblom, 1959).
Furthermore, Reason (1997) argues that the stage
reached in an organizations life history will influence
the opportunities for feed forward and feedback control of activities. The Norwegian railway system was
old enough to have the necessary experience to develop
prescriptive rules in accordance with feed forward
control principles.

843

An additional argument was that the prescriptive

rules were perceived as important and valuable elements in the Norwegian railway systems organizational memory of safe performance, such as discussed
by Stein (1995).
Instead, the cases decided to apply another type
of rationality in their developmental work than the
rationalistic rationality of the deductive top-down
approach. This was based on an inductive bottom-up
strategy where the existing prescriptive rules were used
as the starting point for rule development. This made
it possible to build upon the accumulated knowledge
associated with the prescriptive rules and in particular knowledge about their former and existing context,
their intended function and the experiences of their
efficiency to fulfil their intention. In Marchs terminology, the rationality that the rule developers applied
in the modification processes became dominated by
a rule- and identity based type of decision-making
(March, 1994). The existing rules and associated
knowledge describing safe actions and states served
as a fundament to judge what was considered to be
appropriate outcomes of the regulated activities. In
this way, knowledge associated with existing rules
was brought forth into the new descriptions of wanted
outcomes and the work did not have to start from
scratch. This can be seen as a conservative and precautionary strategy to fulfil the requirement of outcomeoriented rules and a rule hierarchy in a prescriptive
rule tradition.
Furthermore, the inquiries that the cases made into
railway knowledge made them cautious to replace
existing prescriptive rules with outcome-oriented rules
or change them.
Accordingly, the processes got a conservative touch
and prescriptive rules appeared to be more persistent
than expected. The inductive bottom-up approach of
reverse invention made the cases able to build upon
existing knowledge of the prescriptive rules context and function and rule specific knowledge, i.e.
knowledge that resembles Ellstrms descriptions of
intuitively-contextual knowledge (Ellstrm, 1996).
One can say that the decision strategy of the cases
resembled that of Mixed scanning presented by
Etzioni (1967). Building upon his metaphor, predominantly intuitively-contextual railway knowledge
represented a broad angled camera. This was used to
scan the situation or in other words to get an overview
of railway activities, related risks and experiences with
existing means to prevent accidents, including safety
rules. Then rule- and identity based rationality, applying the same knowledge, was used to zero in on those
areas that required more in-depth examination. The
risk analyses contributed to this work. Like Krte et al.
(2002) discuss, the operational environments provided
updated process knowledge and experience data that
served as input to the analytic process.

Accordingly, the inductive bottom-up strategy

created a transition period when higher order rules
were developed from lower level rules. The process
led to inquiries that activated intuitively-contextual
railway knowledge and made it more explicit. Furthermore, attempts to formulate intended outcomes based
on the experiences behind the existing rules made the
intentions of their prescriptions more explicit. Baumard (1999) argues that to make knowledge more
explicit might be an advantage when a system has to
handle organizational transformations. The ongoing
deregulation process of the Norwegian railway system
can be seen as a transformation period. Accordingly,
processes that stimulate articulation of knowledge
might be an advantage for the safety of the system.
All together, this reveals that the predominantly
intuitively-contextual railway knowledge of different sources became revived and integrated into the
form of the higher order outcome oriented rules and
the forms of the chosen risk analytic method. Furthermore, the revived knowledge became selected in
accordance with the rule developers perception of risk
and transformed into the more abstract and contextfree forms of outcomes and the structure and schema
of the chosen risk analytic methods. Such knowledge
can be labelled rationalistic knowledge (Schn, 1991;
Ellstrm, 1996). In this way the system developed the
ability to take advantage of railway knowledge in both
intuitively-contextual and rationalistic forms in its
safety work. However, the inherent conservatism of the
processes might make it difficult to foresee new and
unexpected dangers, such as Turner & Pidgeon (1997)
discuss. The ongoing transition period of the system
creates changes from which it might be difficult to
foresee all consequences for safety.
4.2 Is the revived railway knowledge endangered?
The fact that some of the revived railway knowledge,
and mainly the rationalistic elements, is transformed
into the written form and stored in rules and documents
does not mean that the knowledge is stored in organizational memory. As knowledge is relational and
context specific, data and information that is transferred into written form is not the same as knowledge
for the reader of the written documentation (Baumard, 1999; Nonaka & Takeuchi, 1995; Stein, 1995).
Written documentation cannot communicate the rich
mental maps that might be necessary to decode written texts and to understand the complex dynamics of
reality. Also, written texts stored in databases require
that its storing can be located and that access is
given to it when it is necessary or useful to retrieve
it (Stein, 1995). In addition, mental maps are also
often important to foresee consequences of different
actions and choices that are necessary for accident
prevention (Perrow, 1984/1999; Rasmussen, 1997;

844

Rasmussen & Svedung, 2000). The results of the study

also revealed that such knowledge was experienced as
being important for the understanding of the rules and
their intended function, for judging their relevance and
for rule-followers motivation for compliance.
Accordingly, to keep the revived railway knowledge alive for the future, the written text has to be
retrievable. Furthermore, it has to be supplemented
with elements of social interaction and context. Therefore, it is essential that intuitively-contextual railway
knowledge is stored in organizational memory.
The differences between the cases in their organizing regarding participation, communication and
interaction imply that they differed regarding degree
of social interaction and relation to the rules context. This created differences regarding articulation of
knowledge and how this knowledge was distributed
among the involved actors. However, the cases did not
provide examples of systematic storing of intuitivelycontextual railway knowledge.
Furthermore, the increased emphasis upon risk
analyses as the fundament for rule development and
intentions of increased use of outcome oriented rules
might strengthen this development. In addition, rationalistic knowledge generally holds a higher status than
intuitively-contextual knowledge (Perby, 1995; Schn,
1991). Therefore, the status of the most intuitively contextual elements of railway knowledge might become
reduced in the future. The status of knowledge might
also influence the motivation to retrieve information
(Stein, 1995).
The ongoing deregulation of the Norwegian railway
system might weaken the conditions for developing
railway knowledge that holds an extensive overview of
the complex interactions of the system. The deregulation process may also weaken the systems traditional
conditions for socialization and existing communities
of practice that were considered particularly important for transfer of tacit knowledge (Baumard, 1999;
Lave & Wenger, 1991; Nonaka & Takeuchi, 1995;
Wenger, 1998). The deregulation processes has also
caused a work force reduction. Baumard (1999) warns
against the danger that the need to renew knowledge
might lead the firm to remove the representatives of
the old knowledge. By doing this, they remove the tacit
knowledge of the firm.
The deregulation process also implies increased
complexity, uncertainty and ambiguity. Etzioni (1967)
argues that under such conditions it might be required
to increase investments in thorough studies of the situation. Accordingly, rich intuitively-contextual railway
knowledge will be necessary to provide a good picture.
However, with reference to the discussions of
Turner and Reason of disaster incubation or latent
conditions, there is a danger that even such railway
knowledge is not sufficient to check for dangers that
can not be easily discovered (Turner & Pidgeon, 1997;

Reason, 1997). Therefore it might be useful to search

for the inclusion of alternative approaches to the traditional railway knowledge. In the framework for safety
rule development that is applied for the SAMRAIL
research (European Commission. 2004a), this might
imply to increase investments in the first step of this
framework. This step requires that the rule developers
define the processes to be regulated, related accident
scenarios and means to control the activities.
Seen together, if the rich intuitively-contextual railway knowledge is not stored in organizational memory
by other means than those revealed in the study, the
benefit of revived knowledge might be lost in the
future. Also, the ongoing changes of the Norwegian
railway system require judgements of relevance of the
experience based railway knowledge for the current
context and organizational learning.
There are already ongoing discussions in European
railways about establishing learning agencies to further the development of organizational knowledge
and its storing in organizational memory (European
Commission, 2004b). With reference to the differences in the organizing of the studied cases the
organization of rule modifications can either be given
the status as a learning agency or be linked to such
agencies. Also, the inquiries into railway knowledge
revealed that there are existing communities of practice within the Norwegian railway system that can be
stimulated such as Wenger (1998) has discussed. Furthermore, there might be a potential for establishing
useful communities of practice within the system, such
as revealed in the Dutch railways (European Commission, 2004b). However, the results and discussions
reveal that it is important to further elaborate solutions
for storing and evaluation of railway knowledge.
5

CONCLUSIONS AND PRACTICAL

IMPLICATIONS

The study reveals that the cases met the challenge

of the deductive and risk based top-down approach
to safety rule development with a strategy given the
name reverse invention. This strategy can be seen as
an inductive bottom-up approach to rule development
where existing prescriptive rules and railway knowledge served as the core fundament for the development
of the outcome oriented rules.
The introduction of the deductive and risk based
top-down approach and the revealed process of reverse
invention raised questions that initiated inquiries
into railway knowledge. These inquiries made tacit
knowledge more explicit and knowledge became
gathered and systematized, i.e. railway knowledge
became revived. However, the revived knowledge
became reduced into lean rationalistic forms. It
remained uncertain whether the potential of inquires

845

for organizational learning resulted in actual new

knowledge.
The results and discussions of the study have
practical implications:
Safety rules can serve an important function as
knowledge carriers about operations of a system and
associated dangers. This function should be taken
into consideration when modifying such rules.
Traditional prescriptive safety rules and associated
knowledge can serve as knowledge base for a transformation of rules into outcome oriented rules,
i.e. from rules expressing knowledge about what
to do under certain conditions towards knowledge
expressing intended outcomes. However, this strategy should take ongoing changes with potential for
new and unexpected dangers into consideration.
Introduction of a deductive and risk based approach
in an experience based, prescriptive rule tradition
can stimulate inquiries into existing knowledge. The
inquiries can contribute to a revival and validation of
this knowledge. However, the approach might also
exclude knowledge that does not fit into the frameworks of chosen rule solutions and risk analytic
methods.
Accordingly, organizations should judge the need
for measures to protect safety relevant, endangered
knowledge.
These practical implications are based on only a
few cases in one particular context. Accordingly, they
should be critically judged before applied to other
contexts. To extend the generalizability, studies of
modification processes in other contexts are required.
The authors want to thank The Research Council of
Norway that financed the work.
REFERENCES
Baumard, P. 1999. Tacit Knowledge in Organizations.
London: Sage Publications Ltd.
Blakstad, H.C. 2006. Revising Rules and Reviving Knowledge. Adapting hierarchical and risk-based approaches
to safety rule modifications in the Norwegian railway
system. Trondheim: Doctoral thesis for the degree of
doctor ingenir. Norwegian University of Science and
Technology (NTNU).
Ellstrm, P.E. 1996. Report: Operatrkompetansvad den
er och hur den kan utvecklas. DUP-resultat. Stockholm:
NUTEK. (In Swedish)
European Commission. 2003a. Safety culture in nuclear and
process control. Fifth Framework Program SAMRAIL.
Appendix 10: WP 2.1.9. August 5, 2003.
European Commission. 2003b. SAMNET Glossary. Fifth
Framework Program SAMNET thematic Network. April
09, 2003.
Etzioni, A. 1967. Mixed-Scanning: A Third Approach
To Decision-Making. Public Administration Review
385392. December 1967.

Gullowsen & Ryggvik, 2004. Jernbanen i Norge 18542004.

Nye tider og gamle spor. Bergen: Vigmostad og Bjrke
AS. (In Norwegian)
Hale, A.R. 1990. Safety rules O.K.? Journal of Occupational
Accidents 12, 320.
Hale, A.R., Heming, B.H.J., Carthey, J. & Kirwan, B. 1997.
Modelling of safety management systems. Safety Science
26(1/2), 121140.
Hale, A.R., Heijer, F. & Koornneef, F. 2003. Management of
safety rules: The case of railways. Safety Science Monitor
7, Article III-2, 111.
Hovden, J. 1998. Models of Organizations versus Safety
Management Approaches: A Discussion Based on Studies of the Internal control of SHE Reform in Norway.
In: Hale, A.R. & Baram, M. (Eds.). Safety management:
the challenge of change. Oxford: Pergamon.
Krte, J., Aven, T. & Rosness, R. 2002. On the use of risk
analyses in different decision settings. Paper presented at
ESREL 2002. Lyon. March 1921, 2002.
Lave, J. & Wenger, E. 1991. Situated Learning. Legitimate peripheral participation. Cambridge: Cambridge
University Press.
Lindblom, C. 1959. The Science of Muddling Through.
Public Administration Review 19, 7988.
March, J.G. 1994. A Primer on Decision Making. New York:
The Free Press.
Miles, M.B. & Huberman, A.M. 1994. Qualitative Data
Analysis. London: Sage Publications Ltd.
Nonaka, I., Takeuchi, H. 1995. The Knowledge-Creating
Company. New York: Oxford University Press.
Perby, M.L. 1995. Konsten att bemstra en process. Om att
frvalta yrkeskunnande. Hedemora: Gidlunds Frlag. (In
Swedish)
Perrow, C. 1999. Normal Accidents. Living with HighRisk Technologies. Princeton: Princeton University Press.
(First issued in 1984)
Rasmussen, J. 1997. Risk management in a dynamic society:
A modelling problem. Safety Science 27(2/3), 183213.
Rasmussen, J. & Svedung, I. 2000. Proactive Risk Management in a Dynamic Society. Karlstad: Rddningsverket.
Reason, J. 1997. Managing the risks of organizational
accidents. Aldershot: Ashgate Publishing Limited.
Reason, J., Parker, D. & Lawton, R. 1998. Organizational controls and safety: The varieties of rule-related
behaviour. Journal of Occupational and Organizational
Psychology 71, 189304.
Ryggvik, H. 2004. Jernbanen, oljen, sikkerheten og historien.
In: Lydersen, S. (Ed.). Fra flis i fingeren til ragnarok.
Trondheim: Tapir Akademisk Forlag. (In Norwegian)
Schn, D. 1991. The Reflective Practitioner. Aldershot:
Arena, Ashgate Publishing Limited. (First issued in 1983)
Stein, E.W. 1995. Organizational Memory: Review of
Concepts and Recommendations for Management. International Journal of Information Management. Vol. 15,
No. 2, 1732.
Strauss, A. & Corbin, J. 1998. Basics of Qualitative Research.
California: Sage Publications, Inc.
Turner, B.A., Pidgeon, N.F. 1997. Man-made disasters.
Oxford: Butterworth-Heinemann.
Wenger, E. 1998. Communities of Practice: Learning as a
Social System. Systems Thinker. Vol. 9, No. 5, 15.
Yin, R.K. 1994. Case Study Research. California: Sage
Publications, Inc.

846

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk Management in systems: Learning to recognize and respond

to weak signals
E. Guillaume
Safety Science Group, Technological University of Delft, The Netherlands

ABSTRACT: The prevention of Major Accidents is the core of high hazard industries activities. Despite
the increasing safety level, industrial sites are asking for innovative tools. The notion of weak signals will
enable the industries to anticipate danger and improve their safety management. Our preliminary results show
the huge interest and relevance of the weak signals but also the difficulty to, concretely, treat them within
the Safety Management. We found out that organizational features are weak signals blockers: Bureaucratic
management of Safety, linear and bottom-up communication, and a reactive Safety Management. In order to
favor weak signals treatment, we should act on these organizational factors. This is the main objective of this
PhD research.

INTRODUCTION

Accident prevention is a central issue in high-risk

industries such as nuclear power plants, aircrafts and
petrochemical plants. These industries have set up
safer equipments, technical safety barriers but also
organizational and human barriers in order to manage risk and improve their capacity to prevent the
accidents. Some researches admit that industries
capacities to learn from their own experiences will
enable them to improve accident prevention. Reporting systems such as Learning from Experience (Retour
dExprience in French) aim at learning from failures and negative experiences. They are implanted
to collect, analyze and share data on the accidents
which occurred on sites. Despite the high relevance of
Rex System, many researchers (Bourrier, 2002, Dien,
2006 and Amalberti and Barriquault, 1999) showed
two main weaknesses: the limitsRex would provide
mainly technical and direct causes of the accidents,
and biasesRex would be more used as an enormous
data base than an opportunity to share the lessons
learnt from the accidents. The goal of this research
is trying to overcome these limits by exploring new
research areas. The issue of weak signals is emerging in industrial companies like EDF (Electricit de
France1 ) and academic researches might provide relevant ideas. Defined as accidents precursors, the weak
signals would enable to identify unsafe situations and
degradation of the system. In that respect, identifying

and taking the weak signals into account would favor

proactive approaches and better accidents prevention.
This research is a partnership agreed with a petrochemical company and a steel company, both situated
in France. To collect data, case studies will be carried out in both sites. First, failure scenarios will be
carefully studied. By looking from the end point of the
accident, we will try to identify what weak signals were
picked up by operational crew/shifts, the actions taken
to take them into account (technical and organizational
responses). Then, normal functioning will provide
data on the human and technical tools, methodology
and procedures used to manage, everyday, risk in both
sites.
This document is composed of two sections. The
first will expose the main definitions of the weak signals. The second one will try to describe the first data
collected in Arcelor site.
2

DEFINITION

2.1 Strategic management

The notion of weak signals has already been studied
in several disciplines: history, geology and medicine
(the later uses more frequently the notion of forerunners). Among these works, strategic management
provides interesting views that are described here.
Several studies have tried to define the nature
and the role of the weak signals. In literature dealing with strategic surveillance, weak signals are
more qualitative than quantitative, uncertain, fragmented and ambiguous (Mevel, 2004, p. 2021).

847

The interest of weak signals lies in the role they play in

strategic management. Derived from Ansoff and Dowell (1975), Lesca and Castagnos (2004), and Lesca
and Blancos (2002), researches argue that, threatened
by an uncertain and changing environment, companies have to remain competitive and to capture early
warning signals like weak signals. The more a company develops surveillance, the more it will be able to
detect these signals and anticipate changes, ruptures
or unexpected events. In that respect, weak signals are
defined as anticipation information (Caron-Fasan,
2001).
We would like to discuss the adjective weak
which is the conventional term always used. Weak
implies that strong signals exist and emerge inside
and outside organizations. Ansoff and Mc Donell2
(1975) propose the following definition: weak signals are imprecise early indications about impending
impactful events ( . . . ). Such signals mature over
time and become strong signals. (p. 2021). The
most appropriate adjective would actually be early
signals. In fact, as we will see later, a weak signal
is a signal which is emerging long before an accident occurs. Despite its uncertainty, its fuzziness, the
challenge is to detect them early in order to implement a strategic response. Over time, weak signals
become strong signals and the strategy will be too
late to response correctly to the threat. Therefore,
we assume that the nature of the weak signals is less
important than time evolution aspect. We will try to
develop this idea later on.
The following section is an uncompleted study of
the main studies dealing with the weak signals in
Safety Management.

2.2

The main definitions in safety management field

This document presents the main contributions and

tries to set atemporary- definition of the weak signals. The following section is composed of three parts:
the nature of weak signals, the way they emerge and
the way they are detected and treated.
First of all, D. Vaughan (2003) defines a weak signal as a subjective, intuitive argument, and mainly
ambiguous information. It is supported by informal
information, which means that the threat to safety is
not really clear for the members of the organization.
According to M. Llory (1996) a weak signal is a forerunner, and a repetitive precursor warning of serious
danger.

2 They devoted an entire chapter to the weak signals enti-

tled Using weak signals in which they describe how to

respond to the issue of weak signals.

Then, in the context of Safety Management, two

theoretical schools disagree. The first oneTurner,
Vaughan- explain the occurrence of accidents with
the existence of precursors. The second one, Perrow,
considers the accidents as a normal consequence of
complex and coupled systems. This PhD research
fits in the first stream, which brings us to underline
the following statement: weak signals emerge before
the accident meaning that they could be captured and
could avoid the accident. Many authors attempted to
describe the appearance or the emergence of the weak
signals. B. Turner and N. Pidgeon (1997) proposed the
notion of incubation period (taken from the medical
field) defined as a chain of discrepant events develop
and accumulate unnoticed (B. Turner, 1997, p. 381).
During this period, many alarming signs are emerging
but not detected by the members of the organization.
Some other authors use notions similar to the idea of
incubation. Roux-Dufort (2000) writes that a crisis
is the product of a long gestation, during which organizational dysfunctions accumulate. Dien and Perlot
(2006) state that an accident is not a fortuitous event,
but would be the last stage of a process of damage to
safety.
Despite the interest of these signalsthe possibility to prevent an accident thanks to those accident
precursors- many studies have pointed out the difficulty to treat them. Turner and Pigdeon, (1997)
Vaughan (2003) and Llory (1996) underline the real
difficulty to treat weak signals in time, and therefore to prevent a disaster. In the chapter entitled the
Bhopal precursors Llory (1996) stresses the role of
whistleblowers, Bhopal industrys workers, who complained about very bad safety condition (e.g. damaged
defenses barriers). These workers tried to warn the
management and Madhya Pradesh, Bhopal state capital. But these whistleblowers were not listened; they
messages were not taken into account by the management. D. Vaughan (2003) underlined the role of the
Thiokol engineer who tried to warn NASA. She quotes
the engineer we shouldnt ship anymore rocket until
we got it fixed (O-ring) (2003, p. 254). Although
engineers sent memos to warn Challenger launch
deciders, these messages were ignored. At that time,
NASA believed more in formal and quantitative procedures than messages based on intuition, qualitative
and informal (which was the nature of memos). She
writes the context of these memos made them weak
signals to insiders at that time (2003, p. 255). These
initiatives are often described as failures. The detectors, or whistleblowers, did not send the message
to the right person or suffered from communication pathology (M. Llory used a concept described
by C. Desjours) which means that the communication between workers and deciders, and in that
respect the amplification of the warning messages, is
blocked.

848

We assume that the weak signals emerge before

the accident, during the incubation period. They are
detected (by people we call whistleblowers) but not
treated and taken into account. The PhD research
assumes that, in the pathway detection-treatment,
some factors do not allow the treatment of these signals. In fact, the organization would not accept or
recognize the relevance of these signals and the legitimacy of the members who decided to support and
amplify them. My PhD research will focus on the identification of these factors to understand the difficulty
in treating them.
The weak signals, in the field of Safety Management, are obviously accident precursors (if we admit
that such signals exist before an accident). Detected
on time, they could prevent and even stop an accident.
The studies quoted previously exposed the interesting role of the whistleblowers. Despite their actions,
their messages were, in the cases quoted, ignored.
Thus, the authors acknowledge how difficult the treatment of such signals within the organizations is. The
next section will focus on the treatment of the weak
signals.

3
3.1

TREATING THE WEAK SIGNALS

Communication channels

The issue of weak signals reveals new issues related

to the interpretation of the information emerging
from outside and inside the organization. Turner and
Pidgeon (1997) propose to go back to the general properties of information by paying attention to the manner
in which the information is dealt with in theory
communication. To the authors, if the information
transmitted falls into the available sets of categories,
then the information will be received. If the message
falls outside these recognized categories, this information will be regarded as error. We assume that
the weak signals, regarding their nature (ambiguous,
uncertain) and their discontinuous way of emerging,
are ignored, partly because they are incompatible with
a closed communication system. No category is available to interpret these signals. In other words, three
options can be taken into account. First, the weak
signals are ignored because people have got no tools
to interpret them. The signals are merely seen as
irrelevant information. Then, the signals can be intentionally ignored because they increase uncertainty and
unsafe work conditions. Turner and Pidgeon (1997)
write we may thus regard this kind of event, which
was not assigned a place in the relevant system of
classes (p. ). This is highly relevant for understanding
the issue of weak signals. Finally, these signals were
ignored because of the difficulty to pick the relevant
ones up.

The following section deals with the factors which

may block or amplify the treatment of weak signals,
the main assumption of this PhD research.

3.2 Cognitive and organizational frames: Weak

signals blockers and amplifiers
Based on several accident analyses, Turner (Weick,
1998) pointed out several features to explain why
the signals emerging during the incubation period
weremostly- ignored. The rigidities in perception
and beliefs in organizational settings are the first features. It means that the possibility to detect disasters
can be inhibited by cultural and organizational factors.
Culture can lead in this case to a collective blindness to important issues. Then, in the same book, B.
Turner mentions organizational exclusivity which
means that organization can disregard non members
who try to amplify warning information. Finally, he
points out the capacity to minimize the emergent dangers. The organizational features seem important to
understand why such events are unnoticed. A number of French studies point out the communication
system underlying its cognitive aspect. Bourrier and
Laroche (2000) describe a similar phenomenon. They
state that organizations have some difficulties to treat
correctly information because of cognitive categories
set a priori. These filters can lead the organization
to blindness. Finally, Vaughan observed that culture
could be a factor of weak signals ignorance. Based on
the analysis of Challenger accident (2003) Vaughan
writes that NASA tended to accept anomalies as the
normal functioning leading to a process of deviance
normalization.
After describing the phenomenon and trying to
identify the main reasons for weak signals blocking
in an organizational context, some authors attempt to
explore solutions in order to amplify them.
The main improvements proposed by the researchers
deal with on organizations capacity to be surprised.
Ansoff and Mc Donnell (1975) emphasize environmental surveillance, awareness, and internal flexibility. Roux-Dufort (2000) asserts that weak signals do
not fit in any preconceived coding process. To capture
such precursor signals, he writes that the organizations should accept information or events which do
not fit in the knowledge and technology which is
already implanted. As Bourrier and Laroche (2000)
agree on the fact that these categories a priori have
to be reviewed to treat correctly the information and
to be able to capture the unexpected events (Turner
and Pidgeon, 1997). Finally, stberg (2006) argues
that human intelligence would be a way to take the
weak signals into account. He writes that intelligence
refers to have a bearing on the dealing with surprise (p. 19) He writes obviously, on the one hand

849

human qualifications for intelligent performance are

decisive for the outcome of efforts to cope with surprises and, on the other hand, the preparedness for
surprising events by intelligence activities enhances
the proper reaction in the prevailing circumstances.
(p. 19).
The previous section is an attempt to define weak
signals in academic literature. Although based on
empirical works, the quoted studies reveal generic
definitions: relevance and use to studying weak signals. But what are, concretely, the weak signals?
What do they refer to? What do they say about the
system studied (technical and social), risks and accidents? These points will be discussed in the following
section.

4.1

PRELIMINARY RESULTS: WEAK

SIGNALS AND TREATMENT, EXAMPLES
IN A STEEL PLANT AND A REFINERY

These case studies provided interesting findings on

these four items. The following section will try to
describe these first results.
4.2 From a generic definition to a more
specific definition
Weak signals could be theoretically defined as accident precursors alerting of an imminent event. The
case studies appeared as an important step in the PhD
research. As a matter of fact, we came up with a more
specific and practical- definition of weak signals. They
are defined in a specific context (an organization, a
site), related to specific activity (steel making) and
specific risks.
The main findings concern three items: the difficulty to pick the weak signals on time, signals
interpretation and the channels of communication used
by people trying to transmit them. Lessons learnt
4.3 The difficulty to pick up the weak signals

Methodology

The research project planned to collect data on weak

signals by carrying out case studies on failure scenarios, success stories and normal functioning.
This research aims at studying weak signals of
major accidents. INES scale defines major accidents
as events with major effects outside the site, implying
consequences on environment and health. Fortunately, these events are very rare. The scenarios
investigated would more concern near-misses it is
to say events that could have been more serious in other
circumstances.
We carried out five case studies on failure scenarios. These accidents occurred in three operational
departments: cooking plant, steel making plant and
energy dispatching department. The objective of these
case studies was to look from the end point of the accident, and trying to identify whether weak signals were
picked up by operational crew/shifts on operational
shop floor? What actions were taken to take them
into account (technical and organizational responses)?
Why they did not take any actions? What lessons were
learnt from these weak signals afterwards?
Then, some observations on normal functioning
have been carried. They provided data on the tools
(learning processes like Learning from Experience,
risk analysis, audits), safety barriers (technical and
organizational), but also management, implanted in
both sites to manage risk. Finally, we came up against
the difficulty to investigate success stories. In fact,
success stories are not, by that very fact, based on
critical events. They are not recorded in reporting systems and, regarding to accidents prevention, they do
not deserve any further analysis. This idea must be
deepen in petrochemical plant.

The scenarios investigated refer to standard operations

but the events described appeared as real surprises.
Consequently, the signal emerging before the accident
(during the incubation period as Turner showed) were
detected but not interpreted as accidents precursors.
The difficulty of weak signals lies in the capacity of
people to combine several pieces of information dispatched in the organization and give sense to them.
As Turner wrote, the information were already in the
organization but not interpreted. Then, the scenarios
investigated showed that the accident was the result
of a long process of safety degradation. During this
period, many signals had been detected, recorded in
reporting systems but rarely connected.
Example 1: Explosion of a calcium carbide bunker
Steelmaking plant uses calcium carbide. It is meant to
take sulphur away from pig iron. Acetylene, produced
by a contact between calcium carbide and water, is a
well known risk and the site is preventing the bunker
against water filtering. However, 16th of December 2005, the bunker exploded. A presence of water
explained clearly the accident. People involved in the
investigation discovered (thanks to an external expert)
that the roof of the bunker was corroded. In fact,
the roof had been painted for maintenance reasons,
except in an unreachable part of it. Birds were living
there, attacking and scaring maintenance operators.
Dirt favored the corrosion and degraded, little by little
the roof and enabled water to soak in the bunker.
Signals appeared obvious and relevant to people
after the event. Indeed, analysis revealed a long process
of safety degradation which led to the bunker explosion: maintenance problems (corrosion), mal-adjusted
roof (liable to rust) and the possibility that water

850

seeps through bunkers roof. Though these pieces

of information were detected, known and recorded,
the combination between them were not taken into
account. The objective is now to understand what
factors blocked the possibility to give sense to these
signals. The PhD research is still in the data collection
stage, and these factors have not been identified yet.
Signals interpretation
The scenarios investigated refer to standard operations
but the events described appeared as real surprises.
Consequently, the signal emerging before the accident
(during the incubation period as Turner showed) were
detected but not interpreted as accidents precursors.
The difficulty of weak signals lies in the capacity of
people to combine several pieces of information dispatched in the organization and give sense to them.
As Turner wrote, the information were already in the
organization but not interpreted. Then, the scenarios
investigated showed that the accident was the result
of a long process of safety degradation. During this
period, many signals had been detected, recorded in
reporting systems but rarely connected.
4.4

Channels of signals transmission

The case studies showed clearly that weak signals

were detected but not taken into account. We believe
that the problem lies in the translation of the signals
detected and transmission to the right person/target.
Some academic studies have tried to identify ways to
bring the relevant information to the organization (and
particularly people in position to make decision on its
base).
We identified three ways to transmit the signals
detected in the organization
Whistleblowers
Safety and health committees
Operational visits
This paper will stress the first whistleblowers
because pieces of data on the others are still missing.
Whistleblowers are considered as a channel to transmit
information related to safety [16]. In failure scenarios,
whistleblowers obviously failed.
Example 2: Explosion of an oxygen pipe
Fluids expert of the Energy dispatching department
explained in the interview that in 2005 he advised
department manager to train operators about the dangers of oxygen. His alert was ignored and, in February
2005, few months later, an oxygen pipe exploded.
The issue of whistleblowers is interesting. On the
basis of the interviews carried out, the whistleblowers
would be experts. They are educated (PhD, trainings,
master graduation) and have knowledge on techniques
and safety. The resources they used to detect dangerous situations are based on their role (they work

are experts and give advises to management), their

knowledge and their experience. But obviously, their
message was ignored. At that stage of the research, we
can only suppose that the signals they tried to transmit were more based on their intuition and qualitative
data than on formal and quantitative data. Although
these signals were already considered as strong for
the whistleblowers, they remain weak for the receivers
because on an organizational culture based tangible
proves. Consequently, their message was not taken into
account.
5

LESSONS LEARNT FROM WEAK SIGNALS

As we mentioned previously, weak signals are obvious after the accident. Once identified, people learnt
lessons from weak signals, particularly in the new
design of damaged installations. For instance, Design
Department, in charge of the bunker building, took
Expert recommendations into account for implanting new safety barriers and better detection systems
(acetylene and temperature). However, we must admit
that weak signals are still ignored before they lead to
an accident. The explanation seems to be found on the
organizational playing a role of blockers. This hypothesis will be tested in the last case studies period we will
carry out (from April 2008).
CONCLUSION
As a conclusion, weak signals seem to have a practical
relevance in such a site and considered as difficult to
pick up. This difficulty does not lie on detection but
on the capacity to give sense to several pieces of information and possibility to transmit them. We found out
that weak signals were indeed detected, and identified
channels which enable to bring the information to relevant people. However, factors would block this process
and impede the opportunities to learn. The next stage
of the research will be precisely to analyze the data and
reveal what factors (organizational, cultural, individual) block the opportunities to prevent the accidents
investigated.
ACKNOWLEDGEMENTS
I grandly thank the French Foundation for a Safety
Culture (FonCSI) which provides financial sponsors
and fieldwork.
REFERENCES
Amalberti R. et Barriquault C., 1999, Fondements et limites du Retour dExprience, in Annales des Ponts et
Chausses: Retours dExprience, n. 91, pp. 6775.

851

Ansoff I. and Mc Donnell E., 1990, Implanting strategic

management, Second edition, Prentice Hall International,
United Kingdom.
Bourrier M. et Laroche H., 2000, Risques de dfaillance:
les approches organisationnelles, in Risques, erreurs et
dfaillances, Actes de la premire sance du Sminaire
Le risque de dfaillance et son contrle par les individus et les organisations dans les activits hauts risques,
publications MSH Alpes, pp. 1551.
Bourrier M., 2002, Bridging research and practice: the challenge of normal operations, in Journal of Contingencies
and Crisis Management, vol.10, n.4, pp 173180.
Caron-Fasan ML., 2001, Une mthode de gestion
de lattention des signaux faibles, in Systmes
dInformation et Management, vol. 6, n. 4.
Chateauraynaud F., 1999, Les sombres prcurseurs. Une
sociologie pragmatique de lalerte et du risque, Editions
EHESS, Paris.
Dien Y. et Perlot S., 2006, Cassandre au pays des risques
modernes, 29ime Congrs National de Mdecine et
Sant au Travail, Lyon.
Lesca H. and Blanco S., 2002, Contribution la capacit
des entreprises par la sensibilisation aux signaux faibles,
6eme Congrs International Francophone sur la PME,
HEC Montral.
Lesca H. and Castagnos J-C., 2004, Capter les signaux
faibles: comment amorcer le processus?, Economica e
Gestao, Brsil, vol. 4, n. 7, pp. 1534.

Llory M., 1996, Accidents industriels, le cot du silence,

LHarmattan, Paris.
Mevel O., 2004, Du rle des signaux faibles sur la
reconfiguration des processus de la chane de valeur
de lorganisation : lexemple dune centrale dachats de
la grande distribution franaise, Thse de doctorat en
sciences de gestion, Ecole doctorale Lettres, Langues,
Socit et Gestion et Ecole Nationale Suprieure des
Tlcommunications de Bretagne.
Ostberg G., 2006, An unassorted collection of remarks on
aspects, perspectives and dimensions of weak signals,
University of Lund, Sweden.
Roux Dufort C., 2000, Aspects socio et culturels des signaux
faibles dans les organisations, Association ECRIN, 18
mai, Paris.
Turner A.B. and Pidgeon N.F., 1997, Man-Made disasters,
Second edition, Wycheham Publications, London.
Vaughan D., 1996, The Challenger launch decision risky
technology, culture and deviance at NASA, University of
Chicago Press, United States.
Weick K.E, 1998, Foresights and failure: an appreciation
of Barry Turner, in Journal of Contingencies and Crisis
Management, vol. 6, n. 2, pp. 7275.

852

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Author index

Aamo, O.M. 3311

Aase, K. 1385
Abramowicz-Gerigk, T.
3343
Achermann, D. 2061
Adel-Aissanou, K. 611
Adjabi, S. 611
Affeltranger, B. 3093
Agnello, P. 137
Ait-Kadi, D. 641
At-Kadi, D. 1001
Akiba, T. 1839
Albert, I. 2609
Albrechtsen, E. 407, 2649
Alcock, R.E. 415, 993
Ale, B.J.M. 2223, 2715
Algora, C. 1949
Alkali, B.M. 515
All, R. 391
Allaix, D.L. 1621
Almeida Jr., J.R. 3177
Althaus, D. 2239
Altiok, T. 3257
Alzbutas, R. 1819
Amaral Netto, J.D. 2587
Amari, S.V. 1763
Amendola, L. 2409, 2415
Amyotte, P.R. 1147
Ancione, G. 3143
Andersen, T.K. 259
Andrade Herrera, I. 733
Andrews, J.D. 1739, 1873
Aneziris, O.N. 767, 777,
787
Angelo Bragatto, P. 2701
Ansaldi, S. 137
Anto, P. 3265
Antoni, M. 3231
Antonioni, G. 2397, 2749
Antonsen, S. 1377
Arbaretier, E. 1937
Arezes, P.M. 761
Arizmendi, M. 205
Arnaiz, A. 2175, 3223
Arnaldos, J. 1073, 2421
Artacho, M.A. 2409, 2415
Artiba, A. 641

Asada, Y. 33
Astapenko, D. 2021
Astarita, G. 27
Aubry, J.F. 2549
Auder, B. 2107
Augutis, J. 1867, 2569,
2575, 3101
Ault, G.W. 2601
Aven, T. 323, 365, 1207,
1335, 2081
Avram, D. 477
Azi, M. 611
Bada, F.G. 575
Baggen, J.H. 1519
Baicu, F. 1027
Baksteen, H. 767, 777
Balakrishnan, N. 1915
Bank, M. 351, 2675
Baraldi, P. 2101
Barbarini, P. 1049
Barker, C. 619
Barnert, T. 1463
Barone, S. 2251
Barontini, F. 2345
Barros, A. 2003, 3125
Bartlett, L.M. 2021
Basco, A. 3085
Basnyat, S. 45
Bayart, M. 3245
Beard, A.N. 2765
Becerril, M. 1415
Bedford, T. 515, 987
Belhadaoui, H. 1829,
2549
Bellamy, L.J. 767, 777,
2223
Benard, V. 3245
Benjelloun, F. 2369,
3067
Berenguer, C. 3125
Brenguer, C. 469, 531,
593, 2003
Berg, H.P. 1439
Bernhaupt, R. 45
Berntsen, P.I.B. 3311
Berrade, M.D. 575

853

Bertsche, B. 875,
2233
Beyer, S. 1539
Bhattacharya, D. 1915
Bianco, C. 2727
Bigham, J. 3109
Birkeland, G. 365
Bischof, N. 2789
Bladh, K. 227
Blakstad, H.C. 839
Bloemhoff, A. 777
Blokus-Roszkowska, A.
2269
Bockarjova, M. 1585,
2781, 2817
Boek, F. 2613
Bolado Lavn, R. 2899
Bonanni, G. 2501
Bonvicini, S. 1199
Bordes, L. 593
Borell, J. 83, 3061
Borgia, O. 211
Borsky, S. 973
Bosworth, D.A. 2353
Bouissou, C. 3191
Bouissou, M. 1779
Boussouf, L. 2135
Bveda, D. 1533
Brndl, M. 2773, 2789
Braarud, P.. 267
Braasch, A. 2239, 2245
Bragatto, P.A. 137
Brandowski, A. 3331
Brandt, U.S. 3031
Bri, R. 489
Brik, Z. 1937
Brissaud, F. 2003
Brunet, S. 3007
Buchheit, G. 2549
Buderath, M. 2175
Bnzli, E. 2641
Burciu, Z. 3337
Burgazzi, L. 1787, 2899
Burgherr, P. 129
Busby, J.S. 415, 993, 1251,
1325
Bye, R.J. 1377

Cabarbaye, A. 2185, 2217

Cabrera, E. 2447
Cadini, F. 477
Calixto, E. 957, 1273
Calle, E.O.F. 2807
Camargo Jr, J.B. 2207
Campedel, M. 2749
Campos, J. 1217
Caamn, I. 163
Carbone, V.I. 1621
Carfagna, E. 3217
Carl, B. 89
Carlos, S. 2827, 2837
Carr, M.J. 523
Carrin, A. 2447
Carvalho, M. 587
Casal, J. 1073, 1119
Castanier, B. 469, 3171
Castillo, C. 2473
Castillo, E. 2473, 2689
Castro, I.T. 463
Cauffriez, L. 3245
Cavalcante, C.A.V. 423,
627, 1165
Chang, D. 703
Chang, K.P. 703
Charpentier, D. 2003
Chatelet, E. 1731, 3093
Chen, J.R. 2757
Chen, K.Y. 39
Chen, X. 863, 1663
Chiu, C.-H. 1651
Cho, S. 2851
Choi, Y. 2913
Chojnacki, E. 697, 905
Chou, Y.-P. 2405
Christley, R.M. 2317
Christou, M.D. 2389
Chung, P.W.H. 1739
Ciancamerla, E. 2501
Clarhaut, J. 3199
Clavareau, J. 455
Clemente, G. 505
Clemente, R. 2501
Clmaco, J. 2627
Clough, H.E. 2317
Cocquempot, V. 3199
Cojazzi, G.G.M. 3135
Colli, A. 341, 2715
Collins, A. 1251
Conejo, A.J. 2689
Conrard, B. 3199
Contini, S. 1009, 3135
Cooke, R.M. 2223
Cooper, J. 2223
Cordella, M. 2345
Cornil, N. 2369, 3067

Costescu, M. 99
Coulibaly, A. 1001
Courage, W.M.G. 2807
Cozzani, V. 1147, 1199,
2345, 2397, 2749, 3153
Craveirinha, J. 2627
Crespo Mrquez, A. 669,
929
Crespo, A. 687, 829
Cugnasca, P.S. 1503
DAuria, F. 2899
Damaso, V.C. 497
Damen, M. 767, 777
Dandache, A. 2549
da Silva, S.A. 243
David, J.-F. 981
David, P. 2259
de Almeida, A.T. 627, 1165
De Ambroggi, M. 1431
De Carlo, F. 211
de M. Brito, A.J. 1165
De Minicis, M. 1495
De Souza, D.I. 919
De Valk, H. 2609
de Wit, M.S. 1585
Debn, A. 2447
Debray, B. 3191
Dehghanbaghi, M. 2379
Dehombreux, P. 2117
Deleuze, G. 1309, 3093
Deloux, E. 469
Delvenne, P. 3007
Delvosalle, C. 2369, 3067
Denis, J.-B. 2609
Depool, T. 2409, 2415
Dersin, P. 2117, 3163
Despujols, A. 531
Destercke, S. 697, 905
Deust, C. 3191
Di Baldassarre, G. 2749
Di Gravio, G. 1495
Di Maio, F. 2873
Dien, Y. 63
Dijoux, Y. 1901
Diou, C. 2549
Dohnal, G. 1847
Domnech, E. 2275,
2289
Dondi, C. 2397
Dong, X.L. 2845
Doudakmani, O. 787
Downes, C.G. 1739, 1873
Driessen, P.P.J. 369
Duckett, D.G. 1325
Duffey, R.B. 941, 1351
Dunj, J. 2421

854

Dutfoy, A. 2093
Dutta, B.B. 3323
Dutuit, Y. 1173
Dvork, J. 2613
Dwight, R.W. 423
Ebrahimipour, V. 1125,
2379
Egidi, D. 2397
Eide, K.A. 1747, 2029
Eisinger, S. 365, 2937
El-Koujok, M. 191
Engen, O.A. 1423
Erdos, G. 291
Eriksson, K. 83, 3061
Esch, S. 1705
Escriche, I. 2275, 2289
Escrig, A. 2743
Espern, J. 3, 121
Espi, E. 2609
Espluga, J. 1301, 1371,
2867
Eusgeld, I. 2541
Eustquio Beraldo, J. 1273
Expsito, A. 3, 121
Eymard, R. 155
Faber, M.H. 1567
Faertes, D. 2587
Fallon, C. 1609, 3007
Fan, K.S. 2757
Faragona, B. 3217
Farr, J. 1301
Fako, P. 1671
Faure, J. 2185, 2217
Fechner, B. 147
Fernndez, A. 1533
Fernndez, I. 3, 121
Fernndez, J. 205, 1395
Fernndez-Villodre, G.
1755
Fernandez Bacarizo, H. 559
Ferreira, R.J.P. 1165
Ferreiro, S. 2175
Feuillard, V. 2135
Fivez, C. 2369, 3067
Figueiredo, F.A. 627
Finkelstein, M.S. 1909
Flage, R. 1335, 2081
Flammini, F. 105
Fleurquin, G. 2117
Fodor, F. 1309
Forseth, U. 3039, 3047
Fouladirad, M. 567, 593,
2003
Frackowiak, W. 3331
Franzoni, G. 1049

Fraser, S.J. 2353

Frenkel, I. 483, 551
Frimannslund, L. 2963
Froihofer, L. 1539
Frutuoso e Melo, P.F. 497
Fuchs, P. 635
Fugas, C. 243
Furuta, K. 33
Fuster, V. 1395, 1401
Gagliardi, R.V. 27
Gaglione, A. 105
Galarza, N. 727
Galassi, G. 2899
Galdmez, P. 1539
Gallay, A. 2609
Gmiz, M.L. 2013
Gmez Fernndez, J. 929
Gmez Fernndez, J.F. 669
Gmez, J.F. 687, 829
Gmez-Mares, M. 1119
Gng, J. 2233
Gsemyr, J. 1747, 2029
Gamiz, M.L. 2447
Gamo, L. 3, 121
Ganapini, S. 1199
Garca Ortiz, J.C. 1539
Garca, B. 757
Garca-Daz, J.C. 201
Garcia, P.A.A. 497
Garca-Bertrand, R. 2689
Gayen, J.-T. 1283
Gerbec, M. 1473, 2157
Gerigk, M. 3303
Geurts, P.A.T.M. 2781
Gil, A. 205
Gil, J. 3, 121
Gillon, P. 3007
Gimnez, M. 2899
Giner-Bosch, V. 2735
Ginestar, D. 175
Giorgio, M. 2251
Girard, Ph. 331
Giraud, J.-B. 2987
Glor, M. 1217
Goeschka, K.M. 1539
Gomes, T. 2627
Gonzlez Daz, V. 929
Gonzlez, J.R. 1949
Gonzlez, P. 3, 121
Gonzlez, V. 669, 687, 829
Gonzalo, J. 1301
Gordon, P. 423
Goti, A. 2707
Gouriveau, R. 191
Goyal, S. 949
Grachorloo, N. 2151

Grall, A. 531, 567, 3125

Grande, . 2937
Grande, O. 1431, 3265
Grazia Gnoni, M. 2701
Grenier, E. 2609
Groth, K.M. 113
Gruber, M. 2675
Guaglio, G. 2541
Gucma, L. 3285
Guedes Soares, C. 881,
3265
Gugliermetti, F. 1341
Guida, M. 2251
Guidi, G. 1341
Guillaume, E. 847
Guo, B. 649
Gurley, K. 2453
Gutteling, J.M. 1317, 1585
Guttormsen, G. 813
Hbrekke, S. 805
Hagen, J.M. 407, 2649
Hamid, S. 2453
Hnle, A. 1547, 1555
Hansson, L. 733
Hardeman, F. 89
Hardman, G. 987
Hring, I. 1547, 1555
Harrami, O. 391, 399
Harvey, J. 291, 299, 1447
Hauge, S. 2921
Haugen, K.-E. 1489
Haugland, D. 2963
Hauschild, J. 2245
Hausken, K. 1157
Haver, K. 2929
Hayat, S. 3199
Helland, A. 361
Heps, V. 813, 1407
Hernndez-Simn, L.M. 11
Herrera, I.A. 19
Herrero, R. 121
Heslop, S. 299
Hildebrandt, M. 267
Holicky, M. 1629
Holmberg, J.-E. 227
Hong, Y. 1943
Hoon Han, S. 2619
Hoppe, G. 2037
Horlick-Jones, T. 1301,
1371, 1601, 2867
Hortal, J. 3, 121, 379
Hossein Mohammadian M.,
S. 1001
Hou, H.-Y. 2405
Hovden, J. 839
Hyland, S. 1385

855

Hryniewicz, O. 581
Hsieh, C.C. 1267
Huang, W.-T. 1651
Hurrell, A.C. 749
Huseby, A.B. 1747, 2029,
2199
Hwang, M. 2861
Iacomini, A. 2501
Ibez, M.J. 2743
Ibez-Llano, C. 2051
Idasiak, V. 2259
Ide, E. 1901
Innal, F. 1173
Iooss, B. 2107, 2135,
2899
Ipia, J.L. 727
Isaksen, S.L. 1747, 1891,
2029, 2937
Izquierdo, J.M. 121, 163
Jallouli, M. 2549
Jamieson, R. 1447
Jammes, L. 305
Janilionis, V. 1819
Jarl Ringstad, A. 813
Jeong, J. 2619
Jiang, Y. 863, 1663
Jo, K.T. 913
Jzwiak, I.J. 1929
Jodejko, A. 1065
Joffe, H. 1293
Johansson, J. 2491
Johnsen, S.O. 805
Jongejan, R.B. 1259
Jnsson, H. 2491
Jord, L. 727
Jore, S.H. 3077
Joris, G. 1609
Jzwiak, I.J. 1455
Jzwiak, K. 1455
Jun, L. 1943
Jung, K. 1629, 1635
Jung, W. 221
Jung, W.S. 2913
Juocevicius, Virg. 1641
Juocevicius, Virm. 1641
Juocevicius, V. 1677
Kalusche, W. 2431
Kamenick, J. 891
Kangur, K. 797
Kanno, T. 33
Kar, A.R. 3323
Karlsen, J.E. 1595
Kastenholz, H. 361
Kayrbekova, D. 2955

Kazeminia, A. 2245
Kellner, J. 2613
Kermisch, C. 1357
Khan, F.I. 1147
Khatab, A. 641
Khvatskin, L. 483
Kim, K.Y. 2913
Kim, M.C. 2909
Kim, S. 2851
Kiranoudis, C. 281
Kleyner, A.V. 1961
Kloos, M. 2125
Kobelsky, S. 1141
Kohda, T. 1035
Koivisto, R. 2511
Kollmann, E. 2641
Kolowrocki, K. 1969, 1985
Konak, A. 2657
Kongsvik, T. 733
Konstandinidou, M. 281,
767, 777
Kontic, B. 2157
Korczak, E. 1795
Kortner, H. 1489
Kosmowski, K.T. 249, 1463
Kosugi, M. 2305, 2311
Koucky, M. 1807, 1813
Koutras, V.P. 1525
Kovacs, S.G. 99
Kowalczyk, G. 449
Kratz, F. 2259
Kriktolaitis, R. 2575, 3101
Krger, W. 2541
Krummenacher, B. 2773
Kubota, H. 2305, 2311
Kudzys, A. 1677, 1685
Kuiper, J. 767, 777
Kujawski, K. 1929
Kulot, E. 1049
Kulturel-Konak, S. 2657
Kurowicka, D. 2223
Kuttschreuter, M. 1317
Kvernberg Andersen, T.
3039, 3047
Labeau, P.E. 455
Labeau, P.-E. 559, 1357
Laclemence, P. 3093
Laheij, G.M.H. 1191
Lamvik, G.M. 2981
Landucci, G. 3153
Langbecker, U. 3275
Langeron, Y. 3125
Larisch, M. 1547
Laulheret, R. 2185, 2217
Le Bot, P. 275
Le Guen, Y. 2987

Lbre La Rovere, E. 957,

1273
Lebrun, R. 2093
Lecoze, J.C. 3191
Lei, H.T. 649
Leira, B.J. 3311
Leito, A.F. 675
Lejette, F. 3231
Lemes, M.J.R. 2207
Leopold, T. 875
Lerena, P. 1217
Lettera, G. 2701
Levitin, G. 1157, 1723
Li, Pan 79
Li, Z.Z. 2845
Limbourg, P. 1705
Limnios, N. 2167
Lin, P.H. 2223
Linde, P.H. 1595
Lindhe, A. 1041
Lins, I.D. 541
Lirussi, M. 2727
Lisi, R. 1019, 3143
Lisnianski, A. 483, 551
Lizakowski, P. 3319
LLovera, P. 1401
Loizzo, M. 2987
Lonchampt, J. 531
Lopes, I.S. 675
Lpez Droguett, E. 541
Lorenzo, G. 2899
Lukoeviciene, O. 1685
Lundteigen, M.A. 2921
MacGillivray, B.H. 415
Maftei, E. 2333
Magott, J. 1055
Mai Van, C. 2797
Makin, A.-M. 739
Malass, O. 1829, 2549
Malich, G. 1081
Mancini, G. 1621
Manuel, H.J. 1113
Marais, K.B. 659
Marcos, J. 1533
Maris, U. 2325
Markatos, N. 281
Markeset, T. 2945, 2955
Markov, J. 1635
Marqus, M. 2899
Marrel, A. 2135
Martn, J. 869
Martinez-Alzamora, N. 441
Martorell, S. 175, 441, 505,
1881, 2275, 2289, 2827,
2837, 2873, 2971
Maschio, G. 1019, 3143

856

Massaiu, S. 267
Mateusz, Z. 3237
Matuzas, V. 2569
Matuziene, V. 2575
Matuziene, V. 3101
Mavko, B. 1771
Mazri, C. 3191
Mazzocca, N. 105
McClure, P. 2295
McGillivray, B.H. 993
McMillan, D. 2601
Mearns, K. 1415
Medina, H. 1073
Medonos, S. 1239
Medromi, H. 2549
Mehers, J.P. 2317
Mehicic Eberhardt, S. 2431
Meier-Hirmer, C. 3183,
3231
Melndez, E. 121, 2051
Meli, J.L. 243, 1415
Membr, J.-M. 2295
Mendes, J.M. 1577
Mendizbal, R. 379, 2827,
2837, 2891
Meneghetti, A. 2727
Menoni, S. 3023
Mercier, S. 155, 603
Merz, H.M. 2773
Meyer, P. 275
Meyna, A. 2239
Mikulov, K. 1671
Milazzo, M.F. 1019, 3143
Miles, R. 1251
Mnguez, R. 2473, 2689
Minichino, M. 2501
Missler-Behr, M. 2431
Mlynczak, M. 57
Mock, R. 2641
Moeller, S. 2431
Molag, M. 3153
Moltu, B. 813
Monfort, E. 2743
Monteiro, F. 2549
Montoro-Cazorla, D. 1955
Montoya, M.I. 1089
Moonis, M. 2353
Morales, O. 2223
Moreno, J. 3
Moreu de Len, P. 669,
687, 829, 929
Morra, P. 2345
Mosleh, A. 113
Motoyoshi, T. 2685
Muoz, M. 1119
Muoz-Esco, F.D. 1539
Mud, M. 767, 777

Mulley, C. 299
Mullor, R. 441
Muslewski, L. 2037
Mutel, B. 1001
Nsje, P. 259, 821, 1407
Nsje, P.C. 2981
Nkland, T.E. 1207, 2929
Naked Haddad, A. 919
Napolitano, N. 1495
Natvig, B. 1747, 2029
Navajas, J. 1301, 1371,
2867
Navarro, E. 757
Navarro, J. 1915
Navarro-Esbr, J. 175
Navrtil, J. 2613
Nebot, Y. 2827, 2873
Nedelec, B. 3191
Neto, H.V. 761
Newby, M. 619
Nguyen, H. 3331
Nicholls, J. 291
Nicol, A.-M. 749
Nieto, F. 2051
Niezgoda, T. 449
Nivolianitou, Z. 281
Nj, O. 3077
Nogueira Daz, E. 899
Norberg, T. 1041
Nordgrd, D.E. 2561
Nowakowski, T. 1055,
1065, 1455, 1929
Nunes, E. 587
Nez Mc Leod, J.E. 1129,
1135
Nez, N. 1949
Nuti, C. 2519
Oh, J. 767, 777
Oliveira, A. 3177
Oliveira, L.F.S. 1919, 2587
Olmos-Pea, S. 11
Oltedal, H.A 1423
Oltra, C. 1301, 1371, 2867
Or, I. 3257
Osrael, J. 1539
zbas, B. 3257
Palacios, A. 1119
Palanque, P. 45
Pandey, M.D. 431
Pantanali, C. 2727
Papazoglou, I.A. 767, 777,
787
Paridaens, J. 89
Park, J. 221, 2909

Park, J.J. 1481

Park, S.D. 913
Park, S.J. 913
Parra, C. 687, 829
Parra Mrquez, C. 669, 929
Pashazadeh, S. 2151
Peciulyte, S. 2575, 3101
Pearce, K. 1447
Pecho, J. 1671
Pedersen, L.M. 2581
Pedroni, N. 709
Peiretti, A. 1119
Pelayo, F. 379, 2827, 2837
Pelaz, A. 727
Penalva, M.L. 205
Pereira, G.A.B. 675
Prez, C.J. 869
Prez-Ocn, R. 1755,
1955
Peschke, J. 2125
Pesme, H. 275
Pey, A. 1217
Pierlot, S. 63
Pierro, F. 2899
Pinelli, J.-P. 2453
Pita, G.L. 2453
Pittiglio, P. 137
Piwowar, J. 3093
Planas, E. 1089
Platis, A.N. 1525
Pock, M. 1829
Podofillini, L. 233
Podsiadlo, A. 3289, 3331
Polic, M. 3015
Ponchet, A. 567
Pop, P. 2333
Popentiu Vladicescu, F.
2333
Popoviciu, N. 1027
Post, J.G. 767, 777
Postgrd, U. 391, 399
Pouligny, Ph. 3183
Poupard, O. 2987
Poupart, E. 45
Prades, A. 1301, 1371,
2867
Pragliola, C. 105
Praks, P. 559
Prescott, D.R. 1873
Prosen, R. 2883
Proske, D. 2441
Pulcini, G. 2251
Puuronen, S. 1995
Pyy, P. 227
Quayzin, X. 1937
Queral, C. 3, 121

857

Quigley, J. 987
Quijano, A. 1395, 1401
Rabbe, M. 2199
Rachel, F.M. 1503
Raffetti, A. 3217
Rajabalinejad, M. 717
Rakowsky, U.K. 2045,
3055
Raman, R. 1239
Raschky, P.A. 965, 973
Rasulo, A. 2519
Rauzy, A. 1173, 1937, 2051
Real, A. 175
Reer, B. 233
Reinders, J. 3153
Remenyte-Prescott, R. 1739
Renaux, D. 3245
Renda, G. 3135
Revilla, O. 3223
Rey-Stolle, I. 1949
Rezaie, K. 1125, 2379
Rhee, T.J. 703
Rheinberger, C.M. 1365
Riedstra, D. 1191
Rietveld, P. 2817
Rimas, J. 1819
Rivera, S.S. 1129, 1135
Robin, V. 331
Rocco S., C.M. 1803
Rocha Fonseca, D. 919
Rodrguez, G. 3, 121
Rodrguez, V. 2707
Rodrguez Cano, D. 899
Red, W. 2929
Roelen, A.L.C. 2223
Rohrmann, R. 1567
Romn, Y. 2013
Romang, H. 2789
Rosn, L. 1041
Rosness, R. 839
Roussignol, M. 155
Rowbotham, A.L. 2353
Rubio, B. 727
Rubio, G. 757
Rubio, J. 727
Rcker, W. 1567
Rudolf Mller, J. 2665
Ruiz-Castro, J.E. 1755
Runhaar, H.A.C. 369
Stre, F. 2635
Sabatini, M. 1199
Sadovsk, Z. 1671
Sagasti, D. 727
Saleh, J.H. 659
Salzano, E. 3085

Samaniego, F.J. 1915

Samrout, M. 1731
San Matas, S. 2735
Snchez, M. 121
Snchez, A. 441, 505, 2707
Sand, K. 2561
Sansavini, G. 1861
Santamara, C. 757
SantAna, M.C. 497
Santos-Reyes, J.R. 11, 2765
Sarshar, S. 183
Saull, J.W. 1351
Savic, R. 1513
Saw, J.L. 2353
Scarf, P.A. 423
Scarlatti, A. 2501
Schbe, H. 1283
Schiefloe, P.M. 2997
Schmitz, W. 2511
Schnieder, E. 2665
Schrder, R.W. 1097
Schweckendiek, T. 2807
Schwindt, M. 965
Segovia, M.C. 881, 1955
Serbanescu, D. 341, 2593,
2715
Serradell, V. 2827, 2837
Servranckx, L. 2369, 3067
Shehata, S. 2285
Shingyochi, K. 1715, 1839
Shokravi, S. 1125
Shu, C.M. 71, 1267
Shu, C.-M. 39, 2405
Siebold, U. 1547
Siegrist, M. 361
Signoret, J.-P. 1173
Silva, S.A. 1415
Simes, C. 2627
Simos, G. 281
Singh, M. 2945
Sipa, J. 57
Skarholt, K. 259, 821,
1407, 2981
Skjerve, A.B. 941
Skjong, R. 3275
Skorupski, J. 2191
Skrobanek, P. 1055
Sliwinski, M. 1463
Smalko, Z. 1231, 3337
Smith, N.W. 1293
Smolarek, L. 3295
Sniedovich, M. 2071
Solano, H. 2447
Solberg, G. 733
Soliwoda, J. 3295
Son, K.S. 1481
Soria, A. 1401

Soriano, M.L. 1395

Soszynska, J. 1985
Soto, E. 1533
Sousa, S.D. 761
Spadoni, G. 2397
Sperandio, S. 331
Spitsa, R. 1141
Spouge, J. 2223
Stamenkovic, B.B. 3163,
3209
Steen, R. 323
Steinka, I. 2269
Stelmach, A.W. 2191
Sterkenburg, R.P. 2363
Stevens, I. 3117
Stian strem, J. 1335
Stoop, J.A. 1519
Strmgren, M. 391, 399
Su, J.L. 2757
Subramanian, C.S. 2453
Sunde, L. 1489
Susperregui, L. 3223
Suter, G. 1217
Sykora, M. 1629
Szpytko, J. 1231
Takai, J. 2685
Tambour, F. 2369, 3067
Tao, J. 1663
Tarelko, W. 3289
Tavares, A.T. 1577
Tchrzewska-Cieslak, B.
2463
Telhada, J. 587
Terruggia, R. 2501
Thns, S. 1567
Thevik, H. 1335
Thompson, H.A. 1881,
2971
Thorpe, N. 299
Thorstad, H.H. 2581
Tian, Z. 1723
Todinov, M.T. 1655, 2143
Torres-Echeverria, A.C.
1881, 2971
Torvatn, H. 259, 2981,
3039, 3047
Trainor, M.T. 2353
Trijssenaar-Buhre, I.J.M.
2363
Tronci, M. 1495
Trucco, P. 1431, 3265
Tseng, J.M. 71
Tsujimura, Y. 1839
Tucci, M. 211
Tugnoli, A. 1147, 2345
Turcanu, C. 89

858

Tveiten, C.K. 2997

Tymoteusz, B. 3237
Upuras, E. 1867
Ulmeanu, A.P. 2167
Ulusu, O.S. 3257
Unagami, T. 2685
Uusitalo, T. 2511
Vzquez Lpez, M. 899
Vzquez, M. 1949
Vaccaro, F. 3217
Vaidogas, E.R. 1641
Valis, D. 1807, 1813
van den Berg, A. 1113
van der Boom, R.P. 2223
van der Most, H. 1585
van der Sluijs, J.P. 369
van der Veen, A. 2781
van der Weide, J.A.M. 431
van Erp, N. 717
van Gelder, P.H.A.J.M. 717,
2797
van Mierlo, M.C.L.M.
2807
van Noortwijk, J.M. 431
van Vliet, A.A.C. 1191
Vanem, E. 3275
vant Sant, J.P. 1113
Vanzi, I. 2519
Vaquero, C. 727
Vaurio, J.K. 1103

Cepin,
M. 1771, 2883
Veiga, F. 205
Verga, S. 315
Verhoef, E.T. 2817
Verleye, G. 3117
Vetere Arellano, A.L. 341,
2593
Vikland, K.M. 1377
Vlchez, J.A. 2421
Viles, E. 205
Villamizar, M. 505
Villanueva, J.F. 2827,
2837
Vinnem, J.E. 1181
Vintr, Z. 1813
Vivalda, C. 305
Vleugel, J.M. 1519
Voirin, M. 63
Vojtek, M. 1671
Volkanovski, A. 1771
Volovoi, V. 1961
Vrancken, J.L.M. 1519
Vrijling, J.K. 1259, 2797
Vrouwenvelder, A.C.W.M.
2807

Wagner, S. 2541
Walls, L. 987
Walter, M. 1705, 1829
Wang, C. 113
Wang, J. 3109
Wang, W. 523
Wang, Y. 863
Wemmenhove, E. 2295
Weng, W.P. 1267
Wenguo Weng, B. 79
Werbinska, S. 1055,
1851
Wiencke, H.S. 2929
Wiersma, T. 1223
Wiesner, R. 351
Wijnant-Timmerman, S.I.
1223, 2363
Wilday, A.J. 2353
Wilson, S.P. 949
Winder, C. 739, 1081
Winther, R. 183, 2635

Woltjer, R. 19
Woropay, M. 2037
Wu, C.C. 2757
Wu, S.H. 71, 1267
Wu, S.-H. 39, 2405
Xu, S. 2845
Xuewei Ji, A. 79
Yamaguchi, T. 1839
Yamamoto, H. 1715, 1839
Yang, J.-E. 2861
Yannart, B. 2369, 3067
Yeung, T.G. 3171
Yoon, C. 2861
Yu, L.Q. 2845
Yufang, Z. 1943
Yukhymets, P. 1141
Zaitseva, E. 1995
Zajicek, J. 635

859

Zanelli, S. 1199
Zanocco, P. 2899
Zendri, E. 2501
Zerhouni, N. 191
eleznik, N. 3015
Zhang, C. 863, 1663
Zhang, T. 649
Zhao, X. 593
Zhu, D. 113
Zieja, M. 449
Zilber, N. 3231
Zille, V. 531
Zio, E. 477, 703, 709,
1861, 2081, 2101, 2873
Zubeldia, U. 3223
Zuo, M.J. 1723

Zurek,
J. 449
utautaite-eputiene, I.
1867
Zwetkoff, C. 1609

SAFETY, RELIABILITY AND RISK ANALYSIS: THEORY, METHODS

AND APPLICATIONS

PROCEEDINGS OF THE EUROPEAN SAFETY AND RELIABILITY CONFERENCE, ESREL 2008,

AND 17TH SRA-EUROPE, VALENCIA, SPAIN, SEPTEMBER, 2225, 2008

Safety, Reliability and Risk Analysis:

Theory, Methods and Applications

Editors

Sebastin Martorell
Department of Chemical and Nuclear Engineering,
Universidad Politcnica de Valencia, Spain

C. Guedes Soares
Instituto Superior Tcnico, Technical University of Lisbon, Lisbon, Portugal

Julie Barnett
Department of Psychology, University of Surrey, UK

VOLUME 2

Cover picture designed by Centro de Formacin Permanente - Universidad Politcnica de Valencia

CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business
2009 Taylor & Francis Group, London, UK
Typeset by Vikatan Publishing Solutions (P) Ltd., Chennai, India
Printed and bound in Great Britain by Antony Rowe (A CPI-group Company), Chippenham, Wiltshire.
All rights reserved. No part of this publication or the information contained herein may be reproduced, stored
in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by photocopying,
recording or otherwise, without written prior permission from the publisher.
Although all care is taken to ensure integrity and the quality of this publication and the information herein, no
responsibility is assumed by the publishers nor the author for any damage to the property or persons as a result
of operation or use of this publication and/or the information contained herein.
Published by: CRC Press/Balkema
P.O. Box 447, 2300 AK Leiden, The Netherlands
e-mail: [email protected]
www.crcpress.com www.taylorandfrancis.co.uk www.balkema.nl
ISBN: 978-0-415-48513-5 (set of 4 volumes + CD-ROM)
ISBN: 978-0-415-48514-2 (vol 1)
ISBN: 978-0-415-48515-9 (vol 2)
ISBN: 978-0-415-48516-6 (vol 3)
ISBN: 978-0-415-48792-4 (vol 4)
ISBN: 978-0-203-88297-9 (e-book)

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Table of contents

Preface

XXIV

Organization

XXXI

Acknowledgment

XXXV

Introduction

XXXVII

VOLUME 1

Thematic areas
Accident and incident investigation
A code for the simulation of human failure events in nuclear power plants: SIMPROC
J. Gil, J. Espern, L. Gamo, I. Fernndez, P. Gonzlez, J. Moreno, A. Expsito,
C. Queral, G. Rodrguez & J. Hortal

A preliminary analysis of the Tlahuac incident by applying the MORT technique

J.R. Santos-Reyes, S. Olmos-Pea & L.M. Hernndez-Simn

11

Comparing a multi-linear (STEP) and systemic (FRAM) method for accident analysis
I.A. Herrera & R. Woltjer

19

Development of a database for reporting and analysis of near misses in the Italian
chemical industry
R.V. Gagliardi & G. Astarita

27

Development of incident report analysis system based on m-SHEL ontology

Y. Asada, T. Kanno & K. Furuta

33

Forklifts overturn incidents and prevention in Taiwan

K.Y. Chen, S.-H. Wu & C.-M. Shu

39

Formal modelling of incidents and accidents as a means for enriching training material
for satellite control operations
S. Basnyat, P. Palanque, R. Bernhaupt & E. Poupart

45

Hazard factors analysis in regional traffic records

M. Mlynczak & J. Sipa

57

Organizational analysis of availability: What are the lessons for a high risk industrial company?
M. Voirin, S. Pierlot & Y. Dien

63

Thermal explosion analysis of methyl ethyl ketone peroxide by non-isothermal

and isothermal calorimetry application
S.H. Wu, J.M. Tseng & C.M. Shu

71

Crisis and emergency management

A mathematical model for risk analysis of disaster chains
A. Xuewei Ji, B. Wenguo Weng & Pan Li

79

Effective learning from emergency responses

K. Eriksson & J. Borell

83

On the constructive role of multi-criteria analysis in complex decision-making:

An application in radiological emergency management
C. Turcanu, B. Carl, J. Paridaens & F. Hardeman

89

Decision support systems and software tools for safety and reliability
Complex, expert based multi-role assessment system for small and medium enterprises
S.G. Kovacs & M. Costescu

99

DETECT: A novel framework for the detection of attacks to critical infrastructures

F. Flammini, A. Gaglione, N. Mazzocca & C. Pragliola

105

Methodology and software platform for multi-layer causal modeling

K.M. Groth, C. Wang, D. Zhu & A. Mosleh

113

SCAIS (Simulation Code System for Integrated Safety Assessment): Current

status and applications
J.M. Izquierdo, J. Hortal, M. Snchez, E. Melndez, R. Herrero, J. Gil, L. Gamo,
I. Fernndez, J. Espern, P. Gonzlez, C. Queral, A. Expsito & G. Rodrguez
Using GIS and multivariate analyses to visualize risk levels and spatial patterns
of severe accidents in the energy sector
P. Burgherr
Weak signals of potential accidents at Seveso establishments
P.A. Bragatto, P. Agnello, S. Ansaldi & P. Pittiglio

121

129
137

Dynamic reliability
A dynamic fault classification scheme
B. Fechner

147

Importance factors in dynamic reliability

R. Eymard, S. Mercier & M. Roussignol

155

TSD, a SCAIS suitable variant of the SDTPD

J.M. Izquierdo & I. Caamn

163

Fault identification and diagnostics

Application of a vapour compression chiller lumped model for fault detection
J. Navarro-Esbr, A. Real, D. Ginestar & S. Martorell

175

Automatic source code analysis of failure modes causing error propagation

S. Sarshar & R. Winther

183

Development of a prognostic tool to perform reliability analysis

M. El-Koujok, R. Gouriveau & N. Zerhouni

191

Fault detection and diagnosis in monitoring a hot dip galvanizing line using
multivariate statistical process control
J.C. Garca-Daz

201

Fault identification, diagnosis and compensation of flatness errors in hard turning

of tool steels
F. Veiga, J. Fernndez, E. Viles, M. Arizmendi, A. Gil & M.L. Penalva

205

VI

From diagnosis to prognosis: A maintenance experience for an electric locomotive

O. Borgia, F. De Carlo & M. Tucci

211

Human factors
A study on the validity of R-TACOM measure by comparing operator response
time data
J. Park & W. Jung
An evaluation of the Enhanced Bayesian THERP method using simulator data
K. Bladh, J.-E. Holmberg & P. Pyy

221
227

Comparing CESA-Q human reliability analysis with evidence from simulator:

A first attempt
L. Podofillini & B. Reer

233

Exploratory and confirmatory analysis of the relationship between social norms

and safety behavior
C. Fugas, S.A. da Silva & J.L. Meli

243

Functional safety and layer of protection analysis with regard to human factors
K.T. Kosmowski

249

How employees use of information technology systems shape reliable operations

of large scale technological systems
T.K. Andersen, P. Nsje, H. Torvatn & K. Skarholt

259

Incorporating simulator evidence into HRA: Insights from the data analysis of the
international HRA empirical study
S. Massaiu, P.. Braarud & M. Hildebrandt

267

Insights from the HRA international empirical study: How to link data
and HRA with MERMOS
H. Pesme, P. Le Bot & P. Meyer

275

Operators response time estimation for a critical task using the fuzzy logic theory
M. Konstandinidou, Z. Nivolianitou, G. Simos, C. Kiranoudis & N. Markatos

281

The concept of organizational supportiveness

J. Nicholls, J. Harvey & G. Erdos

291

The influence of personal variables on changes in driver behaviour

S. Heslop, J. Harvey, N. Thorpe & C. Mulley

299

The key role of expert judgment in CO2 underground storage projects

C. Vivalda & L. Jammes

305

Integrated risk management and riskinformed decision-making

All-hazards risk frameworkAn architecture model
S. Verga

315

Comparisons and discussion of different integrated risk approaches

R. Steen & T. Aven

323

Management of risk caused by domino effect resulting from design

system dysfunctions
S. Sperandio, V. Robin & Ph. Girard

331

On some aspects related to the use of integrated risk analyses for the decision
making process, including its use in the non-nuclear applications
D. Serbanescu, A.L. Vetere Arellano & A. Colli

341

On the usage of weather derivatives in AustriaAn empirical study

M. Bank & R. Wiesner

VII

351

Precaution in practice? The case of nanomaterial industry

H. Kastenholz, A. Helland & M. Siegrist

361

Risk based maintenance prioritisation

G. Birkeland, S. Eisinger & T. Aven

365

Shifts in environmental health risk governance: An analytical framework

H.A.C. Runhaar, J.P. van der Sluijs & P.P.J. Driessen

369

What does safety margin really mean?

J. Hortal, R. Mendizbal & F. Pelayo

379

Legislative dimensions of risk management

Accidents, risk analysis and safety managementDifferent perspective at a
Swedish safety authority
O. Harrami, M. Strmgren, U. Postgrd & R. All
Evaluation of risk and safety issues at the Swedish Rescue Services Agency
O. Harrami, U. Postgrd & M. Strmgren
Regulation of information security and the impact on top management commitment
A comparative study of the electric power supply sector and the finance sector
J.M. Hagen & E. Albrechtsen
The unintended consequences of risk regulation
B.H. MacGillivray, R.E. Alcock & J.S. Busby

391
399

407
415

Maintenance modelling and optimisation

A hybrid age-based maintenance policy for heterogeneous items
P.A. Scarf, C.A.V. Cavalcante, R.W. Dwight & P. Gordon

423

A stochastic process model for computing the cost of a condition-based maintenance plan
J.A.M. van der Weide, M.D. Pandey & J.M. van Noortwijk

431

A study about influence of uncertain distribution inputs in maintenance optimization

R. Mullor, S. Martorell, A. Snchez & N. Martinez-Alzamora

441

Aging processes as a primary aspect of predicting reliability and life of aeronautical hardware

J. Zurek,
M. Zieja, G. Kowalczyk & T. Niezgoda

449

An alternative imperfect preventive maintenance model

J. Clavareau & P.E. Labeau

455

An imperfect preventive maintenance model with dependent failure modes

I.T. Castro

463

Condition-based maintenance approaches for deteriorating system influenced

by environmental conditions
E. Deloux, B. Castanier & C. Brenguer

469

Condition-based maintenance by particle filtering

F. Cadini, E. Zio & D. Avram

477

Corrective maintenance for aging air conditioning systems

I. Frenkel, L. Khvatskin & A. Lisnianski

483

Exact reliability quantification of highly reliable systems with maintenance

R. Bri

489

Genetic algorithm optimization of preventive maintenance scheduling for repairable

systems modeled by generalized renewal process
P.A.A. Garcia, M.C. SantAna, V.C. Damaso & P.F. Frutuoso e Melo

VIII

497

Maintenance modelling integrating human and material resources

S. Martorell, M. Villamizar, A. Snchez & G. Clemente

505

Modelling competing risks and opportunistic maintenance with expert judgement

T. Bedford & B.M. Alkali

515

Modelling different types of failure and residual life estimation for condition-based maintenance
M.J. Carr & W. Wang

523

Multi-component systems modeling for quantifying complex maintenance strategies

V. Zille, C. Brenguer, A. Grall, A. Despujols & J. Lonchampt

531

Multiobjective optimization of redundancy allocation in systems with imperfect repairs via

ant colony and discrete event simulation
I.D. Lins & E. Lpez Droguett

541

Non-homogeneous Markov reward model for aging multi-state system under corrective
maintenance
A. Lisnianski & I. Frenkel

551

On the modeling of ageing using Weibull models: Case studies

P. Praks, H. Fernandez Bacarizo & P.-E. Labeau

559

On-line condition-based maintenance for systems with several modes of degradation

A. Ponchet, M. Fouladirad & A. Grall

567

Opportunity-based age replacement for a system under two types of failures

F.G. Bada & M.D. Berrade

575

Optimal inspection intervals for maintainable equipment

O. Hryniewicz

581

Optimal periodic inspection of series systems with revealed and unrevealed failures
M. Carvalho, E. Nunes & J. Telhada

587

Optimal periodic inspection/replacement policy for deteriorating systems with explanatory

variables
X. Zhao, M. Fouladirad, C. Brenguer & L. Bordes

593

Optimal replacement policy for components with general failure rates submitted to obsolescence
S. Mercier

603

Optimization of the maintenance function at a company

S. Adjabi, K. Adel-Aissanou & M. Azi

611

Planning and scheduling maintenance resources in a complex system

M. Newby & C. Barker

619

Preventive maintenance planning using prior expert knowledge and multicriteria method
PROMETHEE III
F.A. Figueiredo, C.A.V. Cavalcante & A.T. de Almeida
Profitability assessment of outsourcing maintenance from the producer (big rotary machine study)
P. Fuchs & J. Zajicek

627
635

Simulated annealing method for the selective maintenance optimization of multi-mission

series-parallel systems
A. Khatab, D. Ait-Kadi & A. Artiba

641

Study on the availability of a k-out-of-N System given limited spares under (m, NG )
maintenance policy
T. Zhang, H.T. Lei & B. Guo

649

System value trajectories, maintenance, and its present value

K.B. Marais & J.H. Saleh

IX

659

The maintenance management framework: A practical view to maintenance management

A. Crespo Mrquez, P. Moreu de Len, J.F. Gmez Fernndez, C. Parra Mrquez & V. Gonzlez
Workplace occupation and equipment availability and utilization, in the context of maintenance
float systems
I.S. Lopes, A.F. Leito & G.A.B. Pereira

669

675

Monte Carlo methods in system safety and reliability

Availability and reliability assessment of industrial complex systems: A practical view
applied on a bioethanol plant simulation
V. Gonzlez, C. Parra, J.F. Gmez, A. Crespo & P. Moreu de Len
Handling dependencies between variables with imprecise probabilistic models
S. Destercke & E. Chojnacki
Monte Carlo simulation for investigating the influence of maintenance strategies on the production
availability of offshore installations
K.P. Chang, D. Chang, T.J. Rhee & E. Zio

687
697

703

Reliability analysis of discrete multi-state systems by means of subset simulation

E. Zio & N. Pedroni

709

The application of Bayesian interpolation in Monte Carlo simulations

M. Rajabalinejad, P.H.A.J.M. van Gelder & N. van Erp

717

Occupational safety
Application of virtual reality technologies to improve occupational & industrial safety
in industrial processes
J. Rubio, B. Rubio, C. Vaquero, N. Galarza, A. Pelaz, J.L. Ipia, D. Sagasti & L. Jord
Applying the resilience concept in practice: A case study from the oil and gas industry
L. Hansson, I. Andrade Herrera, T. Kongsvik & G. Solberg

727
733

Development of an assessment tool to facilitate OHS management based upon the safe
place, safe person, safe systems framework
A.-M. Makin & C. Winder

739

Exploring knowledge translation in occupational health using the mental models approach:
A case study of machine shops
A.-M. Nicol & A.C. Hurrell

749

Mathematical modelling of risk factors concerning work-related traffic accidents

C. Santamara, G. Rubio, B. Garca & E. Navarro

757

New performance indicators for the health and safety domain: A benchmarking use perspective
H.V. Neto, P.M. Arezes & S.D. Sousa

761

Occupational risk management for fall from height

O.N. Aneziris, M. Konstandinidou, I.A. Papazoglou, M. Mud, M. Damen, J. Kuiper, H. Baksteen,
L.J. Bellamy, J.G. Post & J. Oh

767

Occupational risk management for vapour/gas explosions

I.A. Papazoglou, O.N. Aneziris, M. Konstandinidou, M. Mud, M. Damen, J. Kuiper, A. Bloemhoff,
H. Baksteen, L.J. Bellamy, J.G. Post & J. Oh

777

Occupational risk of an aluminium industry

O.N. Aneziris, I.A. Papazoglou & O. Doudakmani

787

Risk regulation bureaucracies in EU accession states: Drinking water safety in Estonia

K. Kangur

797

Organization learning
Can organisational learning improve safety and resilience during changes?
S.O. Johnsen & S. Hbrekke

805

Consequence analysis as organizational development

B. Moltu, A. Jarl Ringstad & G. Guttormsen

813

Integrated operations and leadershipHow virtual cooperation influences leadership practice

K. Skarholt, P. Nsje, V. Heps & A.S. Bye

821

Outsourcing maintenance in services providers

J.F. Gmez, C. Parra, V. Gonzlez, A. Crespo & P. Moreu de Len

829

Revising rules and reviving knowledge in the Norwegian railway system

H.C. Blakstad, R. Rosness & J. Hovden

839

Risk Management in systems: Learning to recognize and respond to weak signals

E. Guillaume

847

Author index

853

VOLUME 2
Reliability and safety data collection and analysis
A new step-stress Accelerated Life Testing approach: Step-Down-Stress
C. Zhang, Y. Wang, X. Chen & Y. Jiang

863

Application of a generalized lognormal distribution to engineering data fitting

J. Martn & C.J. Prez

869

Collection and analysis of reliability data over the whole product lifetime of vehicles
T. Leopold & B. Bertsche

875

Comparison of phase-type distributions with mixed and additive Weibull models

M.C. Segovia & C. Guedes Soares

881

Evaluation methodology of industry equipment functional reliability

J. Kamenick

891

Evaluation of device reliability based on accelerated tests

E. Nogueira Daz, M. Vzquez Lpez & D. Rodrguez Cano

899

Evaluation, analysis and synthesis of multiple source information: An application to nuclear

computer codes
S. Destercke & E. Chojnacki

905

Improving reliability using new processes and methods

S.J. Park, S.D. Park & K.T. Jo

913

Life test applied to Brazilian friction-resistant low alloy-high strength steel rails
D.I. De Souza, A. Naked Haddad & D. Rocha Fonseca

919

Non-homogeneous Poisson Process (NHPP), stochastic model applied to evaluate the economic
impact of the failure in the Life Cycle Cost Analysis (LCCA)
C. Parra Mrquez, A. Crespo Mrquez, P. Moreu de Len, J. Gmez Fernndez & V. Gonzlez Daz

929

Risk trends, indicators and learning rates: A new case study of North sea oil and gas
R.B. Duffey & A.B. Skjerve

941

Robust estimation for an imperfect test and repair model using Gaussian mixtures
S.P. Wilson & S. Goyal

949

XI

Risk and evidence based policy making

Environmental reliability as a requirement for defining environmental impact limits
in critical areas
E. Calixto & E. Lbre La Rovere

957

Hazardous aid? The crowding-out effect of international charity

P.A. Raschky & M. Schwindt

965

Individual risk-taking and external effectsAn empirical examination

S. Borsky & P.A. Raschky

973

Licensing a Biofuel plan transforming animal fats

J.-F. David

981

Modelling incident escalation in explosives storage

G. Hardman, T. Bedford, J. Quigley & L. Walls

987

The measurement and management of Deca-BDEWhy the continued certainty of uncertainty?

R.E. Alcock, B.H. McGillivray & J.S. Busby

993

Risk and hazard analysis

A contribution to accelerated testing implementation
S. Hossein Mohammadian M., D. At-Kadi, A. Coulibaly & B. Mutel

1001

A decomposition method to analyze complex fault trees

S. Contini

1009

A quantitative methodology for risk assessment of explosive atmospheres according to the

ATEX directive
R. Lisi, M.F. Milazzo & G. Maschio
A risk theory based on hyperbolic parallel curves and risk assessment in time
N. Popoviciu & F. Baicu
Accident occurrence evaluation of phased-mission systems composed of components
with multiple failure modes
T. Kohda

1019
1027

1035

Added value in fault tree analyses

T. Norberg, L. Rosn & A. Lindhe

1041

Alarm prioritization at plant design stageA simplified approach

P. Barbarini, G. Franzoni & E. Kulot

1049

Analysis of possibilities of timing dependencies modelingExample of logistic support system

J. Magott, T. Nowakowski, P. Skrobanek & S. Werbinska

1055

Applications of supply process reliability model

A. Jodejko & T. Nowakowski

1065

Applying optimization criteria to risk analysis

H. Medina, J. Arnaldos & J. Casal

1073

Chemical risk assessment for inspection teams during CTBT on-site inspections of sites
potentially contaminated with industrial chemicals
G. Malich & C. Winder

1081

Comparison of different methodologies to estimate the evacuation radius in the case

of a toxic release
M.I. Montoya & E. Planas

1089

Conceptualizing and managing risk networks. New insights for risk management
R.W. Schrder

XII

1097

Developments in fault tree techniques and importance measures

J.K. Vaurio

1103

Dutch registration of risk situations

J.P. vant Sant, H.J. Manuel & A. van den Berg

1113

Experimental study of jet fires

M. Gmez-Mares, A. Palacios, A. Peiretti, M. Muoz & J. Casal

1119

Failure mode and effect analysis algorithm for tunneling projects

K. Rezaie, V. Ebrahimipour & S. Shokravi

1125

Fuzzy FMEA: A study case on a discontinuous distillation plant

S.S. Rivera & J.E. Nez Mc Leod

1129

Risk analysis in extreme environmental conditions for Aconcagua Mountain station

J.E. Nez Mc Leod & S.S. Rivera

1135

Geographic information system for evaluation of technical condition and residual life of pipelines
P. Yukhymets, R. Spitsa & S. Kobelsky

1141

Inherent safety indices for the design of layout plans

A. Tugnoli, V. Cozzani, F.I. Khan & P.R. Amyotte

1147

Minmax defense strategy for multi-state systems

G. Levitin & K. Hausken

1157

Multicriteria risk assessment for risk ranking of natural gas pipelines

A.J. de M. Brito, C.A.V. Cavalcante, R.J.P. Ferreira & A.T. de Almeida

1165

New insight into PFDavg and PFH

F. Innal, Y. Dutuit, A. Rauzy & J.-P. Signoret

1173

On causes and dependencies of errors in human and organizational barriers against major
accidents
J.E. Vinnem

1181

Quantitative risk analysis method for warehouses with packaged hazardous materials
D. Riedstra, G.M.H. Laheij & A.A.C. van Vliet

1191

Ranking the attractiveness of industrial plants to external acts of interference

M. Sabatini, S. Zanelli, S. Ganapini, S. Bonvicini & V. Cozzani

1199

Review and discussion of uncertainty taxonomies used in risk analysis

T.E. Nkland & T. Aven

1207

Risk analysis in the frame of the ATEX Directive and the preparation of an Explosion Protection
Document
A. Pey, G. Suter, M. Glor, P. Lerena & J. Campos

1217

Risk reduction by use of a buffer zone

S.I. Wijnant-Timmerman & T. Wiersma

1223

Safety in engineering practice

Z. Smalko & J. Szpytko

1231

Why ISO 13702 and NFPA 15 standards may lead to unsafe design
S. Medonos & R. Raman

1239

Risk control in complex environments

Is there an optimal type for high reliability organization? A study of the UK offshore industry
J.S. Busby, A. Collins & R. Miles

1251

The optimization of system safety: Rationality, insurance, and optimal protection

R.B. Jongejan & J.K. Vrijling

1259

XIII

Thermal characteristic analysis of Y type zeolite by differential scanning calorimetry

S.H. Wu, W.P. Weng, C.C. Hsieh & C.M. Shu
Using network methodology to define emergency response team location: The Brazilian
refinery case study
E. Calixto, E. Lbre La Rovere & J. Eustquio Beraldo

1267

1273

Risk perception and communication

(Mis-)conceptions of safety principles
J.-T. Gayen & H. Schbe

1283

Climate change in the British press: The role of the visual

N.W. Smith & H. Joffe

1293

Do the people exposed to a technological risk always want more information about it?
Some observations on cases of rejection
J. Espluga, J. Farr, J. Gonzalo, T. Horlick-Jones, A. Prades, C. Oltra & J. Navajas

1301

Media coverage, imaginary of risks and technological organizations

F. Fodor & G. Deleuze

1309

Media disaster coverage over time: Methodological issues and results

M. Kuttschreuter & J.M. Gutteling

1317

Risk amplification and zoonosis

D.G. Duckett & J.S. Busby

1325

Risk communication and addressing uncertainties in risk assessmentsPresentation of a framework

J. Stian strem, H. Thevik, R. Flage & T. Aven

1335

Risk communication for industrial plants and radioactive waste repositories

F. Gugliermetti & G. Guidi

1341

Risk management measurement methodology: Practical procedures and approaches for risk
assessment and prediction
R.B. Duffey & J.W. Saull

1351

Risk perception and cultural theory: Criticism and methodological orientation

C. Kermisch & P.-E. Labeau

1357

Standing in the shoes of hazard managers: An experiment on avalanche risk perception

C.M. Rheinberger

1365

The social perception of nuclear fusion: Investigating lay understanding and reasoning about
the technology
A. Prades, C. Oltra, J. Navajas, T. Horlick-Jones & J. Espluga

1371

Safety culture
Us and Them: The impact of group identity on safety critical behaviour
R.J. Bye, S. Antonsen & K.M. Vikland

1377

Does change challenge safety? Complexity in the civil aviation transport system
S. Hyland & K. Aase

1385

Electromagnetic fields in the industrial enviroment

J. Fernndez, A. Quijano, M.L. Soriano & V. Fuster

1395

Electrostatic charges in industrial environments

P. LLovera, A. Quijano, A. Soria & V. Fuster

1401

Empowering operations and maintenance: Safe operations with the one directed team
organizational model at the Kristin asset
P. Nsje, K. Skarholt, V. Heps & A.S. Bye

XIV

1407

Leadership and safety climate in the construction industry

J.L. Meli, M. Becerril, S.A. Silva & K. Mearns

1415

Local management and its impact on safety culture and safety within Norwegian shipping
H.A Oltedal & O.A. Engen

1423

Quantitative analysis of the anatomy and effectiveness of occupational safety culture

P. Trucco, M. De Ambroggi & O. Grande

1431

Safety management and safety culture assessment in Germany

H.P. Berg

1439

The potential for error in communications between engineering designers

J. Harvey, R. Jamieson & K. Pearce

1447

Safety management systems

Designing the safety policy of IT system on the example of a chosen company
I.J. Jzwiak, T. Nowakowski & K. Jzwiak
Determining and verifying the safety integrity level of the control and protection systems
under uncertainty
T. Barnert, K.T. Kosmowski & M. Sliwinski

1455

1463

Drawing up and running a Security Plan in an SME type companyAn easy task?
M. Gerbec

1473

Efficient safety management for subcontractor at construction sites

K.S. Son & J.J. Park

1481

Production assurance and reliability managementA new international standard

H. Kortner, K.-E. Haugen & L. Sunde

1489

Risk management model for industrial plants maintenance

N. Napolitano, M. De Minicis, G. Di Gravio & M. Tronci

1495

Some safety aspects on multi-agent and CBTC implementation for subway control systems
F.M. Rachel & P.S. Cugnasca

1503

Software reliability
Assessment of software reliability and the efficiency of corrective actions during the software
development process
R. Savic

1513

ERTMS, deals on wheels? An inquiry into a major railway project

J.A. Stoop, J.H. Baggen, J.M. Vleugel & J.L.M. Vrancken

1519

Guaranteed resource availability in a website

V.P. Koutras & A.N. Platis

1525

Reliability oriented electronic design automation tool

J. Marcos, D. Bveda, A. Fernndez & E. Soto

1533

Reliable software for partitionable networked environmentsAn experience report

S. Beyer, J.C. Garca Ortiz, F.D. Muoz-Esco, P. Galdmez, L. Froihofer,
K.M. Goeschka & J. Osrael

1539

SysML aided functional safety assessment

M. Larisch, A. Hnle, U. Siebold & I. Hring

1547

UML safety requirement specification and verification

A. Hnle & I. Hring

1555

XV

Stakeholder and public involvement in risk governance

Assessment and monitoring of reliability and robustness of offshore wind energy converters
S. Thns, M.H. Faber, W. Rcker & R. Rohrmann

1567

Building resilience to natural hazards. Practices and policies on governance and mitigation
in the central region of Portugal
J.M. Mendes & A.T. Tavares

1577

Governance of flood risks in The Netherlands: Interdisciplinary research into the role and
meaning of risk perception
M.S. de Wit, H. van der Most, J.M. Gutteling & M. Bockarjova

1585

Public intervention for better governanceDoes it matter? A study of the Leros Strength case
P.H. Linde & J.E. Karlsen

1595

Reasoning about safety management policy in everyday terms

T. Horlick-Jones

1601

Using stakeholders expertise in EMF and soil contamination to improve the management
of public policies dealing with modern risk: When uncertainty is on the agenda
C. Fallon, G. Joris & C. Zwetkoff

1609

Structural reliability and design codes

Adaptive discretization of 1D homogeneous random fields
D.L. Allaix, V.I. Carbone & G. Mancini

1621

Comparison of methods for estimation of concrete strength

M. Holicky, K. Jung & M. Sykora

1629

Design of structures for accidental design situations

J. Markov & K. Jung

1635

Developing fragility function for a timber structure subjected to fire

E.R. Vaidogas, Virm. Juocevicius & Virg. Juocevicius

1641

Estimations in the random fatigue-limit model

C.-H. Chiu & W.-T. Huang

1651

Limitations of the Weibull distribution related to predicting the probability of failure

initiated by flaws
M.T. Todinov

1655

Simulation techniques of non-gaussian random loadings in structural reliability analysis

Y. Jiang, C. Zhang, X. Chen & J. Tao

1663

Special features of the collection and analysis of snow loads

Z. Sadovsk, P. Fako, K. Mikulov, J. Pecho & M. Vojtek

1671

Structural safety under extreme construction loads

V. Juocevicius & A. Kudzys

1677

The modeling of time-dependent reliability of deteriorating structures

A. Kudzys & O. Lukoeviciene

1685

Author index

1695

VOLUME 3
System reliability analysis
A copula-based approach for dependability analyses of fault-tolerant systems with
interdependent basic events
M. Walter, S. Esch & P. Limbourg

XVI

1705

A depth first search algorithm for optimal arrangements in a circular

consecutive-k-out-of-n:F system
K. Shingyochi & H. Yamamoto

1715

A joint reliability-redundancy optimization approach for multi-state series-parallel systems

Z. Tian, G. Levitin & M.J. Zuo

1723

A new approach to assess the reliability of a multi-state system with dependent components
M. Samrout & E. Chatelet

1731

A reliability analysis and decision making process for autonomous systems

R. Remenyte-Prescott, J.D. Andrews, P.W.H. Chung & C.G. Downes

1739

Advanced discrete event simulation methods with application to importance measure

estimation
A.B. Huseby, K.A. Eide, S.L. Isaksen, B. Natvig & J. Gsemyr

1747

Algorithmic and computational analysis of a multi-component complex system

J.E. Ruiz-Castro, R. Prez-Ocn & G. Fernndez-Villodre

1755

An efficient reliability computation of generalized multi-state k-out-of-n systems

S.V. Amari

1763

Application of the fault tree analysis for assessment of the power system reliability

A. Volkanovski, M. Cepin
& B. Mavko

1771

BDMP (Boolean logic driven Markov processes) as an alternative to event trees

M. Bouissou

1779

Bivariate distribution based passive system performance assessment

L. Burgazzi

1787

Calculating steady state reliability indices of multi-state systems using dual number algebra
E. Korczak

1795

Concordance analysis of importance measure

C.M. Rocco S.

1803

Contribution to availability assessment of systems with one shot items

D. Valis & M. Koucky

1807

Contribution to modeling of complex weapon systems reliability

D. Valis, Z. Vintr & M. Koucky

1813

Delayed system reliability and uncertainty analysis

R. Alzbutas, V. Janilionis & J. Rimas

1819

Efficient generation and representation of failure lists out of an information flux model
for modeling safety critical systems
M. Pock, H. Belhadaoui, O. Malass & M. Walter

1829

Evaluating algorithms for the system state distribution of multi-state k-out-of-n:F system
T. Akiba, H. Yamamoto, T. Yamaguchi, K. Shingyochi & Y. Tsujimura

1839

First-passage time analysis for Markovian deteriorating model

G. Dohnal

1847

Model of logistic support system with time dependency

S. Werbinska

1851

Modeling failure cascades in network systems due to distributed random disturbances

E. Zio & G. Sansavini

1861

Modeling of the changes of graphite bore in RBMK-1500 type nuclear reactor

I. utautaite-eputiene, J. Augutis & E. Upuras

1867

XVII

Modelling multi-platform phased mission system reliability

D.R. Prescott, J.D. Andrews & C.G. Downes
Modelling test strategies effects on the probability of failure on demand for safety
instrumented systems
A.C. Torres-Echeverria, S. Martorell & H.A. Thompson

1873

1881

New insight into measures of component importance in production systems

S.L. Isaksen

1891

New virtual age models for bathtub shaped failure intensities

Y. Dijoux & E. Ide

1901

On some approaches to defining virtual age of non-repairable objects

M.S. Finkelstein

1909

On the application and extension of system signatures in engineering reliability

J. Navarro, F.J. Samaniego, N. Balakrishnan & D. Bhattacharya

1915

PFD of higher-order configurations of SIS with partial stroke testing capability

L.F.S. Oliveira

1919

Power quality as accompanying factor in reliability research of electric engines

I.J. Jzwiak, K. Kujawski & T. Nowakowski

1929

RAMS and performance analysis

X. Quayzin, E. Arbaretier, Z. Brik & A. Rauzy

1937

Reliability evaluation of complex system based on equivalent fault tree

Z. Yufang, Y. Hong & L. Jun

1943

Reliability evaluation of III-V Concentrator solar cells

N. Nez, J.R. Gonzlez, M. Vzquez, C. Algora & I. Rey-Stolle

1949

Reliability of a degrading system under inspections

D. Montoro-Cazorla, R. Prez-Ocn & M.C. Segovia

1955

Reliability prediction using petri nets for on-demand safety systems with fault detection
A.V. Kleyner & V. Volovoi

1961

Reliability, availability and cost analysis of large multi-state systems with ageing components
K. Kolowrocki

1969

Reliability, availability and risk evaluation of technical systems in variable operation conditions
K. Kolowrocki & J. Soszynska

1985

Representation and estimation of multi-state system reliability by decision diagrams

E. Zaitseva & S. Puuronen

1995

Safety instrumented system reliability evaluation with influencing factors

F. Brissaud, D. Charpentier, M. Fouladirad, A. Barros & C. Brenguer

2003

Smooth estimation of the availability function of a repairable system

M.L. Gmiz & Y. Romn

2013

System design optimisation involving phased missions

D. Astapenko & L.M. Bartlett

2021

The Natvig measures of component importance in repairable systems applied to an offshore

oil and gas production system
B. Natvig, K.A. Eide, J. Gsemyr, A.B. Huseby & S.L. Isaksen

2029

The operation quality assessment as an initial part of reliability improvement and low cost
automation of the system
L. Muslewski, M. Woropay & G. Hoppe

2037

XVIII

Three-state modelling of dependent component failures with domino effects

U.K. Rakowsky

2045

Variable ordering techniques for the application of Binary Decision Diagrams on PSA
linked Fault Tree models
C. Ibez-Llano, A. Rauzy, E. Melndez & F. Nieto

2051

Weaknesses of classic availability calculations for interlinked production systems

and their overcoming
D. Achermann

2061

Uncertainty and sensitivity analysis

A critique of Info-Gaps robustness model
M. Sniedovich

2071

Alternative representations of uncertainty in system reliability and risk analysisReview

and discussion
R. Flage, T. Aven & E. Zio

2081

Dependence modelling with copula in probabilistic studies, a practical approach

based on numerical experiments
A. Dutfoy & R. Lebrun

2093

Event tree uncertainty analysis by Monte Carlo and possibility theory

P. Baraldi & E. Zio

2101

Global sensitivity analysis based on entropy

B. Auder & B. Iooss

2107

Impact of uncertainty affecting reliability models on warranty contracts

G. Fleurquin, P. Dehombreux & P. Dersin

2117

Influence of epistemic uncertainties on the probabilistic assessment of an emergency operating

procedure in a nuclear power plant
M. Kloos & J. Peschke

2125

Numerical study of algorithms for metamodel construction and validation

B. Iooss, L. Boussouf, A. Marrel & V. Feuillard

2135

On the variance upper bound theorem and its applications

M.T. Todinov

2143

Reliability assessment under Uncertainty Using Dempster-Shafer and Vague Set Theories
S. Pashazadeh & N. Grachorloo

2151

Types and sources of uncertainties in environmental accidental risk assessment: A case study
for a chemical factory in the Alpine region of Slovenia
M. Gerbec & B. Kontic
Uncertainty estimation for monotone and binary systems
A.P. Ulmeanu & N. Limnios

2157
2167

Industrial and service sectors

Aeronautics and aerospace
Condition based operational risk assessment for improved aircraft operability
A. Arnaiz, M. Buderath & S. Ferreiro

2175

Is optimized design of satellites possible?

J. Faure, R. Laulheret & A. Cabarbaye

2185

XIX

Model of air traffic in terminal area for ATFM safety analysis

J. Skorupski & A.W. Stelmach

2191

Predicting airport runway conditions based on weather data

A.B. Huseby & M. Rabbe

2199

Safety considerations in complex airborne systems

M.J.R. Lemes & J.B. Camargo Jr

2207

The Preliminary Risk Analysis approach: Merging space and aeronautics methods
J. Faure, R. Laulheret & A. Cabarbaye

2217

Using a Causal model for Air Transport Safety (CATS) for the evaluation of alternatives
B.J.M. Ale, L.J. Bellamy, R.P. van der Boom, J. Cooper, R.M. Cooke, D. Kurowicka, P.H. Lin,
O. Morales, A.L.C. Roelen & J. Spouge

2223

Automotive engineering
An approach to describe interactions in and between mechatronic systems
J. Gng & B. Bertsche

2233

Influence of the mileage distribution on reliability prognosis models

A. Braasch, D. Althaus & A. Meyna

2239

Reliability prediction for automotive components using Real-Parameter Genetic Algorithm

J. Hauschild, A. Kazeminia & A. Braasch

2245

Stochastic modeling and prediction of catalytic converters degradation

S. Barone, M. Giorgio, M. Guida & G. Pulcini

2251

Towards a better interaction between design and dependability analysis: FMEA derived from
UML/SysML models
P. David, V. Idasiak & F. Kratz

2259

Biotechnology and food industry

Application of tertiary mathematical models for evaluating the presence of staphylococcal
enterotoxin in lactic acid cheese
I. Steinka & A. Blokus-Roszkowska

2269

Assessment of the risk to company revenue due to deviations in honey quality

E. Domnech, I. Escriche & S. Martorell

2275

Attitudes of Japanese and Hawaiian toward labeling genetically modified fruits

S. Shehata

2285

Ensuring honey quality by means of effective pasteurization

E. Domnech, I. Escriche & S. Martorell

2289

Exposure assessment model to combine thermal inactivation (log reduction) and thermal injury
(heat-treated spore lag time) effects on non-proteolytic Clostridium botulinum
J.-M. Membr, E. Wemmenhove & P. McClure

2295

Public information requirements on health risk of mercury in fish (1): Perception

and knowledge of the public about food safety and the risk of mercury
M. Kosugi & H. Kubota

2305

Public information requirements on health risks of mercury in fish (2): A comparison of mental
models of experts and public in Japan
H. Kubota & M. Kosugi

2311

Review of diffusion models for the social amplification of risk of food-borne zoonoses
J.P. Mehers, H.E. Clough & R.M. Christley

XX

2317

Risk perception and communication of food safety and food technologies in Flanders,
The Netherlands, and the United Kingdom
U. Maris
Synthesis of reliable digital microfluidic biochips using Monte Carlo simulation
E. Maftei, P. Pop & F. Popentiu Vladicescu

2325
2333

Chemical process industry

Accidental scenarios in the loss of control of chemical processes: Screening the impact
profile of secondary substances
M. Cordella, A. Tugnoli, P. Morra, V. Cozzani & F. Barontini
Adapting the EU Seveso II Directive for GHS: Initial UK study on acute toxicity to people
M.T. Trainor, A.J. Wilday, M. Moonis, A.L. Rowbotham, S.J. Fraser, J.L. Saw & D.A. Bosworth

2345
2353

An advanced model for spreading and evaporation of accidentally released hazardous

liquids on land
I.J.M. Trijssenaar-Buhre, R.P. Sterkenburg & S.I. Wijnant-Timmerman

2363

Influence of safety systems on land use planning around seveso sites; example of measures
chosen for a fertiliser company located close to a village
C. Fivez, C. Delvosalle, N. Cornil, L. Servranckx, F. Tambour, B. Yannart & F. Benjelloun

2369

Performance evaluation of manufacturing systems based on dependability management

indicators-case study: Chemical industry
K. Rezaie, M. Dehghanbaghi & V. Ebrahimipour

2379

Protection of chemical industrial installations from intentional adversary acts: Comparison

of the new security challenges with the existing safety practices in Europe
M.D. Christou

2389

Quantitative assessment of domino effect in an extended industrial area

G. Antonioni, G. Spadoni, V. Cozzani, C. Dondi & D. Egidi

2397

Reaction hazard of cumene hydroperoxide with sodium hydroxide by isothermal calorimetry

Y.-P. Chou, S.-H. Wu, C.-M. Shu & H.-Y. Hou

2405

Reliability study of shutdown process through the analysis of decision making in chemical plants.
Case of study: South America, Spain and Portugal
L. Amendola, M.A. Artacho & T. Depool

2409

Study of the application of risk management practices in shutdown chemical process

L. Amendola, M.A. Artacho & T. Depool

2415

Thirty years after the first HAZOP guideline publication. Considerations

J. Dunj, J.A. Vlchez & J. Arnaldos

2421

Civil engineering
Decision tools for risk management support in construction industry
S. Mehicic Eberhardt, S. Moeller, M. Missler-Behr & W. Kalusche

2431

Definition of safety and the existence of optimal safety

D. Proske

2441

Failure risk analysis in Water Supply Networks

A. Carrin, A. Debn, E. Cabrera, M.L. Gamiz & H. Solano

2447

Hurricane vulnerability of multi-story residential buildings in Florida

G.L. Pita, J.-P. Pinelli, C.S. Subramanian, K. Gurley & S. Hamid

2453

Risk management system in water-pipe network functioning

B. Tchrzewska-Cieslak

2463

XXI

Use of extreme value theory in engineering design

E. Castillo, C. Castillo & R. Mnguez

2473

Critical infrastructures
A model for vulnerability analysis of interdependent infrastructure networks
J. Johansson & H. Jnsson
Exploiting stochastic indicators of interdependent infrastructures: The service availability of
interconnected networks
G. Bonanni, E. Ciancamerla, M. Minichino, R. Clemente, A. Iacomini, A. Scarlatti,
E. Zendri & R. Terruggia
Proactive risk assessment of critical infrastructures
T. Uusitalo, R. Koivisto & W. Schmitz

2491

2501

2511

Seismic assessment of utility systems: Application to water, electric power and transportation
networks
C. Nuti, A. Rasulo & I. Vanzi

2519

Author index

2531

VOLUME 4
Electrical and electronic engineering
Balancing safety and availability for an electronic protection system
S. Wagner, I. Eusgeld, W. Krger & G. Guaglio
Evaluation of important reliability parameters using VHDL-RTL modelling and information
flow approach
M. Jallouli, C. Diou, F. Monteiro, A. Dandache, H. Belhadaoui, O. Malass, G. Buchheit,
J.F. Aubry & H. Medromi

2541

2549

Energy production and distribution

Application of Bayesian networks for risk assessment in electricity distribution system
maintenance management
D.E. Nordgrd & K. Sand

2561

Incorporation of ageing effects into reliability model for power transmission network
V. Matuzas & J. Augutis

2569

Mathematical simulation of energy supply disturbances

J. Augutis, R. Kriktolaitis, V. Matuziene & S. Peciulyte

2575

Risk analysis of the electric power transmission grid

L.M. Pedersen & H.H. Thorstad

2581

Security of gas supply to a gas plant from cave storage using discrete-event simulation
J.D. Amaral Netto, L.F.S. Oliveira & D. Faertes

2587

SES RISK a new framework to support decisions on energy supply

D. Serbanescu & A.L. Vetere Arellano

2593

Specification of reliability benchmarks for offshore wind farms

D. McMillan & G.W. Ault

2601

Health and medicine

Bayesian statistical meta-analysis of epidemiological data for QRA
I. Albert, E. Espi, A. Gallay, H. De Valk, E. Grenier & J.-B. Denis

XXII

2609

Cyanotoxins and health risk assessment

J. Kellner, F. Boek, J. Navrtil & J. Dvork

2613

The estimation of health effect risks based on different sampling intervals of meteorological data
J. Jeong & S. Hoon Han

2619

Information technology and telecommunications

A bi-objective model for routing and wavelength assignment in resilient WDM networks
T. Gomes, J. Craveirinha, C. Simes & J. Clmaco

2627

Formal reasoning regarding error propagation in multi-process software architectures

F. Stre & R. Winther

2635

Implementation of risk and reliability analysis techniques in ICT

R. Mock, E. Kollmann & E. Bnzli

2641

Information security measures influencing user performance

E. Albrechtsen & J.M. Hagen

2649

Reliable network server assignment using an ant colony approach

S. Kulturel-Konak & A. Konak

2657

Risk and safety as system-theoretic conceptsA formal view on system-theory

by means of petri-nets
J. Rudolf Mller & E. Schnieder

2665

Insurance and finance

Behaviouristic approaches to insurance decisions in the context of natural hazards
M. Bank & M. Gruber
Gaming tool as a method of natural disaster risk education: Educating the relationship
between risk and insurance
T. Unagami, T. Motoyoshi & J. Takai
Reliability-based risk-metric computation for energy trading
R. Mnguez, A.J. Conejo, R. Garca-Bertrand & E. Castillo

2675

2685
2689

Manufacturing
A decision model for preventing knock-on risk inside industrial plant
M. Grazia Gnoni, G. Lettera & P. Angelo Bragatto

2701

Condition based maintenance optimization under cost and profit criteria for manufacturing
equipment
A. Snchez, A. Goti & V. Rodrguez

2707

PRA-type study adapted to the multi-crystalline silicon photovoltaic cells manufacture

process
A. Colli, D. Serbanescu & B.J.M. Ale

2715

Mechanical engineering
Developing a new methodology for OHS assessment in small and medium enterprises
C. Pantanali, A. Meneghetti, C. Bianco & M. Lirussi

2727

Optimal Pre-control as a tool to monitor the reliability of a manufacturing system

S. San Matas & V. Giner-Bosch

2735

The respirable crystalline silica in the ceramic industriesSampling, exposure

and toxicology
E. Monfort, M.J. Ibez & A. Escrig

XXIII

2743

Natural hazards
A framework for the assessment of the industrial risk caused by floods
M. Campedel, G. Antonioni, V. Cozzani & G. Di Baldassarre

2749

A simple method of risk potential analysis for post-earthquake fires

J.L. Su, C.C. Wu, K.S. Fan & J.R. Chen

2757

Applying the SDMS model to manage natural disasters in Mexico

J.R. Santos-Reyes & A.N. Beard

2765

Decision making tools for natural hazard risk managementExamples from Switzerland
M. Brndl, B. Krummenacher & H.M. Merz

2773

How to motivate people to assume responsibility and act upon their own protection from flood
risk in The Netherlands if they think they are perfectly safe?
M. Bockarjova, A. van der Veen & P.A.T.M. Geurts

2781

Integral risk management of natural hazardsA system analysis of operational application

to rapid mass movements
N. Bischof, H. Romang & M. Brndl

2789

Risk based approach for a long-term solution of coastal flood defencesA Vietnam case
C. Mai Van, P.H.A.J.M. van Gelder & J.K. Vrijling

2797

River system behaviour effects on flood risk

T. Schweckendiek, A.C.W.M. Vrouwenvelder, M.C.L.M. van Mierlo, E.O.F. Calle & W.M.G. Courage

2807

Valuation of flood risk in The Netherlands: Some preliminary results

M. Bockarjova, P. Rietveld & E.T. Verhoef

2817

Nuclear engineering
An approach to integrate thermal-hydraulic and probabilistic analyses in addressing
safety margins estimation accounting for uncertainties
S. Martorell, Y. Nebot, J.F. Villanueva, S. Carlos, V. Serradell, F. Pelayo & R. Mendizbal

2827

Availability of alternative sources for heat removal in case of failure of the RHRS during
midloop conditions addressed in LPSA
J.F. Villanueva, S. Carlos, S. Martorell, V. Serradell, F. Pelayo & R. Mendizbal

2837

Complexity measures of emergency operating procedures: A comparison study with data

from a simulated computerized procedure experiment
L.Q. Yu, Z.Z. Li, X.L. Dong & S. Xu

2845

Distinction impossible!: Comparing risks between Radioactive Wastes Facilities and Nuclear
Power Stations
S. Kim & S. Cho

2851

Heat-up calculation to screen out the room cooling failure function from a PSA model
M. Hwang, C. Yoon & J.-E. Yang
Investigating the material limits on social construction: Practical reasoning about nuclear
fusion and other technologies
T. Horlick-Jones, A. Prades, C. Oltra, J. Navajas & J. Espluga

2861

2867

Neural networks and order statistics for quantifying nuclear power plants safety margins
E. Zio, F. Di Maio, S. Martorell & Y. Nebot

2873

Probabilistic safety assessment for other modes than power operation

M. Cepin
& R. Prosen

2883

Probabilistic safety margins: Definition and calculation

R. Mendizbal

2891

XXIV

Reliability assessment of the thermal hydraulic phenomena related to a CAREM-like

passive RHR System
G. Lorenzo, P. Zanocco, M. Gimnez, M. Marqus, B. Iooss, R. Bolado Lavn, F. Pierro,
G. Galassi, F. DAuria & L. Burgazzi
Some insights from the observation of nuclear power plant operators management of simulated
abnormal situations
M.C. Kim & J. Park
Vital area identification using fire PRA and RI-ISI results in UCN 4 nuclear power plant
K.Y. Kim, Y. Choi & W.S. Jung

2899

2909
2913

Offshore oil and gas

A new approach for follow-up of safety instrumented systems in the oil and gas industry
S. Hauge & M.A. Lundteigen
Consequence based methodology to determine acceptable leakage rate through closed safety
critical valves
W. Red, K. Haver, H.S. Wiencke & T.E. Nkland

2921

2929

FAMUS: Applying a new tool for integrating flow assurance and RAM analysis
. Grande, S. Eisinger & S.L. Isaksen

2937

Fuzzy reliability analysis of corroded oil and gas pipes

M. Singh & T. Markeset

2945

Life cycle cost analysis in design of oil and gas production facilities to be used in harsh,
remote and sensitive environments
D. Kayrbekova & T. Markeset
Line pack management for improved regularity in pipeline gas transportation networks
L. Frimannslund & D. Haugland

2955
2963

Optimization of proof test policies for safety instrumented systems using multi-objective
genetic algorithms
A.C. Torres-Echeverria, S. Martorell & H.A. Thompson

2971

Paperwork, management, and safety: Towards a bureaucratization of working life

and a lack of hands-on supervision
G.M. Lamvik, P.C. Nsje, K. Skarholt & H. Torvatn

2981

Preliminary probabilistic study for risk management associated to casing long-term integrity
in the context of CO2 geological sequestrationRecommendations for cement plug geometry
Y. Le Guen, O. Poupard, J.-B. Giraud & M. Loizzo

2987

Risk images in integrated operations

C.K. Tveiten & P.M. Schiefloe

2997

Policy decisions
Dealing with nanotechnology: Do the boundaries matter?
S. Brunet, P. Delvenne, C. Fallon & P. Gillon

3007

Factors influencing the public acceptability of the LILW repository

N. eleznik, M. Polic & D. Kos

3015

Risk futures in Europe: Perspectives for future research and governance. Insights from a EU
funded project
S. Menoni
Risk management strategies under climatic uncertainties
U.S. Brandt

XXV

3023
3031

Safety representative and managers: Partners in health and safety?

T. Kvernberg Andersen, H. Torvatn & U. Forseth

3039

Stop in the name of safetyThe right of the safety representative to halt dangerous work
U. Forseth, H. Torvatn & T. Kvernberg Andersen

3047

The VDI guideline on requirements for the qualification of reliability engineersCurriculum

and certification process
U.K. Rakowsky

3055

Public planning
Analysing analysesAn approach to combining several risk and vulnerability analyses
J. Borell & K. Eriksson
Land use planning methodology used in Walloon region (Belgium) for tank farms of gasoline
and diesel oil
F. Tambour, N. Cornil, C. Delvosalle, C. Fivez, L. Servranckx, B. Yannart & F. Benjelloun

3061

3067

Security and protection

Protection from half-criminal windows breakers to mass murderers with nuclear weapons:
Changes in the Norwegian authorities discourses on the terrorism threat
S.H. Jore & O. Nj

3077

A preliminary analysis of volcanic Na-Tech risks in the Vesuvius area

E. Salzano & A. Basco

3085

Are safety and security in industrial systems antagonistic or complementary issues?

G. Deleuze, E. Chatelet, P. Laclemence, J. Piwowar & B. Affeltranger

3093

Assesment of energy supply security indicators for Lithuania

J. Augutis, R. Kriktolaitis, V. Matuziene & S. Peciulyte

3101

Enforcing application securityFixing vulnerabilities with aspect oriented programming

J. Wang & J. Bigham

3109

Governmental risk communication: Communication guidelines in the context of terrorism

as a new risk
I. Stevens & G. Verleye

3117

On combination of Safety Integrity Levels (SILs) according to IEC61508 merging rules

Y. Langeron, A. Barros, A. Grall & C. Brenguer

3125

On the methods to model and analyze attack scenarios with Fault Trees
G. Renda, S. Contini & G.G.M. Cojazzi

3135

Risk management for terrorist actions using geoevents

G. Maschio, M.F. Milazzo, G. Ancione & R. Lisi

3143

Surface transportation (road and train)

A modelling approach to assess the effectiveness of BLEVE prevention measures on LPG tanks
G. Landucci, M. Molag, J. Reinders & V. Cozzani

3153

Availability assessment of ALSTOMs safety-relevant trainborne odometry sub-system

B.B. Stamenkovic & P. Dersin

3163

Dynamic maintenance policies for civil infrastructure to minimize cost and manage safety risk
T.G. Yeung & B. Castanier

3171

FAI: Model of business intelligence for projects in metrorailway system

A. Oliveira & J.R. Almeida Jr.

3177

XXVI

Impact of preventive grinding on maintenance costs and determination of an optimal grinding cycle
C. Meier-Hirmer & Ph. Pouligny

3183

Logistics of dangerous goods: A GLOBAL risk assessment approach

C. Mazri, C. Deust, B. Nedelec, C. Bouissou, J.C. Lecoze & B. Debray

3191

Optimal design of control systems using a dependability criteria and temporal sequences
evaluationApplication to a railroad transportation system
J. Clarhaut, S. Hayat, B. Conrard & V. Cocquempot

3199

RAM assurance programme carried out by the Swiss Federal Railways SA-NBS project
B.B. Stamenkovic

3209

RAMS specification for an urban transit Maglev system

A. Raffetti, B. Faragona, E. Carfagna & F. Vaccaro

3217

Safety analysis methodology application into two industrial cases: A new mechatronical system
and during the life cycle of a CAFs high speed train
O. Revilla, A. Arnaiz, L. Susperregui & U. Zubeldia

3223

The ageing of signalling equipment and the impact on maintenance strategies

M. Antoni, N. Zilber, F. Lejette & C. Meier-Hirmer

3231

The development of semi-Markov transportation model

Z. Mateusz & B. Tymoteusz

3237

Valuation of operational architecture dependability using Safe-SADT formalism: Application

to a railway braking system
D. Renaux, L. Cauffriez, M. Bayart & V. Benard

3245

Waterborne transportation
A simulation based risk analysis study of maritime traffic in the Strait of Istanbul
B. zbas, I. Or, T. Altiok & O.S. Ulusu

3257

Analysis of maritime accident data with BBN models

P. Anto, C. Guedes Soares, O. Grande & P. Trucco

3265

Collision risk analyses of waterborne transportation

E. Vanem, R. Skjong & U. Langbecker

3275

Complex model of navigational accident probability assessment based on real time

simulation and manoeuvring cycle concept
L. Gucma

3285

Design of the ship power plant with regard to the operator safety
A. Podsiadlo & W. Tarelko

3289

Human fatigue model at maritime transport

L. Smolarek & J. Soliwoda

3295

Modeling of hazards, consequences and risk for safety assessment of ships in damaged
conditions in operation
M. Gerigk

3303

Numerical and experimental study of a reliability measure for dynamic control of floating vessels
B.J. Leira, P.I.B. Berntsen & O.M. Aamo

3311

Reliability of overtaking maneuvers between ships in restricted area

P. Lizakowski

3319

Risk analysis of ports and harborsApplication of reliability engineering techniques

B.B. Dutta & A.R. Kar

3323

XXVII

Subjective propulsion risk of a seagoing ship estimation

A. Brandowski, W. Frackowiak, H. Nguyen & A. Podsiadlo

3331

The analysis of SAR action effectiveness parameters with respect to drifting search area model
Z. Smalko & Z. Burciu

3337

The risk analysis of harbour operations

T. Abramowicz-Gerigk

3343

Author index

3351

XXVIII

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Preface

This Conference stems from a European initiative merging the ESRA (European Safety and Reliability
Association) and SRA-Europe (Society for Risk AnalysisEurope) annual conferences into the major safety,
reliability and risk analysis conference in Europe during 2008. This is the second joint ESREL (European Safety
and Reliability) and SRA-Europe Conference after the 2000 event held in Edinburg, Scotland.
ESREL is an annual conference series promoted by the European Safety and Reliability Association. The
conference dates back to 1989, but was not referred to as an ESREL conference before 1992. The Conference
has become well established in the international community, attracting a good mix of academics and industry
participants that present and discuss subjects of interest and application across various industries in the fields of
Safety and Reliability.
The Society for Risk AnalysisEurope (SRA-E) was founded in 1987, as a section of SRA international
founded in 1981, to develop a special focus on risk related issues in Europe. SRA-E aims to bring together
individuals and organisations with an academic interest in risk assessment, risk management and risk communication in Europe and emphasises the European dimension in the promotion of interdisciplinary approaches of
risk analysis in science. The annual conferences take place in various countries in Europe in order to enhance the
access to SRA-E for both members and other interested parties. Recent conferences have been held in Stockholm,
Paris, Rotterdam, Lisbon, Berlin, Como, Ljubljana and the Hague.
These conferences come for the first time to Spain and the venue is Valencia, situated in the East coast close
to the Mediterranean Sea, which represents a meeting point of many cultures. The host of the conference is the
Universidad Politcnica de Valencia.
This year the theme of the Conference is "Safety, Reliability and Risk Analysis. Theory, Methods and
Applications". The Conference covers a number of topics within safety, reliability and risk, and provides a
forum for presentation and discussion of scientific papers covering theory, methods and applications to a wide
range of sectors and problem areas. Special focus has been placed on strengthening the bonds between the safety,
reliability and risk analysis communities with an aim at learning from the past building the future.
The Conferences have been growing with time and this year the program of the Joint Conference includes 416
papers from prestigious authors coming from all over the world. Originally, about 890 abstracts were submitted.
After the review by the Technical Programme Committee of the full papers, 416 have been selected and included
in these Proceedings. The effort of authors and the peers guarantee the quality of the work. The initiative and
planning carried out by Technical Area Coordinators have resulted in a number of interesting sessions covering
a broad spectre of topics.
Sebastin Martorell
C. Guedes Soares
Julie Barnett
Editors

XXIX

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Organization

Conference Chairman
Dr. Sebastin Martorell Alsina

Universidad Politcnica de Valencia, Spain

Conference Co-Chairman
Dr. Bls Galvn Gonzlez

University of Las Palmas de Gran Canaria, Spain

Conference Technical Chairs

Prof. Carlos Guedes Soares
Dr. Julie Barnett

Technical University of LisbonIST, Portugal

University of Surrey, United Kingdom

Board of Institution Representatives

Prof. Gumersindo Verd
Vice-Rector for International Actions
Universidad Politcnica de Valencia, Spain
Dr. Ioanis Papazoglou
ESRA Chairman
Dr. Roberto Bubbico
SRA-Europe Chairman
Technical Area Coordinators
Aven, TerjeNorway
Bedford, TimUnited Kingdom
Berenguer, ChristopheFrance
Bubbico, RobertoItaly
Cepin, MarcoSlovenia
Christou, MichalisItaly
Colombo, SimoneItaly
Dien, YvesFrance
Domnech, EvaSpain
Eisinger, SiegfriedNorway
Enander, AnnSweden
Felici, MassimoUnited Kingdom
Finkelstein, MaximSouth Africa
Goossens, LouisThe Netherlands
Hessami, AliUnited Kingdom
Johnson, ChrisUnited Kingdom
Kirchsteiger, ChristianLuxembourg
Technical Programme Committee
Ale B, The Netherlands
Alemano A, Luxembourg
Amari S, United States
Andersen H, Denmark
Aneziris O, Greece
Antao P, Portugal
Arnaiz A, Spain

Leira, BertNorway
Levitin, GregoryIsrael
Merad, MyriamFrance
Palanque, PhilippeFrance
Papazoglou, IoannisGreece
Preyssl, ChristianThe Netherlands
Rackwitz, RuedigerGermany
Rosqvist, TonyFinland
Salvi, OlivierGermany
Skjong, RolfNorway
Spadoni, GigliolaItaly
Tarantola, StefanoItaly
Thalmann, AndreaGermany
Thunem, Atoosa P-JNorway
Van Gelder, PieterThe Netherlands
Vrouwenvelder, TonThe Netherlands
Wolfgang, KrgerSwitzerland

Badia G, Spain
Barros A, France
Bartlett L, United Kingdom
Basnyat S, France
Birkeland G, Norway
Bladh K, Sweden
Boehm G, Norway

XXXI

Bris R, Czech Republic

Brndl M, Switzerland
Burgherr P, Switzerland
Bye R, Norway
Carlos S, Spain
Castanier B, France
Castillo E, Spain
Cojazzi G, Italy
Contini S, Italy
Cozzani V, Italy
Cha J, Korea
Chozos N, United Kingdom
De Wit S, The Netherlands
Droguett E, Brazil
Drottz-Sjoberg B, Norway
Dutuit Y, France
Escriche I, Spain
Faber M, Switzerland
Fouladirad M, France
Garbatov Y, Portugal
Ginestar D, Spain
Grall A, France
Gucma L, Poland
Hardman G, United Kingdom
Harvey J, United Kingdom
Hokstad P, Norway
Holicky M, Czech Republic
Holloway M, United States
Iooss B, France
Iung B, France
Jonkman B, The Netherlands
Kafka P, Germany
Kahle W, Germany
Kleyner A, United States
Kolowrocki K, Poland
Konak A, United States
Korczak E, Poland
Kortner H, Norway
Kosmowski K, Poland
Kozine I, Denmark
Kulturel-Konak S, United States
Kurowicka D, The Netherlands
Labeau P, Belgium

Webpage Administration
Alexandre Janeiro

Le Bot P, France
Limbourg P, Germany
Lisnianski A, Israel
Lucas D, United Kingdom
Luxhoj J, United States
Ma T, United Kingdom
Makin A, Australia
Massaiu S, Norway
Mercier S, France
Navarre D, France
Navarro J, Spain
Nelson W, United States
Newby M, United Kingdom
Nikulin M, France
Nivolianitou Z, Greece
Prez-Ocn R, Spain
Pesme H, France
Piero B, Italy
Pierson J, France
Podofillini L, Italy
Proske D, Austria
Re A, Italy
Revie M, United Kingdom
Rocco C, Venezuela
Rouhiainen V, Finland
Roussignol M, France
Sadovsky Z, Slovakia
Salzano E, Italy
Sanchez A, Spain
Sanchez-Arcilla A, Spain
Scarf P, United Kingdom
Siegrist M, Switzerland
Srensen J, Denmark
Storer T, United Kingdom
Sudret B, France
Teixeira A, Portugal
Tian Z, Canada
Tint P, Estonia
Trbojevic V, United Kingdom
Valis D, Czech Republic
Vaurio J, Finland
Yeh W, Taiwan
Zaitseva E, Slovakia
Zio E, Italy

Instituto Superior Tcnico, Portugal

Local Organizing Committee

Sofa Carlos Alberola
Eva Ma Domnech Antich
Antonio Jos Fernandez
Bls Galvn Gonzlez
Aitor Goti Elordi
Sebastin Martorell Alsina
Rubn Mullor Ibaez

Universidad Politcnica de Valencia

Universidad Politcnica de Valencia
Iberinco, Chairman Reliability Committee AEC
Universidad de Las Palmas de Gran Canaria
Universidad de Mondragn
Universidad Politcnica de Valencia
Universidad de Alicante

XXXII

Rafael Prez Ocn

Ana Isabel Snchez Galdn
Vicente Serradell Garca
Gabriel Winter Althaus

Universidad de Granada
Universidad Politcnica de Valencia
Universidad Politcnica de Valencia
Universidad de Las Palmas de Gran Canaria

Conference Secretariat and Technical Support at Universidad Politcnica de Valencia

Gemma Cabrelles Lpez
Teresa Casquero Garca
Luisa Cerezuela Bravo
Fanny Collado Lpez
Mara Luca Ferreres Alba
Angeles Garzn Salas
Mara De Rus Fuentes Manzanero
Beatriz Gmez Martnez
Jos Luis Pitarch Catal
Ester Srougi Ramn
Isabel Martn Lluch
Alfredo Moreno Manteca
Maryory Villamizar Leon
Jos Felipe Villanueva Lpez
Sponsored by
Ajuntament de Valencia
Asociacin Espaola para la Calidad (Comit de Fiabilidad)
CEANI
Generalitat Valenciana
Iberdrola
Ministerio de Educacin y Ciencia
PMM Institute for Learning
Tekniker
Universidad de Las Palmas de Gran Canaria
Universidad Politcnica de Valencia

XXXIII

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Acknowledgements

The conference is organized jointly by Universidad Politcnica de Valencia, ESRA (European Safety and
Reliability Association) and SRA-Europe (Society for Risk AnalysisEurope), under the high patronage of
the Ministerio de Educacin y Ciencia, Generalitat Valenciana and Ajuntament de Valencia.
Thanks also to the support of our sponsors Iberdrola, PMM Institute for Learning, Tekniker, Asociacin
Espaola para la Calidad (Comit de Fiabilidad), CEANI and Universidad de Las Palmas de Gran Canaria. The
support of all is greatly appreciated.
The work and effort of the peers involved in the Technical Program Committee in helping the authors to
improve their papers are greatly appreciated. Special thanks go to the Technical Area Coordinators and organisers
of the Special Sessions of the Conference, for their initiative and planning which have resulted in a number of
interesting sessions. Thanks to authors as well as reviewers for their contributions in the review process. The
review process has been conducted electronically through the Conference web page. The support to the web
page was provided by the Instituto Superior Tcnico.
We would like to acknowledge specially the local organising committee and the conference secretariat and technical support at the Universidad Politcnica de Valencia for their careful planning of the practical arrangements.
Their many hours of work are greatly appreciated.
These conference proceedings have been partially financed by the Ministerio de Educacin y Ciencia
de Espaa (DPI2007-29009-E), the Generalitat Valenciana (AORG/2007/091 and AORG/2008/135) and the
Universidad Politcnica de Valencia (PAID-03-07-2499).

XXXV

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Introduction

The Conference covers a number of topics within safety, reliability and risk, and provides a forum for presentation
and discussion of scientific papers covering theory, methods and applications to a wide range of sectors and
problem areas.
Thematic Areas

Accident and Incident Investigation

Crisis and Emergency Management
Decision Support Systems and Software Tools for Safety and Reliability
Dynamic Reliability
Fault Identification and Diagnostics
Human Factors
Integrated Risk Management and Risk-Informed Decision-making
Legislative dimensions of risk management
Maintenance Modelling and Optimisation
Monte Carlo Methods in System Safety and Reliability
Occupational Safety
Organizational Learning
Reliability and Safety Data Collection and Analysis
Risk and Evidence Based Policy Making
Risk and Hazard Analysis
Risk Control in Complex Environments
Risk Perception and Communication
Safety Culture
Safety Management Systems
Software Reliability
Stakeholder and public involvement in risk governance
Structural Reliability and Design Codes
System Reliability Analysis
Uncertainty and Sensitivity Analysis

Industrial and Service Sectors

Aeronautics and Aerospace

Automotive Engineering
Biotechnology and Food Industry
Chemical Process Industry
Civil Engineering
Critical Infrastructures
Electrical and Electronic Engineering
Energy Production and Distribution
Health and Medicine
Information Technology and Telecommunications
Insurance and Finance
Manufacturing
Mechanical Engineering
Natural Hazards

XXXVII

Nuclear Engineering
Offshore Oil and Gas
Policy Decisions
Public Planning
Security and Protection
Surface Transportation (road and train)
Waterborne Transportation

XXXVIII

Reliability and safety data collection and analysis

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A new step-stress Accelerated Life Testing approach: Step-Down-Stress

Chunhua Zhang, Yashun Wang, Xun Chen & Yu Jiang
College of Mechatronics and Automation, National University of Defense Technology, Changsha, China

ABSTRACT: Step-stress ALT is a widely used method in life validation of products with high reliability. This
paper presents a new step-stress ALT approach with an opposite exerting sequence of stress levels in contrast to
traditional step-stress, so-called Step-Down-Stress (SDS). The testing efficiency of SDS ALT is compared with
step-stress ALT by Monte-Carlo simulation and contrastive experiment. This paper also presents a statistical
analysis procedure for SDS ALT under Weibull distribution. A practical ALT on bulb is given in the end to
illustrate the approach. SDS ALT may advance testing efficiency of traditional step-stress ALT remarkably
when applied in life validation of products with high reliability. It consumes less time for same failure number,
and gets more failures in same testing time than traditional step-stress ALT with identical testing plan. It also
helps to introduce the statistical analysis procedure accordingly which establishes a uniform analysis procedure
and can be applied to different acceleration equation easily.

constant-stress ALT, step-stress ALT and progressive

stress ALT as shown in Figure 1. Sometimes ALT
also takes a lengthy period of time until the censored
failure number or gains no failure until the censored
time, especially for life validation of products with
high reliability.
Step-stress ALT is applied more widely in engineering, because it is more efficient than constant
stress ALT, and easier to perform than progressive
stress ALT. In step-stress ALT, specimens are first

ACRONYMS
simplies statistical(ly)
ALT
accelerated life testing
AST
accelerated stress testing
SDS
step-down-stress
CEM
cumulative exposure model
CDF
cumulative distribution function
IPL
inverse power law
NOTATION
k
number of stress levels
Si
stress levels, i = 0, . . ., k
X
X = f (S)
Fi (t)
CDF of failures under Si
m
shape parameter of Weibull distribution

scale parameter of Weibull distribution

ti
time to failure under Si
ri
censored number under Si
n
number of specimens
Kij
acceleration factor between Si and Sj
xi
equivalent time to failure under Si
converted from other stress levels
ln(x)
natural logarithm of x

a. Constant-stress ALT

b. Step-stress ALT

INTRODUCTION

Accelerated life testing (ALT) is a commonly used

AST which is used to get information quickly on
life distribution of a product. Specimens are tested
under severe conditions in ALT and fail sooner than
under use conditions. According to the exerting way
of stress, ALT commonly falls into three categories:

c. Progressive stress ALT d. Step-down-stress ALT

Figure 1.

863

Methods of ALT.

subjected to a specified constant stress for a specified length of time; after that, they are subjected to a
higher stress level for another specified time; the stress
on specimens is thus increased step by step[18] .
This paper presents a new step-stress ALT approach
with an opposite exerting sequence of stress levels in
contrast to traditional step-stress, so-called step-downstress (SDS), from the assumption that the change in
exerting sequence of stress levels will improve testing efficiency. The validity of SDS ALT is discussed
through comparison with tradition step-stress ALT by
Monte-Carlo simulation and contrastive experiment,
and it concludes that SDS ALT takes less time for same
failure number and gets more failures in same time.
A new s-analysis procedure is constructed for SDS
ALT, which is applicable to different acceleration
equation and may be programmed easily.
The rest of this paper is organized as follows:
section 2 describes SDS ALT including basic assumptions, the definition, s-analysis model, and MonteCarlo Simulation; section 3 presents an s-analysis
procedure for SDS ALT; section 4 gives a practical
example; section 5 concludes the paper.
2

Figure 2.

CEM for traditional step-stress ALT.

Figure 3.

CEM for SDS ALT.

STEP-DOWN-STRESS ALT MODEL

2.1 Basic assumptions

a. Time to failure at each stress level follows Weibull
distribution
F(t) = 1 exp[(t/)m ],

t>0

(1)

The wear-out failures occur at an increasing failure

rate with m > 1.
b. i and Si meet the acceleration equation
ln() =

n


aj Xj

(2)

j=0

which is the Arrhenius model when X = 1/S and

the IPL model when X = ln(S).
c. Failure mechanism is identical at different stress
levels.
2.2

Definition

SDS ALT is shown in Figure 1d. In SDS ALT, specimens are initially subjected to the highest stress level
Sk for rk failures, then stress is stepped down to Sk1
for rk1 failures, and then stress is stepped down
again to Sk2 for rk2 failures, and so on. The test is
terminated at the lowest stress S1 until r1 failures occur.
From the description above, SDS ALT is symmetrical to traditional step-stress ALT with different varying

direction of stress levels. The following text will prove

that the change in exerting sequence of stress levels
will possibly improve the test efficiency remarkably.

864

Table 1.

Monte-Carlo simulations for step-stress and SDS ALT.

Test plan (n, r)

No.

(m, )

Step-down stress

Step-stress

1
2
3
4
5
6
7
8

Model 1: 2.4, [188 115 81 63]

50, [10 10 10 10]

50, [25 5 5 5]
60, [10 10 10 10]
50, [10 10 10 10]
40, [10 10 10 10]
40, [20 5 5 5]
40, [10 10 10 10]
40, [20 5 5 5]

50, [10 10 10 10]

50, [5 5 5 25]
60, [10 10 10 10]
50, [10 10 10 10]
40, [10 10 10 10]
40, [5 5 5 20]
40, [10 10 10 10]
40, [5 5 5 20]

1.318
1.338
1.399
3.916
2.276
4.699
1.696
2.902

2.3

Model 2: 4.5, [58678 16748 5223 1759]

Model 3: 3.8, [311 108 47 26]

= [1 , 2 , 3 , 4 ]; r = [r4 , r3 , r2 , r1 ] for SDS ALT, r = [r1 , r2 , r3 , r4 ] for step-stress ALT.

s-Analysis model
[2]

According to CEM , the remaining life of units under

a step-stress pattern depends only on the current cumulative fraction failed and current stress regardless how
the fraction accumulated.
As shown in Figure 2, if stress levels in step-stress
ALT are S1 , S2 , . . . , Sk , where the duration at Si is
ti ti1 and the corresponding CDF is Fi (t), the population CDF F0 (t) of specimens in step-stress ALT
equals to F1 (t) at first and then steps to F2 (t) at t1 .
F0 (t) steps up successively by that way until Sk .
As shown in Figure 3, if apply CEM to SDS ALT,
the population CDF F0 (t) of specimens in SDS ALT
equals to Fk (t) at first and then steps to Fk1 (t) at tk .
F0 (t) steps successively by that way until S1 .
2.4 Monte-carlo simulation
In order to discuss the efficiency of SDS ALT further, contrastive analysis is performed through MonteCarlo simulation shown in Table 1.
Let the use stress be S0 , the k accelerated stresses
be S1 , S2 , . . . , Sk , and the scale parameter corresponding to Si be 0 , 1 , . . . , k according to acceleration
equation (2). In Monte-Carlo Simulation, the sampling size of units is n, and k censored failure numbers
corresponding to Si are r1 , r2 , . . . , rk .
Monte-Carlo Simulations, see Table 1, are performed respectively based on three commonly-used
acceleration equations, which come from practical
products. The result of Monte-Carlo simulation is
expressed as efficiency index e, which is the average
ratio of the total testing time in traditional step-stress
ALT to that in SDS ALT. If e > 1, the efficiency of
SDS ALT is higher than traditional step-stress ALT.
From Table 1, the following rules can be drawn:
a. The efficiency of SDS ALT in these occasions are
commonly higher than traditional step-stress ALT.
The highest e = 4.699 in No. 6 simulation, which

means that the total testing time for SDS ALT is

only 21.28% of traditional step-stress ALT.
b. The longer the life of specimen is, the greater
advantage of SDS ALT has, for example, e of
Model 2 is commonly higher than other two models, which means SDS ALT can be adopted better
in long life validation.

STATISTICAL ANALYSIS

3.1 Description of the problem

If stress levels in SDS ALT are Sk , Sk1 , . . .S1 (Sk >
Sk1 > > S1 ), the specimen size is n and the
censored failure number is ri for Si . The stress steps
down to Si1 until ri failures occur at Si . The time
to failure in such an SDS ALT can be described as
follows:
Sk : tk,1 , tk,2 , . . . , tk,rk
Sk1 : tk1,1 , tk1,2 , . . . , tk1,rk1
......
S1 : t1,1 , t1,2 , . . . , t1,r1

(3)

where ti,j means the No. j time to failure under Si timed

from the beginning of Si .
3.2

Estimation of distribution parameters

Failures in SDS ALT are cumulative effect of several

accelerated stresses except for Sk , so it is a key problem in s-analysis for SDS ALT that how to convert
testing time between stress levels to obtain population
life information. Data in s-analysis for traditional stepstress ALT are converted mainly through acceleration
equations [14] , and the solution to this problem will
becomes very complex accordingly and sometimes
diverges.

865

If CDF of specimen is Fi (ti ) at ti under Si , according

to CEM, tj can be found at Sj to satisfy
Fi (ti ) = Fj (tj )

then the inverse moment estimates of Weibull parameters under Si is from

n
i 1

(4)

ln[uni (m
i , i )/uj (m
i , i )] = ni 1

(12)

j=1

which means that cumulative degradation of life at

Si equals to that at Sj in given time respectively. So
acceleration factor can be defined as Kij = tj /ti . For
Weibull distribution,

1/m i
ni

m

xj i ( i ) + (n ni )xnmii ( i )/ni
i =
j=1

(13)
mi = mj
Kij = j /i

(5)

So mi & i can be obtained by solving equation (12)

& (13). See APPENDIX B for a numerical algorithm.

See APPENDIX A for the proof of (5). Because

Ki,i+1 = ti+1 /ti = i+1 /i , the time to failure ti+1 at
Si+1 can be converted into equivalent data under Si for
Weibull distribution

3.3 Estimation of acceleration equation

xi = i (ti+1 /i+1 )

Sk : tk,1 , tk,2 , . . . , tk,rk

Sk1 : xk1,1 , xk1,2 , . . . , xk1,rk +rk1
......
S1 : x1,1 , x1,2 , . . . , x1,rk +rk1 ++r1

By the procedure above, the s-analysis for SDS ALT

is transferred to s-analysis for an equivalent constantstress ALT, which can be described as follows:

(6)

If all the equivalent data at Si+1 are

xi+1,1 , xi+1,2 , . . . , xi+1,ni +1

(7)

where ni+1 = rk + rk1 + + ri+1 , the former ni+1

equivalent time-to-failures at Si are
xi,j = i (xi+1,j /i+1 )

(j = 1, 2, . . . , ni+1 )

(j = 1, 2, . . . , ri )

(9)

So ni (ni = rk + rk1 + + ri ) equivalent time-tofailures at Si are obtained, which form quasi samples
xi (i ) under Weibull distribution
x1 (i ) < x2 (i ) < xni (i )

To estimate parameters of acceleration equation, an

estimation method for constant-stress ALT under
Weibull distribution can be applied[9] .

EXPERIMENT RESULT

(8)

The succeeding ri equivalent time-to-failures at Si

can be obtained by adding cumulative testing time to
ti,j at Si
xi,ni+1 +j = xi,ni+1 + ti,j

To illustrate the efficiency of the proposed SDS ALT

approach, an SDS ALT and a traditional step-stress
ALT on bulb were performed with identical testing
plan but inverse sequence of stress levels shown in
Table 2. The accelerated stress is the voltage, given
by S1 = 250 V, S2 = 270 V, S3 = 287 V, and S4 =
300 V. n = 40 and (r1 , r2 , r3 , r4 ) = (5, 5, 5, 20). It
shows that total testing time of SDS ALT is only 30.7%
of traditional step-stress ALT.
Data from this SDS ALT are s-analyzed with proposed procedure and the distribution parameters are

(10)

Table 2. Contrastive experiment of SDS ALT and traditional step-stress ALT.

Approach Si (V)

(n, r)

Testing time
at each stress
level (h)

Step up

40,
[5 5 20]
40, [20
5 5 5]

192.118 10.183 213.519

3.690 7.528
13.342 7.092
65.502
9.325 35.743

which contains unknown variable i . Let



(14)

Total
testing
time (h)

uj (mi , i ) =

xkmi (i ) + (n j)xjmi (i )

k=1

( j = 1, 2, . . . , ni )

(11)

866

Step
down

250 270
287 300
300 287
270 250

Table 3.

which is identical equation for arbitrary ti . So

Distribution parameters under Si .

mi = mj

Si (V)

300

287

270

250

m
(h)

3.827
14.690

3.836
83.294

3.827
97.248

3.821
285.160

Kij = j /i
B Numerical solution algorithm of (12) & (13).
To simplify the procedure, (12) & (13) is transformed as

shown in Table 3. The relation between and S

satisfies IPL model
ln = 81.43918 13.71378 ln S

1 + (ni 1) ln uni (mi , i )

CONCLUSION

An SDS ALT and its s-analysis procedure are presented in this paper for life validation by AST. The
validity of this new step-stress approach is discussed through Monte-Carlo simulation and contrastive experiment. It shows that SDS ALT may
advance testing efficiency remarkably and accordingly
decrease the cost of test. The s-analysis procedure for
SDS ALT under Weibull distribution is constructed,
and the validity is also demonstrated through ALT
on bulb.
Future efforts in the research of SDS ALT may
include further discussion on efficiency, improvement
of s-analysis procedure, and optimal design of test plan
for SDS ALT.

ln ui (mi , i ) = ni

i=1

(15)

Parameters of acceleration equation (15) approach

nearly to those from constant-stress ALT, which
demonstrates the validity of SDS ALT and its
s-analysis procedure.

n
i 1

(19)
i = [uni (mi , i )/ni ]

1/mi

(20)

Equation (20) can be written as

i (l + 1) = [uni (mi (l), i (l))/ni ]1/mi (l)

(21)

by which the convergent solution of i corresponding

to mi (l) can be computed iteratively.
The solution of (19) can be obtained with a tentative
method. Let
ni (mi , i ) = 1 + (ni 1) ln uni (mi , i )

n
i 1

ln ui (mi , i )

(22)

i=1

by which ni (mi (l), i (l + 1)) can be computed for a

given mi (l) and i (l + 1).

APPENDIX
A Proof of (5)
The CDF under Weibull distribution is
F(t) = 1 exp[(t/)m ],

t>0

(16)

where m, (m, > 0) are shape parameter and scale

parameter respectively. From equation (4),
1 exp[(ti /i )mi ] = 1 exp[(tj /j )mj ]

(17)

Because exp() is strictly monotone and tj = Kij ti ,

(ti /i )mi = (Kij ti /j )mj

(18)

Figure A1.

867

Numerical solution algorithm of (12) & (13).

If ni (mi (l), i (l + 1)) = ni ,

mi (l + 1) = mi (l)

(23)

If ni (mi (l), i (l + 1)) > ni , mi (l + 1) < mi (l) and

try again by
mi (l + 1) = mi (l) m

(24)

If ni (mi (l), i (l + 1)) < ni , mi (l + 1) > mi (l) and

try again by
mi (l + 1) = mi (l) + m

(25)

Figure A1 shows the algorithm above.

REFERENCES
[1] Nelson, W. 1990. Accelerated Testing: Statistical Models, Test Plans, and Data Analyses. New York: John
Willey & Sons: 1822.

868

[2] Nelson, W. 1980. Accelerated Life TestingStepstress Models and Data Analysis. IEEE Trans. on
Reliability R-29: 103108.
[3] Tyoskin, O. & Krivolapov, S. 1996. Nonparametric
Model for Step-Stress Accelerated Life Testing. IEEE
Trans. on Reliability 45(2): 346350.
[4] Tang, L. & Sun, Y. 1996. Analysis of Step-Stress
Accelerated-Life-Test Data: A New Approach. IEEE
Trans. on Reliability 45(1): 6974.
[5] Miller, R. & Nelson, W. 1983. Optimum Simple Stepstress Plans for Accelerated Life Testing. IEEE Trans.
on Reliability. 32(1): 5965.
[6] Bai, D. & Kim, M. & Lee, S. 1989 Optimum Simple
Step-stress Accelerated Life Tests with censoring. IEEE
Trans. on Reliability. 38(5): 528532.
[7] Khamis, I. & Higgins, J. 1996. Optimum 3-Step
Step-Stress Tests. IEEE Trans. on Reliability. 45(2):
341345.
[8] Yeo, K. & Tang, L. 1999. Planning Step-Stress LifeTest with a Target Acceleration-Factor. IEEE Trans. on
Reliability. 48(1): 6167.
[9] Zhang, C. & Chen, X. 2002. Analysis for Constantstress Accelerated Life Testing Data under Weibull Life
Distribution. Journal of National University of Defense
Technology (Chinese). 24(2): 8184.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Application of a generalized lognormal distribution to engineering

data fitting
J. Martn & C.J. Prez
Departamento de Matemticas, Universidad de Extremadura, Cceres, Spain

ABSTRACT: The lognormal distribution is commonly used to model certain types of data that arise in several
fields of engineering as, for example, different types of lifetime data or coefficients of wear and friction.
However, a generalized form of the lognormal distribution can be used to provide better fits for many types of
experimental or observational data. In this paper, a Bayesian analysis of a generalized form of the lognormal
distribution is developed. Bayesian inference offers the possibility of taking expert opinions into account. This
makes this approach appealing in practical problems concerning many fields of knowledge, including reliability
of technical systems. The full Bayesian analysis includes a Gibbs sampling algorithm to obtain the samples from
the posterior distribution of the parameters of interest. Empirical proofs over a wide range of engineering data
sets have shown that the generalized lognormal distribution can outperform the lognormal one in this Bayesian
context.
Keywords: Bayesian analysis, Generalized normal distribution, Engineering data, Lognormal distribution,
Markov chain Monte Carlo methods.
1

INTRODUCTION

The lognormal distribution is commonly used to model

certain types of data that arise in several fields of engineering as, for example, different types of lifetime data
(see, e.g., Meeker and Escobar (1998)) or coefficients
of wear and friction (see, e.g., Steele (2008)). Particular properties of the lognormal random variable (as the
non-negativeness and the skewness) and of the lognormal hazard function (which increases initially and then
decreases) make lognormal distribution a suitable fit
for some engineering data sets. However, a generalized
lognormal distribution can be used to provide better
fits for many types of experimental or observational
data.
If a random variable X has a lognormal distribution, the random variable Y = log X is normally
distributed. This allows the more well-known analysis
techniques for the normal distribution to be applied to
the lognormal data through transformation. The previous relationship suggests a possible generalization of
the lognormal distribution by using a similar transformation for the generalized normal distribution. The
generalized normal distribution (GN for short) is a
generalization of the normal distribution that also has
the t and Laplace distributions as particular cases (see
Nadarajah (2005)). In fact, it is a re-parametrization
of the exponential power (EP) distribution. The first

formulation of the EP distribution could be attributed

to Subbotin (1923). Since then, several different
parameterizations can be found in the literature (see,
for example, Box and Tiao (1973), Gmez et al.,
(1998), and Mineo and Ruggieri (2005)). This family provides distributions with both heavier and lighter
tails compared to the normal ones. The GN distributions allow the modeling of kurtosis providing, in
general, a more flexible fitting to experimental data
than the normal distributions. The main reason why
the GN or EP distributions have not been used as
often as desirable has been purely computational, i.e.,
because most standard statistical software did not contain procedures using GN distributions. Currently, the
GN distribution is considered as a factible alternative
to the normal one as a general distribution for random
errors.
In this paper, a generalized lognormal distribution
(logGN for short) is analyzed from a Bayesian viewpoint. If a random variable X has a logGN distribution,
the random variable Y = log X is distributed as a
GN. The logGN distribution has the lognormal one as
a particular case. Bayesian inference offers the possibility of taking expert opinions into account. This
makes this approach appealing in practical problems
concerning many fields of knowledge, including reliability of technical systems. The Bayesian approach is
also interesting when no prior information is obtained,

869

in this case a noninformative prior distribution is used.

The full Bayesian analysis includes a Gibbs sampling
algorithm to obtain the samples from the posterior
distribution of the parameters of interest. Then, the
predictive distribution can be easily obtained. Empirical proofs over a wide range of engineering data sets
have shown that the generalized lognormal distribution can outperform the lognormal one in this Bayesian
context.
The outline of this work is as follows. In Section 2,
the logGN distribution is described. Section 3 presents
the Bayesian analysis with both noninformative and
informative prior distributions. An example with friction data illustrates the application of the proposed
approach in Section 4. Finally, Section 5 presents the
main conclusions.
2

THE GENERALIZED LOGNORMAL MODEL

If a random variable X has a logGN distribution, the

random variable Y = log X is distributed as a GN.
Therefore, the probability density function of a logGN
distribution with parameters , , and s is given by:

 
 log x s
s
 ,

f (x) =
exp




2 x ( 1s )
with x > 0, < < +, > 0 and s 1. Note
that  denotes the gamma function.
This distribution has the lognormal distribution as
a particular
case by taking s = 2 and changing

to 2 . The log-Laplace distribution is recovered

when s = 1. Figure ?? shows the probability density
functions for some values of s with = 0 and = 1.

Proposition 1 Let X and U be two random variables

such that f (x|u) = 2x1u1/s I [exp ( u1/s ) < x <
exp ( + u1/s )] and f (u) = Gamma (shape = 1 +
1/s, scale = 1), then X log GN(, , s).
This result can also be used to generate random
variates from a logGN(, , s). Generating from U
is standard, and generating from X |U is obvious
through the inverse transformation method. Then, the
algorithm to generate random values is given by the
following steps:
1. Generate W Gamma (1 + 1/s, 1)
2. Generate V Uniform (1, 1)
3. Set X = exp { W 1/s V + }
The next section presents a Bayesian analysis for
the logGN distribution.

BAYESIAN ANALYSIS

0.4

Bayesian analyses with both noninformative and informative prior distributions are addressed in this section.
0.2

Density

0.6

0.8

s=1.0
s=1.5
s=2.0
s=3.0

The capacity of a distribution to provide an accurate

fit to data depends on its shape. The shape can be
defined by the third and fourth moments and they
represent the asymmetry and flatness coefficients of
a given distribution. The logGN distributions allow
the modeling of kurtosis providing, in general, a more
flexible fit to experimental data than the lognormal
distribution.
Random variates from the logGN distribution can
be generated from random variates of the GN distribution via exponentiation. Since the GN distribution is
a reparameterization of the EP distribution, the techniques for random generation of these distributions
can be used for the GN distribution (see, for example,
Devroye (1986), Jhonson (1987) and Barabesi (1993)).
Walker and Gutirrez-Pea (1999) suggested a mixture representation for the EP distribution that is
adapted here to be valid for the logGN distribution.
The following result will be used in the next section
to determine the full conditional distributions necessary to apply the Gibbs sampling method. The proof
is immediate.

0.0

3.1 Noninformative case

Figure 1. Probability density functions for logGN distributions with = 0 and = 1, and several values of s.

Following the suggestions in Box and Tiao (1973)

and Portela and Gmez-Villegas (2004) for the EP
distribution, independence between parameters is considered.
Jeffreys noninformative prior distributions are
considered here. Jeffreys
choice for the noninformative density is () I (), where I () is the Fisher

870

information for (see, e.g., Box and Tiao (1973) or

Gelman et al. (2004)). This prior distribution is noninformative in the sense that it maximizes the entropy. In
order to obtain the expressions for the noninformative
prior distributions of the parameters, the calculation
of the Fisher information matrix is required. The
Fisher information matrix for the logGN distribution
is given by:

2 (1/s)

0
0

L(, , s, u|x)
=

I (, , s)
(s1)s(11/s)

Proposition 1 is used here to obtain the likelihood.

Then, the likelihood of a sample x = (x1 , x2 , . . . ,
xn ), given the vector of mixing parameters u =
(u1 , u2 , . . . , un ), is:

A
s11/s
,
2
(1+1/s)  (1+1/s)+A2 1
A
s11/s
s3

Therefore, the posterior distribution is given by:

s1+2/s

where is the digamma function and A = log(s) +

(1 + 1/s).
Noninformative prior distributions are derived for
the parameters, i.e.,:

n
1/s
1/s
1  1
I [e ui < xi < e+ ui ].
(2 )n i=1 xi ui1/s

f (, , s, u|x)

n

sn1
eui
I
n+1  n ( 1s ) i=1 xi

1/s

1/s

[e ui < xi < e+ ui ].

() 1,
The full conditional distributions are derived:

1
,


(1 + 1/s)  (1 + 1/s) + A2 1
(s)
,
s3
( )

f (| , s, u, x) 1,
1/s

1/s

max{log(xi ) ui } < < min{log(xi ) + ui }

with < < +, > 0 and s 1.

Since the expression for (s) is very involved, a
simple and similar distribution is used, i.e., (s)
1/s, see Figure ??.
A Markov Chain Monte Carlo (MCMC) method
is applied to generate samples from the posterior
distribution. Specifically, a Gibbs sampling algorithm is derived. The mixture representation given in

f ( |, s, u, x)

1
, > max
i
n+1

| log(xi )|
1/s

ui

(2)
f (s|, , u, x)

n1

s
,
 n (1/s)

max{1, ai } < s < min ai

iS

iS +

f (ui |, , s, x) eui ,


| log(xi ) | s
, i = 1, 2, . . . , n,
ui >

(3)

(4)

0 .4

0 .5

0 .6

I s
2 3s

(1)


0 .2

0 .3

where S = {i : log(| log(xi )|/ ) < 0}, S + = {i :

log(| log(xi )|/ ) > 0} and

0 .1

ai =

Figure 2.

Comparison of functions

I (s) and 2/(3s).

10

log(ui )
, i = 1, 2, . . . , n.
log(| log(xi )|/ )

Random variates from these densities can be generated by using standard methods. Note that the densities
given in (??), (2) and (??) are uniform, Pareto and
exponential, respectively, and they are generated by
using the inverse transformation method. The density

871

given in (??) is non-standard, but it can also be easily generated by using the rejection method (see, e.g.,
Devroye (1986)).
Iterative generations from the above conditional
distributions produce a posterior sample of (, , s).

The full conditional distributions are:

f (| , s, u, x) p(),
1/s

1/s

max{log(xi ) ui } < < min{log(xi ) + ui }

i

(8)
(a0 +n+1) b0 /

3.2 Informative case

In many situations, the data analyst is interested in
including relevant initial information in the inference process. The choice of the prior distribution
must be carefully determined to allow the inclusion
of this information. Since the prior distribution choice
depends on the problem in hand, there are multiple references related to this topic in the literature (see, e.g.,
DeGroot (1970), Berger (1985), Ibrahim et al. (2001)),
and Akman and Huwang (2001). Kadane and Wolfson
(1998) present an interesting review on elicitation of
expert opinion. OHagan (1998) considers the elicitation of engineers prior beliefs. Gutirrez-Pulido et al.
(2005) present a comprehensive methodology to specify prior distributions for commonly used models in
reliability.
The following prior distributions have been proposed because they can accommodate many possible
shapes for the kind of parameters involved in the
logGN distribution. Besides, they allow to make efficient posterior calculations and recover the noninformative distribution for each parameter. The proposed
prior distributions are given by:
() p(),

< < +,

( ) (a0 +1) eb0 / ,

(s) s

(c0 +1) d0 /s

(5)

>0

(6)

s 1,

n

eui
i=1

xi

p() snc0 1 ed0 /s

a0 +n+1 eb0 /  n (1/s)
1/s

f (s|, , u, x)

snc0 1 ed0 /s
,
 n (1/s)

max{1, ai } < s < min ai

iS

(9)

(10)

iS +

f (ui |, , s, x) eui ,


| log(xi ) | s
ui >
,

i = 1, 2, . . . , n.

(11)

where S , S + and ai , i = 1, 2, . . . , n, are defined as

in the previous subsection.
In this case, generating from these truncated densities is also easy to perform. Generating from (??)
depends on the chosen density. Note that the density given in (??) is a left truncated inverse-gamma.
Random variates from this distribution are obtained
by taking the reciprocal of variates from a truncated
gamma distribution. The density for the conditional
distribution given in (??) is again non-standard. Similarly to the noninformative case, a rejection method
is implemented. Finally, note that (??) is the same
as (??).

(7)

where (??) is any distribution in its support, (??) is

an inverse-gamma distribution and (??) is a truncated
inverse gamma distribution. Note that the noninformative prior distributions are recovered when p() is
constant and a0 = b0 = c0 = d0 = 0. This fact is
specially interesting because it allows to use a noninformative prior distribution for one or two parameters
and informative prior distributions for the remaining
parameters.
Analogously to the noninformative case, the posterior distribution is derived:
f (, , s, u|x)

f ( |, s, u, x)
e


| log(xi )|
> max
1/s
i
ui

1/s

I [e ui < xi < e+ ui ].

APPLICATION TO FRICTION DATA

In basic tribology, dimensional issues and the central

limit theorem are used to argue that the distribution
of the coefficients of friction and wear are typically
lognormal. Moreover, empirical evidences from many
data sources support this argument (see, e.g., Wallbridge and Dowson (1987) and Steele (2008)). Steele
(2008) recommends that engineers, without evidence
to suggest differently, allocate a lognormal distribution
to the coefficients of friction and wear.
The use of the proposed approach is illustrated with
a data set (presented in Nica (1969)) containing the
coefficients of friction of clean steel in a vacuum.
The size of this data set is 23. These data are used
as historical data to extract knowledge on the predictive distribution of the friction coefficient. The most
usual case in engineering studies is to have some prior
information about the process on which they are trying

872

150
0 50

Posterior density

60
40
20
0

Posterior density

Posterior density

0.06

0.07

0.08

0.09

0.10

8
6
Density
4
2
0.7

0.8

3.0

3.5
s

Figure 3.

95% HDR for , , and s.

4.0

1.1

1.2

hyperparameters obtained are 0 = 0.0460, 0 =

0.0017, a0 = 80.8737, b0 = 4.0744, c0 = 60.2888,
and d0 = 220.4130.
After it has been considered that the chain convergence has been achieved, a sample of size 10,000 for
the parameters of the posterior distribution is generated. The 95% Highest Density Regions (HDR) for ,
, and s are (0.0490, 0.0422), (0.0595, 0.0855),
and (2.2468, 3.3496), respectively (see Figure ??).
Note that the HDR for s does not contain the value
s = 2, that recovers lognormality.
In order to make a performance comparison, a
similar procedure is implemented to obtain the hyperparameters in the lognormal case. Then, a posterior sample is generated by using the lognormal
distribution instead of the logGN distribution. Here
the 95% HDR for and are (0.0494, 0.0427)
and (0.0923, 0.1416). The posterior predictive distributions are presented in Figure ??.
The comparison between both performances is presented based on the generated posterior predictive
distributions. The criterion used to compare them is
based on the use of the logarithmic score as a utility function in a statistical decision framework. This
was proposed by Bernardo (1979) and used, for example, by Walker and Gutirrez-Pea (1999) in a similar
context. The expected utilities for the lognormal and
logGN models can be estimated as:

2.5

1.0

Figure 4. Posterior predictive distributions for friction data.

Solid line: logGN, and dashed line: lognormal.

1
U 0 =
log(p0 (xi ))
n i=1
2.0

0.9
Data

0.0 0.4 0.8 1.2

0.05

to make inferences. This corresponds to the informative case. Then, the historical information and the prior
information provided by the engineer are embedded in
the posterior distribution.
The following step is to choose the hyperparameter values for the prior distributions of the parameters
of interest , , and s. There are many ways to elicit
these values. One possibility is to specify the values
according to previous direct knowledge on the parameters (see, e.g., Berger (1985) and Akman and Huwang
(2001)). Another one consists in using partial information elicited by the expert. In this case, there are
many criteria to obtain the hyperparameters values as,
for example, maximum entropy and maximum posterior risk (see Savchuk and Martz (1994)). A third
possibility considered here is to use expert information on the expected data and not on the parameters.
This is easier for engineers who are not familiarized
with parameters but have an approximate knowledge
of the process. Finally, it is remarkable that noninformative prior distributions can be used for any of the
parameters and informative prior distributions for the
remaining ones.
In this application, the hyperparameters are obtained by using a method similar to the one presented by Gutirrez-Pulido et al. (2005). In this
case the expert is asked to provide occurrence
intervals for some usual quantities as the mode,
median and third quartile. The expert considered
that these quantities should be in the following intervals: [LMo , UMo ] = [0.935, 0.955], [LMe , UMe ] =
[0.95, 0.96], and [LQ3 , UQ3 ] = [0.97, 0.985]. By using
the informative prior distributions presented in subsection 3.2, with N (0 , 0 ), and following the
development in Gutirrez-Pulido et al. (2005), the

1
U 1 =
log(p1 (xi )),
n i=1
n

873

where p0 and p1 are the posterior predictive

distributions of the lognormal and logGN models,
respectively. The estimated values are U 0 = 1.4118
and U 1 = 1.4373, so the logGN model performs
better than the lognormal one. The same happens
for the noninformative model (U 0 = 1.5533 and
U 1 = 1.5881).
5

CONCLUSION

The generalized form of the lognormal distribution,

presented and analyzed from a Bayesian viewpoint,
offers the possibility of taking expert opinions into
account. The proposed approach represents a viable
alternative to analyze data that are supposed to follow
a lognormal distribution and provides flexible fits to
many types of experimental or observational data. The
technical development is based on a mixture representation that allows to perform inferences via Gibbs
sampling. It is remarkable that the logGN family provides very flexible distributions that can empirically
fit many types of experimental or observational data
obtained from engineering studies.
ACKNOWLEDGEMENTS
This research has been partially supported by Ministerio de Educacin y Ciencia, Spain (Project TSI200766706-C04-02).
REFERENCES
Akman, O. and L. Huwang (2001). Bayes computation
for reliability estimation. IEEE Transaction on Reliability 46(1), 5255.
Barabesi, L. (1993). Optimized ratio-of-uniform method for
generating exponential power variates. Statistica Applicata 5(2), 149155.
Berger, J.O. (1985). Statistical Decision Theory and Bayesian
Analysis. Springer.
Bernardo, J.M. (1979). Reference posterior distributions
for Bayesian inference. Journal of the Royal Statistical
Society B 41, 113147.
Box, G. and G. Tiao (1973). Bayesian Inference in Statistical
Analysis. Addison-Wesley, Reading.

DeGroot, M.H. (1970). Optimal Statistical Decisions.

McGraw-Hill, New York.
Devroye, L. (1986). Non-Uniform Random Variate Generation. Springer-Verlag.
Gelman, A., J.B. Carlin, H.S. Stern, and D.B. Rubin (2004).
Bayesian Data Analysis. Chapman & Hall-CRC.
Gmez, E., M.A. Gmez-Villegas, and J.M. Marn (1998).
A multivariate generalization of the power exponential
family of distributions. Communications in StatisticsTheory and Methods 27(3), 589600.
Gutirrez-Pulido, H., V. Aguirre-Torres, and J.A. Christen
(2005). A practical method for obtaining prior distributions in reliability. IEEE Transaction on Reliability 54(2),
262269.
Ibrahim, J.G., M.H. Chen, and D. Sinha (2001). Bayesian
Survival Analysis. Springer-Verlag.
Johnson, M.E. (1987). Multivariate Statistical Simulation.
John Wiley and Sons.
Kadane, J.B. and L.J. Wolfson (1998). Experiences in
elicitation. The Statistician 47, 319.
Meeker, W.Q. and L.A. Escobar (1998). Statistical Methods
for Reliability Data. John Wiley and Sons.
Mineo, A.M. and M. Ruggieri (2005). A software tool for
the exponential power distribution: The normal package.
Journal of Statistical Software 12(4), 124.
Nadarajah, S. (2005). A generalized normal distribution.
Journal of Applied Statistics 32(7), 685694.
Nica, A. (1969). Theory and Practice of lubrication systems.
Scientific Publications.
OHagan, A. (1998). Eliciting expert beliefs in substantial
practical applications. The Statistician 47, 2135.
Portela, J. and M.A. Gmez-Villegas (2004). Implementation of a robust bayesian method. Journal of Statistical
Computation and Simulation 74(4), 235248.
Savchuk, V.P. and H.F. Martz (1994). Bayes reliability estimation using multiple sources fo prior information: Binomial sampling. IEEE Transactions on Reliability 43(I),
138144.
Steele, C. (2008). The use of the lognormal distribution
for the coefficients of friction and wear. Reliability
Engineering and System Safety To appear.
Subbotin, M. (1923). On the law of frecuency errors.
Mathematicheskii Sbornik 31, 296301.
Walker, S.G. and E. Gutirrez-Pea (1999). Robustifying bayesian procedures. In J.M. Bernardo, J.O. Berger,
A.P. Dawid, and A.F.M. Smith (Eds.), Bayesian Statistics
6, pp. 685710. Oxford University Press.
Wallbridge, N.C. and D. Dowson (1987). Distribution of wear
rate data and a statistical approach to sliding wear theory.
Wear 119, 295312.

874

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Collection and analysis of reliability data over the whole product

lifetime of vehicles
T. Leopold
TTI GmbH an der Universitt Stuttgart, Germany

B. Bertsche
Institute of Machine Components, Universitt Stuttgart, Germany

ABSTRACT: One of the biggest challenges in quantitative reliability analyses is a complete and significant
data basis, which describes the complete product lifetime. Today, this data basis, in regards to the demanded
and needed information, is not given. Especially the circumstances, that lead to a failure while the field usage,
e.g. operational and ambient information before and while failure occurrence, are of highest importance. In the
development phase of products much more detailed data is collected and documented, compared to the amount
in general customer field usage. Today, one of the most important data basis to describe the real field behavior
are warrantee and goodwill data. For an optimal correlation between failures that occur while testing and while
field usage, these data are not sufficient. In order to improve this situation, an approach was developed, with
which the collection of reliability relevant data during costumer usage is enabled over the whole product lifetime
of vehicles. The basis of this reliability orientated data collection is the consideration of already available CAN
signals in modern vehicles.

INTRODUCTION

In general, field data of systems (e.g., vehicles) and

its components, which are collected or generated over
the complete life time of the product, is a very important data basis for companies. This data is an essential
precondition for analyzing reliability and availability
of the systems, with which reflections or forecasts of
the real behavior in field use can be done.
An outstanding characteristic of quantitative reliability data is the mapping of failures and its ambient
and operational circumstances that lead to a failure.
The possibility to compare the operational and ambient conditions of intact and failed vehicles is another
important characteristic.
The conclusion of these two characteristics is that
such a data base has to be built up systematically
in forward-looking companies and thus represents an
extension of the present common way of collecting
failure data. Todays warrantee and goodwill data do
not offer that detailed failure information over the
whole product lifetime of vehicles, not to mention the
missing information about intact products while field
usage. To generate dependable failure data while testing, a permanent feedback of field data is necessary for
a continuous improvement of testing conditions and
its assumptions, e.g., correlation between test results

and field behavior. Thus, an observation of the field

behavior can not be substituted by extensive testing.
In fact, testing and observation of the field behavior
of products complement one another.

DATA COLLECTION IN THE VEHICLE

The rising number of ECUs (Electronic Control Unit)

in modern vehicles requires an intensive communication between the ECUs among themselves and with
sensors and actuators. The result is an enormous number of messages, which are communicated by the
communication networks in the vehicles to ensure all
the functions. Some of these messages contain information about the ambient and operational conditions
of the vehicle and its systems as well as information about occurred failures. Unfortunately, these
important and very useful information are not used
consequently for collecting field data over the whole
product lifetime and for following reliability analyses
until now.
In the following, the idea of using messages of the
CAN-bus for collecting field data of vehicles is illustrated. Due to the fact, that the CAN-bus corresponds
to the actual state-of-the-art to realize communication

875

The Controller Area Network (CAN) belongs to the

Local Area Networks (LAN). The CAN-bus is a serial
bus system, whose nodes of the network can receive
all available messages that are present on the CANbus. Such a system is called multi master system. The
principal configuration of a CAN-bus is illustrated in
Figure 1.
Each node has the right to access the CAN-bus if
required to broadcast a message for other nodes. The
broadcasting of the messages is not time triggered but
event controlled. To operate the order of the messages,
different priorities of importance are allocated to the
messages. These broadcasted messages are available
for all other nodes of the CAN-bus. The nodes decide
whether they need the message or not.
The nodes that are illustrated in Figure 1 can be
control units, gateways to other bus systems or a
interface to diagnostics. The interface to diagnostics
has the purpose of connecting an external diagnostics device, which in turn accords a control unit. The
line termination of the CAN-bus is a resistor to avoid
reflections.
Besides the technology of the CAN-bus there
are numerous other possibilities to realize the communication network in vehicles, e.g. flexray, LIN
(Local Interconnect Network) or MOST (Media Oriented Systems Transport). These technologies already
found or will find its way to communication networks
of vehicles, especially in sub zones of the vehicle
network, compare Felbinger (2005).
Consequently, the collection of vehicle bus signals
on basis of CAN technology is just one possibility to
collect such kind of data. Nevertheless, if the CAN

CAN
node 2
CAN
node 1

CAN
node 3

CAN-bus line

CAN
node 4

Figure 1.

line
termination
CAN
node 5

Principal configuration of a CAN-bus.

Connection via
diagnostics interface

resolution of the
data elements

Configuration of a CAN-bus

Enhancement of an
existing control unit

number of the
data elements

2.1

low effort
of implementation

networks in vehicles, a short description of the

fundamentals of the CAN-technology is given first.

Additional control unit

Figure 2. Characteristics of the possibilities of implementing the data collecting device.

technology is replaced sometime in the future, the

basic principal of the data collection and the afterwards
reliability consideration will still be applicable then.
2.2 Development of data collection device
First of all, the data elements to describe the operational and ambient conditions, that have impact on the
reliability of the product, have to be selected out of the
numerous data elements of the CAN-bus. For a procedure to perform the underlying assessment of the data
as well as the possibilities to integrate the hardware
of the collection device into the vehicle, see Leopold
(2007). The three described possibilities of integration
and its main characteristics are illustrated in Figure 2.
The low effort of implementation of the enhanced
control unit and the sufficient number and resolution
of the data elements are the reasons for choosing this
solution for further considerations.
2.3 Data processing in the vehicle
Besides the reduction of the huge amount of data by
selecting the most important messages of the CANbus, a further minimization of the data volume is
necessary in order to minimize the required storage
space in the vehicle and to reduce the amount of data
along the whole flow of information up to the company.
Therefore, possibilities of data classing of the
messages that take place already in the vehicle, see
Bertsche (2008), DIN 45667, in addition to the definition of the frequency of data recording, have to be
defined.
A further reduction of the needed storage capacity
in the vehicle appears to be possible by using compression algorithms of the information technology. An
adaptation to the requirements of the implementation
into an ECU of a vehicle has to be performed. However, the advantage of further reduction of the required
storage capacity has to be faced with the disadvantage
of rising computing power. Therefore, the potential of
using compression algorithms is not investigated more
detailed.
In addition to the selection of the required messages of the CAN-bus and the classing of the data, the
generated classed data has to be stored and transferred

876

The resulting files contain classed data with different

time scales.
Another task of the data management is to name
the files according to a determined syntax. The file
names include the serial number of the ECU, a short
cut for the time scale, the kind of data and the date.
This consistent naming enables a determination of a
prioritized data transfer to the company, which is done
by the file management.

Monitor CAN-Bus

Translate identifier to identify necessary messages

Decide on recording (right event, timeframe)

Classing (counter, time, mileage,)

Data management

File management for data transfer

Figure 3.
vehicle.

Software structure for data processing in the

to the company. The structure of the software with its

main tasks, that executes the data processing in the
vehicle, is shown in Figure 3.
Taking the potentials of data reduction into account,
a data volume of only a few hundred kilobytes per
month is possible. That means an enormous reduction
of the data volume, compared to many hundreds of
megabytes per day when recording the whole CAN
communication. As mentioned, a further reduction can
be derived by using data compression algorithms.
2.3.1 Data management
The data collection has to regard different time scales.
That offers the possibility to differ between average
values over the whole product lifetime as well as single months, days or trips. The higher the time scale,
the more general conclusions can be derived of them.
Lower time scales are especially interesting to analyze
the circumstances very short before or while failure
occurrence. For example, failures as a result of misuse
can be distinguished from failures because of fatigue.
In addition to the more or less general differentiation
of failure causes, the classed data of few data elements
can be used for more detailed analyses.
Therefore, the classing has to be executed for different time scales. To enable different files with different
durations of classing, a data management system is
needed that works simultaneously to the data classing.

2.3.2 File management

Taking into account that the data transfer to the company is realized by radio networks, see Leopold
(2007), and therefore a data transfer is not always
possible for broadcasting all existing files, a useful
sequence of the data transfer has to be managed. The
priority is rising with increasing age of the files and
increasing time scale of the contained classed data.
The reason for rising priority with increasing age of the
files is that the history of the ambient and operational
conditions is available in the company without interruption. To get an overview of the field behavior, the
files containing long time scales are of higher priority
than the files containing low time scales. The information with low time scales are used for detailed analyses
in a second step after analyzing the data with long time
scales.
The history of the broadcasted files and its status of
successful transfer are stored in a log file, see Figure 4.
The exemplary files of Figure 4 describe the classed
data of ABS Activity (Anti-lock braking system) of the
vehicle with the VIN (Vehicle Identification Number)
ADB9340. . . . In Detail, the entry 1 2 1 means 1 ABS
Activity in class 1, 2 ABS activities in class 2 and 1
ABS Activity in class 3. This kind of saving the classed
data in data fields causes only very little storage space.
The log file includes examples of files containing
trip (t1), day (d) and month (m) data. The status of 110
of the file containing the classed data of one month
describes the conditions of completeness of the file,
the execution of the transfer and the status whether the
file has been deleted on the data classing device or not.
By using the serial number of the ECU within the
file name and transferring the VIN as a part of the nonrecurring data, an easy allocation of files and vehicle
is possible.
2.4 Transfer of the data to the company
There are different technologies available, which appear to be suitable for broadcasting the data from the
vehicle to the company.
One possible realization is the usage of wireless
connections within the data collecting devices, e.g.,
Bluetooth or WLAN (Wireless Local Area Network).
In the first step, the transfer of the data of the

877

Time scale

ABS Active
#121
#201

Costs per year

Data of month

Log file

Priority of file transfer

719270_t1_20071017 110
719270_d_20071025 111
719270_m_20071031 11
110

Indirect Data Transfer with outstations

Direct Data Transfer GSM/GPRS

Data of day
ABS Active
#121
#201

Data of trip

Number of vehicles

Figure 5.

Non-recurring
N
data
Ki of vehicle: Truck
Kind
Vehicle Identification No.
V
W
WDB9340

ABS Active
#121
#201

Priority of file transfer

Age
Figure 4.

Priority of file transfer within file management.

vehicles and outstations is realized with these short

range technologies. In the second step, the data of
the outstations have to be transferred to the company, e.g., via radio networks like GSM (Global
System for Mobile Communications), GPRS (General
Packet Radio Service) or UMTS (Universal Mobile
Telecommunications System).
Another possibility is the direct transfer of the data
from the vehicle to the company via radio networks.
This kind of data transfer is already used successfully for assistance of fleet management of truckage
companies.
Depending on the number of vehicles, which are
equipped with a data collection device, the amount of
data and the available storage capacity in the vehicle,
one of the two possible realizations for broadcasting
the data has to be selected.

software as well as additional unit costs and recurring costs for the data transfer from the vehicle to the
company.
The fixed costs for the development of the detailed
investigated concepts, the direct data transfer via radio
networks and the indirect data transfer via outstations, are nearly the same. In both cases, an existing
ECU has to be enhanced with sufficient storage capacity, interfaces and the software, which manages the
data processing in the vehicle. The unit costs for
the broadcasting modules differ only slightly. The
initial investment of the concept with the indirect
data transfer is higher because of the costs for the
outstations.
The main economic differences between the two
concepts are the recurring costs for broadcasting the
data. With rising number of vehicles, which directly
transfer the data from the vehicle to the company, the
broadcasting costs are also increasing. In contrast to
the broadcasting costs of the direct data transfer, the
costs for the indirect data transfer are not rising with
higher number of equipped vehicles. The reason is that
costs causing data transfers via radio network arise
only between outstations and the company. The number of outstations is nearly constant, because the single
outstations just have to broadcast more data volume.
The principal trend of the total costs for both
concepts can be seen in Figure 5.
It is to mention that the graphical visualization of the
costs is a first estimation. The very important conclusion is that a field data collection with many vehicles
causes fewer costs when the data transfer is realized
with the indirect concept.

4
3

Cumulative costs of data collection.

DATA PROTECTION

ECONOMIC CONSIDERATION

The total costs of the data collection system includes

costs for the development of the required hard- and

A comprehensive examination of the data collection

based on the messages of the CAN-bus demands an
assessment of legal aspects. Therefore, the question

878

Fraction of total operating time [%]

35
30
25
20
15
10
5
0
<0

0-9

9-18

18-27 27-36 36-45 45-54 54-63 63-72 72-81 81-90

> 90

Vehicle speed [km/ h]

Figure 6.
vehicles.

Idealized curve of vehicle speed of long-haul

25

Fraction
action of total operating time [%]

has to be answered, whether issues of the law, which

contains the data protection, is affected or not.
The federal data protection act of Germany (BDSG,
Bundesdatenschutzgesetz) rules the matter of personal
data. According to 3 BDSG personal data are particulars of individual or factual circumstances of a defined
or definable person.
The classed data of the field data collection does
not contain the name of the driver. Therefore, the
driver is not defined obviously. In addition to that,
the selection and the classing of the data in the vehicle
reduce the original data content. After the data transfer from the vehicle to the company, no additional
data about the driver are available than the classed
data of the observed vehicle. The classed data can not
be referred to one single driver of a certain truckage
company. Thus, the data of the vehicle CAN-bus are
modified by the data selection and data classing, that
the driver is neither defined nor definable.
The result is, that the federal data protection act of
Germany does not restrict the data collection of the
CAN-bus because of the anonymized data and consequently because of missing relation of classed data to
individuals.

20

15

10

0
<0

0-9

9-18 18-27 27-36 36-45 45-54 54-63 63-72 72-81 81-90 > 90

Vehicle speed [km/h]

Figure 7.
vehicles.

EXAMPLES OF CLASSED DATA

The collection of operational and ambient conditions

in connection with occurred failures of failed vehicles offers many possibilities to analyze the real field
behavior. To compare the classed load collectives with
occurred failures in combination with damage accumulation offers the great possibility to verify and
optimize testing conditions and assumptions. Other
analyses are the comparison of the operational and
ambient conditions of failed and non-failed products, e.g. to derive the most significant factors and
combination of factors, which cause a product to fail.
It is very important to do all these analyses for products, which are comparable. An example is the very
different usage of commercial vehicles, e.g., for longhaul transport, distribution work or construction-site
work. It is very difficult to differ the usage of these
vehicles on basis of common field data. The new
approach on basis of the CAN-bus enables an easy
differentiation of the different kinds of usage. The fist
step for the differentiation is to define idealized curves
for the vehicle speed for the usage of the vehicles, see
Figures 6, 7 and 8.
To derive the kind of usage of a vehicle, the idealized curves of the vehicle speed have to be compared
with the real curve of the classed vehicle speed.
Figure 9 shows the real curve of a vehicle. The
comparison of the vehicle speed of Figure 9 with
Figures 6, 7 and 8 shows that the curve of Figure 9

Fraction of total operating time [%]

Idealized curve of vehicle speed of distribution

30
25
20
15
10
5

0
<0

0-9

9-18

18-27 27-36 36-45 45-54 54-63 63-72 72-81 81-90

> 90

Vehicle speed [km/ h]

Figure 8. Idealized curve of vehicle speed of constructionsite vehicles.

belongs to a distribution vehicle, except the high fraction of the standstill period and interurban operation.
The typical character of a distribution vehicle is the
high fraction of middle-sized vehicle speeds.
Another example of the analysis of the CANbus messages is a comparison between the differences of high and low time scales of the
classed data. A very impressive comparison is the
investigation of braking activities of a commercial vehicle. The collection of the classed data
was done for one month (high time scale) and for

879

The shown approach of using the existing information, that are already available in modern vehicles, is
an appropriate solution to a collection of field data for
reliability analyses. Data about failed and non-failed
parts are available as well as information about operational and ambient conditions over the whole lifetime
of vehicles.
The most important steps of the data processing
in the vehicle lead to a significant reduction of the
huge amount of data and enable a reasonable data
transfer between vehicle and company. The consideration of different time scales has to be considered,
as it is realized by the data management and shown
by an example. The order of transferring the data files
to the company is executed by the file management
according to a defined logic of prioritization.
A short economic consideration shows the principal
trend of the total costs for two concepts. The concept of
indirect data transfer via outstations is more economic
than the direct data transfer via radio networks.
Finally, legal aspects of the federal data protection act of Germany do not appear to be an estoppel,
although only a short summary of the investigation is
shown.
Therefore, a collection and analysis of reliability
data over the whole product lifetime of vehicles is
possible and has to be realized in forward-looking
companies.

Fraction of total operating time [%]

30

25

20

15

10

0
<0

0-9

9-18

18-27 27-36 36-45 45-54 54-63 63-72 72-81 81-90

> 90

Vehicle speed [km/h]

Figure 9.
vehicle.

Real curve of vehicle speed of a distribution

1800

Relative difference [%]

1600
1400
1200
1000
800
600
400

se
n[
io

D
ur

9 01 00

8 090

6 070

[%]

7 080

4 050

5 060

2 030

3 040

0 -1
0

1 020

Brake peda
l position

at

2 0-4 0
1 5-2 0
1 0-1 5
5 -10
0 -5

c]

200

Figure 10. Difference of high and low time scales for brake
pedal position.

REFERENCES

one day (low time scale). The brake pedal position describes the braking activity. The differences
of the results of the classed data are shown in
Figure 10.
The differences are recognizable clearly, partly up
to a factor of 16, which can be seen as an evidence
for regarding unequal time scales. The differences can
get quite higher if the differences of the loading of the
vehicles increase within different days.
6

SUMMARY

Bertsche, B.: Reliability in Automotive and Mechanical

Engineering. Berlin, Springer, 2008.
DIN 45667: Data Classing (in German). 1969.
Felbinger, L.; Schaal, H.: Trucks under Control (in German).
Elektronik Automotive 05/2005.
ISO/DIS 14220-1: Road Vehicles Unified Diagnostics Services (UDS) Part 1: Specification and Requirements.
draft standard, 2005.
Leopold, T.; Pickard, K.; Bertsche, B.: Development of
a Data Collection of Vehicle Bus Signals for Reliability Analyses. Proc. ESREL 2007, 25th27th June 2007,
Stavanger, Norway.
SAE J1939/71: Vehicle Application Layer. 2006.
VDA (ed.): Reliability Assurance in Automobile Industry and
Suppliers (in German). Vol. 3 Part 2, VDA, Frankfurt,
2000.

Field data of products over the whole product lifetime

is an essential precondition for quantitative reliability
analyses. That enables the possibility to derive analyses and forecasts of the real behavior of the products
in field usage.

880

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Comparison of phase-type distributions with mixed and additive Weibull

models
M.C. Segovia
Facultad de Ciencias, University of Granada, Spain

C. Guedes Soares
CENTEC, Instituto Superior Tcnico, Technical University of Lisbon, Portugal

ABSTRACT: The parametric estimation is performed for two models based on the Weibull distribution: the
mixture of Weibull distributions and the additive Weibull model. The estimation is carried out by a method
which use the Kolmogorov-Smirnov distance as the objective. Phase-type distributions are introduced and a
comparison is made between the fitted Weibull models and the phase-type fit to various sets of life data.

INTRODUCTION

Obtaining probabilistic models plays an important role

in the study of lifetime data of any component or system. Given a failure time data set, a usual problem
is to select the lifetime distribution and estimate its
parameters.
Because of its interesting properties, the Weibull
distribution is widely used for modelling lifetimes.
Frequently, several failure mechanisms are present
and as a result the lifetime distribution is composed for
more than one model. In these cases the mixture and
composite Weibull distributions are alternatives to the
simple model.
The mixture of two Weibull distributions provides
a rather flexible model to be fitted to data and is also
able to depict non-monotonous hazard rates.
Supposing the system structure is unknown and
there is no further information about the given system to select the best reliability model, the application
of Weibull mixture distribution as a reliability model
is always possible. Its application is feasible also in
case of complex structure systems where reliability
function is difficult to derive exactly.
On the other hand, to find distributions with bathtub
shaped failure is a frequent problem in practical situations. The usual distributions do not have this type
of failure rate. Then, when a dataset has an empirical bathtub shaped failure, it is usual to construct new
distributions from the usual ones in order to fit the
new distribution to the set of data. To construct a suitable model for the dataset, one of the operations for
this fitting, that are commonly used, is the mixtures of

distributions (Weibull, truncated normal, and others)

see Navarro & Hernndez (2004).
In this paper two models are considered to represent complex life data resulting from more than
one failure mechanism: the mixture of Weibull distributions, extending the paper of Ling and Pan
(1998); and the additive model whose failure rate
function corresponds to the sum of several failure
rate functions of Weibull form. Xie and Lai (1995)
studied the additive model with two components
to apply to data that has a bathtub-shaped failure
rate.
A method whose purpose is to minimize the
Kolmogorov-Smirnov distance is used to estimate the
parameters of the previous models. This method can
be seen in Ling and Pan (1998).
These models lead to expressions that are of difficult treatment. An alternative to these complex
expressions is to consider phase-type distributions.
This kind of distributions presents the following
advantages: they are weakly dense; hence any distribution on [0, ] can be approximated by a phasetype distribution, they can be fitted to any dataset
or parametric distribution, the obtained expressions
have algorithm treatment, their Laplace transform is
rational.
The phase-type distributions were introduced by
Neuts (1981). The versatility of the phase-type distribution is shown in Prez-Ocn and Segovia (2007)
where a particular class formed by mixtures of phasetype distributions is presented in order to calculate
special distributions with properties of interest in
reliability.

881

Phase-type distributions can be an alternative way

of describing failure rates. The purpose of this paper
is to compare the fit of complex data by phase-type
distributions with the more traditional approximations
of mixture of Weibulls or composite Weibulls.
To have control of the properties of the data used in
the analysis, simulated data was adopted. Various data
sets were generated with different levels of mixture
of different basic Weibull distributions. The simulated data are fitted with different methods allowing
conclusions about their relative usefulness.
The fit of a phase-type distribution to the set of
failure time data is carried out, with the EMpht software, based on the EM algorithm. This is an iterative
method to find the maximum likelihood estimation,
see Asmussen (1996).
In order to compare the proposed methods the
Kolmogorov-Smirnov test is used.
The paper is organized as follows; in section 2
the mixture of Weibull distributions and the additive
model are introduced. In section 3, phase-type distributions are defined and in section 4 the minimax algorithm for the parametric estimation is described. For
the proposed models, in section 5, several numerical
applications are shown.

THE MIXED AND THE ADDITIVE WEIBULL

MODEL

where i and i are, respectively, the shape and scale

parameters of the component distributions; pi representsthe weight of every component in the mixture
and ni=1 pi = 1.
The particular cases where n = 2 and n = 3 are
considered, because with a greater number of components the estimation of the parameters is much more
difficult with the minimax algorithm.
2.2 The additive model
The additive model can be interpreted as the lifetime of
a system wich consist in several independent Weibull
components that are arranged in a series. If T is the
lifetime of the system, then,
T = Min(T1 , T2 , . . . , Tn )

(4)

where Ti denote the lifetime of the component i, which

is a Weibull with parameters i , i .
Xie and Lai (1995) presented the additive model
that combines two Weibull distributions (1), one of
them with decreasing failure rate and the other with
increasing failure rate. The combined effect represents
a bathub-shaped failure rate.
The distribution function of this model is given by,
  
 2
t 1
t
F (t) = 1 exp

,
(5)
1
2

Because the Weibull distribution is used to study the

ageing and the opertional and burn-in time of a device,
two models, based on this distribution, are studied.
These models can represent in a better way the failure
time data.
The two parameter Weibull distribution function is
given by,

where i and i are the shape and scale parameters,

respectively, of the Weibull distributions.
The additive model that combines three Weibull distributions can be represented by adding a third term to
the exponential function in eqn (5).



F (t) = 1 exp (t/) , t 0,

3
(1)

where is the scale parameter and is the shape

parameter. The failure rate function has the next form
r (t) = (/) (t/)1

(2)

2.1 The mixture of Weibull distributions

In general, the mixture of Weibull distributions (1),
can be defined as a weighted sum of these distributions:
F (t) =

n

i=1


pi

  

t i
1 exp
,
i

(3)

PHASE-TYPE DISTRIBUTIONS

One of the main purposes of this paper is to compare

the fit obtained for the mixed and additive Weibull
models by the minimax algorithm, with the phase
distribution fit.
The phase-type distributions considered in this
paper are defined in the continuous case.
The continuous distribution F() on [0, ] is a
phase-type distribution (PH-distribution) with representation (, T), if it is the distribution of the time
until absorption in a Markov process on the states
{1, . . . , m, m + 1} with generator


T T0
Q=
,
(6)
0 0
and initial probability vector (, m+1 ), where is a
row m-vector. The states {1, . . . , m} are all transient.

882

of failure.

The matrix T of order m is non-singular with negative

diagonal entries and non-negative off-diagonal entries
and satisfies Te = T0 0. The distribution F() is
given by
F(x) = 1 exp(Tx)e, x 0

(7)

It will be denoted that F() follows a PH(,T)

distribution.

ESTIMATION METHOD

The selected method to estimate the parameters of

these models is the minimax algorithm, based on the
Kolmogorov-Smirnov distance.
The Kolmogorov-Smirnov test allows to check if the
distribution selected for the data set is the most appropriate, no matter what kind of parameters estimation
is used. The Kolmogorov-Smirnov test compares the
empirical distribution of the sample with the proposed
distribution, determining the maximum absolute difference between the observed probability of failure
and the expected probability of failure under the distribution, if the difference is less than a determined
critical value, the proposed distribution is accepted as
a reasonable distribution for the sample.
The minimax algorithm obtains the parameters estimation of the selected distribution minimizing the
maximum absolute difference between the observed
probability of failure and the expected probability of
failure under the selected distribution.
The steps to apply this procedure are:
1. Arrange the data in ascending order.
2. Calculate the observed probability of failure. The
median rank is used to obtain this probability:
F0 (Ti ) =

i 0.3
, i = 1, . . . , n,
n + 0.4

(8)

where Ti is the failure time of i and n is the total

number of samples.
3. Select a suitable probability distribution. Distributions that belong to the mixed Weibull model
and the additive Weibull model are the selected
distributions in this paper.
4. Determine the inital values of the parameters for the
selected distribution, but, if the initial probability
values are not appropriate, the method might cause
problems in the algorithm convergence.
5. Obtain a set of parameters optimized for the
selected distribution by minimizing the maximum
absolute difference between the observed probability of failure and the expected probability

Min Max |Fe (Ti ) F0 (Ti )| ,

i=1,2,... ,n

(9)

where Fe (Ti ) is the expected probability of

failure, which can be calculated from the selected
distribution.
To obtain the optimized parameters for the distribution, solving the previous equation, the function
fminimax in Matlab is used.
6. Compare the maximum absolute difference between the observed probability of failure and the
expected probability of failure with the allowable
critical value.

NUMERICAL APPLICATIONS

In this section the parameters of the different models

are estimated with the minimax algorithm, in order to
perform this estimation 100 values are obtained by the
Monte Carlo method. These values correspond to the
failure time data.
Furthermore, the versatility of PH-distributions is
displayed, showing their fit to the empirical distribution of the sample.
5.1 Mixture of two Weibull distributions
For a mixture of two Weibull distributions, with parameters, 1 = 0.5, 1 = 100, p1 = 0.25, 2 = 2,
2 = 1000 y p2 = 0.75 (2), 100 values are generated
by the Monte Carlo method.
If the adequate distribution for this sample is a mixture of two Weibull distributions, using the minimax
algorithm, the estimation of the parameters is indicated
in Table 1.
Figure 1 shows the fit of this distribution to the
empirical distribution.
Figure 2 represents the failure rate function that corresponds to the original distribution and the failure rate
function of the estimated distribution
Table 2 shows the estimation of the parameters,
with the minimax algorithm, for different examples
of mixtures of two Weibull distributions, and the K-S
test.
PH fit
Now the phase-type distribution, indicated in eqn (7)
that provides the best fit to the data set is obtained.

883

Table 1.

Estimated parameters.

1 = 0.3072,
2 = 2.0816,

1 = 113.1296,
2 = 903.0699,

p1 = 0.211
p2 = 0.789

As is known, this kind of distributions can be fitted to

any dataset.
To find the PH-representation of this distribution
the EMPht software is used. This representation is
given by,
= (0 0 0 1),

0.003184
0
0
0.005175

L = 0.003184
0
0
0.134512

0
0.003462
0.003184
0

0
0.000947

0
0.144241

The fit provided by the PH-distribution can be seen

in Figures 3 and 4.
Figure 1.

Fit of a Weibull mixture the sample.

Figure 2.

Failure rate of the Weibull mixture.

Table 2.

PH fit to the sample.

Parametric estimation of the Weibull mixture.

Weibull mixture
1 = 0.5
1 = 100
p1 = 0.5
1 = 0.5
1 = 100
p1 = 0.75
1 = 0.5
1 = 100
p1 = 0.5
1 = 0.5
1 = 100
p1 = 0.5
1 = 0.5
1 = 500
p1 = 0.5
1 = 0.5
1 = 1000
p1 = 0.5

Figure 3.

Estimation

K-S test

2 = 2
2 = 1000

1 = 0.475
1 = 79.25
p1 = 0.4388

2 = 2.433
2 = 937.05

D0.95 = 0.136
0, 136 > 0, 0419+

2 = 2
2 = 1000

1 = 0.542
1 = 110.6
p1 = 0.7036
1 = 0.822
1 = 97.65
p1 = 0.4488

2 = 1.416
2 = 894.9

D0.95 = 0.136
0, 136 > 0, 0441+

2 = 4.197
2 = 1021

D0.95 = 0.136
0, 136 > 0, 0393+

2 = 1
2 = 1000

1 = 0.62
1 = 225.8
p1 = 0.5452

2 = 1.039
2 = 894.9

D0.95 = 0.136
0, 136 > 0, 0497+

2 = 2
2 = 1000

1 = 0.231
1 = 346.5
p1 = 0.1998
1 = 0.492
1 = 968.04
p1 = 0.655

2 = 1.264
2 = 854.67

D0.95 = 0.136
0, 136 > 0, 0535+

2 = 2.642
2 = 946.64

D0.95 = 0.136
0, 136 > 0, 0465+

2 = 4
2 = 1000

2 = 2
2 = 1000

Critical value, + K-S experimental.

884

Table 3 shows the PH-fit for the examples given in

Table 2 and the K-S test.
5.2 Mixture of three Weibull distribution

Figure 4.

One hundred of values of a mixture of three Weibull

distributions with parameters 1 = 0.5, 1 = 100,
p1 = 0.3, 2 = 1, 2 = 500, p2 = 0.3, 3 = 2,
3 = 1000, p3 = 0.4, are simulated.
The estimation of these parameters is carried out
with the minimax algorithm. Table 4 shows the
estimated values of the parameters.
Figure 5 displays the fit to the empirical distribution, which is good.
Figure 6 represents the failure rate function corresponding to the previous distributions.

Failure rate (sample & PH).

Table 3.

Phase-type fit to the sample.

PH-representation


= 0 0.924658 0.075342 0

0.003365
0
0
0.057758
0.005832
0.036542
T =
0
0
0.004333
0
0
0.003365

K-S test

0.003365
0

0
0.003365



= 0.036284 0.963716


0.005016
0.004900
T =
0.029753
0.048887


= 0 1 0 0

0.003733
0
0.013203
0.008083
T =
0
0.000378
0
0

0, 136 > 0, 0745+

D0.95 = 0.136
0, 136 > 0, 0461+

0
0
0.003733
0.003733

0.003733
0

0
0.003733



= 0.121536 0.160842 0.717622

0.085995
0.002930
0.082800

0.002107
0.000659
T = 0.001061
0.515296
0.019454
0.571630



= 0 0.208189 0 0.791811

0.000210
0
0
0
0.027253
0.000325

T =
0.000331
0.000405
0.002275
0
0.000002
0.002268

D0.95 = 0.136
0, 136 > 0, 1039+

D0.95 = 0.136
0, 136 > 0, 0525+



= 0.025961 0.974039


0.025961
0.000187
T =
0.000067
0.001484

D0.95 = 0.136

D0.95 = 0.136
0, 136 > 0, 0830+

0.000210
0

0
0.002270

Critical value, + K-S experimental.

885

D0.95 = 0.136
0, 136 > 0, 0630+

Table 4.

Estimated parameters.

1 = 0.2682,
2 = 1.8272,
3 = 2.533,

Figure 5.

Figure 6.

1 = 10.5641,
2 = 150.2917,
3 = 919.6259,

p1 = 0.2006
p2 = 0.2444
p3 = 0.555

= 0

Distribution (sample & PH).

Figure 8.

Failure rate (sample & PH).

Fit a Weibull mixture to the sample.

Failure rate of the Weibull mixture.

The experimental value of the K-S test is 0.0418 <

D0.95 = 0.136, then the obtained estimation is
acceptable.
PH fit
The representation of the PH-distribution that provides
a better fit to the data set is


Figure 7.

0.249664

0.024926
0.996349

T = 0.007600
0.002870
0.011298

0.001439
0.005432
9.29289
0.003714
0.003963
0.000688

0.406906


0.341991 ,

0.003831
0.728256
0.014657
0.007491
0.001906

0.009449
1.778256
0.001057
0.109279
0.000456

0.006022
3.402062

0.001943
0.072219
0.014351

The phase-type fit to the original distribution can

be seen in Figures 7 and 8.
The K-S test provides us 0.1233 as experimental
value, because D0.95 = 0.136 > 0.1233 the fit is
acceptable.

5.3 Additive model Weibull with two components

Once 100 values of an Additive Weibull model with
parameters 1 = 0.5, 1 = 100, 2 = 2, 2 = 1000
(3), have been generated, the minimax algorithm is
applied to estimate these parameters. The results are
shown in Table 5.
Figure 9 represents the fit of the estimated distribution to the empirical distribution.

886

Table 5.

Estimated parameters.

1 = 0.6486,
2 = 5.6458,

Figure 9.

1 = 103.8854
2 = 849.1117

Figure 11.

PH fit to the sample.

Figure 12.

Failure rate (Weibull & PH).

Fit of an Additive Weibull to the sample.

Table 6.

Figure 10.

1 = 0.5555,
2 = 2.042,
3 = 2.0373,

Failure rate of the Weibull Additive model.

The fit seems acceptable and the experimental value

of the K-S test is lower than the critical value, D0.95 =
0.136 > 0.0554.
PH fit
The EMPht software provides the following PHrepresentation for the PH-distribution,


= 0.154338 0.027336 0.818326

0.040346 0.003227
0.028734
L = 0.002439 0.004663 0.002056
0.488996
0.037781 0.621015

Estimated parameters.
1 = 100.1662
2 = 628.2805,
3 = 628.2824,

The fit is quite good as can be seen in Figure 11.

Furthermore D0.95 = 0.136 > 0.052.
5.4 Additive model Weibull with three components
For the additive model of three components, where
1 = 0.5, 1 = 100, 2 = 1, 2 = 500, 3 = 2 and
3 = 1000, a hundred of data are simulated.
The parameters estimated are shown in Table 6.
Figures 13 and 14 show the fit provided by the
previous estimation to the data set.

887

Figure 13.

Fit an Additive Weibull to the sample.

Figure 14.

Failure rate of the Weibull Additive model.

As in the previous examples, the experimental

value for the K-S test is lower than the critical value,
D0.95 = 0.136 > 0.0507, then it can be concluded
that the estimated distribution is the distribution of the
sample.
PH fit
Finally, the PH-distribution that better fits to the
sample has this PH-representation


= 0.071507 0.928493 ,


0.019868 0.018978
T =
0.067853 0.110555
Figure 15 represents the empirical distribution and
the phase-type distribution and Figure 16 shows the
failure rate.
This is an acceptable fit, as the K-S test shows
D0.95 = 0.136 > 0.0459.

Figure 15.

PH fit to the sample.

Figure 16.

Failure rate (sample & PH).

DISCUSSION OF RESULTS

As can be seen in the previous figures, the fits provided

by minimax algorithm and the PH-distributions to the
empirical distributions are quite similar.
In figures 6 and 10 significant differences can be
observed between the estimated failure rate and
the original one. Minor variations in the estimation of
the shape parameter can cause a very different shape
of the failure rate function. However, the failure rate
of the PH-distributions is more or less similar to the
failure rate of the original distribution.
It seems that there is no relation between the number
of phases of the PH-distributions that provided a better
fit and the number of components of the distribution
that is estimated.

888

CONCLUSIONS

The PH-distributions are introduced as an useful tool

to fit any set of failure data, as it is shown in the
examples given above. The advantage of this kind
of distributions is that it is not necessary to select a
parametric distribution to fit to the data set.
On the other hand, to estimate the parameters with
the minimax algorithm, the appropriate initial value of
these parameters have to be determined, and it can take
a long time, increasing when the number of parameters
of the distribution is increased.
Finally, in all of the examples presented in the paper,
it is verified, with the K-S test, that the fits obtained,
with both methods, are all adequate, and the goodness
of fit in the previous examples is more or less similar.
REFERENCES
[1] Asmussen S, Nerman O, & Olsson M. 1996. Fitting
phase-type distributions via the em algorithm. Scand.
J. Statist. 23: 419441.
[2] Bucar, T. Nagode, M. & Fajdiga, M. 2004. Reliability
approximation using finite Weibull mixture distributions. Reliab. Engng. Syst. Safety 84: 241251.
[3] Jiang, R. & Murthy, D.N.P. 1997. Parametric study of
multiplicative model involving two Weibull distributions. Reliab. Engng. Syst. Safety 55: 217226.
[4] Jiang, R. & Murthy, D.N.P. 1998. Mixture of Weibull
distributions-parametric characterization of faiulre
rate function. Appl. Stochastic models data anal 14:
4765.
[5] Jiang, R. & Murthy, D.N.P. 2001. Models involving
two inverse Weibull distributions. Reliab. Engng. Syst.
Safety 73: 7381.

889

[6] Jiang, R. & Murthy, D.N.P. 2001. n-fold Weibull multiplicative model. Reliab. Engng. Syst. Safety 74:
211219.
[7] Ling, J. & Pan, J. 1998. A new method for selection
of population distribution and parameter estimation.
Reliab. Engng. Syst. Safety 60: 247255.
[8] Navarro, J. & Hernndez, P.J. 2004. How to obtain
bathub-shaped failure models from normal mixtures.
Probability in the Engineering and Informational
Sciences 18: 511531.
[9] Montoro-Cazorla, D., Prez-Ocn, R. & Segovia
M.C. 2007. Shock and wear models under policy N
using phase-type distributions. Applied Mathematical
Modelling. Article in press.
[10] Murthy, D.N.P. & Jiang, R. 1997. Parametric study of
sectional models involving two Weibulls distributions.
Reliab. Engng. Syst. Safety 56: 151159.
[11] Neuts, M.F. 1981. Matrix geometric solutions in
stochastic models. An algorithmic approach. Univ.
Press, Baltimore.
[12] Prez-Ocn, R. & Segovia M.C. 2007. Modeling lifetimes using phase-type distributions. Risk, Reliability
and Societal Safety. Terje Aven & Jan Erik Vinnem,
Taylor & Francis. Stavanguer, Norway 1: 463469.
[13] Sun, Y.S., Xie, M., Goh, T.N & Ong, H.L. 1993.
Development and applications of a three-parameter
Weibull distribution with load-dependent location and
scale parameters Reliab. Engng. Syst. Safety 40:
133137.
[14] Xie, M. & Lai, C.D. 1995. Reliability analysis using
an additive Weibull model with bathub-shaped failure
rate function. Reliab Engng. Syst. Safety 52: 8793.
[15] Xie, M., Tang, Y. & Goh, T.N. 2002. A modified
Weibull extension with bathub-shaped failure rate
function. Reliab Engng. Syst. Safety 76: 279285.

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Evaluation methodology of industry equipment functional reliability

J. Kamenick
Technical University of Liberec, Liberec, Czech Republic

ABSTRACT: Electric power is an essential condition of every modern economy. But it is not enough only to
maintain existing power plants, it is also necessary to develop the new ones. Companies need to make machines
with the highest possible reliability (the lowest failure rate, lowest repair timelowest unavailability) by this
development. It is very complicated to estimate availability of developed engine which have not worked yet.
We can use an estimation of reliability parameters for older machines with similar design. It is also possible to
allocate weak parts of these older machines and make some improvements.
To find such parts of machine we have to analyze it. It is usually done by statistics methods. So the problem
is where to get the input data for the analysis? The presented methodology tells us how to collect relevant data,
how to differentiate and remove unwanted data and of course how to solve the meaningful rest of data. There is
shown making of failure frequency histogram, appointment of exponential distribution for mean time between
failures of equipment inclusive of chi-square test of this hypothesis. In addition to these parts there is shown
how to perform ABC analysis of failure consequences. Very important part of this paper is Appendix 1, where
it is shown an example of methodology application.

SUMMARY OF USED ACRONYMS

A
CR
MTBF
MTBFL
MTBFU
MTTR
MTTRL
MTTRU
U

L
U
T
t
TTR
r
ri
Tp
L

expected number of failures

Czech Republic
mean time between failures
lower confidence interval limit of
real value of mean time between
failures
upper confidence interval limit of real
value of mean time between failures
mean time to repair
lower confidence interval limit of real
value of mean time to repair
upper confidence interval limit of real
value of mean time to repair
asymptotical unavailability
failure rate
lower confidence interval limit of real
value of failure rate
upper confidence interval limit of real
value of failure rate
total cumulated processing time
length of time interval of test
time to repair
total number of failures
empirical failure frequency in reflected
interval
total cumulated failure time
lower confidence interval limit of real
value of failure rate

U
2
2

upper confidence interval limit of real value

of failure rate
chi-square distribution

MOTIVATION

Electric power is essential condition of every modern

economy growth. But it is not enough only to maintain
existing power plants, it is also necessary to develop
and build the new ones. By this work it is useful to learn
form previous mistakes and weaknesses of processed
machines and equipment. Data databases are available
in most cases, but the data are not processed because
of no immediate wealth effect. This data processing
is economically profitable from long-term point of
view. Evaluation methodology of industry equipment
functional reliability grows up as a supportive document for reliability parameters calculation and weak
places finding. All examples, used in appendix 1, are
taken from analysis of huge industry pumps, which are
already finished.

DATA PART

3.1 Data collection

It is inevitable to find and collect data about
operational conditions and maintenance politics for

891

establishment of machine reliability. However, the

most important part in data collection is still failure number, failure causes and their consequences
location. Power industry in Czech Republic is split
into several branches. Nuclear area looks as separate
section. This is because of massive medial pressure.
This reason makes nuclear area being one step before
the rest of power industry. Nuclear power plants have
their own system for registration of failures and repairs
of equipment, but presented methodology takes this
also into account. Second researched area is named
classic power plantsit means black and brown
coal power plants. There is a failure monitoring software for classic power stations. Its (dis)advantages
are mentioned in the example in Appendix 1. However, recorded data are not processed on adequate level
neither in nuclear nor in classic area.
Better failure rate evaluation needs more data. So
the best evaluation uses all the data about one type
of machine. Then we sort that data regarding on what
surrounding machine is processing, what quality and
maintenance politics is applied etc. Analyst could not
do it himself; he needs help of machine operator. Getting informed is the key question of all analysis, not
only reliability ones. The best case occurs when there
exists some kind of central database of maintenance,
which is possible to simply get the relevant information from. If there is no such a database, data could be
collected in paper forms, so the analyst has to rewrite
information into electronic representation. The worst,
but still solvable, case is that the maintenance history
is saved only in operators memory. Data must be than
rewritten into tabular mode. It is recommended to work
in groups and make a record of work. Analysis could
be repeated then. When no information is available,
it is necessary to make tests or estimate parameters.
We assume data are available for this methodology
purpose.
3.2

Operational time estimation

Total cumulated operational time of equipment is one

of basest item for reliability parameters counting. We
count it as sum of all operational times of one type
equipment. Many times there is not available time
of equipment kickoff. Then we consider date of first
failure as starting time.
Great assistant of motor hours counting is MS
Excel. There is function year360, which counts
differences between two dates. This function reflects
every month as 30 days long, it means year has only
360 days in it. This is no problem regarding the fact
we do not think over scheduled outages.
Total cumulated processing time is just a sum of
all operating times. Analyst has to distinguish whether
equipment was only on position or was really working mostly in case of backup systems.

3.3

Reliability parameters calculation

This chapter handles about failure rate of equipment

type, not about each unit. It is required to erase data
about scheduled overhauls, preventive maintenance
tasks and keep only records about failures and breakdowns. These data should be sorted chronologically
so the histogram of failure numbers in each year could
be done. Then number of failures in one year is the
sum of all failures which happened on all machines of
one type.
Industrial machinery is usually operating still in the
same conditions, equipment etc. Because of that it
seems to us that their failure rate should be also the
same at the time of their operating life. This presumption responds to the exponential distribution of mean
time between failures. The only parameter of that is ,
which is reciprocal to the MTBF.
Lets do chi-square test for confirmation/negation
of exponential distribution usage validity. Exponential distribution has constant failure rate, so we can
presuppose the same number of failures in time intervals of the same length T . Splitting the test period into
m similar intervals we expect the similar number of
failures A.
A=w

d
T

[1]

(1)

where w is the length of the interval. This length should

be chosen so that there are at least 5 failures in each
interval.
d is then number of failures in tested period. Lets
count test statistics:
2 =

m

(ri A)2
i=1

[1]

(2)

To confirm hypothesis of exponential distribution

of mean time between failures the counted value of
2 fractile should be lower than its theoretical value
2 (v) for v = m 1 degrees of freedom. We can do
for example one-sided test on 10%1 significance level.
Point estimation of mean time between failures of
all machineries could be counted by [2], point 5.2:
MTBF =

1 10%

T
=
r

n

i=1 ti

[h]

(3)

significant level of hypothesis testing means that

with probability of 10% we do failure of 1st typewe
do not confirm exponential distribution although the data
really are exponentially distributed.

892

Failure rate for exponential distribution is then counted as:

1
=
MTBF

[h1 ]

(4)

Confidence intervals are done by standard [2],

paragraph 5.1.2.1. Lower confidence interval limit is
counted by form (5a), upper one by (5b):
L =
U =

0,2 05 (2r)

[h1 ]

2T

0,2 95 (2r + 2)
2T

(5a)

[h1 ]

(5b)

Where 2 (v) means fractile of distribution function of 2 distribution with v degrees of freedom.
1
U
1
MTBFU =
L
MTBFL =

[h]

(6a)

[h]

(6b)

Lower and upper limits of mean time between failures with 90% confidence interval indicate that with
90% probability the real MTBF will lay inside of
interval MTBFL , MTBFU .
The next observed reliability indicator is mean time
to repair of equipment function (MTTR). This parameter could be estimated as sum of all times to repair
divided by number of failures. Of course data must be
available.
Tp
MTTR =
r

[h]

(7)

Lower and upper confidence interval limit of mean

time to repair is counted as:
MTTRL =
MTTRU =

2T
0,2 95 (2r + 2)
2T
0,2 05 (2r)

[h]

[h]

(8a)
(8b)

Lower and upper limit of mean time to repair on

90% confidence limit says what is the times to repair
range, which are in estimated interval with 90%
probability.
Based on knowledge of these reliability parameters
we can calculate asymptotical availability (unavailability) of technological equipment:
MTTR
U =
MTTR + MTBF

[1]

3.4 Paret analysis of failure modes

Paret analysis was developed as quantitative analysis
of the most common failure modes and their effects.
It can separate substantial factors from minor ones
and show where maintenance focus effort should by
removing lacks in equipment maintenance process. So
called Paret analysis presupposes that approximately
80% of consequences are caused by only 20% of causations. This analysis accents the fact that it is not
necessary to deal with all causes but that for satisfactory effect is enough to solve only some of the most
important of them. For purpose of this study is adequate to reason about operator experiences and follow
his advices. It is possible to predict which causes will
be the most common from these experiences. Most of
failures are caused by only three failure mode groups
for the methodology purpose. These groups could be,
divided into several sub-modes, as shown in following
Table 1.
There are not all failure modes which occurred on
the equipment in the tablet. Analyst has to diagnose
all failure modes and write them down; e.g., also into
the tablet or by Paret graph, see Fig. 1.

ANALYSIS FLOW-CHART

Following flow-chart summarizes previous chapters of

methodology. It is sufficient base for skilled analyst to
make a study.

RESULTS OF THE STUDY

Analyst should summarize records of the study at

the end of the paper. Summarization should contain processing conditions, operators experiences,
modifications of equipment (if were done). As next
part of the summarization there should be emphasized weak parts of the equipment and as the last
but not least part of paper there should be written
proposals for change of machine construction. Such
study could serve as basis for new equipment development with similar parameters. Secondary outcome
of submitted procedure is estimation of reliability
parameters of produced (or prediction of reliability

(9)

893

Table 1.

List of most common failure modes.

Failure
mode

Sum
of subs

Submode 1

Submode 2

Submode 3

Submode 4

Mode 1
Mode 2
Mode 3

74
11
23

27
5
9

23
3
7

22
3
5

2
2

ACKNOWLEDGEMENT
140
120

Number of failures

This research was supported by the Ministry of Education, Youth and Sports of the Czech Republic, project
No. 1M06059Advanced Technologies and Systems
for Power Engineering and it was solved by Technical
University of Liberec, Faculty of Mechatronics and
Interdisciplinary Engineering Studies.

Cumulate number of failures

100
80
60
40

REFERENCES

20

CSN
IEC 50(191) (010102) Mezinrodn elektrotechnick
slovnkkapitola 191: Spolehlivost a jakost slueb.

CSN
IEC 60605-4 (01 0644-4) Zkouen bezporuchovosti

zarzenCst
4: Statistick postupy pro exponenciln rozdelenBodov odhady, konfidencn intervaly,
predpovedn intervaly a tolerancn intervaly.

0
Mode1 Mode 3 Mode 2 Mode 4 Mode 6 Mode 8 Mode 5 Mode 7

Figure 1.

Graph of failure frequency.

APPENDIX: METHODOLOGY APPLICATION

Experiences with data collecting

Figure 2.

Analysis proceeding flow-chart.

parameters of developed) machine. Knowledge of

reliability parameters is already standard customers
request.

Submitted methodology of technological equipment

analysis operation was applied on huge industrial
pumps made by Sigma Group, used in power industry
on Czech Republic. This appendix collects experiences from finished studies.
First step of analysis was available data collecting.
I though there was a central database of pump failures
and their repairs. So I contacted several companies,
which deal with power-plant pumps maintenance,
to get this database. I obtained extensive tablets
reduced and translated example is shown in
Table A1.
This example shows that it is possible to filter data
by pump location (column PWR_P). Column BS_C
tells us type of the pumpwhether the pump is feeding, condensation or cooling pump. Data in column
DAY mean date of failure and we are able to find time
of repair in column TIME. Unfortunately we can not
locate concrete piece of pump, so when there are more
than one (almost always) machine types in the power
plant for e.g., cooling, we can not find out what type
the data belongs to. This is the reason why I had to
get the data by a little bit more difficult way. I traveled
from one power plant to the other and collect data in
each of them.
Data in pwr-plants are saved at SHIM department.
Staff of this department also knows where to find relevant records for reliability analysis. The example of
that data is shown in Table A2. Data are translated
and some columns are hidden because of the paper
width.
There is no uniform format of equipment operating
data saving in Czech Republic. This is why analyst
has to transfer obtained data into standardized form.

894

Table A1.

Chosen columns of central pumps failure rate database.

PUMP

PWR_P

HTC

DAY

CNSQ

RSN

BS_C

AD_C

SKR_C

ST

MWH

GJ

TIME

290
317
322
339
357
360
361

ECH
ECH
ECH
ECH
ECH
ECH
ECH

B3
B3
B4
B2
B4
B2
B2

1.7.82
9.7.82
15.7.82
23.7.82
30.7.82
31.7.82
31.7.82

RS20
RE10
RH10
RS20
RE11
RE11
RX10

90
61
61
64
64
61
19

4320
2222
4320
2222
4320
2222
2222

4900
4900
4600
4900
4600
4900
4800

312
314
311
312
311
314

2
2
2
2
2
2
2

55
120
162
12
260
600
0

0
0
0
0
0
0
0

0
0,6
0,9
0
1,3
3
0

Table A2.

List of maintenance tasks.

Description

Pr. system

WRK

EQ type

Equipment

Date

TTR

Req. date

RELAY CHANGE EZH-112V-T

1VC01L051
PUMP BODY CLEANING-OIL
PUMP BEARING LEAKAGE
GDZ/485/00RUN-UP
MEASUREMENT
OIL DRAINAGE ARMOUR
BLANKING
NEW SENSOR WEC 1VC01F001
NEW SEAL ON RADIAL
BEARING

SOVP

SO

CER

1VC01D00 1

17.04.00

200

17.12.99

SOVP
SOVP
SOVP

SO
SO
SO

CER
CER
CER

1VC01D00 1
1VC01D00 1
1VC01D00 1

04.10.00
16.10.00
23.11.00

12
24
20

12.06.00
11.07.00
11.08.00

SOVP

SO

CER

1VC01D00 1

01.05.01

01.01.00

29.01.01

SOVP
SOVP

SO
SO

CER
CER

1VC01D00 1
1VC01D00 1

22.05.01
30.04.01

24.01.00
17.04.00

Table A3.
failures.
ID number

Corrected, chronologically sorted data about

Date

Repair description

Table A4.

TTR

Unique ID

1VC01D001 17.04.00 Change of relay EZH200

112VT 1VC01L051
1VC01D001 04.10.00 Pump body cleaningoil
12
1VC01D001 16.10.00 Repair of pump bearing
24
leakage
1VC01D001 23.11.00 Motor run-up measurement
20
1VC01D001 19.03.01 Oil relay checking
8
1VC01D001 30.04.01 New seal on pump radial
108
bearing
1VC01D001 30.04.01 Pump sensors disconnecting
6

Operating times of each pump.

Start of
operating

P_P1_B1_P1
3.8.1985
P_P1_B1_P2
18.12.1986
P_P1_B1_P3
15.2.1990
P_P1_B1_P4
25.4.1986
P_P1_B2_P1
18.9.1987
P_P1_B2_P2
6.4.1988
P_P1_B2_P3
25.9.1987
P_P1_B2_P4
17.8.1987
Total cumulated operational time
Total number of failures

For analysis purpose it is sufficient to get data about

unique machine identification (e.g., production number, position number), failure description and the date
of failure. Information about time to repair, which is
not always feasible, adds necessary data for availability
point estimation. Table A3 shows example of corrected
data, as written above.
Proper analysis could start when corrected data
are available. This analysis is described in following
steps.

Est. operational
time [h]
184 248
172 368
145 080
177 960
165 888
161 136
165 720
166 632
1 339 032
135

Reliability parameters estimation

Cumulated operating time is counted and shown in
following table. All machines were in operational state
all the time, so Total cumulated operating time is just
a sum of estimated operating times for each machine.
Number of failures in separate years could be mentioned in table. However graphical form is more
transparent, we can see failure rate trends. An example
of failure rate histogram in separate years is shown in
following figure.

895

Table A5.
modes.

25

Seal

74

1
1
Air
Oil
venting change
1
1

Bearing

Drain
Seal
plugged clean

27
23
22
ReseaTotal
Cooling Revision
ling
11
5
2
2

WithTotal Resea- Top-up Cover

leakage drawal
ling
23
9
5
5
2

15
10

Total Set-up Change Mass

top-up

Oil
system

Number of failures

20

Example of the three most common pump failure

O ring
montage
2

0
2006
2005
2004
2003
2002
2001
2000
1999
1998
1997
1996
1995
1994
1993
1992
1991
1990
1989
1988
1987
1986
1985

Test statistic value by (2):

Year

Figure A3.
operating.

Number of failures in separate years of

Counted value 2 is 9, 71, theoretical value 0,2 9 (7)

is equal to 12, 02. Hypothesis of mean time to pump
failure exponential distribution is confirmed on 10%
level of significance for years 19852000. In the year
2001 there were done bigger amount of maintenance
tasks, which decreases failure rate in following years.
Point estimation of mean time between failures is
counted by (3):

140
120

2 = 9, 71

Number of failures
Cumulated number of failures

100
80
60
40
20

Figure A4.
rence.

Design

Cooling spiral

Alignment

Prevention

Noisiness

Shaft

Flange

Clutch

Gearbox

Chain

Sensors

Bearing

Oil system

Seal system

0
failure mode

MTBF =

Paret graph of failure modes percent occur-

It is evident from the picture that fault current

was relatively constant for whole test period, only in
the year 2001 there was massive growing of failure
number. Lets focus on this year during failure cause
location. In this case there was a try of seal system
change, what increased number of maintenance tasks.
There were fewer failures in years after modification,
so we can say that the change was successful and pump
failure rate decreased.
Due to the fact, that in year 2001 were 24 failures
(this fact is explained, so we are allowed to do simplification) will be hypothesis of exponential distribution
of mean time to failure tested for years before 2001.
Tested time is 16 years (19852000 included), during
this time there occurred 107 failures, length of testing
period was determined to 2 years. Expected number
of failures, counted by (1), is in every 2-years interval:
A=2

107
= 13
16

1339032
[h] = 9900h
135

Failure rate, counted by (4), is then:

= 1104 h1
Confidence levels of failure rate could be obtained
by (5a) and (5b).
232, 95
[h1 ] = 8, 7 105 h1
2 1339032
311, 5
[h1 ] = 1, 2 104 h1
U =
2 1339032
L =

Confidence levels of mean time between failures

could be obtained by (6a) and (6b):
2 1339032
[h] = 8600h
311, 5
2 1339032
[h] = 11500h
MTBFU =
232, 95

896

MTBFL =

Mean time to repair is counted by (7):

MTTR =

3270
[h] = 24h
135

Confidence limits of mean time to repair are determined by (8a) and (8b):
2 3270
[h] = 21h
311, 5
2 3270
[h] = 28h
MTTRU =
232, 95
MTTRL =

Now we can finish reliability part of work by

counting of asymptotical unavailability by the form
(9):
U =

24, 2
= 2, 4 103
24, 2 + 9919

and bearing failure. That is why these failure modes

were divided out to more sub-modes, called root
causes. This subdivision is shown in Table A5.
These three dominant failure modes really covered
exactly 80% of all failures (it is example from industry, based on real data, not school one)108 from
135. Number of failures of one failure mode, sorted
by percent occurrence is charted in following Paret
graph.
Results from Paret analysis shows that the most
troubled place in the pumps operational history was
seal system. However in the year 2001 seals were modified and their failure rate decreased. Because of that
fact analyst recommends to focus on group of second most common failures, on the oil system failures.
It is recommended to observe modified seal system
to ensure that the modification really erase problems
with seals.

Paret analysis
Operators experiences show us that three most
common failures are seal leakage, oil system failure

897

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Evaluation of device reliability based on accelerated tests

E. Nogueira Daz
Telefnica I +D, UPM, Spain

M. Vzquez Lpez & D. Rodrguez Cano

EUITT-Universidad Politcnica Madrid, Spain

ABSTRACT: Reliability evaluation based on degradation is very useful in systems with scarce failures. In
this paper a new degradation model based on Weibull distribution is proposed. The model is applied to the
degradation of Light Emitting Diodes (LEDs) under different accelerated tests. The results of these tests are in
agreement with the proposed model and reliability function is evaluated.

INTRODUCTION

Reliability evaluation based on degradation models[1]

is commonly applied in highly reliable products as a
cost effective and confident way of evaluating their
reliability. In this paper a devices degradation model
is presented and subsequently applied in the quantitative analysis of LEDs reliability. With this model
the different parameters related to module reliability,
such as the reliability function, failure rate function,
the Mean Time to Failure (MTTF) or the warranty
period can be assessed based on LED degradation. In
order to obtain reliability data in a suitable period of
time degradation is measured in climatic chamber in
accelerated tests.
The classical degradation model determines the
number of failures at any time based on degradation
data. This model assumes that functionality parameter, light output in the case of LEDs, of a group of
devices follows a normal distribution in each instant of
time, whose parameters (mean and standard deviation)
change as a function of time.
In this paper the classical model limitations, from
theoretical and practical point of view, are analysed.
The calculations were performed in order to see the
temporal limitation of the classic model, using mean
and standard deviation linear variation with time. The
standard deviation trend limits are also analysed in
order to avoid non real results from the degradation
point of view as LEDS that improve with time and
light output values lower than zero.
Finally, we propose a model using the Weibull distribution to solve the classical model limitations.

CLASSIC MODEL

In degradation models it is assumed that a component

fails when one of its functional parameters (power,
voltage, light output etc) degrades enough that does
not allow it to carry out its functionality successfully.
Degradation failure is usually defined as a percentage of the nominal value from which the component
is considered to be unable to perform its function.
For example in LEDs case, the failure is considered
when the light output falls below 70% of nominal
value [2].
Classical model assumes that:
The functionality parameter is distributed following a normal distribution with an average, , and
standard deviation, .
Average and standard deviation are functions of
time, (t) and (t).
For the mean a linear variation is usually used by
several authors [34].
(t) = 0 A t

(1)

Where:
0 mean initial value
A constant that indicates the speed of degradation.
t time.
Linear trend presents a problem for t 0 /A
because, in this period of time, functionality parameter
takes values lower than zero.

899

Other authors [5] propose an exponential trend:

(t) = 0 et/C

(2)

being:
0 mean initial value
C Constant that represents the time for which the
parameter has degraded to a 36.7% of its initial
value.

or in a simpler way (t) 3(t), will pass through the

failure limit, and in that moment degradation failures
will appear.
Based on this model it is possible to evaluate the
reliability as the probability in any instant of time
that functionality parameter is within the non-failure
parameter limits.
LS
R(t) =
LL

For the time variation of the standard deviation is often

used a linear variation:
(t) = 0 + B t

f (p, t) =

1
1
e 2
(t) 2


p(t) 2
(t)

(4)

Figure 1 shows the previous model assuming a linear variation of both the average and the standard
deviation.
Due the standard deviation increases with time, the
normal distribution will be flatting with time, and
therefore in a certain instant of time the normal curve,


p(t) 2
(t)

dp

(5)

p Parameter that is being analysed

mean
standard deviation
LL and LS are the lower and upper failures limits.

Where:

In general this model assumes that the parameter

distribution in any instant of time follows a normal distribution with an average (t) and standard deviation
(t).

Where:

(3)

0 initial standard deviation.

B constant that indicates the speed of degradation
of standard deviation.
t time.

2
1
e
(t) 2

There are some device manufacturers that provide

degradation data but they are scarce. In order to obtain
data from degradation in a suitable period of time it is
necessary to use accelerate tests as it will be explained
in this paper.
Reliability from degradation data can be estimated
using the equation (5). One time parameter that
is easily evaluated with this model following the
Figure 1 is the time at which 50% of the devices failed
R(t50 ) = 0.5.
In the linear parameter trend case t50 will be:
t50 =

0 p F
A

(6)

pF failure limit parameter.

In the exponential parameter trend case t50 will be:
t50 = C ln

pF
0

(7)

It is also easily evaluated the time at which reliability is

practically zero. This time can be calculated by means
of the following equation:
(t) = pF 3(t)
3

Figure 1.

Mean and standard deviation evolution with time.

(8)

CLASSIC MODEL LIMITATIONS

Using classic model, and depending on the degradation parameter, it is possible to obtain results without
any physical sense. As an example, it is possible that
using the classic model a percentage of the devices
improves their performance, functionality parameter,
s it can be seen in Figure 2.
As can be seen in Figure 2 there is a percentage
of devices (calculated as (t) + 3(t)) that improves
their performance with time that is not possible in a

900

Figure 2. Normal distribution power values (+, , )

with mean and standard deviation linear trend according to
the classical model.

Figure 4. Normal distribution power values (+, , )

with average exponential trend and standard deviation linear
trend according to the classical model.

because depending on the parameters is possible to

approximate very different degradation trends.
In Weibull distribution functionality parameter
depends on time in the following way:


(t) = 0 e

tt0

(11)

being
Figure 3. Normal distribution power values (+, , )
with mean and standard deviation linear trend according to
the classical mode (A 3B).

degradation model. In order to avoid this situation it

is necessary that degradation trend follows the next
equation:
(t) + 3(t) 0 + 30

(10)

In the next figure it can be seen a case in which

A 3B.
In the exponential average degradation case the
analysis is very similar to the previous case as it can
be seen in Figure 4.

PROPOSED MODEL

In the paper we propose a model that is based on

the assumptions that functionality parameter decays
with time following a Weibull distribution function[6].
Weibull function is very useful due its versatility

t+t

(t + t)
0 e
  = et
=
(t)
t
0 e

(9)

In the case of mean and standard linear trend it is

necessary that:
3BA

t0 location parameter.
scale parameter.
shape parameter (or slope).
Related scale parameter:
If = 1 functionality parameter varies with respect
time following an exponential. It means that degradation rate is constant in the whole period of time:

(12)

If < 1 degradation rate decreases with time.

If > 1 degradation rate increases with time.
is the scale parameter and is the time at which the
functionality parameter has reduced to e1 (0.368).
t0 location parameter, defines the degradation starting point.
For the common case where t0 = 0 and = 1
the figure shows three curves for the three types of
described in the preceding paragraph.
Main advantages of Weibull function are:
It takes values between 0 in t = 0 and 0 (t =
infinite) according theoretical degradation models.
Although takes zero value for t equal to infinite it
is possible to model practical zero value at finite
times.

901

Pressure cooker

Figure 6. Test circuit.

Figure 5. Weibull distribution for t0 = 0, = 1 and
different values (0.2, 1 and 5).
Table 1. Power luminosity values for the days 10, 17 and
26. Pressure Cooker 110 C/85% RH.

Different parameters degradations reported in the

literature as linear degradation, exponential or others can be modeled with a Weibull distribution.
Weibull parameters give us information about the
degradation rate (constant, decreases with time or
increases with time).
5

Day

EXPERIMENTAL TESTS

Degradation model has been applied to experimental

results obtained in LEDs that has been working in a
Pressure Cooker chamber under different temperature
and humidity conditions. LEDs were working during
the tests in the pressure cooker chamber according to
the next figure.
Test procedure was:
1. 15 LEDs were introduced in the Presure Cooker
applying a bias volage according to the scheme
shown in the figure. Polarization resistance and
power supply were outside of the camera.
2. Periodically, once each day, the test was interrupted
and all the LEDs were characterized with an optical
power meter.
3. Luminosity power has been represented and compared with initial power luminosity.
In the 110 C/85% case the test lasted 29 days and
in the 19 day began to appear catastrophic failures.
Catastrophic failures have not been taken into account
in calculating means and standard deviations.
6

DEGRADATION MODEL VERIFICATION

In order to analyse the model validity we have evaluated different subjects:

1. Power luminosity of the 15 LEDs follows a normal distribution in any instant of time. Average and
standard deviation have been calculated by means
of normal representation.

LED

10

17

26

L1
L2
L3
L4
L5
L6
L7
L8
L9
L10
L11
L12
L13
L14
L15

0,552
0,574
0,613
0,591
0,616
0,606
0,634
0,606
0,610
0,623
0,604
0,628
0,629
0,575
0,565

0,322
0,488
0,549
0,538
0,560
0,524
0,384
0,623
0,567
0,385
0,489
0,570
0,600
0,519
0,535

0,183
0,316
0,476
0,294
0,534
0,400
0,265
0,551
0,481
0,171
0,000
0,462
0,521
0,000
0,410

2. Power luminosity degrades following a Weibull

function. Weibull representation has been used for
this purpose.

6.1 LEDs Power luminosity distribution

We have Table 1 shows the 15 LEDs power luminosity
results for the days ten, seventeen and twenty-six in
the test 110 C/85% RH in the Pressure Cooker.
As can be seen in Table 1 on day 26, two LEDs, 11
and 14, have failed catastrophically. These Power luminosity data have bee represented in a normal paper as
it can be seen in Figure 7. From these normal representations average and standard deviation from different
instant of time have been evaluated.
From normal plot average and standard deviation
values are easily evaluated. In the following figures
the average and standard deviation power luminosity
with respect time have been represented.

902

Weibull Pm/Po - t

Power luminosity values in normal plot

for 10, 17 y 26 days

1
y = 2,1642x - 7,2746
R2 = 0,9088

Ln (-Ln P/Po)

2
1
0
-1

0,2

0,4

0,6

0,8

-1
-2
-3
-4
-5
-6

-2

0,5

1,5

-3

Figure 7. Normal distribution representation in three different instant of time (day 10-blue and right, day 17pink-middle and day 26-yellow-left). Pressure cooker test
(110 C/85% RH).

Pm(t) Time (110 C 85 % H)

0,8
0,6

Pm

2,5

3,5

Ln (t)

Figure 10.

Power luminosity vs time in a Weibull plot.

from the beginning till eleven day and the standard

deviation is almost constant. Second period is from
eleven to eighteen day and in this period the standard
deviation increases following a linear trend, according
to the classical model. In the last period that starts in
19 day catastrophic failures appear and therefore it is
not easy to find the standard deviation trend.

0,4
0,2
0

6.2 Power luminosity Weibull function

0

10

20

30

40

Based on the figure 7 we have evaluated average power

luminosity with respect time. It can be seen that degradation does not start until day 4. In figure 10 we
have represented the relative power luminosity (relative to power luminosity at fourth day) with respect
time in a Weibull representation concluding that can
be modelled with a Weibull function.
From Weibull representation we have obtained
Weibull parameters for this specific test (110
C/85% RH).
As can be seen following the proposed law with
= 2.1642, = 28.8276days = 691, 86hours, and
therefore power luminosity evolutions with time in the
following way.

Ti m e

Figure 8.

Average Power luminosity respect time.

Standard deviation
0,35
0,3
0,25
0,2
0,15
0,1
0,05
0
1

11 13 15 17 19 21 23

Pm (t) = 0, 62e

Time

Figure 9.
time.

Standard deviation Power luminosity respect

From Figure 8 it can be seen the behaviour of power

luminosity versus time. During the first days power
luminosity increases with respect initial power luminosity and after that the device starts to degrade. This
evolution has been reported in the literature by other
authors [7,8].
From standard deviation evolution it is possible to
distinguish three different periods: the first period is

t96
691,86

2.16

(13)

RELIABILITY EVALUATION

We have evaluated reliability function and MTTF

assuming that device failure limit is 70% of nominal
luminosity. Table 2 shows in which days the failures
appear.
We have represented the failures in Weibull plot as
it can be seen in Figure 8.
From Weibull plot it is possible to evaluate reliability function obtaining a value higher than one for all
the tests indicating a degradation mechanism.

903

Table 2. Accumulated
failures at different days
(110 C/85% RH).

Days

Accumulated failures

12
13
16
20
22
23
24
25

1
3
4
8
9
11
14
15

We have proposed a degradation model based on

Weibull function that fits with several degradation
models for different devices reported in the literature.
From this model it is possible to evaluate the reliability function of any device by means of analyzing
degradation data.
First results in accelerated tests in AlGaInP LEDs
shows that:

Main conclusions of this paper are:

LEDs degradation follows a Weibull function

respect time in agreement with the proposed
model.
Reliability in all pressure cooker tests follows a
Weibull function with a shape parameter higher
than one in agreement with the degradation
mechanism.

Weibull plot

F(t)

2
1

y = 4,6621x - 14,219
R2 = 0,9309

0
-1
-2
-3
-4

Accelerated tests at different conditions are on going

in order to extrapolate reliability data at normal
working conditions.
4

REFERENCES

Ln time

Figure 11. Weibull plot for 110 C/85% RH pressure

cooker test.

CONCLUSIONS

RESULTS AND FUTURE WORK

Once average and standard deviation power luminosity has been evaluated it is easy to evaluate reliability based on the proposed model. First results
show that:
All the experiments done at different conditions (110 C/85%RH, 130 C/85% RH and
140 C/85RH) inside the pressure cooker chamber
can be fitted by the proposed model.
First results related with reliability evaluation by
the proposed model show that all of them have
similar degradation behaviour but at different rates
depending on the accelerated tests.
After several hours of operation, this time depends
on the acceleration factor, LEDs start degradation
and failure rate increases with time.
Tests at different temperature/humidity conditions
are on going in order to analyse the reliability
influence of the different parameters: temperature,
humidity and pressure.
Reliability at normal working conditions will be
evaluated when all the tests will be finished.

904

[1] Coit D.W., Evans J.L., Vogt N.T., Thomson J.R. A

method for correlating field life degradation with
reliability prediction for electronic modules. Quality
and Reliability Engineering International 2005; 21:
715726.
[2] ASSIST: Alliance for Solid-State Illumination Systems and Technologies. LED life for general lighting.
Assist Recommends 2005;1(1):113.
[3] Osterwald C.R., Benner J.P., Pruett J., Anderberg A.,
Rummeland S., Ottoson L. Degradation in weathered
crystalline-silicon PV modules apparently caused by
UV radiation. 3rd World Conference on Photovoltaic
Energy Conversion.
[4] Jia-Sheng Huang. Reliability-Extrapolation Methodology of Semiconductor Laser Diodes. IEEE
Transactions on Device and Materials Reliability, Vol.
6, NO 1, (2006).
[5] Xie J., Pecht M. Reliability Predicition Modeling of
Semiconductor Light Emitting Device. IEEE Transactions on Device and Materials Reliability Vol 6,
No3. 218222.
[6] Weibull W. Fatigue Testing and Analysis of results
Pergamon Press New York.
[7] Kish F.A., Vanderwater D.A., DeFevere D.C., Steigerwald D.A., Hofler G.E., Park K.G., Steranka F.M.
Highly reliable and efficient semiconductor waferbonded AlGaInP/GaP light-emitting diode. Electronics Letters Vol. 32, No. 2; 132134 (1996).
[8] Grillot P.N., Krames M.R., Zhao H., Teoh S.H. Sixty
Thousand Hour Light Output Reliability of AlGaInP
Light Emitting Diodes. IEEE Transactions on Device
and Materials Reliability. Vol. 6, No. 4; 564574
(2006).

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Evaluation, analysis and synthesis of multiple source information:

An application to nuclear computer codes
Sebastien Destercke & Eric Chojnacki
Institut de Radioprotection et de Sret Nuclaire, Cadarache, France

ABSTRACT: In this paper, were interested in the problem of evaluating, analyzing and synthesizing
information delivered by multiple sources about the same badly known variable. We focus on two approaches that
can be used to solve the problem, a probabilistic and a possibilistic one. They are first described and then applied
to the results of uncertainty studies performed in the framework of the OECD BEMUSE project. Usefulness and
advantages of the proposed methods are discussed and emphasized in the lights of obtained results.
1

INTRODUCTION

In the field of nuclear safety, the value of many variables are tainted with uncertainty. This uncertainty can
be due to a lack of knowledge, of experimental value or
simply because a variable cannot be directly observed
and must be evaluated by some mathematical models.
Two common problems encountered in such situations
are the following:
1. The difficulty to build synthetic representations of
our knowledge of a variable;
2. The need to compare, analyse and synthesize
results coming from different mathematical models
modeling a common physical phenomenon.
Both issues can be viewed as problems of information fusion in presence of multiple sources. In the first
case, the information can come from multiple experts,
sensors, or from different experimental results. Taking into account these multiple sources to model input
uncertainty is therefore desirable. In the second case,
the output of each single mathematical model or computer code can be considered as a single source of
information. The synthesis and analysis of the different outputs can then be treated as an information fusion
problem.
Both probability and possibility theories offer formal frameworks to evaluate, analyze and synthesize
multiple sources of information. In this paper, we
remind the basics of each methodology derived from
these theories and then apply them to the results of the
BEMUSE (Best Estimate MethodsUncertainty and
Sensitivity Evaluation) OECD/CSNI program (OCDE
2007), in which the IRSN participated.
The rest of the paper is divided in two main sections.
Section 2 details the ideas on which are based the
methods and then gives some basics about the formal

settings of the two approaches. Section 3 describes

the BEMUSE program and the application of each
methodology to its results. The benefits of using these
approaches are then discussed in the lights of the
obtained results.
2

METHODS

Most of the formal approaches proposing to handle

information provided by multiple sources consist in
three main steps: modeling the information, evaluating the sources by criteria measuring the quality
of the provided information, and synthesizing the
information.
In this paper, we will only remind the basic
ideas of each methodology, and will focus on the
results obtained with the BEMUSE program. More
details are given in an extended paper (Destercke and
Chojnacki 2007), while the probabilistic and possibilistic approaches are fully motivated respectively
in (Cooke 1991) and in (Sandri, Dubois, and Kalfsbeek
1995), respectively.
2.1

Modeling information

In this paper, we consider that the information is

provided in term of percentiles, which are surely
the commonest type of probabilistic information
encountered in safety studies. Other kinds of information cover characteristics of distribution (mean,
median, . . . ), comparative assessments, (see (Cooke
1991) and (Walley 1991, ch 4.) for extensive reviews).
Classical approaches consist of singling out a probability distribution that correspond to the information, usually by maximizing an information measure
(e.g. the entropy). Nevertheless, many arguments

905

q100%

q95 %

0.9
45%
%

q0%
0

q5%
%

5%
50 0 K 60 0 K

Figure 1.

45%
%

0.5

q50 %
5%
|
80 0 K

0.1

90 0 K 1000 K

500 K

Examples of probabilistic modeling.

Figure 2.

600 K

700 K

800 K

900 K

1000 K

Examples of possibilistic modeling.

be interpreted as lower and upper probability measures (Dubois and Prade 1992), thus defining a set
P of probability distributions such that

converge to the fact that single probabilities cannot

adequately account for incompleteness, imprecision
or unreliability in the information (See (Ferson and
Ginzburg 1996) for a short discussion). Other uncertainty theories, such as possibility theory, allows to
explicitly account for such features of the information.
Such theories are less precise than probability theory,
but ensure that no extra assumptions are added to the
available information.
The probability distribution that fits a set of percentiles qk% 1 and maximize entropy simply corresponds to a linear interpolation between percentiles.
Figure 1 represents a cumulative distribution function (CDF) corresponding to the peak clad temperature
(Maximal temperature value reached during an accidental transient phase) temperature of a fuel rode in
a nuclear reactor core for which available information
is q0% = 500 K, q5% = 600 K, q50% = 800 K,
q95% = 900 K, q100% = 1000 K. The corresponding
probability density is pictured in dashed lines.
A possibility distribution (Dubois and Prade 1988)
over the reals is formally defined as a mapping :
R [0, 1]. For a given value [0, 1], the (strict) cut of is defined as the set = {x R|(x) > }.
Given a possibility distribution , possibility  and
necessity N measures of an event A are respectively
defined as:

This relation indicates that possibility distributions

allows to model information given in terms of nested
intervals associated to confidence levels (the narrower
the interval, the less the confidence in it).
It can thus model information given by a finite
number of percentiles, as well as cases where we
have partial information about characteristics of an
unknown distribution (e.g. mean, percentiles, mode,
. . . , see (Baudrit and Dubois 2006). Figure 2 represents a possibility distribution corresponding to the
peak clad temperature of a fuel rode in a nuclear reactor core where information consists of four intervals
[750 K, 850 K], [650 K, 900 K], [600 K, 950 K],
[500 K, 1000 K] which have respective confidence
levels of 10%, 50%, 90% and 100%.

(A) = max (x) and N (A) = 1 (Ac )

2.2 Evaluating sources

xA

with Ac the complements of A. We have, for any event

A, N (A) (A), and possibility and necessity measure are respectively interpreted as upper and lower
confidence level given to an event. They can be compared to classical probabilities, where the confidence
in an event is given by a single (precise) measure.
Actually, the possibility and necessity measures can

1 the

percentile qk% of the probability distribution

P of a variable X is the deterministic value x s.t.
P(X x) = k%.

P = {P|A RN (A) P(A) (A)}

where P are probability measures over R. This set of
probabilities is also related to -cuts in the following
sense
P = {P| [0, 1], P( ) 1 }.

Once information has been given by a source, it is

desirable to evaluate the quality of this information
and of the sources. In each approach, this quality is
given by two numerical values computed on the basis
of rationality requirements:
Informativeness (Inf): evaluate the precision of the
information, by comparing it to the model representing ignorance. The more informative is a source, the
more useful the information is, and the higher is the
associated score.
Calibration (Cal): evaluate the coherence between
the provided information and some observed experimental values. The higher this coherence with
observed values, the higher the calibration score.

906

Variables on which are computed calibration are called

seed variables (that is, variables for which sources have
given information and for which experimental data are
or will be available)
In the probabilistic approach, informativeness
and calibration are computed by the means of the
Kullbach-Leibler (KL) divergence, which can be interpreted as a distance between two probabilities. The
informativeness is obtained by comparing the probability distribution pX derived from the source information to the uniform probability distribution uX defined
on the whole variation domain of the variable. Calibration is obtained by comparing probability pX to an
empirical distribution rX built from the observations.
If distributions are discretized in B elements, then the
KL divergence used to compute informativeness and
calibration of a source respectively read:
I (p, u) =

B



pi log

i=1

pi
ui

ql q5%

q50%

ql

q50%

q5%

source 2
weight 0.5
q95%
qu

1
synthesis
0

ql

q
q3 . 75%8 . 75%

Figure 3.

q42 . 5%

q65% q96 . 25%

q83 . 75%
qu

Probabilistic synthesis illustration.


1

and
I (r, p) =

source 1
weight 0.5
q95% qu

B

i=1


ri log

ri
pi

1
h

And are then transformed to obtain, for all sources,

non-negative scores summing up to one. In the probabilistic approach, calibration is based on a convergence
argument and requires about 10 experiment to ensure
a good stability. It is argued by Sandri et al. (Sandri, Dubois, and Kalfsbeek 1995) that the probabilistic
approach tends to confuse variability and imprecision.
In the possibilistic approach, informativeness is
evaluated by comparing the distribution built from the
source information to the interval covering the whole
variation domain of a variable. Calibration is simply
the extent to which experimental value are judged plausible by the built distribution. In this case, no argument
of convergence is used. Let Xr denote the variation
domain of a variable X , IXr the indicator function of
Xr (i.e. has value one in Xr , zero elsewhere), and
X the possibility distribution built from the source
information. Informativeness is given by:

X (IXr X )dx
I (X ) = r 
Xr IXr dx
and if x denote the observed value for X , calibration
score C(X ) is simply given by the value X (x ) (the
upper confidence degree given to x .
Once calibration and informativeness scores for
every source and for all variables are computed,
these scores are then normalized so that they are
non-negative and sums up to one.

Figure 4.

mean

Possibilistic synthesis illustration.

2.3 Synthesizing the information

Synthesizing the information consists of aggregating
multiple models built from the information given by
different sources to get a single models. This model
can be used in subsequent treatments or analyzed to
get information about the sources. Three main kinds
of operators are usually used:
Conjunction: equivalent to set intersection. Supposes that all sources are reliable. Conjunction gives
poorly reliable results in case of disagreement between
sources, but allows to detect such disagreement.
Disjunction: equivalent to set union. Supposes that
at least one source is reliable. Disjunction gives reliable results that are often very imprecise (hence of
limited usefulness).
Arithmetic mean: equivalent to a statistical counting of the sources. Supposes that sources are independent, and gives a result that is between conjunction
and disjunction. With this operator, sources can also
be weighted by scores obtained during the evaluation
phase.
Disjunctive and conjunctive operators are not applicable to the probabilistic approach, and it is commonly
recognized that the weighted arithmetic mean is the

907

best approach to aggregate probability distributions.

We dont consider Bayesian methods here, because we
do not assume we have prior information (see (Clemen
and Winkler 1999) for a recent review of such methods). Let p1 , . . . , pN be the probability distributions
corresponding to the information delivered by N different sources, and 1 , . . . , N be the non-negative
weights summing to one attached to these sources
(possibly provided by the evaluation procedure briefly
described in Section 2.2). The probability distribution
p obtained by arithmetic weighted mean is:
p =

N


i pi

i=1

This is not the case for the possibilistic approach, for

which conjunctive ( ), disjunctive operators ( ) and
the arithmetic mean (mean ) are well defined, allowing
for a greater flexibility in the synthesis and analysis.
Let 1 , . . . , N be the probability distributions corresponding to the information delivered by N different
sources, and 1 , . . . , N be the non-negative weights
summing to one attached to these sources (possibly
provided by the evaluation procedure briefly described
in Section 2.2). Then, classical conjunctions, disjunctions and arithmetic mean are given, for all x R, by:
(x) = min i (x)

(1)

(x) = max i (x)

(2)

i=1,... ,N

i=1,... ,N

 (x) =

N


i i (x)

(3)

i=1

Note that the above conjunctive and disjunctive

operators belongs to a broad family of mathematical

Table 1.

APPLICATION TO BEMUSE BENCHMARK

To show the usefulness and potential applications of

the methodology, we apply them to the results of the
BEMUSE (Best Estimate MethodsUncertainty and
Sensitivity Evaluation) programme (OCDE 2007) performed by the NEA (Nuclear Energy Agency). Our
study will focus on the results of the first step of the
programme, in which nine organisations were brought
together in order to compare their respective uncertainty analysis with experimental data coming from
the experiment L2-5 performed on the loss-of-fluid
test (LOFT) facility, for which an accidental transient
was simulated.
We will focus on four scalar variables for which
each participant had to provide a lower bound (Low), a
reference value (Ref) and an upper bound (Upp). These
variables are the first (PCT1) and second (PCT2) peak
clad temperature (respectively corresponding to the
peak of the blowdown and of the reflood phase), the
time of accumulator injection (Tinj ) and the time of
complete quenching (Tq ). These four variables are
amongst the more critical values that have to be surveyed in case of nuclear accident (this is particularly
true for the peak clad temperatures). Values resulting from the uncertainty studies achieved by each
participant are summarized in Table 1
For each participant and each variable, the chosen
probabilistic model was to take the lower bound as q1% ,
the reference value as q50% (median) and the upper
bound as q99% . The possibilistic model was taken as
(Low) = (Upp) = 0.02 (98% confidence interval), (Ref ) = 1 (most plausible value). Figure 5
illustrates both models built from the information of
NRI2 concerning the second PCT.

Results of the BEMUSE program.

1PCT (K )

CEA
GRS
IRSN
KAERI
KINS
NRI1
NRI2
PSI
UNIPI
UPC
Exp. val.

operators respectively called t-norms and t-conorms

(Klement, Mesiar and Pap 2000).

PCT (K )

Tinj (s)

Tq (s)

Low

Ref

Up

Low

Ref

Up

Low

Ref

Up

Low

Ref

Up

919
969
872
759
626
913
903
961
992
1103

1107
1058
1069
1040
1063
1058
1041
1026
1099
1177
1062

1255
1107
1233
1217
1097
1208
1165
1100
1197
1249

674
955
805
598
608
845
628
887
708
989

993
1143
1014
1024
1068
1012
970
972
944
1157
1077

1176
1171
1152
1197
1108
1167
1177
1014
1118
1222

14.8
14
15.8
12.7
13.1
13.7
12.8
15.2
8.0
12

16.2
15.6
16.8
13.5
13.8
14.7
15.3
15.6
16.0
13.5
16.8

16.8
17.6
17.3
16.6
13.8
17.7
17.8
16.2
23.5
16.5

30
62.9
41.9
60.9
47.7
51.5
47.4
55.1
41.4
56.5

69.7
80.5
50
73.2
66.9
66.9
62.7
78.5
62.0
63.5
64.9

98
103.3
120
100
100
87.5
82.6
88.4
81.5
66.5

908

and informal observations confirms that using formal

methods to analyze information is meaningful.
Another noticeable result is that participants using
the same code can have very different scores (both
high and low, e.g. global scores of RELAP5 users
can range from 0.025 to 0.59), which illustrates and
confirms the well-known user influence on the result
of a given computer code. Also note that, since scores
are built to be directly comparable between them, they
can also be used as code validation tools (the better
the global result, the better the information delivered
by the code). We will see in the next section that using
the results of the evaluation can improve the results of
the synthesis.

592

845

1012

1167

1228

592

845

1012

1167

1228

T(K)

T(K)

Figure 5.
2PCT.

3.1

3.2

Probability and possibility dist. of NRI1 for the

Evaluation

Table 2 shows the results of the evaluation steps performed on the results of the BEMUSE programme,
with the models described above. From a methodological point of view, we can notice that the scores and the
ranking between sources are globally in agreement,
even if there are some differences coming from the
differences between formalisms.
From a practical standpoint, interesting things can
be said from the analysis of results. First, our results are
in accordance with informal observations made in previous reports (OCDE 2007): PSI and UNIPI have high
informative scores, which reflects their narrow uncertainty bands, and have very low calibration scores, due
to the fact that, for each of them, two experimental
values are outside interval [Low, Upp]. This consistency between conclusions drawn from our methods
Table 2.

Synthesis

Figure 6 shows some results of the synthesis for the

PCT2. Since this variable is of critical importance in
accidental transient and is difficult to estimate, it is of
particular interest in the current problem.
Figure 6.A shows the synthetic probabilities when
we consider subgroup of participant using the same
code. This figures indicate that, while CATHARE
and RELAP5 users seem to underestimate the experimental value, ATHLET users tend to overestimate it.
Figure 6.B shows the benefits of weighting sources
or of selecting a subgroup of sources judged better
by the evaluation step. Such a selection and weighting shift the curves towards the experimental value
(resulting in a better global calibration) and tighten
their uncertainty bounds (resulting in a better global
informativeness). We also see that the arithmetic
mean tends to average the result, and that using
probabilistic modeling do not allow us to see possible disagreements between sources. This can be
problematic, since it is often desirable to detect and
investigate the sources of such disagreements, particularly when synthesis tools are used to analyze the
information.

Scores resulting from evaluation (Inf.: informativeness ; Cal.: Calibration).

Prob. approach

Poss. approach

Participant

Used code

Inf.

Cal.

Global

Inf.

Cal.

Global

CEA
GRS
IRSN
KAERI
KINS
NRI1
NRI2
PSI
UNIPI
UPC

CATHARE
ATHLET
CATHARE
MARS
RELAP5
RELAP5
ATHLET
TRACE
RELAP5
RELAP5

0.77
1.23
0.98
0.68
1.29
0.79
0.79
1.6
0.53
1.44

0.16
0.98
0.75
0.16
0.16
0.75
0.13
0.004
0.75
0.02

0.12
1.21
0.73
0.11
0.21
0.59
0.10
0.008
0.4
0.025

0.71
0.84
0.73
0.70
0.72
0.75
0.78
0.88
0.69
0.87

0.55
0.52
0.83
0.48
0.67
0.63
0.72
0.25
0.67
0.28

0.40
0.44
0.60
0.34
0.49
0.47
0.56
0.22
0.46
0.24

909

0.8

0.9
0.8

0.7

0.7

0.6

0.6
F(X)

F(X)

RELAP5 (KINS,NRI1,UNIPI,UPC)
CATHARE (IRSN,CEA)
ATHLET (GRS,NRI2)

0.9

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1
600

700

800

900
T(K)

1000

1100

1200

0
500

1300

6.A : Probabilities by used codes

1

1
0.9
0.8

0.7

0.7

0.6

0.6

(x)

0.8

0.5

700

800

900
T(K)

1000

1100

1200

1300

0.4

0.3

0.3

0.2

0.2

0.1

All sou rces

First Four (IRSN,KINS,NRI1,NRI2)
IRSN,KINS,NRI1,UNIPI

0.5

0.4

600

6.B : Probabilities by scores

RELAP5 (KINS,NRI1,UNIPI,UPC)
ATHLET (GRS,NRI2)
CATHARE (IRSN,CEA)

0.9

(x)

0.5

0.4

0
500

(GRS,IRSN,NRI1,UNIPI)
IRSN,KINS,NRI1,UNIPI

0.1
500

600

700

800

900
T(K)

1000

1100

1200

0
600

1300

700

800

900

T(K)

1000

1100

1200

1300

6.C : Possibilities conjunctions by used codes 6.D:Possibilities conjunctions by scores

1

All sou rces : maxi mum

0.9

0.8

0.8

0.7

0.7

0.6

0.6

(x)

(x)

0.9

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0
500

600

700

800

900

T(K)

1000

1100

1200

0
300

1300

6.E : Possibilities disjunction

Figure 6.

0.5

0.4

400

500

600

700
T(K)

800

900

1000

1100

1200

1300

6.F:Possibilities arithmetic weighted mean

Results of synthesis for PCT2: probabilistic and possibilistic approaches (- - -: experimental value).

Figure 6.C and 6.D show synthetic possibility distributions resulting from the application of a conjunctive
operator (Equation (1)). In this case, the disagreement between sources of a particular subgroup is
directly visible, both graphically and quantitatively
(disagreement is measured by the maximal height of
a distribution: the lower the distribution, the higher
the disagreement). We can thus see that information given by ATHLET users are more conflicting
than those given by CATHARE users (this could be
explained by the higher number of input data parameters in ATHLET code). Similarly, Figure 6.D shows

that all sources strongly disagreeing when considered

as a whole, but that the best sources globally agree
together, and that taking only their information into
account gives a more reliable synthesis.
Figure 6.E and 6.F respectively illustrate the synthetic possibility distributions resulting from the application of the disjunction (Equation (2)) and of the
arithmetic weighted mean (Equation (3)) over all
sources. Anew, we can see on Figure 6.F that the arithmetic mean averages the result, thus smoothing the
resulting curves. Figure 6.E well illustrates the potential high imprecision resulting from the disjunction.

910

Although the resulting uncertainty model is reliable,

its informative content appears of poor interest (e.g.
the 50% confidence interval for the 2PCT temperature
is [800, 1200], which is very broad).

are needed. This is why the IRSN is working on

methods that are more complex but remain tractable
and interpretable (Destercke, Dubolis, and Chojnacki
2007).
REFERENCES

CONCLUSIONS

We have applied methods to evaluate, synthesize and

analyze information coming from multiple sources
to results of uncertainty studies on various computer
codes. By using formal methods based on rational
requirements, evaluations are made as objective as
possible.
Proposed methods allow to take uncertainty (either
aleatory or coming from imprecision in the data)
explicitly into account in the evaluation process. They
provide interesting tools to evaluate sources. In the particular case of computer codes, they give new instrumental tools for code validation procedures (Trucano,
Swiler, Igusa, Oberkampf, and Pilch 2006), a problem
particularly important for nuclear safety institute as
the IRSN. The consistency between conclusions drawn
from our results and informal observations confirms
that using formal methods to analyze information is
meaningful and can be useful. Compared to such
informal observations, presented methods allow for
a more subtle analysis, allowing to quantify disagreement among sources, to detect biases, underestimated
uncertainty, . . .
We have also illustrated the potential advantages
offered by the use of possibility theory. In terms of
information evaluation, probabilistic and possibilistic approaches have comparable results (which is not
surprising, since they are based on similar rational
requirements). However, the possibilistic approach has
more flexibility to synthesis and analyze the information, offering a wider range of tools. The fact that both
probabilities and possibilities can be seen as special
cases of imprecise probabilities could be used to build
a generalized approach, possibly by using some recent
research results about measure of divergence for sets
of probabilities (Abellan and Gomez 2006). Such a
generalization remains the subject of further research.
Also, since results given by basic synthesizing operators can sometimes be found too rough, sometimes
more complex tools that allow for a finer analysis

Abellan, J. and M. Gomez (2006). Measures of divergence

on credal sets. Fuzzy Sets and System 157 (11).
Baudrit, C. and D. Dubois (2006). Practical representations
of incomplete probabilistic knowledge. Computational
Statistics and Data Analysis 51 (1), 86108.
Clemen, R. and R. Winkler (1999). Combining probability
distributions from experts in risk analysis. Risk Analysis
19 (2), 187203.
Cooke, R. (1991). Experts in uncertainty. Oxford, UK:
Oxford University Press.
Destercke, S. and E. Chojnacki (2007). Methods for the
evaluation and synthesis of multiple sources of information applied to nuclear computer codes. Accepted for
publication in Nuclear Eng. and Design.
Destercke, S., D. Dubois, and E. Chojnacki (2007). Possibilistic information fusion using maximal coherent
subsets. In Proc. IEEE Int. Conf. On Fuzzy Systems
(FUZZIEEE).
Dubois, D. and H. Prade (1988). Possibility Theory: An
Approach to Computerized Processing of Uncertainty.
New York: Plenum Press.
Dubois, D. and H. Prade (1992). On the relevance of nonstandard theories of uncertainty in modeling and pooling expert opinions. Reliability Engineering and System
Safety 36, 95107.
Ferson, S. and L.R. Ginzburg (1996). Different methods are
needed to propagate ignorance and variability. Reliability
Engineering and System Safety 54, 133144.
Klement, E., R. Mesiar, and E. Pap (2000). Triangular
Norms. Dordrecht: Kluwer Academic Publisher.
OCDE (2007, May). Bemuse phase iii report: Uncertainty
and sensitivity analysis of the loft l2-5 test. Technical
Report NEA/NCIS/R(2007)4, NEA.
Sandri, S., D. Dubois, and H. Kalfsbeek (1995, August).
Elicitation, assessment and pooling of expert judgments
using possibility theory. IEEE Trans. on Fuzzy Systems 3
(3), 313335.
Trucano, T., L. Swiler, T. Igusa, W. Oberkampf, and M. Pilch
(2006). Calibration, validation, and sensitivity analysis:
Whats what. Reliability Engineering and System Safety
91, 13311357.
Walley, P. (1991). Statistical reasoning with imprecise Probabilities. New York: Chapman and Hall.

911

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Improving reliability using new processes and methods

S.J. Park & S.D. Park
CS Management Center, Samsung Electronics, Suwon, Korea

K.T. Jo
Digital Printing Division, Samsung Electronics, Suwon, Korea

ABSTRACT: To improve reliability of a newly designed product, we introduced new processes and methods.
The reliability definitions contain probability, intended function, specified period and stated conditions. Therefore, it is inevitable to research the operating condition, current state of probability and target of
probability, potential failure mode and mechanism of the product for the specified period.
We conducted a 4-step test program that is, Architecture and Failure Mode Analysis, Potential Failure Mechanism Analysis, Dominant Failure Extraction and Compliance Test. Based upon the Architecture and Failure
Mode Analysis, we selected the stress factors in an ALT and reproduced the field failure mode by an ALT.
Based on the results of researches, test plans are designed to satisfy the target of reliability. Stress analysis is
also useful tool to improve the reliability of a printed circuit assembly. HALT and HASS are respectively used
to improve reliability by finding root causes of latent defects at the stage of development and screening weak
products at the stage of pre-mass production in a short time. We conducted all these kinds of processes and
methods to improve the reliability of electronic devices developed for the first time.

INSTRUCTIONS

It is well known that the shape of failure rate of product in the field shows like Figure 1. The bathtub curve
consists of three periods: an infant mortality period
with a decreasing failure rate followed by a normal
life period (also known as useful life) with a relatively constant and low failure rate and concluding
with a wear-out period that exhibits an increasing
failure rate
The most appropriate shape of failure rate is illustrated as target curve in Figure 1. Ryu called it hockey
stick line (Ryu, 2003). He mentioned that the failure rate of a bad product proceeds along the bathtub

Screening/HASS

Current

HALT
Target
ALT/Stress analysis

t
Figure 1.

Bathtub curve and reliability actions.

curve, while the failure rate of a good product follows

the hockey stick line.
The main causes of initial failures are quality defects such as workmanship error or mal-manufacturing
etc. These earlier failures are unacceptable from a
viewpoint of customer satisfaction and result in change
of their royalty to the company. Therefore, stress
screening is usually used to avoid infant mortalities. Burn-in, ESS (Environmental Stress Screening),
HASS (Highly Accelerated Stress Screening) are the
most representative methods that can eliminate the
defects. Although, appropriate specifications, adequate design tolerance and stress analysis for sufficient
component derating could decrease the initial failure, it is impossible to cover all possible interactions
between components in operation. ALT (Accelerated Life Test) and HALT (Highly Accelerated Life
Test) are used to reduce the level of failure rate during a normal operating period. HALT, quite different
from standard life testing, design verification testing
and end-of-production testing, are becoming recognized as a powerful tool to improve product reliability,
reducing warranty costs and increasing customer satisfaction. HALT has been widely used in industry
to find out the weak point in a product in a short
time (Hobbs, 2000). Although HALT is a powerful tool to find out the latent defects that can be
eliminated or reduced prior to the life tests at the

913

balance between stress and strength. We also conducted transient analysis to protect abnormal problem
in the field. For example, the stress ratio of main components is analyzed in case of mode change, on/off
testing of power, motor, heater, etc. More than 20
parts of the original design were changed to assure
the reliability of the product through stress analysis

M ea n
f

Stress

Strength

Varian ce

F ailure s

Stress/Strength

Product reliability highly depends on development

process. Most companies have their own reliability assurance process. The definition of reliability
involves probability, intended function, specified period and stated conditions. Therefore, it
is inevitable to study the operating condition, current
state of probability and target of probability, potential failure mode and mechanism of the product for
the specified period. We used to do set level reliability test such as environmental test, electrical test, and
shipping test in the past time. But nowadays we are
trying to do unit level life test and parts stress analysis
to estimate the lifetime of the designed product.

Stress
analysis

Strength

Stress

RELIABILITY ASSURANCE PROCESS

Stress/Strength

2.1
Figure 2.

Part stress vs. strength relationship.

development stage, however, it is difficult to estimate

the lifetime of the product. Therefore, an ALT is conducted for the purpose of estimating the lifetime of the
product as soon as possible in an economical way. An
ALT employs higher-than-usual levels of a stress to
quickly obtain reliability-related quantities (e.g., the
qth quantile of the lifetime distribution). In the case
of printed circuit board assembly, parts stress analysis (derating) is a basic tool of assuring that stresses,
either environmental or operational, are applied below
rated values to enhance reliability by decreasing failure
rates. It essentially prevents small changes in operating
characteristics from creating large increases in failure
rate (Reliability toolkit, 2001).
As shown in Figure 2, there are four basic stress
derating approaches to consider are:
1. Decrease average stress (i.e., using fan or heat
sink).
2. Increase average strength (i.e., using higher rated
part if size and weight increases are not a problem).
3. Decrease stress variation (i.e., using noise filter
clamping higher stresses).
4. Decrease strength variation (i.e., holding tighter
control over the process or by tests to eliminate
defects).
The purpose of these stress analysis is to improve
the reliability of the design by achieving the optimum

Define the operating environments

The first step of reliability assurance process is to

define exactly the operating condition. It may include
temperature, humidity, vibration level, voltage, frequency and altitude etc. Most industry may have
some experiences of failures resulted from insufficient
information of the use condition. Operating hours or
cycles per year are also important to set up the test plan.
These environmental profiles are directly related to all
kinds of reliability test.
2.2

Extraction of potential failure modes

and mechanisms

The second step is to find out the potential failure

modes and mechanisms. FMEA (Failure Mode and
Effect Analysis) is traditional tool by which each
potential failure mode in a system is analyzed to determine the results or effects on the system and to classify
each potential failure mode regarding its severity. But
FMEA has some limitations when finding out the
patent failure mode and mechanism. Therefore we use
4-step test program explained in detail in Section 3.
2.3

Allocation reliability goal to unit level

The third step is to determine the target of reliability

of the system and to allocate it to the unit level. This
target is based on the failure rate of the previous similar
product in the field or the level of competitor. Figure 3
shows an example of the failure rate allocation.

914

Product

Set

Annual failure rate

n%

Table 1.

Unit #1

Unit #1
#2

Unit #3

1%

1%

1%

Figure 3.

2.4

Allocation of annual failure rate to the unit level.

Determine the sample size and test time

2
(,
2r + 2)

1
T

r=0

r=1

r=2

29

41

63

n (r + 1)

1
t arg et h AF

n (r + 1)

LB
h


(4)

where LB is Bx life and x is the probability of failure

until LB .
3

4-STEP TEST PROGRAMS

We use a 4-step test program that is, architecture and

failure mode analysis, potential failure mechanism

(1)

(r + 1)

Step 1

Architecture,
failure mode analysis

Step 2

Potential failure
mechanism analysis

Step 3

Dominant failure
extraction

Step 4

Compliance test

(2)

where n = the sample size; h = the test hour.

For example, under the assumption that the target
of annual failure rate of fuser unit is 1% and the worst
case operating time is 100 hrs per year, the sample size
to be tested is calculated as follows

(3)

Therefore, we can reduce the sample size by AF times.

In the case of a decreasing or increasing failure rate,
the number of test samples would be calculated with
about a 60% confidence level as follows (Ryu & Jang,
2005).

where = failure rate; 2 = 100 (1 ) percentile

of the chi-square distribution with degrees of freedom
2r + 2; r = the number of failures; T = the total time
on test.
The failure rate should be less than or equal to the
target. Therefore, the sample size under 60% confidence level can be determined as follows (Ryu et al.
2003).
1
t arg et
nh
1
n (r + 1)
t arg et h

h = 500 hrs

Accelerated tests (ATs) have been widely used in

industry to reduce test time. If AF (Acceleration Factor) was known in an AT, equation (2) is changed as
follows.

We will describe how to make test plan for assuring

the target of failure rate.
First, we will explain the case of a constant failure
rate. If the target of failure rate is determined, the test
plan to prove it can be set up. Tests can be censored
(stopped) at either a pre-planed number of hours or a
pre-planned number of failures. Censoring at a predetermined amount of time allows for scheduling when
the test will be completed and is called Type I censoring. Censoring at a predetermined number of failures
allows for planning the maximum number of units that
will be required for testing and is referred to as Type II
censoring (Nelson, 1990). In real situation, Type I
censoring is usually adopted as a censoring scheme
to meet the due date of a product development. For a
one-sided upper confidence limit in a Type I censoring
situation with a few failure, calculate:

Determination of the sample size.

Figure 4.

915

4-step test program.

Table 2.

Table 3.

Key parts and its material.

Extraction potential failure mechanisms.

Key part

Material

Key part

Potential failure mechanism

1. Gear cap
2. Ball bearing
3. Outer pipe
4. Mica sheet
5. E-coil
6. Pi film
7. Inner pipe

Plastic
Stainless steel
Aluminum
Mica
Ni-Cr wire
500 NH
Aluminum

1. Gear cap
2. Ball bearing
3. Outer pipe
4. Mica sheet
5. E-coil
6. Pi film
7. Inner pipe

Melt/Crack
Wear
Deformation
Tear off
Breakdown
Creep
Deformation

To extract potential failure mode in the field, we

should analyze the special structure and its material.
Table 2 shows key part and its material in the heater
roller module.

(a) Heater roller.

3.2 Potential failure mechanism
Based upon the results of structure and material analysis we extracted potential failure mechanism. Table 3
shows potential failure mechanisms of each part.

(b) e-coil inside of heater roller.

Figure 5.

Components of an e-coil type heater roller.

3.3 Dominant failure mechanism extraction

analysis, dominant failure extraction and compliance

test to prevent field failure at the early stage of development. Based upon the architecture and failure mode
analysis, we selected the stress factors in an ALT and
reproduce the potential failure modes by ALTs. In this
section, we will explain the 4-step test program step
by step with the e-coil type heater roller of a printer
developed in our company.

In the real field, failure time of the product is determined by the dominant failure mechanism stimulated
by environmental stress factors such as temperature,
humidity, voltage and so on. Table 4 shows the dominant failure mechanism. As shown in Table 4, the
dominant failure mechanism is breakdown of e-coil
wire.
3.4 Compliance test

3.1

Architecture and failure mode

First step is architecture and failure mode analysis. As

shown in Table 2, heater roller is composed of 7 parts.
Figure 5 shows these items more detail.
Table 4.

Based on the relationship between potential failure

mechanism and environmental stress factor, we found
the best compliance test for the module as shown in
Table 5. The main stress factors in an ALT are thermal
cycle under power on/off.

Dominant failure mechanisms.

Environmental condition

Key
parts

Potential
failure
mechanism

Temp
(0 50 C)

Humidity
(0 85%RH)

Voltage
(21 26V)

Point

Gear cap
Ball bearing
Outer pipe
Mica sheet
E-coil
Pi film
Inner pipe

Melt/Crack
Wear
Deformation
Tear off
Breakdown
Creep
Deformation













6
5
1
6
9
6
1

: 5,  : 3, : 1.
Point 

916

Table 5.

1 Cycle = 5 hours

Compliance test.
Test types

32 C

Potential
failure
mechanism

Point

Thermal
cycle

Temp/
humidity

Power
on/off

Melt/Crack
Wear
Deformation
Tear off
Breakdown
Creep
Deformation

6
5
1
6
9
6
1













Total

126

53

57

Ranking

Table 6.

Normal environmental condition.

Environmental
condition
( C)

Temperature
Humidity (%RH)

22 C

60

90

Min

Mean

Max

Remark

0
10

30
65

50
85

Operating

Figure 6.

Test profiles for an accelerated life test.

30

Figure 7.

Test system.

ACCELERATED LIFE TEST

Test results of the accelerated tests.

Test plans

Based upon the architecture and failure mode analysis,

we selected the stress factors in an ALT and determined
test plans as follows. Table 6 shows normal environmental condition and Figure 6 shows test profiles for
ALTs of the module.
We conducted a test set and unit level using the
above thermal cycle profiles. The purpose of unit level
accelerated test is to save time and papers used in
the set level test. Figure 7 shows the unit level test
system. It is composed by fuser unit of the printer,
continuous monitoring module, rotating motor and
temperature/humidity chamber.
4.2

90

-5 C

Table 7.

4.1

30

Figure 8.

Test results

We observed if any failures occurred, and obtained

the results as shown in Table 7 and Figure 8. The main
failure mode was breakdown of the e-coil.
The failure is reproduced as same as set level failure. We analyzed the above data using the Weibull
distribution. Its pdf is given by,
   
   1

t
t
f (t) =
exp

Failure times

(5)

Set level

Humidity (%RH)

72, 612, 923

24, 45, 143, 338

Failure mode breakdown of e-coil.

where t > 0, > 0, > 0 and its reliability function

is given by
   
t
R(t) = exp

(6)

The Maximum likelihood estimates (MLEs) of

Weibull distribution parameters for the lifetime data
are presented in Table 8 and Figure 9.

917

Table 8.

calculated AF and equation (4). The annual printing

time in the field is 104 hours. The number of sample
size to assure B5 = 5 years, that is 5 % probability of
failure until 5 years, under confidence level 60 % is
calculated as Table 9.
The newly designed assembly is conducted accelerated test to evaluate the reliability. We tested 7 fusers
until 300 hrs and couldnt find any failure. Therefore, the cumulated failure probability of the newly
designed one for 5 years would be less than 5%.

Weibull parameters.

Test condition

Set level
Unit level

1.15
1.15

556.41
145.53

Figure 9.
Table 9.

Test plan for assurance test.

200

300

400

Sample size

12

4.3

In this paper, we introduced reliability assurance process and related tools such as 4-step test programs. In
particular, parts stress analysis and ALTs at the unit
level had excellent effects decreasing initial service
call rate in recent years. We explained it using e-coil
type heater roller in a laser beam printer. The unit
level accelerated test was developed to save time and
money for set level testing. In a unit level test, we
reproduced the same failure of the heater roller in a
set level test. We also developed ALT plans to assure
the goal of reliability of the module. The heater roller
was redesigned to reduce failure rate and evaluated by
ALT plan suggested in this paper.
The most important thing is to analyze and try to
decrease the gap between the estimated Bx life and the
one obtained from the field real data.

Weibull probability plot.

Test time (hrs)

The number of allowable failures is 0.

Acceleration factor

Suppose that qth quantile of a lifetime distribution

is ts for the set level test and is tu for the unit level
test. Then the acceleration factor (AF) for those two
conditions is
AF = ts /tu = 556.41/145.53 = 3.82

(7)

Therefore, we can save the test time 3.82 times by

using unit level test compared with set level.
5

CONCLUSIONS

DESIGN CHANGE AND EVALUATION

5.1 Design improvement

The heat roller is redesigned to improve the reliability
of it. We adopt a new structure and thickness of inner
pipe that can protect the breakdown of the e-coil wire
by thermal cycle.

REFERENCES
ALTA reference manual, 1998, Reliasoft.
Kececioglu, D. & Sun, F.B. 1995. Environmental Stress
Screening: Prentice Hall.
Nelson, W. 1990. Accelerated testing: John Wiley & Sons.
Park, S.J. 2000. ALT + + reference manual. SAMSUNG
ELECTRONICS CO., LTD.
Park S.J., Park S.D. & Kim K.S. 2006. Evaluation of the Reliability of a Newly Designed Duct-Scroll by Accelerated
Life Test, ESREL 2006.
Reliability toolkit. 2001. RAC Publication, CPE. Commercial practices edition.
Ryu D.S., Park S.J. & Jang S.W. 2003. The Novel Concepts
for Reliability Technology, 11th Asia-Pacific Conference
on Non-Destructive Testing, Nov. 37.
Ryu D.S, & Jang S.W. 2005. The Novel Concepts for
Reliability Technology, Microelectronics Reliability 45:
611622.

5.2 Reliability evaluation

We made accelerated test plans to assure reliability
of the newly designed heat roller based upon the

918

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Life test applied to Brazilian friction-resistant low alloy-high strength

steel rails
Daniel I. De Souza
North Fluminense University & Fluminense Federal University, Campos & Niteri, RJ, Brazil

Assed Naked Haddad

Rio de Janeiro Federal University, Rio de Janeiro, RJ, Brazil

Daniele Rocha Fonseca

North Fluminense State University, Campos, RJ, Brazil

ABSTRACT: In this work we will apply a combined approach of a sequential life testing and an accelerated
life testing to friction-resistant low alloy-high strength steel rails used in Brazil. One possible way to translate test
results obtained under accelerated conditions to normal using conditions could be through the application of the
Maxwell Distribution Law. To estimate the three parameters of the underlying Inverse Weibull sampling model
we will use a maximum likelihood approach for censored failure data. We will be assuming a linear acceleration
condition. To evaluate the accuracy (significance) of the parameter values obtained under normal conditions for
the underlying Inverse Weibull model we will apply to the expected normal failure times a sequential life testing
using a truncation mechanism developed by De Souza (2005). An example will illustrate the application of this
procedure.

INTRODUCTION

The sequential life testing approach is an attractive

alternative to that of predetermined, fixed sample size
hypothesis testing because of the fewer observations
required for its use, especially when the underlying
sampling distribution is the three-parameter Inverse
Weibull model. It happens that even with the use of
a sequential life testing mechanism, sometimes the
number of items necessary to reach a decision about
accepting or rejecting a null hypothesis is quite large
(De Souza 2000, De Souza 2001). Then, for a threeparameter underlying Inverse Weibull Distribution, a
truncation mechanism for this life-testing situation
was developed by De Souza (2005) and an application of this mechanism was presented by De Souza
& Haddad (2007). But it happens that sometimes
the amount of time available for testing could be
considerably less than the expected lifetime of the
component. To overcome such a problem, there is
the accelerated life-testing alternative aimed at forcing
components to fail by testing them at much higherthan-intended application conditions. These models
are known as acceleration models. One possible
way to translate test results obtained under accelerated conditions to normal using conditions could be

through the application of the Maxwell Distribution

Law.
The Inverse Weibull model was developed by Erto
(1982). It has been used in Bayesian reliability estimation to represent the information available about the
shape parameter of an underlying Weibull sampling
distribution, as in (Erto 1982, De Souza & Lamberson 1995). It has a location (or minimum life), a
scale and a shape parameter. It has also been used in
reliability estimation of electronic products where it
seems to have a better answer to the accuracy problem presented by the Weibull model, as shown by
De Souza (2005). It happens that when the shape
parameter of the Weibull model is greater than 7, the
Weibull curve becomes highly pointed, resulting in
some computational difficulty (accuracy) in calculating the components characteristics of interest values.
The three-parameter Inverse Weibull distribution has
a location, a scale and a shape parameter. Its density
function is given by:

f (t) =

+1

t 0; , , > 0.

919

 
 

exp
t

THE ACCELERATING CONDITION

The Maxwell Distribution Law, which expresses the

distribution of kinetic energies of molecules, is given
by the following equation:
MTE = Mtot eE/KT

(1)

MTE represents the number of molecules at a particular absolute Kelvin temperature T (Kelvin = 273.16
plus the temperature in Centigrade), that passes a
kinetic energy greater than E among the total number
of molecules present, Mtot ; E is the energy of activation of the reaction and K represents the gas constant
(1.986 calories per mole). Equation 1 expresses the
probability of a molecule having energy in excess of
E. The acceleration factor AF2/1 at two different stress
temperatures, T2 and T1 , will be given by the ratio of
the number of molecules having energy E at these two
different temperatures, that is:
eE/KT2
MTE (2)
= E/KT
1
MTE (1)
e
 

E 1
1
= exp

K T1
T2

AF2/1 =
AF2/1

(2)

Applying natural logarithm to both sides of Equation 1 and after some algebraic manipulation, we will
obtain:






E 1
MTE (2)
1
=

(3)
ln AF2/1 = ln
MTE (1)
K T1
T2
From Equation 3 we can estimate the term E/K by
testing at two different stress temperatures and computing the acceleration factor on the basis of the fitted
distributions. Then:


ln AF2/1
E

(4)
=

1
K
1
T1

T2

The acceleration factor AF2/1 will be given by the

relationship 1 /2 , with i representing a scale parameter or a percentile at a stress level corresponding to
Ti . Once the term E/K is determined, the acceleration
factor AF2/n to be applied at the normal stress temperature is obtained from Equation 2 by replacing the
stress temperature T1 with the temperature at normal
condition of use Tn . Then:
 

E 1
1

(5)
AF2/n = exp
K Tn
T2

Weibull model represents the life distribution at one

stress level, a three-parameter Inverse Weibull model
also represents the life distribution at any other stress
level. We will be assuming a linear acceleration condition. In general, the scale parameter and the minimum
life can be estimated by using two different stress levels
(temperature or cycles or miles, etc.), and their ratios
will provide the desired value for the acceleration
factors AF and AF . Then:
n
a
n
AF =
a
AF =

(6)
(7)

According to De Souza (2005), for the Inverse

Weibull model the cumulative distribution function
at normal testing condition Fn (tn n ) for a certain
testing time t = tn , will be given by:

n
 
n
t
AF
(8)

= exp
Fn (t) = Fa
n
AF
AF t AF
Equation 8 tells us that, under a linear acceleration assumption, if the life distribution at one stress
level is Inverse Weibull, the life distribution at any
other stress level is also an Inverse Weibull model.
The shape parameter remains the same while the
accelerated scale parameter and the accelerated minimum life are multiplied by the acceleration factor.
The equal shape parameter is a necessary mathematical consequence to the other two assumptions;
assuming a linear acceleration model and an Inverse
Weibull sampling distribution. If different stress levels
yield data with very different shape parameters, then
either the Inverse Weibull sampling distribution is the
wrong model for the data or we do not have a linear
acceleration condition.
3

HYPOTHESIS TESTING SITUATIONS

The hypothesis testing situations will be given by:

1. For the scale parameter :
H0 : 0 ;

H1 : < 0

The probability of accepting H0 will be set at

(1) if = 0 . If = 1 where 1 < 0 , the probability of accepting H0 will be set at a low level .
2. For the shape parameter :

De Souza (2005) has shown that under a linear

acceleration assumption, if a three-parameter Inverse

920

H0 : 0 ;

H1 : < 0

The probability of accepting H0 will be set at

(1 ) if = 0 . If = 1 where 1 < 0 , the

probability of accepting H0 will also be set at a low

level .
3. For the location parameter :
H0 : 0 ;

EXPECTED SAMPLE SIZE

OF A SEQUENYIAL LIFE TESTING

According to Mood & Graybill (1963), an approximate expression for the expected sample size E(n) of
a sequential life testing will be given by:

H1 : < 0

Again, the probability of accepting H0 will be set

at (1 ) if = 0 . Now, if = 1 where < 0 ,
then the probability of accepting H0 will be once more
set at a low level .

E(n) =

E(Wn )
E(w)

(11)

Here, w is given by:

4

SEQUENTIAL TESTING

According to (Kapur & Lamberson 1977, De Souza

2004), the development of a sequential test uses the
likelihood ratio given by the following relationship:
L1;n /L0;n
The sequential probability ratio (SPR) will be given
by SPR = L1,1,1,n /L0,0,0,n , or yet, according to De
Souza (2004), for the Inverse Weibull model, the
sequential probability ratio (SPR) will be:

w = ln

f (t; 1 , 1 , 1 )
f (t; 0 , 0 , 0 )

(12)

The variate Wn takes on only values in which Wn

exceeds ln (A) or falls short of ln (B). When the true
distribution is f (t; , , ), the probability that Wn
takes the value ln (A) is P(, , ), while the probability that it takes the value ln (B) is 1 P(, , ). Then,
according to Mood & Graybill (1963), the expression for the expected value of the variate Wn will
be given by:

n n 


11  (ti 0 )0 +1
SPR =
0
(ti 1 )1 +1
00
i=1

 n 


00
1 1

exp
(ti 1 )1
(ti 0 )0
i=1


E Wn P (, , ) ln (A) + [1 P (, , )] ln (B)
(13)

So, the continue region becomes A < SPR < B,

where A = /(1 ) and B = (1 )/. We will
accept the null hypothesis H0 if SPR B and we will
reject H0 if SPR A. Now, if A < SPR < B, we
will take one more observation. Then, by taking the
natural logarithm of each term in the above inequality
and rearranging them, we get:

E (n)


(1 )
n ln
<X
ln

00





1 1
1
(1 )
+ ln
< n ln
0

00

X=

n

i=1

n

i=1

1
1
0

1 1

(ti 1 )1

Hence, with A = /(1 ) and B = (1 )/,

Equation 11 becomes:

E (w) = ln (C) + (0 + 1) E [ln (ti 0 )] (1 + 1)


1
1
1
E [ln (ti 1 )] 1 E
(ti 1 )

0
1

+ 0 0 E
(15)

(ti 0 )

ln (ti 1 ) (10)

The solution for the components of Equation 15

can be found in De Souza (2004).

(ti 0 )0

ln (ti 0 ) + (1 + 1)

n

i=1

Equation 14 enables one to compare sequential tests

with fixed sample size tests. The proofs of the existence of Equations 11 to 14 can be found in Mood &
Graybill (1963), pp. 391392.
For a three-parameter Inverse Weibull sampling distribution, the expected value of Equation 12 will be
given by:

(9)

00

P (, , ) ln (A) + [1 P (, , )] ln (B)
E (w)
(14)

(0 + 1)

921


dL r
= + r ln ()
ln (ti )
d
r

MAXIMUM LIKELIHOOD ESTIMATION

FOR THE INVERSE WEIBULL MODEL
FOR CENSORED TYPE II DATA (FAILURE
CENSORED)

The maximum likelihood estimator for the shape, scale

and minimum life parameters of an Inverse Weibull
sampling distribution for censored Type II data (failure
censored) will be given by:

r 

i=1

i=1

ti


(n r)


ln

tr

ti


ln

tr


=0

(21)


1
dL
= ( + 1)

d
(ti )
i=1
 r 
+1

1

ti
r

L(; ; ) = k!

 r



f (ti ) [1 F(tr )]nr , or yet:

i=1

L (; ; ) = k!

 r



f (ti ) [R(tr )]nr ; t > 0

i=1

f (ti ) =

ti
 

+1

 
exp
 

tr
 r

r r
L (; ; ) = k!

R(tr ) = exp

i=1

r


ti

i=1

(16)

1
+ (n r)
tr

 

(/ti )

i=1

e(/tr )

nr
(19)

The log likelihood function L = ln [L (; ; )] will

be given by:

(n r)

i=1

tr

1
ti

r
+ (n r)

1
tr

(23)

Notice that, when = 1, Equation 23 reduces

to the maximum likelihood estimator for the inverse
two-parameter exponential distribution. Using Equation 23 for in Equations 21 and 22 and applying some
algebra, Equations 21 and 22 reduce to:

L = ln (k!) + r ln () + r ln () ( + 1)

r
r 



ln (ti )

ti
i=1

1/

i=i

(22)

From Equation 20 we obtain:

+1

=0

(17)

(18)
1
(ti )

+1 

r 
ln (ti )

i=1
 r

 1 
ln (ti )
r
ti
i=1



1
ln (tr )
+ (n r) tr
=0
+




1
1
+

r)
(n
ti
tr
r

i=1

(24)
To find the value of and that maximizes the log
likelihood function, we take the , and derivatives
and make them equal to zero. Then, we will have:

dL r
= 1
d

r 

i=1

(n r) 1

1
ti

( + 1)

i=1

1
tr

r


=0

(20)

922

1
(ti )

1
ti

i=1
r


i=1

+1

1
ti

+ (n r)

+ (n r)

1
tr

1
tr

+1 

=0
(25)

Equations 24 and 25 must be solved iteratively.

The problem was reduced to the simultaneous solution of the two iterative Equations 24 and 25. The
simultaneous solution of two iterative equations can
be seen as relatively simple when compared to the
arduous task of solving three simultaneous iterative
Equations (20, 21, 22) as outlined by Harter (Harter et al. 1965). Even though this is the present case,
one possible simplification in solving for estimates
when all three parameters are unknown could be the
following approach proposed by Bain (1978).
For example, let us suppose that and represent
the good linear unbiased estimators (GLUEs) of the
shape parameter and of the scale parameter for a
fixed value of the minimum life . We could choose an
and
initial value for to obtain the estimators and ,
then apply these two values in Equation 18, that is, the
maximum likelihood equation for the minimum life .
An estimate can then be obtained from Equation 25,
then the GLUEs of and of can be recalculated for
the new estimate ,
and a second estimate for the minimum life obtained from Equation 25. Continuing
this iteration would lead to approximate values of the
maximum likelihood estimators. As we can notice, the
advantage of using the GLUEs in this iteration is that
only one equation must be solved implicitly. The existence of solutions to the above set of Equations 24
and 25 has been frequently addressed by researchers
as there can be more than one solution or none at all;
see Zanakis & Kyparisis (1986).
The standard maximum likelihood method for estimating the parameters of the three-parameter Weibull
model can have problems since the regularity conditions are not met, see (Murthy et al. 2004, Blischke
1974, Zanakis & Kyparisis 1986). To overcome this
regularity problem, one of the approaches proposed by
Cohen (Cohen et al. 1984) is to replace Equation 25
with the equation

we can determine a 0 value which should make the

right side of Equation 26 equal to the first failure time
t1 . When the decisions about these quantities 0 , 1 ,
0 , 1 , 0 , 1 , , and P(, ) are made, and after the
E(n) is calculated, the sequential test is totally defined.

We are trying to determine the values of the shape,

scale and minimum life parameters of an underlying
three-parameter Inverse Weibull model, representing
the life cycle of a new friction-resistant low alloy-high
strength steel rail. Once a life curve for this steel rail
is determined, we will be able to verify using sequential life testing, if new units produced will have the
necessary required characteristics. It happens that the
amount of time available for testing is considerably
less than the expected lifetime of the component. So,
we will have to rely on an accelerated life testing
procedure to obtain failure times used on the parameters estimation procedure. The steel rail has a normal
operating temperature of 296 K (about 23 degrees
Centigrade). Under stress testing at 480 K, 16 steel
rail items were subjected to testing, with the testing
being truncated at the moment of occurrence of the
twelfth failure. Table 1 shows these failure time data
(hours).
Now, under stress testing at 520 K, 16 steel rail
items were again subjected to testing, with the testing
being truncated at the moment of occurrence of the
twelfth failure. Table 2 shows these failure time data
(hours).
Using the maximum likelihood estimator approach
for the shape parameter , for the scale parameter
and for the minimum life of the Inverse Weibull
model for censored Type II data (failure censored), we
obtain the following values for these three parameters

k+1


 
1/  

g
Ui n
Ui
1e
n
(1, 2or4)
3

Table 1. Failure times (hours) of steel rail items tested under

accelerated temperature conditions (480 K).

i=1

+ n j

EXAMPLE

k+1


 
n
g

1 eUi (1, 2or4) = t1

3
i=1

(26)

Here, t1 is the first order statistic in a sample of

size n. In solving the maximum likelihood equations,
we will use this approach proposed by Cohen (Cohen
et al. 1984). Appendix 1 shows the derivation of
Equation 26.
Some preliminarily life testing is then performed in
order to determine an estimated value for the first failure time x1 of the underlying Inverse Weibull sampling
distribution. Then, using a simple search procedure,

923

765.1
862.2
909.4
973.2

843.6
877.3
930.9
1,014.7

850.4
891.0
952.4
1,123.6

Table 2. Failure times (hours) of steel rail items tested under

accelerated temperature conditions (520 K).
652.5
692.9
738.2
784.9

673.6
705.1
769.2
816.0

683.1
725.4
776.6
981.9

under accelerated conditions of testing:

normal stress temperature AF2/n , will be:

At 480 K. 1 = n = = 8.38; 1 = 642.3 hours;




1
1
AF2/n = exp 1, 015.1

= 4.38
296 520

1 = 117.9 hours
At 520 K.2 = n = = 8.41; 2 = 548.0 hours;
2 = 100.2 hours
The shape parameter did not change with 8.4.
The acceleration factor for the scale parameter AF2/1
will be given by:
AF2/1 = 1 /2 = 642.3/548.0
Using Equation 4, we can estimate the term E/K.


ln AF2/1
ln (642.3/548.0)
E
 =  1
= 990.8
=

1
1
1
K
480 520
T1 T2
Using now Equation 5, the acceleration factor for
the scale parameter, to be applied at the normal stress
temperature AF2/n , will be:



1
1

Tn
T2



1
1

= 4.23
= exp 990.8
296 520

AF2/n = exp
AF2/n

E
K

Therefore, the scale parameter of the component at

normal operating temperatures is estimated to be:
n = AF2/n 2 = 4.23 548.0 = 2, 318.0 hours
The acceleration factor for the minimum life parameter AF2/1 will be given by:
AF2/1

Then, as we expected, AF = 4.23 AF =

4.38 AF = 4.3. Finally, the minimum life parameter of the component at normal operating temperatures
is estimated to be:
n = AF2/n 2 = 4.3 100.2 = 430.9 hours
Then, the steel rail life when operating at normal use
conditions could be represented by a three-parameter
Inverse Weibull model having a shape parameter of
8.4; a scale parameter of 2,318.0 hours and a minimum life of 430.9 hours. To evaluate the accuracy
(significance) of the three-parameter values obtained
under normal conditions for the underlying Inverse
Weibull model we will apply, to the expected normal
failure times, a sequential life testing using a truncation mechanism developed by De Souza (2004). These
expected normal failure times will be acquired by multiplying the twelve failure times obtained under accelerated testing conditions at 520 K given by Table 2 by
the accelerating factor AF of 4.3. It was decided that
the value of was 0.05 and was 0.10. In this example,
the following values for the alternative and null parameters were chosen: alternative scale parameter 1 =
2, 100 hours, alternative shape parameter 1 = 7.8
and alternative location parameter 1 = 380 hours;
null scale parameter 0 = 2, 320 hours, null shape
parameter 0 = 8.4 and null minimum life parameter
0 = 430 hours. Now electing P(, , ) to be 0.01,
we can calculate the expected sample size E(n) of this
sequential life testing under analysis. Using now Equation 15, the expression for the expected sample size of
the sequential life testing for truncation purpose E(n),
we will have:
E (w) = 5.501 + 9.4 7.802 8.8 7.749

1
117.9
=
=
2
100.2

0.033 + 1.0 = 0.611

Now, with P(, , ) = 0.01, with

Again applying Equation 4, we can again estimate

the term E/K. Then:


ln AF2/1
ln (117.9/100.2)
E
 =  1
= 1, 015.1
=

1
1
1
K
480 520
T1 T2


ln (B) = ln




(1 )
(1 0.10)
= ln
= 2.8904,

0.05

and also with



Using once more Equation 5, the acceleration factor

for the minimum life parameter, to be applied at the

ln (A) = ln

924


= ln

0.10
1 0.05


= 2.2513,

NUMBER OF ITEMS
0
0

V
A
L
U
E
S

-5
-10
ACCEPT Ho

-15
-20
-25

O
F
X

-30

REJECT Ho

-35
-40

Figure 1. Sequential test graph for the three-parameter

Inverse Weibull model.

we will have:
P (, ) ln (A) + [1 P (, )] ln (B)
= 0.01 2.2513 + 0.99 2.8904 = 2.8390
Then: E (n) = 2.8390
0.6115 = 4.6427 5 items.
So, we could make a decision about accepting or
rejecting the null hypothesis H0 after the analysis of
observation number 5. Using Equations 9 and 10 and
the twelve failure times obtained under accelerated
conditions at 520 K given by Table 2, multiplied by
the accelerating factor AF of 4.3, we calculate the
sequential life testing limits. Figure 1 below shows the
sequential life-testing for the three-parameter Inverse
Weibull model.
Then, since we were able to make a decision
about accepting or rejecting the null hypothesis H0
after the analysis of observation number 4, we do not
have to analyze a number of observations corresponding to the truncation point (5 observations). As we
can see in Figure 1, the null hypothesis H0 should be
accepted since the final observation (observation number 4) lays on the region related to the acceptance of H0 .

CONCLUSIONS

There are two key limitations to the use of the Arrhenius equation: first, at all the temperatures used, linear specific rates of change must be obtained. This
requires that the rate of reaction, regardless of whether
or not it is measured or represented, must be constant
over the period of time at which the aging process is
evaluated. Now, if the expected rate of reaction should
vary over the time of the test, then one would not be

able to identify a specific rate that is assignable to a

specific temperature. If the mechanism of reaction at
higher or lower temperatures should differ, this, too,
would alter the slope of the curve. Second, it is necessary that the energy activation be independent of
temperature, that is, constant over the range of temperatures of interest. It happens that, according to
Chornet & Roy (1980), the apparent energy of activation is not always constant, particularly when there
is more than one process going on. Further comments
on the limitations of the use of the Arrhenius equation
can be found in Feller (1994). In this work we lifetested a new industrial product using an accelerated
mechanism. We assumed a linear acceleration condition. To estimate the parameters of the three-parameter
Inverse Weibull model we used a maximum likelihood
approach for censored failure data, since the lifetesting will be terminated at the moment the truncation
point is reached. The shape parameter remained the
same while the accelerated scale parameter and the
accelerated minimum life parameter were multiplied
by the acceleration factor. The equal shape parameter is a necessary mathematical consequence of the
other two assumptions; that is, assuming a linear acceleration model and a three-parameter Inverse Weibull
sampling distribution. If different stress levels yield
data with very different shape parameters, then either
the three-parameter Inverse Weibull sampling distribution is the wrong model for the data or we do not
have a linear acceleration condition. In order to translate test results obtained under accelerated conditions
to normal using conditions we applied some reasoning
given by the Maxwell Distribution Law. To evaluate the accuracy (significance) of the three-parameter
values estimated under normal conditions for the
underlying Inverse Weibull model we employed, to the
expected normal failure times, a sequential life testing using a truncation mechanism developed by De
Souza (2004). These expected normal failure times
were acquired by multiplying the twelve failure times
obtained under accelerated testing conditions at 520 K
given by Table II, by the accelerating factor AF of
4.3. Since we were able to make a decision about
accepting or rejecting the null hypothesis H0 after the
analysis of observation number 4, we did not have to
analyze a number of observations corresponding to
the truncation point (5 observations). As we saw in
Figure 1, the null hypothesis H0 should be accepted
since the final observation (observation number 4)
lays on the region related to the acceptance of H0 .
Therefore, we accept the hypothesis that the frictionresistant low alloy-high strength steel rails life when
operating at normal use conditions could be represented by a three-parameter Inverse Weibull model
having a shape parameter of 8.4; a scale parameter of 2,320 hours and a minimum life of 430
hours.

925

REFERENCES
Bain, Lee J. 1978. Statistical Analysis of Reliability and
Life-Testing Models, Theory and Method. Marcel Dekker,
Inc., New York, NY, USA.
Blischke, W.R. 1974. On non-regular estimation II. Estimation of the Location Parameter of the Gamma and
Weibull Distributions, Communications in Statistics, 3,
11091129.
Chornet & Roy. 1980. Compensation of Temperature on Peroxide Initiated Cross linking of Polypropylene, European
Polymer Journal, 20, 8184.
Cohen, A.C.; Whitten, B.J. & Ding, Y. 1984. Modified Moment Estimation for the Three-Parameter
Weibull Distribution, Journal of Quality Technology 16,
159167.
De Souza & Daniel I. 2000. Further thoughts on a sequential life testing approach using a Weibull model. In
Cottam, Harvey, Pape & Tait (eds.), Foresight and Precaution, ESREL 2000 Congress. 2: 16411647, Edinburgh,
Scotland: Balkema.
De Souza & Daniel I. 2001. Sequential Life Testing with
a Truncation Mechanism for an Underlying Weibull
Model. In Zio, Demichela & Piccinini (eds.), Towards a
Safer World, ESREL 2001 Conference, 1620 September.
3:15391546. Politecnico Di Torino. Italy.
De Souza & Daniel I. 2004. Sequential Life-Testing with
Truncation Mechanisms for Underlying Three-Parameter
Weibull and Inverse Weibull Models.,In Raj B.K. Rao,
B.E. Jones & R.I. Grosvenor Eds.; COMADEM Conference, Cambridge, U.K., August 2004260271, Comadem
International, Birmingham, U.K.
De Souza & Daniel I. 2005. A Maximum Likelihood
Approach Applied to an Accelerated Life Testing with
an Underlying Three-Parameter Inverse Weibull Model
In: Raj B.K. Rao & David U Mba Eds. COMADEM
2005 Condition Monitoring and Diagnostic Engineering Management, University Press, 2005. v.01. p.63 72.
Cranfield, Bedfordshire, UK.
De Souza, Daniel I. & Addad, Assed N. 2007. Sequential
Life-Testing with an Underlying Three-Parameter Inverse
Weibull Model A Maximum Likelihood Approach
In: IIE Annual Conference and Exposition. Nashville,
TN: The Institute of Industrial Engineering, 2007. V.01.
pp. 907 912. USA.
De Souza, Daniel I. & Lamberson, Leonard R. 1995.
Bayesian Weibull Reliability Estimation, IIE Transactions
27 (3), 311320.
Erto & Pasquale. 1982. New Practical Bayes Estimators for
the 2-Parameter Weibull Distribution, IEEE Transactions
on Reliability, R-31, (2), 194197.
Feller & Robert L. 1994. Accelerated Aging, Photochemical
and Thermal Aspects. The Getty Conservation Institute,
Eds.; Printer: Edwards Bross., Ann Harbor, Michigan.
Harter, H. et al. 1965, Maximum Likelihood Estimation of
the Parameters of Gamma and Weibull Populations from
Complete and from Censored Samples, Technometrics,
No 7, pp. 639643; erratum, 15 (1973), pp. 431.
Kapur, K. & Lamberson, L.R. 1977. Reliability in Engineering Design, John Willey & Sons, Inc., New York.
Mood, A.M. & Graybill, F.A. 1963. Introduction to the Theory of Statistics. Second Edition, McGraw-Hill, New
York.

Murthy, D.N.P., Xie, M. & Hang, R. 2004. Weibull Models. In

Wiley Series in Probability and Statistics, John Wiley &
Sons, Inc., New Jersey.
Zanakis, S.H. & Kyparisis, J. 1986. A Review of Maximum
Likelihood Estimation Methods for the Three Parameter
Weibull Distribution. Journal of Statistical Computation
and Simulation, 25, 5373.

APPENDIX 1. DETERMINING AN INITIAL

ESTIMATE TO THE MINIMUM LIFE
The pdf of t1 will be given by:
f (t1 ) = n [1 F(t1 )]n1 f (t1 ). Since
F(t1 ) = 1R(t1 ), we will have
f (t1 ) = n [R(t1 )]n1 f (t1 )
For the three-parameter Inverse Weibull sampling
distribution, we will have:


+1 
 n

1 exp
f (t1 ) =
t
t
The expected value of x1 is given by:

E(t1 ) =



+1

n
t

t


1 exp
Letting U =


du =



 n
dt

, we will have:

+1

dt; t =

+
U1/

When t , U 0; When t , U . Then:


E (t1 ) =

!n


n U1/ + 1 eU du


E (t1 ) = n

U1/ 1 eU

!n

du


+ n

1 eU

!n

du

The above integrals have to be solved by using a

numerical integration procedure, such as Simpsons

926

1/3 rule. Remembering that Simpsons 1/3 rule is

given by:
b


E (t1 ) = n

g
(f1 + 4f2 + 2f3 + + 4fk + fk+1 )
3

f (x)dx =

Using Equations A and B, we will have:


+ n

error
Making the error = 0; and with i = 1, 2, . . ., k + 1,
we will have:
U

n
0

1/

1e

 k+1
 

!
U n

1/
Ui

1e

(1, 2 or 4)

1 eU

!n

du = n j

1e

Ui n

(1, 2 or 4)

du.

Finally:

g
3

 k+1
 

!n

(A)
n

1 eU

i=1

i=1

du

1
k+1

g 
1/  
n
2
Ui
1e
E (t1 ) = n
or
Ui

3
i=1

4
k+1


 
n
g
1 eUi (1, 2 or 4) . (27)
+ n j
3

g
du = n
3

Ui n

!n

U1/ 1 eU


(B)

i=1

927

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Non-homogeneous Poisson Process (NHPP), stochastic model

applied to evaluate the economic impact of the failure in the Life Cycle
Cost Analysis (LCCA)
Carlos Parra Mrquez, Adolfo Crespo Mrquez, Pedro Moreu de Len, Juan Gmez Fernndez &
Vicente Gonzlez Daz
Department of Industrial Management School of Engineering, University of Seville, Spain

ABSTRACT: This paper aims to explore different aspects related with the failure costs (non reliability costs)
within the Life Cycle Cost Analysis (LCCA) of a production asset. Life cycle costing is a well-established
method used to evaluate alternative asset options. This methodology takes into account all costs arising during
the life cycle of the asset. These costs can be classified as the capital expenditure (CAPEX) incurred when
the asset is purchased and the operating expenditure (OPEX) incurred throughout the assets life. In this paper
we explore different aspects related with the failure costs within the life cycle cost analysis, and we describe
the most important aspects of the stochastic model called: Non-homogeneous Poisson Process (NHPP). This
model will be used to estimate the frequency failures and the impact that could cause the diverse failures in the
total costs of a production asset. The paper also contains a case study where we applied the above mentioned
concepts. Finally, the model presented provides maintenance managers with a decision tool that optimizes the
life cycle cost analysis of an asset and will increase the efficiency of the decision-making process related with
the control of failures.
Keywords: Asset; Failures; Life Cycle Cost Analysis (LCCA); Non-homogeneous Poisson Process (NHPP);
Maintenance; Reliability; Repairable Systems

INTRODUCTION

With the purpose of optimizing the costs and to

improve the profitability of the productive processes,
the denominated organizations of World Class category (Mackenzie, 1997), dedicate enormous efforts
to visualize, analyze, implant and execute strategies
for the solution of problems, that involve decisions
in high impact areas: security, environment, production goals, products quality, operation costs and
maintenance. In the last years, specialists in the
areas of value engineering and operations direction
have improved the quantification process of the costs,
including the use of techniques that quantify the Reliability factor and the impact of the failure events
on the total costs of a production system along their
life cycle (Woodhouse, 1993). These improvements
have allowed diminishing the uncertainty in the process of decisions making in vital importance areas
such as: design, development, maintenance, substitution and acquisition of production assets. It is
important to clear up that, in this whole process,

many decisions and actions exist, technical as much

as not technical, that should be adopted through the
whole use period of an industrial asset. Product support and maintenance needs of systems are more
or less decided during the design and manufacturing phase (Markeset and Kumar, 2001). Outlines
that most of these actions, particularly those that
correspond to the design phase of the production system, have a high impact in the total life cycle of
the asset, being of particular interest, those decisions related with the improvement process of the
Reliability factor (quality of the design, used technology, technical complexity, frequency of failures,
costs of preventive/corrective maintenance, maintainability levels and accessibility), since these aspects,
have a great influence on the total cost of the assets
life cycle, and they influence in great measure on
the possible expectations to extend the useful life
of the production systems to reasonable costs (see,
e.g. Blanchard, 2001; Blanchard and Fabrycky, 1998;
Goffin, 2000; Markeset and Kumar, 2001; Smith and
Knezevic, 1996 and Woodward, 1997).

929

ANTECEDENTS OF THE LCCA

TECHNIQUES

In the last years, the investigation area related with the

Life cycle Costs Analysis, has continued its development, as much in the academic level as to the industrial
level. It is important to mention the existence of other
methodologies that have emerged in the area of LCCA,
such as: Life cycle Costs Analysis and Environmental Impact, Total Costs Analysis of Production Assets,
among other (Durairaj and Ong, 2002). These methodologies have their particular characteristics, although
regarding the estimation process of the costs for failure
events impact, they propose Reliability analysis usually based on rate of constant failures. The antecedents
of the LCCA are shown next (Kirt and Dellisola,
1996):
1930, one of the first records that are known of
the LCCA techniques is found in the book named
Principles of Engineering Economics by Eugene L.
Grant.
1933, the first reference of Life cycle Analysis by
the Government of the United States shows up carried out by part of the federal department: General
Accounting Office (GAO), which is related to the
purchase of a series of tractors.
1950, Lawrence D. Miles originated the concept
of Value Engineering at General Electric, incorporating aspects related with the techniques of
LCCA.
1960, Stone (1975) began to work in England,
giving as a result in the decade of the 70s the publication of two of the biggest texts developed in
Europe in relation to costs engineering.
1960, the Logistics Management Institute of the
United States developed an investigation in the area
of Obsolescence Engineering for the Ministry of
Defense. The final result of this investigation was
the publication of the first Life cycle Cost Manual
in the year of 1970.
1972, the Ministry of Defense of the United States,
promoted the development of a group of Manuals
with the purpose of applying the LCCA Methodology, in all the Logistics areas.
1974, the Department of Energy of the United
States, decided to develop its expansion and energy
consumption plans supported by the analysis of Life
cycle.
1975, the Federal Department of Supplies and Services of the United States developed a Logistics and
Acquisition technique based on the LCCA.
1979, the Department of Energy introduced a proposal (44 FR 25366, April 30 1979) which intended
that evaluations of LCCA were included in all
the new constructions and mayor modifications in
government facilities.

1980, the American Society for Testing and

Materials (ASTM) developed a series of standards
and database oriented to ease the search of necessary
information for the application of the LCCA.
1992, two investigators of the University of
Virginia, Wolter Fabrycky and B.S. Blanchard,
developed a model of LCCAsee details in
(Fabrycky and Blanchard, 1993), in which they
include a structured process to calculate the costs of
Non Reliability starting from the estimate of constant values of failures per year (constant rate of
failures).
1994, Woodward (1997), from the School of Business of the University of Staffordshire (England,
Great Britain), developed an investigation line in
which included basic aspects of analysis of the Reliability factor and their impact on the Costs of Life
cycle.
1998, David Willians and Robert Scott of the consulting firm RM-Reliability Group, developed a
model of LCCA based on the Weibull Distribution
to estimate the frequency of failures and the impact
of the Reliability Costs, see details in (Zohrul Kabil,
1987, Ebeling, 1997 and Willians and Scott, 2000).
1999, the Woodhouse Partnership consulting group
participated in the European Project EUREKA,
specifically inside the line of investigation
denominated MACRO (Maintenance Cost/Risk
Optimization Project) and they developed an LCCA
commercial software of denominated APT Lifespan, see details in (Roca, 1987, Barlow, Clarotti and
Spizzichino, 1993, Woodhouse, 1991 and Riddell
and Jennings, 2001).
2001, the Woodhouse Partnership consulting firm
and the Venezuelan Oil Technological Institute
(INTEVEP), put on test this model, evaluating the
Total Costs of Life cycle for 56 gas compression
systems, used for the extraction of heavy oil in the
San Tom District (Venezuela), see details in (Parra
and Omaa, 2003).
3

BASIC ASPECTS OF THE LCCA

To evaluate the costs associated to the life cycle of

a production system, a collection of procedures that
group together exists in the denominated: Techniques
of Life cycle Costs Analysis. The early implementation of the costs analysis techniques allows to evaluate
in advance the potential design problems and to quantify the potential impact in the costs along the life cycle
of the industrial assets (Durairaj and Ong, 2002). Next,
some basic definitions of Life cycle Cost Analysis are
presented:
Kirt and Dellisolla (1996) defines the LCCA as
a technique of economic calculation that allows
to optimize the making of decisions associated to

930

the design processes, selection, development and

substitution of the assets that conform a production system. It intends to evaluate in a quantitative
way all the costs associated to the economic period
of expected useful life, expressed in yearly equivalent monetary units (Dollars/year, Euros/year,
Pesos/year).
Woodhouse (1991) defines the LCCA like a systematic process of technical-economical evaluation,
applied in the selection and replacement process of
production systems that allows to consider in simultaneous way economic and Reliability aspects, with
the purpose of quantifying the real impact of all the
costs along the life cycle of the assets ($/year), and
in this way, be able to select the asset that contributes
the largest benefits to the productive system.
The great quantity of variables that must be managed when estimating the real costs of an asset along
its useful life generates a scenario of high uncertainty
(Durairaj and Ong, 2002). The combination among
inflation, rise/decrease of the costs, reduction/increase
of the purchasing power, budget limitations, increase
of the competition and other similar characteristics,
has generated a restlessness and interest about the total
cost of the assets. Often the total cost of the production
system is not visible, in particular those costs associated with: operation, maintenance, installation tests,
personnels training, among others.
Additionally, the dynamics of the economic scenario generate problems related to the real determination of the assets cost. Some of them are (Fabrycky,
1997):
The factors of costs are usually applied incorrectly.
The individual costs are inadequately identified
and, many times, they are included in the wrong
category: the variable costs are treated as fixed
(and vice versa); the indirect costs are treated as
direct, etc.
The countable procedures do not always allow a
realistic and timely evaluation of the total cost.
Besides, it is often difficult (if not impossible) to
determine the costs, according to a functional base.
Many times the budgetary practices are inflexible
with regard to the change of funds from a category
to another, or, from one year to another.
To avoid the uncertainty in the costs analysis, the
studies of economic viability should approach all the
aspects of the life cycle cost. The tendency to the variability of the main economic factors, together with the
additional problems already enunciated, have driven
to erroneous estimates, causing designs and developments of production systems that are not suitable from
the point of view of cost-benefit (Fabrycky, 1997). It
can be anticipated that these conditions will worsen,
unless the design engineers assume a bigger grade of

consideration of the costs. Inside the dynamic process

of change, the acquisition costs associated to the new
systems are not the only ones to increase, but rather
the operation and maintenance costs of the systems
already in use also do it in a quick way. This is due
mainly to a combination of such factors as (Fabrycky,
1997):
Inaccuracies in the estimates, predictions and
forecasts of the events of failures (Reliability),
ignorance of the probability of occurrence of the different failure events inside the production systems
in evaluation.
Ignorance of the deterioration processes behavior.
Lack of forecast in the maintenance processes and
ignorance of the modern techniques of maintenance
management.
Engineering changes during the design and development.
Changes in the own construction of the system.
Changes in the expected production patterns.
Changes during the acquisition of system components.
Setbacks and unexpected problems.

3.1

Characteristics of the costs

in a production asset

The cost of a life cycle is determined identifying the

applicable functions in each one of its phases, calculating the cost of these functions and applying the
appropriate costs during the whole extension of the
life cycle. So that it is complete, the cost of the life
cycle should include all the costs of design, fabrication and production (Ahmed, 1995). In the following
paragraphs the characteristics of the costs in the different phases of an assets life cycle are summarized
(Levi and Sarnat, 1990):
Investigation, design and development costs: initial
planning, market analysis, product investigation,
design and engineering requirements, etc.
Production, acquisition and construction costs:
industrial engineering and analysis of operations,
production (manufacturing, assembly and tests),
construction of facilities, process development,
production operations, quality control and initial
requirements of logistics support.
Operation and support costs: operations inputs of
the production system, planned maintenance, corrective maintenance (it depends on the Reliability
Factor) and costs of logistical support during the
systems life cycle.
Remove and elimination costs: elimination of non
repairable elements along the life cycle, retirement
of the system and recycling material.

931

From the financial point of view, the costs generated

along the life cycle of the asset are classified in two
types of costs:
CAPEX: Capital costs (design, development, acquisition, installation, staff training, manuals, documentation, tools and facilities for maintenance,
replacement parts for assurance, withdrawal).
OPEX: Operational costs: (manpower, operations,
planned maintenance, storage, recruiting and
corrective maintenancepenalizations for failure
events/low Reliability).
3.2

Costs of Non
Reliability

Costs for
penalization

Figure 1.

Costs for
corrective
maintenance

Economic impact of the Reliability.

Impact of the reliability in the LCCA

Woodhouse (1991) outlines that to be able to design

an efficient and competitive productive system in the
modern industrial environment, it is necessary to evaluate and to quantify in a detailed way the following two
aspects:
Costs: aspect that is related with all the costs
associated to the expected total life cycle of
the production system. Including: design costs,
production, logistics, development, construction,
operation, preventive/corrective maintenance, withdrawal.
Reliability: factor that allows to predict the form in
which the production processes can lose their operational continuity due to events of accidental failures
and to evaluate the impact in the costs that the failures cause in security, environment, operations and
production.
The key aspect of the term Reliability is related
to the operational continuity. In other words, we can
affirm that a production system is Reliable when it is
able to accomplish its function in a secure and efficient
way along its life cycle. Now, when the production
process begins to be affected by a great quantity of
accidental failure events(low Reliability), this scenario causes high costs, associated mainly with the
recovery of the function (direct costs) and with growing impact in the production process (penalization
costs). See Figure 1:
The totals costs of Non Reliability are described
next (Barlow, Clarotti and Spizzichino, 1993, Ruff and
Paasch, 1993 and Woodhouse, 1993):
Costs for penalization:
Downtime: opportunity losses/deferred production, production losses (unavailability), operational losses, impact in the quality, impact in
security and environment.

Materials and replacement parts: direct costs

related with the consumable parts and the replacements used in the event of an unplanned action.
The impact in the costs that an asset of low Reliability generates is associated directly with the behavior
of the following index:
(t) = expected number of failures in a time
interval [0, t]
According to Woodhouse (1991), the increase of
the costs is caused in its great majority, for the lack of
forecast in case of unexpected failures appearances,
scenario basically provoked by ignorance and lack of
analysis in the design phase of the aspects related with
the Reliability. This situation brings as a result an
increase in the operation costs (costs that were not
considered in a beginning) affecting in this way the
profitability of the production process.
It is important to mention that the results obtained
from the LCCA, reach their maximum effectiveness
during the phases of: initial development, visualization, and conceptual, basic and details engineering.
Once the design has been completed, it is substantially
difficult to modify the economic results. Also, the economic considerations related with the life cycle should
be specifically outlined during the phases previously
mentioned, if it is that one wants to totally exploit the
possibilities of an effective economic engineering. It
is necessary to keep in mind that almost two thirds of
the life cycle cost of an asset or system are already
determined in the preliminary conceptual and design
phase (7085% of value creation and costs reduction
opportunities), according to (Dowlatshahi, 1992).
4

Costs for corrective maintenance:

Manpower: direct costs related with the manpower (own or hired) in the event of a non planned
action.

STOCHASTIC MODELS CONSIDERED

FOR THE ANALYSIS OF THE RELIABILITY

The following sections discuss the different stochastic

models considered for the analysis of the frequency of
failures in repairable units and systems.

932

4.3 Generalized renewal process (GRP)

A repairable system may end up in one of the five
possible states after a repair:
a.
b.
c.
d.
e.
Figure 2.

4.1

Basic notation for a stochastic point process.

Ordinary Renewal Process (ORP)

This model assumes that, following a repair, the

unit returns to an as good as new (AGAN) condition. In this process, the interarrival times, xi,
between successive failures (see Figure 2) are considered independently and identically distributed random
variables. It is a generalization of a Homogeneous
Poisson Process (HPP). This model represents an ideal
situation; it is only appropriate for replaceable items
and hence has very limited applications in the analysis
of repairable components and systems. Variations of
the ORP can also be defined. The modified renewal
process, where the first interarrival time differs from
the others, and the superimposed renewal process
(union of many independent ORPs) are examples of
these possible variations (Ascher and Feingold, 1984).
4.2

As good as new
As bad as old
Better than old, but worse than new
Better than new
Worse than old

The two models described before, ordinary renewal

process and NHPP, account for the first two states
respectively. However, the last three repair states have
received less attention since they involve more complex mathematical models. Kijima and Sumita (1987)
proposed a probabilistic model for all the after-repair
states called Generalized Renewal Process (GRP).
According to this approach, the ordinary renewal process and the NHPP are considered specific cases of
the generalized model. The GRP theory of repairable
items introduces the concept of virtual age (An). This
value represents the calculated age of the element
immediately after the nth repair occurs. For An = y the
system has a time to the (n + 1)th failure, xn+1 , which
is distributed according to the following cumulative
distribution function (cdf ):
F(x |A n = y) =

t2
mean =

(t)dt

(1)

t1

where (t) is the rate of occurrence of failures

(ROCOF) defined as the inverse of the expected interarrival times, 1/E[xi] (Ascher and Feinfold, 1984 and
Crow, 1974 ).

(2)

where F(x) is the cdf of the time to the first failure

(TTFF) distribution of a new component or system.
The summation:

Non-Homogeneous Poisson Process (NHPP)

This model is also called minimal repair and it

assumes that the unit returns to an as bad as old
(ABAO) condition after a repair. So that, after the
restoration the item is assumed to be operative but
as old as it was before the failure. The NHPP differs
from the HPP in that the rate of occurrence of failures
varies with time rather the being constant (Ascher and
Feingold, 1984). Unlike the previous model, in this
process the interarrival times are neither independent
nor identically distributed. The NHPP is a stochastic
point process in which the probability of occurrence
of n failures in any interval [t1, t2] has a Poisson
distribution with:

F(x + y) F(y)
1 F(y)

Sn =

n


xi

(3)

i=1

with S = 0, is called the real age of the element. The

model assumes that the nth repair only compensates
for the damage accumulated during the time between
the (n 1)th and the nth failure. With this assumption,
the virtual age of the component or system after the
nth repair is:
An = An1 + qxn = qSn

(4)

where q is the repair effectiveness (or rejuvenation)

parameter and A0 = 0. According to this model, the
result of assuming a value of q = 0 leads to an ordinary
renewal process (as good as new), while the assumption of q = 1 corresponds to a non-homogeneous
Poisson process (as bad as old). The values of q that
fall in the interval 0 < q < 1 represent the after-repair
states in which the condition of the element is better
than old but worse than new, whereas the cases where
q > 1 correspond to a condition worse than old. Similarly, cases with q < 0 would suggest a component

933

or system restored to a state better than new. Therefore, physically speaking, q can be seen as an index
for representing the effectiveness and quality of repairs
(Yaez et al., 2002). Even though the q value of the
GRP model constitutes a realistic approach to simulate
the quality of maintenance, it is important to point out
that the model assumes an identical q for every repair
in the item life. A constant q may not be the case for
some equipment and maintenance process, but it is a
reasonable approach for most repairable components
and systems.
The three models described above have advantages
and limitations. In general, the more realistic is the
model, the more complex are the mathematical expression involved. The NHPP model has been proved to
provide good results even for realistic situations with
better-than-old but worse-than-new repairs (Yaez
et al., 2002). Based on this, and given its conservative
nature and manageable mathematical expressions, the
NHPP was selected for this particular work. The specific analytical modeling is discussed in the following
section.

4.4

Non-homogeneous Poisson process

analytical modeling

The NHPP is a stochastic point process in which the

probability of occurrence of n failures in any interval
[t1, t2] has a Poisson distribution with the mean:

=

t2

(t)dt

(5)

t1

where (t) is the rate of occurrence of failures

(ROCOF).
Therefore, according to the Poisson process:
Pr[N (t2 ) N (t1 ) = n]

n
 

t2
t
exp t12 (t)dt
t1 (t)dt
=
n!

(t) =

(t)dt

Conditional probability of occurrence of failure.

Law Model (Ascher and Feinfold, 1984 and Crow,

1974 ):

(t) =

 1
t

(6)

(7)

F(t) F(t1 )
R(t1 )
1 R(t) 1 + R(t)
R(t1 )

=1

R(t)
R(t1 )

(9)

where F() and R() are the probability of component failure and the reliability at the respective times.
Assuming a Weibull distribution, Eq. (9) yields:

One of the most common forms of ROCOF used in

reliability analysis of repairable systems is the Power

(8)

This form comes from the assumption that the interarrival times between successive failures follow a
conditional Weibull probability density function, with
parameters and . The Weibull distribution is typically used in maintenance area due to its flexibility
and applicability to various failure processes, however,
solutions to Gamma and Log-normal distributions are
also possible. This model implies that the arrival of
the ith failure is conditional on the cumulative operating time up to the (i 1)th failure. Figure 3 shows a
schematic of this conditionality (Yaez et al., 2002).
This conditionality also arises from the fact that the
system retains the condition of as bad as old after
the (i 1)th repair. Thus, the repair process does not
restore any added life to the component or system.
In order to obtain the maximum likelihood (ML)
estimators of the parameters of the power law model,
consider the following definition of conditional
probability:
P(T t |T > t1 ) =

where n = 0, 1, 2, . . . are the total expected number of

failures in the time interval [t1, t2]. The total expected
number of failures is given by the cumulative intensity
function:


Figure 3.

F(ti ) = 1 exp

934

ti1

 

ti

(10)

Therefore, the conditional Weibull density function is:



 

 
ti1
ti
ti 1
f (ti ) =
. exp

(11)

For the case of the NHPP, different expressions for
the likelihood function may be obtained. We will use
expression based on estimation at a time t after the
occurrence of the last failure and before the occurrence
of the next failure, see details on these expressions in
(Modarres et al., 1999).
4.4.1 Time terminated NHPP maximum
likelihood estimators
In the case of time terminated repairable components, the maximum likelihood function L can be
expressed as:
L=

f (ti ) = f (t1 )

i=1

f (ti )R(tn |t )

(12)

i=2

Therefore:
  
 

t1 1
t1
L=
exp

 


n
n1  t1 1

i=2
 n 

 


ti1
ti
exp

i=2

 
 

t
tn

exp

(13)

Again, the ML estimators for the parameters are

calculated. The results are (Ascher and Feinfold, 1984
and Crow, 1974):
=
=

tn

(14)

n

i=1

n
 
ln ttni

(15)

where ti is the time at which the ith failure occurs, tn

is the total time where the last failure occurred, and
n is the total number of failures. The total expected
number of failures in the time interval [tn, tn + s]
by the Weibull cumulative intensity function is
(Modarres et al., 1999):

(tn , tn+s ) =


1 
(tn + ts ) (tn )

(16)

where ts is the time after the last failure occurred in

the one which needs to be considered the number of
failures and
tn =

n


ti

(17)

i=1

NHPP MODEL PROPOSED

FOR THE EVALUATION OF THE COSTS
PER FAILURE

Asiedu and Gu (1998) have published a state of the art

review on LCCA. Most of the methodologies proposed
in the last years, include basic analysis that allow to
quantify the economic impact that the failures inside a
production system generate. In relation to the quantification of the costs for non Reliability in the LCCA, we
recommend to use NHPP model. This model proposes
to evaluate the impact of the main failures on the costs
structure of a production system, starting from a simple process, which is summarized next: first, the more
important types of failures are determined, then, it is
assigned to each failure type a constant value of occurrence frequency per year (this value will not change
along the expected useful life), later on, the impact
in costs per year is estimated, generated by the failures to the production, operations, environment and
security, and finally, the total impact in costs of the
failures for the years of expected useful life is considered in present value to a specific discount rate. Next,
are detailed the steps to estimate the costs for failures
according to NHPP model:
1. Identify for each alternative to evaluate the main
types of failures. This way for certain equipment
there will be f = 1. . . . . . . F types of failures.
2. Determine for the n (total of failures), the times
to failures tf . This information will be gathered by
the designer based on records of failures, databases
and/or experience of maintenance and operations
personnel.
3. Calculate the Costs for failures Cf ($/failure).
These costs include: costs of replacement parts,
manpower, penalization for production loss and
operational impact.
4. Define the expected frequency of failures per year
(tn , tn+s ). This frequency is assumed as a constant value per year for the expected cycle of useful
life. The (tn , tn+s ) is calculated starting from the
expression (16). This process is carried out starting from the times to failures registered tf by failure
type (step 2). The parameters and , are set starting from the following expressions (14) and (15).

935

In the expression (16), ts it will be a year (1 year) or

equivalent units (8760 hours, 365 days, 12 months,
etc.). This time ts represents the value for estimate
de frequency of failures per year.
5. Calculate the total costs per failures per year TCPf ,
generated by the different events of stops in the
production, operations, environment and security,
with the following expression:

TCPf =

F


 (tn , tn+s ) Cf

Define the types of failures (f ). Where f = 1. . .F

for F types of failures:
F = 1 types of failures
Calculate the Costs per failure Cf (these costs
include: costs of replacement parts, manpower,
penalization for production loss and operational
impact):

(18)

Cf = 5000

The obtained equivalent annual total cost, represents the probable value of money that will be
needed every year to pay the problems of reliability caused by the event of Failure, during the years
of expected useful life.
6. Calculate the total costs per failures in present value
PTCPf . Given a yearly value TCPf , the quantity
of money in the present (today) that needs to be
saved, to be able to pay this annuity for the expected
number of years of useful life (T), for a discount rate
(i). The expression used to estimate the PTCPf is
shown next:
PTCPf = TCPf

(1 + i)T 1
i (1 + i)T

(19)

$
failure

Define the expected frequency of failures per year

(tn , tn+s ), use expression (16):
(tn , tn+s ) =

1
[(tn + ts ) (tn ) ]

Where:
n = 24 failures
n

ti = 5 + 7 + 3. . . . . . .4 + 7 + 4 = 117
tn =
i=1

months
ts = 12 months
tn+s = 119 months
The parameters and , are calculated from the
expressions (14) and (15):

Later on, to the costs calculated by non reliability,

the rest of the evaluated costs (investment, planned
maintenance, operations, etc.) are added, the total
cost is calculated in present value for the selected discount rate and the expected years of useful life and the
obtained result is compared with the total costs of the
other evaluated options.

= 6.829314945
= 1.11865901
The expected frequency of failures per year:
(tn , tn+s ) = 2.769896307

CASE STUDY

The following data of failures will be used for the three

models previously explained. This information was
gathered by records of failures of a Gas Compressor
from Venezuelan National Oil Company. In this equipment have occurred 24 events of failures in 10 years of
useful life. Next, the times to failures tf are presented
in months:
This model proposes to evaluate the impact of the
failures in the following way:
Table 1.

Times to failures.

failures
,
year

this frequency is assumed as a constant value per

year for the expected cycle of useful life.
Calculate the total Costs per failures per year TCPf ,
use expression (18):
TCPf = 2.769896307
= 13849.48154

936

$
failures
5000
year
failures
$
year

 Calculate the total cost per failure in present value

PTCPf , use expression (19), for a period T = 10
years and discount rate i = 10%:
PTCPf = 73734, 805$,
value that represents the quantity of money (today)
that the organization needs to be able to cover
the annual expenses projected by failures in the
next 10 years, with a discount factor of 10%. For
this example, the total expected number of failures in the time interval [tn, tn + s] is estimated
by the NHPP stochastic model (Weibull cumulative
intensity function), see Modarres et al., 1999.
6.1

Limitations of the model evaluated

The NHPP model has been proved to provide good

results even for realistic situations with better-thanold but worse-than-new repairs (Hurtado et al., 2005).
Based on this, and given its conservative nature and
manageable mathematical expressions, the NHPP was
selected for this particular work. The model described
above has advantages and limitations. In general, the
more realistic is the model, the more complex are the
mathematical expression involved. The main strengths
and weakness of this model are summarized next:
Strengths:
It is a useful and quite simple model to represent
equipment under aging (deterioration).
Involves relatively simple mathematical expressions.
It is a conservative approach and in most cases provides results very similar to those of more complex
models like GRP (Hurtado et al., 2005).
Weakness:
Is not adequate to simulate repair actions that restore
the unit to conditions better than new or worse
than old.
7

FUTURE DIRECTIONS

The specific orientation of this work toward the analysis of the Reliability factor and its impact in the costs, is
due to, that great part of the increment of the total costs
during the expected cycle of useful life of a production
system, is caused in its majority, for the lack of prevision in the face of unexpected appearance of failure
events, scenario basically provoked by ignorance and
by the absence of a technical evaluation in the design
phase of the aspects related with the Reliability. This
situation brings as a result an increment in the total
costs of operation (costs that were not considered in
the beginning) affecting in this way the profitability of
the production process.

In the process of analysis of the costs along the life

cycle of an asset, many decisions and actions exist
that should be taken, being of particular interest for
this work, those aspects related with the process of
improvement of the Reliability (quality of the design,
used technology, technical complexity, frequency of
failures, costs of preventive/corrective maintenance,
maintainability levels and accessibility), since these,
have a great impact on the total cost of the life cycle
of the asset, and they influence in great measure on
the possible expectations to extend the useful life of
the assets to reasonable costs. For these reasons, it
is of supreme importance inside the process to estimate the life cycle of the assets, to evaluate and to
analyze detailedly the aspects related with the failure
rate. According to Ascher (1984), the following points
should be considered in failure rate trend analyses:
Failure of a component may be partial, and repair
work done on a failed component may be imperfect.
Therefore, the time periods between successive failures are not necessarily independent. This is a major
source of trend in the failure rate.
Imperfect repairs performed following failures do
not renew the system, i.e., the component will
not be as good as new; only then can the statistical inference methods using a Rate Of Occurrence
Of Failures (ROCOF) assumption be used.
Repairs made by adjusting, lubricating, or otherwise treating component parts that are wearing out
provide only a small additional capability for further operation, and do not renew the component or
system. These types of repair may result in a trend
of a increasing ROCOF.
A component may fail more frequently due to aging
and wearing out.
It is important to mention that inside the LCCA
techniques a potential area of optimization related
with the evaluation of the Reliability impact exists.
In the near future the new proposals of evaluation of
the costs generated by aspects of low Reliability will
use advanced mathematical methods such as:
Stochastic methods see (Tejms, 1986, Karyagina
et al., 1998, Yaez et al., 2002, Hurtado et al.,
2005 and Vasiliy, 2007). Table 2 shows the stochastic processes used in reliability investigations of
repairable systems, with their possibilities and
limits (Modarres et al., 1999).
Advanced maintenance optimization using genetic
algorithms see (Martorell et al., 2000 and Martorell
et al., 2005).
Monte Carlo simulation techniques see (Barringer,
1997, Barringer and Webber, 1996, and Kaminskiy
and Krivtsov, 1998).
Advanced Reliability distribution analysis see
(Elsayed, 1982, Barlow, Clarotti and Spizzichino,

937

Table 2. Stochastic processes used in reliability analysis of

repairable systems.
Stochastic
process

Can be used

Renewal process Spare parts provisioning

in the case of
arbitrary failure rates
and negligible
replacement or repair
time (Poisson
process)
Alternating
One-item repairable
renewal
(renewable) structure
process
with arbitrary failure
and repair rates
Markov process Systems of arbitrary
(MP)
structure whose
elements have constant
failure and repair rates
during the stay
time (sojourn time) in
every state (not
necessarily at a state
change, e.g. because
of load sharing)
Semi-Markov
Some systems whose
process (SMP)
elements have constant
or Erlangian failure
rates (Erlang
distributed failurefree times) and
arbitrary repair rates
SemiSystems with only one
Regenerative
repair crew, arbitrary
process
structure, and whose
elements have constant
failure rates and
arbitrary repair rates
Nonregenerative Systems of arbitrary
process
structure whose
elements have arbitrary
failure and repair
rates

Background/
Difficulty
Renewal
theory/
Medium

Renewal
theory/
Medium
Differential
equations
or integral
equations/
Low

Integral
equations/
Medium

Integral
equations/
High

Partial diff.
eq.; case by
base sol./
High to
very high

1993, Ireson, et al., 1996, Elsayed, 1996, Scarf,

1997, Ebeling, 1997 and Dhillon, 1999).
Markov simulation methods see (Roca, 1987,
Kijima and Sumita, 1987 Kijima, 1997 and BlochMercier, 2000).
These methods will have their particular characteristics and their main objective will be to diminish the
uncertainty inside the estimation process of the total
costs of an asset along the expected useful life cycle.
Finally, it is not feasible to develop a unique LCCA
model, which suits all the requirements. However, it is
possible to develop more elaborate models to address
specific needs such as a Reliability cost-effective asset
development.

REFERENCES
Ahmed, N.U. 1995. A design and implementation model
for life cycle cost management system, Information and
Management, 28, pp. 261269.
Asiedu, Y. and Gu, P. 1998. Product lifecycle cost analysis:
state of art review, International Journal of Production
Research, Vol. 36 No. 4, pp. 883908.
Ascher, H. and Feingold, H. Repairable System Reliability:
Modeling, Inference, Misconceptions and their Causes,
New York, Marcel Dekker, 1984.
Barlow, R.E., Clarotti, C.A. and Spizzichino, F. 1993. Reliability and Decision Making, Chapman & Hall, London.
Barringer, H. Paul and David P. Weber. 1996. Life
Cycle Cost Tutorial, Fifth International Conference
on Process Plant Reliability, Gulf Publishing Company,
Houston, TX.
Barringer, H. Paul and David P. Weber. 1997. Life Cycle
Cost & Reliability for Process Equipment, 8th Annual
ENERGY WEEK Conference & Exhibition, George R.
Brown Convention Center, Houston, Texas, Organized by
American Petroleum Institute.
Barroeta, C. 2005. Risk and economic estimation of inspection policy for periodically tested repairable components,
Thesis for the Master of Science, University of Maryland,
Faculty of Graduate School, College Park,Cod.Umi-umd2712, pp. 77, August, Maryland.
Blanchard, B.S. 2001. Maintenance and support: a
critical element in the system life cycle, Proceedings of
the International Conference of Maintenance Societies,
paper 003, May, Melbourne.
Blanchard, B.S. and Fabrycky, W.J. 1998. Systems Engineering and Analysis, 3rd ed., Prentice-Hall, Upper Saddle
River, NJ.
Bloch-Mercier, S. 2000. Stationary availability of a semiMarkov system with random maintenance, Applied
Stochastic Models in Business and Industry, 16,
pp. 219234.
Crow, LH. 1974. Reliability analysis for complex repairable
systems, Reliability and biometry, Proschan F, Serfling
RJ, eds., SIAM, Philadelphia, pp. 379410.
Dhillon, B.S. 1989. Life Cycle Costing: Techniques, Models
and Applications, Gordon and Breach Science Publishers,
New York.
Dhillon, B.S. 1999. Engineering Maintainability: How to
Design for Reliability and Easy Maintenance, Gulf,
Houston, TX.
Dowlatshahi, S. 1992. Product design in a concurrent engineering environment: an optimization approach, Journal
of Production Research, Vol. 30 (8), pp. 18031818.
Durairaj, S. and Ong, S. 2002. Evaluation of Life Cycle
Cost Analysis Methodologies, Corporate Environmental
Strategy, Vol. 9, No. 1, pp. 3039.
DOD Guide LCC-1, DOD Guide LCC-2, DOD Guide LCC3. 1998. Life Cycle Costing Procurement Guide, Life
Cycle Costing Guide for System Acquisitions, Life Cycle
Costing Guide for System Acquisitions, Department of
Defense, Washington, D.C.
Ebeling, C. 1997. Reliability and Maintainability Engineering, McGraw Hill Companies, USA.
Elsayed, E.A. 1982. Reliability Analysis of a container
spreader, Microlelectronics and Reliability, Vol. 22,
No. 4, pp. 723734.

938

Elsayed, E.A. 1996. Reliability Engineering, Addison Wesley

Longman INC, New York.
Fabrycky, W.J. 1997. Anlisis del Coste de Ciclo de Vida de
los Sistemas, ISDEFE, Ingeniera de Sistemas, Madrid,
Espaa.
Fabrycky, W.J. and Blanchard, B.S. 1993. Life Cycle Costing
and Economic Analysis, Prentice Hall, Inc, Englewwod
Cliff, New Jersey.
Goffin, K. 2000. Design for supportability: essential
component of new product development, ResearchTechnology Management, Vol. 43, No. 2, pp. 407.
Hurtado, J.L., Joglar, F. and Modarres, M. 2005. Generalized Renewal Process: Models, Parameter Estimation and
Applications to Maintenance Problems, International
Journal on Performability Engineering, Vol. 1, No. 1,
paper 3, pp. 3750.
Ireson, W. Grant, Clyde F. Coombs Jr., Richard Y. Moss.
1996. Handbook of Reliability Engineering and Management, 2nd edition, McGraw-Hill, New York.
Kaminskiy M, Krivtsov V. 1998. A Monte Carlo
approach to repairable system reliability analysis, Probabilistic safety assessment and management, Springer,
pp. 10631068.
Karyagina, M., Wong, W., Vlacic, L. 1998. Life cycle
cost modelling using marked point processes, Reliability
Engineering & System Safety, Vol. 59, pp. 291298.
Kececioglu, D. 1991. Reliability and Life Testing Handbook, Prentice Hall, Inc, Englewood Cliff, New Jersey.
Kijima, M., Sumita, N. 1987. A useful generalization
of renewal theory: counting process governed by nonnegative Markovian increments, Journal Appl. Prob.,
Vol. 23, pp. 7188.
Kijima, M. 1977. Markov processes for stochastic modeling, Chapman & Hall, London.
Kirk, S. and Dellisola, A. 1996. Life Cycle Costing for Design
Professionals, McGraw Hill, New York, pp. 657.
Levy, H. and Sarnat, M. 1990. Capital Investment and Financial Decisions, 4th Edition, Prentice Hall, New York.
Life Cycle Costing Workbook: A guide for implementation of Life on Life Cycle Costing in the Federal Supply
Services. 1989. U.S. General Services Administration,
Washington.
Life Cycle Analysis as an Aid Decision Making. 1985.
Building Information Circular, Department of Energy,
Office of Facilities Engineering and Property Management, Washington.
Mackenzie, J. 1997. Turn your companys strategy into
reality, Manufacturing Management, January, pp. 68.
Markeset, T. and Kumar, U. 2001. R&M and risk analysis tools in product design to reduce life-cycle cost
and improve product attractiveness, Proceedings of
the Annual Reliability and Maintainability Symposium,
2225 January, Philadelphia, pp. 116122.
Markeset, T. and Kumar, U. 2003. Design and development of product support and maintenance concepts for
industrial systems, Journal of Quality in Maintenance
Engineering, Vol. 9, No. 4, pp. 376392.
Martorell, S., Carlos, S., Sanchez, A. and Serradell, V.
2000. Constrained optimization of test intervals using
a steady-state genetic algorithm. Reliability Engineering & System Safety 67, 215232.

Martorell, S., Villanueva, J.F., Nebot, Y., Carlos, S.,

Snchez, A., Pitarch, J.L. and Serradell, V. 2005.
RAMS+C informed decision-making with application
to multi-objective optimization of technical specifications
and maintenance using genetic algorithms. Reliability
Engineering & System Safety 87, 6575.
Modarres, M., Kaminskiy, M. and Krivtsov V. 1999. Reliability engineering and risk analysis. Marcel Dekker Inc.,
New York.
Nachlas, J. 1995. Fiabilidad, ISDEFE, Ingeniera de Sistemas, Madrid, Espaa.
Navas, J. 1997. Ingeniera de Mantenimiento, Universidad
de los Andes, Mrida, Venezuela.
Parra, C. 2002. Evaluacin de la Influencia de la Confiabilidad en el Ciclo de Vida de 16 Motocompresores del
Distrito Maturn, Informe Tcnico INT-9680, PDVSA
INTEVEP, Venezuela.
Parra, C. y Omaa C. 2003. Anlisis determinstico del
Ciclo de Vida y evaluacin del factor Confiabilidad en
Motocompresores de gas para extraccin de petrleo,
Congreso Internacional de Mantenimiento, Colombia,
Bogot, Colombia.
Riddell, H., Jennings, A. 2001. Asset Investment & Life
Cycle Costing, The Woodhouse Partnership, Technical
paper, London.
Roca, J.L. 1987. An approach in the life cycle costing
of an integrated logistic support, Microelectronics and
Reliability, Vol. 27, No. 1, pp. 2527.
Ruff, D.N., and Paasch, R.K. 1993. Consideration of failure
diagnosis in conceptual design of mechanical systems,
Design Theory and Methodology, ASME, New York,
pp. 175187.
Scarf, P.A. 1997. On the application of mathematical models in maintenance, European Journal of Operational
Research, Vol. 99, No. 3, pp. 493506.
Smith, C. and Knezevic, J. 1996. Achieving quality through
supportability: part 1: concepts and principles, Journal
of Quality in Maintenance Engineering, Vol. 2, No. 2,
pp. 219.
Tejms, H.C. 1986. Stochastic Modelling and Analysis, Wiley
and Sons, New York, NY.
Vasiliy, V. 2007. Recent advances in theory and applications
of stochastic point process model in reliability engineering, Reliability Engineering & System Safety, Vol. 92,
No. 5, pp. 549551.
Willians, D., Scott R. 2000. Reliability and Life Cycle
Costs, RM-Reliability Group, Technical Paper, Texas,
TX, November.
Woodhouse, J. 1991. Turning engineers into businessmen,
14th National Maintenance Conference, London.
Woodhouse, J. 1993. Managing Industrial Risk, Chapman
Hill Inc, London.
Woodward, D.G. 1997. Life Cycle CostingTheory, Information Acquisition and Application, International Journal of Project Management, Vol. 15, No. 6, pp. 335344.
Yaez, M., Joglar, F., Mohammad, M. 2002. Generalized
renewal process for analysis of repairable systems with
limited failure experience, Reliability Engineering &
System Safety, Vol. 77, pp. 167180.

939

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk trends, indicators and learning rates: A new case study

of North sea oil and gas
R.B. Duffey
Atomic Energy of Canada Limited, Chalk River, Ontario, Canada

A.B. Skjerve
Institute for Energy Technology, Norway

ABSTRACT: Industrial accidents, explosions and fires have a depressingly familiar habit of re-occurring, with
similar if not identical causes. There is a continual stream of major losses that commonly are ascribed to poor
operating and management practices. The safety risks associated with modern technological enterprises make it
pertinent to consciously monitor the risk level. A comprehensive approach in this respect is being taken by the
Petroleum Safety Authority Norway (PSA)s program Trends in Risk Levels Norwegian Continental Shelf.
We analyse the publicly available data provided by this program using the DuffeySaull Method. The purpose of
the analysis is to discern the learning trends, and to determine the learning rates for construction, maintenance,
operation and administrative activities in the North Sea oil and gas industry. This outcome of this analysis allows
risk predictions, and workers, management and safety authorities to focus on the most meaningful trends and
high-risk activities.

INTRODUCTION

The offshore oil and gas industry is a huge and

modern technological enterprise in which vast quantities of oil are pumped, shipped, and stored. These large
socio-technological facilities pose major hazards, with
potential for spills, fires, sinkings and explosions in
a hazardous and sensitive sea environment with risk
of accident, injury and death to the operators. In his
review of major oil industry failures and engineering practices (Moan 2004) assesses that: the main
cause of actual structural failures are due to human
errors and omissions . . . and cause 8090% of the
failures of buildings, bridges and other engineering
structures. This is also true with what is found in all
other industries and technological systems world wide,
and the same types of mistakes and common errors
appear (Duffey & Saull 2002). Industrial accidents,
explosions and fires have a depressingly familiar habit
of re-occurring, with similar if not identical causes.
There is a continual stream of major losses that commonly are ascribed to poor operating and management
practices, as in recent examples of sometimes billion
dollar damage and losses at offshore oil rigs (e.g. the
P-36 platform that sank following three explosions on
March 2001), oil storage facilities (e.g. the Buncefield
Oil Depot explosions and fire of December 2005), and
refineries (e.g. the catastrophic process accident on

the BP Texas City refinery of March 2005). This has

extreme significance in risk management, and great
importance in the development of safety management
systems, not to mention the large impact on insurance costs, corporate liability and losses, and the threat
to worker safety. This is important not only for such
large disasters but also for everyday accidents, where
the issue is: How can a large facility loss be predicted
using the everyday events and operation of that facility.
The safety risks associated with modern technological enterprises make it pertinent to consciously
monitor the risk level, to assess the extent to which
safety improving initiatives are required. A comprehensive approach in this respect is being taken by
the Norwegian oil and gas regulator, the Petroleum
Safety Authority Norway (PSA), who are responsible for overseeing the safety of many of the massive
deep sea platforms operating in the storm swept North
Sea. The PSAs newly developed program called
on Trends in Risk Levels Norwegian Continental
Shelf, where key measures have been defined for
the purpose of tracking and evaluating both relative
safety improvements in specific areas plus defining
an overall risk level, with the objective to create
a reliable decision making platform for industry and
authorities (PSA 2003). In a wide-range approach to
measure risk for an entire industrial sector, twentyone (21) risk indicators define situations of hazard and

941

accident (called DFUs), covering many known major

and minor outcomes. These indicators include data
for the differing activity segments of the oil and
gas offshore and onshore work (shipping, transport,
maintenance . . .) and include events and abnormal
activities (leaks, accidents and incidents . . .), plus
the effectiveness of barriers (systems, practices
and procedures . . .). The yearly trends of the quantitative data are analyzed as to whether these show
change (increase, decrease, or not) of both the numbers and rates of indicator outcomes; and whether there
is any relation to more qualitative measures based on
attitudinal surveys.
Determining the safety level based on this type
of calculation can, however, be a difficult task. The
recent PSA report states: On the basis of the data
and indicators used in this project, no clear positive
or negative trends can be observed in risk level. Most
major accident indicators show an improvement in
2003 in relation to 2002. Serious injuries to personnel also show a decrease in 2003. The position is now
on a level with the average for the previous 10 years.
Cooperation and trust between the parties are seen as
good.
Since the common factor and major cause in industrial accidents everywhere is the human involvement,
it is postulated here that by understanding the prior
outcomes, human learning and error correction, we
can predict the probability of observing any outcome.
The key questions to answer when looking at trends
are: Are we learning from our past mistakes? What is
the rate of learning now? What is it predicted to be in
the future?
Precisely to quantify such issues, Duffey & Saull
(2002) have derived measures and methods for the
analysis of learning rates as direct indicators of safety
improvement using existing worldwide outcome data
for some 200 years and covering over 60 examples. The approach, called the Duffey-Saull Method
(DSM), uses the Learning Hypothesis to analyze and
predict errors, accidents, injuries and all other such

Table 1.

risk outcomes as a function of experience. It assumes

that with continuous exposure to a given operational
setting humans will learn to master task performance,
and that the manifest effect of learning will be lower
accident/incident rates because humans as a starting
point is assumed to be the key contributing factor to accident/incidents. The present Case Study
applies these techniques and approaches to analyze
the new and publicly available North Sea outcome
data. Using the experience-based DSM, we try to discern the learning trends, and determine the learning
rates for construction, maintenance, operation and
drilling activities in the North Sea oil and gas industry. In our Case Study, we provide a basis to determine,
prioritize, and compare the learning rates and injury
trends between different key work phases. This analysis allows risk predictions, and provides guidance
for workers, management and safety authorities to
focus on the most meaningful trends and high-risk
activities.
2

RISK INDICATOR DATA ANALYSIS

The procedure we use is to first determine the risk

outcomes, rates and numbers, and their distribution
with experience. The basic prior data for Norway for
19962005 are reported by the PSA in both graphical
and tabular form (PSA 2007). Some equivalent data
for the UK for 19922002 are tabulated in Yang &
Trbojevic 2007 (Table 6.16 p 195). All the data are typically given and are analyzed by calendar year, such as
the number of injuries to workers, broken down by different sub-categories of severity (e.g., major or total),
and work location and/or activity type (e.g., fixed or
mobile facility, drilling or maintenance).
To convert to a learning basis for analysis, we use
the relevant measure of experience as the accumulated worker-hours, summing the year-by-year numbers reported. A typical xls. spreadsheet tabulation
and analysis of the rates is shown in Table 1 for a

Typical data subsetNorway well drilling injuries 19962005.

Well drilling (hours)

Injuries, n

AccMh

Entropy, H

Injury rate/Mh

Year

4670117
4913477
4967799
4418068
4696224
5168486
5506589
5827360
6248973
6273504

145
141
133
117
121
110
103
90
54
59

4.670117
9.583595
14.551394
18.969462
23.665686
28.834172
34.340761
40.168122
46.417095
52.690599

0.088633
0.181884
0.276167
0.360016
0.449144
0.547236
0.651744
0.762339
0.880937
1

0.27047
0.266685
0.258794
0.241637
0.246107
0.233505
0.224957
0.207881
0.150437
0.159497

31.04847
28.69658
26.77242
26.48216
25.76538
21.28283
18.70486
15.44438
8.64142
9.404633

1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

942

subset in our observational interval. In this case, the

data are for injuries in drilling at fixed facilities for
Norway, and similar tables were made for all the
various sets where numbers were available.
This Table is in general for the jth observation interval, with the sub-intervals within it. Such a tabulation
is not by itself very informative, apart from illustrating
the manipulations and steps in the necessary arithmetic
for each experience increment:
1. adding up prior worker-hours to obtain the running total of the accumulated millions of hours of
experience, (AccMh) for each ith sub-interval;
2. turning the injury numbers, ni , into risk Rates per
Mh by straightforward division;
3. calculating the non-dimensional experience, N ,
by dividing each AccMh interval, i , by the
total accumulated experience, T (T = AccMh =
53Mh); and
4. calculating the entropy (Hi = pi ln pj ) in each
ith sub-interval from the probability, where pi =
ni /Nj , where, Nj , is the total number of injuries
(Nj = ni = 1073).
To clarify the trends, typical results of such analysis of the raw data are then plotted in Figure 1. The
figure also shows some UK data alongside the Norway
data. By grouping the data together in this way, several key comparisons and observations are possible.
In addition, to simplify the presentation, simple exponential fits are shown to the data, since we expect such
a curve to crudely represent the improvement effects
of learning (Duffey & Saull 2002).
Firstly, learning is evident in most of the data, but
the absolute risk indicator rates are higher for some
activities, and some task areas are clearly learning
slower than others (the slope is half). We may predict that they will all reach some asymptotic but slow
learning state, by about twice the present experience
if learning continues. The most hazardous (highest
Offshore Risk Rates
(Injuries Data: Norway 1996-2005 and UK 1992-2001)
90
Norway administration
Norway production

80

Norway drilling
Norway construction

70

Norway maintenance

Rate/ Mh

UK Major Injuries

-0.017Mh

60

Rate (UK) = 63e

UK Major injuries
Norway Maintenance injuries

50

Norway drilling
Norway administration

40
Rate (Norway)= 52e

-0.014Mh

30
20

-0.0271Mh

Rate= 41.e
2
R = 0.8883

10

-0.0289Mh

Rate = 21e
2
R = 0.8865

0
0

10

20

30

40

50

60

70

80

Work Experience, AccMh

Figure 1.

Typical data plot and simplified curve fits.

risk) activities are clearly maintenance and drilling,

and must be the areas of most safety importance and
management attention.
Secondly, the lowest rates attained so far (in administration and production) are 5/Mh, or about 1 in
200,000 experience hours, in complete accord with
the lowest risk found in any other industry (Duffey &
Saull 2002). However, the highest rates are ten times
more, or 1 in 20,000 experience hours, which is also
comparable to other industries.
Thirdly, the UK has less experience, but a simple
extrapolation forward of the rough fit to the major
injury (MI) data, and backward extrapolation of the
Norway maintenance data shows similar event rate
magnitudes and learning rates. The implication is that
learning of similar effectiveness in Norway and the
UK suggests that further research is needed into the
influencing factors and causes of this similarity. Thus,
we may predict and expect the UK rates to fall further
and track down towards the Norway risk rates if such
international learning continues. Similar convergence
trends are observed with differing experience in, say,
commercial aircraft near-misses and marine shipping
accidents.
3

ARE WE DOING ENOUGH TO ENSURE

SAFETY?

A key question for managers of hazardous industries

is: Are we doing enough to ensure safety? In this
section we will take a closer look at learning in a workplace setting, and suggest that this question may also
be answered based on an assessment of the effectiveness of the joint initiatives taken by an organization
(or an entire industry) to ensure safety.
Individuals learn as they gain experience (Ebbinghause 1885). Employees in petroleum companies will
learn from participation in the formal education and
training programs offered by their organization. The
aim of these programs is to ensure that all employees possess the competence, i.e. the skills, knowledge
and attitudes required to efficiently perform their jobs
to the specified standard (IAEA 2002; Skjerve &
Torgersen 2007). As part of their engagement in the
every-day work activities, the employees will moreover face a range of learning opportunities resulting
from the myriad of different situations that arise
from interactions between humans, technology and
administrative systems. The employees will need both
the competence acquired from the formal education/
training sessions and the competence acquired based
on the more informal experiences gained on-the-job,
to be able to perform their tasks efficiently (Johnston &
Hawke 2002). With increased experience, employees will obtain still more refined insights into the
task performance process and their task performance

943

environment,1 and gradually they will be able to

perform the routine part of their tasks in a highly
automated manner (Rasmussen 1986).
Observation, imitation, reflection, discussion, and
repetition may all constitute important elements in
employees learning processes. Handling of situations where unexpected occurrences happen in relation to task performance provides an important basis
for learning. Such unexpected occurrences may be
caused by human errors (e.g. errors of the particular
employee, errors of colleagues in some situations
the errors may even be consciously introduced for the
employees to learn something). Unexpected occurrences may also be caused by breakdowns in technology or administrative systems, or by any combination
of the above factors. When unexpected occurrences
arise things will not progress according to plan, and
this will spur the employees to develop a more comprehensive understanding of the task performance
process and the work environment. This, in turn, will
improve their ability to perform safely in future situations. Accidents constitute important, but highly
unwarranted, learning opportunities. When accidents
happen, they will tend to challenge the organizations
model of the risks it faces and the effectiveness of its
countermeasure (Woods 2006). For this reason, radical changes may be implemented in the organization
following an accident investigation. This suggests that
not only individuals but also the organization as such
may learn from experience.
Organizational learning may be defined as . . . the
capacity or processes within an organization to maintain or improve the performance based on experience
(DiBella 2001, Duffey & Saull 2002). A key element in organizational learning is the transformation
of experiences gained by employees to the organizational level. In this process, however, the organization
needs to be aware that not all the experiences gained
by employees will contribute to increase the likelihood for safe performance: Employees are engaged in
a continuous learning process. Misunderstandings of
factors in the work environment, misunderstandings of
the inter-relationship between these factors, inaccurate
risk perception, etc. can all be expected to be (intermediate) elements or states in a learning process. In
addition, to the experiences of the employees, experiences obtained by other organizations or by other
industries may also prove valuable to organizational
learning. Organizational learning should be manifest
in the structures and processes of the organization

1 The extent to which this process involves deduction

based on inference rules or the development of mental
models is still a matter of debate (cf., e.g. Johnson-Laird &
Byrne 2000).

(Svenson 2006). Concretely, organizational learning

may result in the introduction of new work practices,
revisions of operational procedures, refinements of
training programs, improvement in the safety management approach, etc. That is, in initiatives that jointly
aim at ensuring safe and efficient production.
To facilitate learning processes at all levels in the
organization it is important to ensure that a learning culture is engineered (Reason 1997). A learning
culture can be defined as . . . an environment in
which opportunities for learning are openly valued and
supported and are built, where possible, into all activities (DEST 2005). It has been suggested that effective
high-reliability organizations is characterised by their
ability to learn as much as possible for the failures that
occur (Weick & Sutcliffe 2001).
Finally, the importance of ensuring a sound safety
culture is generally reckoned as a prerequisite for
safe production in high-risk industries. The agenda
among most actors in the Norwegian petroleum sector is to improve the safety culture both within and
across the industry (Hoholm 2003). Safety culture can
be defined as . . . that assembly of characteristics
and attitudes in organizations and individuals which
establishes that, as an overriding priority, safety issues
receive the attention warranted by their significance.
(Adapted from IAEA 1991, Yang & Trbojevic 2007).
A sound safety culture means that the structures and
process of the organization should work together to
ensure safety. Thus, deviations caused by the activities
in one part of the organization should be compensated by the activity in other parts of the organization
so that safety is always ensured (Weick & Sutcliffe
2001). A sound safety culture, moreover, implies that
the attitudes and behaviours of employees should promote safety. In the context of the Norwegian petroleum
industry, the impact of colleagues and managers attitudes to safety on the individual employee at was
demonstrated in two recent studies (Aase et al. 2005;
Skjerve, in press).
One way to answer the question: Are we doing
enough to ensure safety? could be to calculate the
effectiveness of the joint initiatives taken by an organization (or an entire industry) to ensure safety. In the
next section, we introduce H as one such possible
measure.
4

SAFETY CULTURE, RISK MANAGEMENT

AND PREDICTION

The emergence of order from chaos is a precise

analogy to that postulated for microscopic physical and chemical systems (Prigogine 1984), and is
the expressed intent of safety management and risk
indicators for macroscopic socio-technological systems. In particular, we determine the Information

944

Entropy risk measure, H, which Duffey & Saull (2007)

suggest is the objective and quantitative measure of
safety culture, management systems, organizational
learning and risk perception. Thus, since we may
regard, H, as a measure of the disorder this, of
course, is the converse of order, and hence is an indication of the effectiveness of these safety management
processes.
The statistical theory that determines the outcome
risk distribution yields an explicit expression for the
Information Entropy, H, using the probability of the
outcomes (Pierce 1980, Jaynes 2003, Duffey & Saull
2004, Duffey & Saull 2008). The degree of order is
a function of the depth of experience based on the
frequency of error state occupation, ni = i Nj .
The classic result for the Information Entropy,
H, is a measure of the uncertainty, or the missing
information or the degree of order given by:
Hj = pi ln pi

(1)

Substituting in the expression for the Information

Entropy, H, in the companion paper (Duffey & Saull
2008), we obtain:

2
Hj = 1/2 p0 eaN {aN + 1/2}

behavior on the degree of order and on the risk trends

with experience that are attained.

COMPARISON OF THEORY AND DATA

For the present Case Study, we can now compare this

theory to the overall trends of a subset of the present
risk indicator data, noting that we have evaluated the
entropy already as part of the initial data analysis
(see Table 1). To simplify, the data are normalized to
the initial probability at the initial or lowest experience,
where we take, p0 = 1, by definition. Figure 2 shows
the Norway (injury) and UK (major injury and >3 day
injury) data compared to the theory (SEST) prediction,
but adopting a value of, a = 1, for the shape or slope
parameter in the entropy distribution.
Rather satisfyingly, the theory and data easily
appear side-by-side on the same graph, lending some
credence to this analysis. The other data shown
for comparison purposes are the commercial aircraft
near-misses (NMACs), because of the significant and
traditional airline emphasis on safety (Duffey & Saull
2002). The NMAC line up rather better with the

(2)

Information Entropy
Offshore Injuries UK and Norway
0.4

0.35

Entropy, plnp

0.3

0.25

0.2

0.15
Norway entropy
UK entropy MI

0.1

UK entropy 3d
US NMACs entropy

0.05

Theoretical entropy, a=1

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Non-dimensional Experience, N*

Entropy
Norway Offshore Injury Data 1996-2005
0.35

0.3

0.25

Entropy, p lnp

The relative value of the information entropy, H,

at any experience depth is also an objective measure
of the cultural aspect of modern technologies called
organizational learning since it reflects the degree
of learning and the extent of management order
or effectiveness. The so-called organizational learning and safety culture attributes of a HTS, and its
managements ability to respond effectively to the
demands for continuous safety improvement. The
resulting structure and probability of observed outcomes are a direct reflection of the internal organizational and skill acquisition caused by the innumerable
human learning and unlearning interactions occurring
within.
These statistical fluctuations, due to human decision making and actions, also cause the uncertainty
of precisely when an event will actually occur, which
uncertainty is determined and measured by the information entropy. The unobserved random fluctuations produced by unpredictable human behavior (the
chaos) are reflected in the emergent order at the system
level: namely, the observed predictable trend of learning and error reduction with increasing experience.
Thus the influence of safety management can be quantified, and predictions made about the effectiveness of
learning and management systems. By providing the
information entropy, H-measure (which we can also
refer to as the Learning Entropy), we can not only
describe but also can predict the impact of the learning

0.2

0.15

0.1
Norway injury data
Theory, a=1

0.05

E = 0.28 - 0.11N*
2
R = 0.478

Linear data fit

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Non-dimensional Experience, N*

Figure 2.

945

Comparisons of theory and data.

0.8

0.9

theoretical prediction, but there are clearly some

key differences between the oil and gas data set
trends. Despite the scatter, we note for this data subset
that the:
1. entropy distribution with experience lies above the
theory line;
2. slope trend is less than the theory, indicating
insufficient attainment of order;
3. data lie above the best aircraft practices (aircraft
near-misses); and
4. best (but still far from perfect) fit to all the injury
data is a straight line, not an exponential as we
should expect.
The approximate straight line fit shown is, H =
0.28 1.11N , which actually corresponds to the first
two terms of the series expansion of the rather slowlydecaying exponential. Therefore, the implied firstorder approximate value is a 1.11 for the distribution
exponent.
All these trends and comparisons suggest symptoms of potentially insufficient learning, giving inadequate reduction in risk compared both to the expected
ideal, and other industries. This adverse trend was
confirmed by plotting the rates against the Universal
Learning Curve, and finding a similar value of, k 1,
for the learning rate constant.
6

CONCLUSIONS AND OBSERVATIONS

We are interested in predicting safety performance

and accident occurrences utilizing quantitative analysis of prior data. These predictions should serve to
inform the industry to facilitate decision making with
respect to when more emphasis on safety initiatives
is required. As the predictions express learning rates,
they will allow companies to readily compare their
learning rate with other companies in the same domain
to establish whether they are on the right track. Likewise, the predictions allow for comparisons between
entire industrial sectors, and they may in this way contribute to decisions of the national safety authorities
when defining requirements to the various industrial
sectors.
The two types of representations presented in this
paper in Figure 1 and Figure 2 invites different interpretations of what spurs the learning rate. The representation used in Figure 1 may invite the interpretation
that further accidents/incidents are necessary for learning to take place whereas Figure 2 suggests that we can
intercompare learning and progress using the existing
knowledge.
Still, even if handling of unexpected events is a key
element in the learning process (as discussed above),
this does not imply that accidents/incidents will have to
take place for people to learn. Accidents/incidents will

only occur when the organization is not sufficiently

robust to prevent that human performance will have
adverse implications. The fact that the learning rate
seems to decrease only when incidents and accidents
occur is caused by the fact that accidents/incidents
(rather than e.g. successful outcomes) serves as input
data for the model. In general accidents/incidents can
be expected to occur in lower frequencies as an organization or an entire industrial section gains experience.
This point of view is emphasised by the representation
contained in Figure 2. It shows the learning rate based
on the level of control an organisation or and industrial
section have over the production processes.
Based on our Case Study of the observed and published safety indicator data for some ten years of
operation of Norway and UK North Sea oil and gas
facilities, and the trends shown in a subset of the data,
we observe that:
Learning is occurring in the major risk indicators
as experience is gained, and this trend is similar
between the Norway and UK indicating some commonality in approach and safety standards (after
correction for differing experience);
Certain activities, notably maintenance and drilling,
apparently have much higher risk than others, both
in numbers and rates, and suggest themselves as
priority areas for management emphasis; and
Evaluation of the Learning Entropy as a measure of
the degree of order attained by safety management
(suggested to represent organizational learning and
safety culture) also indicate symptoms of potentially
insufficient learning.
Extension of this Case Study to the complete set
of risk indicators would be desirable; as also would
revising the indicator choices to reflect priority of
risk-related importance; and changing the conventional purely time-series manner of data reporting and
analysis.

REFERENCES
Aase, K., Skjerve, A.B.M. & Rosness, R. 2005. Why Good
Luck has a Reason: Mindful Practices in Offshore Oil and
Gas Drilling. In: S. Gherardi & D. Nicolini (eds.), The
Passion for Learning and Knowing. Proceedings of the
6th International Conference on Organizational Learning
and Knowledge, vol. 1.: 193210. Trento: University of
Trento e-books.
DEST, 2005. The website of the Department of Education,
Science and Training of Australia. http://www.dest.gov.
au / sectors / training _ skills / policy_issues_reviews/key _
issues/nts/glo/ftol.htm#Glossary_-_L (Accessed January
2008)
DiBella, A.K. 2001. Learning practices: Assessment and
Action for Organizational Improvement. Upper Saddle
River, N.J: Prentice-Hall.

946

Duffey, R.B. & Saull, J.W. 2002. Know the Risk, First Edition,
Boston, USA, Butterworth and Heinemann.
Duffey, R.B. & Saull J.W. 2004. Reliability and Failures of
Engineering Systems Due to Human Errors, Proc. The
First Cappadocia Int. Mechanical Engineering Symposium (CMES-04), Cappadocia, Turkey.
Duffey, R.B. & Saull, J.W. 2007. Risk Perception in Society: Quantification and Management for Modern Technologies, Proc. Safety and Reliability Conference, Risk
Reliability & Societal Safety (ESREL 2007), Stavanger,
Norway, 2427 June.
Duffey, R.B. & Saull, J.W. 2008. Risk Management Measurement Methodology: Practical Procedures and Approaches
for Risk Assessment and Prediction, Proc. ESREL 2008
and 17th SRA Europe Annual Conference, Valencia,
Spain, 2225 September.
Ebbinghaus, H. 1885. Memory: A Contribution to Experimental Psychology. (Translated from: "ber das Gedchtnis"). http://psy.ed.asu.edu/classics/Ebbinghaus/index.
htm (Accessed January 2008).
Hoholm, T. 2003. Safety Culture in the Norwegian
Petroleum Industry: Towards an Understanding of interorganisational culture development as network learning.
Arbeidsnotat nr. 23/2003. Oslo: Center for Technology,
Innovation and Culture, University of Oslo.
IAEA, 1991. Safety Culture, Safety Series no. 75-INSAG-4,
Vienna: International Atomic Energy Agency.
IAEA, 2002. Recruitment, Qualification and Training of
Personnel for Nuclear Power Plants, Safety Guide no.
NS-G-2.8, Vienna: International Atomic Energy Agency.
Jaynes, E.T. 2003. Probability Theory: The Logic of Science, First Edition, Edited by G.L. Bretthorst, Cambridge
University Press, Cambridge, UK.
Johnston, R. & Hawke, G. 2002. Case studies of organisations with established learning cultures, The National
Centre for Vocational Education Research (NCVER),
Adelaide, Australia. http://www.ncver.edu.au/research/
proj/nr9014.pdf (Accessed January 2008)
Johnson-Laird, P. & Byrne, R., 2000. Mental Models Website. http://www.tcd.ie/Psychology/Ruth_Byrne/mental_
models/ (Accessed January 2008).
Moan, T. 2004. Safety of Offshore Structures, Second Keppel
Offshore and Marine Lecture, CORE Report No. 200504,
National University of Singapore.

Petroleum Safety Authority Norway (PSA) 2003. Trends

in risk levels-Norwegian Continental Shelf, Summary
Report, Phase 42003, Ptil-04-04, p.11, Norway.
Petroleum Safety Authority Norway (PSA) 2007. Supervision and Facts, Annual Report 2006, Stavanger, Norway,
26 April, available at www.ptil.no.
Pierce, J.R. 1980. An Introduction to Information Theory,
Dover, New York.
Prigogine, I. and Stengers, I. 1984. Order Out of
Chaos: Mans New Dialogue with Nature, Toronto,
Bantam Books.
Rasmussen, J. 1986, Information Processing and HumanMachine Interaction. An Approach to Cognitive Engineering, System Science and Engineering, vol. 12, New York:
North-Holland.
Reason, J. 1997. Managing the Risks of Organizational
Accidents. Aldershot, UK: Ashgate.
Skjerve, A.B. (in press). The Use of Mindful Safety Practices
at Norwegian Petroleum Installations. To be published in
Safety Science.
Skjerve, A.B. & Torgersen, G.E. 2007. An OrganizationalPedagogical Framework to Support Competence Assurance Activities. In: T. Aven, J.E.Vinnem (Eds.), Risk,
Reliability and Societal Safety: 19251932. London, UK:
Taylor & Francis Group.
Svenson, O. 2006. A Frame of Reference for Studies of Safety
Management. In (Ed.) O. Svenson, I. Salo, P. Oedewald, T.
Reiman, A.B. Skjerve, Nordic Perspectives on Safety
Management in High Reliability Organizations. Theory
and Applications: 17. Valdemarsvik, Sweden: Stockholm University.
Weick, K.E. & Sutcliffe, K.M. 2001. Managing the Unexpected. Assuring High Performance in an Age of Complexity. San Francisco, CA: Jossey Bass.
Woods, D.D. 2006. Essential Characteristics of Resilience.
In E. Hollnagel, D.D. Woods & N. Leveson, Resilience
Engineering. Concepts and Precepts:2134. Aldershot,
UK: Asgate.
Yang, J. & Trbojevic, V. 2007. Design for Safety of Marine
and Offshore Systems, IMarEST Publications, ISBN:
1-902536-58-4.

947

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Robust estimation for an imperfect test and repair model

using Gaussian mixtures
Simon P. Wilson
Centre for Telecommunications Value-Chain Research, Trinity College Dublin, Dublin, Ireland

Suresh Goyal
Bell Labs Ireland, Dublin, Ireland

ABSTRACT: We describe a technique for estimating production test performance parameters from typical
data that are available from past testing. Gaussian mixture models are used for the data because it is often
multi-modal, and the inference is implemented via a Bayesian approach. An approximation to the posterior
distribution of the Gaussian mixture parameters is used to facilitate a quick computation time. The method is
illustrated with examples.

INTRODUCTION

Many manufacturing processes for electronic equipment involve a complex sequence of tests on components and the system. Statistical imperfect test and
repair models can be used to derive the properties of
the sequence of tests, such as incoming quality, rate
of false positives and negatives, and the success rate
of repair, but require the value of these properties
to be specified. It is recognized that optimal testing
strategies can be highly sensitive to their value (Dick,
Trischler, Dislis, and Ambler 1994).
Fortunately, manufacturers often maintain extensive databases from production testing that should
allow these properties to be estimated. In this paper
we propose a technique to compute the properties of
a test from the test measurement data. It is a robust
technique that is designed to be applied automatically,
with no intervention from the test engineer in most
cases.
This learning process is not as straightforward as
it might appear at first for several reasons. First, the
properties of the test that interest us are not what are
recorded in the database. Rather, the test measurements themselves are what are stored. We address this
by defining a model for the measurements, introduced
in (Fisher et al. 2007a; Fisher et al. 2007b), that implicitly defines the test properties; we fit measurement
data to the model following the Bayesian approach
which then gives us estimates of the test properties.
The Bayesian approach has the advantage that it correctly propagates the uncertainty in the measurement
model parameter estimates, as inferred directly from
the data, through to the estimates of the test model

parameters. Second, the measurement data display

a wide variety of behaviour that are challenging to
model, such as multimodality and extreme outliers.
We address this by using a Gaussian mixture model
for the measurement data. Third, the data are sometimes censored; for example, the database may only
record that a measurement was within acceptable limits, rather than the actual value. In this case inference
with a Gaussian mixture model is very difficult. For
these cases we show that it is possible to fit a single
component Gaussian model. Finally, practical application of the method requires that the computations be
done on-line, so computing time must not be longer
than a few seconds. We address this by using a fast but
rather crude approximation to implement the Bayesian
inference.
The paper is organised as follows. Section 2
describe the model for what is observed in the test, the
model for the test properties that interest us and the
relationship between them. It also includes a description of the data. Section 3 describes the statistical
inference procedure that obtains estimates of the test
model parameters from the data. Section 4 illustrates
the method with some examples and Section 5 is some
concluding remarks.
2
2.1

MODEL
The measurement model

A unit is tested by measuring the value of one property of the unit. The true value of the property being
measured is x. The value is measured with error
and we denote the observed value as y. A Gaussian

949

mixture model is used for x in most circumstances,

since it is able to model a wide range of behaviour
that we have seen in practical examples, such as
extreme outliers, skewness and multi-modality. We let
= {pk , k , k2 | k = 1, . . . , K} denote the Gaussian
mixture component weights, means and variances.
Therefore:
px (x | ) =

K


pk 

k=1

Figure 1.


1
2
exp 2 (x k ) ,
2k

(1)

< x < . We assume in this case that the

measurement error is Gaussian:

< y < .

measurement model for x and y, while in this work we

leave BG to be defined directly.
2.3 Data
There are three sets of measurements available:

1
2
2
py|x (y | x, s2 ) =
e(yx) /2s ,
2
2 s
(2)

Marginally, y is a Gaussian mixture with the same pk

and k as x but with variances k2 + s2 . The model for
x and y in terms of and s2 is called the measurement
model.
2.2

Flow chart of the test and repair model.

2 k2

The test and repair model

A unit is classified to be good if x is in the interval

(L, U ). A unit passes the test if y is in the interval
(L, U ). The parameters of real interest pertain to the
performance of the test. They are:
GI = P(L x U | ), the proportion of good
units;
GG = P(L y U | L x U , , s2 ), the
probability that a good unit passes the test;
GB = 1 GG , the probability that a good unit
fails the test (a false negative);
BB = P(y < L or y > U | x < L or x > U , , s2 ),
the probability that a bad unit fails the test;
BG = 1 BG , the probability that a bad unit
passes the test (a false positive).
BG , the probability that a unit that is bad is repaired
to good. This arises because a unit that fails the test
is sent to be repaired and is then retested. There
is imperfect repair so truly bad units may not be
repaired to good. We do assume, however, that truly
good units that have failed the test cannot be repaired
to be bad.
Figure 1 represents the test and repair process.
There are therefore 4 free parameters of the test and
repair model: GI , GG , BB and BG . It is important
to note that the first three are defined in terms of the

1. Data from a set of one-off tests where a single

unit is tested m times. Such tests are occasionally
carried out by the engineers to learn about the
repeatability of the test results and are clearly an
important source of data to learn about py (y | x, s2 ).
We define z1 , . . . , zm to be the measured values
from the one-off test, which we also refer to as
one-off data in some equations. We also define
x to be the unknown true value of the unit used in
the one-off test.
2. Data from the first-pass test where n different
units are tested. Let y1 , . . . , yn denote the measured values from the first pass test, which are also
denoted first pass data.
3. Data from the second-pass test where n2 units,
that failed the first-pass test and were repaired, were
measured again. In this case we only observe the
number of units ns that pass this test.
While the quantity being measured in the test may
be continuous, sometimes the available one-off and
first pass data only show whether the test was passed
or not e.g. the zj and yi are interval-censored to (L, U ).
It is difficult to fit the Gaussian mixture model to such
interval censored data because they contain very little
information about the number of components. However we will show that it is possible to fit a single
component Gaussian. We have found that it is sufficiently parsimonious to allow identification of the
measurement model parameters and produce sensible
estimates of the test model parameters.

STATISTICAL INFERENCE

A Bayesian approach is adopted, so the goal is to compute the distribution p(GI , GG , BG , BG | data). The
likelihood is most easily written in terms of , s2 , BG

950

and also x, the true value of the unit used in the oneoff tests. The deterministic relationships in Section 2.2
between the measurement model parameters ( , s2 )
and the test model parameters then allow us to compute the posterior distribution of GI , GG , BG and
BG from that of , s2 and BG .
3.1

The likelihood

If the measurements themselves are recorded then the

likelihood for z1 , . . . , zm and y1 , . . . , yn is:
p(one-off data, first pass data | x, , s2 )



m
n


2
2
py (yj | , s ) .
= py|x (zj | x, s )
j=1

3.2

Currently we assume flat prior distributions for all

model parameters, although we recognise that a lot of
useful information could be incorporated into a prior
that could improve the estimation, particularly in the
case of censored data where the information in the data
can be weak. We compute the posterior distribution of
the test parameters p(GI , GG , BB , BG | all data) as
follows. We have found that the second pass data contain little information about GI , GG and BB so we
ignore it and use the approximation:
p(GI , GG , BB , BG | all data)
p(GI , GG , BB | one-off data, first pass data)

(3)

I , GG , BB ).
p(BG | second pass data, G

i=1

For the examples that we consider here, we use

Equation 2 for py|x (zj | x, s2 ), and the Gaussian mixture model of Equation 1 for x, which means that
py (yj | , s2 ) is a Gaussian mixture probability as in
Equation 1 but with variances k2 + s2 instead of k2 . If
interval censored data are recorded then the likelihood
is Bernoulli for both zj and yi with success probabilities
Pz = P(L y U | x, s2 )

(4)

and
Py = P(L y U | , s2 )

(5)

respectively. These probabilities are easily evaluated

under the Gaussian mixture model. Hence:
p(one-off data, first pass data | x, , s2 )
n

= Pznz (1 Pz )mnz Py y (1 Py )nny ,

Computing the posterior distribution

(6)

where ny and nz are the number of units passing in the

first pass and one off tests respectively.
The second-pass test data is always simply pass/fail. The probability of passing is Ps =
P(pass 2nd test | fail 1st test). Applying the partition
law and Bayes law we can show that:

(9)

I , GG and BB are posterior means

where G
that are computed from p(GI , GG , BB | one-off data,
first pass data). Therefore our task is reduced
to the computation of p(GI , GG , BB | one-off data,
I , GG ,
first pass data) and p(BG | second pass data, G
BB ).
I , GG , BB ) is
The term p(BG | second pass data, G
straightforward, being proportional to the likelihood
I , GG =
of Equation 8 with fixed values of GI = G
GG and BB = BB . This is computed directly on
a discrete grid of values of BG , once the posterior
I , GG and BB have been computed from
means G
p(GI , GG , BB | one-off data, first pass data).
The first term is more difficult. It is done by simulating values of and s2 from p(, s2 , x | one-off data,
first pass data). For each sample, we compute
(GI , GG , BB ) via the Equations in Section 2.2. The
sample averages of these are the posterior means
I , GG and BB . The samples are smoothed, via
G
a kernel density estimate, to an approximation of
p(GI , GG , BB | one-off data, first pass data).
The way and s2 are simulated from p(, s2 , x |
one-off data, first pass data) is different according
to whether the data are exact or censored. This is
described in the subsections below.

Ps = [GG (1 GG )GI + GG BG BB (1 GI )
+ (1 BB )(1 BG )BB (1 GI )]
[(1 GG )GI + BB (1 GI )]1 .

(7)

The likelihood for ns units passing from n2 is

Bernoulli:
p(ns | n2 , GG , BB , BG ) = Psns (1 Ps )n2 ns .

(8)

3.2.1 Exact data case

By Bayes law, p(, s2 , x | one-off data, first pass data)
is proportional to Equation 3. For the Gaussian
mixture model that we are proposing, recall that
= {pk , k , k2 | k = 1, . . . , K}. We note that if we
define i2 = i2 + s2 then we can reparameterise
( , s2 ) as = {pk , k , k2 | k = 1, . . . , K} and
s2 , with the restriction 2 > s2 . This allows us to

951

factorise the posterior.

p( , s2 , x | one-off data, first pass data)
p(s2 , x | one-off data) p( | first pass data).
(10)
To compute p( | first pass data), the standard
Bayesian approach to fitting a mixture model is
by Monte Carlo simulation and requires a timeconsuming reversible jump MCMC (Richardson and
Green 1997). This is too slow for the practical implementation of this method, where we expect the test
engineer to interact with the inference algorithm as
a test sequence is designed. We adopt a much faster
although more crude alternative that assumes that each
set of component means and variances are independent, and so the posterior distribution is evaluated
for each separately. This independence assumption
is used in the variational Bayes approximation to
mixtures (Constantinopoulos and Likas 2007). The
number of components in the mixture K is determined
by an initial fit using the fast message length algorithm (Figueiredo and Jain 2002). This method also
gives point estimates of mixture means, variances and
weights which we denote k , k2 and p k . A posterior
distribution on the k , k2 and pk is fitted around these
point estimates by first assigning each first pass observation to the mixture component with the smallest
Mahalonobis distance (yi k )/ k . A posterior distribution for each component mean and variance pair
(k , k2 ) is then computed separately using the observations assigned to it using a Gaussian likelihood;
the result is the standard conjugate normal-inverse
gamma posterior for each (k , k2 ) separately (Gelman
et al. 2003). Components with no observation assigned
are eliminated. The posterior distribution of the pk is
Dirichlet with the parameter given to pk equal to the
number of observations assigned to it; this is again the
conjugate posterior distribution for the pk . Thus we
make the approximation:

3.2.2 Censored Data Case

Here p(, s2 , x | one-off data, first pass data) is proportional to Equation 6, and we only attempt to
fit a single component Gaussian model. Since in
this case = (, 2 ), there are only 4 unknown
parameters, 2 , s2 and xand this distribution
can be evaluated on a discrete grid, from which values
of (, s2 ) are simulated.

EXAMPLES

Two examples are shown. One is simulated data and the

other is a real data example. In both cases the posterior
distribution of the test properties was computed using
the method of Section 3.2.
Data were simulated from the Gaussian mixture
model with 3 components. The means, variances
and weights of the components were (3.0, 3.5, 20.0),
(0.22 , 0.42 , 0.32 ) and (0.2, 0.7, 0.1). The measurement
variance is s2 = 0.0052 and the true value of the
unit used in the one-off tests is x = 3.05. Units are
accepted if they are measured to be in the interval
(2, 4). This leads to true test model parameter values
of GI = 0.826, GG = 0.9992 and BB = 0.9966;
BG was defined to be 0.9. Finally, sample sizes were
n = 5000, m = 500 and n2 = 10. Note that the
data are mainly concentrated in the accept interval but
that there is a group of observations very far from
that interval, centered around 20.0. Also note that the
second pass data size is small; since few units fail
the test, this is also typical. The presence of a group
of extreme outliers like this is common in data from
real tests. Figure 2 shows the marginal posterior distributions of GI , GG , BB and BG . We see that the
method has recovered the true test parameter values
quite well. The much greater posterior variance for

p( | first pass data)

 K


2

p(k , k | first pass data assigned to k)

k=1

p(p1 , . . . , pn ).

(11)

The other distribution p(s2 , x | one-off data) is

another conjugate normal-inverse gamma, since the
data are a normal random sample with mean x and variance s2 . Values of and s2 , and hence and s2 , are
easily simulated from these distributions, as it merely
requires sampling from Gaussian, inverse gamma and
Dirichlet distribution, with the only difficulty being
that the samples must satisfy k2 > s2 .

Figure 2. Posterior distributions of (clockwise from top left)

GI , GG , BG and BB from simulated data. True parameter
values are shown by a vertical line.

952

the censored data. The analysis now gives posterior means and (2.5%, 97.5%) probability intervals as: GI = 0.950 (0.871, 0.997); GG =
0.939 (0.880, 0.993); BB = 0.697 (0.380, 0.962);
BG = 0.55 (0.13, 0.99). We see that the posterior distributions have considerably higher variance,
reflecting the loss of information from the censoring.
5

Figure 3. Histogram of first pass data with fitted Gaussian

mixture model and accept interval limits (vertical lines).

Figure 4. Posterior distributions of (clockwise from top left)

GI , GG , BG and BB for data from the real test data.

BG reflects the far smaller sample size of the second

pass test. Figure 3 shows the first pass data, with the
fitted Gaussian mixture model, showing that the model
has captured the multi-modality of the data easily.
Figure 4 shows the analysis of data from a real test.
The accept limits for this test are (40, 24). Sample
sizes are n = 878 (of which 807 passed), m = 22 (of
which all passed and the sample standard deviation as
0.07) and n2 = 3 (of which 2 passed). The analysis
gives posterior means and (2.5%, 97.5%) probability intervals as: GI = 0.92 (0.89, 0.95); GG =
0.9999 (0.9996, 1.000); BB = 0.999 (0.996, 1.000);
BG = 0.60 (0.19, 0.93). To illustrate the effect
of censoring, we take these data and assume that
they are interval censored to (40, 24). We
then fit a single component Gaussian model to

CONCLUSIONS

We have presented a Bayesian approach to estimating

test properties of an imperfect test and repair model
from test measurement data. For the sort of data that
we describe, many statistical estimation methods have
problems because of the highly varied properties of the
measurement data. Our method can cope with outliers
and multi-modality.
Several issues have come out of the work so far.
First, more work can be done to specify informative
priors that could help analyses, especially in the censored data case. Second, we have employed quite a
crude estimate to the posterior of the mixture model
parameters, by fitting each component separately to
observations that are closest to it. In our experience,
comparing the measurement data to the fitted Gaussian mixture, this appears to work well, particularly
because in most real cases the mixture components
are well separated in the data. The approximation, by
imposing independence between mixture component
parameters, is in the spirit of the variational Bayes
approximation. Nevertheless, a better study to evaluate how well this approximation performs is necessary.
Another issue is that we have left BG to be defined
directly rather than through the measurement model,
as is the case with the other test properties. This means
that we have no way of using the test measurements to
infer anything more than the probability that a repair
succeeds or not. A model for the effect of repair on the
measured value would allow a more detailed understanding of the performance of the repair process.
ACKNOWLEDGEMENTS
This research is work of the Centre for Telecommunications Value-Chain Research http://www.ctvr. ie,
supported by Science Foundation Ireland under grant
number 03/CE3/I405. Bell Labs Ireland research was
also partly funded by a grant from the Irish Development Agency.
REFERENCES
Constantinopoulos, C. and A. Likas (2007). Unsupervised
learning of Gaussian mixtures based on variational com-

953

ponent splitting. IEEE Transactions on Neural Networks

18, 745755.
Dick, J.H., E. Trischler, C. Dislis, and A.P. Ambler (1994).
Sensitivity analysis in economic based test strategy planning. Journal of Electronic Testing: Theory and Applications 5, 239252.
Figueiredo, M. and A.K. Jain (2002). Unsupervised learning
of finite mixture models. IEEE Trans. Pattern Anal. Mach.
Intell. 24, 381396.
Fisher, E., S. Fortune, M. Gladstein, S. Goyal, W. Lyons,
J. Mosher, and G. Wilfong (2007a). Economic modeling
of global test strategy I: mathematical models. Bell Labs
Technical Journal 12, 161174.

Fisher, E., S. Fortune, M. Gladstein, S. Goyal, W. Lyons,

J. Mosher, and G. Wilfong (2007b). Economic modeling
of global test strategy II: software system and examples.
Bell Labs Technical Journal 12, 175186.
Gelman, A., J.B. Carlin, H.S. Stern, and D.B. Rubin (2003).
Bayesian Data Analysis (Second ed.). London: Chapman
and Hall.
Richardson, S. and P. Green (1997). On Bayesian analysis of mixtures with an unknown number of components
(with discussion). Journal of the Royal Statistical Society,
Series B 59, 731792.

954

Risk and evidence based policy making

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Environmental reliability as a requirement for defining environmental

impact limits in critical areas
E. Calixto & Emilio Lbre La Rovere
UFRJ-COPPE, Rio de Janeiro, Brasil

ABSTRACT: The main objective of this study is to define reliability requirements in relation to environmental
impacts in critical areas in terms of environmental resource sensitivity. Nowadays many enterprises in Brazil are
evaluated in this area in term of many different environment requirements, but the environmental impact of the
enterprise or the group of enterprises as a whole are not assessed, and nor are their future modifications.
When the number of enterprises in a specific area increases the risk of accidents is also rises. In other words
reliability over time gets worse. Unfortunately most of cases in Brazil do not take into account the entire enterprise
risk impact in a specific area and the decrease of reliability over time.
The methodology in question takes into account all the critical events which cause a serious environmental
impact for each enterprise in the same area that take place over time. By taking into account all relevant events,
it is possible to produce the Environment Diagram Block which covers all related events and their probability
of occurring over time. This means that failures in any block represent accidents with potential environment
impacts.
The environmental reliability target is associated with the tolerable number of environmental impacts in a
specific area, taking into account all events over a specific period of time. The tolerable number of accidents
depends on social perception and environmental sensitivity.
For this analysis the Monte Carlo simulation has to be carried out over a period of time in order to define the
Environmental Availability and Environmental Reliability related to the number of tolerable events. Moreover, in
the case of any enterprise modifications or an increase in the number of enterprises a new block will be inputted
in the Environmental Block Diagram and the new results will be assessed.

INTRODUCTION

Nowadays many environmental requirements are

imposed on enterprises in order to preserve environmental conditions and avoid serious environmental
impacts. In Brazil there are specific laws which stipulate specific risk analyses, procedures and waste limits
for enterprises depending on their characteristics and
potential environmental impacts.
In the Oil and Gas industry, the law is even more
strict due to previous catastrophic events, such as accidents in Guanabara Bay, the P-36 platform incident,
and so on.
Even regarding the worst possibilities of accidents
and environmental impacts, no methodology takes
into account groups of enterprises and analyzes their
impact on the environment over time. Therefore, there
is no strategy to evaluate a specific area regarding the
overall potential environmental impact.
Tolerable environmental sensitivity is related to
the ability of ecosystems to adapt or react to the
number of potential environmental impacts over time

in relation to all the enterprises in a specific area.

Therefore, based on environmental sensitivity it is possible to define a limit in terms of the number and types
of environment accidents in order to set the reliability target for a specific area taking all enterprises as a
single system. The reliability target is environmental
reliability and the protection level will be stipulated
both for each individual enterprise and for the group
of enterprises in a specific area in order to preserve
the environment over time.
To clarify the methodology a case study will be
carried out of a group of enterprises with an environmental impact in a specific area.
2

ANALYSIS METHODOLOGY

In order to perform this analysis a number of important

steps have to be followed to achieve consistent results
which can provide support for decision making about
enterprise limits and environmental reliability targets
for specific areas.

957

The first step is to discover the sensitivity of the

ecosystem in terms of environmental impacts. For this
the characteristics of the ecosystem have to be analyzed and its limits estimated. This is very difficult
in many cases and almost impossible in others due
to complex features of ecosystems. Therefore, it is
advisable that ecosystems be compared with others to
estimate limits regarding environmental impacts. In
this case, it is important to be conservative as regards
the tolerable limits of events in order to preserve the
environment.
After the environmental impact limits have been
defined, the enterprises and their potential environmental accident impacts have to be studied. In this
case historical accident data has to be analyzed and
the density probability function established to discover
accident frequency over time. In many cases accident frequency is considered constant over time, but
this is not true in all situations. This concept significantly influences the analysis because some accidents
increase in frequency in a specific period of time,
which in turn allows these accidents to be discussed
and in some cases leads to the discovery of their causes
and the consequent proposal of preventive action in
order to avoid future accidents.
The final step is group all the enterprises and simulate events over time. The Monte Carlo simulation
will be used, with the group of enterprises being represented by the Block Diagram Methodology. This
analysis requires that the group of enterprises be taken
as a single system. Each enterprise will be represented by a specific block and all blocks will be in
series. This means that in the case of accidents the
system will impact the environment and the reliability

1 Environment Sensitivity

2 Critical Events

3 Environment Diagram Block

4 Simulation

5 Critical analysis

6 Conclusion

Figure 1.

Environmental reliability analysis methodology.

of system environments will be influenced by

enterprises reliability.
After the simulation the system reliability will be
analyzed and it will be possible to discover if the target
has or has not been achieved and whether the number
of accidents is higher or not. In negative cases, it is also
possible to find out how much improvement is necessary to achieve the reliability target. The methodology
is summarized below in Figure 1.

ENVIRONMENTAL SENSITIVITY

Environmental sensitivity in some specific areas can

involve issues with social, economic and environment
impacts in the case of accidents. To facilitate the understanding of environmental sensitivity ESI maps were
drafted to serve as quick references for oil and chemical spill responders and coastal zone managers. They
contain three kinds of information:
Shorelines are ranked based on their physical and
biological character, then color-coded to indicate their
sensitivity to oiling.
Sensitive biological resources, such as seabird colonies and marine mammal hauling grounds, are
depicted by shaded polygons and symbol icons to
convey their location and extent on the maps.
ESI maps also show sensitive human-use resources, such as water intakes, marinas, and swimming
beaches.
In the USA at present project scientists have created collections of ESI maps, called ESI atlases, for
most coastal areas, including Alaska. To do this, vulnerable coastal locations have to be identified before
a spill happens, so that protection priorities can be
established and cleanup strategies identified. To meet
this need, NOAA OR&R researchers, working with
colleagues in state and federal governments, have produced Environmental Sensitivity Index (ESI) maps.
An example section from an ESI map appears in
Figure 2 below.
The Environmental Sensitivity Index (ESI) project
team has developed a systematic method for creating
ESI maps. Others are welcome to adopt this method
when it proves useful to them. This section gives
an introduction to the basic elements of ESI maps.
ESI maps include three kinds of information, delineated on maps by color-coding, symbols, or other
markings:
Shoreline Rankings: Shorelines are ranked according to their sensitivity, the natural persistence of oil,
and the expected ease of cleanup.
Biological Resources: Oil-sensitive animals, as well
as habitats that either (a) are used by oil-sensitive
animals, or (b) are themselves sensitive to spilled
oil (e.g., coral reefs).

958

Table 1.

Sensitivity rankings of marine habitats.

Sensitivity
ranking
High

Medium High
Medium
Low Medium
Low

Figure 2.

Environmental sensitivity.

Human-Use Resources: Resources and places

important to humans and sensitive to oiling, such as
public beaches and parks, marine sanctuaries, water
intakes, and archaeological sites.
The Shoreline Rankings have been defined on the
basis of factors that influence sensitivity to oiling,
including substrate grain size, permeability, trafficability, and mobility; the slope of the intertidal zone;
relative degree of exposure of the physical setting; ease
of cleanup; and biological productivity and sensitivity.
A ranking of 1 represents the shorelines least susceptible to damage by oiling, and 10 represents the locations
most likely to be damaged.
Habitat is the single most important influence on
impacts of oil in marine ecosystems (API 1985; NAS
1985). Intertidal habitats are exposed to much higher
concentrations of oil than subtidal habitats (Ballou
et al. 1987). Benthic (sea bottom) habitats are generally more affected by contact with oil than pelagic
(open water) habitats. The two key intertidal habitat
variables are exposure to wave action and substrate
(Baker 1991). Table 1 below lists sensitivity rankings
of marine habitats (excluding tropical) to oil augmented from the American Petroleum Institute (API
1985).
Sheltered habitats with fine-grained sediments are
highly sensitive whereas exposed rocky shores have a
relatively low sensitivity to oil pollution. In sheltered,
fine-grained habitats oil tends to linger whereas on
exposed rocky shores oil is subject to rapid removal

Habitat type
Saltmarsh
Sheltered Rocky Intertidal
Sheltered Rocky Intertidal
Special Use (endangered
species/marine protected areas)
Seagrass Meadow (low intertidal to
shallow subtidal)
Open Water Enclosed Bays and
Harbours
Exposed Sand/Gravel/Cobble Intertidal
Exposed Rocky Intertidal
Kelp Forest Subtidal
Open Water, Non-enclosed Nearshore
and Offshore
Soft Bottom to Rocky Subtidal

by wave action. Moreover, rocky intertidal species are

adapted to counteracting the stressful effects of desiccation and these adaptations can help them against oil.
The importance of habitat is also reflected in the Vulnerability Index and habitat recovery generalizations
(excluding tropical habitats) listed in Table 2.
In relation to Biological Resources, certain animal and plant species are especially vulnerable to the
effects of oil spills. Under the ESI method, these
species have been classified into seven categories,
each further divided into sub-categories of species
similar in their sensitivity to spilled oil. Many species
that are vulnerable to oil are wide-ranging, and may be
present over large areas at any time. These species can
be especially vulnerable at particular times and places.
ESI maps show where the most sensitive species, life
stages, and locations exist, but do not necessarily show
the entire area where members of a sensitive species
occur.
In Human-Use Resources exact locations of some
archaeological and cultural resources cannot be disclosed because of the risk of vandalism. Either these
locations are shown within a polygon enclosing a
larger area, or a map symbol is placed near to, but
not at the exact location. People using the ESI method
to map human-use resources are encouraged to denote
not only surface water intakes, but also groundwater
recharge zones and well fields.
Although this methodology defines the most critical areas in the case of accidents, it is not enough
to know the exact environmental limits of accident
impacts. Also necessary is specialist knowledge of
environmental behavior in the case of accidents or to
carry out a simulation of environmental effects. The
most usual means is to look for a similar environment area that was affected by similar accidents and to
evaluate its effects and environment behavior in order

959

of a system, subsystem or equipment to work properly

for a specific period of time. The reliability function requires historical data and uses methods like the
minimum square to establish the PDF (density probability function) that best fits the historical data. The
reliability function is as follows:

Table 2. Vulnerability index and habitat recovery

generalizations.
Vulnerability Shoreline
index 1
type
10
9

2
1

Comments

Marine
Wetlands

Very productive aquatic

ecosystems; oil can
persist for decades
Sheltered Tidal Areas of low wave
Flat
energy-high
Boulder
biological
Barricade
productivity; oil may
Beach
persist for decades
Sheltered
Areas of reduced
Rocky
wave action; oil may
Coast
persist for over a
decade
Gravel Beach
Same as Index 6; if
asphalt pavement
forms at high spring
tide level it will
persist for decades
Mixed
Oil may undergo rapid
Sand/Gravel
penetration/burial
Beach
under moderate to
low-energy conditions;
oil may persist for
decades
Exposed
Most oil not likely to
Compacted
adhere to or penetrate
Tidal Flat
the compacted
sediments
Course Sand
Oil may sink and/or be
Beach
buried rapidly; under
moderate to high-energy
conditions oil likely
removed naturally
within months
Fine Sand
Oil does not usually
Beach
penetrate far into the
sediment; oil may
persist several months
Eroding
Wave-swept; most oil
Wavecut
removed by natural
Platform
processes within weeks
Exposed Rocky Wave reflection keeps
Headland
most oil offshore

t
R(t) = 1

f (t)dt
0

Depending on the PDF function the reliability

format can differ.
The reliability concept can be used in environmental analysis in order to establish the probability of an
environment impact not occurring in a specific period
of time. It is possible to stipulate environmental reliability targets to limit the quantity of environmental
impacts and increase the level of safety protection in
one or more enterprises. Figure 3 represents the environment reliability of oil spills in Japan, in relation to
the worst events. Most of the events have occurred in
the last 30 years, due to the increase in oil transport.
The best PDF that represents the events is Gumbel,
with a correlation of 0.9484.
The remarkable aspect is that the frequency of this
event is not constant, as is usually assumed in most
risk analyses, but changes over time. The frequency
index is:
(t) =

f (t)
R(t)

Figure 4 below show the frequency index, which

is almost constant until 10 years and then starts to
increase.
Environmental reliability can be used to prioritize
critical areas or enterprises in terms of environmental impact risk, which can provide useful support for
decision making related to the allocation of emergency
resources. It should be noted that priorities will change
over time depending on the PDF of the critical events

Note: 10 = most vulnerable, 1 = least vulnerable, index

is a qualitative rank order.

ReliaSoft Weibull++ 7 - www.ReliaSoft.com

Reliability, R(t)=1-F(t)

to define limits for accident impacts. When doing this

it is necessary to be conservative in defining limits in
order to preserve environments because in most cases
environment limits are not easy to define.

Reliability vs Time Plot

1,000

Reliability

0,801

Derramamento de leo no japo

Gumbel-2P
RRX SRM MED
FMF=16/S=0
Data Points
Reliability Line

0,601

0,402

0,202

ENVIRONMENTAL RELIABILITY

0,003
0,000

The concept of reliability is known in industry and has

many different applications. It means the probability

Figure 3.

960

Eduardo Calixto Calixto

Petrobras
4/12/2007
17:08:12
10,000

20,000

Time, (t)

30,000

40,000

50,000

Environmental reliability of oil spills in Japan.

ReliaSoft Weibull++ 7 - www.ReliaSoft.com

The concept takes into account the number of environmental impacts and their duration as shown in the
equation below (D(t) = availability).

Failure Rate vs Time Plot

4,000

Failure Rate
Derramamento de leo no japo
Gumbel-2P
RRX SRM MED FM
F=16/S=0
Failure Rate Line

Failure Rate, f(t)/R(t)

3,201

n

2,401

D(t) =
1,602

i=1
n


ti
Ti

i=1
0,802

0,003
0,000

Where D(t) is environmental availability, t is the

duration of the event, T lifetime duration and i represents the time when the undesired event occurred.
It is possible in this way to represent the actions of
emergency teams to alleviate environmental impacts
and to assess their performance in relation to various aspects, such as time taken for emergency team
to reach the accident area and the duration of emergency actions to control the accident and eliminate the
environmental impact.
To represent a group of events or enterprises an
environmental block diagram is used covering the
whole system or group of events that cause the environmental impact. The main idea is to take into account
a single group of events that can occur in a specific
area or a group of enterprises which have a potential
environmental impact.
In Brazil the total environment impact of a group of
enterprises in critical environmental areas is not taken
into account. Nevertheless, the ecosystem can only
support a limited number of environmental impacts
and this number has to be taken into account when a
group of enterprises is being evaluated.

Eduardo Calixto Calixto

Petrobras
4/12/2007
17:05:02
10,000

20,000

30,000

40,000

50,000

Time, (t)

Figure 4.

Event frequency.

ReliaSoft Weibull++ 7 - www.ReliaSoft.com

Failure Rate vs Time Plot

4,000

Failure Rate
Derramamento de leo no japo
Gumbel-2P
RRX SRM MED FM
F=16/S=0
Failure Rate Line

3,201

Failure Rate, f(t)/R(t)

2,401

1,602

0,802

Eduardo Calixto Calixto

Petrobras
4/12/2007

Figure 5.

Environmental reliability of oil spills in the UK.

analyzed. Figure 5 below shows the environmental

reliability curve for different oil spill areas in the UK.
In relation to environmental reliability it can be seen
that oil spills in UKCS, represented by the red curve,
are the most reliable event along time. This means
that it is not a critical area requiring the investment of
emergency resource allocations during the first thirty
years.
On the other hand, it is also necessary to analyze what the historical data represents and if the
behavior will remain the same over time. In the UK
oil spill area, the first events are related to accidents with ships, while the final one involve platform
accidents and other types of occurrence. In this particular case, it is most advisable to consider the last
ten years.

ENVIRONMENTAL AVAILABILITY

Environmental availability is the part of total time

when there is no environmental impact in an area.

CASE STUDY

To find out the maximum number of environmental

impacts and their intensity, ecosystems and their characteristics have to be analyzed. The result is related
to environment reliability and can be estimated using
Monte Carlo Simulation Methodology. Figure 6 represents the Santos Basin which is considered to be
a sensitive environmental area. Therefore, no catastrophic events in this area are tolerable and to find out if
enterprises are reliable the Environmental Reliability
Methodology has to be followed.
There are many different types of enterprises and oil
spill resources, but in relation to drilling activities the
worst events are surface, submarine and underground
blowouts. The probability that events are constant over
time is shown in Table 3 below.
When carrying out the Monte Carlo Simulation if
it is intended not to permit any kind of catastrophic
event, the maximum level of drilling in the area has to
keep the number of catastrophic events below one.

961

Submarine
Figure 7.

Drill 1

Figure 8.

Figure 6.

Area of santos basin.

Table 3.

Blowout events.

Blowout

Probability

Underground Blowout
Submarine Blowout
Surface Blowout
Blowout

1,50E-04
1,80E-04
1,62E-06
3,33E-04

The main objective of the methodology is to estimate the period of time when the event analyzed
happen based on its PDF characteristics. In this direct
simulation it is possible to analyze the life of the
system as a whole and to find out which event occurred,
when it happened and to have an idea about its direct
and indirect impacts and the environment reliability
and availability for a specific period of time. For
instance, a group of events involving drilling activity
in the Campos Basin is represented in an Environmental Diagram Block where each block takes into
account blowout events over the lifetime of the drilling.
Each drilling activity has three catastrophic events that
are represented in Figure 7 below. The events are in
series in the Block Diagram because the occurrence of
any event represents an environmental impact on the
system as a whole.
The next step is to represent the group of enterprises
located in same area and to estimate environmental
availability and reliability limits. In the same way that
each event is in a series in an individual enterprise,
groups of enterprises are also in series. Figure 8 below
represents the drilling group of the Campos Basin. The
same block is also used for the analysis.

Surface

Underground

Drill block diagram.

Drill 2

Drill 3

Drill 4

Drill 5

Drill 6

Drill block diagram.

The Direct Simulation will take into account three

catastrophic events for all drilling activities over time.
As a result it is possible to show the number of drilling
activities in this area over a lifetime of 20 years as
illustrated in Table 4 below.
The first column represents the number of drilling
activities and the second environment availability.
Environmental reliability is given in the fourth column, while the number of expected catastrophic events
is stated in column five. Finally, column six gives the
emergency team capacity. The maximum number of
drilling activities in this area is ten because the number of expected catastrophes is lower than one and
there is one hundred percent emergency team availability. The emergency availability capacity represents
the total time that a catastrophic event takes place and
the availability of the emergency team to respond.
Unavailability to respond to events represents worse
damage to the environment.
It is important to state that drilling is only one part
of the oil production chain. Therefore, it is advisable
to take into account all platform and ship transport in
the area and related catastrophic events.
In this case the probabilities were considered to be
constant over time due to the lack of historical data.
This is an important assumption which is very influential in the analysis because it considers that the density
of events is random over time, However, this is actually
not the case, in fact blowout events have a high probability of occurring early in drilling life. Therefore,
the PDF lognormal distribution would represent better this kind of event and indicates that it is advisable
to be more careful at the beginning. Figure 9 shows
the blowout frequency index.
The other important aspect is to know which events
are the most critical to the system as a whole. In the
case in question it is the underground blowout as shown
in Figure 10 below. To be able to identify the most
critical events with the greatest impact on environmental availability, it is essential to assess environmental
availability events. In fact, the event which has the
lowest environment availability will impact more on

962

Table 4.

Monte Carlo simulation results.

Monte Carlo Simulation (20 years)

Drills

Env Avail

Env Unavail

Env Reliab

CE

ETC

EIT

1
5
10
20
40
80

99,99%
99,92%
99,84%
99,69%
99,42%
98,81%

0,01%
0,08%
0,16%
0,31%
0,58%
1,19%

96,80%
82,00%
64,40%
46,00%
25,20%
6,40%

0,03
0,20
0,40
0,74
1,41
2,86

100%
100%
100%
99,47%
99,44%
98,77%

23
141
283
523
1017
2082

ReliaSoft BlockSim 7 - www.ReliaSoft.com

lowest environmental availability will impact more on

the environmental availability system.
In some cases certain events will happen many
times but will not impact over time like others which
happen much less frequently.

Failure Rate vs Time

2,300E-6

Failure Rate
Multiplos 10 pocos
Failure Rate Line

Failure Rate, f (t) / R (t)

1,840E-6

1,380E-6

7
9,200E-7

4,600E-7

0,000
0,000

600000,000

1,200E+6

1,800E+6

2,400E+6

Eduardo Calixto Calixto

Petrobras
9/12/2007
11:18:52

3,000E+6

Time, (t)

Figure 9.

Drill blowout frequency.

ReliaSoft BlockSim 7 - www.ReliaSoft.com

RS FCI

7,071

Availability
100%

50%

5,657
0%
1 Item(s)

4,242

Underground blowout
2,828

1,414

Eduardo Calixto Calixto

Petrobras
27/6/2008
14:14:10

Figure 10.

Critical system events.

the whole system. This happens because in the Block

Diagram when a group of blocks is in a series, the system availability will be lower than the lowest block
availability, in which case the event which has the

CONCLUSION

Environmental reliability is a powerful tool to support decision making related to environmental protection, defining limits for enterprises with reliability
requirements, numbers of enterprises and establishing
the most vulnerable areas for the location of emergency
teams.
Unlike the usual methodology, it is possible
to consider a group of enterprises and critical
events in simulation over a specific period of time.
The difficultly is obtaining historical data about
events and defining environments limits for specific
areas.
In the case of emergency teams it is considered that
they will be in the correct position and that all procedures and actions will happen correctly avoiding any
delay. In real life this does not happen, therefore the
specific model has to be evaluated taking into account
the performance of emergency teams.
The remarkable point about historical data is
understanding why accidents happens and if the data
fits well enough to be used in the current simulation
case.
In this case study only drilling activities which
affected a specific area where taken into account, but in
addition all enterprises and the lifetimes of platforms
and ships have also to be considered.
The next step in the case study is to consider all
enterprise data which has an influence on environmental sensitivity in the area in question. Because of
the environmental effects of other enterprises drilling
limits will probably be reduced in order to keep
the number of catastrophic accidents lower than one
during the lifetime in question.

963

REFERENCES
A.M. Cassula, Evaluation of Distribution System Reliability Considering Generation and Transmission Impacts,
Masters Dissertation, UNIFEI, Nov. 1998.
API (American Petroleum Institute). 1985. Oil spill response:
Options for minimizing ecological impacts. American
Petroleum Institute Publication No. 4398. Washington,
DC: American Petroleum Institute.
Ballou, T.G., R.E. Dodge, S.C. Hess, A.H. Knap and
T.D. Sleeter. 1987. Effects on a dispersed and undispersed
crude oil on mangroves, seagrasses and corals. American
Petroleum Institute Publication No. 4460. Washington,
DC: American Petroleum Institute.
Barber, W.E., L.L. McDonald, W.P. Erickson and
M. Vallario. 1995. Effect of the Exxon Valdez oil spill on
intertidal fish: A field study. Transactions of the American
Fisheries Society 124: 461476.
Calixto, Eduardo; Schimitt, William. Anlise Ram do projeto
Cenpes II. ESREL 2006, Estoril.
Calixto, Eduardo. The enhancement availability methodology: a refinery case study, ESREL 2006, Estoril.
Calixto, Eduardo. Sensitivity analysis in critical equipments: the distillation plant study case in the Brazilian
oil and gas industry. ESREL 2007, Stavanger.
Calixto, Eduardo. Integrated preliminary hazard analysis
methodology regarding environment, safety and social
issues: The platform risk analysis study. ESREL 2007,
Stavanger.
Calixto, Eduardo. The safety integrity level as hazop
risk consistence. the Brazilian risk analysis case study.
ESREL 2007, Stavanger.
Calixto, Eduardo. The non-linear optimization methodology model: the refinery plant availability optimization
case study. ESREL 2007, Stavanger.
Calixto, Eduardo. Dynamic equipments life cycle analysis. 5 International Reliability Symposium SIC 2007,
Brazil.

IEEE Recommended Practice for the Design of Reliable

Industrial and Commercial Power SystemsIEEE Std.
4931997.
Kececioglu, Dimitri, and Sun, Feng-Bin, Environmental
Stress ScreeningIts Quantification, Optimization and
Management, Prentice Hall PTR, New Jersey, 1995.
Lafraia, Joo R. Barusso, Manual de Confiabilidade,
Mantenabilidade e Disponibilidade, Qualimark, Rio de
Janeiro, Petrobras, 2001.
Monteiro, Aline Guimares. Metodologia de avaliao de
custos ambientais provocados por vazamento de leo:
O estudo de caso do complexo REDUC-DTSE. Rio de
Janeiro, 22/12/03COPPE/UFRJ.
Moraes, Giovanni de Araujo. Elementos do Sistema de
Gesto de segurana meio ambiente e sade ocupacional.
Gerenciamento Verde Consultoria Rio de Janeiro: 2004
R. Billinton, and R.N. Allan, Reliability Evaluation of
Engineering Systems: Concepts and Techniques, 1st
Edition, Plenum Press, New York, 1983.
ReliaSoft Corporation, Weibull++ 6.0 Software Package,
Tucson, AZ, www.Weibull.com.
Rolan, R.G. and R. Gallagher. 1991. Recovery of intertidal
biotic communities at Sullam Voe following the Esso
Bernica oil spill of 1978. Proceedings of the 1991 Oil
Spill Conference, San Diego. American Petroleum Institute Publication No. 4529: 461465. Washington, DC:
American Petroleum Institute.
W.F. Schmitt, Distribution System Reliability: Chronological and Analytical Methodologies, Masters.

964

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Hazardous aid? The crowding-out effect of international charity

P.A. Raschky & M. Schwindt
University of Innsbruck, Austria and alpS, Centre for Natural Hazard ManagementLtd., Innsbruck, Austria

ABSTRACT: Research suggests that public support for natural hazard mitigation activities is distorted due to
choice anomalies. For that reason, preparedness measures are often implemented to an insufficient extent. From
an ex-post perspective the lack of mitigation might result in a necessity for a risk-transfer. On the other hand,
based on the conclusions from the Samaritans Dilemma, the anticipation of relief in case of a disaster event
might induce individuals to diminish ex-ante protection activities.
In order to analyze the existence of this phenomenon in an international context, this paper discusses the
impact of expected foreign aid in case of a natural disaster on the level of disaster mitigation activities. The
results suggest that foreign aid in previous disaster years implies future ex-post charity and thus does crowd
out risk-management activities. The paper concludes with propositions on the enlightenment of natural hazards
aiming to counter the crowding-out of prevention.

INTRODUCTION

The recurrent pictures of disaster victims usually stem

from developing or emerging countries and leave the
mark that these countries are more frequently hit by
natural disasters. However, according to Kahn (2005)
developing countries do not experience significantly
more natural disaster events; nonetheless countries
with lower levels of development suffer ceteris paribus
more victims from natural catastrophes than developed countries. Out of all large-scale natural disasters,
epidemics, NA-TECH accidents, famines and technological accidents between 1994 and 2004, 79% general
catastrophes happened in countries receiving OECD
aid (see Table 1). In fact, research suggests that apart
from low GDP per capita, weak quality of institutions
and high income inequality are good indicators for
countries which suffer more casualties after a disaster
(Anbarci et al. 2005, Kahn 2005).
In order to limit both, the financial as well as
the physical consequences of natural hazards the
implementation of ex-ante mitigating strategies i.e.
risk management is important. Since these activities
require detailed knowledge about the extent of the risk
and the technical possibilities to mitigate such a risk,
they are hard to put into action for individuals. But
oftentimes, the failure of sufficient preparedness starts
earlier in the decision process and is based on a lack
of risk awareness (Kunreuther & Pauly 2006).
Foreign aid usually flows to a great extent after
catastrophes occur, aiming to moderate the consequences. Although it is paid ex-post, it can be used for
the catastrophe management, i.e. for medical support

of the people relevant as well as for the reconstruction of infrastructure. It is meant to limit financial as
well as physical losses. On the other hand, based on
the conclusions from the Samaritans Dilemma, the
anticipation of foreign aid in case of a disaster might
induce people to diminish ex-ante protection activities
(Buchanan 1975, Coate 1995), especially in the case of
natural disasters where the probability of occurrence is
relatively low and the danger is underestimated. The
question arising is, whether the provision of foreign
aid can induce adverse effects in form of a reduction
of ex-ante mitigating activities.
In order to build up sustainable development strategies and to support less-developed countries an analysis of a countrys dependency on foreign aid and its
vulnerability against large-scale disasters is essential.
Both, a theoretical as well as an empirical analysis of
this relationship is still missing.
The remainder of the paper is structured as follows: The next section presents the theoretical background of our analysis. The framework incorporates
the ideas of two fields in the literature on the economics of natural hazards. The first string consists
of the ideas of choice heuristics in decisions on lowprobability-high-loss events based on the theories
by Kahneman & Tversky (1974). The second area
touched is the phenomena of charity hazard, an adoption of the Samaritans dilemma (Buchanan 1975,
Coate 1995) to insurance decisions for disaster events.
In Section 3 the connection of charity hazard and
foreign aid is illustrated. In section 4 preliminary
descriptive statistics and a summary of the results of
the study conducted by Raschky & Schwindt (2008) on

965

Table 1.

Large-scale disasters (19942004) and Development AID.

No.

Country

Year

Type

No. Killed

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

Korea Dem P Rep

Indonesia
Sri Lanka
Venezuela
Iran Islam Rep
Italy
India
France
Turkey
India
Spain
Honduras
India
Germany
Thailand
Japan
Afghanistan
Nigeria
Burkina Faso
Viet Nam
China P Rep
Nicaragua
Niger
India
China P Rep
Haiti
Portugal
Haiti
India
Afghanistan
Afghanistan
Somalia
Burkina Faso
Algeria
Taiwan (China)
Papua New Guinea
Tanzania Uni Rep
India
Zaire/Congo Dem Rep
Russia

1995
2004
2004
1999
2003
2003
2001
2003
1999
2004
2003
1998
1999
2003
2004
1995
1998
1996
1996
1997
1998
1998
1995
1998
1996
2004
2003
2004
1998
2002
1998
1997
1997
2003
1999
1998
1997
1994
2002
1995

Famine
Tsunami
Tsunami
Flood
Earthquake
Heat wave
Earthquake
Heat wave
Earthquake
Tsunami
Heat wave
Hurricane
Cyclone
Heat wave
Tsunami
Earthquake
Earthquake
Meningitis
Meningitis
Typhoon
Flood
Hurricane
Meningitis
Cyclone
Flood
Hurricane
Heat wave
Flood
Heat wave
Unknown
Earthquake
Flood
Meningitis
Earthquake
Earthquake
Tsunami
Diarrhoeal
Flood
Respiratory
Earthquake

610,000
165,708
35,399
30,000
26,796
20,089
20,005
19,490
17,127
16,389
15,090
14,600
9,843
9,355
8,345
5,297
4,700
4,346
4,071
3,682
3,656
3,332
3,022
2,871
2,775
2,754
2,696
2,665
2,541
2,500
2,323
2,311
2,274
2,266
2,264
2,182
2,025
2,001
2,000
1,989

OECD
Total Aid

OECD
Emerg. Aid

4.72
1,886.34
540.29
56.58
111.73

0.83
14.49
83.37
4.02
36.9

1,186.97

79.29

822.19
2,427.62

264.65
8.48

222.41
1,165.99

35.37
17.4

581.95

9.4

102.1
51.94
218.5
1,256.58
2,358.79
427.31
131.25
877.98
3,423.18
291.57

59.24
5.08
1.54
5.87
14.02
34.79
2.14
2.64
7.62
56.94

291.57
877.98
1,291.94
102.1
47.85
198.82
317.2

56.94
2.64
654.39
59.24
34.27
3.96
16.39

182.18
640.23
1,984.16

18.31
11.54
1.7

Data Source: EM-DAT, CRED Brussels (2008).

in Mio. USD (2004 PPP).

the relationship of foreign aid and earthquake fatalities

is presented. Finally, section 5 concludes.
2
2.1

THEORETICAL BACKGROUND
Choice anomalies

Whenever natural hazard events cause losses of human

lives or destroy capital, the question of obligation has
to be answered. From an ex-post perspective insufficient early warning systems and mitigation activities are usually in the centre of criticism. But are

preparedness activities, which cause costs, in fact

desired by the society from an ex-ante point of view?
In order to answer this question it is first necessary to
understand how people perceive risk.
Addressing this problem Slovic et al. (1984) ask:
How should a single accident that takes N lives be
weighted relative to N accidents, each of which takes a
single life? In other words: Do people value the severity of an event, measured by the death toll, higher than
the probability of events occurring in their decision
process? The authors show, that in contrast to thenresearch suggestions, it is not the severity, but rather

966

the frequency of occurrence, which dominates the risk

perception of individuals.
Based on this result, a society should undertake
more risk management activities for natural hazards,
which are more probable, and less for the more
unlikely ones. Kahneman & Tversky (1974) point
out the complexity of the problem by arguing that
individuals decisions are subject to choice anomalies. This theory proposes that the standard expected
utility theory does not sufficiently describe and predict individual behaviour under uncertainty (Frey &
Eichenberger 1989). When it comes to natural hazards,
individuals do not base their decisions on calculated
probabilities, but rather use inferential rules known
as heuristics (Kahneman & Tversky 1974). This suggestion has been clearly applied on the market for
natural hazard insurance by Kunreuther (2000) who
defined the situation as natural disaster syndrome
which is a term that links the lack of interest by
those at risk in protecting themselves against hazards and the resulting significant financial burden on
society, property owners, the insurance industry and
municipal, state and federal governments when severe
disasters do occur. He points out that five heuristics
are responsible for anomalies on the natural disaster
insurance market. One main reason is connected to
information biases. Individuals misperceive the risk of
natural disasters, because of extensive media-coverage
(availability bias) or they tend to overestimate risks
of being harmed by a natural hazard that has recently
occurred. A second very typical heuristic in the area
of natural hazard insurance is the common attitude:
It wont happen to me!1 . For example, a mountain
farmer, who has been living his whole live in an area
with high avalanche risk (red zone)2 , where almost
every year avalanches impact next to his farm. Nevertheless, he has no incentives to either move away or
to insure his farm against potential losses. The third
heuristic refers to the role of emotions connected to
catastrophic events. Individuals may purchase insurances because of feelings such as fear instead of the
outcome of a cost-benefit analysis. Heuristic number four originates during the risk communication
process. The form how risks and probabilities are
presented can have a huge impact on individuals decisions. For example, the actual danger to die from an
avalanche is perceived to be moderate at the avalanchelevel 33 . The fact that at this stage the highest number

1 For

additional information see Weinstein (1984).

hazard zone system distinguishes between, blue,
yellow and red zones.
3 The avalanche risk is expressed by a system starting at
level one, meaning nearly no risk and ending at level
five meaning extreme risk.

of avalanche victims occurs, changes the individuals

attitude towards avalanche risks tremendously. The
fifth heuristic is concerned with the ambiguity about
the probability that a natural disaster might occur. This
vague probabilities lead to inefficiencies on the private insurance market. A publication by Kunreuther &
Pauly (2004) included this idea into a formal model
of decision making costs under imperfect information
and showed that individuals still refuse to purchase
natural hazard insurance even if the premiums are
attractive. The authors show that the demand-side inefficiency is a problem of a) transaction costs in order
to obtain information and b) ambiguity about probability estimations by different insurance companies.
The search for the optimal insurance imposes costs
which are high enough to discourage the individual
to engage in any further mitigation activity4 . Additionally the insurance premiums are likely to be much
higher, because of vagueness about the probabilities.
Returning to the initially point of interest, individuals preferences concerning risk mitigation, the above
mentioned results imply that this decision is more complex than expected and dependent on a multitude of
features. The probability of occurrence indeed perceived by the individuals is likely to deviate from
the actual value, which implies a suboptimal level of
mitigation activities.
The above presented results suggest that insufficient mitigation is a consequence of decision anomalies. In order to get a clearer picture of peoples
preferences for risk mitigation, Flynn et al. (1999)
interview 400 inhabitants of Portland/Oregon about
their assessment of earthquake risk and their willingness to support the implementation of more rigorous
building-codes aiming to reduce seismic risk. The
results suggest that contrary to expectations, people are well informed about earthquakes and aware
of seismic risk. Nevertheless, public support is primarily aimed at public facilities. The willingness to
support privately owned buildings in contrast is very
low. Moreover, people favour to protect buildings
with great amount of people at risk and buildings,
where especially vulnerable people are at risk in comparison to buildings, which are most at risk. The
question arising is Why are people willing to protect one building, but not the other, although they
recognize the level of risk? A more precise analysis of their choice suggests that support is especially
given for protecting buildings, where a large number
of people are affected. In other words, people seem
to be more risk averse when a multitude of human
lives are in danger. Else wise they are rather willing to

2 The

4 For

additional information and an illustrative example

see Kunreuther & Pauly (2004).

967

take the risk. From an ex-post perspective this uneven

behaviour towards risk might result in a necessity for
a risk-transfer.
2.2

The Samaritans Dilemma

The question in the centre of this work is, whether

the willingness to take risk in certain situations,
is enforced by the fact, that people anticipate the
possibility of ex-post relief. So far this phenomenon
is mostly mentioned in the context of natural hazard
insurance. Raschky & Weck-Hannemann (2007) use
a theoretical model, to demonstrate that the possibility of governmental ex-post relief influences the
demand for insurance in a negative way. The reason is that ex-post relief is only paid, if individuals
do not have an insurance coverage. Else wise the
insurance company is responsible for the financial
compensation. A rational individual will reduce his
demand for insurance and hence his ex-ante costs, in
order to receive charity. The crowding-out of insurance demand due to the reliance on ex-post relief can
be interpreted as an especial case of the moral hazard
problem and is called charity hazard (Browne & Hoyt
2000). Buchanan (1975) uses the term Samaritans
Dilemma, to point a donors situation up: Since people
trust in altruism, even the commitment not to help, is
implausible. In case of need potential donors might
have a guilty conscience. Hence altruism is expected
to eclipse the threat. In order to evade this problem,
Coate (1995) suggests donations to take the form of
in-kind transfers.
But do people behave in a similar way, when the
point of interest is not insurance coverage, but rather
mitigation activities? Before trying to answer this
question, it is important to classify mitigation activities
in comparison to demand for insurance. The latter can
be categorized as a private good. Individuals decide
to demand insurance subject to their risk perception
and their readiness to assume a risk5 . In contrast, the
former is usually financed by taxes, and shows more
features of a public good, since mitigation activities
are hardly to implement by individuals, because of
insufficient information and budget restrictions. The
public good characteristic also implies a certain degree
of redistribution, resulting from the fact that a) the
benefit from supporting mitigation activities does not
necessarily cover the expenses of each individual and

5 Obligatory insurance systems correct the deviations in

insurance demand, which result from the above mentioned
heuristic as variations in risk perception. For example
several cantons in Switzerland dictate a compulsory insurance in order to protect buildings from the danger of
natural hazards as flood events.

b) income inequality exists6 . Hence, in accordance to

the results of Anbarci et al. (2005) and Kahn (2005),
a lack of support for natural hazard preparedness can
also be the result of limited willingness to redistribute.
Moreover, it is important to notice that a reduction
in insurance coverage is equivalent with the willingness to accept a financial risk. In contrast a reduction
in natural hazard preparedness would imply the willingness to accept a loss in monetary terms as well as
a loss of human lives. Although it seems counter intuitive, that people would undergo such a risk, the above
mentioned results of Flynn et al. (1999) provide some
evidence for such behaviour.
Note, that the first difference mentioned, namely the
public good character of preparedness and the limited
willingness to redistribute, would suggest less support
for mitigation activities in comparison to demand for
insurance, whereas the second feature implies a higher
support for mitigation activities compared to insurance demand, since mitigation does not only reduce
financial risks, but also human risks.
Although the explanations above show significant differences between insurance and mitigation
and therefore suggest, that the substitution between
insurance and mitigation in the charity hazard argumentation is not straightforward, section 3 aims to
explain how the phenomenon of charity hazard can
be connected to the public support for mitigation
activities.

THE ADVERSE EFFECT OF FOREIGN AID

So far, the focus of analysis has been on a national

level. Individuals decide to reduce mitigation activities and to rely on ex-post charity from the government.
But the huge amounts of foreign aid flows, would suggest a rational for charity hazard in an international
context as well. In fact, we believe that countries which
have received foreign aid in the past, feel some kind of
connection to the donating country. Since these countries expect the donating country to help in case of
natural hazards, they reduce or even forgo mitigation
activities. This effect is strengthened by the fact that
a large part of foreign aid flows due to strategic reasons (Alesina & Dollar 2000 and Frey & Schneider
1986). Hence, receiving countries might feel implicitly insured by the strategic connection to a donating
country.
Imagine a country with potential risk of experiencing a natural hazard and assume that preparedness
is financed by the implementation of a proportional

6 We

disregard the possibility of a lump-sum tax in this

context.

968

tax rate. Individuals express their preferences about

mitigation activities by vote. What are the characteristics, which influence the individual in the decision
process? First, we know from section one, that the
perceived probability of risk event influences the individuals decision. Second, the individual will take into
account, that there is no control about the way the tax
revenue is used. This means, that it can be used for the
protection of buildings, which from a taxpayers point
of view might not be worth protecting. Moreover, as
mentioned above, the implementation of taxes implies
some degree of redistribution. Hence richer people
will bear a larger part of the costs in absolute terms than
poorer people. The larger the necessary tax rate chosen by society, the higher is the redistribution ceteris
paribus. And finally, the main feature relevant for the
charity hazard theory, the existence of possible ex-post
charity, which in an international context would stem
from the international community of states. Similar
to the case of insurances, ex-post foreign aid is only
paid, if there is a necessity for it. But the exigency will
probably not arise when sufficient mitigation activities
are implemented. On the other hand the implementation causes costs. Taking into account, that individuals
underestimate the probability of occurrence of natural hazards, there would be an efficiency rational for
reducing protection in anticipation of ex-post relief.
Furthermore, we believe that the willingness to bear
the risk is higher, if that country has already experienced some kind of international aid in the past, which
increases the trust in the community of states.
4

Figure 1. Distribution of disaster fatalities in countries

receiving DAC and in countries not receiving DAC.
Data Source: EM-DAT, CRED Brussels (2008).

EMPIRICS

Thus far, literature on natural hazard management

has delivered a variety of definitions for vulnerability. Mainly for reasons due to available data we define
a nations vulnerability in the broad way by limiting the death toll from natural disasters. Figure 1
reveals that about 92% of all disaster deaths happen
in aid-receiving countries.
One key issue for an econometric analysis of the
impact of foreign aid on a countrys vulnerability is
gaining data on the magnitude of the natural process itself. In contrast to climatological disasters (e.g.
floods or hurricanes), the magnitude of earthquakes
is captured by the Richter scale. This allows us to
identify and separate the impact of the disasters magnitude on the death toll and the effect of man-made
institutional rules and the resulting incentives on the
ultimate extent of a catastrophe. In addition, earthquakes take place in numerous countries with varying
levels of GDP, development aid and institutional rules.
Figure 2 is the graphical representation of data
from a recent study by Columbia University and the
Worldbank (Dilley et al. 2005). It represents the

Figure 2. Distribution of earthquake mortality.

Data Source: Dilley et al. 2005.

geographical distribution of mortality risk from earthquakes worldwide. The darker the grid cells the larger
the mortality risk on a 10-point scale. Earthquakes
are already used in the existing economic literature to
analyse the effects of income and institutions (Kahn
2005) as well as inequality (Anbarci et al. 2005) on a
countrys vulnerability.
In a subsequent study, based on this paper,
Raschky & Schwindt (2008) augment existing theoretical and empirical literature on the determinants a
nations vulnerability against quakes by identifying the
effect of foreign aid. They collected data on the magnitude and outcomes of 335 major earthquakes occurring
worldwide between 1980 and 2002. Table 2 presents
the geographical distribution of the earthquakes in the
sample.
Raschky & Schwindt (2008) apply standard OLS
and 2SLStechniques to estimate the effect of foreign aid on the extent of earthquakes. They regress

969

Table 2.

this issue, they instrument the lagged foreign aid using

the countrys colonial past and the lag of openness for
trade. These instruments are commonly used in the
literature and were introduced by Alesina & Dollar
(2000). The significance, sign and size of the controls
do not largely change the size of the coefficient for
total aid changes marginally and stays robust and is
even larger than in the OLS estimates.

List of countries and earthquakes in sample.

Country

No. of
quakes

Country

No. of
quakes

Algeria
Argentina
Australia
Bangladesh
Bolivia
Brazil
Chile
China, P Rep.
Colombia
Costa Rica
Cuba
Ecuador
Egypt
El Salvador
Fiji
Greece
Guatemala
Honduras
India
Indonesia
Iran, Isl. Rep.

8
2
3
5
3
1
5
52
11
7
1
7
3
4
2
16
7
1
14
44
43

Italy
Japan
Kenya
Malawi
Mexico
Nepal
New Zealand
Nicaragua
Pakistan
Panama
Papua New Guinea
Peru
Philippines
South Africa
Spain
Taiwan (China)
Tanzania, Uni Rep.
Turkey
United Kingdom
United States
Venezuela

15
19
1
1
16
1
1
1
15
1
7
18
10
3
1
2
3
22
1
18
5

the number of disaster victims on foreign aid, GDP

p.c., inequality (measured by the GINI-coefficient)
and a number of a number of institutional variables
(Government effectiveness, regulatory quality, political stability and ethnic fractionalization). In addition they control for the quakes magnitude and the
countrys exposure to earthquakes.
They find that past foreign aid c. p. significantly
increases the number of earthquake fatalities. A 10%
increase in foreign aid increase the death toll by 1.3%.
The estimated coefficients of their control variables
largely accord to the econometric results in the literature. The major driving forces are GDP and the
magnitude of the quake. The empirical results largely
confirm the hypothesis that foreign aid in previous
periods has an increasing effect on the number of
earthquake fatalities. A higher level of GDP and institutional quality decreases the death toll from quakes.
The magnitude of the geo-physical process has a highly
significant impact and is the key variable determining
fatalities; the results suggest that a countrys overall
exposure to quakes does not seem to have a significant
impact on the death toll.
Regulatory quality as an indicator for institutions
becomes significant. After the OLS estimates they
address the potential endogenity of foreign aid. It is
likely that a country that has suffered from major
earthquakes in the past received more foreign aid in
previous periods. In a first step towards addressing

CONCLUSION

The aim of this work was to illustrate that the altruistic idea of foreign aid can result in an aggravation
for aid receiving countries, if the incentives created
by aid distort the decision about prevention. This
change to the worse increases the vulnerability of
recipient countries and can have disastrous effects
if the country indeed experiences a natural hazard
event.
In order to counter the reduction in prevention
the causes of the distortion should be identified and
combated. The analysis of the present work suggests
three features of the decision process, which could be
improved.
First, the five heuristics imply, that it is important
to inform people about the risk of natural hazards and
possible consequences. To make sure, that people are
not overstrained with the information, it seems plausible to limit the information to risk types a society is
really confronted with. The wrong perception of the
probability of occurrence is one reason, why people,
even without the consideration of foreign aid, dont
support mitigation activities to a sufficient extent.
Enlightenment about hazard risks might be a first step
to counteract the misperception.
Second, the survey results of Flynn et al. (1999)
suggest, that the reduction in prevention can result
from a mistrust of taxpayers in governmental decisions. Politicians could try to decrease the lack of
reliance by getting more information about tax payers
preferences and if possible consider this information
in their decision process.
The first two points just described are valid for the
overall problem of insufficient support for risk mitigation activities. In order to reduce the adverse effects of
foreign aid in special as a direct result of our approach,
the way of foreign aid distribution should we reconsidered. Important questions which need to be answered
in this context are whether to favour ex-ante or
ex-post financial aid and whether to choose financial
or in-kind transfer. These questions are hardly new for
the foreign aid literature, but the connection to natural
hazards is missing so far.

970

REFERENCES
Alesina, A.F. & Dollar, D. 2000. Who gives foreign aid
to whom and why? Journal of Economic Growth 5 (1):
3363.
Anbarci, N., Escaleras, M. & Register, C.A. 2005.
Earthquake Fatalities: The Interaction of Nature and Political Economy. Journal of Public Economics 89 (910):
19071933.
Browne, M.J. & Hoyt, R.E. 2000. The demand for flood insurance: Empirical evidence. Journal of Risk and Uncertainty
20 (3): 291306.
Buchanan, J.M. 1975. The Samaritans dilemma. In
E.S. Phelps, ed., Altruism, morality and economic theory:
7185. Russell Sage Foundation.
Coate, S. 1995. Altruism, the Samaritans dilemma, and government transfer policy. American Economic Review 85
(1): 4657.
Dilley, M., Chen, R.S., Deichmann, U., Lerner-Lam, A.L.,
Arnold, M., Agwe, J., Buys, P., Kjekstad, O., Lyon, B. &
Yetman, G. 2005. Natural disaster hotspots: A global
risk analysis, Disaster risk management series no. 5. The
World Bank and Columbia University.
Flynn, J., Slovic, P., Mertz, C.K. & Carlisle, C. 1999. Public support for earthquake risk mitigation in Portland,
Oregon. Risk Analysis 19: 205216.
Frey, B.S. & Schneider, F. 1986. Competing models for
international lending activity. Journal of Development
Economics 20 (2): 225245.
Frey, B.S. & Eichenberger, R. 1989. How important are
choice anomalies for economics? Jahrbuch der Nationaloekonomie und Statistik 206 (2): 81101.
Kahn, M.E. 2005. The death toll from natural disasters: The
role of income, geography and institutions. The Review of
Economics and Statistics 87 (2): 271284.

Kahneman, D. & Tversky, A. 1974. Judgement under

Uncertainty: Heuristics and Biases. Science 185:
11241131.
Kunreuther, H. 2000. Strategies for dealing with largescale and environmental risks. In H. Folmer, L. Gabel,
S. Gerking & A. Rose, (ed.), Frontiers in Environmental
Economics: 293318. Edward Elgar.
Kunreuther, H. & Pauly, M. 2004. Neglecting disaster: Why
dont people insure against large losses? The Journal of
Risk and Uncertainty 28 (1): 521.
Kunreuther, H. & Pauly, M. 2006. Rules rather than discretion: Lessons from Hurricane Katrina. Journal of Risk and
Uncertainty 33: 101116.
Raschky, P.A. & Weck-Hannemann, H. 2007. Charity hazarda real hazard to natural disaster insurance? Environmental Hazards 7 (4): 321329.
Raschky, P.A. & Schwindt, M. 2008. Aid, catastrophes
and the Samaritans DilemmaThe Humanitarian way
of paving the road to the apocalyptic horsemen? mimeo.
University of Innsbruck.
Slovic, P., Lichtenstein, S., & Fischhoff, B. 1984. Modeling the societal impact of fatal accidents. Management
Science 30: 464474.
Weinstein, N.D. 1984. Why it wont happen to me: Perceptions of risk factors and susceptibility. Health Psychology
3: 431457.

971

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Individual risk-taking and external effectsAn empirical examination

S. Borsky & P.A. Raschky
University of Innsbruck, Austria and alpS, Centre for Natural Hazard Management, Ltd., Innsbruck, Austria

ABSTRACT: Thousands of ski enthusiast surge into alpine ski resorts every year to indulge their passion and
seek sensation and thrill. This sensation-seeking by the individual skiers, an increasing density on the slope and
heterogeneity of skills results in around 60.000 accidents per annum on Austrian ski runs alone. The recent trend
towards adventure seeking activities in the leisure time and the accompanied adverse effects for individuals and
society call for appropriate adoption strategies by public policymakers and the relevant private industries alike.
An appropriate approach, from an economic perspective, is to internalize the external costs of private risk-taking
behaviour. We use data of 71 Austrian Ski resorts and 3,637 reported ski accidents to estimate the marginal
effect of one additional skier on the amount of accidents using standard OLS-methods. These estimates build the
basis for the evaluation of the external costs of risk-taking behaviour in alpine skiing. Based on this evaluation
instruments can be designed to reduce socially sub-optimal demand for risk-taking activities.

INTRODUCTION

The tendency of the population for participating in

thrill and adventure seeking activities in the leisure
time has been constantly growing over the last decades.
More and more people decide to seek fulfilment in
activities like rock-climbing, sky-diving or mountainbiking. This rather young development in leisure
activity can be related to two major driving forces:
Social acceptance and technological innovation. During the first half of the last century risk-taking activity
was regarded as counterphobic behaviour and an
expression of death wish. The predominant opinion of
psychologists in the 1930s was that risky-sports helped
the participants to cope with fears about masculine
inadequacy (Slangler & Rudestam, 1997). By the late
1950s the dominant explanation for high-risk activity was sensation-seeking and the individual need for
excitement and thus provide impetus for more acceptance within society (Slangler & Rudestam, 1997).
In recent years high-risk sports received increasing
attention by the media. Sport shows dedicate more
airtime to events in rock-climbing, sky-diving or
surfing. Commercials increasingly use either highrisk sportsmen or images of high-risk activity as an
advertisement medium.
This development has gone hand-in-hand with a
remarkable series of technological improvements in
medicine that have reduced the cost of risk-taking.
In other words what once was a fatal, sometimes
disabling, sport injury can now be relatively easily
repaired. In the view of risk, extreme sports

today are not really so extreme than in former days

(Johnston, 2003). Figures on the number of people
exercising high-risk sports underpin these findings in
the literature and reveal the booming trend in risky
sports: For example the number of active skydivers in
the United States Parachuting Association increased
from about 20,250 in 1991 to about 34,000 in 2005,
an increase of about 68%1 . Experts estimate that in
1959 there were about 5,000 wave-riders worldwide.
This number apparently escalated to about 2 million
by 1962 and it is estimated that today there are about
17 to 23 million surfers worldwide2 .
Psychologists argue that risk-taking activities such
as sky-diving, skiing or speeding with your car are
motivated by an individual need for excitement and
stimulation (Slangler & Rudestam, 1997). From an
economists point of view, risk-taking and sensationseeking increase individual utility and thus result in
a positive demand for these goods. This individual
demand to feel excitement through risk-taking activities, however, can result in adverse side-effects. For
example, around 60,000 accidents (resulting in some
form of injury) on Austrian ski runs are recorded
every year3 . In comparison in the same season about
1 http://www.parapublishing.com/sites/parachute/resour-

ces/statistics.cfm.
2 http://www.fff.org/freedom/fd0607f.asp.
3 Freizeitunfallstatistik

2006, Kuratorium fr Verkehrssicherheit (Statistic of accidents in spare time 2006),

http://www.kfv.at/

973

40,000 car accidents occurred on Austrians roads4 .

In the case of negative externalities (e.g. the probability of causing an accident) a socially sub-optimal
amount of the good is demanded, in our case risktaking behaviour. In this context risk taking behaviour
can be defined as an activity, which contains for example riding with high-speed, riding on ski runs with a
higher difficult level than normally chosen and ignoring the fading out of their personal skills. This shows
the relevance of the external effects of risk-taking
behaviour and the accompanying issue on safety on
Austrian ski runs. The growing number of participants
in risky sports and its related negative effects has drawn
a lot of attention from social sciences in general and
economics in particular on risk-taking behaviour5 .
This work is divided into three parts. The first
section deals with the concept of risk taking from
a psychological as well as an economic viewpoint.
Beyond that the external effect from risk-taking is
being discussed. In the second part of this work the
external effect of such behaviour is being investigated.
Using data on collision accidents in 71 major Austrian
ski resorts we estimate the marginal effect of one additional skier on the amount accidents using standard
OLS-methods. The third part concludes.

CONCEPT OF RISK TAKING

In its general meaning risk is associated with potential

harm and uncertainty as a consequence of a certain
behaviour or event. Basically in most contexts risk
is seen as something negative. Only few would willingly choose greater exposure to risk of a loss, where
loss can be defined in a broad range from the lost of
personal wealth to the lost of health. In this context
risk is something that should be managed, mitigated
or avoided. But in some contexts risk can have a positive meaning as well. In cases of adventure, bravery,
heroics and entrepreneurship risk is willingly taken
and risk-taking is high esteemed by the society. Schtz,
Wiedemann and Gray (2000) differentiate in their work
between life-style risks and environmental risks. Lifestyle risks are described as risks, which are under
control of the acting individual and which are depending on the decisions of the individual. These are risks
associated with car-driving, smoking, consumption of
alcohol or risk-sports. Environmental risks are related

4 Verkehrsunfallstatistik 2006, Kuratorium fr Verkehrssicherheit (Statistic of road accidents 2006),

http://www.kfv.at/.
5 For a general overview see Zuckerman (1979); on the
economic analysis of risk-taking behaviour see for example Schoemaker (1993).

to technological developments, like nuclear energy or

chemical industry, and their products, like pesticides,
or are potential risks in context with natural processes,
like earthquakes.
The concept of risk-taking is defined as behaviour
toward a risky choice or situation in which the risk
of a potential negative outcome is willingly taken.
Risk-taking can be divided into two different concepts.
One the one hand risk-taking behaviour can be seen
as a socially unacceptable behaviour, like speeding
or driving under influence, in which no precautions
are taken against the potential negative outcome. In
another context risk-taking can be seen as a socially
accepted behaviour, like competitive sports, skydiving or skiing, in which the risk of a potential negative
outcome is appreciated and willingly taken (Turner,
McClure and Pirozzo 2004). The main focus of this
work lies on the concept of risk-taking as a socially
accepted behaviour.
The question why people are willingly choosing
activities with a potential negative outcome and how
to observe risk-taking from a persons actual activities
is a widely discussed subject in the scientific communities. In psychological theory risk-taking is highly
context depending and is determined by the function
containing task, peoples decision frames and their
information process (Kahneman & Tversky 1974). In
search for motives why the individual is doing a risky
activity the psychological literature demonstrates a
broad range of different approaches. Fenichel (1939)
makes a distinction between counterphobic and phobic personalities. He describes the counterphobic as
an individual, who tries to overcome his fears by
incurring situations he is in fear of. In contrary the
phobic individual tries to avoid such situations. In the
case of sports this means that a counterphobic individual, who is in fear of appearing weak, chooses a
sport in which a lot of strength is needed. Though
you can find examples, which back this approach, it
is more an exemption (Rettenwander, 2003). Eyseneck (1973) differentiates between introverted and
extroverted individuals. In his work he describes the
introvert as an individual, who doesnt like excitement
and tries to avoid risk. The extrovert, in contrary, is
an individual, who is unworried, unconcerned and
optimistic and likes to take the risk. But empirical
investigations of the extroversion/introversion theory
dont provide clear results and show that the relation
between extroversion and tendency to take risk is not
as strong as assumed (Rettenwander, 2003). A widely
used approach to explain risk-taking is the concept of
sensation-seeking by Marvin Zuckerman6 . Sensation
Seeking is defined as a personal trait. It means that

6 see

974

for example Zuckerman (1979).

every individual needs a certain amount of stimulus

of body and mind. Individuals with a need of a high
stimulus are called sensation seekers. They search
for a satisfaction of this need in unusual situations,
like skydiving, and dislike repetitions and monotony.
In other words you can say that sensation seekers
are individuals, who are in search of the so called
kick. The results of empirical investigations of this
concept in relation to risk-taking behaviour are quite
divergent. Some studies can demonstrate an empirical
relationship between risk-taking and sensation seeking
(Jack & Ronan 1997, Cherpitel et al 2001) whereas
others show that there is no significant relationship
(Horvath & Zuckerman 1992). Weinstein (1984) has
formulated a concept, called unrealistic optimism,
which could help to explain the motives for risk-taking.
He has worked out that people tend to underestimate
the potential negative outcome of health-risks. This
attitude is also called It wont happen to me attitude
and implies that people think from themselves that
they are less exposed to health risks than other. This
approach is used among other to explain the motives
for speeding or not using the seatbelt.
In the economic theory the concept of risk-taking
is mainly investigated in relation to economic situations, i.e. gambles, insurance and market behaviour.
The aspect of risk-taking in high-risk sports is a nearly
neglected one from an economic viewpoint. Traditional economists assume that risk-taking is specified
by the underlying individual utility function, which
can be seen as a trait that determines behaviour in
risky situations. This approach has its origin in the
expected utility theory by von Neumann & Morgenstern (1947) and by Friedman & Savage (1952). The
application of the expected utility theory on risky decisions is based on three assumptions. The first is that the
overall utility is the expected utility of its outcomes.
The second assumption is that the final state is the
domain measured by the utility function rather than
gains or losses of the single bundles in the utility function. The perhaps best known generalization of the
expected utility theory is the assumption of the concavity of the utility function, i.e. risk aversion. This
means that a risk averse individual prefers the certain
outcome to any risky outcome with the same expected
value. Basing on the concept of utility function, generalized utility models have been constructed to derive
similar results. The advantage of these models are that
they allow nonlinear probability as a second, complex and controversial source of risk-taking. Nonlinear
probability enables to take a variety of psychological factors, like personal inclinations toward optimism or pessimism, into account7 . Recent economic

7 see

Shoemaker (1993).

literature provides an approach which accepts that risk

taking differs across contexts, but that the contexts
are strongly correlated (Dohmen et al. 2006). Decision analysts criticise the assumption of the traditional
economists that individuals naturally act in accordance
to the expected utility function. So the model got transformed over time. Markowitz (1952) was one of the
first who proposed that the utility is defined by the
gains and loses and not by the final asset as it was
assumed by the expected utility model. Furthermore
he noticed the importance of risk-seeking and so proposed utility functions, which can be concave as well
as convex. After a couple of alterations of the expected
utility theory, Kahneman & Tversky (1979) published
their seminal work about prospect theory. In general
the equations of the prospect theory stay in the same
form as of the expected utility models. Perhaps the
most important differences are first, that the changes
are important and not the final states and second that
the weights of the decision process dont coincide
with the stated probabilities.
In summary the concept of risk-taking is a widely
discussed issue in the psychological as well as in the
economic literature. The purpose of this chapter is to
provide the reader a short overview of the research in
these areas. Because of the vast amount of literature
and concepts in this field of research we cannot make
the attempt to reflect all of these research areas.
2.1 Risk-taking and injuries
There is a general consensus among scholars analysing
risk-taking behaviour that there exists a causal relationship between individual differences in risk-taking
and the occurrence of injuries. In general they come
to the conclusion that more risk-taking increases the
probability of an injury. While there is a comprehensive body of work examining the link between
risk-taking and road traffic crashes there are just a few
studies available which investigate this relationship in
other circumstances, especially in the case of alpine
downhill skiing (Bouter et al, 1998, Cherpitel et al,
1998, Goulet et al, 2000). These studies have in common that they have their theoretical foundations on the
approach of sensation seeking based on the works of
(Zuckerman, 1979) and that they determine the level of
risk-taking behaviour mainly with self-reported data
based on interview questions. The findings of these
works are quite divergent from those investigating the
relationship of risk-taking and injuries in other contexts. The results of these studies suggest that the score
on the sensation seeking scale has no positive influence on the probability to get injured while skiing.
This means that for a risk-taking skier with a higher
sensation seeking score the probability to get injured
is lower. Additionally they show that risk-takers in
alpine skiing are normally more skilled riders than

975

the other, which probably helps them to anticipate

potential dangerous situations. Their concluding is
that injuries are not a factor of risk-taking, but they
appreciate the weaknesses of their technique to measure the level of risk taking with the potential problem
of biases. Whereas risk-taking in alpine skiing has
no influence on getting injured they show that going
skiing is a decision based on risk-taking behaviour.
They investigate with the concept of sensation seeking that the general skier has a higher score on the
sensation seeking scale than the general non-skier.
In other words the general skier wants to satisfy his
demand for thrill and adventure when practising his
leisure time activity. This fits to the work of (Eitzinger
& Wiedemann, 2007), who analyse risk perceptions
of alpine tourists. They demonstrate that the probability of having a skiing accident is overestimated by
the factor of 50. To sum up the possibility of having
a ski accident is a common knowledge in the population. Individuals, who purchase a ski ticket, do this
with the intension to satisfy their demand for thrill and
adventure. Rephrased this means that the decision to
purchase a ski lift ticket is the risk-taking action.
In this work risk taking in the concept of socially
accepted behaviour serves as base for our estimation
of the marginal effect of an additional skier. In the
case of alpine skiing it is very difficult to observe
risk-taking behaviour directly. One possibility to measure risk-taking behaviour lies in the observation of
ski accidents. Assuming that skiing is a risk-taking
behaviour, which implicates ski accidents, the amount
of ski accidents can be taken as a proxy for risk-taking
behaviour.
2.2

Risk-taking and externalities

Buchanan & Stubblebine (1962) define an externality

to be present when:
U A = U A (X1 , X2 , . . . , Xm , Y1 )

(1)

This means that the utility of an individual A is composed of activities (X1 , . . ., Xm ), which are exclusively
under his control or authority and a single activity, Y1 ,
which is by definition under control of a second individual. It is assumed that A will maximize his utility in
respect to X1 , . . ., Xm and the externality Y1 and that
he modify the values for X1 , . . ., Xm as Y1 changes to
maintain in a state of equilibrium.
For the economic analysis of the external effect
of risk-taking we use the model employed by Turvey
(1963) expanded by the case of accidents introduced by
Williamson, Olson and Ralston (1967). Let us suppose
we have two classes of individuals, A and B, who interact in a non-market situation. In the case of risk-taking
this means that B is experiencing physical injury as a
result of his interaction with A. In other words the class

of A is composed of damage-producing agents and the

class of B is composed of damage-sustaining agents.
The interaction between A and B happens not through a
specific pair-wise association but incidentally, i.e. the
accident. Additionally B doesnt know the expected
losses which arises out of As activity. This makes
negotiations between the individuals of the two classes
from an economic viewpoint unfeasible. The level of
As activity, i.e. risk-taking, is the principal decision
variable. So we distinguish between to cases: A may
or may not be liable for the damages to B (Williamson
et al. 1967). In the case where A is not liable to damages for B, A will act without considering the losses
his activity imposes on B. This means that A will maximize his demand for risk-taking without considering
appropriate protective measures to protect B. To sum
up the outcome will remain socially inefficient whenever A is not liable for the damages to B. In the second
case in which A is liable to B Williamson et al. (1967)
differentiate between two alternatives: A has assets
against B can satisfy his claims and A has no assets.
First led us consider the case in which A is liable for the
damages he imposed to B and has assets against B can
satisfy. In this case A will have an interest to minimize
the potential losses and so will consider appropriate
protective measures and restrict his risk-taking to a
social optimal level. The case in which A is liable to
B and has no assets will lead to the same outcome as
in the case he is not liable to B. This means that B
will maximize his risk-taking without considering the
damages occurring to B.
As we have shown, if the demand for risk-taking is
unregulated, the individual has the incentive to maximize his demand for risk-taking which lies above the
social optimal level. So the question is, how the system
should be arranged to bring the demand for risk-taking
and so the amount of the externality to a social optimal
level. A couple of possibilities are available, which we
want to introduce in its basic thoughts. First a tax may
be introduced, which reduces As marginal net gain for
satisfying his demand for risk-taking by the amount
of costs he imposes on B. So As marginal net gain
becomes zero at the social optimal level and he has no
incentive to consume more risk-taking. The difficulty
with this instrument is that the tax has to be at the
same level as the external costs, which are sometimes
difficult to determine.
Another possibility is the introduction of command
and control policies, for example constraints which
regulate the amount of risk-taking demanded by the
individual. In this case the problem of prohibitive high
monitoring cost has to be regarded.
To sum up an unregulated activity, in our case risktaking in sports, which produces an externality, here
accidents, is consumed on a level above the social optimal level. Regulating instruments often need to valuate
the external effect by determining the marginal effect

976

of an additional consumption on the society and so

calculating the external costs of such behaviour. For
this reason in the next chapter the marginal effect of
an additional skier with his risk-taking on the ski-run
is estimated to serve as a base for the calculation of
the external costs.
2.3

Estimating the marginal effect of an

additional skier

The first step consists of calculating the marginal

effect of one additional skier on the amount of accidents. As the costs of accidents vary by severity, this
exercise needs to be done for different injury levels.
The dataset from the Austrian Kuratorium for alpine
Security only differs between light, severe and deadly
accidents. We thus analyse the impact of the number of
visitors in a ski resort on overall accidents and severe
accidents (incl. fatal ones).
Figures 1 and 2 present the relationship between
the number of skiers and the number of accidents and
severe accidents respectively. Both regression lines
show a positive slope, however the difference between
the two accident categories becomes obvious when one
takes into account the scale of the y-axis.
In order to quantify the external effect we have to
estimate the slope of the regression line. For any skier
the probability of having an accident depends on external factors that increase the risk of an accident (e.g.
fraction of hard pists, congestion) and the number of
other skiers in the resort. By simply aggregating the
function for an individual skier across all skiers we
can derive an econometric function for the number of
ski-accidents in resort j:
ACCj = 0 + 1 SKIERSj + 2 Xj + j

(2)

ACC is the number of (severe) accidents in skiresort j, SKIERS is the number of skiers (in thousand)

Figure 1. Relationship between number of skiers and number of accidents.

Figure 2. Relationship between number of skiers and number of severe accidents.

Table 1. Marginal effect of additional skiers on number
of accidents.
Model 1
Coeff.

Std. Err. Coeff.

2.088

Skier (in 1,000)

Density
0.132
Hard pists %
0.330
Cable cars %
47.951
Constant
14.591
Obs.
Prob > F
R2

Model 2

0.298
12.960
34.221
17.205
11.419

71
0,000
0.735

Std. Err.

0,627

0.228
3.233
8.185
1.944

0.103
5.346
12.888
6.388
3.665

71
0,000
0.676

Notes: OLS estimates; Dependent variable: Number of accidents (Model 2); Robust standard errors. , , indicate
significance at the 1,5 and 10% level.

and X is a vector of control-variables including the

fraction of hard pists, the density in the ski-resort and
the fraction of cable cars. We assume the number of
skiers to increase the number of accidents. Density and
the number of hard pists should also have a positive
sign. Cable cars are assumed to increase the number
of accidents due to the fact that resorts with higher
amount of cable cars have a higher fraction of skitourists and thus a lower average skill-level. i are
coefficients to be estimated, where 1 is the coefficient of main interest, the marginal effect of additional
skiers and is the error term.
The results of the OLS-estimates are summarized
in Table 1. Model 1 estimates the effects of additional
skiers on the number of overall accidents. One additional skier increases the number of accidents by 0.002.
The difficulty-level of a resort does not seem to have
a significant effect on the number of accidents. This
confirms the assumption that skiers self select themselves into resorts that challenges their skill level and

977

that the cumulative distribution of skills is equal across

resorts, but the mean might differ. Density also seems
to have no significant effect on accidents.
However, the fraction of cable cars and chair lifts
in overall lifts has a highly significant effect on
accidents. This could indicate that more comfortable
transportation attracts more inexperienced skiers. The
results of Model 2 show similar results for the regression of severe accidents on the number of visitors and
the other controls. One skier increases the amount of
severe accident by 0.0006.
After estimating the average marginal effect of one
skier, the next step is to determine the external accident costs this skier incurs. These costs would then
be multiplied with the coefficients estimated in this
section.
2.4

Calculating external costs

Figure 3 illustrates the relationship between private

marginal costs of the consumption of risk-taking of
A (MCprivate ) (e.g. injuries from an accident) and its
social marginal costs (MCsocial ) (e.g. endangering or
injuring other skiers). Social marginal costs are the
product of MCprivate and the marginal costs of the
externality (MCexternal ), which are in our case the damage occurring to B. The abscissa in this figure shows
the amount of risk-taking consumed by A (Q). The
ordinate represents the notional price of risk-taking
measured in monetary terms (P). Because risk-taking
is not traded on a market and so has no market price,
this axis pictures the willingness to pay for risk-taking.
The downward sloping curve (D) is the demand curve
for risk-taking, i.e. if the price of risk-taking is high,
less risk-taking is demanded and vice versa. If the negative externality of the consummation of risk-taking is
not internalized A demands risk-taking to this level,
where MCprivate intersects D, point R1 with a quantity

MCsocial

R2

MCprivate

R1

MCexternal

demanded of Q. MCprivate is upward sloping, because

next to private costs like price of a day-ticket, opportunity costs, etc. risk-taking generates costs by himself as
well. So if he is consuming more risk-taking the probability of having a severe accident increases. The social
optimal level of the consummation of risk-taking is
the point where MCsocial intersects D, point R2 with a
quantity demanded of Q . It can be seen, if the consummation of risk-taking is not regulated the amount
of risk-taking exceeds the social optimal level.
2.5

Evaluate willingness to pay for risk-taking

activity

An alternative approach to design instruments using

monetary incentives could be via the individual WTP
for risk-taking behaviour. A monetary value for the
option to exercise sensation-seeking could allow to
possibly compensating individuals for the decrease in
utility from reducing excessive risk-taking.
Borsky & Raschky (2008) applied the hedonic market approach to estimate the individual WTP for the
option to exercise risk-taking. The hedonic market
approach tries to explain the value of a commodity
as a function of valuable characteristics. The strength
of the hedonic market approach is that it is an indirect
valuation method based on observed behaviour and
not on stated preferences. They used data presented
in this paper to examine the impact of various factors
on the price of ski day passes. Ski Resorts in Austria
differ in their characteristics like total length of their
ski runs, altitude, transport capacity, types of transport
facilities, level of difficulty of the ski area, snow conditions and scenery as well as driving distance to the
next population centres, proximity of other ski resorts
and the number and types of accommodations and so
does the price.
Individual data on risk-taking is not available and
not possible to collect with reasonable means. However, risk-taking behaviour can result in accidents.
Thus accidents could serve as an empirical proxy for
risk-taking. Given that they do not know the mathematical relationship between risk-taking behaviour and
accidents they compared varying levels of accidents,
as performed in this paper. This enabled the authors
to calculate the willingness to pay for a hypothetical reduction of the risk-taking-behaviour resulting
in severe accidents to the level of overall accidents.
Borsky & Raschky calculate a WTP for this difference in risk-taking of around 15 % of the overall day
pass price.

D
Q*
Figure 3.

Q1

The concept of external costs.

CONCLUDING REMARKS

To our knowledge this is the first attempt to empirically investigate the concept of risk-taking in high

978

risk-sports, especially alpine skiing, from an economic

viewpoint. We have shown that unregulated risk-taking
in high-risk sports lead to a consumption above the
social optimal level. Furthermore this consumption of
risk-taking produces a negative externality, i.e. accidents. Policies to internalize the external effect of
socially unaccepted risk-taking behaviour (e.g. speeding, driving under influence) are already in place,
while instruments to reduce the adverse effects of
socially accepted risk-taking, for example in high-risk
sports, regulations are more the exemption. One cause
of this undersupply can be, that regulations for risktaking in a social accepted manner are not favoured
to implement by politicians. Another cause is that
it is very difficult to calculate the external costs for
risk taking. The purpose of this paper was to provide explorative analysis in this direction approach
by determining the marginal effect of individual risktaking behaviour. We applied data from Austrian
ski-resorts and examined the determinants of skiaccidents. In a next step, the willingness to pay for
risk-taking should be estimated in order to define an
appropriate compensation for the decrease of utility
from reducing excessive risk-taking. This compensation can then be used to design instruments, such
as alternative lift-ticket-pricing, insurances or liability regulations that reduce the amount of excessive
risk-taking to a social optimal level.
REFERENCES
Borsky, S. & Raschky, P.A. (2008). Estimating risk-taking
behaviour in risky sports: A hedonic market approach.
mimeo. University of Innsbruck
Bouter, L.M., Knipshild, P.G., Feij, J.A. & Volovics, A.
1998. Sensation seeking and injury risk in downhill skiing.
Personal Individual Differences 9(3): 667673.
Buchanan, J.M. & Stubblebine, W.C. 1962. Externality.
Economica 9 (3): 371384.
Cherpitel, C.J., Meyers, A.R. & Perinne, M.W. 1998. Alcohol
consumption, sensation seeking and ski injury: A casecontrol study. Journal of Studies on Alcohol and Drugs
58(2): 216221.
Dohmen, T., Falk, A., Huffman, D., Schupp, J., Sunde, U. &
Wagner, G. 2006. Individual risk attitudes: new evidence
from a large, representative, experimentally-validated survey. Cepr discussion paper series no. 5517, Center for
Economic Policy Research (CEPR).
Eitzinger, C. & Wiedemann, P. 2007. Risk Perceptions
in the Alpine Tourist Destination Tyrol-An Exploratory
Analysis of Residents Views. Tourism Management 28:
911916.
Esyneck, H.J. 1973. Conditioning, introversion-extraversion
and the strength of the nervous system. In Esyneck, H.J.
(ed.), Esyneck on extraversion, London: Crosby Lockwood Staples.

Fenichel, O. 1939. The counterphobic attitude. International

Journal of Psychoanalysis 20: 263274.
Friedman, M. & Savage, L.J. 1952. The expected-utility
hypothesis and the measurability of utility. Journal of
Political Economy 60: 463474.
Goulet, C., Regnier, G., Valois, P. & Ouellet, G. 2000.
Injuries and risk taking in alpine skiing. Skiing Trauma
and Safety 13: 139146.
Horvath, P. & Zuckerman, M. 1992. Sensation seeking, risk
appraisal and behaviour. Personality and Individual Differences 14: 4153.
Jack, S.J. & Ronan, K.R. 1997. Sensation seeking among
high- and low-risk sports participants. Personality and
Individual Differences 25: 10631083.
Johnston, J.S. 2003. Paradoxes of the safe society: A rational
actor approach to the reconceptualization of risk and the
reformation of risk regulation. University of Pennsylvania
Law Review 151 (3): 747786.
Kahneman, D. & Tversky, A. 1974. Judgement under uncertainty: Heuristics and biases. Science 185: 11241131.
Kahneman, D. & Tversky, A. 1979. Prospect theory: An
analysis of decisions under risk. Econometrica 47 (2):
263292.
Markowitz, H. 1952. The utility of wealth. Journal of Political Economy 60: 151158.
Rettenwander, A. 2003. Risiko und Extremsport. OEAV
Study, Innsbruck: OEAV.
Schoemaker, P. 1993. Determinants of risk-taking: Behavioral and economic views. Journal of Risk and Uncertainty
6: 493501.
Schuetz, H., Wiedemann, P.M. & Gray P.C.R. 2000. Risk perceptionBeyond the psychometric paradigm. Works on
Risk Perception 78 (2). Jlich, Forschungszentrum Jlich.
Slangler, E. & Rudestam, K.E. 1997. Motivation and disinhibition in high risk sports: Sensation seeking and
self-efficacy, Journal of Research in Personality 31:
355374.
Turner, C., McClure, R. & Pirozzo, S. 2004. Injury and
risk-taking behaviorA systematic review. Accident
Analyses and Prevention 36: 93101.
Turvey, R. 1963. On divergences between social cost and
private cost. Economica 30 (119): 309313.
Von Neumann, J. & Morgenstern, O. (2 ed.) 1947. The theory of games and economic behaviour. Princton: Princton
University Press.
Weinstein, N.D. 1984. Why it wont happen to me: Perceptions of risk factors and susceptibility. Health Psychology
3: 431457.
Williamson, O.E., Olson, D. & Ralston, A. 1967. Externalities, insurance and disability analysis. Economica 34
(135): 235253.
Zuckerman, M. 1979. Sensation seeking: Beyond the optimal level of arousal. Hillsdale: Lawrence Erlbaum
Association.

979

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Licensing a Biofuel plan transforming animal fats

Jean-Franois David
Compagnie Nationale des Experts de Justice en Environnement, France

ABSTRACT: As specific and new industry, biofuel industry mixes both a new processing for agricultural and
animal raw materials, in a rather chemical looking plant beyond rather classical industrial processes.
A controversy rises about stakes in biofuel industry: sustainability, indirect effects of allowance of agricultural
raw materials to produce industrial commodities, so the debate reaches unexpected levels;
Presentation will tackle:
Requirements for application drafting and public debate in France for a specific plant, transforming animal
fats into biofuels,
Developments of debate, management of application-specific arguments and industry or policy specific
discussions to reach a final licensing decision.
Adaptation of public regulation to such a wide span in public expectations about an environmental and energy
issue.
Consideration upon environment management inside project management.

to contribute to the European target of 5% biofuel in

car fuel.

INTRODUCTION

Due to fear of shortage in automotive fuel EU decided

in xx a minimum of biofuel had to be added to normal
automotive fuels.
In France, those provisions are enforced through
a state led initiative granting tax alleviation to industrial projects. The tax alleviation decision (agreement
fiscal) is a prerequisite for every industrial project.
That presentation will tackle licensing process downstream that tax alleviation decision, and focus on
environmental aspects.
It led to a generation of several industrial projects
based widely upon alcohol (sugar beets), and vegetable
oil (rapeseed, sunflower), which are mainly known.
Their use as industrial raw material and their impact
upon food prices is a subject of a wide debate.
An other source of fats is to be explored: animal
fats, from rendering plants. That source of fats is
rather destined to lipo chemistry or burnt in rendering
plants.

2
2.1

APPLICATION DRAFTING
Project content

The project content includes collecting animal fats in

a wide part of France and treating them in a unique
location in order to produce biofuel destined to be
eligible in the biofuel ax exemption program, in order

2.1.1 Improving animal fats valorization

Rendering industry has two main products: fats and
proteins.
Fats are currently used either in lipo chemistry or
as common fuels for rendering plants.
Biofuels are an opportunity for improving valorization by rendering industry itself.
The acronym of FAME (Fatty Acids Methyl Ester)
describes both the result of a process and the steps:
trans etherification.
2.1.2 Actions and reactions
Animal fats are subject to hydrolysis producing glycerin and fatty acids.
A the end of the process, fatty acids are subject
to an other reaction with methanol to get the Fatty
Acids Methyl Ester (FAME) which can be mixed with
fuel oil.
2.1.3 The project
The project is sized to produce yearly 100 000 tons of
biofuel through treatment of 275 t/day of animal fats
and eventually additional raw fats.
It is located close to a rendering plant, belonging
to the same shareholder and existing for more than
50 years.

981

An average of 22 people will operate the new

=, 15% of which devoted
plant. Investment is 20 MC
to environment protection.
The project will produce biofuels (FAME) and
glycerin.

in the dpartement, of the local authorities, of the

professions concerned, of environmental protection
associations and competent persons.
Environment
Impact statement
Riskstudy

2.2

Biofuel national
program

Tax exemption
application

January September

Tax
ax
exemption
Decision

2006

Environmental requirements

2.3.1 Main issues and topics

Air pollution, water pollution and control, waste management are main subjects for licensing process. In
fact in a rather small plant those discharge limit are
already decides at national level, according to European directives, and the local prfet may have a stricter
enforcement if for example the sensitivity of rivers or
the local needs of air pollution demands it.
Those steps will be described here after, with legal
requirements and their achievements in what can be
called true life.
2.4

Commissaire
Enquteur
opinion

The administrative background

The project is first to be selected at national level to get

a tax exemption. That is a compulsory prerequisite for
such a project: less that eligibility to tax exemption,
the investment would not even have been considered
by share holder.
In fact the first application for tax exemption was
issued as an answer to a call for tender for several biofuel plants, the call for tender had specified,
end of 2005, that beyond vegetable products, animal
fats were also eligible raw materials. In September
2006 the project was granted a tax exemption on a
35000 t/year basis. That decision was the kick off for
the project.
Then, administrative environmental procedure has
to be conducted at department level, under the
leadership of the prfet, and technical civil
servants.

2.3

Public enquiry

Licensing steps

The main steps are as follow: filing application, public inquiry, and opinion of a commissaire enquteur,
then presentation to a dpartement-level commission
is also consulted; it may vary in accordance with
the type of facilities in question and its composition, includes notably representatives of the State

8 months

Licensing conditions

Commission
For environment
and technological
hazards

Ruling of Prfet
Granting license
or refusal

4 months

That presentation is done by classified installation inspectorate, and is backed by provisions for
licensing: conditions for supervision, discharge limits, ways and means of informing public and stake
holders.
2.4.1 Filing the application
When sent to the prfet, the applicant is declared entitled to proceed with its project within a two month
time span.
Then, due may be to the innovative nature of the
project, which entitlement lasted four months.
The content of the application lies mainly around
following key documents, beyond administrative identification of the applicant and its financial and technical reliability:
An environment impact statement,
A risk study,
A declaration of compliance with occupational and
health rules,
A declaration of the future use of land (land planning outlook) in order to foresee conditions for land
remediation at the end of operations.
2.4.2 Risk study
Hazards are mainly linked to explosion and fire.
The risk study described the range of influence of
various types of explosions likely to occur.
For the various classes of accident examined, it
appeared that effects would be limited inside the plant
precinct.
Anyway, solutions where described and organization was shaped in order to grant an acceptable
prevention level.
Due to the accuracy with which several hazards
scenario had been studied, no questions were raised
by the public.

982

2.4.3 Environment impact statement

2.4.3.1 2.4.3.1 Water
Discharge had to be treated inside the neighbouring
treatment plantfortunately oversized; the characteristics of raw liquid discharged was not very different
of other effluent; the pollution abatement is for COD
from 7 000 mg/l to a range of 110 mg/l to 300 mg/l.
The additional flow is 500 m3 /day of an effluent
with a concentration of Muck spreading is used mainly
in summer to spread sludge and recycle nitrogen in
relation with surrounding farmers.
River Meuse at that point has strict needs at low
water level (usually beginning of autumn); respect
of those quality requirements had to be fulfilled, as
described through here after tables.
Table 1.

Effluents concentration at low water level

TSS
COD
BOD5
Nitrogen NTK
Phosphorus

Table 2.

River Meuse
quality at
Bras-surMeuse (mg/l)

150
450
450
300
15
Discharge
500 m3 /j

11,5
10,33
2,12
3,67
0,11

0,263
0,789
0,789
0,526
0,026
River Meuse
flow at low
level:
3,3 m3 /s

Resulting river quality at low water level.

River quality at low water level

TSS
COD
BOD5
Nitrogen
NTK
Phosphorus

River Meuse
quality at
Bras-surMeuse (mg/l)

Resulting
quality at
discharge
point

River Meuse
quality
objective

11,5
10,33
2,12
3,67

11,763
11,119
2,909
4,196

< 25
20 25
35

0,11

0,136

0,1 to 0,3

River Meuse quality objectives are respected.

2.4.4 Air pollutions
The plant air pollution comes from three main
sources:
a high pressure boiler (3 MW),

Anyway, some peopled considered the fact of neihgbouring rendering plant to fear some ill-smelling
phenomenon, those concerns were dropped after a
simple presentation of the differences between processes (chemical processing of fats versus a thermic
process upon raw animal waste).

PUBLIC ENQUIRY AND PUBLIC OPINIONS

3.1.1 Public enquiry

Public enquiry lasts one month, every one, citizen
French or foreigner, association city council may
express his opinion. It is announced through two local
newspapers fifteen days before the enquiry starts.
A commissaire enquteur, independent person
appointed by the Administrative Tribunaloften a
recently retired professionaldelivers his opinion
about the project, the objections raised by environment
activists, public groups. The commissaire enquteur may also postpone the end of the enquiry by
fifteen days if he has an insight of public needs for
additional information.
As a by-law, public enquiry is preceded by an explanation issued by future plant operator to various group
of opinion in order to trace and understand what are
needs and expectations from each group front to such
a project.
Let us say that no objection was filed, neither any
asking for additional explanation or information.
So the commissaire-enquteur was in a mood to
deliver a favorable opinion both on the project and
environmental provisions and technological precautions.

Impact of plant discharge at low water level.

Discharge
Concentrations Assessed
out of sewage Impact
plant (mg/l)
(mg/l)

a coolant and fluid heating boiler (10 MW), emitting particulates, CO, SO2 , NOx in compliance with
general regulation and discharge level,
a scrubber emitting methanol.

3.1.2 What could have been said

A debate about biofuel legitimacy, general product
lifecycle, and influence on cost of foods could have
been triggered. Anyway, the raw material, animal fats
which are not aimed at any food consumption, was not
likely to generate any controversy.
However water pollution could be tackled and fairly
explained it was a strong subject for anglers.
Air pollution is not a subject of interest in a rural
area so far as customary parameters (CO, SO2 , and
NOx ) are at stake, but the past history of neighboring
rendering plant.
The prospect of having a new plant in a rather small
municipality was of high interest for city council.
In fact the major subject of concern was the traffic
of trucks induced by the plant. That major concern

983

content, but the delays resulting additional cross

examination by public agencies are questionable.

for population and their representatives was tackled

through:
a strong participation to a public debate related to a
road project bypassing the city centre
a commitment to carry on a study of reopening a
nearby railway for plant use, and other industrial
users if any.

An important notice: classical industrial impacts

air pollution, water discharge and waste management
were exclusively driven by state civil servants,
without echo among local population.

Technological innovation and environment impact

assessment are going along, feeding one another;
so additional innovative steps and associate decisions need an environmental update; that will affect
licensing duration.

3.1

5.1

Environment management inside project

management

5.2

That industrial project related to public goals

achievementincorporating more biofuel in automotive fuelshas lasted two years before investment
being considered by its share holder because of the
need for environmental procedure.
First, those procedures are essential both for public
acceptation and to have best achievable technologies
fully adopted.
Then, the length of procedure does not grant a better
ownership, by stake holders, of long term and long
range environmental issues.
One must notice that the length of time spent by
state services in order to prepare procedures delays
project achievement without providing the conditions
of a sound public debates which has been by passed,
so far as global issues are at stake .
Should public debate and public inquiry be included
in the rule of the game as compulsory steps of project
management but carried on by project manager.

SUMMARYFIVE KEY POINTS

An innovative industrial process can be no public

concern when developed in an area keen to accept
prospect of additional employment.
Legitimacy of those innovations was fully admitted
in an unchallenged acceptance of a step on the way
of general progress.
Public expectations:
are targeted to limited drawbacks such as noise
pollution related to truck traffic,
but do not consider water or air pollution at a
higher level,
Public agencies are in charge of enforcement of
existing legislation which is an easy task.
Duration of licensing: public enquiry and
commissaire-enquteur opinion adds a value to public understanding and then ownership of application

LESSONS LEARNED
Innovation management

Project management

Licensing procedure and licensing achievements is a

key milestone in project management: environment
license is a prerequisite for shareholders decision to
allow the funding.
5.3

Improvements

Explanation to public stake holders, environmental

activists is a fully understood prerequisite.
Beyond content of environment impact and hazard
assessment in application file a third party assessment
has to be initiated by the applicant to provide reliable opinion to back public debate, if any but also
discussions with public agencies.

ANNEX MAIN LEGAL REFERENCES

Environmental Code
Article L512-1
(Act no. 2003699 of 30 July 2003 Article 4, Article
25 Official Journal of 31 July 2003).
(Act no. 200611 of 5 January 2006 Article 77 Official
Journal of 6 January 2006).
The facilities that present serious hazards or drawbacks for the interests referred to in Article L. 511-1
are subject to authorization by the Prfecture.
The authorisation may be granted only if these hazards or drawbacks can be prevented by measures which
are specified in the ruling of the Prfecture.
The applicant supplies a risk study specifying the
risks to which the facility may expose the interests
referred to in Article L. 511-1, directly or indirectly,
in case of an accident for whatever reason, be it internal
or external to the facility.
The content of the risk study must relate to the scale
of the risks caused by the facility. Where needs be,
this study gives rise to a risk analysis which takes into
consideration the probability of an accident occurring
and the kinetics and gravity of potential accidents, in

984

accordance with a methodology which is explained in

the said analysis.
It defines and justifies appropriate measures to
reduce the probability and effects of such accidents.
The granting of the authorisation, for these facilities, may be subject in particular to their distance
from dwellings, from buildings habitually occupied
by third parties, establishments receiving the public,
waterways, communication routes, water catchment
areas, or zones destined for dwellings by binding planning documents. It takes into account the technical
and financial capacities of the applicant to conduct
his or her project in compliance with the interests
referred to in Article L. 511-1 and to fulfil the
obligations of Article L. 512-17 when the activity
ceases.
Article L512-2
(Order no. 2005-1527 of 8 December 2005 Article 28
Official Journal of 9 December 2005, in force on 1st
July 2007).
The authorisation provided for in Article L. 5121 is granted by the Prfet, after a public enquiry
into any possible incidences the project might have
on the interests mentioned in Article L. 511-1 and
after the municipal councils concerned have given
their opinion. A dpartement-level commission is also

consulted; it may vary in accordance with the type

of facilities in question and its composition, set by
a Conseil dEtat decree, includes notably representatives of the State, of the local authorities, of the
professions concerned, of environmental protection
associations and competent persons. The authorisation
is granted by the Minister responsible for classified
facilities, after an opinion has been given by the
Higher Council of Classified Facilities, in any case
in which the risks may concern several dpartements
or regions.
A Conseil dEtat decree sets the terms of application
of the preceding paragraph. It also sets the conditions
under which the Conseils Gnraux and the Conseils
Rgionaux must be consulted and the form of that consultation. If a building permit has been applied for, it
may be granted but may not be executed before the
closure of the public enquiry.
Article L512-3
The installation and operation conditions considered
indispensable for the protection of the interests mentioned in Article L. 511-1, the means of analysis and
measurement and the means of intervention in case of
an incident are set out in the authorisation ruling and,
should the occasion arise, by complementary rulings
issued after the said ruling.

985

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Modelling incident escalation in explosives storage

Gavin Hardman, Tim Bedford, John Quigley & Lesley Walls
Department of Management Science, University of Strathclyde, Glasgow, UK

ABSTRACT: We consider the problem of modelling incident escalation in explosive storage. Escalation is
a key stage in the process of transition between minor and major incidents, and therefore is worthy of careful
consideration. Our aim is to develop a model that can be used to inform the risk assessment process, by providing
supporting information (in this case its contribution to incident escalation) about the properties of the explosive
to be stored. Scarcity of data (especially for model validation) is often a significant difficulty in these problems,
so we focus on developing a model that is less data-dependent than a full physical model.

INTRODUCTION

The development of models to aid in quantitative

risk assessment (QRA) for explosives storage is a
problem faced by a number of organisations. Commercially, explosives are used extensively in mining
and quarrying and have many uses in civil engineering. The storage of military explosives, which are
often more specialised than those used in industrial
applications due to the need for different operational
behaviour (Marder 1998), also requires quantitative
risk assessment (HSE 1999).
The motivation for conducting quantitative risk
assessments for explosives storage is the potential
severity of the consequences of incidents involving
explosives. Table 1 gives details of 4 incidents from
2001 that illustrate the severe consequences of incidents at explosives stores. By undertaking activities
that pose risks, organisations are also responsible for
protecting both their workers and the general public
from the consequences of these risks. The role of QRA
is to allow those responsible for explosives storage
to demonstrate that they understand the nature of the
associated risks and what controls are needed to ensure
the risks are as low as reasonably practicable (HSE
2001).
Such incidents can be extremely costly, in financial terms and in terms of injuries and fatalities. In an
high-profile incident at Camp Doha in July 1991, over
$15million worth of ordnance was lost as a result of a
fire within the camps storage compound, which triggered a sequence of sympathetic reactions that lasted
several hours. The incident left unexploded ordnance
scattered around the compound which took months to
clear (USDoD 1998).
Although such incidents in storage are often serious, they are also rare. For example, (Merrifield

and Moreton 1998) show there were 79 incidents

on the UK mainland between 19501997. Of these
79, just 16 occurred in storage, a rate of one every
3 years. Consequently, there is a lack of available data to aid model development. In addition,
explosives are often costly to test, so testing is concentrated on assessing design mode functionality,
which does not necessarily provide the information
required to assess in-storage safety. In such situations,
expert judgement can be used to support the sparse
raw data.
There is a large volume of work on explosion modelling in the physics and engineering literature. (Kapila
et al. 1999) and (Valone 2000) provide detailed physical models for describing the response of explosive
materials to specific threats. (Goldfarb and Weber
2006) and (Marder 1998) give accounts on the historical development of explosion models and comment
on the need for experimental validation, and (Wu and
Hao 2006) develop a numerical method for damage
assessment.
(Azhar et al. 2006) and (Stewart and Netherton
2008) consider risk assessment case studies for a commercial applications and a series of technical reports
(Hardwick et al. 2007), (APTResearch 2002) outline
the SAFER model architectureas used by the US
military. (Pfitzer et al. 2004) compares QRA methods for explosives storage with those used in other
industries.
In the next section we look at the difficulties of
explosives storage risk assessment and the particular problem under consideration in this paper. In
Section 3 we describe the model structure and how
to simulate from the model. We review our simulation
results in Section 4 and discuss the potential developments and current limitations of our approach in
Section 5.

987

Table 1.

Major incidents involving stored explosives 2001.

Location

Incident

11 July,
Afghanistan

Three people are injured in an

explosion at an ammunition depot at
Darulaman on the southern outskirts
of the Afghan capital, Kabul.
Three people are killed and four others
are injured in a blaze at an army
ammunition depot in the Buryat
region of eastern Siberia. The
accident was reportedly caused by
lightning.
At least four people are injured and
100 homes are damaged in an
explosion at an army base in central
Vietnam. Reports say 3.5 tons of
explosives and ammunition blew up
at an army warehouse in Hoa They,
south of the Vietnamese capital,
Hanoi.
At least 10 people are killed after a fire
starts a series of explosions at an
ammunition dump at army base in
the Guinean capital, Conakry.

21 July, Russia

8 June, Vietnam

3 March, Guinea

Source: BBC.

EXPLOSIVES STORAGE PROBLEMS

In this section we describe the problem we wish to

address, and illustrate the need for pragmatic modelling approaches to these problems by considering
current practice and other potential solutions.
The main problem is that of assessing the level
of risk associated with a particular article, where we
define an article to be the generic object being stored,
and the items to be the specific individual objects.
Further, we define a unit to be one box of an article
as packaged for storage, so the number of items contained within the unit will depend on the article being
stored. In the UK, explosives storage is subject to the
Control of Materials and Hazards (COMAH) policies,
as laid out by the HSE (HSE 1999). These recommend
that for any activity in which the individual risks are
perceived to be high (the HSE use 1 in 1,000,000 and 1
in 10,000 as heuristics for activities of negligible risk
and unacceptable risk), a quantitative risk assessment
(QRA) is carried out. Such QRA are used to demonstrate that the individual and societal risks of the hazard
are, in fact, tolerable and ALARP.
A case study (van Donegen et al. 2000) on a related
problem examined the (QRA) models of 7 NATO
member states. The case used to compare the QRA
models was a 40 tonne donor trial. Each participant
in the QRA comparison was required to provide their
risk assessment procedure given the trials data. It
was concluded that the nations use the same basic

formula that risk is equal to the probability of an event

multiplied by its consequences. There is broad agreement on the order of magnitude of event probabilities,
although different methods are used to derive these.
Examples of current practice in explosives storage are
given by (Hardwick et al. 2007) and (APTResearch
2002), which describe the US SAFER model, and by
(Bienz 2003), which outlines a different, events based
approach.
Methods for determining the consequences of an

incident follow the same approach (van Donegen

et al.
2000). This is to determine the area affected by the
incident and to assess the consequences for each acceptor site within that area. Costs are assigned and the
total consequences given as the sum of the costs from
individual sites, where these costs are given as the
sum of the costs of expected structural damage and
expected number of injuries/fatalities. The models
make different assumptions about how to define these
expectations, but the basic structure is consistent.
Current practice for determining incident probabilities is less clearly defined. In the Swiss model (Bienz
2003), incidents occur as a result of a sequence of
events, beginning with a trigger and undergoing an
escalation process. We focus on modelling the escalation process under the assumption that the triggering
events have already led to incident initiation. For this
account we do not consider the problem of assigning probabilities to the trigger events, although this is
clearly a significant part of building a full model.
We choose to focus on escalation as this is the stage
that determines the scale of the incident. If there is
no form of escalation then there is no transition from
small triggering events to major incidents.
The problem under consideration can be formulated as follows: We have an article, A, for which
we must assign a risk level. This article is stored
in 3-dimensional stacks of size n1 n2 n3 = N
units, and individual units of an item are denoted ui ,
i = 1, . . . , N . The severity of an incident involving
A is measured by the number of units that respond
during the incident and the duration of the incident.
The latter is based on the assumption that a more protracted incident will be more easily controlled. Units
can respond in one of two ways. They can detonate,
or burn. These response modes occur as a result of 2
initiation mechanismsshock initiation or heat initiation. We assume that either response can be achieved
from each initiation mechanism, and that these are the
only response modes/initiation mechanisms.

988

ESCALATION MODEL

We tackle this problem by treating it as a stochastic

process. We choose not to implement a full physical
explosion model due to the difficulties in obtaining

sufficient data to estimate the parameters accurately.

Instead, we aim to develop a model that adequately
characterises the underlying process whilst remaining
easily interpretable. In addition, we make use of expert
judgement techniques to help negate the impact of data
unavailability.
From the problem description in Section 2 we
have:
a 3-dimensional object consisting of N units of
article A
2 response modesdetonation response (d) and
burning response (b)
2 initiation mechanismsshock (m1 ) and heat/
energy determined initiation (m2 )
The stack of units is modelled as a 3-d lattice with
a unit at each corner, where we assume edge length
is equal to unit size. The following approach aims to
assign values to the probability of there being a major
incident whilst the articles are in storage, through
modelling of the different initiation mechanisms.
We consider the two initiation mechanisms as distinct processes. Shock initiations are consider to be
an instantaneous process that occurs as a consequence
of a detonation response and heat related initiations
to occur at random as a consequence of the prevailing energetic conditions. Escalation is modelled as
the combined effect of these processes, which we
introduced separately in 3.1 and 3.2.
Given N units within a stack, we assume these can
be split into two sets, those at higher risk of detonation, and those with lower risk. It is not unreasonable
to assume that some units would be at greater risk
than others due to their position in the stack, and these
sets are intended to represent the difference in physical
location between units without building a full spatial
model. These risk sets will have to be defined for each
responding unit, ur . The higher risk set is defined as
G + (ur ) = {ui |(ui ur ) < K} where is a distance measure (we use the Cartesian distance) and K is
known constant. The lower risk set, G (ur ) consists
of all yet to respond units not in G + (ur ).

We assume each unit can be assigned one and only

one response per shock, and ps + pd + p0 = 1.
The branching process has parent ur , a single unit
undergoing a detonation response. On detonation, ur
yields C1 children. These C1 are, in principle, shared
between G (ur ) and G + (ur ). We define p+ , the probability of response for a yet to respond unit ui G + (ur ),
to be p+ = pb + pd and p , the probability of
response for a yet to respond unit ui G (ur ), as
p = (pb + pd ). Using these values, and letting the
size of G + (ur ) = x the pdf of C1 is given by:
P(C1 = y) =



y 

x
pz+ (1 p+ )xz
z
z=0



N 1x
yz

(1)

Of the C1 children, C1d will be detonation

responses, and C1b will be burning responses. The
number
responses in C1 is distributed as
 of burning

b
Bin C1 , pbp+p
.
Burn
responses do not yield further
d
children as part of the branching process, so these
branches terminate at this stage. However, each of
the C1d detonation response children produces C2j
(j = 1, . . . , C1d ) on detonation. For each of these
responding units it is necessary to redefine the higher
and lower risk sets, so that they are centred on the
correct unit.
The first generation children will produce at total
of C2 children, and the detonations responses amongst
these could yield further children. The branching process is run until it either dies out (i.e. when Ci = 0) or
all units have undergone some form of response. The
pdf for the total number of children in generation i is
can be derived from (1):
P(Ci = y) =



y 

x
pz+ (1 p+ )xz
z


3.1 Shock initiation

1. no response, with probability p0 ;

2. burning response, with probability pb ;
3. detonation response, with probability pd ;

yz


(1 p )N 1xy+z .

z=0

Shock initiation is considered to give an instantaneous

response, with potential to propagate multiple sympathetic responses. A branching process is used to
describe this initiation mechanism. Shock initiation
occurs as a consequence of the impact on the unit during a detonation. There are three potential responses
to a shock for each unit, these are:

i1

Cj x
yz
0

(1 p )

i1
0


yz

Cj xy+z


.

(2)

3.2 Heat determined initiation

Responses due to heat occur randomly in time. We
use a Poisson process to model this initiation mechanism and aim to capture the effects of heat by linking
the intensity of the Poisson process to the heat energy
emitted during the unit responses.

989

We model the time to response of a yet to respond

unit (TTRi ) as:

(3)
exp hi (y)dy

where hi (t) is the hazard rate of unit ui at time t.

As more units respond, and therefore release energy,
the value of hi (t) (and consequently the expectation of
TTRi ) changes for those ui near to the responding
units (we define near as being in G + (ur ) for a given
T
ur ). The value of 0 hi (y)dy is taken to be an accumulation of the effects of all responding sites on the
unresponded site ui . That is,
T
hi (y)dy =

R(t) T


hr (y sr )dy

(4)

r=1 0

where R(t) is the number of units to have responded

so far, r is an index over the responded units and sr is
the response time of ur .
The effect of a detonation response is given by:
T
Wd (sr ) =

the point at which the stacks response was completed.

The response is deemed to be complete when R(t) = N
or if the heat energy emitted by the responses falls back
to the initial level. Clearly, the larger the value of R(T ),
the more severe the incident. We also record the duration of the incident (as the difference between the time
of the last response and the first). This allows to distinguish between incidents in which the same number
of units have responded.
3.3 Populating the model
In order to run the model, we must specify values for:
1. pb , pd probabilities for observing response b or d
to an initiation.
2. [0, 1] weight for reducing response probabilities in G (u)
3. K the range of the at risk region when defining
G + (u).
4. hd , hb function describing energy release behaviour
for each response mode.

4
hd (y sr )dy,

(5)

To simulate from the model, we use the following

procedure:

and a burning response by

T
Wb (sr ) =

hb (y sr )dy,

(6)

so the TTRi values are given by:




TTRi exp
Wd (sr ) + Wb (sr ) .

(7)

SIMULATION

This assumes that the integrated hazard rate is

proportional to the heat energy released during unit
response. The functional form of hb and hd can be
chosen to reflect the energy release behaviour of
the detonation response and the burning response.
Units undergoing heat/energy initiation can take
either response mode. Theproportion
taking burning

b
response is given by Ber pbp+p
,
similar
to shock
d
initiation.
The model is defined by these two processes running sequentially. The heat initiation determines the
TTR values for each component, and, if required, the
branching process determines the sympathetic reactions due to shock initiation of some units (effectively
censoring the TTR values in the shock initiated units).
Ultimately, we are interested in the severity of the
incident. This can be measured in terms of R(T ),
the total number units that underwent reaction by T ,

990

1. Define parameters: pb , pd , , K, hd , hb , stack

dimensions n1 n2 n3 = N , and background
hazard rate, h0 , for all units.
2. Generate the initial time to response estimate TTRi
for all ui as a realisation from a random variable
Y Exp(h0 ).
3. Identify next response location, defined as: ui s.t.
TTRi < TTRj j  = i.
4. Determine the response

type at ui , by sampling
b
from X Ber pbp+p
. If X = 1 we have a
d
burning response, and if X = 0 a detonation
response.
5. Define G + (ui ) and G (ui ) as described in
Section 3.
6. If X = 1, go to step 7, else; if X = 0, run
branching to completion to determine the number
of sympathetic shock responses.
a. Identify (n+1)th level detonation responses;
b. Identify (n + 1)th level burn responses;
c. Repeat for all children until process dies out.
T
7. Update 0 hj (y)dy values for all yet to respond
units, uj .
a. Determine effect of each detonation on uj using
Wd (sr )
b. Determine effect of each burn on uj using
Wb (sr )

8. Generate proposed TTRj as a realisation of

T
Exp( 0 hj (y)dy) for all yet to respond uj .
9. Define TTRj = min{TTRj , TTRj } for all remaining uj .
10. Iterate steps 39 until reaction complete. Where
complete means all units have responded or the
energy level falls below specified threshold K.
Results

Simulations were run for a stack with dimensions 3

4 5. 1000 realisations of the model were generated
for each set of parameters under test and we considered
the mean behaviour of R(t), the number of units to have
responded by time t. Parameters were assigned values
that fall within expected bounds to allow us to explore
model performance.
Figures 1 and 2 illustrate the effect of increasing the
value of parameter K, which defines the radius of the
high risk set. The range of values tested are taken to
be either side of points at which the number of units
included in the set increases. The plots show the value
of R(t) against t. It can be seen that as the value of K
is increased the rate of response increases, which is as
we would expect.
Figures 3 and 4 shows typical behaviour as we
increase the values associated with Wd (sr ) and Wb (sr ),

Figure 3. Effects of changing the Poisson process intensity

in the higher risk set.

60
55
50

R(t)

4.1

45
40
35

Intensit y = 0
Intensit y = 10
Intensit y = 10 2

30

Intensit y = 10 1
25
0

Figure 4. Effects of changing the Poisson process intensity

in the lower risk set.

60

50
K = 0.5
K = 1.2
K = 1.6
K = 2.1
K = 2.7
K = 3.2
K = 3.5

R(t)

40

30

20

10

0
0

0.5

Figure 1.

1.5

2
t

2.5

3.5

Effects of varying radius of higher risk region.

Figure 5. Effects of varying the probability of shock initiation within branching process.

60

50

R(t)

40

30

K = 0.5

20

10

0
0

Figure 2.

500

1000

1500

2000
t

2500

Reaction rate for K = 0.5.

3000

3500

4000

which determine the intensity of the Poisson process,

by accounting for variations introduced by energy
released by reacting units. Accurate values for this at
not known to us, so we test the effect of variation in
the range [0,1]. We look at how increasing the intensity
affects R(t). As expected more units respond earlier as
the intensity is increased.
Figure 5 shows the typical response behaviour as
the probability of unit response (of either type) during the branching process is increased. We can see
that there is a noticeable increase in the initial rate

991

REFERENCES

Figure 6. Effects of varying the ratio of shock initiations

within branching process.

of response as the probability of reaction increases,

which is to be expected. Figure 6 illustrates the effect of
changing the ratio of detonations to burning responses.
Again, behaviour is as expected, with the rate of
response being lower when the proportion of burnd
ing responses is increased ( pbp+p
= .25, so we would
d
expect to see only 1 detonation response in every
4 responses), and increasing as the proportion of
detonation responses is increased.
5

CONCLUSIONS

We have discussed the development of a model for

the escalation phase of incidents in explosive storage.
The model has been tested for functionality through
simulation. By considering basic behaviours, we can
see that the simulation behaves as we would expect.
At this stage we need to consider obtaining data to
allow us to parameterise the simulation model more
realistically.
The behaviour of the model is driven by the factors
stated in Section 3.3. This means we need to elicit and
validate two probabilities, two quantities and two functions. This a comparatively light elicitation burden,
compared with building and populating a full physical model. Subject to further testing and validation,
this may present a feasible model for supporting risk
decisions.
ACKNOWLEDGEMENT
The authors acknowledge the support and funding of
DOSG, part of UK MOD TES.

APTResearch (2002). Risk-based explosives safety analysis. Technical report, Department of Defense Explosives
Safety Board.
Azhar, S.S., M.A. Rahman, M. Saari, K.F. Hafiz, and
D. Suhardy (2006). Risk assessment study for storage
explosive. American Journal of Applied Sciences 3(1),
16851689.
Bienz, A. (2003). The new Swiss model for the probability of
an explosion in an ammunition storage. Technical report,
NATO/Bienz, Kummer & Partner Ltd.
Goldfarb, I. and R. Weber (2006). Thermal explosion:
modern developments in modelling, computation and
applications. Journal of Engineering Mathematics 56,
101104.
Hardwick, M.J., N. Donath, and J.W. Tatom (2007). Users
Reference Manual for Saftey Assessment for Explosive
Risk (SAFER) software.
HSE (1999). The Control of Major Accident Hazards Regulations 1999. The Health and Saftey Executive.
HSE (2001). Reducing risks, protecting people. Norwich,
UK: The Health and Safety Executive.
Kapila, A.K., R. Menikoff, J.B. Bdzil, S.F. Son, and
D.S. Stewart (1999). Two-phase modelling of ddt in
granular materials: Reduced equations. Technical Report
LA-UR-99-3329, Los Alamos National Laboratory.
Mader, C.L. (1998). Numerical Modeling of Explosives and
Propellants. Boca Raton, Florida: CRC Press.
Merrifield, R. and P. Moreton (1998). An examination of the
major-accident record for explosives manufacturing and
storage in the UK. Journal of Hazardous Materials A63,
107111.
Pfitzer, B., M. Hardwick, and T. Pfitzer (2004). A comparison of QRA methods used by DOD for explosives
and range safety with methods used by NRC and EPA.
Technical report, APT Research.
Stewart, M. G. and M. D. Netherton (2008). Security risks and
probabilistic risk assessment of glazing subject to explosive blast loading. Reliability Engineering and System
Safety 93, 627638.
USDoD (1998). Environmental exposure report. online,
http://www.gulflink.osd.mil/du/.
Valone, S.M. (2000). Development of a shocked materialsresponse description for simulations. Technical Report
LA-13665-MS, Los Alamos National Laboratory.
van Donegen, P., M.J. Hardwick, D. Hewkin, P. Kummer,
and H. Oiom (2000). Comparison of international QRA
models on the basis of the set-up and results of the joint
UK/Australian 40 tonne donor/acceptor trial. Technical
report, APT Research/NATO.
Wu, C. and H. Hao (2006). Numerical prediction of rock
mass damage due to accidental explosions in an underground ammunition storage chamber. Shock Waves 15(1),
4354.

992

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The measurement and management of Deca-BDEWhy the continued

certainty of uncertainty?
R.E. Alcock, B.H. McGillivray & J.S. Busby
Department of Management Science, Lancaster University, UK

ABSTRACT: The widely used brominated flame retardant Deca-BDE is currently the subject of a controversial debate regarding its environmental persistence and human health safety. As an effective inhibitor of fire in
numerous plastics and textiles is also has a measured and well characterized risk benefit to society. An inconclusive ten year EU risk assessment has fueled an ongoing search for evidence of harm. Within this context, the
need for accurate measurement of the chemical is necessary to determine if risk reduction efforts are reducing
environmental burdens and if the chemical could degrade to more toxic components. An analysis of interlaboratory comparison data has shown that many competent laboratories remain unable to accurately determine trace
levels of the compound in a variety of media. This inability to measure a chemical risk is problematic and the
associated uncertainty does little to support effective risk management.

1
1.1

INTRODUCTION
Polybrominated diphenyl ethers

There are numerous brominated chemicals which are

used extensively as fire retardants. Amongst these, a
group called polybrominated diphenyl ethers (PBDEs)
have been the focus of concern as environmental contaminants for over 25 years. Recent research interest
has focused on one particular product-Deca-BDE-one
of the most widely used brominated flame retardants.
Deca-BDE is added to a wide range of plastics, interior
textile materials for home, office and transport applications and reduces both the chances of the material
igniting and slows the rate of combustion. Its use in
the manufacture of electronic equipment, upholstered
furniture and textiles has an important benefit to society in that it has saved thousands of lives by reducing
the risk of fire (DTI 2000).
A 10 year EU risk assessment highlighted no direct
risks to human health (ECB 2004) but the inconclusive
environmental risk assessment meant Member States
agreed a further 10 year research agenda to answer
some key questions in relation to environmental fate,
behaviour and possible toxic effects. This phase is well
underway and the scientific community, including the
bromine industry is gathering evidence on environmental trends, toxicological impacts and the extent
and significance of debromination of Deca-BDE to
potentially more toxic constituents. Within this context
the chemical continues to be the subject of controversy, intense pressure group activity and industrial
lobbying.

Key to the current evidence-gathering phase for

Deca-BDE is accurate measurement of ultra-trace levels in a variety of environmental media and human
tissues. The need to find evidence for reduced emission inventories, to define trends in a variety of
environmental media and human tissues and to carefully characterize metabolic breakdown pathways and
products will be critical factors in any subsequent
political decisions regarding Europe-wide restrictions
on the use of Deca-BDE or a complete production
phase-out. This evidence from numerous scientific
laboratories will form the backbone of the informational input upon which the political decision-making
process will synthesise different viewpoints, perspectives and interests and represents the starting point
from which the science and elements of judgment
integrate together.
The need for reliable measurement results that can
be compared across space and time has never been
greater. But analysis of Deca-BDE is very expensive, time consuming and cannot be carried out in
tandem with other more traditional compounds or
incorporated within conventional multi-residue methods because it requires far higher analytical specificity
and occurs at lower concentrations than traditional
organic contaminants. It has been described as one
of the most problematic chemicals to determine accurately. Reliable measurements depend critically on
validated and tested methods and one such measure
is the use of interlaboratory exercises to provide a
measure of accuracy across a range of competent laboratories. In this study, the available data on such

993

exercises has been carefully studied to provide an

indication of the historical and recent ability of key
laboratories to meet the demands of the regulators. In
addition, interviews with key scientists have provided
important background material.
2

MEASUREMENT AND UNCERTAINTY

Uncertainty is an unavoidable part of any measurement. It is a parameter that defines the range of
the values that could reasonably be attributed to the
measured quantity and when uncertainty is evaluated and reported in a specified way it indicates the
level of confidence that the value actually lies within
the range. The uncertainty is a quantitative indication of the quality of the result. It gives an answer
to the question, how well does the result represent
the value of the quantity being measured? Because
uncertainty clouds analytical results, it is one of the
most critical parameters characterizing an analytical
determination.
3

elements. Several review papers highlight the degree

of this variability (see Covavi et al., 2007; Stapleton
et al., 2004; Stapleton et al., 2007).
BDE-209 is extremely difficult to analyse for a
variety of reasons. Data may both overestimate true
values as a result of laboratory contamination and
underestimate as a result of losses during clean-up
and analysis. For example, the compound is not stable and so disappears at higher temperatures during
gas chromatographic separation and is also sensitive
to degradation by UV light (Sellstrm, 1999). Early
reports of the compound in environmental samples
were initially thought to be the result of laboratory contamination and therefore dismissed . As one analytical
chemist recalled:
People have only really seriously started to
measure Deca in the last year or twothere
was lots of data beforebut most of it likely
to be incorrect . . . you need a whole separate
method and measurement regime (Analytical
Laboratory Manager. UK)
Current methods use carbon-13 isotope labeled recovery standards which allow quantification of possible
loses during extraction, clean-up and analysis. As one
scientist recalls:

THE ANALYTICAL CHALLENGE

OF DECA-BDE

Within the last 3 years global PBDE research has

focused on Deca-BDE and in particular the search for
evidence related to debromination of the parent compound to more toxic constituents and sources/impacts
of human exposure (TSI, 2007). One of the greatest challenges for scientists has been the development
of methods to accurately quantify the principal component of the Deca-BDE mixtureBDE-209. While
analytical methods are readily available for quantifying lower brominated congenersakin to more
traditional organic chemical analysis, the analysis of
higher brominated compounds such as BDE-209 has
proven to be very different and consequently difficult
to master. As a result, a limited number of research
groups worldwide dominate the data generation and
consequently the literature.
The analytical procedures developed for BDE-209
analysis must be specifically targeted, multi-stepped
and carefully controlled to minimise both contamination from external sources and losses from within
the sample itself. Briefly, at least 15 individual stages
define the methodology including extraction of the
compound from its environmental matrix, concentration of the extract, tedious purification steps to remove
unwanted chemical elements within an extract and
then careful fractionation prior to chromatographic
separation and detection. Whilst several elements are
common to this outline methodology there is little harmonization of analytical steps between laboratories,
with each preferring their own optimized features and

Around 2004, the analytical technology for

Deca got better, and now everybody is interested in measuring it everywherein animals
like polar bears and penguins. If you went to conferences in 2004 on brominated flame retardants,
there wouldnt have been much discussion about
deca BDE, but at the last conference I went to in
April of 2007, at least half the talks were about
it. (Senior Research Scientist, US)
Laboratory contamination has been highlighted as an
important factor by several of those interviewed. For
example:
Deca has been difficult to analyze, mainly
because a lot of laboratories are contaminated
with this compound. You need new glassware
for each batch of samples which really adds
to your costs. Weve been fortunate enough to
work in a newly renovated building, which meant
we could control all the materials being used
in the construction. So weve had very low
blank levelsno contamination problems at all.
(Analytical Laboratory Manager, Sweden)
Most scientists interviewed were committed to further
research on Deca-BDE for a variety of different reasons including ongoing production, ongoing indoor
exposure issues, concerns surrounding health effects

994

Table 1.

Summary of levels of BDE-209 in human blood collected from general populations.

Survey

Total serum samples

No. of positive
detections

WWF 2003
(general public in the UK)

155 (105 females/50 males)

11 (7%)

WWF 2004a
(seven families in the UK)

33 (6 grandmothers, 7 mothers,
6 fathers and 14 children)

7 (21%)

WWF 2004b
(MPs from 13 European Union
countries)

14 (3 females /11 males)

3 (21%)

WWF 2004c
(general public in Europe, 17
countries)

47 (23 females/24 males)

16 (34%)

Fngstrm et al., 2005

(pregnant mothers and children
from the Faroe Islands)

(samples from children were

taken at age 7)

Fischer et al., 2006

(family of four from California, US)

8 (mother, father, daughter and son)

8 (100%)

Link et al., 2005

(German children)

8 pooled samples from 437 children

Schecter et al., 2005

(US blood donors)

39 (17 females/22 males)

18 (46%)

SFT, 2005
(pregnant women in Russia and Norway)

10 (pregnant women)

0 (0%)

Sjdin et al., 2001

(US blood donors)

12 (gender unknown)

5 (42%)

Thomas et al., 2006

(general public in the UK)

154 (gender unknown)

11 (7%)

Weiss et al., 2006

(wives of fishermen, Sweden)

50 females

40 (80%)

Veith et al., 2005

(mothers from Germany

7 mothers

0 (0%)

and problems with global transport, recycling and

recovery of Deca-BDE containing plastics in the Far
East. Others felt that there had been an excess of data
generation over the last 10 years which was a little
over the top.
Notwithstanding some progress in the quality of
data generated, Table 1 demonstrates how infrequently
BDE-209 has been positively confirmed in a literature
search of levels in human serum samples. Presence of
the chemical in human tissues has been an important
driver for lobbying against its continued use. Within a
given sampling pool, frequency of positive detection
was variable and often very low, ranging from 0 and
100% for human serum, suggesting a combination of
uneven exposure routes and analytical difficulties with
the compound itself.

Despite the low detection rates in some studies,

the focus of each of the papers tabulated here was
the positive presence of BDE-209 in human tissues and subsequent heightened risk concerns this
prompted. Often, the positive presence of the chemical is barely above the background noise in the
sample. Poor positive detection rates and their associated closeness to the analytical limit of detection should ensure laboratories are careful with
the language chosen when reporting their findings.
This rarely happens in practice as does subsequent
correction and/or withdrawal of information which
is erroneous or invalid. The result is a pool of
non uniform, open textured numerical information that is difficult to subsequently characterize and
evaluate.

995

INTERLABORATORY COMPARISONS

Reliable measurements depend critically on validated

and tested methods, comprehensive quality systems,
and traceability to appropriate measurement references. A key step is method validationthe process
of proving that an analytical method is acceptable for
its intended purpose. It can be a time consuming and
tedious process and demands detailed elements such
as specificity (i.e. the ability of the method to accurately measure the analyte response in the presence of
all potential sample components), accuracy (i.e. closeness of the measured value to the true value for the sample), precision (i.e. the amount of scatter in the results
obtained from multiple analyses of a homogeneous
sample), range (i.e. the concentration interval over
which acceptable accuracy, linearity, and precision
are obtained) and detection limit (i.e. the lowest analyte concentration that produces a response detectable
above the noise level of the system). Although there is
general agreement about what type of studies should
be done, there is great diversity between laboratories
in how they are performed.
4.1

Key interlaboratory studies

A detailed historical analysis of results from a

variety of recent checks on world-wide laboratory
performance of PBDEs over the last 10 years has indicated consistent poor performance from many participants. This has occurred despite detailed optimisation
information repeated to all participants. Laboratories with poor performance suffered from multiple
difficulties throughout their chosen methods which
hindered clear identification of the sources of error.
The European studies are briefly outlined below.
The first world-wide interlaboratory study on
PBDEs was organised in 1999 and funded by the
bromine industry. It involved five biological samples,
two sediments and two standard solutions with materials sent to 26 participant laboratories in nine different
countries (de Boer, 2001). Results were returned from
18 laboratories (70% of those that agreed to participate) from 9 countries. The emphasis of the study
was on the between-laboratory agreement so only one
result per determinant per sample was required. The
participants were asked to use their own laboratory
methods and some advice was given on how to avoid
specific errors during the determination. A period of
4.5 months was given to the participants to complete
this work.
The results for only one single congener-BDE 47
were considered satisfactory with a range of Relative Standard Deviations (Rsd) of 1740% but the vast
majority of congeners were below the detection limit
of most laboratories that returned results. At the time
of this first intercomparison, the capacity to measure

BDE-209 was poorly established and interlaboratory

variation varied by over 2 orders of magnitude with
the authors concluding that the analysis of BDE-209
was not under control in most laboratories.
A second international interlaboratory study was
organized in March 2002. A wider set of sample analytes were selected in addition to a broader range
of sample types including lake trout, mussels, sediment, and human milk. The study was a collaborative
project of the Bromine Science and Environmental
Forum (BSEF), and The Netherlands Institute for Fisheries Research (RIVO). A third interlaboratory study
was organised in 2003. This study was carried out
by QUASIMEME (Quality Assurance of Information
for Marine Environmental Monitoring in Europe) in
collaboration with RIVO. The target compounds were
identical to those of the second study. A fourth study,
also organised by QUASIMEME and RIVO, was conducted in 2004. The most recent QUASIMEME study
in 2007 allows trends in performance data over 8
years to be evaluated. These are difficult to identify
in this type of dataset because there are a number of
parameters that vary from round to round including
the number and the identity of the participating laboratories, the concentrations of the determinants in
the test material, and the complexity and variations in
the analytical methods used. In summary, the following conclusions could be drawn: With the exception
of BDE 209, most congeners could be analysed in a
variety of matrices with realistic concentrations in the
range 0.1>10 ng/g, resulting in Coefficient of Variation (CV) values of about 530%. As expected, overall
lower analyte concentrations led to much higher CVs
so that at concentrations below 0.1 ng/g, CV values
tended to increase to between 30% and 100%. There
was poor agreement with the relatively simple task
of evaluating the content of an unknown ready prepared solution, highlighting the gas chromatography
quantification steptraditionally one of the more simple elements to controlwas problematic (Galban and
Ubide, 2007).
BDE-209 continues to be problematic for most
laboratories choosing to participate in these studies,
despite detailed advice aimed at minimizing potential problems being repeated to those taking part. The
most recent data (QUASIMEME 2007) report Rsds of
between 50 and 60% within the group of 18 laboratories that chose to return results (70% of the total).
The ongoing efforts of these unique interlaboratory
studies over a period of several years highlights the
continued problems that many of those participating
have in achieving an acceptable level of accuracy. As
one key scientist stated recently we still produce a
lot of rubbish data (Analytical Laboratory Manager,
Netherlands).
Similar poor results have been repeated in each
worldwide interlaboratory study that has included

996

BDE-209. For example, in 2004, NIST (US National

Institute of Standards and Technology) organized
an interlaboratory exercise for the determination of
PBDEs in two sediment standard reference materials
(Stapleton et al., 2005). Again, many of the participating laboratories struggled to attain acceptable
values for BDE-209. Another recent intercomparison study in Japan has also yielded similar results
(Takahashi et al., 2006). Unlike previous studies,
guidelines were given to participants to reduce specific known errors and meetings were held to improve
and standardize the methodology. RDS values after
the round of improvements were considered reasonable at between 10 and 38%. Within Europe, the
need for reliable Deca-BDE analysis in air has been
investigated as part of the NORMAN co-ordination
Network (see Duffek, 2007). Seven laboratories
from six European countries participated and two
other laboratories who were new to PBDE analysis also took part. Each laboratory applied their
own preferred analytical method and possible pitfalls and problems with the analysis of Deca-BDE
were discussed carefully with all participants at the
start of the project.
Despite this careful planning and attention to problem areas, one of the seven laboratories put forward completely incorrect resultsthe laboratory had
added different quantities of internal standard to calibration solutions/control extracts and sample extracts,
respectively, which resulted in a false quantification result. All participants used separation conditions
specifically optimised for the analysis of Deca-BDE
in accordance with the recommendations given in the
literature (e.g. Covaci et al., 2007). After elimination of outliers, the reproducibility and repeatability
variation coefficients were less than 10%, demonstrating that laboratories experienced in the analysis
of PBDEs are able to determine Deca-BDE accurately even though they applied a variety of methods.
Therefore, the choice of the analytical method is
less important than the experience of the laboratories and the careful control of critical factors like
thermal and photochemical degradation of Deca-BDE
as well as attention to blanks. However, the detailed
instruction protocol provided with advice on all critical aspects of the methodology was not sufficient to
enable inexperienced users to establish a fully validated method for the determination of Deca-BDE
within the given time frame. These results taken
together provide evidence that several years of careful method development are needed before a sufficient
number of laboratories can produce reliable results for
a relatively new contaminant of interest. Laboratories
experienced in other organic contaminants would need
a sustained period of refinement to be able to attain the
required levels of accuracy for a difficult analyte such
as BDE-209.

4.2

Other evidence of problems

Poor analytical reproducibility between laboratories is

not confined to BDE-209. Data from other recent interlaboratory comparison studies suggest the problem is
widespread, particularly for the newer compounds
of interest. For example, measurement of toxaphene
in air (Bidleman et al., 2004) and polychlorinated
naphthalenes (Harner et al., 2003) have given Rsds of
between 20 and 65%. A first inter laboratory comparison for perfluorinated compounds (Van Leewen et al.,
2006) demonstrated that a pool of 36 participating laboratories, several with a history of analyzing these
compounds for several years were unable to produce
consistent data.
A comparison with dioxin analysisnow essentially a mature method with well characterized guidelines for optimization of performancedemonstrates
the continued problems with method performance over
time. Over a 10 year period from 1994 to 2005, Rsds
have increasedrather than decreasedfrom 25% to
75% as the number of laboratories participating in
global comparison exercises increased from 5 to 76
(UNEP 2005). Significantly, little is made of these
continued problems in the report although they are
acknowledged and partially blamed on the inability
of analytical laboratories in developing countries to
compete with those in Europe and the US which have
had a long evolving history of dioxin analysis. Such
sustained uncertainty between data produced from different laboratories makes the interpretation of pooled
data sets within risk assessment extremely problematic. Equal weighting is routinely given to all analytical
supporting information put forward as input to regulatory decision-making and a weight-of-evidence
approach applied. For example, based on our analysis defining environmental trends of BDE-209-an
important measure of recent efforts to limit chemical releaseswill be extremely difficult using data
from a contemporary pool of competent laboratories.
At current levels of dispersion around mean values
(between 50 and 60%), trends would be impossible to
see unless environmental concentrations were either
declining or rising over a given period by at least 60%.

CONCLUSIONS

The management of Deca-BDE will rely to a large

extent on the scientific evidence related to its potential
for harm in the environment. Coefficients of variation within studies designed to measure uncertainty
in measurement methods are higher than they need
to be at between 30100%. Importantly, the general inability to quantify levels accurately is rarely
cited in the literature or taken into account within the
risk assessment process. This makes comparisons of

997

data between laboratories and the search for subtle

environmental trends virtually impossible and fraught
with uncertainty at a time when risk managers demand
definitive answers to make regulatory decisions based
on complete information. Given the political interest in proposing a ban on Deca-BDE within Europe
and several US States, the inability of many laboratories to produce reliable and accurate numerical
concentrations is problematic. Evidence from interlaboratory studies provides a picture clouded with
uncertainty, an uncertainty that is rarely identified in
individual studies, suggesting analytical science is far
from providing definitive answers. Including uncertainties several orders of magnitude greater associated
with the measurements of chemical effects and the
science of toxicology adds an even greater haze to any
subsequent risk calculation. Until uncertainties surrounding the measurement of risk become resolved
the management of Deca-BDE will be compromised.

ACKNOWLEDGEMENTS
The authors would like to thank The Leverhulme Trust
for funding research on characterising the nature of
risk migration processes. www.riskmigration.org

REFERENCES
Bidleman, T.F., S. Cussion and J.M. Jantunen. 2004. Interlaboratory study on toxaphene in ambient air. Atmos.
Environ. 38: 37133722.
Covacci, A., S. Voorspoels and L. Ramos. 2007. Recent
developments in the analysis of brominated flame retardants and brominated natural compounds. J. Chromat A.
1153: 145171.
De Boer, J. 1999. Capillary gas chromatography for the
determinatio of halogenated micropollutants. J. Chromatography A. 843: 179198.
De Boer, J., C. Allchin, R. Law, Bart Zegers and J.P. Boon.
2001. Method for the analysis of PBDEs in sediment and
biota. Trends in Analytical Chemistry, 20: 591602.
De Boer, J. and W.P. Cofino. 2002. First world-wide interlaboratory study on PBDEs. Chemosphere, 46: 625633.
De Boer, J. and W. Wells. 2006. Pitfalls in the analysis
of brominated flame retadants in environmental, human

and food samples. Trends in Analytical Chemistry, 25:

364372.
Duffek, A. 2007. NORMAN Network of reference laboratories and related organisations for monitoring and
bio monitoring of emerging environmental pollutants
Deliverable number C3.2 Report on the 1st interlaboratory study C3-I.
DTI 2000. Effectiveness of the Furniture and Furnishings
Fire Safety Regulations 1988 second report of UK
Department of Trade and Industry June 2000.
Galban, J. and C. Ubide. 2007. Uncertainty due to the
quantification step in analytical methods. Talanta 71:
13391344.
Harner, T. and J. Kucklick. 2003. Interlaboratory study for
polychlorinated naphthalenes (PCNs). Chemosphere: 51:
555562.
QUASIMEME Laboratory Performance Studies. 2007.
Round 48Exercise 748. BS-1 Brominated flame retardanats.
Stapleton, H.M., Keller, M., Schantz, M. M. and
Kucklick, J.R. 2007. Determination of polybrominated
diphenyl ethers in environmental standard reference materials. Anal. Bioanal. Chem., 387, 23652379.
Stapleton, H.M. 2006. Instrumental methods and challenges in quantifying PBDEs in environment al extracts:
A review. Bioanal. Chem., 386, 807817.
Stapleton, H.M., Keller, J.M., Schantz, M.M., Kucklick, J.R.,
and Wise, S.A., NIST Intercomparison Exercise Program for Polybrominated Diphenyl Ethers (PBDEs)
in Marine Sediment: Description and Results of the
2004 Intercomparison Exercise, NISTIR 7278, NIST,
Gaithersburg MD (2006).
Takahashi, S. 2006. An intercalibration study on
organobromine compounds: Results on PBDEs and
dioxin-related compounds. Chemosphere, 64: 234244.
Thomson Scientific (TSI) 2007. Essential Science IndicatorsPolybrominated Diphenyl Ethers.http://www.esitopics.com/pbde/index.html
UNEP Chemicals, 2005. International Intercalibration Studies: a global QA/QC tool for the analysis of Pops under
the Stockholm Convention.
US-EPA, 2003. Draft Method 1614, Brominated diphenyl
ethers in water, soil, sediment, and tissue by
HRGC/HRMS, EPA Office of Science and Technology,
August 2003.
Van Leeuwen, S. A. Karrman, B. Van Bavel, J. De Boer and
G. Lindstrom. 2006. Struggle for quality in determination
of PCFs in environmental and human samples. Environ,
Sci Tcehnol., 40. 78547860.

998

Risk and hazard analysis

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A contribution to accelerated testing implementation

S. Hossein Mohammadian M. & Daoud At-Kadi
Department of Mechanical Engineering, University of Laval, Quebec, Canada

Amadou Coulibaly & Bernard Mutel

INSA, Strasbourg, France

ABSTRACT: Accelerated testing is an effort to quickly fail products in order to obtain failure-related data. This
paper presents an implementation approach of accelerated testing in industry. It introduces test samples, stressors
and performance factors as basic elements. Sample size calculation, failure acceleration strategies and statistical
concepts as main auxiliary tools are elaborated. In order to prepare an estimation of service life to compare
with test data, different types of field failure data with their expected results are studied. Finally, available and
possible qualitative and quantitative accelerated testing methods are presented according to required results.
They are categorized in design modification methods, extrapolation methods, using rate methods, performance
testing-related method, and comparison method.

INTRODUCTION

From the economic point of view and customers

expectations, durability and quality (performance) of
products are as important as their functionalities. The
estimation of products life and performance could be
conventionally performed by life testing that aims to
present failure-related information via testing. Time
limitation and severe competitive market lead companies to carry out accelerated testing (AT) that takes
shorter time compared to conventional life testing.
AT is a multipurpose approach in which the main
reason of testing is necessary to be identified before
testing. It is not a pure practical approach. In fact,
it should be supported by theoretical aspects of statistical analysis (e.g. Nelson (1990)). Depending on
stressors, the relationship between stress level and life
might be formulated according to available models
(e.g. Escobar and Meeker (2006)). In real applications, such models are applied to extrapolate life test
data (obtained by AT), to normal level in order to estimate service life (e.g. Koo and Kim (2005), Zhang
et al. (2002), Wu et al. (2003)). Residual life for
service-exposed components could be estimated by
performing an accelerated performance testing. For
example, Marahleh et al. (2006) used stress-rupture
test to mathematically relate life of samples with their
time to rupture. If there are not enough samples, such
testing could be performed on some specimens taken
from the product in service to compare or validate

residual performance over operating time in service

(e.g. Drew et al. (2006)). A successful accelerated
testing is able to detect failure modes, and to verify
the positive effect of modification on components if
there is no field data (e.g. Murphy and Baxter (2004)).
If field failure data is available, AT might be used to
estimate products lifetime before and after any modifications (e.g. Park et al. (2006)). The comparison of
field and test data may result a correlation between
service and testing time which have same effect on
failure (e.g. G. Ranganathan et al. (2004)). AT is
also capable to compare several aspects of design (e.g.
G. Ranganathan et al. (2004)), materials (e.g. Knox
and Cowling (2000), Chou and Lamers (2005)), and
operating or environmental stressors (e.g. Yang and
Ding (2006)). Confidence level of expected reliability could be validated by testing time and number of
required samples according to cumulative distribution
functions obtained by accelerated testing (e.g. Park
et al. (2006), Nagel et al. (2004)).
Variety of above applications and the need of a
comprehensive reference for testers and manufacturers motivated us to propose a practical approach which
is supported by statistical and practical methods. The
paper is organized as follows: section 2 introduces
required definitions and main concepts of AT, and it
has an overview of auxiliary tools. The main part of
this paper, i.e. accelerated testing methods is discussed
in section 3. As conclusion, a general procedure of our
approach is presented in section 4.

1001

BASIC ELEMENTS AND TOOLS

In simple words, accelerated testing is defined as any

life testing in which failure-related data is expected
to be obtained by the strategy of decreasing time of
testing.
A function of a product is its ability to perform a
specified mission. In contrast, related failure mode
(or briefly failure) is defined as the disability of performing that mission as expected. Failures are divided
into two main types which are Catastrophic failure and
degrading failure. A catastrophic failure causes stopping in operation and could be recognized by simple
apparatus and our senses, whereas a degrading failure
results a decrease in quality of operation and needs to
be verified by testing in different time intervals.
For a product, selecting an appropriate accelerated
testing methodology strongly depends on its failure
type. If the expected failure of the product is catastrophic, required accelerated testing is called accelerated life testing (ALT) that aims to investigate the
time in which the sample is not able to perform its function any longer. Unlike ALT, accelerated stress testing
(AST) is intended to obtain the rate of degradation of
the product during testing.

2.1 Test sample

Test samples (or briefly samples) as one of the most
important elements for each AT, must be identified
from the viewpoint of their levels, types and their life
as follows.
Samples are recognized according to their levels
which are system level, component level and material
level. The selection of a proper sample level strongly
depends on expected results. For example, accelerated testing at system level (i.e. a machine including
many components) could present latent failure modes
and possibility of verification of positive effect of redesign on its components, whereas the aim of testing
at component level may result some quantitative information about their life. At material level, deeper study
is done to obtain degradation rate of different materials under various environmental stressors. One of the
major challenges at material level is to select suitable
shape and size of test samples (specimens) that could
represent a typical pattern of field samples.
From the viewpoint of age, test samples are divided
into new samples, aged samples and field samples.
A new sample could be made as a final product to market or as a prototype. Aged samples are new samples
which have already passed an artificial aging process.
Field samples refer to the ones which have been used
in service. They are used to obtain field related data. In
some literatures, field samples refer to used samples
and service-exposed samples.

Age of a product might be expressed in terms of

time, cycle or any other parameters related to the quantity of using in field or test. Life (L) is defined as time
period (lifetime) or number of cycles (lifecycle) from
starting the operation (in field or testing) until failure.
The most popular explanation of life for a group of
samples are B10 (reliability = 90%), B50 (median or
reliability = 50%) and mean time to failure (MTTF).
There are four types of lifetime data which are failure
time (complete data), right censored, left censored and
interval censored (ReliaSoft (2001)).
2.2 Stress factor
Design, material and manufacturing process affect life
and degradation rate of a product and they are almost
unique for all manufactured products of a particular model, but on the other hand, stress factors or
simply stressors which are defined as operating or
environmental physical forces on a product in the field
could have different effect because of difference in
customers using manners, applications and environment. As mentioned in above definition, stress factors
are divided into operating stressor and environmental
stressor. For instance, temperature and UV irradiance
are environmental stressors and abrasion is operating
stressor for a car seat belt webbing (e.g. Koo and Kim
(2005)). Any combination of stress factors either in
normal or high level is called stress configuration.
If all stressors are in normal level, its called normal stress configuration and if at least one of them
is in accelerated level, it is called accelerated stress
configuration.
2.3 Acceleration Factor (AF)
The acceleration factor is a unitless number that
relates a products life at an accelerated stress level
to the life at the use stress level (ReliaSoft (2001)):
AF =

Lu
La

(1)

where Lu and La are service life in use conditions and

testing life in accelerated level of stress respectively.
It is concluded that this factor must always be much
higher than 1 and indicates how effective the test is
to reduce the time to meet required failure. AF is a
characteristic of AT in that stress configuration, and it
is assumed that it has same value for the products in
different versions of design or manufactured by other
companies.
2.4 Performance factor
The main assumption of each AST is that the quality of samples under test decreases over time. In

1002

order to reveal the rate of degradation, such quality

must be quantified. The factor which expresses the
performance of products is called performance factor (PF). Each performance factor must provide three
requirements: firstly it should be a reasonable representative of quality in long-term, i.e. as it has already
mentioned it has to be decreased continuously over
time of testing, secondly it must be measurable, and
finally the reduction in quality should be sensitive
enough during testing time in order to observe its variation in different time intervals. As a substitution of PF,
deterioration of materials could be defined by degradation factor (DF) which increases (unlike PF) during
testing time.
Performance factors are measured by performance
testing (PT) which should be done during accelerated
stress testing in different time/cycle intervals. PT could
be performed in each level of aging process; at the
beginning, in the middle or at the end of AST. From
the viewpoint of testing nature and required number
of samples, PT is divided into destructive and nondestructive test types.
Thespecimenthathaspassedadestructiveperformance
test is no longer acceptable to continue aging process
because destructive nature of such testing causes a
damage that affects on the endurance of the specimen.
So the main conclusion of such testing is to prepare
more specimens in parallel to pass the aging process.
Accelerated creep test in order to obtain time to rupture
(e.g. Drew et al. (2006)) and scrape test in order
to obtain shear failure load (e.g. Knox and Cowling
(2000)) are considered as destructive performance test.
Unlike destructive performance test, each specimen
that has passed a non-destructive performance test is
able to continue other aging processes. Measurement
of wear (e.g. Chou and Lamers (2005)), pressure and
volume (e.g. B et al. (2007)) are some examples of
non-destructive performance test.
Each AST could be transformed to ALT by defining
an acceptable minimum level of PF ((PF)min ) in which
the operation is satisfied. The simplest estimation of
life of an aged product in terms of its performance
factor is to consider a linear relationship as shown in
Fig. 1 as follows:
(Age)2 (Age)1
(PF)2 (PF)min
+
=1
L (Age)1
(PF)1 (PF)min

(2)

where (Age)1 and (PF)1 are age and performance factor of first aged sample, and (Age)2 and (PF)2 are
those ones for second aged sample, and L is estimated
life of samples.
2.5 Statistical analysis
Cumulative Distribution Functions (CDF): Consider
there are some discrete life data as presented in

Figure 1.

Linear relationship between age and PF.

Figure 2.

CDF diagrams for test and field data.

subsection 2.1. By choosing a specified probability distribution function (PDF) like weibull, normal,
or lognormal distribution, and by using graphical
method, least square method or maximum likelihood
estimation (e.g. ReliaSoft (2001)), unknown parameters of PDF function and consequently CDF function
could be obtained. For a product, CDF diagram for test
locates on the left side of field CDF diagram and they
are assumed to be parallel as shown in Fig. 2. In fact,
AF is the relation between these two diagrams and its
simply expressed as:
AF =

(B10 )field
(B50 )field
=
(B10 )test
(B50 )test

(3)

Confidence level: If the target is to achieve reliability of R in time t, the problem is to obtain
minimum required sample size (N ) needed for testing, if number of maximum allowable failed samples
is r, so N r samples must survive after testing

1003

time t. The relationship between above parameters is

mathematically expressed as (Park et al. (2006)):
1C =

r

(Nx ) R(t)N x (1 R(t))x

(4)

x=0

where C is confidence level and its most applicable

values in life testing are 90% and 95%. Validation
of reliability is a challenge between sample size (N )
and testing time (t), i.e. if there are not many samples
available, the test must be done for longer time, vice
versa, if the aim is to validate required reliability as
soon as possible, more sample size is needed. In every
case, the value of R(t) for above formula is obtained
by CDF diagram.

A sample could be considered as a source of

acceleration by modifying its geometry or creating a
defect on it before testing. For example, small specimens result rapid degradation and economic benefits
(e.g. Apostolopoulos and Papadopoulos (2007)), and
changing their geometry could enable a rapid exposure and consequently a quick aging (e.g. Knox and
Cowling (2000)). In catastrophic failure a small crack
may accelerate the propagation process of the crack.
Any other approaches, which are indirectly stimulants
for failure mechanism, might be considered as artificial methods. For instance, the form of test equipment
could stimulate degradation mechanism (e.g. Chou
and Lamers (2005)), and a talc may spoil lubrication
(e.g. G. Ranganathan et al. (2004)).
2.8 Field-related analysis

2.6

Sample size

Sample size refers to the total number of samples

required for testing. Sample size estimation is usually a complicated problem, because it is a function of
many related items in AT.
In AST, the number of stressors, number of their
levels, number of time intervals (only for destructive
performance test), and outdoor exposure testing definitely increase sample size, whereas, the possibility
of cutting materials to smaller pieces than real one
(in material level) in order to economize on materials,
considerably decreases sample size. Replicates, which
are a group of samples which pass same test to make
a CDF diagram and to verify a confidence level, also
increase required number of samples. If the aim of
AT is to compare different samples in terms of their
design, manufacturing processes, and vendors, more
samples must be prepared for each item.
2.7

Acceleration strategies

Each accelerated testing method must include an

approach to decrease the time of testing. Regardless
of the approach which is used for acceleration, it must
result relevant failure modes.
The lifetime of a product depends on total time
duration of its operation. For the ones which are used
less frequent in fields, continuous operation in testing
could considerably decrease testing time to failure. On
the other hand, increasing the level of stressors causes
a higher failure rate. This is the most known source
of acceleration in AT and the base of extrapolation
methods.
Especially for destructive PT, the testing time might
be decreased by employing rapid destructive methods.
For instance, instead of usual creep test, a stressrupture test is used to perform the test in shorter time
(e.g. Marahleh et al. (2006)).

Estimation of service failure data is vital for many

AT methods. They have different formats and must be
transformed to useful diagrams.
2.8.1 Catastrophic Failure Analysis
2.8.1.1 Random samples
If there are many field-returned samples, for a group
of random samples, regardless warranty period, their
life make a CDF diagram. This analysis would not
be accurate because the distributions of field-returned
samples in warranty and out of warranty period are not
the same.
2.8.1.2 Warranty and out of warranty period
This analysis is based on obtaining two CDF diagrams
which must be fitted. To obtain the first diagram,
failure data for warranty period are considered as complete data (n1 ), and all other failures are considered as
right censored. For the second diagram, all available
failure (in and out of warranty period until a specified
time) are considered as complete data (n1 + n2 ). Other
unavailable field samples are considered as unreported
failure before specified time (n3 ) and survivors, i.e.
right censored (n4 ) after specified time as shown in
Fig. 3. The CDF diagrams obtained by these two
approaches must be equal, so the value of n3 and n4
could be estimated by:
n = n1 + n2 + n3 + n4

(5)

where n is total number of production for this particular

product.
2.8.2 Degradation failure analysis
If service-exposed samples or aged samples are available, by performing PT, their PF are measured in
terms of their ages as shown in Fig. 1. For better
estimation, PF for some new samples has to be
measured.

1004

to obtain a correlation between required substitution

level of mechanical load (e.g. XIE (2000)).

Figure 3.
period.

Availability of failure during and after warranty

2.8.2.1 PF-Age diagram

The PF data can be analyzed by least square method
for given function (e.g. (PF) = A.eB.(Age) + C) to
present unknown parameters. Such diagram might be
used to compared with the results obtained from AT.
2.8.2.2 CDF diagram
To transform a PFAge diagram to CDF diagram,
minimum value for PF as failure point must be
defined. Scatter data for life of samples would be
obtained by using a linear relationship as eq. (2) for
new sample and every field sample.
2.9

Finite element method

Finite element method (FEM), as a theoretical aspect

of mechanical stress-strain analysis, by using computer aided design (CAD) has great distinctions in
accelerated testing in which stress factors are mechanical forces, temperature and vibration.
For newly-designed components, the location of
failure is difficult to estimate because of new aspects
of design and lack of field data. Static finite element
analysis can present the distribution of stress in the
component due to a particular load. Hot spots, which
are some local elements under high stress/strain in
finite element models, could be discovered by using
Von-Mises criterion or other theories. The hot spot(s)
could verify and validate the failure location obtained
by accelerated testing (e.g. Su et al. (2004), Wang et al.
(2000), Wu et al. (2002)).
The main assumption in static FEA is to consider
a constant level of stressor, however, this level varies
between a maximum and a minimum value, so static
FEA would not be sufficient for such conditions. The
ability of software to perform dynamic analysis could
result the distribution of stress in hot spots (obtained
by static analysis) during time in order to find out
real maximum stress and crucial hot spot. If life-stress
material diagram is available, the life of the component
could be estimated (e.g. Su et al. (2004)).
Thermal cycling in life testing relatively takes long
time than mechanical load cycling to meet required
failure, so the possibility of substitution of mechanical
loads instead of temperature in test could considerably
decrease time of testing. FEA is able to validate such
replacement by presenting stress-strain distribution of
crucial hot spot under thermal and mechanical loads

AT METHODS AND ANALYSIS

This section is devoted to introduce AT methods and

analysis which are categorized in using rate methods,
design modification methods, extrapolation methods,
PF-related methods and comparison methods.
3.1

Using rate methods

The most popular way to increase the rate of failure is

to raise its using rate which does not need any verification of relevant failure modes. Fatigue testing is a kind
of such testing. The main concept of this method is
based on calculating AF which is simply expressed as:
AF =

(Usage)a
(Usage)u

(6)

where (Usage)a and (Usage)u are the estimation usage

rate of operation in accelerated testing and in service respectively, and they could be expressed in
hour/week, cycle/day, etc.
3.1.1 Reliability calculation
If n samples are tested in same conditions as actual
ones in field, their test life make a CDF diagram, then
(B10 )field could be simply calculated by:
(B10 )field = AF.(B10 )test

(7)

Consequently CDF diagram for field is easily obtained

by drawing a parallel line.
3.1.2 Verification of reliability
In order to have field reliability of R in time tfield ,
with confidence level of C, accelerated testing must
field
be done for the period of time equals to ttest = tAF
for N new samples with r failures according to eq. (4).
The problem is to obtain sample size N needed for
testing, so N r samples must survive after time ttest .
If test CDF diagram is available, suitable sample size
could be calculated in terms of testing time.
3.2

Design modification methods

In the process of design when the products are expected

to be manufactured in near future, modification of
design to remove their latent failure modes is a vital
matter to be considered. On the other hand, after production and marketing, regarding existing failure types
in the field and customers expectations, the need of
re-design is arisen. In both cases, life testing must be
performed in order to validate the positive effect of

1005

new design on products performance. Most of design

modification methods are two-stage testing, namely
the test should be done in two stages, before modification (on recently-developed samples) and after
modification (on newly-developed samples).
3.2.1 Qualitative two-stage method
The main concept of this method is firstly to test a
machine (system level) in order to get its failure modes.
The test might be preferably done in normal conditions
and in high using rate. If there is any failure during
test, related component should be replaced by a new
one. After the end of first stage, root cause analysis must be performed for every failure to remove it
from the machine. Second stage of test should be performed like first one, and in same period of time. At
the end of this stage, a comparative analysis about
number of replaced components is done to verify the
modifications that have been done.

the test has to be continued. For mechanical machines,

the value of stress levels and testing time are unknown
and they could be identified by preliminary testing and
literatures.
3.3 Extrapolation methods
If the failure acceleration method is based on increasing operating or environmental stressor, the possibility
of estimating service life by available test life could be
done by life-stress models which intend to extrapolate the data obtained from high stress level to normal
stress level. For this purpose, the test must be done in
some different high levels of stress, and for each level,
several new samples are required as replicates.
3.3.1 Life-stress models
Among many available life-stress models (e.g. Escobar and Meeker (2006)), Arrhenius model and inverse
power law are presented here.

3.2.2 Validation of modification

At component level, where there is no field data, accelerated testing on both versions of design results failure
time and consequently CDF diagrams (which are
assumed to be parallel). The comparison between their
life reveals the effect of modification and increasing
rate of service life as follows:

Arrhenius model: This is the most important model

which is used only for temperature as stressor to predict
service life with high accuracy. This model is simply
expressed by:

(Lu )2
(B10 )2
=
(Lu )1
(B10 )1

where T is absolute temperature (K), L(T ) is estimated

life in temperature T and, B and C are two unknown
constants to be obtained. Acceleration factor for each
level of temperature is expressed by eq. (1):

(8)

where indexes 1 and 2 relate to last and new versions

of design, Lu and B10 are service and test life.
3.2.3 Quantitative two-stage method
At component level, if field data is available, comparison of test CDF diagram with field CDF diagram in
first stage would result AF. Here, the main assumption is that AF before and after modification has same
value: (AF)1 = (AF)2 , so by performing AT for
second stage (after modification) and obtaining its
test CDF diagram and knowing AF, new field CDF
diagram would be estimated.
3.2.4 Highly Accelerated Life Testing (HALT)
HALT, as a multi-stage modification method at system level, is done in some different accelerated stress
configurations of temperature, vibration, and humidity. For each stress configuration, the level of stressor
should be increased gradually to meet failure. Because
the test is not in normal conditions, the obtained failures must be investigated according to experience in
order to differ relevant and irrelevant failure modes.
In HALT, testing must be stopped in the case of any
failure, root cause analysis must be done to remove
the failure from the machine, new version of design
must be applied to produce new version of sample, and

L(T ) = Ce T

(9)

AF(Ta ) = eB( Tu Ta )
1

(10)

where Ta and Tu are high level and normal level of

temperature respectively.
Inverse power law: This model is presented for nonthermal stressors and simply relates life of products to
levels of stressors in normal and accelerated levels:
Sa
Lu
= ( )m
La
Su

(11)

where Sa and La are the value of the stressor and life

in accelerated test, Su and Lu are those ones in normal
conditions, and m is an unknown parameter must be
identified. The left-hand side of this equation clearly
expresses acceleration factor according to definition
of AF in eq. (1).
3.3.2 Life estimation
If the data are available in accelerated stress, service
life could be estimated by regression methods (like
least square method). As it is shown in Fig. 4, for
each high level of stress, there is a CDF diagram. The
data for two different life (e.g. B10 and B50 ) must be

1006

Figure 4.

Extrapolation of test diagrams to field diagram.

obtained separately by the regression method, and the

line passes these two points is field CDF diagram.
3.4

Figure 5.

Residual life methods.

Figure 6.

Correlation between testing and field time.

PF-related methods

3.4.1 Residual life method

If catastrophic failure of an ALT is inaccessible or it
takes long time to meet required failure, and there is a
possibility to define a suitable performance factor and
respected PT, residual life method could be employed
to estimate service life. For this purpose, new samples
must be aged in different time (or cycle) intervals. For
each interval, there must be some replicates to draw
their CDF diagram as shown in Fig. 5. As failure point,
minimum value of PF must be defined. One way of
estimating life (L) is to implicitly obtain it according
to eq. (2) by replacing (Age)1 = 0. The other way is to
assume PF as a function of age with some unknown
parameters which could be calculated by regression
analysis.
3.4.2 Degradation method
This is the most popular AST method in which the
specimens (material level) must be artificially aged to
record their PF values in different intervals. PF should
be also measured at the beginning, and the test must
be continued until specified time. If field degradation
data is available, they could be compared with test data
to obtain a correlation between test and field time as
shown in Fig. 6.
3.5

Comparison methods

If the aim of AT is to comparatively judge about different aspects of design and material to select the best
option, it is considered as comparison method. The

product may be also tested in different levels of stressors in order to decide the effect of each stressor on
life and to remove its effect. It is based on performing
same test for all available products in order to obtain
their life data. Such data could be drawn on CDF diagram in order to decide which one is better from the
view of particular life. The diagrams are not necessarily parallel, and may intersect each other. Accordingly,
for example, if a diagram has the highest B10 , there is
no guarantee to have the highest B50 .

1007

CONCLUSION

This paper presented a comprehensive approach for

testers and manufacturers to obtain the reliability
of their products by available tools like field failure data, probability distribution functions and finite
element method. Accelerated testing methods were
categorized in using rate methods, design modification
methods, extrapolation methods, performance factorrelated methods and comparison methods. The study
of multi-function products will be the continuation
of this research. Degradation of each function and
the interaction between different failure modes during accelerated testing will be considered in order to
achieve a general accelerated testing approach.
REFERENCES
Apostolopoulos, C. and M. Papadopoulos (2007). Tensile
and low cycle fatigue behavior of corroded reinforcing
steel bars s400. Construction and building materials 21,
855864.
B, M.C., J.P. Gerofi, L.L.Y. Visconte, and R.C.R. Nunes
(2007). Prediction of shelf life of natural rubber male
condoms-a necessity. Polymer testing 26, 306314.
Chou, P. and M. Lamers (2005). Quantitative study of
magnetic tape abrasivity using accelerated wear testing.
Microsystem technologies 11, 901906.
Drew, M., S. Humphries, K. Thorogood, and N. Barnett
(2006). Remaining life assessment of carbon steel boiler
headers by repeated creep testing. International journal
of pressure vessels and piping 83, 343348.
Escobar, L.A. and W.Q. Meeker (2006). A review of accelerated test models. Statistical science 21(4), 552577.
G. Ranganathan, T.H.S. Raj, and P.M. Ram (2004). Wear
characteristics of small pm rotors and oil pump bearings.
Tribology international 37, 19.
Knox, E. and M. Cowling (2000). A rapid durability test
method for adhesives. International journal of adhesion
and adhesives 20, 201208.
Koo, H.-J. and Y.-K. Kim (2005). Reliability assessment
of seat belt webbings through accelerated life testing.
Polymer Testing 24, 309315.

Marahleh, G., A. Kheder, and H. Hamad (2006). Creep life

prediction of service-exposed turbine blades. Materials
science and engineering 433, 305309.
Murphy, R. and V. Baxter (2004). Accelerated life test and
field test performance results for an intergral heat pump
water heater. Proceedings of IMECE, ASME International
Mechanical Engineering Congress and Expostion.
Nagel, T., J. Kramer, M. Presti, A. Schatz, R. Salzman, J. A.
Scaparo, and A. J. Montalbano (2004). A new approach
of accelerated life testing for metallic catalytic converters.
SAE transactions 113, 362375.
Nelson, W. (1990). Accelerated testing : statistical models,
test plans and data analyses. Wiley.
Park, S.-J., S.-D. Park, K.-S. Kim, and J.-H. Cho (2006).
Reliability evaluation for the pump assembly using an
accelerated test. International journal of pressure vessels
and piping 83, 283286.
ReliaSoft (2001). Acceleration life testing analysis reference.
Su, H., M. Ma, and D. Olson (2004). Accelerated tests of
wiper motor retainers using cae durability and reliability
techniques. SAE SP, 103110.
Wang, G., D. Taylor, B. Bouquin, J. Devlukia, and
A. Ciepalowicz (2000). Predition of fatigue failure in a
camshaft using the crack modeling method. Engineering
failure analysis (7), 189197.
Wu, C., N.H. Hwang, and Y. Lin (2002). Measurement
of dynamic stresses in a mechanical heart valve during
accelerated testing. Cardiovascular engineering 2(3).
Wu, J., R.R. Zhang, Q. Wu, and K.K. Stevens (2003).
Environmental vibration assessment and its application
in accelerated tests for medical devices. Journal of sound
and vibration 267, 371383.
XIE, D. (2000). A new experimental method to evaluate
creep fatigue life of flip-chip solder joints with underfill.
Microelectronics reliability 40, 11911198.
Yang, X. and X. Ding (2006). Prediction of outdoor weathering performance of polypropylene filaments by accelerated weathering tests. Geotextiles and geomembrances 24,
103109.
Zhang, C., M. Le, B. Seth, and S. Liang (2002). Bearing life prognosis under environmental effects based on
accelerated life testing. IMechE 216, 509516.

1008

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A decomposition method to analyze complex fault trees

S. Contini
European Commission, Joint Research Centre,
Institute for the Protection and Security of the Citizen, Ispra (VA), Italy

ABSTRACT: This paper describes a new method for decomposing a coherent fault tree (X) of n variables
into a set of simpler functions. Each simple function is analyzed independently to determine the minimal cut sets
(MCS) and the probabilistic parameters of interest, i.e. top event failure probability and importance measures of
basic events. These results are suitably re-combined to obtain the results at (X) level such that i) the MCS of
(X) is the union of the MCS of simple functions and ii) the probabilistic parameters at (X) level are obtained
by simply combining the results of the probabilistic analysis of all simple functions. It is shown that in applying
the cut-off techniques, for determining the most important MCS, to a decomposed fault tree the truncation error
can be correctly determined. An example is described to show the potential of the proposed method.

INTRODUCTION

The efficient analysis of large fault trees is a complex

problem. Since the early seventies, several methods
have been defined and implemented in computer codes
to determine the Minimal Cut Sets (MCS), which are
needed to perform the probabilistic analysis. Due to
the typical huge number of MCS, the analysis has
to necessarily be limited to the determination of the
most important ones. This is conducted by applying
(logical-probabilistic) cut-off techniques which, however, imply the estimation of the truncation error, i.e.
of the cumulative probability of neglected MCS.
The calculation capability of Fault Tree Analysis
tools has greatly improved since the introduction of the
Binary Decision Diagrams (BDD) approach, which is
based on a compact graph representation of a Boolean
function (Bryant, 1996). Compared to the previous
approach based on MCS the main advantage of the
BDD approach relies on the possibility to exactly perform the probabilistic quantification without the need
of explicitly determining the MCS. Many complex
fault trees, which could be analysed with the MCS
approach only using the cut off technique (with the
consequent problem of the truncation error estimate),
could be successfully analysed with the BDD approach
without approximations.
The main problem with the BDD is, however, the
exponential increase in the number of nodes of the
graph embedding the system failure modes, a problem which is an important subject of research. Many
techniques have been proposed to solve this problem
as e.g. the fault tree re-writing to make it more suitable
to BDD representation, and the identification of more

effective variables ordering (see e.g. Bouissou & al.

(1997); Nusbaumer (2007)).
Independently from the method adopted for the
analysis, it is also interesting to explore the potential of
performing a functional decomposition of the complex
fault tree. The functional decomposition aims at representing a function (X) of n variables by means of two
or more (simple) functions containing a lower number
of variables. In fault tree analysis the modularisation
is a classical example of functional decomposition.
The fault tree decomposition is the subject of this
paper. A new method is described to decompose a
fault tree (X) into a set of simple fault trees i (Y)
and a set of conjunctions BCi (S) with X = Y S and
Y S = , such that: i) the MCS of (X) is the union
of the MCS of BCi (S) i (Y) and ii) the probabilistic
parameters at (X) level can be simply obtained from
the probabilistic analysis of all i (Y) and BCi (S).
The decomposition of a modularised tree of a simplified chemical reactor is shown as an example of
application to show the potential of the proposed
method.
2

BASIC CONCEPTS AND DEFINITIONS

Let (X) be the binary logical function of n variables

to be decomposed. Let S be a subset of X containing
m binary variables (m << n). The m-elements of S
can be represented in a truth-table, called BC-table,
where BC stands for Boundary Conditions. Each row
of the table represents a particular set of values (0, 1)
assumed by the m-variables in S. BCi (S) identifies the
i-th row of the table; BCi,j identifies the value assumed

1009

by the j-th element of S in the i-th row. The order of

BCi (S), represented as wi , is defined as the number of
BCi,j = 1.
Rows of the BC-table are rearranged into m + 1
groups according to their order. A group of order w,
where 0 w m, contains a number of rows equal to:


m!
.
w!(m w)!
For instance, if S = (a, b, c), the BC-table is as
follows:
Each row of the table contains the values assumed
by the variables in S. For instance, the 3rd row is
represented as BC3 (a, b, c) = (0, 1, 0).
We say that BCk (S) subsumes BCh (S), which is
represented as BCk > BCh , iff BCk,j BCh,j for
any j(j = 1, . . . , m). For instance, in Table 1, BC5
subsumes BC2 and BC3 ; BC7 subsumes BC3 and BC4 .
Associated with each BCi (S) is the simple function
resulting from the application of BCi (S) to (X), i.e.
i (Y) = (X)|BCi(S) .
With reference to the generic gate of a fault tree,
represented as G = E H where E is the basic event
to which the BC is applied, H is a generic logical function and the logical operator AND/OR, the rules for
determining i (Y) = (X)|BCi(S) are given in Table 2.
For instance, the fault tree in Figure 1a contains the
following BC set:

Figure 1a.

Function Top.

Figure 1b.

Function Top |(BC(B4) = 0, BC(B3) .

Therefore, the function (X) can be written as:

m

BC(B4) = 0(status = Good);

(X) =

BC(B3) = 1(status = Failed).

2


BCi (S) i (Y)

(1)

i=1

With the application of the rules given in Table 1,

the fault tree in Figure 1a, is reduced to the simple
fault tree shown in Figure 1b.

where:
X = (x1 , x2 , . . ., xn ); S = (x1 , x2 , . . ., xm );
Y = (xm+1 , xm+2 , . . ., xn ); X = S Y; S Y = .

Table 1.

1
2
3
4
5
6
7
8

Table 2.

Example of BC-table for S = (a, b, c).

A

i (Y) = (X)|BCi(S)

0
1
0
0
1
1
0
1

0
0
1
0
1
0
1
1

0
0
0
1
0
1
1
1

1
2
3
4
5
6
7
8

Since BCi (S)(i = 1, . . ., 2m ) are pair-wise mutually

exclusive:
m

Pr((X)) =

2


Pr(BCi (S)) Pr(i (Y))

(2)

i=1

Depending on the way the BC-table is constructed,

two types of decomposition procedures could be
conceived, addressing:
1. the exact probabilistic quantification (Top occurrence probability and importance analysis);
2. the determination of Minimal Cut Sets.

Rules for determining i (Y) = (X)|BCi(S) .

Boundary condition

Operator AND

Operator OR

BC(E) = 1 or Failed
BC(E) = 0 or Good

G=H
G=0

G=1
G=H

In the first case, being m the cardinality of S,

it is necessary to generate and quantify 2m simple
functions. Each simple function has to be exactly
quantified by means of e.g the BDD technique.

1010

In the second case the number of simple functions

is smaller, but the probabilistic quantification can give
only upper and lower bounds, i.e.

(MCS of (X)) =
(MCS of i (Y))

More generally, let S = (x1 , x2 , . . . , xm ) be an MPS

of (X). The BC-table can be constructed observing,
first of all, that:
(X)|BC0 (S) = 0

that is to say:

and
Pr LB ((X)) Pr((X)) Pr UB ((X))

(3)

This paper deals with the latter case applied to a

coherent fault tree.
In practice, in the fault tree analysis procedure, the
proposed decomposition technique should be applied
to those modules that cannot be analyzed with the
available resources.
From what has been shown, it appears that the
important problem to be solved is the choice of the
set S. Indeed, the number of simple functions and their
dimensions depends on the basic events of (X) that
make up the set S.
3

THE DECOMPOSITION METHOD

It is shown, in this section, the advantage of selecting

S as a Minimum Path Set (MPS) of the fault tree.
In system reliability a MPS is defined as the minimum set of components such that if all components
are working (state 0) the Top event is 0, i.e. all MCS
are 0. Hence, each MCS contains at least an element
of the MPS. Therefore, it is possible to partition the
set of MCS according to the elements of the MPS.
To explain how the proposed method works, consider for instance the following function, simply
expressed as the disjunction of its MCS:
 = a c + b c + a de + be.
It can be easily recognised that (a b) is one of the
path sets, since variables (a, b) are contained in all failure combinations of . By setting a = 1 and b = 0 in
 the function 1 = c+d e is obtained. Setting a = 0
and b = 1 in  the function 2 = c + e is obtained.
Therefore,  can be expressed as the disjunction of
two terms:  = a 1 + b 2 .
Suppose now that two failure combinations (a b d)
and (a b c), containing both a and b, are added to , i.e.
 = a c + b c + a d e + b e + a b d + a b c.
To extract from  a function containing these two
extra combinations it is necessary to set a = 1 and
b = 1 in . The new function is 3 = c + d + e. Thus,
 = a 1 + b 2 + a b 3 .
Note that 3 contains the cut set c, which is contained into 1 and 2 , and the cut set e that is also in
2 . Therefore 3 must be minimized with respect to
the subsumed simple functions 1 and 2 .

(x1 = 0, . . ., xm = 0, xm+1 , . . ., xn ) = 0
Moreover:
The group with w = 1 is composed of m rows;
the MCS of the corresponding simple functions
i (Y)(i = 1, . . . , m) are also MCS of (X).
If i (Y) = (X)|BCi(S) = 1 then the variables of
BCi set to 1 constitute a MCS of (X) of order w.
Consequently, for all BCk > BCi it is useless to construct the corresponding simple functions, because
they would contain only non-minimal cut sets;
If the generic BCi is of order w (2 w m)
and its w variables at 1 do not combine each other
in (X) (i.e. they do not belong to any MCS of
), then it is useless to construct the corresponding
simple function, because it would contain only nonminimal cut sets;
If the generic BCi is of order w (2 w m)
and its w variables at 1 combine each other in
(X) then the corresponding simple function has
to be generated. Such function may or may not
contain MCS of (X). It is easy to recognise that
the MCS of i (Y) can be obtained by minimising
i (Y) only with respect to those j (Y) for which
BCi > BCj . The following pseudo code describes
the main operations for constructing i (Y).
If BCi > BCj then
if i j then i = 0
else
i = i |Ej=0
j = j |Cj=0

i = i j
endif
endif
If i (Y) j (Y) then all cut sets of i (Y) are nonminimal cut sets of (X) as BCi subsumes BCj . The
check whether i (Y) j (Y) can be performed either
on the fault tree structure of i (Y) and j (Y), or on
their BDD. In the first case the check can be limited to
gates that are modified by the use of the BC set, which
are those gates on the path from events in S to Top.
When i (Y)  j (Y), the MCS of i (Y) are those
that do not verify j (Y). The complexity of this new
function can be reduced by removing the sets Ej and Cj .
Ej is the set of MCS of order 1 of j (Y) that can be

1011

set to 0 in i (Y) (they are contained into non-minimal

cuts of i (Y)). Cj is the set of variables of j (Y) that
are not contained into i (Y), i.e. variables that are not
useful for removing non minimal cut sets from i (Y).
The analysis of the resulting function i (Y) can give
either no MCS or MCS that are also MCS of (X).
From these considerations it follows that the best set
S (of order m) is the one in which none of its events
combine each other, which implies that (X) can be
decomposed into m simple functions containing MCS
of (X).
For each i (Y) the probabilistic analysis and the
importance measures of basic events is performed
using a method based either on MCS or on BDD.

RECOMBINATION OF RESULTS

After determining the unavailability and importance

measures of basic events for all simple functions
i (Y), the probabilistic parameters at (X) level can
easily be calculated as described in this section.
In order to simplify the notation BCi (S), (X) and
i (Y) will be simply represented as BCi ,  and i
when this does not generate confusion.

4.2

Importance measures of basic events

The simple functions are dependent, since they are

derived from a common tree. The equations hereinafter described are applied to separately determine
the following importance measures for events in Y and
in S: 1) Birnbaum, 2) Criticality, 3) Risk Achievement
Worth, 4) Risk Reduction Worth.
Other measures can be derived, e.g. Fussell-Vesely,
which are not, however, considered in this paper.
4.2.1 Birnbaum importance index
Let (X) be a coherent fault tree. The Birnbaum
importance index IjB is the probability that a generic
component xj is critical, i.e. the system fails if the
component fails (xj = 1 implies (1i , X) = 1);
the system works if the component works (xj = 0
implies (0i , X) = 0). The Birnbaum importance of
the generic event in (X), say xj is given by:
IjB =

Pr((X))
Pr(xj )

which can also be written as:

IjB = Pr((1j , X)) Pr((0j , X))

4.1

Upper and lower bounds of (X)

The Upper Bound Pr UB () is given by:

Pr UB () =

Ns


Pr(BCi ) Pr(i )

(4)

Let xj be a variable that belongs to simple functions

i (Y).
IjiB = Pr(i (1j , Y)) Pr(i (0j , Y)) is the importance index of the j-th event in i (Y).

i=1

Pr(BCi ) =

m


Pr(xj )BCi,j

(5)

IjB

= Pr


BCi i (1j , Y) Pr

 Ns


i=1

j=1

where:

Ns



BCi i (0j , Y)

i=1

Pr(BCi ) Pr(i (1j , Y))

i=1

Ns is the number of simple functions;

m is the number of elements of S;
BCi,j is the value (0, 1) assumed by the j-th variable
xj of S in the i-th BC set;
Pr(i ) is the top-event of the i-th simple function;
Pr(xj) is the probability of the j-th variable of S.

m


Ns


Pr(BCi ) Pr(i (0j , Y))

i=1

Ns


Pr BCi )(Pr(i (1j , Y))

i=1

The Lower Bound, Pr LB () is given by equation (4)

in which:
Pr(BCi ) =

 Ns


Pr(i (0j , Y))) =

Ns


Pr(BCi )IjiB

i=1

Pr(xj )BCi,j (1 Pr(xj ))1BCi,j

(6)

Therefore:

j=1

Again, Pr(i ) is calculated on the BDD using

equation (3)

IjB

1012

Ns

i=1

Pr(BCi )IjiB

(7a)

Consider now the case in which xj belongs to S, i.e.

to one or more BCi . Since:

4.2.3 Risk Achievement Worth

The Risk Achievement Worth RAWj of xj is
defined as:

BCi = xik

RAWj =

k=1

it is easy to recognize that:

IjB

Ns

i=1

Pr(BCi )
Pr(i )
Pr(xj )

(7b)

where i = 1 if xj BCi , otherwise i = 0.

4.2.2 Criticality importance index
The Criticality index is defined as:
IjC =

Pr((1j , X))
Pr((X))

and it is a measure of the change of the system failure

probability when xj is supposed to be failed.
The RAW can be easily derived considering again
two possible cases.
If xj belongs to one or more simple functions i (Y),
then:
Ns

Pr(BCi ) Pr(i (1j , Y))
i=1
RAWj
Pr UB ((X))

Pr(xj ) B
I
Pr((X)) j

Multiplying by

Pr(i (Y))
Pr(i (Y))

we obtain:


1
Pr(BCi )RAWji Pr(i )
Pr UB ()
Ns

It represents the probability that the event xj

occurred at time t and it is at the same time critical
to the occurrence of system failure, given that the system has failed. In other words, given the system failure,
the Criticality index is the probability that xi occurred
last.
If xj belongs to one or more functions i , then the
following equation can be derived for determining the
Criticality importance. Given:
1
IjC =
IB Pr(xj )
Pr UB () j

and

RAWj

If xj belongs to one or more BC sets, then:

It is possible to write:

Where i =

Ns

i=1

Pr(i )
Pr(i )

the following equation is


1
IjiC Pr(i )
Pr UB ()
Ns

IjC

(8a)

Pr((X))
Pr((0j , X))

If xj belongs to one or more BC sets, then:

Ns

(8b)

RRWj

i=1

where has the meaning defined in eq. (7b).

if xj BCi , otherwise i = 1

RRWj is a measure of the change of the system

failure probability when xj is supposed to be perfectly
working.
If xj belongs to one or more simple functions i (Y),
then:
Pr UB ()
RRWj
(10a)
Ns

1
RRWji Pr(BCi ) Pr(i )

If xj is an element of S, then:

1
i Pr(BCi ) Pr(i )
Pr UB ()

1
Pr(xj )

i=1

i=1

IjC

(9b)

i=1

RRWj =


1
IjiB Pr(BCi ) Pr(xj )
Pr UB ()

Multiplying by
derived:


1
i Pr(BCi ) Pr(i )
Pr UB ()
Ns

RAWj

4.2.4 Risk Reduction Worth

The equations for the determination of the Risk Reduction Worth can be derived in a similar way. The Risk
Reduction Worth RRWj of xj is defined as:

1 B
I Pr(xj ),
IjiC =
Pr(i ) ji

IjC

(9a)

i=1

Pr(UB (X))
Ns

(1 i ) Pr(BCi ) Pr(i (Y))

i=1

where has the meaning defined in eq. (7b).

1013

(10b)

DETERMINATION OF SIGNIFICANT MCS

It has been shown in section 2 that:


(MCS of (X)) =
(MCS of i (Y))

(11)

Since the number of MCS can be very high, cut-off

thresholdsbased respectively on: a) the maximum
order of the minimal cut-sets nmax and b) the minimum
probability of the minimal cut-sets Pmin are applied
to determine the most important failure combinations.
Consequently, it becomes important to estimate the
truncation error Pe , i.e. the degree of underestimation
of the Top event probability due to the neglected MCS.
The estimation of Pe is indeed an important problem
(see e.g. Jung & al. 2005; Cepin 2005; Duflot &
al 2007). The decomposition procedure allows solving this problem, since each simple function can be
analysed without applying the cut-off technique.
To extract the significant MCS from the generic
i (Y), the following cut-off values are used:
nmax (i) > (nmax wi )

and/or

Pmin (i) < Pmin / Pr(BCi )

Since Pr(i (Y)) is known, the estimation of Pe
becomes trivial.
Generally the top event probabilities of simple functions decrease as the order w of the corresponding BC
increases. Hence, by calculating for i (Y) the minimum order of the MCS, say zi , and the maximum
probability of the MCS, say Pi , it is possible to cut the
whole functions if:
zi > nmax (i)
6

Figure 2.
Table 3.

or Pi < Pmin (i).

EXAMPLE

The decomposition procedure is described in this

section with reference to a simple example. Consider
the modularised fault tree represented in Figure 2.
Modules M1, M2, M7, and M8 are characterised by
their probability and number of MCS. In this example
it is assumed that all events and modules have equal
probability q = 0.1.
Each basic event is associated with the repetition
factor R, i.e the number of times the event is replicated
in the tree. For an OR gate the value R is obtained by
summing up the R of its descendants; for an AND gates
the minimum value of R is considered. In Figure 2
the R values are the numbers indicated by arrows. The
MPS is obtained by visiting the tree from Top to bottom
according to the following rules:
AND gates: select the descendant with minimum
R; if it is a basic event, add it to S.
OR
gates: select all descendants.

1
2
3
4
5
6

Modularized fault tree to be decomposed.

BC-table for the example in Figure 2.
E15

M8

M7

E20

0
0
0
1
1
0

0
0
1
0
0
1

0
1
0
0
0
1

1
0
0
0
1
0

In Figure 2 gates indicated with ( ) are those that

do not change with the application of the boundary
conditions.
The application of this algorithm gives the following MPS: S = (E15M8M7E20).
6.1 Construction of the BC-table
In order to construct the BC-table it is necessary to
determine the combinations of the elements of S.
An algorithm has been developed to this purpose.
This algorithm, applied to the considered example
gives the elements of S that combine each other; they
are: (E20 E15) and (M7 M8).

1014

The 6 BC sets to be considered for determining the

simple functions are given in Table 3.
6.2

Function 5. BC5 : E15 = E20 = 1; M7 = M8 = 0,

hence G_TOP5 = Top|BC5 :

Determination of simple functions

The original modularised tree is decomposed into 6

simple functions, one for each BC set.
Function 1. BC1 : E20 = 1; M7 = M8 = E15 = 0,
hence G_TOP1 = Top|BC1 :

Function 2. BC2 : M7 = 1; M8 = E20 = E15 = 0,

hence G_TOP2 = Top|BC2 :

Function 5. needs to be minimized, with respect

to G_TOP1 and G_TOP4, since BC5 > BC1 and
BC5 > BC4. The check of the gates definition
leads to the following result: G_TOP5  G_TOP4
and G_TOP5  G_TOP1. Consequently, G_TOP5 is
given by:
TOP5 = G_TOP5 G_TOP1 G_TOP4.

T0P5
Function 3. BC3 : M8 = 1; M7 = E20 = E15 = 0,
hence G_TOP3 = Top|BC3 :

Function 4. BC4 : E15 = 1; M7 = M8 = E20 = 0,

hence G_TOP4 = Top|BC4 :

The analysis of this function gives TOP5 = 0, i.e.

TOP5 does not contain any MCS of TOP.
Function 6. BC6 : M8 = M7 = 1; E15 = E20 = 0,
hence G_TOP6 = Top|BC6 = 1, which means that M7
M8 is a MCS of TOP.

1015

Table 4.

Unvailability upper bound of simple functions.

Pr(BC)

Pr(i)( )

1
2
3
4
5
6

0.1
0.1
0.1
0.1
0.01
0.01

2.782 102
0.1981
0.1089
3.439 102
0.0
1

Table 6.

Pr(BC) Pr(i)

Function

MCS

2.782 103
1.981 102
1.089 102
3.439 103
0.0
1.0 102

E20 E2 E1
E20 E2 M1
E20 E2 M3
E20 E2 E9 M2
M7 M3
M7 M1
M7 E9 M2
M8 M3
M8 E9 M2
E15 E1 M2
E15 M1 M2
E15 M2 M3
E15 E9 M2
M7 M8

( ) Pr(i) is calculated on the BDD using equation (3).

3
4

Table 5.

Unavailability lower bound of simple functions.

6

Pr(BC)

Pr(i)( )

Pr(BC) Pr(i)

1
2
3
4
5
6

7.29 102
7.29 102
7.29 102
7.29 102
8.10 103
8.10 103

2.782 102
0.1981
0.109
3.439 102
0.0
1

2.028 103
1.444 102
7.946 103
2.507 103
0.0
8.10 103

6.3

MCS of the fault tree of Figure 2.

Probabilistic analysis of simple functions

6.3.1 Unavailability Upper Bound

The following results of the analysis are obtained using
the UB equations (1) and (2).
The Total unavailability upper bound is the sum of
the values in the above table: Pr UB = 4.692 102
6.3.2 Unavailability Lower Bound
The following results are obtained using the Lower
Bound equations (4) and (5).
The Total unavailability lower bound is the sum of
the values in the above table: Pr LB = 3.502 102
The analysis of the original fault tree gives the following exact result for the unavailability value: Pr EX =
4.307 102 , which is in between the determined
upper and lower bounds.
The importance analysis has been performed for
all events. Due to space limitations results are not
reported. The values obtained are in good agreement
with the exact values determined on the fault tree in
Figure 2.

6.4 Determination of MCS

Table 6 contains the MCS of the original tree shown
in Figure 2 given by the union of the MCS determined
for the five functions above analysed.

CONCLUSIONS AND FURTHER

DEVELOPMENTS

This paper presented a method to decompose a coherent fault tree (X)based on the events of a path
set Sinto simple fault trees i (Y). The probabilistic parameters at (X) level have been obtained by
recomposing the results of the probabilistic analysis
of all i (Y) and BCi (S). The MCS of (X) are the
union of the MCS of i (Y). The number of simple
fault trees and their dimensions (number of events and
gates) depends on the number m of events making up
the path set S and on the structure of the fault tree to be
decomposed. Generally, the greater is m, the smaller
are the i (Y).
A decomposition method is useful for the analysis
of complex fault trees when the available FTA tool
is not able to perform the analysis without applying
the (logical and/or probabilistic) cut-off techniques.
In these cases the decomposition allows the analyst to
overcome the difficulties of the problem of estimating
the truncation error.
The fault tree to be decomposed has been supposed
to be already modularised. Indeed, the modularisation
is in itself a decomposition method. Some interesting
considerations could be drawn considering the different type of modules. Type I module contains only non
repeated basic events; type II contains type I modules as well as all occurrences of one or more repeated
basic events; finally, type III contains the previous two
types of modules and all occurrences of one or more
repeated sub-trees. It would be interesting to study the
integration of the third modularisation type with the
decomposition method.
It should be stressed that the proposed method
is sensitive to the algorithms for determining the
BC-table. Experiments will be undertaken to compare
the different alternatives on real fault trees.

1016

At the time of writing the software implementation has not yet been completed. The available partial
experimental results are promising. It is important to
complete the software and to perform the experimental verification, in order to verify the capability of the
method to decompose very large fault trees.
If the experimental results will be positive, then
the method can easily be extended to deal also with
non-coherent fault trees.
ACKNOWLEDGEMENTS
This work was carried out by the Joint Research Centre within the action: Assessment Methodologies for
Nuclear Security of the European Commission Seventh multi-annual Framework Programme, as direct
research, in the area of nuclear security.
Thanks are due to G.G.M. Cojazzi, G. de Cola, and
L. Fabbri for the support given.
REFERENCES
Bouissou M, Bruyere F, Rauzy A., (1997) BDD based fault
tree processing: a comparison of variable ordering heuristics, C. Guedes Soares, ed., Proceedings of ESREL 97
Conference, Vol. 3, Pergamon Press.

Bryant R.E, 1986, Graph Based Algorithms for Boolean

Functions Manipulation, IEEE Transactions on Computers, Vol. C35.
Cepin M. (2005) Analysis of truncation limit in PSA,
Reliability Engineering and System Safety, vol. 87.
Contini S., Fault tree decomposition methods, JRC Internal
report, 2008.
Duflot N., Berenguer C., Dieulle L., Vasseur D. (2006) How
to build an adequate set of MCS in PSA importance measures calculation. Proc. of 8th International Conference
on Probabilistic Safety Assessment and ManagementPSAM8, New Orleans, Louisiana, USA.
Kumamoto H., Henley E.J., 1996, Probabilistic Risk Assessment and Management for Engineers and Scientists, IEEE
Press, New York.
Jung W.S., Yang J.E., Ha J. (2005) Development of measures to estimate truncation error in fault tree analysis,
Reliability Engineering and System Safety, vol. 90.
Nusbaumer O. Analytical solutions of linked fault tree
probabilistic risk assessments using binary decision diagrams with emphasis on nuclear safety applications. ETH
Zrich, Nr. 17286, 2007.
Rauzy A. (1996) A brief introduction to Binary Decision
Diagrams, European Journal of Automation, Vol. 30.
Rauzy, A. (1993). New Algorithms for Fault Trees Analysis,
Reliability Engineering and System Safety, Vol 40.
Van der Borst M., Schoonakker, H. (2001). An overview
of PSA importance measures, Reliability Enineering and
System Safety, Vol. 72.

1017

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A quantitative methodology for risk assessment of explosive atmospheres

according to the ATEX directive
R. Lisi & M.F. Milazzo
University of Messina-Dipartimento di Chimica Industriale e Ingegneria dei Materiali, Messina, Italy

G. Maschio
University of Padova-Dipartimento di Principi e Impianti di Ingegneria Chimica, Padova, Italy

ABSTRACT: The Directive 99/92/EC (ATEX 137) deals with the safety and health protection of workers
potentially exposed at explosive atmospheres. The application of the ATEX Directive requires the assessment of
specific risks due to the presence of potentially explosive atmospheres. The potential development of explosive
atmospheres is generally typical of industries classified at major hazard, but it is also possible in other industries
where flammable materials are handling. Since the assess of the risk due to the formation of explosive atmospheres
is required in both the cases and it is fundamental for the safety in the workplaces, a quantitative approach has been
proposed. The objective of the paper is to describe the main aspects of a quantitative methodology based on a
probabilistic risk assessment starting from a detailed knowledge of the analyzed system.

GENERAL OVERVIEW OF EU DIRECTIVES

In the framework of the General Directive 89/391/CE,

concerning the application of measures to promote the
improvement of the safety and health of the workers, the places potential characterized by explosion
hazards have assumed particular attention. In this
context the safety associated with the presence of
explosive atmospheres requires a new legislation.
Explosive atmospheres can be originated by releases
of flammable gases, mists or vapours or by combustible dusts. If there is enough substance, mixed with
air, then a source of ignition is sufficient to cause an
explosion.
Explosions can cause loss of life and serious injuries
as well as significant damage. Preventing releases
of dangerous substances, which can create explosive
atmospheres, and preventing sources of ignition are
two widely used ways of reducing the risk. The use
of the correct equipment can help greatly in achieving
these aims.
To increase the safety of the workers and prevent the
explosions, during 2003, the European Committee has
realized and, on purpose, compiled an official guide
defined of good practice for the correct application
of the new directives related to the safety of the workers
regarding the risk due to the presence of explosive
atmospheres.
The term ATEX derives from the French language,
means Atmospheres Explosibles and is the name

commonly given to the framework for controlling

explosive atmospheres and the standards of equipment
and protective systems used in them. It is based on the
requirements of two European Directives:
1. Directive 99/92/EC (also known as ATEX 137
or the ATEX Workplace Directive) on minimum
requirements for improving the health and safety
protection of workers potentially exposed at risk
from explosive atmospheres.
2. Directive 94/9/EC (also known as ATEX 95 or the
ATEX Equipment Directive) on the approximation
of the laws of Members States concerning equipment and protective systems intended for use in
potentially explosive atmospheres.
The Directive 99/92/EC has been adopted in Italy.
It has been modified and integrated in the main Italian
regulations concerning the safety and health protection
in the workplace (D.L. 626/94, D.P.R. 126/98 and D.L.
233/03).

1.1 The Italian legislation concerning safety

and health protection in the workplace
The D.L. 626 is the Italian regulation concerning safety
and health protection in the workplace, the part VIII of
the legislation includes the norms of protection against
the explosions.

1019

The normative states that the Employer has the

obligation to prevent the formation of explosive atmospheres adopting all the technical-organizational measures required, this can be done on the base of an
evaluation of the presence of explosive atmospheres
in the workplace. If the formation of such flammable
clouds can not be avoided, the ignition must be prevented and the damages potentially caused by an
explosion must be reduced to the least.
Furthermore, on the base of an appropriate criterion, the Employer has to classify the areas in which
is possible the formation of explosive atmospheres
and, then, the Document of evaluation of the risk
due to explosive atmospheres (in this work named
Document of evaluation of the ATEX risk) has to be
redacted and, periodically, updated. Such document
must undergo to the least requisites fixed by the decree
law mentioned before.
The classification of hazardous areas must be done
on the base of the use of harmonized norms. The
document must include a section in which the risk is
evaluated using specific algorithms; measure to avoid
the formation of explosive atmospheres and ignition
must also be indicated; finally the equipments used in
the workplace must be specifies.

Description of the
workplace and the
activity

Identification of the
substances and of the
emission source

Definition of the
cloud formation

Calculation of the
release probability

Classification of Hazardous
Areas

Identification of the
presence of ignition sources
and estimation of ignition
probabilities

Estimation of the effects of

the explosion (impact zone)

Calculation of probability
of presence of workers

METHODOLOGICAL APPROACH

Safety Reports describe the risk assessment related to

the explosions of great magnitude for industries classified at major hazard. The estimation of the risk due
to explosions, characterized by minor magnitude, but,
potentially, involving workers, is described in the document of the risk assessment in the workplace and in
the document of evaluation of the ATEX risk.
The approach applied for the risk assessment in the
workplaces is generally qualitative (Benintendi et al.
2006) or semi-quantitative (Pezzo et al. 2006). Unfortunately, the application of a qualitative method often
causes an underestimation of the risk associated to
the explosion of clouds originated by small releases,
particularly, in confined places.
The use of a quantitative approach for risk analysis allows an improvement of the overall safety levels.
At the same time, the quantitative method for industries classified at minor hazard permits to avoid the
underestimation of the risks associated with explosive
atmospheres and identify the correct preventive and
protection measures for each case.
In this paper a quantitative method has been proposed, it is based on a probabilistic risk assessment.
The analysis starts with detailed knowledge of the
analyzed system.
In Figure 1 a flow-chart of the overall procedure has
been drawn.

Measures of
prevention and
protection

ATEX Risk Calculation

No
Risk Acceptability
Yes
Document of evaluation of
the ATEX risk

Figure 1. Flow-chart for the evaluation of risk due to the

presence of explosive atmospheres.

The first step of the procedure is the classification of hazardous areas which takes into account
the characterization of all the emission sources of
flammable substances. In order to evaluate the risk,
the calculation of the effects of the explosion, the
assessment of probability of presence of ignition
sources and the number of exposed workers are also
necessary.
The flow-chart for the evaluation of risk due to the
presence of explosive atmospheres, shown in Figure 1,
can be divided in two part: the census phase and the
risk evaluation phase.

1020

The fundamental steps are:

The study of industrial process;
The identification of the explosion risks: it is necessary to know the level of flammability and the
chemical-physical characteristics of the substances
(gas, liquid, dust);
The evaluation of the risk due to the presence of
explosive atmospheres:
The probability of formation an explosive atmosphere and the quantity of the released substance
must be determinate, it is necessary the classification of the areas characterized by explosion
risks;
The ignition sources have to be identified;
The effects of the potential explosion must be
quantified;
In order to calculate the number of exposed workers the probability of presence of people in the
workplaces has to be determinate;
Finally the risk index can be obtained according
to the consideration of the next paragraph.
The preliminary phase of the risk evaluation consists of verifying if it is possible to previously
exclude the presence of explosive atmospheres in some
unit of the establishment. In such case, in fact, the
place is not classified at explosion hazard and the analysis is not required. In the other cases it is necessary
to proceed to the risk assessment.
The procedure is completed when the evaluation
of the acceptability of the risk has been made. The
analysis gives an risk index which, eventually, permits
to define the necessary changes to reduce the risk level.
The risk has to be done analyzing each process,
that takes place in the establishment relatively to the
following phases:
under the normal conditions of operation;
during the starting and out of service phases;
during malfunction of the plant.

RISK INDEX

According the flow-chart described above, it is possible to calculate the risk due to the presence of explosive
atmospheres. The risk can be expressed through the
risk index.
The risk index, Rae is given by equation (1):

RISK ANALYSIS PHASES

The estimation of the risk index comprises the following phases.

4.1 Explosive atmosphere formation
and classification of the areas
An explosive atmosphere is defined as a mixture of
flammable substances with air, under atmospheric
conditions, in the form of gases, vapours, mist or
dust in which, after ignition has occurred, combustion
spreads to the entire unburned mixture. Atmospheric
conditions are commonly referred to ambient temperatures and pressures. This means temperatures of
20 C to 40 C and pressures of 0.8 to 1.1 bar.
Many workplaces may contain, or have activities
that produce, explosive or potentially explosive atmospheres. Examples include places where work activities create or release flammable gases or vapours, such
as vehicle paint spraying, or in workplaces handling
fine organic dusts such as grain flour or wood.
The flammable substances can be grouped into four
categories:
a. Liquids
Liquids (such as petrol and other fuels) and solvents
in industrial products (such as paint, ink, adhesives and
cleaning fluids) give off flammable vapor. Flammable
liquids are classified as:
Extremely flammable: Liquids which have a flash
point lower than 0 C and a boiling point (or, in the
case of a boiling range, the initial boiling point)
lower than or equal to 35 C.
Highly flammable: Liquids which have a flash
point below 21 C, but which are not extremely
flammable.
Flammable
b. Dusts
Dusts can be produced from many materials such as
coal, wood, grain, sugar, certain metals and synthetic
organic chemicals. A cloud of combustible dust in the
air can explode violently if ignited.

(1)

c. Gases
This class comprises gases such as liquefied
petroleum gas or methane. These are usually stored
under pressure in cylinders and bulk containers.
Uncontrolled releases can readily ignite or cause the
cylinder to become a missile.

where: pe = probability of release of flammable substance from an emission source; pa = probability of

presence of an ignition source; D = number of people
in impact area.

d. Solids
Solids include materials such as plastic foam, packaging, and textiles which can burn fiercely and give
off dense black smoke, sometimes poisonous.

Rae = pe pa D

1021

The classification of the areas has the purpose

to establish the presence of zones characterized by
explosion hazard, in which technical and organizational provisions must be adopts with the aim to make
negligible the risk due to the presence of explosive
atmospheres.
In order to classify the areas, the establishment
must be divided in unit and it is necessary to
define the zones where flammable substances can
be release due to the normal operation of the plant,
deviations of process or during the maintenance
activities.
According to the standards EN 60079-10 (gas/
vapour) and EN 50281-3 (dust), process plants can
be divided into zones. These zones are defined taking
into account the likelihood of a potentially explosive atmosphere being present (European and IEC
method).
4.2

Estimation of the probability of presence

of explosive atmospheres

4.3 Estimation of the probability of the presence

of effective ignition sources
The third phase of the risk evaluation is the identification of the ignition sources that are able to cause the
explosion. The presence of an explosive atmosphere
is not enough to burn an ignition is necessary.
According to the standards UNI EN 1127-1, there
are many causes of ignition of flammable atmospheres:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.

The definition of the areas at risk of explosion at

normal operating conditions is based on the area
classification according to CEI-31-30 and relative
Guidelines.
The directive CEI-31-30 identifies the following
hazardous zones:
1. Zone 0: area with a continuously present explosive
atmosphere due to the presence of gas.
2. Zone 1: area with a likely to occur explosive atmosphere in normal operating conditions, due to the
presence of gas.
3. Zone 2: area where explosive atmosphere due to
the presence of gas is unlikely to occur in normal
operating conditions or occurs infrequently of for
short periods of time.
The CEI 3135 guidelines define the explosive
atmosphere probabilities and durations shown in
Table 1.
The probability of the presence of explosive atmosphere can directly be derived from the Classification
of Hazardous Areas as described in the directive CEI
EN 60079-10.

Table 1.

Explosive atmosphere probabilities and durations.

Zone

Explosive Atmosphere
Probability P
(year1 )

Overall Explosive Atmosphere Duration D

(hours/year)

Zone 0
Zone 1
Zone 2

P > 0.1
0.1 P > 1 E-03
1 E-03 P > 1 E-05

D > 1000 hrs

10 hrs < D 1000 hrs
0.1 hrs < D 10 hrs

12.
13.
14.
15.

Hot surfaces;
Flames and hot gases or particles;
Mechanical sparks;
Electrical network;
Wandering electrical currents;
Cathode protection;
Static electricity;
Lightning;
Heating cables;
Waves with radio-frequency (RF) from 104 Hz to
3 1012 Hz;
Electromagnetic waves with frequency from
3 1011 Hz to 3 1015 Hz;
Ionizing radiations;
Ultrasounds;
Adiabatic compressions and bump waves,
Exothermic reactions (also the spontaneous ignition of dusts).

In order to quantify the probability of effectiveness for each type of ignition source it is possible
to use literature data or, preferably, specific studies
for the plant under analysis. Assessment methods,
such as historical analysis, fault tree analysis, FMEA
or FMECA, or specific analytic procedures can be
used.
4.4

Explosion consequences

The consequences must be estimated for each emission

source identified through the classification of the areas
and for each unit of the establishment.
This phase consists in the calculation of the effects
of the explosion, which means the estimation of the
overpressure vs. the distance from source.
An explosion is a release of energy during a sufficiently small time, following the release of energy, a
pressure wave (perturbation) starts to propagate in the
space.
The complexity of the phenomenon would require
a fluidynamic study through appropriate simulation
code. The purposes of the work and the time available for the risk analysis force to the use of simplified
methods/models which give the pressure peak vs. the
distance.
In literature (Lees, 1996) many simplified models
for the estimation of the overpressure originated by an

1022

explosion are available. The most diffused methods

are the following:
1. the equivalent TNT model;
2. the equivalent piston.
The equivalent TNT method
The equivalent TNT method approaches the effects of a
deflagration to those of an ideal detonation. It is based
on the correlation between overpressure and released
energy which is used for solid explosives (such as
TNT). The relationship overpressure/distance is ideal
and, due to the high density of solids, must be adapted
to liquid and/or gaseous explosives.
This method, using the data available for TNT
explosions, predicts the consequences of explosions
resulting from other sources. The first step is to establish the energy released by the explosion. This can
be done by combining the total available energy with
an estimation of the combustion yield, sometimes
expressed in terms of the efficiency of the explosion.
This energy is then used to calculate an equivalent
mass of TNT. The consequence of the explosion of this
mass of TNT can then be found from various graphs
and tables of data or using appropriate correlation.
The combustion enthalpy of TNT is 4196 kJ/kg,
thus the equivalent mass of TNT for a certain explosion
is given by equation (2):
mTNT =

HC
mcloud
4,196 106

(2)

where: mTNT = equivalent mass of TNT (kg); =

yield factor; HC = enthalpy of combustion of the
explosive (kJ/kg); mcloud = mass of the explosive (kg).
The scaled distance is then estimated using equation (3). In the correlation overpressure is a function of
the distance from the centre of the explosion (scaled
distance) and the magnitude of the TNT equivalent
mass (Yellow Book, 1997).

x = (mTNT )1/3 exp 3.5031 0.724 ln(Op )

+ 0.0398 (ln(Op ))2
(3)
where: x = scaled distance (feet); Op = overpressure
(psi).

expands to a hemisphere volume V1 . In this process the

energy added to the unburnt gas is assumed to be added
by a hemispherical piston, this process effectively
replacing the combustion process.
The overpressure and the duration time are both
function of the distance from the centre of the explosion and can be calculated using graphs or the
equations reported in Lees (1996).
Damage zones calculation
The damage zones of a potential accidental scenario
can be calculated using both the methods described
above. The models allow to quantify the impact distance, it is the distance where the pressure wave
reaches the value of 0.03 bar. Such value of overpressure is the threshold limit causing reversing lesions.
Over such distance the effects on the people are
negligible.
The consequences estimation can be done using a
simulation code based on the models described before.
Since the application of the TNT model is very
simple and the algorithm of the simulation code, available in the laboratory of the Department of Industrial
Chemistry and Material Engineering of the University of Messina, is based on equation equations (2) e
(3), the TNT method has been applied in the phase of
validation of the procedure.
The information required for the consequences
assessment, using both the method, are the flow rate
Q (kg/s) of the released substance and the distance of
permanence of the cloud dz (m). To define this last
parameter it is necessary to note that after the release
the cloud exists for a certain time in the area and, in
presence of ignition sources, can potential cause an
explosion. The distance of permanence indicates the
maximum dimension of the explosive cloud.
4.5 Number of people in the damage zone
The number of the people involved is calculated taking
into account the frequency of presence of personnel in the workplace. It is necessary the knowledge
of the type activities of the workers (shift workers,
head-shifts, maintenance staff).
The number of workers involved in a potential
explosion can be calculated according the equation (4).

The equivalent piston model (TNO shock wave

model)

D = (Ai /AImp )pp

The models of shock wave model have been developed by TNO and is described in the Yellow Book
(1997). It is also known as the expanding piston or
piston blast model and allows the peak overpressure
and the duration time of explosion to be estimated.
This model is based on the assumption that the
cloud is a hemisphere of unburnt gas volume V0 which

Where: Ai = impact zone of the explosion; AImp = the

whole area of the establishment; pp = probability of
presence of personal in the establishment.
The impact zone of a potential accidental scenario
is a circle whose ray is the distance from the centre
of the explosion. The pressure wave inside this area
reaches 0.03 bar.

1023

(4)

4.6

Assessment of the risk

The calculation of the ATEX risk index is the most

important step of the whole procedure. In this work
it has been proposed to apply the equation (1), each
parameter of the correlation has been described in the
previous paragraphs. The knowledge of them permits
to determine the risk index.
The estimation of the ATEX risk does not complete
the analysis, it must be combines with the evaluation
of the risk acceptability.
In this paper it has been proposed to refer to the
same threshold values used in industries at major risk
to judge the risk acceptability. Safety Reports of such
industries analyze also explosions, these accidental
scenarios are different from those described in the document of the evaluation of risk due to the presence of
explosive atmospheres.
The type of explosions analyzed in the Safety
Reports are originated by great releases of flammable
substances and, consequently, impact on large areas.
For establishments not at major risk, the explosions
have a smaller impact area and involve only the
workers.
Since both the type of explosions can be studied in
the same way it is opportune to uniform the approaches
of risk evaluation. This work represents a first attempt
to uniform the approaches.
In order to adopt measures of prevention and protection, the calculation of the risk must be combines
with the judgment about the risk level.
The same criteria for the evaluation of the risk
acceptability used in the industries at major risk can
be applied. The Italian normative does not defined
acceptability criteria, the risk judgment is made referring to the threshold values of frequency and consequences of the D.P.C.M. of the 25th of February
2005.
Concerning the ATEX risk, in this work it has
been proposed to refer to the risk acceptability
criterion adopted in the United Kingdom and described
by LUPACS (1999) and ESCIS (1998). On the
base of the English criterion the threshold values of
risk are:
Rae < 1065
106 < Rae < 104
Rae > 104

4.7

the risk is acceptable

it is necessary to reduce the
risk as low as technically
and economically possible
the risk is not acceptable

Definition of protective measures

The proposed approach allows to verify the effects

of the adopted measures of protection to reduce the
risk. Based on the risk assessment and through on-site
inspections the necessity of further implementation of

protective and preventive measures can be verified. In

order to achieve this aim a detailed check-list has been
drawn.

CONCLUSIONS

The proposed methodology permits to identify the critical points in the system (technical and procedural)
and decrease the exposure of the workers as low as
possible.
In particular the quantitative approach allows to not
underestimate the risks for the exposed workers.
The quantitative evaluation of the risk ATEX allows
to obtain an improved effective in the prevention and
protection interventions adopted by the company.
A tool for the risk assessment has been created, it
allows to repeat the calculations and permits a faster
verifications of the possible improvement of the measures of risk prevention and mitigation for the system
under analysis.
Finally the quantitative analysis permits a detailed
study of the accidental scenarios due to small releases.
For industries at major risk a detailed analysis of
such events is necessary because they can represent
potential sources of domino effects.

REFERENCES
Benintendi, R., Dolcetti G., Grimaz Stefano, Iorio C.S.
A comprehensive approach to modelling explosion phenomena within ATEX 1999/92/CE directive, Chemical
Engineering Transactions. Chemical Engineering Transactions, proceeding of 2nd International Conference
on Safety & Environment in Process Industry CISAP2.
(2006).
Chems Plus v.2 Manual (1991).
Council Directive 89/391/EEC on the introduction of measures to encourage improvements in the safety and health
of workers at work (1989).
Council Directive 94/9/EEC on the approximation of the
laws of the Member States concerning equipment and protective systems intended for use in potentially explosive
atmospheres (1994).
Council Directive 99/92//EEC on the approximation of the
laws of the Member States concerning equipment and protective systems intended for use in potentially explosive
atmospheres (1999).
Decreto Legislativo 626/94 riguardante il miglioramento
della salute e sicurezza dei lavoratori durante il lavoro
(1994).
D.P.C.M. 25 febbraio 2005 recante le Linee Guida per
la predisposizione del piano demergenza esterna di cui
allarticolo 20, comma 4, del decreto legislativo 17 agosto
1999, n. 334 (2005).
Decreto Legislativo 233/03. Regolamento recante norme
per lattuazione della direttiva 99/92/CE relativa alle

1024

prescrizioni minime per il miglioramento della tutela della

sicurezza e della salute dei lavoratori che possono essere
esposti al rischio di atmosfere esplosive (2003).
DPR 126 del 23 marzo 1998. Regolamento recante norme per
lattuazione della direttiva 94/9/CE in materia di apparecchi e sistemi di protezione destinati ad essere utilizzati in
atmosfera potenzialmente esplosiva (1998).
ESCIS. Introduction to Risk Analysis, Approaches and
Methods. Booklet 4 (1998).
Lees, F.P. Loss Prevention in the Process Industries. Elsevier
(1996).

LUPACS. Methodology & Case Studies. (1999).

Pezzo, T., Vannucci, D. & Uguccioni, G. Semiquantitative Analysis for Assessment of Explosion Risk for
WorkersApplication of the Requirements of the ATEX
Directive. Chemical Engineering Transactions, proceeding of 2nd International Conference on Safety & Environment in Process Industry CISAP2. (2006).
TNO Yellow Book. Methods for the calculation of physical effects due to releases of hazardous materials (liquids
and gases). Eds: van den Bosch, C.J.H. and Weterings,
R.A.P.M. (1997).

1025

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

A risk theory based on hyperbolic parallel curves and risk assessment

in time
Nicolae Popoviciu & Floarea Baicu
Hyperion University of Bucharest, Bucharest, Romania

ABSTRACT: The work deals with the risk assessment theory. An unitary risk algorithm is elaborated. This
algorithm is based on parallel curves. The basic risk curve is of hyperbolic type, obtained as a multiplication
between the probability of occurrence of certain event and its impact (expected loss). Section 1 refers the problem
formulation. Section 2 contains some specific notations and the mathematical background of risk algorithm. A
numerical application based on risk algorithm is presented in section 3. Risk evolution in time is described in
section 4. A generalization of the curve of risk and other possibilities to define the risk measure are given in the
last part of the work.

PROBLEM STATEMENT

The topic of the work is the following: notations

and specific definitions, the mathematical background
of risk theory and the risk computation algorithm,
followed by a numerical example.
1.1

Specific definitions

The risk is defined by the international standard

ISO/CEI 17799 2005 as being the combination
between the probability of occurrence of certain event
and its consequences (expected loss).
The risk level is an arbitrary indicator, denoted
L, which allows grouping certain risks into equivalency classes. These classes include risks which
are placed between two limit levelsacceptable and
unacceptableconventionally established.
The acceptable risk level is the risk level conventionally admitted by the organization management,
regarded as not causing undesirable impacts on its
activity. This is determined by methodic and specific
evaluations.
The risk value is a positive real number R which may
be represented by the area of a quadrangle surface,
having as one side the Probability of occurrence of
certain event, noted with P, and as the other side the
consequences of that event occurrence, respectively
the Impact of the undesirable event, noted with I , upon
the security of the studied organization.
Mathematically speaking, the same area may be
obtained through various combinations between P
and I , of which the preferred one is quite the product
between probability and impact.

Figure 1. Graph representation for equivalency of risks

defined by different probabilityimpact couples.

There are a lot of Probability Impact couples generating the same Risk R, defining quadrangles of same
area as illustrated in figure 1.
If the vertexes of such quadrangles, which are not
on axes, are linked through a continuous line it results
a hyperbolic curve C, named the curve of risk. This
curve allows the differentiation between the acceptable
risk (TolerableT) and the unacceptable one (NonTolerableNT).

1027

Thus, the risk of occurrence of a certain event

A, with high impact and serious consequences but
low probability of occurrence, defined by coordinates placed below the represented acceptability curve
is considered acceptable, while the risk of event B,
with less serious consequences but high probability of
occurrence, of which coordinates are placed upwards
the curve, is considered unacceptable.
The hyperbolic curve of risk based on couples
(Probability, Impact) is illustrated in figure 1.
This idea is in accordance with wikipedia/risk
assessment 2008 and wikipedia/risk 2008, because
the quantitative risk assessment requires calculations
of two components of risk: the magnitude of the
potential loss and the probability that the loss will
occur.
Any placement of an assessed risk on a particulate
risk level L represents a major and critical decision
for any manager of security system and raises minimum three heavy and sensitive problems, respectively
(Baicu & Baicu 2006a, b):
1. How to define the risk coordinates, respectively the
ProbabilityImpact couple?
2. How to set out the Curve of Risk for assessing
which risk is tolerable (acceptable) and which is
not?
3. How many curves are to be set out, respectively
how many risk levels are to be defined and which
is the distance h between different curves?

1.2

Classes of impact and probability

A set of 7 levels I1 , I2 , . . ., I7 is proposed in relation

to impact, respectively: insignificant, very low, low,
medium, high, very high and critical. These impact
levels are established considering the values of losses
that may occur as a risk consequence (Baicu & Baicu
2006a).
As to the event occurrence probability the proposal
is for 9 (nine) classes, by assigning a probability class
to each probability of occurrence, starting with the
most rare event, class 1, and up to the most frequent
event, the class 9, counted to be the highest one, class
9, as is presented in the table 1, column 3.
For the lowest probability of occurrence selected
for this work, class of probability 1, the assumption
for the occurrence of a certain event is once at every
20 years, respectively an event at 175200 hours, gives
a probability of 1/175200 5 106 , mentioned in
column 3, line 1 of table 1.
For the second probability of occurrence mentioned in column 3 of table 1, the assumption for
the occurrence of a certain event is once at every
2 years, respectively a probability of 1/17520
5 105 .

Table 1.

With the classes of probabilities.

No.

Frequency
of occurrence

Probability

Class of
probability

One event at 20 years

One event at 2 years
One event per year
One event at 6 months
One event per month
Two events per month
One event per week
One event per day
One event at each hour

5 106

1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
9

5 105
104
2 104
1.4 103
3 103
6 103 h
4 102 h1
1 h1

For all the other probabilities mentioned in column

3 of table 1, the assumption is of more frequent occurrences, namely those mentioned in column 2 of this
table. The calculation is made similarly, the probabilities being based on the occurrence frequency measured
in hours (Baicu & Baicu 2006a).
It is assumed that the most frequent event occurs
during each functioning hour, its probability being 1,
line 9 of table 1.
Now we have an important remark. Very often the
risk theory uses the FN-curves (Evans 2003). FNcurves are a graphical presentation of information
about the frequency of fatal accidents in a system
and the distribution of the number N of fatalities in
such accidents. They are decreasing functions. The
FN-graphs are usually drawn with logarithmic scales
because the values of both N and F(N) have a big order
of magnitude. The FN-curves are similar to histograms
and indeed present the same information as histograms
in a different way.
The remark is that the FN-curves are calculated
from empirical frequency data on past events. In this
work we do not use the FN-curves because we are interested in the future activity, namely we have to help the
system manager to optimize the future activity of the
system.
1.3 Risk levels
The classes of probabilities and the impacts generate a set of 9 7 = 63 points which represents the
states of a system. A system manager needs a rule to
classify these points. The rule is based on levels or
risk.
The section delimitated between two consecutive
curves Cj represent the level of risk, denoted Lj. We
consider the case (Cj , Cj+1 ), which generates the levels
L1, L2 and so on.

1028

the following functions of three variables (Popoviciu

2002).

MATHEMATICAL SUPPORT

For the orthogonal system xOy we use a set of known

points:
Aij = A(xi , yj ),
(xi , yj ) N N

i = 1, m;

y=

j = 1, n

c
;
x

ch
X (x, c, h) = x +
c2 + x 4

Y (x, c, h) =

or (xi , yj ) R+ R+

where m = number of the probability classes; n =

number of the impact classes.
Explicitly, the points Aij are
(x1 , y1 )(x2 , y1 ) . . . (xm , y1 ) impact of level 1;
(x1 , y2 )(x2 , y2 ) . . . (xm , y2 ) impact of level 2;

hx2
c
+
x
c2 + x 4

(2)
(3)

and activate them by Insert, Graph, X-Y Plot.

Remark 2 Also it is possible to formulate the inverse
problem: having the equation of and the known point
N (a, b), we look for the value of h and the equation of
 so that N  and // .

.................................

Proposition 2 In the above conditions we obtain the

algebraic equation:

(x1 , yn )(x2 , yn ) . . . (xm , yn ) impact of level n.

x4 ax3 + bcx c2 = 0,

We propose that the point (xm , y1 ) belongs to the

hyperbolic curve
C1 : xy = R1 , where R1 = xm y1 is the risk value (or
risk measure, measure of risk) and so C1 is the curve
of risk for the risk value R1 .
Let r be a natural number that represents the total
number of curves of risk. The value of r is at users
disposal. Hence there exist the risk curves C1 , . . . , Cr .
Now the problem is: how to define and construct
them? There exist several possibilities. In this work
we chose the case of parallel curves.

(proof, see Popoviciu & Baicu 2007).

By a numerical method we solve the equation (4)
and if x be the chosen solution, then

2.1 The general theory of parallel curves

A local problem formulation is the following: given a
known curve , let us construct a new curve  so that
//  ( parallel with ). In our case is hyperbolic
curve. Let (x, y) be a current point on and (X , Y ) a
current point on .
Proposition 1 If has the explicit equation y =
c
x , x  = 0, c > 0 then the parallel curve  has the
parametric equations:
X =x+

ch
c2 + x 4

Y =

hx2
c
+
x
c2 + x 4

(1)

h=

a x 2
c + x4 .
c

Having the value h we can construct  by (2),(3).

2.2 The construction of r parallel curves of risk
2.2.1 The risk algorithm
We recall that the parameter r defines the total number
of parallel curves of risk and r is known i.e. it is given
by the user. So we have to construct r parallel curves
of risk Cj , j = 1, r where C1 is a known hyperbolic
curve. The others curves C2 , . . . , Cr are constructed
in the manner shown below and called the algorithm p
c (algorithm of parallel curves).
1. We denote V the point A(xm , yn ) and the equation
of line (OV ) is y = (yn x)/xm .
2. Let B1 be the point defined as B1 = (OV ) C1 and
denote the coordinates by B1 (a1 , b1 ).
3. Write the normal in B1 for the hyperbolic curve
f (x) = c/x, x  = 0. We obtain successively the
slopes (for tangent and normal) and the equation:

where h = d(M , N ), M , N  is the constant distance between any two points on the common
normal. (proof, see Popoviciu & Baicu 2007).
Remark 1 It is very difficult to eliminate the variable
x from (1) and to obtain the explicit equation Y =
F(X ) for the curve . In numerical applications we
set c = R1 .
In order to draw the parallel curves and , let
us say by a Mathcad Professional subroutine, we use

a, b, c known values (4)

mtan =

c
,
a21

mnorm =

a21
c

y b1 = mnorm (x a1 ) (the normal in B1 ).

4. Let Br+1 be the point defined as the intersection

1029

Br+1 = (normal in B1 ) (y = yn ),
Br+1 (ar+1 , br+1 )

5. Compute the distance B1 Br+1  and denote the step

between the parallel curves by h = B1 Br+1  /r,
where h appears in parametric equations (1).
6. We use the following correspondence for curves
C2 with h2 = h,

C3 with h3 = 2h an so on

Cr with hr = (r 1)h

(5)

The values hj , j = 2, r are used in equations (2).

7. Compute the risk values Rj corresponding to each
curve Cj , j = 2, r. Up to now we know, by computation, the coordinates of the points B1 (the first
point) and Br+1 (the last point).
7.1. We also need the coordinates of the points
B2 (a2 , b2 ), . . . , Br (ar , br ).
For that one uses the formulas


B1 Bj 

 = kj , j = 2, r, kj Q
Bj Br+1 
a1 + kj ar+1
b1 + kj br+1
, bj =
, j = 2, r
aj =
1 + kj
1 + kj
7.2. The value of risk Rj is defined as Rj = aj bj and
the corresponding parallel curve Cj is given by
parametric equations (2), (3).
8. On a single orthogonal system xOy we draw: all the
parallel curves Cj , j = 1, r, all the points Ai j and all
the horizontal and vertical lines x = xi , i = 1, m;
y = yj , j = 1, n
We remark that the space R2 could be considered
a affined space in this drawing, namely the unite
measures on Ox and Oy could be different.
9. The level of risk is defined as the set of points (P, I )
settled between two successive curves of risk but
so that PI Rj We consider the cases (Cj , Cj+1 ).
So we obtain:
Level L1 contains the points Aij settled under C1
and on C1 ;
Level L2 contains the points Aij settled between
C1 , C2 and on C2 ;
Level L3 the points between C2 and C3 ; and so on.

m = the number of classes of probabilities; m = 9 on

Ox axis;
n = the number of levels of impact; n = 7 on Oy axis;
r = the number of risk levels; r = 6;
xi = pi the class of probability number i;
yj = Ij the class of impact number j;
xi = i, i = 1, 9; yj = j, j = 1, 7; Aij = A(Pi , Ij ).
For simplicity we suppress the lower index and denote
hj = hj, Cj = Cj,

Rj = Rj,

Bj = Bj.

So we use the notations:

h2, h3, . . . , h6; C1, C2, . . . , C6;
R1, R2, . . . , R6; B1, B2, . . . , B6, B7.
By numerical computations we obtain successively

c = 9, (OV ) (xy = 9) B1(9 7/7; 7)

(normal in B1): 1.286x y 1.729 = 0
(normal in B1) (y = 7) B7(6.778; 7)
B1B7 = 5.517; h = 5.517/6 = 0.919
h2 = h = 0.919, h3 = 2h = 1.838
h4 = 3h = 2.757, h5 = 4h = 3.676
h6 = 5h = 4.595.
The curve Cj has the parametric equations
Xj(x, c, hj) = x +
Yj(x, c, hj) =

c hj
c2 + x 4

, j = 2, 6,

x2 hj
c
+
, j = 2, 6.
x
c2 + x 4

In order to draw the parallel curves Cj we use the

Mathcad library. The unit of measure on Ox is different
of the unit on Oy axis.
In figure 2, on axis Ox and Oy is written the
information used by Mathcad, respectively

We call the risk algorithm p c (risk algorithm of

parallel curves) the initial data introduction and all the
above steps 19.

x, X 2(x, c, h2), . . . , X 6(x, c, h6)

c
, Y 2(x, c, h2), . . . , Y 6(x, c, h6), 1, 2, . . . , 6.
x

The user must draw by himself the parallel lines

having the equations

NUMERICAL APPLICATION BASED

ON THE RISK ALGORITHM P C

Remark 3 In applications, the mathematical notations from section 2 have the following meanings:

x = 1, x = 2, . . . , x = 9 in figure 2.

1030

Table 2.

R1
R2
R3
R4
R5
R6

4
Figure 2.

Six parallel curves of risk Cj.

The values of risks are obtained from 7.1 and 7.2 of

the above algorithm:
B1(9; 1),
B2(3.966; 3.372),
B3(4.531; 4.097),
B4(5.095; 4.823),
B5(5.659; 5.549),
B6(6.224; 6.274),

R1 = 9
R2 = a2 b2
R3 = a3 b3
R4 = a4 b4
R5 = a5 b5
R6 = a6 b6

The risk values in time.

t=0

t=1

t=2

t=3

t=4

9.00
13.37
18.56
24.37
31.40
39.07

24.40
36.26
50.34
66.63
85.15
105.95

66.18
98.34
136.51
180.70
230.92
287.33

179.48
266.68
370.20
490.04
626.22
779.20

486.71
723.20
1004.00
1329.00
1698.00
2113.00

EVOLUTION AND RISK ASSESSEMENT

IN TIME

Generally, the value of risk increases in time. We consider the value Rj from step 7.2 as the corresponding
value for time t = 0. For time t we propose the formula
Rj (t) = Rj eat ,

= 13.373
= 18.564
= 24.573
= 31.402
= 39.037.

a > 0, j = 1, 2, . . ..

Level 1: {(P, I )/0 PI R1}

(1,1), (1,2), (1,3), (1,4), (1,5), (1,6), (1,7), (2,1),
(2,2), (2,3), (2,4), (3,1), (3,2), (3,3), (4,1), (4,2),
(5,1), (6,1), (7,1), (8,1), (9,1);

In order to obtain an appropriate set of rules S

(level j) and an appropriate parameter a or even formula Rj (t), a simulation should be performed on the
real system under investigation.
The setting of rules S(level j) and the formula Rj (t)
have to be simulated
For simplicity, we take the parameter a = 1 and
t = 1, 2, 3, 4, . . . So, in time, the values of Rj generate
the table 2. The used value of e was e = 2.71, but a
better value could be recommended.
If the rules S(level j) are not applied immediately after the event occurrence, then there was the
possibility for an extended risk, denoted ER and
defined as

Level 2: {(P, I )/R1 < PI R2}

(2,5), (2,6), (3,4), (4,3), (5,2), (6,2);

ER = Pi Ij Pi+1 Ij+1 .

A number of 7 (seven) levels of risk were assessed

according to the figure 2 and to this correspond the
following couples (Probability, Impact):

Level 3: {(P, I )/R2 < PI R3}

(2,7), (3,5), (3,6), (4,4), (5,3), (6,3), (7,2), (8,2),
(9,2);

Example.

ER = 4 4 5 5 = 400

or

ER = 4 3 5 4 = 240.

Level 4: {(P, I )/R3 < PI R4}

(3,7), (7,3), (4,5), (4,6), (5,4), (6,4), (8,3);

Now, the extended risk value ER must be interpolate

in table 2 in order to establish the time value t and the
risk level. For ER = 400 we obtain 370.20 < 400 <
490.04 and take t = 3; R3 ; S(level 3).
Hence, in order to avoid a new undesirable impact,
the manager should apply the set of rules S(level 3) in
at most 3 units of time.

Level 5: {(P, I )/R4 < PI R5}

(4,7), (7,4), (9,3), (5,5), (5,6), (6,5);
Level 6: {(P, I )/R5 < PI R6}
(5,7), (7,5), (6,6), (8,4), (9,4);
Level 7: {(P, I )/PI > R4}
(6,7), (7,6), (7,7), (8,5), (8,6), (8,7), (9,5), (9,6),
(9,7).

For each level of risk, the manager formulates and

establishes a set of rules S(level j) to be used when
certain event has happened.

THE GENERALIZATION OF THE RISK

CURVE

We have used up to now an explicit hyperbolic curve

of risk : xy = c.

1031

Now the curve has the general explicit equation

y = f (x) with function f continuous, derivable and
f
(x) = 0 for let us say x 0.
Let M ,N be two points so that
M (x, y)

and

Now we define several measures of risk.

I. RI (X ) = Var(x)

(8)

When the computation is based on estimators we

denote it by R

N (X , Y ) .

Proposition 3 The parametric equations of  are



f (x)
(6)
X = x + h
1 + f
2 (x)


f (x)
1
(7)
Y = f (x) h


f (x) 1 + f
2 (x)

II. We denote by T a threshold of risk and define

RII 1 (X , T ) = M [(T X )+ ]2

(9)

(one eliminates the values xi > T )

RII 2 (X , T ) = M [(X T )+ ]2

(10)

See the proof in Popoviciu & Baicu 2007.

If we put in (6),(7) the function f (x) = c/x, one
obtains the parametric equations (1).

(one eliminates the values xi < T ).

III. Generalization

Remark 4 The risk algorithm p c with the steps 1-9

could be repeated for any function f (x).

RIIIp (X , T ) = M [(T X )+ ]p ,

OTHER POSSIBILITIES TO DEFINE

THE MEASURE OF RISK

SD2 (X ) = S 2 (X ) = SVar(X )
RIV (X ) = SVar(X ) = M [(m X )+ ]2

Ss2 (X ) =

pi = 1

and the measure of risk is a real number R(X ).

There are several possibilities to define R(X ). In
order to do that we recall several usual notations of
probability domain:
n
M (X ) = m =
xi pi is the mean value;
i=1

n
1 
[(X xi )+ ]2
n 1 i=1

RIV (X ) = Ss2 (X ).
V. The Taguchi [6] measure of risk (in industrial
practice) is defined as
RT (X , T ) = k[Var(X ) + (m T )2 ],

(x)+ means (x)+ = max{0; x}.

An estimation of Taguchis measure of risk is

The non-shifted estimators of mean value and

dispersion are
s2 (X ) =

n
2
1 
X xi .
n 1 i=1

k>0
(14)

is the dispersion or the variance of X ;

(13)

The user is interested only in dispersion of values

xi so that xi < X .
An estimator of risk measure is

D2 (X ) = 2 (X ) = Var(X ) = M [(m X )2 ]

1
xi ,
n i=1

(12)

An estimator of semi-dispersion is

i=1

X =

(11)

IV. The risk measure based on semi-variance or semidispersion, denoted by S

In section 2 we have defined the measure of risk by

using a hyperbolic curve and denoted R1 = xm y1 .
From mathematical point of view (Radulescu 2006,
chapters 2 and 3), the risk is a random variable X .
In a finite state space we denote the discrete random
variable


x x2 . . . xi . . . xn
X = 1
, pi 0,
p1 p2 . . . pi . . . pn
n


p N

RT (X , T ) = k[s2 + (X T )2 ]

(15)

Each user should choose the appropriate law of

measuring the risk.
Numerical examples.

1032

1. Let X be a random variable with the probability

repartition

15

X =

12

11

14 18 15
5
4
5

1
12

1
12

1
12

1
12

1
12

1
12
1
12

1
12

1
12
1
12

1
12
1
12

a. Risk measure based on estimators.

The mean value and its estimation are
m = M (X ) = 8.917;

X = 8.917.

The estimation of dispersion and risk measure (8) are s2 (X ) = 32.354; RI (X ) = 32.354.
b. The risk measure based on semi-dispersion is
obtained as it follows
12



(X xi )+

2

= 184.73

i=1

with good results, in any risk assessment, depending

on the probability of occurrence of an undesirable certain event generating the risk and consequences of the
respective event.
In this purpose, the scales for the two axes (Probability and Impact) are conveniently selected and those
two coordinates assessing the risk are represented on
the graphic in figure 2.
According to the tolerable level accepted for the
regarded risk, respectively to the plotted curve of risk,
the positioning of the point in regard to this curve is
to be analyzed. If the point is placed below the plotted
curve, the presumed risk is acceptable and implicitly
does not require for any decreasing measures to be
adopted. Otherwise, if the point is placed upwards
the risk curve, some measures have to be taken for
decreasing the risk to a tolerable level.
The priorities of treatment are determined for each
regarded value of risk according to the assessed risk
level (section 4).
There are also others possibilities to define the value
of risk, based on mean value and variance.

RIV (X ) = Ss2 (X ) = 16.794.

Remark 5 Because we use only the estimators, the
probabilities pi are not necessary
2. X = (21 21 30 22 32 19 3 3 5 8 12 11)
a. The risk measure based on estimators.
X = 15.583;

s2 (X ) = 100.72

RI (X ) = s2 (X ) = 100.72.
b. The risk measure based on semi-dispersion
12



(X xi )+

2

= 520.008

i=1

RIV (X ) = Ss2 (X ) = 47.273

7

CONCLUSIONS

The assessment of risks specific to a certain security

system should be performed on strict and objective
mathematic basis which can be repeated in time and
the results obtained to be reproducible. The assessment of risks on particular levels, mathematically and
objectively determined, allows selecting the options of
treatment, so that the management decision should be
impartial.
The proposed method has been elaborated for
assessing the risk level specific to the information
security within an organization, but it may be applied,

REFERENCES
Anderson, K. 2005. Intelligence-based threat assessments for
information networks and infrastructures. A White Paper.
Baicu, Floarea & Baicu, A.M. 2006a. Audit and security of
information systems. Victor Publishing House, Bucharest.
Baicu, Floarea & Baicu, A.M. 2006b. Risks associated to
security of information systems. Hyperion University
of Bucharest. Annals mathematics-informatics. 5966.
Bucharest.
Ball, D.J. & Floyd, P.J. 1998. Societal risks. Final report to
HSE. HSE Books.
BS ISO/IEC 17799. 2005. Information technology. Security techniques. Code of practice for information security
management.
Evans, W. Andrew. 2003. Transport fatal accidents and
FN-curves: 19672001. United Kingdom. Health &
Safety Executive. HSE Books. United Kingdom.
Kendrick, Tom. 2003. Identifying and managing project
risk: essential tools for failure-proofing your project.
AMACOM. American Management Association.
Lerche, I. 2006. Environmental risk assessment: quantitative measures, anthropogenic influences, human impact.
Berlin, Springer.
Panjer, H. Harry & Willmot, E. Gordon. 1992. Insurance
risk models. Society of actuaries. Actuarial education and
research fund. Library of Congress. USA.
Popoviciu, Nicolae & Baicu, Floarea. 2007. A new approach
for an unitary risk theory. WSEAS Press. Procc. of the 7th
WSEAS intern. conference on signal processing, computational geometry and artificial vision. 218222. August.
Athens.
Radulescu, M. 2006. Modele matematice pentru optimizarea investitiilor financiare. Editura Academiei
Romne. http://en.wikipedia.org/wiki/Risk_assessment.
2008. http://en.wikipedia.org/wiki/Risk. 2008.

1033

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Accident occurrence evaluation of phased-mission systems composed

of components with multiple failure modes
T. Kohda
Kyoto University, Kyoto, Japan

ABSTRACT: General systems such as production plants and aircrafts can be regarded as a phased-mission
system, which perform several different functions depending on their operational stage. The definition of failure
for a component also changes depending on its operation stage, and thus multiple failure modes must be considered for each component. Further, the occurrence of an accident in the system depends on the operation stage
as well as the total operation time. This paper proposes a novel, simple accident occurrence evaluation method
for a phased-mission system composed of components with multiple failure modes. Firstly, based on the system
accident conditions at each phase, the system accident occurrence condition at present time is given analytically
in terms of component state conditions. Evaluating component state conditions, the system accident occurrence
probability can be easily evaluated. An illustrative example of a simple redundant pump system shows the details
and merits of the proposed method, where the obtained results are compared with those obtained by the state
transformation method.

INTRODUCTION

General systems such as production plants and

aircrafts perform several different functions depending on their operation stage, and so they should
be regarded as a phased-mission system (Burdick,
et al., 1977). In applying logical approaches such as
basic event transformation (Esary & Ziehms, 1975) or
sophisticated methods such as Markov chains (Dugan,
1991) and Petri-Nets (Mura & Bondavalli, 1999) to the
evaluation of mission reliability, the system success/
failure conditions at each phase must be given for
each phase. However, the system accident conditions
or minimal cut sets for the phased-mission system
cannot be obtained easily, because normal component condition changes depending on the processing
stage and so multiple failure modes must be considered. Since the definition of failure for a component
also changes depending on its operation stage, the
conventional binary state representation such as failure/success states cannot be applied effectively and
thus multiple failure modes must be considered for
each component.
This paper proposes a novel, simple accident evaluation method for a phased-mission system composed
of components with multiple failure modes. Firstly,
a state variable representation of component conditions with multiple failure modes is introduced, which
represents a component state at a specific time. A
component state can be among failure modes including

its normal state. For each failure mode, one binary

state variable is allocated. Due to the exclusiveness
of failure modes and their monotonic transition from
the normal state with time, the simplification rules are
given for the logical operation of component state variables. Then, based on the system accident conditions
(or minimal combination of component failures leading to a system accident) at each phase, the system
accident occurrence conditions that a system accident occurs at a specific phase can be obtained in
terms of component state conditions. The system accident occurrence conditions can be represented as a
logical OR of logical AND combinations of component state variable expressions. According to the
assumptions on multiple failure modes, probabilities
of component state conditions can be evaluated easily.
The quantitative evaluation of system accident probability can also be easily obtained from component
failure probabilities. An illustrative example shows the
merits of the proposed representation are shown compared with the conventional logical approaches such
as basic event transformation method.
2

SYSTEM ACCIDENT CONDITION

State variable expressions for components and system

are introduced to obtain a simple logical expression of
the system accident occurrence during its operational
stages.

1035

2.1

Assumption (a2) implies that

Assumptions

The following assumptions are made for the system

and components.
a1. A system accident occurs if any system accident condition represented in terms of component
conditions is satisfied at any phase.
a2. Minimal cut sets (minimal combinations of component state conditions) for each phase are given.
a3. The system state is represented in terms of states
of N components.
a4. The system state is as good as new at the start of
a mission, i.e., all the components are as good as
new.
a5. Component failures are statistically independent
a6. Component i has M i + 1 failure modes, which
are mutually exclusive. Failure mode 0 for each
component denotes its normal state.
a7. Without stop of the plant operation, no component
can be repaired or replaced. After component i
comes into failure mode j, it remains until the end
of mission, a shutdown, or a system accident.
a8. Transition rate of component i from its normal
state to failure mode j at time t is represented as,
where t is the elapsed time from the initiation of
the mission.
2.2

Component state representation

The state of component i at time t is represented by

binary variable X (i, j, t) as follows.

X (i, j, t)

1,

0,

if component i is in failure mode j

at time t,
otherwise
(1)

where j = 0, 1, . . ., Mi , Since component i must be in

a state: it can be a failure mode or the normal operating state, the following equation must hold from
assumptions (a6) & (a7):
Mi


X (i, j, t) = 1

(2)

j=0

X (i, 0, t) = 1 and X (i, j, t) = 0 for j  = 0.

Assumptions (a7) & (a8) imply that
X (i, 0, t1 ) X (i, 0, t2 )

(5)

X (i, j, t1 ) X (i, j, t2 ) for t1 t2 & j  = 0

(6)

Using this monotonic property of X (i, j, t) with

respect to time t represented as Eqs. (5) and (6), the
following simplification rules hold.
X (i, 0, t1 )X (i, 0, t2 ) = X (i, 0, t2 )

for t1 t2

X (i, j, t) = 0
j

for t1 t2 & j  = 0
X (i, 0, t1 ) X (i, 0, t2 ) = X (i, 0, t1 )

for t1 t2

(8)
(9)

X (i, j, t1 ) X (i, j, t2 ) = X (i, j, t2 )

for t1 t2 & j  = 0

(10)

where denotes logical OR combination. Due to the

mutually exclusiveness of failure modes, the following
simplification rules hold:
X (i, j1 , t1 )X (i, j2 , t2 ) = 0

for t1 t2 & j1  = j2 > 0

(11)

X (i, j, t1 )X (i, 0, t2 ) = 0

for t1 t2 & j  = 0

X (i, j, t) = X (i, k, t)

(12)
(13)

k =j

From Eqs. (7), (8), (10), and (13), the following

simplification rules holds
X (i, j1 , t1 )X (i, j2 , t2 ) = X (i, j2 , t2 )
(14)

X (i, j1 , t1 )X (i, j2 , t2 ) = X (i, j1 , t1 )

for t1 t2 & j1  = j2

(3)

where X (i, j, t) denotes the negation of X (i, j, t), and

denotes the logical AND combination.

(7)

X (i, j, t1 )X (i, j, t2 ) = X (i, j, t1 )

for t1 t2 & j1  = j2
Applying De Morgans laws (Hoyland & Rausand,
1994) to eq. (1), the following equation holds.

(4)

(15)

Further, considering possible state transitions from the

normal state, the following combination rule for event
sequences holds.

1036

X (i, 0, t1 )X (i, j, t2 ) + X (i, 0, t2 )X (i, j, t3 )

= X (i, 0, t1 )X (i, j, t3 )

for t1 t2 t3 & j = 0
(16)

For t1 = 0 where transitions occur from the initial

operating time, the above equation is simplified as:
X (i, j, t2 ) + X (i, 0, t2 )X (i, j, t3 ) = X (i, j, t3 )
for t2 t3 & j = 0
(17)
Let B(i, j, t1 , t2 ) denote event that component i experiences failure mode j between t1 and t2 can be represented as {component i is not in failure mode j at t1 }
AND {component i is in failure mode j at t2 }. Thus,
the following equation holds.

Now obtain the system accident occurrence condition at phase l. The above condition, Eq. (21), shows a
condition whether the system is suffering from the system accident at phase k, which does not always implies
that the system accident occurs at phase k because
another system accident might occur at the previous
phase. One of the most important characteristic is that
the system accident condition at the later phase does
not always lead to an immediate system accident even
if it is satisfied. A system accident occurs at phase
l requires that {no accident occurs before the phase
l} AND {an accident occurs at phase l}. Thus, system
accident occurrence conditions at time t during phase l
denoted by binary variable Ys(l, t), can be represented
in terms of Y (k, t), k = 1, . . ., l as:

Ys(l, t) =

B(i, j, t1 , t2 ) = {X (i, j, t1 ) = 0}AND{X (i, j, t2 ) = 1}

(18)
Using Eq. (7), this expression can be represented as:
B(i, j, t1 , t2 ) = {X (i, 0, t1 )X (i, j, t2 ) = 1}

(19)

Assumption (a7) on no repair of components until a

system accident implies that no component failure can
be detected before the system failure. Thus, the event
that component i is in failure mode j at time t means
that failure mode j of component i can occur at any time
before time t. This implication is used in obtaining the
system accident condition at time t using minimal cut
sets for the corresponding phase.
2.3 System accident occurrence condition at phase l

l1

k=1



Y (k, t) Y (l, t)

(22)

where each Y (k, t) can be represented in the similar

way as Eq. (21).
Using the simplification rules, Eq. (22) can be
obtained as logical OR of logical AND combinations, each of which corresponds to a minimal cut
set of system accident occurrence at phase l. System accident occurrence probabilities can be obtained
similarly using the inclusion-exclusion method (Hoyland & Rausand, 1994) for conventional minimal
cut sets.

2.4

Component state probabilities

Define state probability of component i at time t,

Q(i, j, t) as:

Similarly to component states, the system state at time

t can be represented as follows:

Y (k, t)

Q(i, j, t) Pr{X (i, j, t) = 1} = E{X (i, j, t)}

1,

if the system is in a system failed

state of phase k at time t,
otherwise
(20)

0,

Here, the existence of a system failed state of phase

k implies the occurring state of a minimal cut set of
phase k. According to assumption (a2), all minimal
cut sets for each phase are given. Using minimal cut
set C(k, m), m = 1, . . ., nk , for phase k, the system
state Y (k, t) can be represented as:
nk

Y (k, t) =

m=1

(i,j)C(k,m)

X (i, j, t)

where E{} denotes the expectation operation.

From assumptions (a5)-(a8), Q(i, j, t) can be
obtained as:

Q(i, 0, t) = exp (i, s)ds

(24)

t
Q(i, j, t) =

(23)

(21)

(i, j, ) exp (i, s)ds d

for j  = 0

1037

(25)

where
(i, t)

Mii


(i, j, t)

(26)

j=1

Note that the value of (i, j, t) mainly depends its

operation condition at each phase. For consistent transition rate (i, j, t) = (i, j), Eq. (24) and (25) can be
simplified into:

Q(i, 0, t) = exp (i)ds

(27)

t
Q(i, j, t) =

(i, j) exp (i)ds d

for j  = 0

(28)

where
(i)

Mii


(i, j)

(29)

j=1

Using these basic probabilities, the probability of event

B(i, j, t1 , t2 ) can be evaluated as follows:
Pr{B(i, j, t1 , t2 )} = Q(i, j, t2 ) Q(i, j, t1 )

(30)

2.5 System accident occurrence probability

ILLUSTRATIVE EXAMPLE

This section considers the phased-mission system composed of two redundant pumps as shown
in Figure 1. Analysis results are compared with the
conventional event transformation method.

3.1 System description

Two identical pumps are connected in parallel as
shown in Figure 1. At their full operation of both
pumps, the maximum flow can be obtained. In this
operational mode (denoted as mode 1), both pumps
must be running at its maximum performance. In the
ordinary operation (denoted as mode 2), the system
flow requirement is assumed as half the maximum
performance. In the ordinary operation, either one
pump can satisfy the operation requirement. Lastly,
in the non-operational or stand-by operation (denoted
as mode 0) where no flow exists, both pumps must be
stopped.
Corresponding to the above operation requirements, system structure functions and normal states
of pumps are defined as follows:
0. Mode 0: The system structure is a series of two
pumps. For each pump, its normal state is defined
as non-operation.
1. Mode 1: The system structure is a series of two
pumps. For each pump, its normal state is defined
as operation with full performance.
2. Mode 2: The system structure is a parallel of two
pumps. For each pump, its normal state is defined
as operation with required performance.

The system accident occurrence probability at time t

during phase l for the phased-mission system, QS (l, t),
can be obtained as:
QS (l, t) = Pr{Ys(l, t) = 1}




l1
=E
Y (k, t) Y (l, t)
k=1

(31)

As shown in equation (21), Y (k, t) are represented as

logical OR of {logical AND combination of X (i, j, t)
at different time t}, and so the negation of Y (k, t)
are represented as logical AND of {logical OR combination of the negation of X (i, j, t) at different
time t}. Expanding the logical expression of the
right term in equation (31) by applying the simplification rules of X (i, j, t), the reduced form can be
expressed as logical OR of logical AND combination
of {X (i, j, t), X (i, 0, t1 )X (i, j, t2 ), or the negation of
X (i, j, t)}. Using the statistical independence of component failures, the system occurrence probability can
be evaluated using component failure probabilities.

Figure 1.

1038

Redundant pump system.

Depending on the system operation mode and its

performance requirement, the following two failure
modes are assumed for pumps.
1. inadvertent operation failure, running when it
should stop,
2. non-operation failure, or failure to operate when it
should.
For operation modes 1 and 2, non-operation failure
mode must be considered, while inadvertent operation failure mode should be considered for operation
mode 0. Thus, in the phased-mission system, failure
modes to be considered at each phase depend on its
operational mode and performance requirement.

3.2

system accident occurrence condition at phase 2 is

given as:


Y (2, t) = X (1, 2, t1e) X (2, 2, t1e)
{X (1, 1, t) X (2, 1, t)}

Finally, using the simplification rule (Eq. (14)), the

above equation can be reduced as:
Y (2, t) = {X (1, 1, t)} X (2, 1, t)}

Y (3, t) = (X (1, 0, t2e) X (1, 2, t) X (2, 1, t1e)

Mission operation

(X (1, 0, t2e) X (1, 0, t1e)

(X (2, 0, t2e) X (2, 2, t) X (1, 1, t2e)
(X (2, 0, t2e) X (2, 0, t)
(36)

Y (4, t) = [X (1, 0, t1e) X (1, 1, t) {X (2, 0, t3e)

X (2, 0, t1e) X (2, 1, t3e)}]
[X (2, 0, t1e) X (2, 1, t) {X (1, 0, t3e)

System accident occurrence conditions

Firstly, system accident occurrence conditions are

obtained by the proposed method as follows:
For phase 1, a system accident occurs if a minimal
cut set occurs, and thus, Y (1, t) can be obtained as:
Y (1, t) = X (1, 2, t)} X (2, 2, t)}

X (1, 1, t1e)) X (2, 2, t)

X (2, 1, t2e)) X (1, 2, t)

Phase 1: {X (1, 2, ) = 1}, {X (2, 2, ) = 1}

Phase 2: {X (1, 1, ) = 1, X (2, 1, ) = 1}
Phase 3: {X (1, 1, ) = 1}, {X (2, 1, ) = 1}
Phase 4: {X (1, 2, ) = 1}, {X (2, 2, ) = 1}

Here time parameter denotes any time during the

corresponding phase, and component i corresponds to
pump i.

3.3

(35)

Similarly, system accident occurrence conditions at

phases 3 and 4, Y (3, t) and Y (4, t), are obtained as:

For simplicity, the mission operation considered in this

example is composed of 4 phases: mode 0, mode 2,
and mode 1 and mode 0. According to the system
structure function considered in the previous section,
minimal cut sets for each phase can be obtained as
follows:
1.
2.
3.
4.

(34)

(32)

For phase 2, a system accident occurs if no minimal

cut set occurs during phase 1 and a minimal cut set
occurs at phase 2. Y (2, t) is obtained as:

X (1, 0, t1e) X (1, 1, t3e)}]

(37)

Secondly, obtain the system accident occurrence

conditions using the basic event transformation
method (Esary & Ziehms, 1975). In this approach,
component i being in failure mode j during phase k
means that component i enters failure mode j at any
one previous phase l, l = 1, 2, . . ., k. Let F(i, j, k)
denote that component i get into failure mode j during
phase k. Compared with the proposed method, the system accident occurring at phases 1 and 2 are obtained
using this expression,
Y (1, t1e) =F(1, 2, 1) F(2, 2, 1)


Y (2, t2e) = X (1, 1, 1) X (2, 1, 1)

(38)

[{X (1, 2, 1) X (1, 2, 2)}

Y (2, t) = Y (1, t1e) {X (1, 1, t)} X (2, 1, t)}

{X (2, 2, 1) X (2, 2, 2)}]

(33)

The first negative term in Eq. (33) denotes that no

system accidents occur during phase 1, and tke
denotes the end time of phase k. Substituting Eq. (32)
into Eq. (33) and using De Morgans law, the

(39)

Thus, as the number of phases increases, the system accident condition becomes complicated. On the
other hand, the proposed method need not increases
the number of basic events.

1039

3.4

System accident occurrence probability

In this section, the transient probability of the system

accident occurring from the initiation of the mission
to phase 2 is analyzed for simplicity.
Assume the following constant failure rate for each
failure mode of pumps 1 and 2, and they are the same
for pumps 1 and 2, as follows:
(1, 1) = (2, 1) = (1)

(40)

(1, 2) = (2, 2) = (2)

(41)

From Eqs. (27)(29), component state probabilities

are obtained as:
Q(i, 0, t) = exp [ {(1) + (2)} t]
Q(i, j, t) =

(43)

2(2)
[1 exp({(1) + (2)}t)]
(1) + (2)

(2)2
[1 exp({(1)
((1) + (2))2
+ (2)} t)]2

(44)

Similarly, system accident occurrence probability

QS (2, t) can be obtained as:

+ (2)}t)]

(45)

Eqs. (44) and (45) show that the system accident probability increases during each phase as time passes. Since
QS (k, t) is a cumulative probability of the system accident during phase k, the cumulative value changes at
the transition from phase 1 to phase 2. Since the start of
phase 2 corresponds to the end of phase 1, the increase
of accident occurrence probability at the start of phase
2 can be obtained as:
QS (2, t1e) =

CONCLUSIONS

This paper proposes a simple novel method for failure analysis of the phased-mission system composed
of components with multi-failure modes. To represent the state transition of component clearly, the
disjoint binary variables are introduced, which can
simplify the logical operation of state variables at the
different phase. Further, failure probability for each
failure mode can be obtained based on the transition rate for each failure mode. Illustrative example
shows the characteristics of the phased-mission system where the latent failure condition lead to a system
failure at the start of a new phase. The failure analysis
system based on the proposed method is now under
development.
Part of this research was performed under the
support of Grant-in-Aid for Scientific Research (C)
19510170.
REFERENCES

(1)2
[1 exp({(1)
QS (2, t) =
((1) + (2))2
2

(47)

This probability equals to the probability that both

components are in failure mode 1 at the end of phase 1,
which leads to a system accident at the start of phase 2.
This implies that {no system failure at phase 1} does
not always guarantee {no failure at the start of phase
2} due to the change of system structure. The latent
system failure will appear at the start of a new phase.
This property is typical for the phased-mission system.
4

During phase 1, system accident occurrence probability QS (1, t) can be obtained from Eq. (32).
QS (1, t) =

QS (2, t1e) ((1)t1e)2

(42)

(j)
[1 exp({(1)
(1) + (2)
+ (2)}t)]

Assuming the approximation: 1- exp(t) t, the

above probability can be obtained as:

Burdick, G.R. Fussell, J.B. Rasmuson, D.M. & Wilson, J.R,

1977, Phased-mission analysis: A review of new developments and an application, IEEE Trans. Reliability, 26(1),
4349.
Dugan, J.B., 1991, Automated analysis of phased-mission
reliability, IEEE Trans. Reliability, 40(1), 4551.
Esary, J.D. & Ziehms, H., 1975, Reliability analysis
of phased-missions, Reliability and fault tree analysis,
SIAM, Philadelphia, 213236.
Hoyland, A. & Rausand, M., 1994, System reliability theory,
models and statistical methods, John Wiley & Sons, INC.
Mura, I. & Bondavalli, A., 1999, Hierarchical modeling &
evaluation of phased-mission systems, IEEE Trans. Reliability, 48(4), 4551

(1)2
((1) + (2))2
[1 exp ( {(1) + (2)} t1e)]2
(46)

1040

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Added value in fault tree analyses

Tommy Norberg
Department of Mathematical Sciences, Chalmers University of Technology, Gteborg, Sweden
Department of Mathematical Sciences, University of Gothenburg, Gteborg, Sweden

Lars Rosn & Andreas Lindhe

Department of Civil and Environmental Engineering, Chalmers University of Technology, Gteborg, Sweden

ABSTRACT: It is recognized that the usual output of a fault tree analysis in some studies is not sufficiently
informative. For added value in a widely used instrument for doing risk analyses, a Markovian approach is
suggested. It is shown how to extend the calculations of the standard fault tree gates, so that information is
available not only on the failure probability at the top level, but also on estimates of its rate of failure and mean
down time. In applying this to an integrated fault tree analysis of a municipal drinking water system, we further
identified the need for gates that are variations of those currently in use.

INTRODUCTION

A major purpose of a fault tree is to assist in the

calculation of the probability of failure, P(F) say, for
systems built by simpler subsystems or components.
A fault tree is constructed deductively until a suitable
level of detail is reached. Input to the fault tree calculations are probabilities P(Fi ) for base events Fi ,
where Fi denotes failure of subsystem or component i.
Often it is assumed, and so will we, that the base
events are independent. If for any i, Fi represents the
event that subsystem i is down at a particular point
in time, then P(F) is the probability that the system
is down at that particular time point. Moreover, in the
stationary case, P(F) as well as all P(Fi ) do not depend
on time.
Assuming ergodicity, P(F) can be thought of as the
ratio between the Mean Down Time (MDT) and the
Mean Time Between Failures (MTBF),
P(F) =

MDT
MTBF

where
MTBF = MTTF + MDT
and MTTF is short for Mean Time To Failure. This fact
tells us that the probability P(F) is not particularly
informative, since two systems with very different
dynamic behavior can share the same P(F).

It is the purpose of this paper to suggest a way to

extend the fault tree calculations so that also estimates
of MTTF and MDT are calculated at the top level
as well as at each intermediate level. This of course
requires knowledge of all MTTFs and all MDTs at the
base level. Estimates are then recursively calculated at
each level of the tree. These estimates are calculated
under the assumption that the input processes are independent and Markovian with only two states, up and
down. That is, the failure rate is assumed to be constant and equal to 1/MTTF. A similar remark applies
to the rate at which the process recovers from failure.
At intermediate levels and at the top, the rates typically
are not constant. The calculations, however, assumes
constant input rates and yield constant output rates.
Thus there are errors that when propagated to the top
level may be quite substantial. We believe, however,
that in applications with gross parameter uncertainties often the main errors in the final result are due
to the errors in the assumed base parameter values. In
such cases the enlargement of the standard fault tree
calculations suggested in this paper will provide a considerable and valuable insight into the dynamics of the
system under study.
We commence in Section 2 by showing how this can
be done for the two basic types of gates in a fault tree.
We then extend the argumentation to two auxiliary
gates that we have found use for. Then, in Section 3, we
study a simple example to some extent. In Section 4 we
briefly discuss how our ideas were used in a fault tree
analysis of a large municipal drinking water system,

1041

see (Lindhe et al. 2008). It was this study that motivated the present work. Finally, in Section 5 we discuss
and draw some general conclusions.

FAULT TREES

The reader is referred to (Rausand and Hjland 2004,

Chapter 3.6) for a basic introduction to fault tree analysis in reliability; (Bedford and Cook 2001, Chapters
6, 7) discuss fault trees from a risk perspective. Typically a fault tree is a combination of two types of logic
gatesthe OR- and the AND-gateand the fault tree is
built by repeated application from the top event and
down until a suitable level of detail is obtained. We
will refer to a gate with top or output event F and base
events Fi as a subsystem consisting of components all
of which are either up or down at a particular point in
time. Below we will consider the two basic types of
gates and two variations of the AND-gate.
A subsystem comprised of an OR-gate is up (functioning)as long as every component is up. That is,
c
c
Fc =
i Fi (A is the logical complement of A).
Hence,
F=

Fi

(1)

Similarly, a subsystem comprised of an AND-gate is up

as longas at least one of its components is up. That is,
F c = i Fic or
F=

Fi

(2)

Cf. Figures 1 and 2. The abbreviations in these figures

will be explained in the example of Section 3 below.
If the base events are independent (which will be
assumed throughout), then (1) and (2) imply

Figure 2. An AND-gate with base events FD1 , FD2 and

output event FDP = FD1 FD2 .

P(F) = 1
P(F) =

(1 P(Fi ))

(3)

P(Fi )

(4)

respectively. These are the basic fault tree calculations

that will be enlarged below.
2.1 The OR-gate
We first study the OR-gate, which is defined by formulae (1) and (3). In reliability, the OR-gate corresponds
to a series structure. See e.g. (Rausand and Hjland
2004).
Replace each base event of the gate by a Markov
Process (MP) having two states, up 1 and down 0. Let
i and 1/i denote the failure rates and the mean down
times, respectively, for these base MPs; (Rausand and
Hjland 2004) contains an introduction to continuous
time Markov chains in reliability. Clearly,
P(Fi ) =

i
i + i

so, by (3),
P(F) = 1


i

i
i + i

(5)

In Figure 3 the state diagram of the combined MP is

shown for the case of two base components. Suppose
the system is
up. Then
the time to the next failure is


exponential i i . The system failure rate is thus
Figure 1. An OR-gate with base events FPS , FSC , FD and
output event FAS = FPS FSC FD .

1042

(6)

Figure 3. State diagram of a Markov Processes representing

an OR-gate. The process is down (i.e., in failed state) if at least
one base process is down.

Let further 1/ be the system mean down time.

Then
P(F) =

(7)

By equaling (7) and (5), and using (6) while solving

for , we obtain


i i

= 
(
+

i)
i i
i i

(8)

We next study the standard AND-gate, the calculation

of which is given by (4); cf. also (2). In reliability, an
AND-gate corresponds to a parallel structure.
Again replace each base event by an MP with only
two states, up 1 and down 0. Let i and 1/i denote
the failure rates and the mean down times, respectively,
for the base MPs. Then, by (4),

i

i
i + i

(10)

Next, introduce a mean up time 1/ so that (7) holds

true. By equaling (9) and (7), and using (10) while
solving for , we obtain


The AND-gate

P(F) =

is the probability that the combined MP is down. In

Figure 4 the state diagram of the combined MP is
shown for the case with two base events.
Write 1/ for the mean down time of the combined
MP. Then, necessarily,

The formulae (6) and (8) enlarge the standard fault

tree calculations for the case of an OR-gate, so that
now we know not only how to calculate the probability
P(F) that the subsystem comprising the gate is in its
failed state, but also its failure rate and mean down
time 1/.
2.2

Figure 4. State diagram of a Markov Process representing

an AND-gate. The process is down when all base processes
are down.

(9)

i

(
+

) i i
i
i
i

(11)

The formulae (10) and (11) extend the standard fault

tree calculations for the case of an AND-gate, so that
now we know not only how to calculate the probability
P(F) that the subsystem comprising the gate fails, but
also its failure rate and mean down time 1/. (We
make here and in similar instances below the simplifying assumption that the failure rate is constant and
equals the reciprocal of the mean time to failure.)
2.3 A 1st variation of the AND-gate
Consider e.g. a subsystem comprising of a power generator and a battery back-up of limited life length and
such that it may fail on demand. The state diagram of
such a process is shown in Figure 5. The combined system is up as long as the main subsystem is up (state 1)
or down while at the same time the back-up is up (state

1043

01); it is down when in state 0; 1 and 1/1 denotes

the failure rate and mean down time of the subsystem
(i.e., power generator), q2 is the probability of failure
on demand for the back-up system and its up-time is
exponential(2 ).
The balance equations for this Markovian system
are

The rate at which the system recovers from its down

state 0 is 1 . Hence
= 1

(13)

Now may be calculated by inserting (13) in (7)

and equating the latter and (12).

p1 1 = p01 1 + p0 1
p01 (1 + 2 ) = p1 1 (1 q2 )

2.4 A 2nd variation of the AND-gate

p0 1 = p1 1 q2 + p01 2
where p1 and p01 are the stationary probabilities for
the system to be in either of its up-states, and p0 is the
probability that the system is in its down state. Solving
for p0 is straightforward. We get
p0 =

1 2 + q2 1
1 + 1 2 + 1

If the life length of the back-up is unlimited, i.e.,

2 = 0, this formula becomes
p0 =

1
q2
1 + 1

Consider next a subsystem comprising a power generator having another power generator as back-up. The
back-up may fail to start on demand. Some people
would refer to such a system with the phrase cold
stand-by. The state diagram of such a process is
shown in Figure 6. The down states are 0 and 00; 1 and
1/1 denotes the failure rate and mean down time of
the main power generator, q2 is the probability of failure on demand for the back-up generator while its upand down times are independent and exponential(2 )
and (2 ), respectively.
A straightforward analysis of the balance equations
shows that
p0 =

Similarly, one may prove that

p00 =

 i + qi 1
1
p0 =
1 + 1 i=1 i + 1

 i + qi 1
1
1 + 1 i=1 i + 1

1 (1 q2 )
2
1 + 1 2 + 1 + 2

(14)
(15)

Hence,

if the subsystem has several independent back-ups.

From the fact that P(F) = p0 , we now conclude
P(F) =

1 q2
1 + 1

P(F) =

1 2 + q2 (1 + 2 )
1 + 1 2 + 1 + 2

(16)

(12)

Figure 5. State diagram of a Markov Processes representing

the variant of the AND-gate studied in Section 2.3. The process
is down while in state 0.

Figure 6. State diagram of a Markov processes representing

the variant of the AND-gate studied in Section 2.4. The process
is down while in state 0 or 00.

1044

The down times are 1/1 and 1/(1 + 2 ), while

in states 0 and 00, respectively. Hence,
p00
1
1
p0
1
+
=
p0 + p00 1 + 2

p0 + p00 1

By solving for , we obtain

=

1 (1 + 2 )(2 + q2 (1 + 2 ))
q2 (2 + 1 + 2 )(1 + 2 ) + (1 q2 )2 1
(17)

Now follows readily from (16) and (2.4) by means

of (7).
It is clear that if 2 = 0, then this 2nd variant of the AND-gate coincides with the one discussed
in Section 2.3. The reason for a separate discussion
of the 1st variant is that it readily extends to cases
with several independent back-ups. This is not so for
the 2nd.
2.5

by construction always is true. This final remark

applies of course only to fault trees that only contain
the standard OR- and AND-gates.

Remarks on our enlarged fault tree

calculations

The calculations in a fault tree are from bottom to

top. The data (i.e., results) at one level form the
input to the calculations on the level immediately
above. Our approach presumes that the input data to
one level consist of Markovian down and up rates. The
calculated rates, however, are not necessarily Markovian. Still, the calculations at the next level presume
that they are. Clearly, this introduces an error making
it wrong to assume that rates calculated at the top or at
any intermediate level are Markovian. We claim, however, that it is in many instances reasonable to assume
that they are Markovian.
In other words, denote by the calculated rate of
failure at the, say, top level. Let T be a typical uptime and denote by (t) its failure rate. Then T is
not necessarily exponential() and it needs not to be
true that (t) = . However, in many instances is a
reasonably good first order constant approximation to
(t). That is,

EXAMPLE

An Alarm System (AS) consists of a Supervising

Computer (SC), its Power Supply (PS) and a Detector Package (DP). In order for the system to function
all these subsystems must. Thus they are coupled in
series and the appropriate fault tree gate is an OR-gate.
See Figure 1.
The detector package consists of two detectors (D1
and D2) that are separately battery powered and it is
regarded enough that one of them is functioning at
each particular time instant. Thus they are coupled in
parallel and the appropriate fault tree gate is an ANDgate. See Figure 2.
3.1 The power supply
We now direct our attention to the reliability of the
power supply, since it is naturally modeled by the 2nd
variation of the AND-gate discussed in Section 2.4. The
back-up generator does not always start on demand,
and that if so then it is not at all likely that it will
be repaired before the main power is up again. On
the other hand, if it does start, it may stop because of
lack of fuel and such stops are likely to be short. The
estimated values of the parameters of the Main Power
system (MP) and the Back-up power system (BP) is
shown in Figure 7. The unit of time is hours.

(t)
A similar remark applies to the mean down time
1/. Notice, however, that the probability P(F)
of failure is calculated exactly at each level and
that (7), i.e.,
P(F) =

Figure 7. Fault tree representation of the Power Supply (PS).

Its components are the Main Power (MP) and the Back-up
Power (BP) systems. The AND-gate is of the kind discussed
in Section 2.4. The time unit is hours.

1045

It is required that the long run unavailability of the

alarm system is less than 0.15%, and that its rate of
failure is less than 1 per month. A risk measure of
interest is the yearly frequency of down periods that
last for more than 12 hours. If this (risk) frequency
is considered to be too large, risk reduction measures
have to be implemented.
Thus, at the top level, the requirements are
100P(F) 0.15 and 8766 12. Our unit of time
is hours. Let Y denote a generic down time. Then
P(Y > y) ey . An estimate of the frequency of
long stops (LS), LS e12 , is required from the
analysis.
Using the formulae of Section 2.4, we get
100P(F) 0.14 0.15
8766 3.7 12
1/ 3.32
8766LS 0.1
(In the next section these values will be referred to as
being FT-calculated.)
A few remarks are in order.
1. The power supply system long run unavailability is
0.14%.
2. The back-up generator changes the failure rate from
2 failures per 1000 hours, to 0.4 which is well below
the stipulated 1.4 for the complete system.
3. The mean down time is 3.3 hours.
4. The frequency of long stops is about 1 per 10 year.
It is then recognized that there are gross uncertainties in the input parameters. Thus an uncertainty
analysis of the system is needed in order to see how it
is translated into the output parameters and especially
into the frequency LS of long stops.
After appropriately combining the available sparse
data with expert opinion, the uncertainty densities of
the parameters were specified. See Table 1.
The reason for choosing Gamma densities for the
Markov rates and are two-fold: (1) the Gamma
density is conjugate to the exponential (or Poisson)
density; and (2) our experience is that it is easier
for experts to state their opinion in terms of failure
Table 1. Specification of the uncertainty densities for the
parameters of the power supply system. The subscripts 1 and
2 refer to the main and the back-up system, respectively. From
left to right: the parameter, the median, the 90th percentile
and the particular choice of density.
1
1/1
2
1/2
q2

0.002
4.0
0.11
0.5
0.10

0.007
39.5
0.39
4.9
0.14

Gamma (0.9,0.0033)
1/Gamma (0.78,0.52)
Gamma (0.87,0.19)
1/Gamma (0.78,4.17)
Beta (11.2,97)

Table 2. Main facts on the uncertainty of the output

parameters of the power supply system. From left to right:
the parameter, estimations of the 10th percentile, the median
and the 90th percentile, the requirement if any and unit.
100P(F)
8766
1/
8766LS

0.01
0.39
0.84

0.12
3.3
2.96
0.04

1.8
16.8
25.5
2.7

0.15
12

%
1/year
hours
1/year

rates and mean down times, compared to in terms of

P(F) and either or . Notice also that the Beta density is conjugate to the Binomial. These conjugacy
facts imply that all uncertainty distributions are easily
updated with hard data if and when such become
available.
The calculations of the fault tree were repeated 106
times with simulated values of the input parameters.
The main facts for the uncertainty densities of the
output parameters are tabulated in Table 2.
Generally there are large uncertainties in the output parameters. This is of course a consequence of
the corresponding fact for the input parameters. Notice
that the medians typically do not coincide with the
calculated values.
Although not in focus here we remark furthermore that 53.6% of the simulations fulfilled both
100P(F) < 0.15 and 8766 12. That is, we are
about 50% certain that the power system simultaneously fulfills the requirements on P(F) and . This is
in accordance with the result of the first calculation.
3.2 Remarks on accuracy
As notice already it is not to be expected that the rates
and at the top level are Markovian. However, we
think that they are approximately so. We did compare the results from the fault tree modeling of the
power supply system with the results from one long
simulation of a Markov process having the state diagram in Figure 6. The following estimates of the output
parameters were obtained:
100P(F) 0.10
8766 6.59
1/ 1.4
8766LS 0.09
There are clear discrepancies between our MPsimulated results above and the FT-calculated ones of
the foregoing section. The FT-calculated rates and
are considerably lower then the MP-simulated rates,
which we consider to be quite near the true values.
The error in P(F) = /( + ) / is small in
the FT-calculation. Thus the errors in and are of

1046

comparable size and have the same sign. It follows

that a negative error in tends to cancel out a positive
error in the probability P(Y > 12) = e12 . Indeed,
the FT-calculated value, 0.10, coincides nearly with
the MP-simulated estimate 0.09.
One long MP-simulation needs a considerable
amount of computing time (this one needed nearly
20 minutes on a two year old desk top using an
interpreting code), while a considerable amount of FTcalculations can be done with almost no computing
time. Thus analysing the effect of parameter uncertainty with, say 104 simulated sets of base parameter
values, is probably not feasible with todays computing
power. In that perspective and recognizing the fact that
quite often the parameter uncertainty is large in FTcalculations, we argue that the in Section 2 suggested
enlargement of the FT-calculations are of a considerable value and give much insight into the dynamic
behavior of the system under investigation.
4

APPLICATION

We now briefly discuss how the above gate constructions were used to analyse a municipal drinking water
system (Lindhe et al. 2008).
The drinking water system was analyzed based on
an integrated approach. This means that the entire system, from source to tap, was considered. The main
reasons for an integrated approach are: (1) interactions between events exist, i.e., chains of events have
to be considered; and (2) failure in one part of the system may be compensated for by other parts, i.e., the
system has an inherent ability to prevent failure.
The top event, supply failure, was defined as
including situations when no water is delivered to the
consumer, quantity failure, as well as situations when
water is delivered, but it does not comply with existing water-quality standards, quality failure. The fault
tree was constructed so that the top event may occur
due to failure in any of the three main subsystems
raw water, treatment or distribution. Within each subsystem quantity or quality failures may occur. Note
that, for example, quantity failure in the treatment
may occur due to the water utility detects unacceptable
water quality and decides to stop the delivery.
To consider the fact that failures in one part of
the system may be compensated for by other parts,
cold stand-by AND-gates (cf. Sections 2.3 and 2.4)
were used. For example, if the supply of raw water
to the treatment plant is interrupted, this may be compensated for by the drinking water reservoirs in the
treatment plant and distribution network. Since only a
limited amount of water is stored the ability to compensate is limited in time. Similar unacceptable raw water
quality may be compensated for in the treatment. In
the case of compensation by means of reservoirs the

1st variation of the AND-gate was used. The probability

of failure on demand represents the probability
The fault tree of the drinking water system was constructed so that failure in one subsystem cause the top
event to occur only if subsequent subsystem are not
able to compensate for failure. The possibility to not
only estimate the probability of failure, but also the
failure rate and mean down time at each intermediate
level provided valuable information on the systems
dynamic behavior. Two subsystems may have the same
probability of failure, but one may have a low failure
rate and long down time while the other have a high
failure rate and a short down time. Properties like these
are important to know about when analyzing a system
and suggesting risk reduction measures. Furthermore,
the cold stand-by AND-gates made it possible to include
the systems inherent ability to compensate for failures, thus the system could be modeled in an adequate
manner.
We refer the reader to (Lindhe et al. 2008) for a
more detailed description of how the new formulae for
doing fault tree calculations were used in a practical
application.
5

DISCUSSION

A major advantage with incorporating Markov Processes into the fault tree calculations is the added
dimensionality of the output as can be seen already
in simple examples such as the one in Section 3.
The application to a municipal drinking water system (cf. Section 4) also showed that the variations of
the AND-gate were of great importance. The information on failure rates and down times for each
intermediate level provided valuable information on
the systems dynamic behavior.
It is moreover worth mentioning here that many
experts prefer to elicit their thoughts on dynamic systems in terms of percentiles for failure rates and mean
down times, making it fairly easy to define uncertainty distributions that in the future may be updated
by hard data.
We finally emphasize the fact that the resulting
binary process at the top level typically is not Markov.
Thus, it is wrong to think of and at the top level as
being parameters of an MP, although we do not hesitate
to do so approximately. For instance, in the example
of Section 3, we approximated the down time distribution using an exponential() density. This is also
implicit in our referring to as the system failure rate.
ACKNOWLEDGMENTS
All authors acknowledge support from the Framework
Programme for Drinking Water Research at Chalmers

1047

(DRICKS), and from the EU Sixth Framework project

Techneau (www.techneau.org).
REFERENCES
Beford, T. and R. Cooke (2001). Probabilistic Risk Analysis:
Foundatation and methods. Cambridge.

Lindhe, A., L. Rosn, T. Norberg, O. Bergstedt, T. Pettersson,

J. strm, and M. Bondelid (2008). Integrated probabilistic risk analysis of a drinking-water system: A fault-tree
analysis. In preparation.
Rausand, M. and A. Hjland (2004). System Reliability
Theory. Model, Statistical Methods, and Applications.
Wiley.

1048

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Alarm prioritization at plant design stageA simplified approach

P. Barbarini, G. Franzoni & E. Kulot
Design HSE Department, Tecnimont S.p.A.

ABSTRACT: Alarm prioritisation helps operators in keeping the full control of the plant and in performing
the appropriate corrective actions while in emergency, when a possible flood of alarms is likely to obscure the
real source of troubles with irrelevant information. Alarm prioritisation is a very complex, time consuming and
iterative process that requires detailed information about the plant and its operations, such as: time available
for taking corrective actions, number of operators and their preparedness in recognizing and facing emergencies, availability of written procedures to be applied, likelihood of simultaneous occurrence of multiple alarms
reducing the operator response time. Most of this information is normally not available at the plant design
stage. Notwithstanding this, a basis for an alarm prioritisation in accordance with the EEMUA guidelines can be
established at the plant design phase, by starting from the results of the safety analysis carried over on all plant
subsystems, and by applying a set of prioritisation general rules. An example of prioritisation in accordance
with EEMUA guidelines is provided for a Petrochemical Plant currently under construction. The results of the
preliminary prioritisation can be the basis for a complete prioritisation process to be performed at a later stage.

INTRODUCTION

Due to the great development of electronic applications, new plants are much more instrumented than in
the past, and a very large number of diagnostics are
usually provided in a plant.
Additionally the wide use of well-established systematic analysis techniques (e.g., FMEA, HAZOP)
allows the identification of the safeguards to be implemented in the design of the plant in order to prevent
or mitigate the consequences of potential accident
scenarios.
Typically, these safeguards include the provision of
new instrumentation for a wider monitoring and control of plant process parameters, as well as the implementation of a number of additional plant alarms.
However, in the absence of prioritisation, it is largely questionable that these additional monitoring and
alarms allow the operators a better and timely reaction
during a major plant upset. In fact, when a plant upset
occurs, a flood of alarms is likely to obscure the
real sources of trouble and to significantly increase the
stress level of operators, thus reducing the probability
that the proper corrective actions are taken.
There is a general consensus that some major accidents occurred in the past years to petrochemical and
nuclear plants were exacerbated by wrong or delayed
decisions, which were taken by operators because
overwhelmed by irrelevant or misleading information.
Typical examples of these case are the LOCA (Loss
of Coolant Accident) occurred at Three Mile Island in

1979, [1], and the explosion of the Texaco Refinery at

Milford Haven in 1994, [2].
Due to wide use of sensors and diagnostics, the
number of alarm messages to be managed by control
room operators is usually very high. According to the
results of a recent study, [3], the average number of
alarms per day expected on an operator station in a
European refinery is about 14000, with peaks up to
26000 alarms per day, while the number of alarms actually seen by operators is estimated to be approximately
one half.
These figures are by far greater than the number of alarms per day that the Guidelines established
by the Engineering Equipment and Materials Users
Association (EEMUA), [4], consider as very likely
to be acceptable (150 alarms per day) or at least
manageable (300 alarms per day).
Accordingly the Alarm System has to be designed
so that the presentation of alarm information allows
operators to easily recognize where alarms are occurring in the plant and what their importance is.
2

ALARM PRIORITISATION

At the time when the preliminary prioritisation has

to be finalized, results of safety studies are being
implemented in the project, minor design changes are
underway, detailed engineering of some packages is
in progress, so that even the alarm list is subjected to
modifications.

1049

Neverthless, the results of the safety Layers of

Protection Analysis (LOPA) studies performed for all
the Plant Units are available. These studies, [5], provide a complete overview of all the expected accident
and/or relevant plant upset scenarios, a measure of
their magnitude (expressed as LOPA target factors),
and a list all the corrective actions (the layer of protections) necessary to avoid their occurrence or to
mitigate their consequences: this information provided
in LOPA studies can be used for a complete preliminary alarm prioritisation on the basis of the EEMUA
Guidelines and a set of assumptions described in the
following.
The prioritisation process consists of the following
steps:
description of the alarm system and characterization
of plant alarms into groups;
definition of priority levels;
definition of the prioritisation criteria;
assignment of the priority level to each alarm.
2.1

Plant alarms characterization

For the reference Petrochemical Project, the following

are main components of the Alarm System, entirely
accessible from the operators consoles installed in the
Main Control Room:
DCS process alarms;
Alarms deriving from other automated systems
(e.g., Machine Monitoring Protection System);
Safety Instrumented Systems alarms processed by
the ESD.
The overall number of DCS, ESD and other systems alarms estimated for the plant is about 8550.
This figure includes about 1800 process alarms, 1000
pre-alarms (450 of ESD, 550 of other systems), 800
interlock activation alarms (250 of ESD, 550 of other
systems), 900 devices malfunction alarms at ESD
(transmitters, switches, on/off valves), 3900 devices
malfunction alarms at DCS (motors, on/off MOV
and control valves, transmitters), and about 150 system general fault alarms (ESD, DCS, other automated
systems).
2.2

Definition of priority levels

Priority levels are intended to provide a guide for determining in which order the operators have to respond.
Following the recommendations provided by EEMUA,
four different priority levels have been defined for the
Project. They are the following:
Emergency Priority Level (1): To this level belong
all alarms indicating a situation for which an operator action is immediately required in order to
prevent immediate hazardous conditions which can

turn, if no operator action is performed, into severe

consequences, such as injury to people, releases to
the environment, damages to critical items or major
production losses. This category includes typically
alarms indicating that is required an operator action,
which is part of a safety related function.
High Priority Level (2): To this level belong all
alarms indicating a situation for which an operator action is immediately required in order to avoid
major process deviations leading, if no operator
action is performed, to major plant upsets. Alarms
indicating faults of the systems dedicated to prevent the occurrence of dangerous situations (ESD)
belong to this level. Alarms indicating an ESD
intervention requiring an operator action in order
to prevent subsequent (cascaded) ESD intervention
belong to this level too.
Medium Priority Level (3): To this level belong all
alarms indicating a situation for which an operator
action is timely required to avoid process deviations
which can lead, if no operator action is performed, to
process minor upsets. DCS alarms, as well as alarms
indicating trips due to the intervention of ESD or
other systems, belong to this level.
Low Priority Level (4): This level encompasses all
the alarms indicating process deviations that do not
directly cause plant significant upsets. Alarms due
to devices general malfunctions belong to this level.

2.3

Definition of prioritisation criteria

The EEMUA Guidelines, [4], require that the alarm

priority depends both on the severity of the expected
consequences (when no operator action is performed)
and the time available to perform the required action.
The greater are the expected consequences and the
shorter is the time available to operators to respond,
the higher is the priority to be assigned to the
alarm(s) identified to indicate the occurrence of a
given upset scenario. A consistent alarm prioritisation
requires therefore the assessment of these two variables for each alarm. As far as the assessment of consequences is concerned, the same criteria adopted for
the all LOPA studies have to be applied to all the plant
alarms. Tab. 1 shows the correspondence between the
expected consequences (in terms of impact on people and environment) and the equivalent LOPA target
factor. An equivalent scheme (not shown) has been
adopted to make a correspondence between economic
losses and the equivalent target factors.
A systematic LOPA analysis has been carried out
over all the plant units, identifying a total of 284 significant accidents or major plant upset scenarios with
a LOPA target factor in the range 15.
Based on these results, an Impact Assessment Table
has been prepared, i.e., a correspondence between the

1050

Table 1.

Correspondence between expected consequences and equivalent LOPA target factors.

Consequence specific safety target factor table

Impact on people
Target factor

On-site

Minor injury

Serious irreversible injury

4
5
6

Fatality
Multiple fatalities

7
8

Table 2.

Off-site

Site contamination,
recovery within 1 year
Local public needs to
take shelters indoors
Public evacuation
Irreversible injury
in the public
Fatality in the public
Multiple fatalities
in the public

Catastrophic event
with many fatalities

Impact assessment.

Table 4.

Expected consequence
Minimal
Equivalent
LOPA target
factor

Table 3.

01

Minor
23

Ground water Contamination

Need for site clean-up

Alarm prioritisation matrix.

Consequence severity

Major
4

Severe
5

Operator
response

Minimal

Minor

Major

Severe

Delayed
Slow
Fast

Low
Low
Low

Medium
High

High
Emergency

Emergency
Emergency

Time response.

based on the Impact Assessment Table and the Time

Response Table.

Operator response
Fast
Slow
Delayed

Impact on environment

Operator action immediately required

Operator action timely required
Operator action not aimed at preventing the
direct occurrence of a plant upset

expected consequence and the equivalent target LOPA

target factor (See Table 2).
For the operators response following an alarm, the
following Time Response Table is used (See Table 3).
At this stage of the project, a precise assessment of
the effective time available to operators to respond to
each alarm is not feasible: as a first tentative assessment one can think that a fast response should require
an operator action to be performed within a few minutes, while a slow response is for those alarms
requiring an action to be performed in a longer time
(indicatively within 1015 minutes).
Finally an Alarm Prioritisation Matrix consistent
with the EEMUA guidelines is defined (Table 4),

2.4 Assignment of priority levels to alarms

2.4.1 Process alarms
For all the process alarms, the assignment of the
priority levels is made in accordance with the proposed Alarm Prioritization Matrix and on the basis of
the results of the LOPA studies. LOPA studies systematically describe all the expected accident and/or
relevant plant upset scenarios, starting from the relevant Initiating Events that have been identified. For
each scenario, the LOPA target factor and the layer
of protections are systematically identified. Layers of
protections include the automatic actions performed
by ESD and those required to the plant operators.
Regarding the sub-group of alarms listed in LOPA
studies, they are typically High/Low alarms of process
parameters, either requiring an operator action (e.g.,
a low flow-rate alarm requiring the start of a standby pump) or forerunner of a subsequent (cascaded)
intervention of an automated system (e.g., a high

1051

temperature alarm, forerunner of the corresponding

high-high temperature alarm requiring ESD to intervene): for each of these alarms, on the basis of Table 2,
the consequence severity is the one corresponding to
the LOPA target factor associated to the relevant scenario. For these alarms, the assessment of the time
available to operators to perform a corrective action
requiresin principlea detailed knowledge of the
plant dynamical behavior following any possible process deviation. For the purpose of this work, i.e., to
find a basis for a preliminary prioritization, it can be
assumed that alarms following pressure and flow deviations generally require a corrective action (if any) to
be done at most within a few minutes (fast operator response, as defined in Table 3), while alarms
following level and temperature deviations generally
allow a longer time for the operators to intervene
(slow operator response). This general assumption
is consistent with the fact that pressure/flow deviations, possible forerunner of plant upsets, propagate
quickly through a plant loop, while level and temperature deviations propagate more slowly, being plant
upsets caused by level deviations basically linked to
emptying/overfilling of vessels and columns, while
plant upsets caused by temperature deviations are basically linked to the thermal inertia of plant sections. It is
to be noted that the occurrence of temperature-driven
runaway reactions is not expected for the reference
Petrochemical Plant. Under these assumptions, and
on the basis of the Prioritization Matrix, it is possible
to associate to any alarm identified in LOPA studies
its proper priority level. This approach can be applied
to all process alarms.
LOPA studies systematically describe all the relevant accident and/or relevant plant upset scenarios
that are expected to occur in the Plant. Since, for
each scenario, all and only the corrective actions (and
the significant related alarms) required to prevent its
occurrence are identified, it is concluded that the process alarms not identified in LOPA studies can be
generally assumed as precursors of plant upsets whose
consequences are not significant, i.e., they correspond
to equivalent LOPA target factors equal or lower than
one. For these alarms, therefore, the associated consequence severity is Minimal (see Table 3), while
the operator response (see Table 4) is defined on
the basis of the physical variable monitored, as for
the sub-group of alarms listed in LOPA studies. The
associated priority level is assessed on the basis of the
Prioritization Matrix (see Table 4).
2.4.2 Other alarms
For all the other alarms, the priority level is defined
on the basis of the following few general rules.
1. An alarm indicating a system fault for ESD, DCS
or other systems is notby itselfforerunner of

an anomalous situation that can directly pose a

significant threat to the plant and/or the operators: nevertheless, the malfunction can turn into
the impossibility for the operators to timely detect
a dangerous situation and/or to intervene whenever
required. Being in any case a maintenance/repair
action required to timely restore these systems, it
is assumed that the priority level associated to the
alarms should be High for ESD system faults, while
the priority level associated to system fault alarms
of DCS or other systems should be slightly lower
(Medium); this assumption is consistent with the
fact that the plant safety instrumented functions,
required to prevent the occurrence of a major plant
upsets, are usually addressed to ESD only.
2. Whilst an ESD intervention alarm is the precursor of a major plant upset, which can require
subsequent (cascade) interventions of the ESD as
a safeguard to avoid significant consequences, an
ESD intervention alarm not triggering subsequent
ESD interventions, and/or for which an operator
corrective action is not feasible, in the worst case
is indicative of a potential plant upset for which the
required corrective actions have been performed.
For these cases, the function of the alarm is merely
to make operators aware of the intervention of such
systems, for the necessary recovery actions (e.g.,
re-start up of one or more plant sections), which
presumably will not require immediate action. It is
assumed therefore that the associated priority level
to these alarms should be Medium.
3. It is assumed that a device malfunction alarm is
notby itselfforerunner of a plant upset. Operators must nevertheless be aware of the fault, in
order to correctly address the maintenance/repair
actions required to restore the system. It is therefore assumed that the associated priority level to
these alarms should be Low for malfunction alarms
addressed to DCS, and slightly higher (Medium)
for those addressed to ESD.
2.5

Results

Based on criteria and assumptions above described,

a complete allocation has been made for all the Plant
alarms. Table 5 shows the preliminary distribution into
the assigned four priority levels.
The resulting preliminary distribution is consistent
with the EEMUA general guidelines, requiring that
the number of alarms should decrease as a function of
the Priority Level, i.e., the higher the Priority Level,
the lesser the number of alarms. This preliminary
result would be revised on the basis of the alarms effectively addressed in the LOPA studies, as well as on the
basis of the interlocks developed for the Plant, in order
to identify the interlock alarms that are precursors of
subsequent ESD interventions.

1052

Table 5.

Alarms preliminary distribution in priority levels.

Priority level

No of alarms

% Overall

Emergency
High
Medium
Low

55
308
1761
6414

0.6
3.6
20.6
75.2

2.6

Update of preliminary results following

design changes

During the project lifetime it is quite common that

design changes are introduced after the completion of
HAZOP and SIL/LOPA studies that are the basis for
the preliminary alarm prioritization. These changes
are managed by the implementation of a Management Of Change (MOC) Procedure that foresees to
carry out, for each design change, a preliminary
evaluation of the potential impact HSE aspects (preliminary screening). The objective of this screening is
to discard from subsequent analysis any design change
clearly having no or negligible HSE impact, while
any critical change i.e. any change having potential HSE impact is subjected to a deeper analysis,
adequate to the extent of the expected impact. Typical examples of critical design changes are process
modifications leading to heavier process conditions in
streams of hazardous materials, modifications of containment conditions for hazardous materials (including modifications in seal arrangement for pumps
and compressors, deletion of block valves, change
of control valve failure positions), modifications of
process control philosophy and/or functionality (addition, deletion modification of process control and
ESD loops). For some critical design changes the
analysis required by the application of the MOC procedure may also imply re-Hazoping or reviewing the
results of the SIL/LOPA analysis. For these particular cases, the preliminary prioritisation is reviewed
and updated accordingly, by applying the same prioritization criteria above described. For the reference
Petrochemical Plant, it turned out that the impact of
design changes on the preliminary prioritization was
negligible.

be assigned different priority levels, so that alarms

related to higher potential consequences and/or lower
intervention time available to operators are associated
to higher priorities, depending on the information they
contain.
Four different priority levels (EmergencyHigh
MediumLow) are defined.
For all the process alarms, the assignment of the
priority level is made on the basis of a prioritization matrix, which considers either the consequence
severity of a plant upset following a process deviation when no corrective action is performed, either the
time available to operators to perform the corrective
action.
For all the process alarms, precursors of accident
and/or major plant upset scenarios following process
deviations, the severity of consequences is assessed
on the basis of equivalence with the LOPA target factors resulting from the LOPA analyses, while the time
available to operators are assessed on the basis of the
type of the physical parameter monitored by the relevant instrumentation. Design changes made after the
completion of HAZOP and SIL/LOPA reviews are
managed by the implementation of a Management
of Change Procedure that may imply re-Hazoping or
reviewing the results of the SIL/LOPA analysis. For
these particular cases, the preliminary prioritisation is
reviewed and updated accordingly.
This approach for the alarms categorization is
applied to the alarms identified in the LOPA analyses,
and it is also applied to all the other process alarms
by assuming that each alarm can be considered, following any process deviation, as a possible precursor
of a minor plant upset characterized by a lower LOPA
target factor.
Finally, the alarm categorization is applied also to
fault and malfunction alarms, on the basis of their
importance, i.e., assigning higher priorities on system faults alarms with respect to single devices
fault alarms, and assigning higher priorities to
alarms addressed to systems performing plant safety
functions (ESD) with respect to alarms addressed
to DCS.
The obtained results can be considered as the basis
for the final prioritization to be performed at a later
stage.

REFERENCES

CONCLUSIONS

For the reference Petrochemical Plant, a preliminary

allocation of the plant alarms into different priority level was performed, based on the information
available at the present design stage.
Prioritisation is based on the EEMUA general
guidelines, requiring that all the plant alarms have to

[1] United States Nuclear Regulatory Commission, Special Inquiry Group, Three Mile Island: A report to
the Commissioners and to the public, Washington
D.C. U.S. Government Printing Office, 1980 (Rogovin
Report)
[2] Health and Safety Executive, The explosion and
fires at the Texaco Refinery, Milford Haven 24 July

1053

1994: A report of the investigation by the Health and

Safety Executive into the explosion and fires on the
Pembroke Cracking Company Plant at the Texaco
Refinery, Milford Haven on 24 July 1994, ISBN 0 7176
1413 1, 1997
[3] P. Grosdidier, P. Connor, B. Hollifield, S. Kulkarni,
A Path Forward for DCS Alarm Management,
Hydrocarbon Processing, November 2003, Vol. 82,
No 11

[4] The Engineering Equipment and Materials Users

Association (EEMUA), publication no. 191 Alarm
SystemsA Guide to Design, Management and
procurement, ISBN 085931 076 0, 1999
[5] Layers of Protection Analysis: Simplified Process
Risk Assessment, Center for Chemical Process Safety,
American Institute of Chemical Engineers, New York,
NY, 2001, ISBN 0-8169-0811-7

1054

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Analysis of possibilities of timing dependencies modelingExample

of logistic support system
J. Magott, T. Nowakowski, P. Skrobanek & S. Werbinska
Wroclaw University of Technology, Wroclaw, Poland

ABSTRACT: In the paper we consider the time dependent system of systems where the system total task must
be executed during the constrained time resource. To model the delay times of operational system maintenance
process performance, there is used Fault Trees with Time Dependency (FTTD) analysis. In the FTTD, events
and gates are characterized by time parameters, which are derived from the time characteristics of system of
systems processes performance. The FTTD analysis for calculations of proper value of time parameters for
system of systems will be shown in this approach.

INTRODUCTION

The main objective of any operational system is to

perform continuously its intended functions. Failure
of any working element affects the operational process
and adds costs for spare parts, labour and maintenance
materials. In addition to the losses on these accounts,
there is also a loss of profit margin caused by system
downtime.
Sufficient inventory level of spare parts has to be
maintained to keep the system in operating condition.
When the spare elements are under stocked, it may lead
to extended system downtime and, as a consequence,
have a negative impact on the system availability level.
On the other hand, maintenance of excessive spares
can lead to large inventory holding costs. Moreover,
the requirements for planning the logistics of spare
elements necessary in maintenance actions of operational systems differ from those of other materials in
several ways: the service requirements are higher, the
effects of stock-outs may be financially remarkable,
the demand for parts may be sporadic and difficult
to forecast, and the prices of individual parts may be
very high.
Consequently, one of the most important decisions
faced by maintenance managers is the determination
of optimal stocking levels which involves finding the
answer to the following questions, such that the total
expected inventory costs are minimized:

base on a classic inventory theory, where the procurement process parameters are usually optimized
taking into account the cost constraints (see e.g.
Sarjusz-Wolski, 1998).
The surveys by Cho & Parlar (1991), Guide Jr &
Srivastava (1997), Kennedy et al. (2002), Nahmias
(1981), and Pierskalla & Voelker (1976) review the
growing body of literature on spare parts inventories
problem.
The stock level of spare parts is dependent on the
chosen maintenance policy. Therefore, maintenance
programs should be designed to reduce both, the maintenance and inventory related costs. As a result, the
problems of system performance optimization can be
solved by joint, rather than separate or sequential
approach to define the maintenance and inventory
policies.
Consequently, there is used a system of systems conception to model the interactions between
operational system and its logistic support system.
This conception is widely used in the defense sector,
where logistic processes are aimed at supporting the
warfighter/mission requirements.
According to the definition (Crossley, 2006),
the system of systems context arises when a need
or a set of needs are met with a mix of multiple
systems, each of which are capable of independent operation but must interact with each other
in order to provide a given capability. More
information can be found in (Werbinska, 2007a,
Werbinska in prep.).

When to (re)order?
How many items to (re)order?
There is a plethora of models in the literature
on reliability theory regarding to the spares supply
process optimization. A significant portion of them

Moreover, when designing such real-time systems like logistic support of operational system performance, the following time parameters: parts

1055

operational time, repair process time, time of spare

parts delivery, redundancy time, have to be analyzed.
In the paper, these times are characterized by minimal
and maximal values.
Taking into account the above considerations, in the
presented paper, hazard is such an event that Operational system fault lasts after time redundancy. The
mentioned time parameters and the hazard occurrence can be expressed in Fault Trees with Time
Dependencies (FTTD).
The FTTD technique is a top-down approach, which
starts with identifying hazards and their duration
time. In the FTTD, event duration times, delay times
between causes and events and the result events can
be described by the time intervals given by its minimal and the maximal lengths. Additionally, there are
generalization gates and causal gates. In the case of
generalization gates the effect is a combination of
causes or is one of the causes. The causal gates are
characterized by the delay time between the causes
and the effect.
Timing analysis of the system of systems using
FTTD gives such time relations that are necessary conditions in order to avoid the hazard. These results can
be applied in the system construction process.
As a result, the paper is aimed at implementation of
the presented FTTD method in analysis of time relations that occur in such a presented below system of
systems.
Following the introduction, the system of systems
model development is presented in Section 2. In the
next point, there is introduced the FTTD technique.
In Section 4, the FTTD for the system of systems
total task performance time and its analysis are given.
Finally, the work ends up with summary.

SYSTEM OF SYSTEMS WITH TIME

DEPENDENCY

In many systems undesired event occurs later than

components failure, if and only if repair is not completed within a grace period. In other words, time
redundant system has the ability to tolerate interruptions in their basic function for a specific period of time
without having the negative impact on the system task
performance.
Moreover, in the case of two or more independent
systems integration problem, time dependency is a
convenient approach used to integrate these systems
(Werbinska, in prep.).
This kind of time relations are especially characterized for logistic support systems, which performs
according to the Just in Time inventory strategy. Moreover, such problems investigations and developments
can be found in the defense sector.

Consider a repairable system of systems under

continuous monitoring, in which are integrated two
independent systems: a single-unit operational system
and its supporting system. Both systems have only
two states: up state, when they are operable and can
perform its specified functions, and down state, when
they are inoperable.
The system of systems total task is defined as
the continuous performance of operational processes.
Lets also assume that the logistic support functions are
narrowed down only to provide the necessary spare
parts during the operational systems maintenance
actions performance.
To model the system of systems performance, there
is used the individual time redundancy (Werbinska,
2007a). Thus, the system of systems reliability is
defined as its ability to correctly complete the task
during the corresponding time resource , which may
be randomly distributed (Werbinska, 2007a).
Consequently, there can be presented the system of
systems with time redundancy model (Fig. 1).
The following assumptions for the developed model
are made:

independency of performed processes,

perfect replacement policy,
reliable and instantaneous information flow,
lead-time independent of the quantity of spares
being ordered,
only regular orders of spare parts performed,
non-negligible replacement time of operational
units,
infinite planning horizon.
According to the scheme, after operational system
having failed, the information about its downstate is
immediately sent to the logistic support system. When
there is available spare element in the remaining stock,
the necessary one is sent to the operational system.
In this situation, the time of supply task performance,
denoted by is assumed to be negligible. When there
are no available spare parts, the time lasts from the
moment of failure till the new delivery arrival. Finally
the operational system is put back into service.
If there is restricted the system of systems total
task completion time, defined as the time of operational system recovery process, the system of systems
remains in upstate if this defined time will be shorter
than time resource. Otherwise, the system of systems
will fail and remain in downstate till the end of the
delivery process.
The chosen stock policy affects the possible periods
of time without spare parts and the way the system of
systems performs. The investigated inventory system
operates under an (s, Q) policy. A replenishment order
of size Q is triggered whenever the inventory level
(stock-on-hand) drops to the reorder level s (s 0)
at the demand epoch. Before the new delivery arrival,

1056

Figure 1. Time dependent system of systems model with a single-unit operable system, non-negligible replacement time,
and an (s, Q) inventory policy.

the operational system can only use the limited amount

of spare elements taken from the remaining stock.
Demand that is not immediately satisfied is backordered and filled when the new order arrives. After the
delivery, new elements are used according to system
demand until the stock level falls again to the critical
point. The time between two orders placing defines a
procurement cycle.
As a result, in the single procurement cycle the
system of systems may fail if:

in prep.). However, when designing time relations

between two independent systems which interact with
each other, the probabilistic analysis cannot simply describe the time dependencies occurred during
the system of systems tasks performance, and other
approaches are required. As a result, the novelty of
this study is to adopt the presented FTTD method to
calculate the time parameters of the described system
of systems.

time of supply task performance lasts longer than

the defined time resource (system of systems
downtime includes the lead-time from the moment
of over crossing the , and the time of operational
system recovery),
time of operational system recovery lasts longer
than the defined time resource (system of systems
downtime encompasses the remaining repair time
from the moment of over crossing).

The main measure, which may define the presented

system of systems performance level, is the probability of its downtime caused by over crossing the
defined time resource by the system total task performance. The analytical model of performance of the
presented system of systems with time dependency is
investigated in (Werbinska, 2008, Werbinska, 2007a,
Werbinska, 2007b). In addition, the results obtained
from simulation model are presented in (Werbinska,

FAULT TREES WITH TIME DEPENDENCY

One of the most often used technique for identifying

and classifying hazards, and calculating system reliability for both simple and complex systems is Fault
Tree Analysis (FTA). The analysis is started with
searching all the potential hazards (failure or an accident). For each of them the fault tree (FT) is being
constructed. Hazard is called the top event. FT evaluates all the contributing events that may ultimately
lead to the occurrence of the top event.
The FTA is often used in many industry sectors,
for examples nuclear, automotive, software, financial,
or medical. The FTA can be used to determine the
following: minimal cut sets of faults (minimal sets of
faults) that can cause a hazard, probabilities of the
hazard and faults occurrence, or hazard failure rate. In
the standard FTA, time dependencies between events
are not considered.

1057

In Fault Trees with Time Dependency (FTTD), the

time dependence between the cause and the hazard or
the other consequence is taken into account. Described
above relationship is useful, e.g. in constructing protections, recovery subsystem, or in evaluation of a fault
reaction time. The FTTD is a complementary approach
to the standard FTA and can be used to determine
(Grski et al. 1995, Grski 1994):
minimal cut sets with time dependencies that cause
a hazard,
fault reaction (exposure) time,
fault detection time,
hazard tolerance time if hazard can be tolerated.
In papers (Magott & Skrobanek 2002, Magott &
Skrobanek 2000, Skrobanek 2005) the method of
FTTD analysis which is based on a system of inequalities and equalities has been given.
The FTTD can be created by experts or/and using a
new version of timed state chart (Magott & Skrobanek
2006).
The FTTD can be expressed on a diagram. The
diagram is a graphical representation of the chain of
events, being built using events and logical gate configurations. The formal notation of an event is given
in Fig. 2, and gates notation is given in Fig. 3.
The time parameter pair < S , S > denotes, respectively, the minimal and the maximal duration time of
the event Event name.
The time parameter pair denoted as <sS , sS >
expresses the earliest, and the latest time instants of
start of the event Event name, and is obligatory only
for hazard. More information is given in (Magott &
Skrobanek 2002, Skrobanek 2005).

<

S
s

S
s

>

<

minimal and maximal duration times for the events,

minimal and maximal delay times between the
causes and effects for the causal gates.
In this paper for construction of the system of systems FTTD, we need only two types of gates: causal
XOR gate and generalization XOR gate.
3.1 Hazard
Graphical representation of the hazard is given in
Fig. 4.
In order to carry over an analysis, one needs to
choose a time instant as zero time. Let assume:
hs = 0

(1)

We start the analysis with calculations for the HAZARD, from formula (see (Magott & Skrobanek 2002)):

S
S

hs = (hs hs )
S
he = hs + he

S
he = hs + hs

(2)

where: <hs , hs > denotes the time interval in which

the hazard has to start (denotes, respectively, the earliest and the latest time moment); <he , he > denotes
the time interval in which the hazard has to end
(denotes, respectively, the earliest and the latest time
moment).
The details of the notations and formulas are given
in (Magott & Skrobanek 2002, Skrobanek 2005).
3.2 Causal XOR Gate

>

The causal XOR gates with one output event z and two
inputs events: x, y is given in Fig. 5.

Event name
Figure 2.

The time parameters of FTTD are:

The notation for an event.

<

S
hs

S
hs

>

<

S
he

S
he

>

Hazard
Figure 4.

<
<

S
d1 ,

S
d,

S
d1 >

S
d >

<
<

S
d2

S
d,

Notation for the hazard.

S
d >

S
d2 >

<

Figure 3. The notations for a) causal AND gate, b) causal

priority AND gate, c) generalization AND gate, c) causal XOR
gate, e) generalization XOR.

S
xe

Figure 5.

1058

S
xe

>

<

S
d1

S
d1

>

The causal XOR gate.

<

S
d2

S
d2

<

>
S
ye

S
ye

>

In this gate:

given by:


S
occur(z) (x (duration(x) d1
S
(zs) (xs) + Sd1 ))
(xs) + d1
S
(y (duration(y) d2

(3)

S
S
(ys) + d2
(zs) (ys) + d2
))
S
d1
,

xs = zs
xe = ze

(6a)


ys = zs ,
ye = ze ,

ys = zs
ye = ze

(6b)

S
S
S
d1
, (d2
, d2
)

where:
represent respectively the
minimal and maximal time delays between the occurrence of the cause x(y) and the effect z.
If we know in which time interval the event z can
start, e.g. <zs , zs >, we can make calculations of
the time intervals for events x and y. As a result,
<xs , xs >(<xe , xe >) denotes the time interval in
which event x must start (end) in order to cause the
hazard. The details of the calculations are given in
(Magott & Skrobanek 2002, Skrobanek 2005).
For causal XOR gates we have the following inequalitiesequalities system:

S
S
xe
dl
S
S
}, xs = zs dl
xs = zs min{dlS , xe

S
xe = zs , xe = xs + xe
(4a)

S
S
d2 ye
S
S
S
ys = zs min{d2
, ye
}, ys = zs d2

S
ye = zs , ye = ys + ye
(4b)

3.3

xs = zs ,
xe = ze ,

We start our analysis with hazard event identification and construction of the fault tree for the defined
top event. For the system of systems presented in
the section 2, we can describe hazard as operational
system fault lasts after time redundancy .
The FTTD for the investigated system of systems
with time dependency is given in Fig. 7. This FTTD is
described in (Magott et al. 2008).
Left branch of the FTTD concerns situations when
the operational system fault has occurred while there
were no available spare parts in the logistic support
system. Right branch of the FTTD concerns situations when the system of systems downtime occurs
due to over crossing the defined time resource by the
operational elements repair times.
The analysis begins with calculations for the hazard.
In the FTTD given in Fig. 7 we have two cases:
hazard is equal to event 2, therefore using formulae
(1), (2) we obtain:
2s = hs = 0

Generalization XOR Gate

2s = hs =

The generalization XOR gates with one output event z

and two inputs events: x, y is given in Fig. 6.
For this gate:

S
(hs

2e = he = hs +

(7)

S
hs
)

S
he

= (0 0) = 0

= hs +

S
2e

(8)

S
2e

= max{0, min + max{0, td min s_max }

occur(z) (occur(x) x = z)
(occur(y) y = z)

THE FTTD ANALYSIS FOR SYSTEM

OF SYSTEMS WITH TIME DEPENDENCY

max }
(5)

2e = he = hs +

S
he

(9)
= hs +

S
2e

S
2e

= max + td max s_min min

The event z is identical to one of the causes. The

equalities system for this generalization XOR gate is

(10)

That can be written as:

2<0, 0>< max{0, min + max{0, td min s_max }
max }, max + td max s_min min >

(11)

hazard is equal to event 3, what gives:

Figure 6.

3s = hs = 0

The generalization XOR gate.

1059

(12)

<0, 0>
< max{0,
max

<

min

min

+ max{0, t d min

+ t d max

max

s _ min

min

s _ max

max

},

>

< max{0,

<

>

<

+ max{0, t d min

min

max

+ t d max

s _ min

s _ max

min

max

min

max

>

},

},

max

<

s _ min

s _ max

>

max

>

>
< 0,0 >

<

min

min

< 0, >

where :

>

< t d min , t d max >

s _ min

= s

min

s _ max

= s

max

+ (s + 1) t min
+ (s + 1) t max

Figure 7. The fault tree with time dependencies for the investigated system of systems.

from formulae (1), (2)

(hazard = event 2)

2(equal event 1) < 0,0 >

< max{0, min + max{0, t d min
max

+ t d max

s _ min

min

s _ max

3(equal event 1) < 0,0 > < max{0,

Figure 8.

>
min

In the next step, we make calculations for the event 2

occurrences and consequently for the remaining gates
in the left branch of the FTTD.
Time intervals for start and end of the event 2 are
given in formula (11).
Let us now analyze the event 4 occurrence possibility. The event 4 can be interpreted as a sequence
of the following events: start of (s + 1)-th operational
system fault, waiting for delivery, and repair process.
Therefore, the minimal and maximal duration times
for the event 4 are:

from
formulae
(1), (2)
(hazard
max },
= event 3)

max

},

max

min

>

Result tree calculations for events 2 and 3.

S
S
3s = hs = (hs
hs
)

3e = he = hs +

S
he

= hs +

(13)
S
3e

= max{0, min max }

3e = he = hs +

S
he

= hs +

(14)

S
4e
= min + max{0, tdmin s_max },

(17)

S
4e

(18)

= max + td max s_min

S
3e

S
= 3e
= max min

(15)

That can be written as:

3<0, 0>< max{0, min max }, max min >

The gate number is equal to the number of its output

event.
Taking into account the time redundancy in maintenance processes performance, the time delays for
gate 2 are:

(16)
The results of the presented calculations are given
in Fig. 8.

S
d1
= min

(19)

S
d1

(20)

1060

= max

from formulae (1), (2)

(hazard = event 2)
2(equalevent 1)
< 0,0 > < max{0,

min

+ max{0, t d min

from formulae 4a
(for gate 2)
only if min max + t d max
4 < min{ max , max + t d max
from formulae 4a
(for gate 5)
only if min max + t d max

s _ max

max

+ t d max

max

3(equal event 1) < 0,0 > < max{0,

s _ min
s _ min

},

min

> < 0,

min

only if min
5 < min{ max ,
s _ min

and

max

max

+ t d max

s _ min

} min{

< min{

max

max

+ t d max

s _ min

},

max

min

+ t d max

max

max

},

> < 0,

min

min

t d _ max

s _ min

6 < min{

min

s _ min

7 < min{

max

min

s _ min

max

max

max

},

min

>
from
formulae 4a
(for gate 5)

>

+ t d max >

min

, t d max },

s _ max

only if

Figure 9.

},

from
formulae
(1), (2)
(hazard
s _ min
min >
= event 3)
},
max
max
min >
from
formulae 4a
s _ min >
(for gate 3)

B
> < min{ max , max }, >

Results of the analysis of the FTTD.

Therefore, from formula (4a) we can obtain:

min max + td max s_min

only if

(21)

only if
7<

and

= min{max , max + td max s_ min }

4s = 2s

S
d1

= min

4e = 2s = 0
4e = 4s +

S
4s

max

max

max

},

min

max

min

only if

max

> <

> < min{

max

max

, >

7<

max

min
max

max

}, >

max

min

max

><

max

, >

Time relations for the minimal cut set B.

(22)
The hazard can occur only if:

(23)
(24)

= min + max + td max s_min

(25)

The presented calculations given by formulae

(22) (25) can be written as:
4< min{max , max + td max s_min }, min >
<0, min + max + td max s_min >

min
max

Figure 10.

S
S
, 4e
}
4s = 2s min{d1

min

7 < min{
max

(26)

Similar reasoning can be performed to obtain the

time parameters for other gates in the FTTD.
The results of the FTTD analysis are given in Fig. 9.
First, the right branch is analyzed. The result of the
analysis is given as minimal cut set B (see Fig. 9).

min max

(27)

We also know, that minimal and maximal delay

time, respectively, between start of the event 7 and
start of the hazard is equal to min and min{max , max },
respectively. This knowledge can be used in e.g.
constructions of the alternative plan of supplies out.
According to the presented model of analysis, we
can describe time relations for the hazard occurrence
too. It can be useful e.g. at planning of alternative kind
of supply as insurances.
For the minimal cut set B we have, depending on
the values min , max , two cases of time relations for
the event 7, see Fig. 10.
Graphical representation of the time relations
between the start of the event 7 and the hazard is given
in Fig. 11. The hazard cannot start earlier than min ,
and later than max (see Fig. 11a) or max (see Fig. 11b)

1061

counted from the start of the event 5, which is equal

to the start of the event 7.
Let us now analyze the left branch of the presented
FTTD. The result of the analysis is given as minimal
cut set A (see Fig. 9).
The hazard can occur only if:

s_min td_max
(28)
min + s_min max + td max

The relation given in formula (28) is equivalent to:

min + (s + 1) tmin + s min max + td max

and represents the condition for hazard occurrence.

The time relations for the minimal cut set A are
given in Fig. 12.
Graphical representation of time relations between
the start of the event 6 and the start of the hazard is
given in Fig. 13.
From these relations we know, e.g. that the hazard
can occur not earlier than:

Therefore, if:
max + td max < min + s_min

(29)

min + s_min

then hazard does not occur.

a)

max

min{max , max + td max s_min }

max

+ min{s_max , td max }

max

(3.e) = (1.e)

min

(3 .s) = (1.s)

(7.s)
(5.s)
max

(7.e)

start of
hazard

max

start of hazard
min{

max

(3.e) = (1.e)

min

min{

(5.e)
(3 .s) = (1.s)

(7.s)
(5.s)

4 < min{

max

max

(7.e)

+ t d max

max

+ t d max

4<

max

min

>

< 0,

min

max

+ t d max

if
6<
<

(6.s)

max
max ,

Figure 12.

max

s _ min

},

max

+ t d max

s _ min

s _ max
min

min

min

> < 0,

>

min
s _ min

s _ min

>

+ t d max >
>

min

s _ min

s _ min

+ t d max >

(2.e)
= (1.e)

, t d max }

(4.e)

(2.s) = (1.s)

(4.s)

start of
event 4

min

if

4<

max

< 0,

min

if
,

s _ max

s _ min

(6.e)

Figure 13. Time relations for the system of systems hazard

event occurrence that result from the events from left branch
of the FTTD illustrated in Fig. 7.

s _ min

t d _ max

s _ max

if s _ max t d _ max
6 < max t d max ,
<

s _ min

Figure 11. Time relations for the system of systems hazard

event occurrence that result from the events from right branch
of the FTTD illustrated in Fig. 7.

max

max

min

max

if

(32)

counted from the start of the event 6. This knowledge

can be used e.g. at planning of the alternative express
supply.

(5.e)

b)

(31)

and not later than:

start of
hazard

max

(30)

max

t d max +
+

max

s _ max

max

max

+ t d max

+ t d max

s _ min

+ t d max

s _ min

s _ min

min

>

s _ min

>

t d _ max

6<

s _ min

max

t d max

s _ max

<

s _ min

max

t d max ,

min

if

s _ max

s _ min

max

t d max

<

s _ min

max

t d max ,

1062

min
s _ min

s _ min

t d max ,
min

>

+ t d max >

t d _ max

6<

Time relations for the minimal cut set A.

>

>

min

s _ min

s _ min

+ t d max >

For example, if the maximal time of express

supply is tes_ max then (using formula (31)) it must
start not later than min + s_min (tes_max + max )
counting from the start of the event 6.
5

CONCLUSIONS

If the conditions (27) and (28) are not satisfied then

the hazard will not occur.
These conditions can be taken into account during
the system of systems design process. For example,
in order to avoid the hazard occurrence, the following
aspects can be taken into account:
selection of suppliers in terms of the required
delivery time,
reliability of operational system (e.g. in terms of
achieved times between failures),
maintainability of operational system (e.g. in terms
of required repair times),
minimal CIL quantity appraisement,
definition of minimal redundancy time.
If the conditions (27) or/and (28) are satisfied, we
can plan any alternative insurances using knowledge
of the time relations between the hazard and its causes.
The knowledge obtained in FTTD analysis allows
proper construction of the system of systems that
satisfies the time safety requirements.
Now there is such a tool that can execute the analysis of FTTD with five types of gates: causal XOR,
causal AND, generalization XOR, generalization AND,
causal priority AND ones for defined values of time
parameters. In the paper symbolic analysis has been
performed. It is planned to prepare the tool for this
symbolic analysis.
REFERENCES
Cho, I.D. & Parlar, M. 1991. A survey of maintenance models
for multi-unit systems. European Journal of Operational,
Research 51, 123.
Grski, J., Magott, J. & Wardzinski, A. 1995. Modelling
Fault Trees Using Petri Nets. In Proc. SAFECOMP95,
Belgirate, Italy, LNCS, Springer-Verlag.
Grski, J. 1994. Extending Safety Analysis Techniques With
Formal Semantics. in: Technology and Assessment of

Safety-Critical Systems. (eds.) Redmill, F.J. & Anderson,

T. Springer-Verlag, 147163.
Guide Jr, V.D.R. & Srivastava, R. 1997. Repairable inventory
theory: Models and applications. European Journal of
Operational Research, Vol. 102, Issue 1, 120.
Kennedy, W.J., Patterson, J.W. & Fredendall, L.D. 2002. An
overview of recent literature on spare parts inventories.
International Journal of Production Economics, Vol. 76,
Issue 2, 201215.
Magott, J., Nowakowski, T., Skrobanek, P. & Werbinska, S.
2008. Analysis of logistic support system using Fault
Trees with Time Dependencies, Archives of Transport,
2007, No. 4, (in print).
Magott, J. & Skrobanek, P. 2006. Partially automatic generation of fault-trees with time dependencies. In Proc.
Dependability of Computer Systems, DepCoS 06, Szklarska Poreba, Poland, IEEE Computer Society Press,
4350.
Magott, J. & Skrobanek, P. 2002. Method of Time Petri
Net Analysis for Analysis of Fault Trees with Time
Dependencies. IEE ProceedingsComputers and Digital
Techniques, Vol. 149, No. 6, 257271.
Magott, J. & Skrobanek, P. 2000. A method of analysis of fault
trees with time dependencies. In Proc. SAFECOMP2000,
Rotterdam, The Netherlands, LNCS, Vol. 1943, SpringerVerlag, 2000, 176186.
Nahmias, S. 1981. Approximation techniques for several
stochastic inventory models. Computer and Operations
Research, Vol. 8, No. 3.
Pierskalla, W.P. & Voelker, J.A. 1976. A survey of maintenance models: the control and surveillance of deteriorating systems. Naval Research Logistics Quarterly 23,
353388.
Sarjusz-Wolski, Z. 1998. Strategy of supply management.
Practice of business logistics. (in Polish). A.W. PLACET,
Warsaw.
Skrobanek, P. 2005. A method of analysis of fault tree with
time dependencies for safety-related systems (in Polish),
Ph. D. Thesis, Technical University of Wroclaw, Poland,
report: PRE. 24/2005.
Werbinska, S. 2008. Problems of availability maintenance of
logistic support system with time dependency (in Polish).
In Proc. of Winter School of Reliability, WITE Radom.
Werbinska, S. 2007a. The reliability model of logistic support
system with Time Dependency (in Polish). Logistics No.
3, 6673.
Werbinska, S. 2007b. The availability model of logistic
support system with time redundancy. Maintenance and
Reliability, No. 3, 2329.
Werbinska S. Model of logistic support system with time
dependency (in prep.).

1063

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Applications of supply process reliability model

A. Jodejko & T. Nowakowski
Wroclaw University of Technology

ABSTRACT: According to the definition of the Council of Logistics Management, logistics is that part of
the supply chain process that plans, implements and controls the efficient, effective flow and storage of goods,
services and related information from the point of origin to the point of consumption in order to meet customers
requirements. Problems of logistic system reliability assessment were discussed in paper (Nowakowski 2006).
The major questions that are discussed concern definition of the logistic system failure and estimation of
simple measures of the system reliability. Reliability assessment of distribution process was discussed in paper
(Nowakowski 2007). This paper deals with two problems: Developing more adequate reliability measure that
the proposed earlier and assessing reliability of supply/procurement process which is critical for operation
system.

INTRODUCTION

The logistics is defined as the part of the supply chain

process that plans, implements and controls the efficient, effective flow and storage of goods, services
and related information from the point of origin to
the point of consumption in order to meet customers
requirements (Ballou 2004). Logistics exists in two
general forms, which are similar: military and managerial logistics. The first one (defense approach)
is concerned with movement of goods and people,
while the second one concentrates primarily on finished products. Both of them are the part of some
supply chain (SC), which encompasses all activities
associated with flow and transformation of goods
from the raw materials stage through to the end
user, as well as the associated information flow
(Nowakowski 2006). Thus the efficiency of a supply chain depends on its logistics system. This fact
induced researchers to look for methods of logistic
process assessment.
One of recently developed concept, important from
business and military point of view, is reliability of
logistic system. The problem was analyzed in few
papers (Nowakowski 2004, 2005, 2006). The major
questions that were discussed, concern definition of
the logistic system failure (what does it mean that
logistic system is reliable or failed?) and estimation of
simple measures of the system reliability (how create
a simple measure of the system reliability?). Further
paper (Nowakowski 2007) discussed reliability assessment of distribution process and the multidimensional

model of the process reliability was proposed. This

paper includes the extension of proposed ideas and
deals with two main problems:
developing more adequate reliability measure that
the proposed earlier,
assessing reliability of supply/procurement process
which is critical for operation system maintenance
(Jodejko 2007).
2

RELIABILITY MEASURES

Reliability of logistic system is defined differently

from military and business-orientated point of view.
According to the glossary published by Defense Systems Management College (DCMS 2006), logistics
reliability recognizes the effects of occurrences that
place a demand on the logistics support structure
without regard to the effect on mission or function.
Commercial approach consider, that correct
functioning of logistics system may be defined by 7R
formula (Ballou 2004): Right product, Right quantity,
Right quality, Right place, Right time, Right customer,
Right price.
Logistic system unreliability may be considered
with the respect of two kinds of consequences. Some
failures influence costs of SC (e.g. a company incurs
additional expenses resulting from logistics system
mistakes). Some faults of the system may be critical
for an SC link functioning (e.g. a company has to stop
production process because delivery of components is
delayed).

1065

In order to estimate the process reliability

Nowakowski (2006, 2007) proposed to find the partial
indexes for each possible error:
Ncpl
Ncpr
Pcpr =
; Pccu =
; Pcpl =
N
N
Ncql
Ncqt
; Pcqt =
; Pcdo =
Pcql =
N
N

Nccu
Ncti
; Pcti =
;
N
N
Ncdo
(1)
N

where P = probability of correct action; N =

number of correct actions; N = number of deliveries; cpr = correct product; cpl = correct place;
ccu = correct customer; cti = correct time interval of delivery; cgl = correct quality of product;
cqt = correct quantity of product; cdo = correct
documentation.
On this base reliability indicator of the logistics
system may be estimated:
Rd1 = Pcpr Pcpl Pccu Pcti Pcql Pcqt Pcdo

(2)

or:
Rd2 = wcpr Pcpr + wcpl Pcpl + wccu Pccu
+ wcti Pcti + wcql Pcql + wcqt Pcqt + wcdo Pcdot
(3)
where w = weight of the partial index P .
System reliability Rd1 (Equation 2) assumes that the
analyzed random events (faults, failures) are statistically independent and the significance of the events
is the same. This assumption is a strong simplification. The second approach (Equation 3) takes into
consideration the weight of potential consequences
of a logistic failure. The indicator is more realistic
because every fault of logistics system causes different
effect on the link of SC. The indicator allows to: identify the importance of a given component (parameter)
with a respect to the system performance of interest
and evaluate the effect of the change in any direction
of any parameter (Do Van et al. 2007). If every deviation from planned values of 7R formula is understood
as a fault of the system, the direction of deviation is
significant.
The discussed problem is known and solved i.e.
dealing with maintenance policy defining in multiunit, heterogeneous system. Bevilacqua & Braglia
(2000) suggest that criticality analysis based on
FMECA technique may be very helpful tool. Authors
present the list of possible criteria of elements assessment: safety, machine importance for the process,
spare machine/parts availability, maintenance cost,
access difficulty, failure frequency, downtime length,
machine type, operating conditions, propagation
effect and production loss cost. They perform the
analysis in the electrical power plant, composed on

thousands various elements. Six chosen criteria are

evaluated with its wages:

safety machine 1, 5,
machine importance for the process 2, 5,
maintenance costs 2,
failure frequency 1,
downtime length 1, 5,
operating conditions 1.

The factor coming from this assessment gives the

base for maintenance program defining.
But assessing logistic process the indicators are difficult to calculate for any real logistic process because
of a lack of sufficient information/data. Companies
rarely decide to publish statistical data concerning
errors that occur during the logistics process. This fact
also hinders researchers to estimate values of partial
indexes weights of Rd2 coefficient (Equation 3). In
practice companies have to elicit that knowledge by
less certain methods, i.e. experts questionings.
In order to eliminate such subjective assessment, the
next chapter show the concept how to attach a measure of importance of partial event/errors as regard
reliability of the logistics system.

IMPORTANCE FACTORS IN RELIABILITY

To develop more adequate reliability measure importance weights should be estimated. The new concept
assumes that the weights may base on importance factor models used in the reliability theory to establish
the importance of a component or an event. Reliability
analysts have proposed different analytical and empirical measures which rank components regarding their
importance in a system. They allow to identify components/events within a system that more significantly
influence a system behavior with respect to reliability,
risk and safety (Espiritu et al. 2007). The short brief
of existing models is presented below.
3.1 Marginal importance factor
(Birnbaum importance factor)
The indicator introduced by Birnbaum (1969) describe
probability that the system will be in operating state
with component i as the critical, when i is operating.
It is the rate at which the system reliability increases
when the reliability of the component i increases
(Villemeur 1992). Analytically it is defined by:
IiB (t) =

Rs (t)
Ri (t)

(4)

where: Rs (t) = system reliability at time t; Ri (t) =

reliability of component i at time t.

1066

The weakness of Birnbaum indicator is that IiB (t)

does not depend on current level of component reliability (in t). Therefore, two components may have
a similar metric IiB (t) value, although these current
levels of reliability could differ substantially (Espiritu
et al. 2007).

caused by a particular component (Espiritu et al.

2007):
IiVF (t) =

Rs (t) Rs (t; Ri (t) = 0)

Rs (t)

(7)

where: Rs (t; Ri (t) = 0) = system reliability at time t

if given component i is failed.
3.2

Critical importance factor

The critical importance factor is an extension of the

previous metric. It is the probability that the component i has caused the failure of the system, knowing
that the system is in failed state (Villemeur 1992).
Component i is not necessarily the only component in
failed state, but if this is the case, it is the most recent
to have failed. The metric includes the component
unreliability:

3.5 The Risk Achievement Worth

The Risk Achievement Worth indicator is the standard
risk importance measure (Vesely et al. 1983):
RAWi =

Wi+
W

(8)

where: Fs (t) = system unreliability at time t; Fi (t) =

unreliability of component i at time t.

where: Wi+ = the increased risk level without feature

i or with feature i assumed failed; W = the present
risk level.
The indicator was converted into a system performance (e.g. reliability) measure PAW (Levitin et al.
2003). It quantifies the maximum possible percentage
increase in system reliability generated by a particular
component (Espiritu et al. 2007):

3.3

PAWi (t) =

IiCR (t) = IiB (t)

Fi (t)
Fs (t)

(5)

The BarlowProschan measure

The ratio describes probability that the failure of ith

component coincides with the failure of the system.
It is a weighted average of the Birnbaum measure,
dependent on unreliability of ith component in t.
A component is more important the more likely is to be
the direct cause of system failure (Natvig & Gasemyr
2007):

IiBP =

IiB (t) fi (t) dt

(6)

3.4

pdf of ith component lifetime

Diagnostic importance factor (VeselyFussell

importance factor)

This importance factor describes probability that the

component i has failed, knowing that the system
is in failed state. It can be useful in diagnosing
system failure causes (Villemeur 1992). According
to this measure, the importance of a component
in the system, depends on the number and on the
order of the cutsets in which it appears. It quantifies the maximum decrement in system reliability

(9)

where: Rs (t; Ri (t) = 1) = system reliability at time t

if given component i is perfectly working.
3.6 The Risk Reduction Worth
The Risk Reduction Worth RRW indicator (Vesely
et al. 1983) may also be adapted to measure a system
performance:
RRWi =

where fi (t) =
distribution.

Rs (t; Ri (t) = 1)
Rs (t)

W
Wi

(10)

where: Wi = the decreased risk level with the ith

feature optimized or assumed to be perfectly reliable.
The measure (converted into reliability goals) quantifies potential damage caused to the system by a
particular component (Levitin et al. 2003):
PRWi (t) =

Rs (t)
Rs (t; Ri (t) = 0)

(11)

where: PRW = Performance Reduction Worth indicator.

All presented factors subtly differently rank elements or event importance for the system. There is
no simple rule which indicator should be used. Similarly, there is no simple rule which model is the best

1067

Table 1.

Table 2. The type of logistic system failure with its criticality for a production process.

Importance factors list.

The influence
of element
reliability
Analyzed reliability level of ith element
on system
reliability
Ri (t) = 0 Ri (t) = function of t Ri (t) = 1
Rs (t) increase
Rs (t) increase IiVF (t)
RAWi (t)
Fs (t) increase

IiB (t)

RAWi (t)

IiCR (t)IiBP (t)

to adopt for a company to assess reliability of its

logistic processes. The greatest problem in the reliability theory is to estimate functions necessary to
obtain values given by Equations 411. In the case
of logistic system reliability, the most problematic
may be to adopt properly above indicators to the case
of logistics. Table 1 summarizes importance factors
according to elements importance influence on system reliability results. Element is higher weighted
when it has stronger influence on particular system
results.

SUPPLY PROCESS RELIABILITY

Every link of SC has to provide oneself with the raw

materials, supplies, and components in order to carry
out its basic functions (e.g. production). The tasks are
realized during supply process, which is composed of
some stages (Ballou 2004):

order preparation and transmittal,

order realization made by supplier,
transportation of goods,
delivery inspection (deviation from the order examination),
supply correcting (e.g. complementary order).
Every step may generate logistic failures understood as discrepancies between values of planned and
real supply process in any 7R formula criterion. Perfectly reliable supply process would have no failures
in every stage of it.
As stated before, logistic system unreliability may
have two types of consequences: cost and critical for
the SC. Table 2 presents the analysis of possible consequences of supply process faults for a production
enterprise.
The above analysis shows explicitly various importance of particular failures of supply process. In
order to develop logistics system reliability model it
seems reasonable to distinguish strength of influence
of different types of logistic failures, as proposed.

The type
of supply
process failure

Direction of
discrepancy
between order
and delivery

incorrect
product

incorrect place

incorrect
customer

incorrect time
interval of
delivery

too early
delivery
too late
delivery

incorrect quality
of product

better quality
than ordered
worse quality
than ordered

incorrect
quantity of
product

bigger delivery
than ordered
lesser delivery
than ordered

incorrect
documentation

Consequences
may cause production
process stoppage
cost, critical
may cause production
process stoppage
cost, critical
may cause production
process stoppage
cost, critical
higher costs of
inventory holding
cost, non critical
may cause production
process stoppage
cost, critical
no consequences
may cause production
process stoppage
cost, critical
no consequences
may cause production
process stoppage
cost, critical
cost

NUMERICAL EXAMPLE

A company realized 100 deliveries of particular

product. There were observed some failures of supply process (deviations from orders). Some of them
were critical for the enterprise production process and
it had to be stopped. The number of errors and its
consequences are presented in Table 3.
When adopting the PAWi factor to estimate importance of logistics system faults, Rs (t) and Rs (t; Ri (t) =
1) have to be redefined and calculated. From the
supplied enterprise point of view, any critical logistic failure caused production stoppage (like in series
reliability structure). That is why Rs (t) may be defined:
Rs =

1068

 Mi
i

99
97 98
11

1 1 = 0, 94
100
100 100
(12)

Table 3.

Failures of the supply process.

Table 4.

The type
of logistics
system failure
(incorrect)

The number
of failures

The number
of failures, which
were critical
for production
process

product
place
customer
time interval of delivery
quality of product
quantity of product
documentation

1
0
0
30
3
7
15

1
0
0
3
2
0
0

Requirements for the suppliers.

Domain

Requirements

Quality

Quality management system ISO-TS

16949, ISO 14001
PPM
DPM
Six Sigma
Product cleanness
Delivery punctuality
Delivery completeness
Multiuse packages
Transport globalisation
Service: 7days/24 hours
SPC use
CPK 1.66
CMK 1.66
Development and production flexibility
Modern production technology
Every year price reduction: 35%
Price independence on raw material costs
fluctuations
Price independence on rate of exchange
Strong dependency of price on quantity
and time of contract
Regular concurrency examination

Logistics

Technology

where: Mi = number of deliveries realized without

any critical failure of ith type.
For ith type of possible logistic errors:
Rs (R1 = 1) = 1 1 1

Rs (R4 = 1) =

Rs (R4 = 1) =

Purchase

97 98

1 1 = 0, 95
100 100
(13)

98
99
111
1 1 = 0, 97
100
100
(14)
98
99
111
1 1 = 0, 97
100
100
(15)

where: Rs (Ri = 1) = system reliability ration when

component i works perfectly.
Hence:
PAW1 =

0, 95
= 1, 01
0, 94

(16)

PAW4 =

0, 97
= 1, 03
0, 94

(17)

PAW5 =

0, 96
= 1, 02
0, 94

(18)

For another types of failures PAWi = 1.

When the importance factor is taken into account,
Rd2 is estimated as it is higher than in the case when
all wi = 1. It means that reliable realized deliveries (with a view to the most critical logistic failures)
have greater meaning. Every next logistic failure will
reduce reliability of the supply process dependently on
the its type.

EXPERIMENTAL EXAMPLE

An example of suppliers assessment is based on data

from one of the piston factory of MAHLE GmbH consortium (Kuprowska 2005). After over taking of one
of the polish factories it was necessary to verify and
to unify the method of suppliers assessment according
to procedures and regulations obligated in the consortium. The group of experts from the consortium
realized the analysis of present method of logistic system reliability improvement. Preliminary evaluation
of the method showed that it doesnt protect reliability
of the logistic system in the factory.
One of tasks of the project (Kuprowska 2005) was
to elaborate new levels of importance factors for all
criterions and to prove/check the values by the experimental way. The assumptions of the importance factor
structure were based on world automotive industry
requirements for the suppliersTable 4.
According to the analysis new assessment system
took into consideration 9 criterions:

1069

material quality,
price,
punctuality of delivery,
completeness of delivery,
service level,
payment conditions,
documentation correctness,

 package quality,
way of controversy solutions.
The criterions were ranged according to the level of
importance by group of 16 experts. Some results of this
expertise is shown in Table 5 for criterions concerning
quality of delivery.
The method was implemented in the factory and
some data concerning reliability/quality of delivery
process of brass sleeves from one supplier are shown
in figures 14.
The analysis concern deliveries between 1999 and
2004 while new assessment method was implemented
in 2002 year.
The rate of deliveries with correct number of
parts changes among particular monthsfrom 35% to
Table 5.

Requirements for the suppliers.

Assessment
criterion

Criterion
wages

Delivery/Punctuality
On time
1 day
2 days
Delivery/Completness
Correct
Deviation 3
Deviation 5
Deviation 10
Delivery/shipment/
packeging instruction
Full, no omissions
Slight omissions,
unimportant
Service
Marshaling warehouse
Buffer warehouse
DDP/DDU

Figure 1.
of time.

85%. Taking annual mean into consideration (correct

amount and right day) one can notice the progress
from 32% in 1999 year to 78% in 2004see Figure 2.
Rate of deliveries realized just in time is shown in
Figure 3 and rate of late deliveriesin figure 4. There
were some differences in interpretation of time of
delivery between the produces and the supplier. Thus,
the producer has estimated the reliability of deliveries

Maximal number
of points
for the criterion

12,5%

100
70
50

5,75%

100
80
50
30

Figure 2.

Rate of deliveriescorrect amount and right day.

Figure 3.

Rate of deliveries just in time.

Figure 4.

Rate of late deliveries.

1,75%
100
50
1,75%

100
70
50

Total number of deliveries in analyzed period

1070

as about 90% while according to the supplier datait

was 95100%. But the progress of the process reliability obtained after the new method of assessment is
significant.
7

SUMMARY

Every logistic (e.g. supply) process is composed of

stages realized in time by different participants of SC.
The process unreliability may generate various, not
necessary economic, consequences for every SC company. Thus there is a need to define adequate reliability
measure of the process. The article takes into consideration reliability models existing earlier and proposes to
broaden them in order to get more adequate measures.
The presented method gives a chance to distinguish
different meaning of factors that have an influence
on process reliability. More precise management of
logistic processes is possible according to results of
the analysis.
The presented concept of process reliability
assessment needs estimation of weights of defined
faults/failures. Different analytical and empirical measures which rank components regarding their importance for reliability of a system are compared. It
allows to identify components/events within a system
that more significantly influence a system behavior
with respect to reliability, risk and safety. Numerical
example shows possibility of RAW model use. Observation from automotive industry allows to discuss the
effectiveness of method of logistic system reliability
improvement.
The samples of reliability assessment of supply
process reliability underline the necessity of further
conceptual work to be carried out.
REFERENCES
Ballou, R.H. 2004. Business logistics/supply chain management. New Jersey: Pearson Education Inc.
Bevilacqua, M. & Braglia, M. 2000. The analytic hierarchy process applied to maintenance strategy selection.
Reliability Engineering and System Safety 70: 7183.
Birnbaum, Z.W. 1969. On the importance of different components in a multi-component system. Multivariate analysis
2. New York: Academic Press.

Do Van, P. & Barros, A. & Berenguer C. 2007. Importance

measure on finite time horizon and application to Markovian multi-state production systems. In Aven & Vinnem
(ed.), Risk, Reliability and Societal Safety; European
Safety and Reliability Conference Proceedings, ESREL,
Stavanger 2007. London: Taylor & Francis Group.
Espiritu, J.F. & Coit, D.W. & Prakash U. 2007. Component
criticality importance measures for the power industry.
Electric Power Systems Research 77: 407420.
Jodejko, A. 2007, The model of procurement process taking
into account the consequences of a technical system reliability. In Aven & Vinnem (ed.), Risk, Reliability and
Societal Safety; European Safety and Reliability Conference Proceedings, ESREL, Stavanger 2007. London:
Taylor & Francis Group.
Kuprowska, I. 2006. Method of logistic system reliability improvement of multiseries production in automotive
industry. Phd Thesis, Wroclaw University of Technology,
Mechanical Engineering Faculty, Wroclaw (in polish).
Levitin G., Podofillini L., Zio E. 2003. Generalised importance measures for multi-state elements based on performance level restrictions. Reliability Engineering and
System Safety 82 (3): 287298.
Natvig B., Gasemyr J. New results on component importance in nonrepairable and repairable systems. In Aven
& Vinnem (ed.), Risk, Reliability and Societal Safety;
European Safety and Reliability Conference Proceedings, ESREL, Stavanger 2007. London: Taylor & Francis
Group.
Nowakowski, T. 2004. Problems of reliable operation of
logistics system (in polish). Inzynieria Maszyn 9 (1):
5056.
Nowakowski, T. 2005. Problems of logistics system reliability modeling (in polish). Systems 10 (1): 122133.
Nowakowski, T. 2006. Analysis of possibilities of logistics
system reliability assessment. In Guedes Soares & Zio
(ed.), Safety and Reliability for Managing Risk; European
Safety and Reliability Conference Proceedings, ESREL
07, Lisbon 2006. London: Taylor & Francis Group.
Nowakowski, T. 2007. Assessment of reliability distribution
process. Case study. In Aven & Vinnem (ed.), Risk, Reliability and Societal Safety; European Safety and Reliability
Conference Proceedings, ESREL 07, Stavanger 2007.
London: Taylor & Francis Group.
Vesely W.E., Davis T.C., Denning R.S., Saltos N. 1983. Measures of risk importance and their applications. Report
NUREG/CR 3385 BMI-2103. Battelle Columbus Laboratories, Ohio.
Villemeur, A. 1992. Reliability, Availability, Maintainability
and Safety Assessment. New York: Wiley.
DSMC Glossary. Defense Systems Management College
http://www.dau.mil/regions/dsmc_spm.asp (30.03.2006).

1071

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Applying optimization criteria to risk analysis

H. Medina, J. Arnaldos & J. Casal
Centre for Studies on Technological Risk (CERTEC), Department of Chemical Engineering
Universitat Politcnica de Catalunya, Barcelona, Spain

ABSTRACT: When performing a risk analysis, the target is usually a certain tolerable risk level. The applied
safety measures have a direct influence on the final value of the risk for a given plant and, thus, increasing these
safety measures is the usual way to reduce the risk to a tolerable level. However, there is also the possibility
of reaching the desired risk value by searching the best design through an optimization approach. When the
design is modified to diminish the riskfor example, by substituting a large storage tank by two or more
smaller tanksthe cost of the plant often increases and the consequences of the hypothetical accident decrease.
Consequently, the cost of the accident decreases as well. Therefore, the possibility of reaching an optimum
design implies a minimum total cost and a tolerable risk level constraint (taking into account both the cost of
the equipment and the cost of the accidents that can occur).

INTRODUCTION

The chemical industry has reachedat least in certain

countriesa relatively high level of safety, especially
when it is compared to several other human activities. The existing safety criteria provide critical limits
for the process parameters. Over the years, important
advances have been done in this area in order to obtain
efficient and profitable processes which, at the same
time, be safe for the environment and people. Nevertheless, despite the efforts that have been done to
reduce the frequency of major accidents, the increasing industrialization continues generating accidents
with an impact on people, environment and properties.
The reduction of risk can be achieved through the
use of protective measures, additional intervention
equipment and automatic controls. Another complementary way is the application of safety criteria from
the first design steps.
However, there is not a systematic methodology or
a protocol that allows combining all the measures to
reduce the risk as much as possible, taking simultaneously into account all costs. Optimization has allowed
the chemical industry to be more competitive and profitable by making a better use of the logistic and material
resources of the plants. As a consequence of this, a
reduction of costs and an increasing of production and
profits, among other benefits, are achieved. Therefore,
the application of optimization criteria to plant design
must be considered as a promising tool to minimize
the risk to people and environment to tolerable values
with a reasonable investment.

SAFETY THROUGH A SAFER DESIGN

The design of a large process plant is a quite complex

activity which is carried out in stages over a period
of time and involves people of many disciplines.
The decisions which are made in the early stages,
particularly those concerning the process route are
influenced by diverse circumstances: research and
development, which define the technical possibilities and constraints; safety, health and environmental
studies which indicate further constraints; economical studies that indicate the productions costs and
sales returns together with the sensitivity to the various
factors (Lees, 1996).
The safer design is a philosophy that considers
safety aspects from the beginning of the designing step
up to the operation of chemical plants. It focuses on
eliminating or reducing significant hazards. Once the
existence of a hazard and the magnitude of the risk
associated to it have been identified, efforts to reduce
it are applied. The inherently safer design of plants
provides more robust and reliable risk management
and has the potential to make the chemical processing technology simpler and more economical in many
cases.
The safer design focuses on the immediate impact of
chemical accidents on people, the environment, property and business. In a chemical plant, this generally
means the immediate impact of fires, explosions and
toxic releases. Of course, these types of events will
also have the potential for long term impacts. The
main purpose of designing safer plants is to reduce

1073

the frequency and potential impact of chemical plants

accidents (Hendershot, 2005).

RISK ANALYSIS AND COST

OPTIMIZATION PROCEDURE

The use of the optimization methodologies is an interesting way to reduce risk to an acceptable level with
a minimum economical investment. The suggested
methodology has been represented in a schematic way
in Figure 2.
4.1 System definition
The system defines the borders of process or set of
equipments to focus calculations on. It can be applied
to a section of a plant and particularly to the equipment
that is more susceptible to risk, either due to the process
itself or to the substances that are processed.
4.2 Selection of variables
The selection of variables refers to the problem of
selecting input variables that are most predictive in risk
analysis; this is related to variables and parameters that
influence the process design and the calculation of consequences. These variables can physically represent
equipment size and operating conditions.
4.3 Establishment of the objective function
This is one of the most important steps in the applications of optimization to any problem. It is often

Investment in safety

Risk analysis is used to evaluate the hazards associated to an industrial installation, a particular activity
or transportation of a hazardous material. The methodologies of risk analysis provide a reasonable estimation
of potential accidents, the frequency of these accidents
and the magnitude of its effects or consequences. The
reasons why organizations are concerned about risk
analysis can be classified in three types:
Moral: related to human life and health (Tweeddale,
2003).
Legal: depending on the specific legal and legislation framework of the community. They are related
to environmental laws and the international safety
standards.
Commercial: loss of benefits, costs associated to
damages to people, buildings, environment and image
of the company.
In order to reduce the risk, investments are made
in chemical plants in areas such as: management,
research, design, process, operational, plant sitting,
plan layout, equipment, instrumentation, fire protection, inspection and emergency planning. Generally
speaking, as the investment in this fields increases,
the risk associated to a given plant or activity decreases
(Figure 1).
There is no doubt about the economic importance of
loss prevention. The various safety factors which are
incorporated in the plant design can increase significantly the cost. These can include design with thicker
vessels walls, use of more costly materials of construction, selection of more expensive equipments and
redundancy of certain items. Some cost arise through
failure to take proper loss prevention measures, others are incurred through uninformed and unnecessarily
expensive measures.

Risk
Figure 1.

Investment in safety implies a reduction of risk.

Figure 2. Steps to optimize the investment required to reach

a tolerable risk level.

1074

expressed in units of currency because the goal of

the enterprise is to minimize cost or maximize profit
subjected to a variety of constraints. In the case of
risk analysis an alternative unit can be the number of
fatalities.
4.4

Selection of significant accidents

Selection of the significant (because of their consequences and frequency) accidents, depending on: the
features of the system, the type of substances that are
handled and the operating conditions of the process.
4.5

Calculation of the intensity of the damage

It is related to the magnitude that causes the damage and the time of exposure to it. It refers to the
effects of an accident, for example, the overpressure,
the radiation or the concentration of a toxic gas.
4.6

Calculation of consequences

The calculation of consequences refers to the actual

damage caused: any degree burns due to radiation,
pulmonary haemorrhage or ruptured eardrum due to
overpressure, intoxication by breathing a toxic gas,
and death for any of the events, in the worst case. Furthermore, damage to equipment, houses, etc. should
be also considered.
4.7

Calculation of the cost of the proposed designs

The cost of the sized equipment will be estimated

using appropriate methodology such as a power law
relationship.
4.8

Determination of the optimal design

The goal of optimization is to find the values of

the variables in the process that yield the best value
of the performance criterion. The variables, process
and models constitute the optimization problem. It is
concerned with selecting the best among the entire
set by efficient quantitative methods. Computers and
associated software make the necessary computations
feasible and cost effective (Edgar et al., 2001).
5

According to a recent historical survey (Darbra and

Casal, 2004), among the diverse major accidents that
occur approximately 60% are fires, 35% explosions
and 6% gas clouds. Nevertheless, although gas clouds
are significantly less frequent than the other accidents,
they can have severe consequences over a large area.
This is why a toxic cloud has been selected to illustrate
the proposed method.
Once the release into the atmosphere takes place,
a plume or a cloud is formed and evolves under the
influence of the prevailing meteorological conditions
(wind speed and direction, atmospheric turbulence,
humidity, temperature and the possible existence of
thermal inversion). Meteorological average parameters are considered during the escape. Appropriate
models predicting the atmospheric dispersion must be
applied to predict this evolution and to know how the
population will be affected by the toxic cloud.

CASE STUDY. RELEASE OF A TOXIC

CLOUD FROM A STORAGE TANK

In industrial plants, a great number of accidents take

place due to the loss of containment in tanks and
pipes that contain or transport liquid and gaseous
products. When a release occurs, these substances can
originate gas clouds depending on the meteorological
conditions.

CALCULATING THE INTENSITY

OF EFFECTS

The intensity of effects in the case of the atmospheric

dispersion of a toxic cloud is determined by the evolution of the gas concentration as a function of time.
Dispersion models are used to predict the evolution of
the pollutant cloud as a function of position and time.
An instantaneous emission produces a cloud that
moves in the wind direction while it disperses in the
atmosphere. The dispersion of a substance in the atmosphere can be influenced by its density, especially at
the beginning of the escape. If the released material is
heavier than air, first it will descend to the ground at the
same time it extends (Casal et al., 1999). Many toxic
substances that can be released are heavier than air, so
it is the case of chlorine, hydrofluoric acid, ammonia
and propane, which behave as heavy gases as a result
of their molecular weight, chemical reactions of the
components during their escape, their volatility or the
tendency to form aerosols.
For light gases, the concentration of the toxic gas
can be predicted by analytical equations with certain
degree of accuracy. Nevertheless, for heavy gases the
prediction is significantly more difficult because of
the complexity of the phenomenon which requires
complex models including many parameters.
In order to determine the evolution of concentration
as a function of time and distance, the Aloha (Areal
Locations of Hazardous Atmospheres) code has been
used. ALOHA was developed by the National Oceanic
and Atmospheric Administration (NOAA) and Environmental Protection Agency (EPA). The following
are features of the software:
Used for heavy gas and neutrally buoyant gas
modelling.

1075

Estimates source strength.

Used in releases from tanks, puddles, and pipes.
Calculates indoor air infiltration.
Contains an extensive chemical library.
Estimates gas cloud area and concentration over
time in varying environmental conditions.
Uses weather data that can be entered by the user or
directly from a meteorological station.

tables or expressions such as (Casal, 2008):




Y 5
% death = 50 1 + erf
20.5
where erf = error function.
8

For heavier than air gases, the code uses AlohaDegadis, a simplified version of the heavy gas model
Degadis, which is one of the more well-known and
accepted heavy gas models. This model computes
pollutant concentrations at ground level, where a dispersing chemical is most likely to contact people,
and produces a footprint, plotted either on a grid
or background map. Each footprint represents the
area within which, at some point during the hour
following the beginning of a release, ground level pollutant concentrations will reach or exceed the level
of concern entered by the user. The model produces
graphs of indoor and outdoor concentration expected
to be experienced by people at any specific location
(Cameo, 2007). The program uses information concerning meteorology, chemical compounds, source,
mass and type of release, etc.

CALCULATION OF CONSEQUENCES

Once the concentration to which people are exposed

has been determined, the consequences can be studied applying vulnerability models, and the best form
to analyze these toxic cloud effects is through probit
equations.
These are expressions that establish the relationship
between the magnitude of the action and the damage
caused to the exposed population, i.e. the percentage of
a population subjected to an effect at a given intensity
which will suffer a given damage. These relations are
very trustworthy for many dangerous accidents.
The vulnerability to inhalation of a toxic substance
is related to its nature and dose, this dose being a function of the concentration and the time of inhalation.
The probit equation for cases of death by inhalation
of a toxic substance is:
Y = a + b ln(cn t)

(1)

where Y = function probit for a release of a toxic

gas; a, b and n are constants depending on each substance; c = concentration of the gas [ppm]; t = time
of exposition to the gas [min].
The percentage of people killed due to the exposition to the gas is related to the probit function through

(2)

8.1

OBJECTIVE FUNCTION
The objective function

The objective function is represented as an equation of

costs, in order to minimize the total cost associated to
the consequences of risk of an instantaneous release
of gas and the cost of the equipment used for storage,
as presented below:
TC =

ni


CIi +

i=1

nj


CCj

(3)

j=1

TC = total cost; CI = cost of the equipment, including the cost of one or more storage tanks along with
direct and indirect costs; CC = costs of the consequences, including costs of fatalities. Damages to
buildings and equipment, damages to the environment
are not considered.
The total capital cost of the equipment is referred
to a stainless steel pressurized tank (Smith, 2005),
updated to 2007 through appropriate indexes:

Equipment cost = 70354

V
5

0.53
N

(4)

where: V = volume of the tank [m3 ]; N = number of

tanks, (all costs expressed in C
=).
One of the major difficulties in this optimization is
the quantification of certain costs. This is especially
difficult when human lives are involved. However,
and even though this is a rather delicate field, some
attempts have been done to allow this quantification
(Ronza et al., 2008). According to certain criteria,
these authors suggested that the penalty for dead
people has a cost (C
=) given by:
Death cost per person = 212, 000

(5)

This is an average value assigned to a fatal victim,

updated to 2007. This attributed cost changes with the
type of society and country.
8.2 Constraint
The optimization is subjected to a constraint. Once the
optimum cost has been selected, the associated risk
must be less or equal than the tolerable risk.

1076

OPTIMIZATION. NUMERICAL EXAMPLE

The selected scenario, taken from an existing installation, considers the storage of 1900 kg of chlorine in
one or more pressurized tanks. The accident consists
in the instantaneous release of the full inventory of the
tank in 10 min. There are inhabited buildings with 18
people indoors at a distance of 300 m downwind.
Weather conditions:
Wind speed = 4 m/s.
Temperature = 12 C.
Stability = D type.
Night.
Humidity = 70%.
Type of gas = chlorine.
Time of release = 10 min.
Pressure in the tank = 5 atm.
Ground roughness coefficient = 10 cm.

Figure 4.

Figure 3. Concentration gas indoors and outdoors vs. time

(one tank).

5
4

Fatalities

The distance or the specific coordinates of the

building close to the release can be used as a reference. For specific cases and to consider the scenario of
greatest risk, adverse weather conditions can be used
or the most likely conditions through the use of frequencies, directions and a wind speed rose with other
data specific to the area. For the location and with this
information, indoors and outdoors concentration can
be calculated, in the event of a complete release of
the total content of chlorine (Figure 3). It is observed
that the maximum indoor concentration would reach
244 ppm when using a single tank.
By applying the vulnerability models, the zone with
a probability of 100% of mortality for people located
within buildings and outside them can be estimated
(Figure 4). For the zones outside these regions, the
consequences must be calculated from the concentration data for every single point.
It can be observed that all the people will die up to a
distance of 90 m if they are confined (darker shadow)
and up to a distance of 310 m if they are outdoors.

Area of 100% mortality, indoors and outdoors.

3
2
1
0
1

Figure 5.

3 N

Fatalities versus number of storage tanks.

These percentages decrease as the distance from the

release point increases. At 300 m, fatalities indoors
are 24%.
If two tanks are usedeach one with half of the
initial massthe involved mass of chlorine in the event
of a release will be significantly smaller. Thus, in the
event of an accidental release of the whole inventory
of a tank, indoor concentration decreases at a given
distance. As a consequence, it can be seen that the
number of fatalities decreases as the number of storage
tanks increases (Figure 5); this implies consequently
a reduction of the consequences.
Taking into account these considerations, particularly in the reduction of consequences, it is clear that
the variation in the number of storage tanks (and, consequently, in their respective volume) is the essential
factor having a definitive influence on the hazard associated to an eventual release of chlorine. Therefore, the
objective function will be based on this variable.
In the objective function, the cost of the equipment increases with the number of tanks (Figure 6),
despite the reduction of the volume of each one
of them.

1077

Consequences Cost
800000

Cost []

Equipment Cost
600000

400000

200000

0
1

Figure 6.

Total cost versus the number of storage tanks.

The overall cost is the sum of both, equipment cost

and consequences cost. For the proposed scenario,
the curve presents a minimum; using optimization
software, this point was calculated and corresponds
to 3,6.
Initially, the total cost decreases considerably due
to the significant reduction of the consequences; by
increasing further the number of tanks, the cost of the
consequences continues to decrease but very slowly,
while the cost of equipment continues increasing
steadily. Thus, for the case of study, the use of three
tanks would result in a reduction of 76% of the total
cost, when compared with that of a single one.
Although this is the theoretical optimum number,
looking at Figure 6 it is clear that increasing the number
of storage tanks from two to three implies only a light
reduction of the total cost when it is compared with the
significant reduction given by increasing from one to
two storage tanks. Therefore, and taking into account
the additional complexity associated to any increase in
N, in this case probably the optimum should be taken
as two storage tanks.
It is important to indicate that there will be conditions in which the objective function will not exhibit
a minimum. This can occur when one of the components of the total cost predominates on the other
one. This is the case of zones very close to the release
point, where the cost of the consequences predominates due to the greater concentration and, therefore,
to the greater mortality. On the other hand, there will
be distant zones from the release where this value will
be negligible.

9.1

0,0000025

Risk [fatalities/year]

Total Cost

1000000

approach. However, the definition of risk must include

the frequency of the accident.
For the accidental scenario considered, the instantaneous release of the full inventory of the tank, a
frequency of 5 107 years1 has been assumed
[CPD18E, 2005]. Now, when risk is calculated as the
product of frequency and consequences, it can be seen
that the risk for a single tank is significantly higher
than that corresponding to several tanks (Figure 7),
risk decreasing with the number of tanks. Thus, for
example, when using 3 tanks, risk is reduced in one
order of magnitude.
Risk can be presented by plotting the frequency
(f ) of the accidental scenarios that cause Nk fatalities (Figure 8). It should be noted that this figure is not
the usual f -N curve used to present societal risk, but
it allows comparing the different cases: it can be seen
that increasing the number of storage tanks reduces the
number of fatalities but increases the frequency of the
event due to the higher number of potential sources.

0,000002

0,0000015

0,000001

0,0000005

0
1

Figure 7.

Risk versus number of storage tanks.

1,0E-05

frequency [1/year]

1200000

1 tank
2 tanks
3 tanks
4 tanks
5 tanks

1,0E-06

1,0E-07
0,01

Frequency and risk

0,1

1
Nk

In the previous sections, the cost associated to the

diverse situations was considered with a deterministic

Figure 8.

1078

Frequency of the accident versus fatalities.

10

10

CONCLUSIONS

ACKNOWLEDGEMENTS

A systematic methodology to reduce the risk using

optimization criteria as a tool is presented. Satisfactory results have been obtained when applying it to
the storage of chlorine assuming the occurrence of
an accidental release. A reduction of the total cost in
76% was reached when increasing the number of tanks
from one to three along with a considerable reduction of risk. Additional considerations showed that
the optimum number could be considered to be two
tanks.
Applying optimization criteria can be an additional
useful and important tool in risks analysis, being
able to play an important role in the design of safer
plants. In this optimization process, however, reaching a tolerable risk level must be always an essential
constraint.
11
a, b
c
CC
CI
erf
f
n
N
ni
nj
Nk
t
TC
V
Y

NOTATION
constants in Eq. (1)
indoors concentration of the gas (ppm)
costs of the consequences (C
=)
cost of the equipment (C
=)
error function ()
frequency (year 1 )
constant in Eq. (1)
number of storage tanks (C
=)
number of equipment cost items ()
number of consequences cost items ()
number of fatalities ()
time of exposure to the gas (min)
total cost (C
=)
volume of the storage tank (m3 )
function probit ()

H. Medina thanks the Universidad de los Andes in

Venezuela, for a doctoral grant.
REFERENCES
Cameo, Aloha users Manual. 2007. Environmental
Protection Agency and National Oceanic and Atmospheric Administration.
Casal, J., Montiel, H., Planas, E., Vlchez, J. 1999. Anlisis
del riesgo en instalaciones industriales. Barcelona. UPC
edicions.
Casal, J. 2008. Evaluation of the effects and consequences of
major accidents in industrial plants. Amsterdam. Elsevier.
CPD18E. 2005. Guideline for Quantitative Risk Assessment.
(The Purple Book). The Hague, SDU. VROM.
Darbra, R.M., Casal, J. 2004. Historical analysis of accidents
in sea ports. Safety Science, 42, 85.98.
Edgar, T., Himmelblau, D., Lasdon, L., 2001. Optimization
of chemical processes. 2nd ed. Singapore. Mcgraw-Hill
Book Co.
Hendershot, D. 2005. An overview of inherent safer design.
20th annual CCPS international conference on risk management. Atlanta. CCPS.
Lees, F. 1996. Loss prevention in the process industries
hazard identification, assessment and control, 2nd ed.
Oxford. Butterworth-Heinemann.
Ronza, A., Lzaro-Touza, L., J. Casal. 2008. Economic valuation of damages originated by major accidents in port
areas. Journal of Loss Prevention in the Process Industries
(in press).
Smith, R. 2005. Chemical Process. Design and Integration.
Chichester. John Wiley & Sons Ltd.
Storch de Gracia, J. 1998. Manual de seguridad industrial en
plantas qumicas y petroleras, fundamentos, evaluacin
de riesgos y diseo. Madrid. Mcgraw-Hill.
Tweeddale, M. 2003. Managing risk and reliability of process
plants. Amsterdam. Elsevier Science.

1079

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Chemical risk assessment for inspection teams during CTBT on-site

inspections of sites potentially contaminated with industrial chemicals
G. Malich
Preparatory Commission for the Comprehensive Nuclear Test Ban Treaty Organisation, Vienna, Austria

C. Winder
School of Safety Science, The University of New South Wales, UNSW Sydney, Sydney, Australia

ABSTRACT: The Comprehensive Nuclear-Test-Ban Treaty (CTBT) bans nuclear weapon test explosions
anywhere in the world in order to constrain the development and qualitative improvement of nuclear weapons
and the development of advanced new types of these weapons. Compliance with the CBCT will be monitored
by a global verification regime with international monitoring and on-site inspections (OSI). As part of normal
operational activities, OSI teams may encounter a range of hazards and risks. One such risk is from exposure
to chemical hazards from investigations of sites where industrial, agricultural or other hazardous chemicals
were manufactured, stored, used or disposed, and where risks to OSI team members from hazardous chemicals
may arise. Conventional risk management procedures can assist in dealing with such exposures through identification of chemical hazards, assessment of chemical by risks into account a range of hazard and exposure
factors, and control of risks assessed as being unacceptable. Chemical risk assessments under the OSI situation
require consideration of a toxicity assessment component and an exposure assessment component to characterise
risk. Further, after using industrial chemicals of known hazard characteristics, practical hypothetical exposure
situations and established IDLH recommendations indicate OSI activities for many industrial chemicals where
chemical exposures may occur, it is possible to conclude that in the absence of information or evidence of presence
of industrial chemicals, Inspection Team members are unlikely to be at risk from exposure to toxic chemicals.

INTRODUCTION

The general principle of liability for negligent conduct has evolved over the last century into the general
concept of duty of care. This concept has been
crystallised into legislation in some countries, placing
mandatory obligations on individuals and organisations (duty holders), such as employers or service
providers. For employers and principal contractors,
this includes obligations to provide:

A safe place of work.

Safe materials, processes and tools.
Knowledge of work hazards.
Competent supervisors and fellow employees.
Rules and procedures for safe work.
Means for enforcement of rules.

Typically, the development and delivery of these

systems is now made through risk assessment and risk
management.
The obligations imposed by duty of care are absolute. That is, while the function of discharge of duty
of care may be shared or delegated, responsibility

remains with the primary duty holder. For this reason, delegation of responsibilities should not be taken
lightly, and checks and balances must be included in
the delegation to assure that duty of care obligations
remain intact. Further, while these obligations apply
to all conventional employer-employee (or indeed, the
principal-subcontractor) relationship, there are especial workplace circumstances where it is not always
possible for a duty holder to meet these obligations.
For example, in military or emergency situations, it is
not always possible to meet the duty of care obligation
to provide a safe workplace. Under such circumstances, duty of care obligations must be met by other
arrangements, including:
A sequence of preliminary, operational and special
risk assessments.
Development and implementation of standardised
working arrangements.
Ensuring all employees have received sufficient
training to be competent in their work activities.
Arrangements for debriefing of employees and
review of working arrangements after each assignment to ensure effective operational readiness.

1081

This paper investigates the delivery of systems for

the safety of inspectors undergoing inspection activities of a potentially dangerous nature, concentrating
on the potential problems from exposure to toxic
industrial chemicals.

(in an uncontaminated area within or near the inspection area) that is to serve as a safe haven/clean area
from where it would conduct deployments to/within
the inspection area.
2.2 Challenges to OSI risk management

COMPREHENSIVE NUCLEAR-TEST-BAN
TREATY (CTBT)

The Comprehensive Nuclear-Test-Ban Treaty (CTBT)

bans nuclear weapon test explosions anywhere in order
to constrain the development and qualitative improvement of nuclear weapons and the development of
advanced new types of these weapons. Preparations
for the effective implementation of the Treaty are
carried out by the Preparatory Commission for the
Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) in Vienna. After the Treaty enters into
force, compliance with the CTBT will be monitored
by a global verification regime.
This global verification regime is currently being
established and mainly comprises an International
Monitoring System (IMS), an International Data
Centre (IDC) and on-site inspection (OSI) capabilities.
The IMS monitors the Earth for evidence of nuclear
explosions and will eventually comprise 321 seismic,
hydroacoustic, infrasound, and radionuclide stations
as well as sixteen radionuclide laboratories. Data generated at these stations are being transmitted to the
IDC in Vienna, where they are processed, reviewed
and transmitted to the States Signatories. In the event
that a suspected nuclear explosion is detected after the
Treaty enters into force either by one of the stations
of the IMS or by national technical means, any State
Party can request an OSI.
2.1

On-Site Inspection (OSI)

An OSI would be regarded as a final verification measure. It may be launched by the CTBTO upon the
request of any State Party and subsequent approval by
the Executive Council. Once an OSI is approved, up to
forty inspectors and inspection assistants would deploy
to the field and utilise approved inspection equipment
in an inspection area of up to 1, 000 km2 , abiding by a
very short time schedule with a launching period of just
six days. This period will also see the formation of the
inspection team from a roster of potential inspectors
since the CTBTO will not have its own inspectorate.
The purpose of an OSI would be to clarify whether
a nuclear explosion has been carried out in violation of
the Treaty and to gather any information which might
assist in identifying the potential violator. For this,
the inspection team, with the expected support of the
inspected State Party, would set up a base of operation

OSIs represent a particular challenge to risk management processes as they pose a variety of health
and safety concerns which are not only consequences
of tasks being performed by the team but also a
function of the dynamic nature of the mission. For
instance, unique conditions of sites where OSIs are
likely to be conducted are associated with phenomena of ambiguous events that might or might not be
a nuclear explosion and are therefore of considerable
concern for the health and safety of the team. These
conditions relate, among other things, to the potential for ionising radiation, radioactive debris, explosive
ordnance, tectonic activity and unstable buildings or
terrain.
Additional challenges relate to those hazards that
are associated with the possible/likely conduct of an
OSI at a remote site and/or in a foreign environment.
Moreover, it is to be expected that site conditions
within the inspection area are unidentified at the outset of an OSI. This not only relates to natural terrain
features which may affect site accessibility but also to
remnants from past or present land use. For instance,
site contamination with unidentified chemical substances as a result of former industrial, agricultural
and/or military activities or uncontrolled waste disposal is a possible scenario that should not result in
unacceptable risks for the inspection team. While the
inspection team will request the inspected State Party
during the launching period to provide information
on site characteristics including previous hazardous
activities and possible contamination that will allow
conducting a preliminary risk assessment, there will
be no guarantee that the provided information is totally
accurate and comprehensive.

RISK ASSESSMENT COMPONENTS

3.1 Purpose of protection regime

OSI should be undertaken within an acceptable risk
envelope with respect to an individual and collective
basis. The present report aims to outline considerations for developing a protection regime for CTBT
OSI that may take place at sites potentially contaminated with (toxic industrial) chemicals. It is based
on the hypothesis that chemical contaminants (in
terms of individual chemicals or mixtures of chemicals assumed to interact additively) and their surface
contamination levels are present, but need only be

1082

assessed qualitatively to comply with risk management

requirements.
3.2

Underlying assumptions

The inspection team has to deal with an immense

amount and variety of tasks to be completed during the initial stages of an inspection. Among other
things, this includes the deployment of personnel and
equipment, establishing infrastructure and logistical
support, familiarisation with site-specifics with regard
to developing a sound search strategy, negotiations
with the inspected State Party with regard to guidance
and support, conduct of preliminary risk assessments
and so forth. Further, OSI team members will be
trained to comply with a minimum level of preventive
measures, such as the use of personal protective equipment. In view of this, it is not part of the OSI strategy
that comprehensive quantification of type or extent of
contamination be undertaken by the inspection team,
if any.
Considering the short timelines and the temporary
nature of an OSI, the OSI is more of an intervention
situation. Therefore, exposures to chemicals during
OSI cannot be compared to conventional occupational
situations. Chemicals in normal workplace activities
arise as a result of those activities, and are usually
produced by active, purposeful, activities. In contrast,
during an OSI, chemicals may be available passively,
that is, by virtue of their presence. Therefore, criteria
for acceptable risks, such as time weighted average (TWA) occupational exposure limits for chemical
substances designed for exposure at workplaces (that
is, eight hours per day, forty hours a week, over an
entire working lifetime) are not applicable. Similarly,
because they have a number of qualifying characteristics such as limited exposure duration and recovery
periods, short term exposure limits (STEL) are also
not considered applicable.
Instead, contamination levels immediately dangerous to life or health (IDLH), such as those
recommended by the US NIOSH would seem more
applicable, depending on the toxicity of the presumed contaminant(s), the relevant exposure situation
(concentration, intensity, duration) and the available
mitigating circumstances (climate, terrain, exposure
controls).
3.3

Methodology

The methodology for this paper covers two main

activities.
1. Toxicity assessment component: Examining types
of chemical hazards, identifying categories of
chemicals relevant for OSIs, and selecting industrial
chemicals with a view to demonstrate that specific

toxicity values exist for them which later will be

used in an illustrative risk assessment.
2. Exposure assessment component: Identifying
parameters that affect and describe exposure situations and discussing their relevance for OSIs taking
into account various realistic operational activities.
These two assessment components provide the
basis for the characterisation of risks. An illustrative
risk characterisation for OSI is outlined below.
3.3.1 Toxicity assessment component
Toxicology and chemical safety relates to the identification, assessment and control of chemical hazards.
There are many properties and factors that relate to
this process including:
Type of airborne chemical hazards: There can be
a wide range of airborne contaminants, including
gases, vapours, aerosols, mists, fogs, fumes, dusts,
and fibres. These can be divided into two main types
of contaminants: (i) those that are dissolved in air,
such as gases and vapours, that can reach, and can
cause effects in all parts of the respiratory system;
and (ii) those that are suspended in air, such as
aerosols, mists, fogs, fumes, dusts, and fibres (also
called particulates).
Effects of inhaled chemical hazards: Inhaled contaminants are inspired into the respiratory system,
although the depth that they can reach is a function
of physical characteristics, such as their size and
solubility. Dissolved contaminants will reach, and
can have effects in all parts of the respiratory system. Suspended particles may be inspired into the
respiratory system, although the depth that they can
reach is a function of their size, measured as the
mean aerodynamic diameter of the particle.
The nature of an OSI is such that with the exception of contained or partially contained chemicals, any
gaseous or volatile chemicals can be presumed to have
evaporated and dispersed well before the inspection
team arrives. Accordingly, it is for assessment of the
chemical risks to an inspection team, it is proposed
that the chemical contaminants be grouped under five
categories.
1. Synthetic inorganic solids or liquids absorbed onto
solids.
2. Synthetic inorganic solids or liquids absorbed
onto solids, but dispersed as a ground plume
and hence much diluted in concentration.
3. Mining and manufacturing residues, such as metal
ores or mine tailings, either in localised or extensive
dispersed form.
4. Semi- or non-volatile organic liquids, present as
surface pools or near surface pools.
5. Organic materials adsorbed on to a particulate
medium.

1083

Using these criteria, a list of candidate chemicals

were selected for toxicity assessment. The selection
and derivation of data and the selection of candidate contaminants must be based on well-documented
(quality) information in the public domain and should
consider both the toxicological properties and the
likely industrial applicability of the chemicals.
Other properties such as reactivity or ambient stability
of the chemical form will also be considered.
Suitable lists of toxic chemicals exist in the public domain and one such list, suitable for this paper,
is the 1994 US NIOSH list of Agents Immediately
Dangerous to Life or Health (IDLH) as recommended
in the US National Institute for Occupational Safety
and Health (NIOSH, 1994). This refers to the single
intense respiratory exposure that specifically poses an
immediate threat of loss of life, immediate or delayed
irreversible adverse effects on health or intense eye
exposure that would prevent escape from a hazardous
atmosphere.
The IDLH list is available on the Internet and was
accessed. IDLH values that were presented in terms of
mg/m3 were converted into ppm. A subset of relevant
chemicals was selected with a view to representing all
five chemical categories presented above:
Paraquat, a bipyridilium herbicide with high toxicity in the pulmonary system on inhalation of mists
(IDLH 1.00 mg/m3 );
Hydrogen selenide, a gas produced on reaction with
selenium compounds and mineral acids with high
toxicity on inhalation of gases (IDLH 3.32 mg/m3 );
Acreolin, a chemical intermediate and combustion
product with irritation potential and high toxicity on
inhalation of vapours (IDLH 4.60 mg/m3 );
PCB containing wastes, with low toxicity in the liver
and blood forming tissues on inhalation of dusts
(IDLH 5.00 mg/m3 );
Mercury vapour and dusts containing mercuric
compounds, with moderate toxicity in the central
nervous system on inhalation of vapours or dusts
(IDLH 10 mg/m3 );
Formaldehyde vapours, with irritation potential to
skin and upper airways on contact to the liquid
and high toxicity in the respiratory and immunological systems on inhalation of vapours (IDLH
24.6 mg/m3 );
Hydroquinone, an aromatic organic chemical, with
moderate toxicity in the skin on contact with the
liquid and to the kidney on inhalation of vapours
(IDLH 50 mg/m3 );
Lead containing dusts, with moderate toxicity in
the central nervous, gastrointestinal, kidney and
reproductive systems on inhalation of dusts and
absorption (IDLH 100 mg/m3 );
Manganese containing dusts, with low toxicity
in the central nervous, and reproductive systems

on inhalation of dusts and absorption (IDLH

500 mg/m3 );
Phosphoric acid, an inorganic acid with irritant and
corrosive properties on contact with the liquid of
inhalation of mists (IDLH 1, 000 mg/m3 );
Carbon black, a pyrolysis product in dust form
of low inhalational hazard, (although some poorly
pyrolised carbon black products may contain carcinogenic impurities (IDLH 1, 750 mg/m3 ).
3.3.2 Exposure Assessment Component
To arrive at a suitable exposure model the following
need to be addressed:
OSI operational activities: There are a number of
proposed activities that OSI teams carry out. Some
require consideration of standing orders and the preliminary risk assessment, other require a more detailed
on the spot risk assessment, with consideration of
new information when it becomes available. These
activities include:
1. Walkover/surveys of known low hazard areas.
2. Vehicle transits within the inspection area. While
team members are offered some protection when
travelling in vehicles, most such vehicles are not
airtight (even with all windows closed). Further,
as significant dust is raised into the air by a single
vehicle on unsealed roads, the risk from a vehicle
transit in a single vehicle is a different risk from a
vehicle travelling in the wake of another vehicle.1
3. Intrusive activities in well ventilated open areas of
unknown or presumed hazard, with water spray for
dust suppression, as needed, and use of respirators
(particulate filter) and other personal protective
equipment such as eye protection.
4. Walkover/surveys in enclosed or poorly ventilated
(possibly subsurface) areas with use of some direct
reading instruments (for example, for O2 , CH4 ,
CO), respirators (combined filter such as particulates/organics) and other personal protective
equipment such as eye protection.2
Exposure mitigation: Mitigation of exposure includes all controls, preventive measures and other
activities that reduce exposure. They include generic
features already discussed, such as team member
training, risk assessment procedures, provision of
equipment, safe working procedures, and personal

1 For

operational activities 1 and 2, a preliminary risk

assessment should be consulted as to whether the risk is
acceptable without use of respirators or eye protection.
2 For operational activities 3 and 4 a further assessment is
required as to whether the protection provided will result
in acceptable risk levels.

1084

protective equipment. Further, they may also include

features specific for each OSI, such as designation of
clean/rest areas, air-conditioned vehicles, and/or dust
suppression.
Standing orders and procedures: Documentation
will be available for any OSI that provides details about
the site, including geographic parameters, previously
suspected or known site specific activities and so on.
This documentation (and any other relevant information) should be used to prepare a preliminary risk
assessment that outlines the hazards, probable risks
and recommends appropriate risk mitigation activities. Based on this, all OSI team members will be
informed about specific risk mitigation activities for
the site being inspected.
Standing orders and procedures include, among
others, requirements for:
use of gloves (of an appropriate material) and other
personal protective clothing, as needed, before
touching potentially contaminated material;
no lying down on the ground;
personal hygiene in the field;
no eating, drinking or smoking before washing hands,
face, and food containers.
Inhalation of contaminants is considered to be the
only credible pathway, since other pathway routes are
essentially eliminated by adherence to standing orders
and procedures, which as a priority, will minimise skin
or eye contact to particulates.
Base parameters: For deploying to the inspection
area, OSI team members will leave the base of operation at a controlled access point where they will put
on coveralls and sturdy footwear. They will also carry
respirators and disposable personal protective equipment including coveralls, eye protection, gloves, and
overboots. At the controlled access point, contamination monitoring and decontamination procedures will
ensure that contaminants are not transmitted back to
the base of operation. Since there may be a potential for dust to be transported by individuals back
into the interior of vehicles, those vehicles are to
be used within the inspection area only. Availability of decontamination showers or vehicle washing
facilities should be considered on a case by case
basis.
Vapour pressure: Saturated Vapour Pressure (SVP)
is an important determinant of the level of vapour from
a material (usually liquid, but rarely, solid) that can
evaporate into air. It represents an equilibrium between
vapours evaporating from the material and vapours
condensing back to the material, and is a function of
temperature. SVP therefore represents the maximum
carrying capacity of air for a particular vapour and
varies from chemical to chemical. However, where
the vapours from volatile chemicals are likely to have

dispersed, the SVP will not be used as a predictor of

risk.
Odour threshold: Exposure of workers to toxic
chemicals is often subjectively described in terms of
the odour of a specific exposure. While this may seem
suitable for chemicals with known odours and well
characterised hazards or risks, it is poor occupational
practice. Many chemicals do not have discernible
odours. In addition, there is little warning potential
for a chemical which has an odour threshold the same
as its exposure standard. About 30% of chemicals for
which a warning property can be calculated fit into
this category. Therefore, odour should not be used as
a measure of exposure, and must not be used in chemicals risk assessments in the absence of information
about suitable warning properties. Further, as many
odour thresholds rely on chemicals with appreciable
vapour pressures, and as chemical exposures in most
OSIs will be from particulates, odour should not be
used as a measure of risk.
Inhalability and respirability: Inhaled particulates are rarely of constant dimensionsmost contain
particles of varying sizes and shapes. The size distribution of many aerosols encountered in occupational
environments approximates to a log normal distribution, which can be statistically characterised by the
median or geometric mean and the geometric standard
deviation.
The relative mass of small particles to large particles
is dependent on the cube of the radius of the particle,
and a toxic effect caused by inhalation of a particulate may be due to a very small proportion of the total
particles present in the aerosol. For this reason, particulates are divided into size specific fractions, based on
their ability to enter and penetrate into the respiratory
system. For a particulate to have a toxic effect in the
respiratory system, it must be inhaled.
The largest inhaled particles in the inspirable fraction, with aerodynamic diameters greater than 30 m,
are deposited in the upper airways of the respiratory
system, which are the air passages from the point of
entry at the nostrils or lips to the larynx (the nasopharyngeal fraction). During nasal breathing, particles
are captured by nasal hairs and by impaction onto the
mucosal lining of the turbinates and upper airways. Of
the particles that fail to deposit in the nasopharyngeal
fraction, the larger particles will deposit in the tracheobronchial airway region and can either be absorbed
(if soluble or capable of dissolution) or be cleared
by mucocilary clearance. This is the tracheobronchial
fraction. The remaining particles not deposited in
the nasopharyngeal and tracheobronchial regions can
penetrate to the alveolar region, the region of gas
exchange with the body. This is the respirable fraction.
There is a discernable range of particle sizes that fall
into nasopharyngeal, tracheobronchial and respirable
fractions (Winder, 2004).

1085

Release, dispersion and re-suspension models:

Dispersion modelling may be necessary to understand what can happen to atmospheric releases of
contaminants and subsequent deposition, combining
information about source of release, nature of chemical hazard and observations of climatic conditions to
develop predictions of how and where releases can be
dispersed. Further, when a relatively insoluble contaminant has been deposited on a surface, it may again
be re-suspended by other activities into the air, thereby
becoming a new contributor to inhalational exposure.
While dust suppression and other exposure mitigation
activities may be in place, OSI team activities may
cause release or re-suspension of contaminants; this
needs to be considered in developing any exposure
models.
Super-imposed effects of air movements: Gases
and vapours are subject to normal molecular movement in still air, and can be moved by the slightest of
air movements. Any such contaminants will be available to workers who are working in the vicinity of
such sources, or if air movements bring contaminated
air into worker breathing zones.
Particulates suspended in air such as mists and
dusts may also be available and may become available
for exposure. For instance, settled particulate materials may become airborne when there is sufficient
wind velocity to lift fine particles from the ground surface. One further source of particulate re-suspension
is the energy imparted to particulates though vehicle
or equipment movement, or in some cases, movement
of people. In general, the susceptibility of the particles
to lift is a function of the presence of any ground cover,
compaction and the moisture content of the ground.
Therefore, wind direction may be a factor in OSI
activities. Wherever possible, team members should
work up wind from known sources of contamination,
and consider other controls and preventive measures
to reduce exposures.
Respiratory rate, Respiratory volume, Minute
volume: Respiratory rate (RR) is one measure of respiratory function, the other being respiratory volume
(RV). Respiratory rate multiplied by respiratory volume gives the minute volume (MV), or the amount of
air inhaled in a minute. The minute volume increases
as work increases and can range from 67 L/min at rest
to over 100 L/min during high work loads (Grandjean,
1988). This can cover a wide range of physiological
functions. Most individuals are unlikely to work at
a very high level, which is seen mainly at an elite
athlete level. However moderate and high levels are
possible in some inspection activities. It should also
be noted that these values are made from individuals
with unimpeded respiration. Work rates in individuals
undergoing moderate or high workloads may require
higher minute volumes than a respirator will allow,
possibly producing physiological stress. Therefore, the

respirator type worn and work rate of team members

should be recorded, so that assessments of exposure
are realistic.
Measures of risk acceptability: An inspection
team will not be subject to routine occupational exposures and hence it is considered that reference to
conventional measures of workplace risk acceptability is not relevant. However, OSI team members will
be subject to recommendations which, as a minimum,
control exposures to a level that do not exceed concentrations considered immediately dangerous to life or
health (IDLH). As a limiting case, the objective should
be a demonstration of arrangements which adequately
prevent against significant harm both during an OSI
and against symptoms/consequences which manifest
themselves after its completion. Accordingly, exposure durations and frequencies need to be modelled and
should reflect the above relevant operational activities
and the contaminated surface area likely to exist. The
mechanism is limited to the inhalation pathway and
so the receptors should include the nasal, bronchial
and pulmonary zones as well as transmission to other
receptor areas within the body.
Factors related on operational activities: These
include
Frequency of Operational Activities The OSI team
activities are dictated by the provisions of the Treaty.
Despite this, their exact execution and frequency
must be formulated on site in response to prevailing or changing operational conditions. Where
team activities have factors that include exposure to
chemical or other risks, it is appropriate that team
members comply with exposure mitigation factors.
One further mitigation factor is that where operational activities include exposure to risks, that the
duration and frequency of operational exposure to
such risks be minimised as much as is reasonably
practicable.
Residence Time This relates to the amount of time
that OSI team members spend exposed to chemical hazards. This can be considered the amount of
time spent in operational activities in contaminated
areas, but may be inaccurate if team members are
moving between areas on a regular basis. It may also
include a residual risk during times spent in areas
of low exposure, where minor exposures may still
occur. As it may not always be apparent that activities have been carried out in previously unknown
contaminated areas, operational logs should contain
enough detail so that where necessary, retrospective
exposure assessments can be conducted.
4

ILLUSTRATIVE RISK CHARACTERISATION

The characterisation of risk requires consideration of

the existing hazard, exposure (including mitigation

1086

factors), and consequence, that is, toxicity/severity.

These are discussed below.
4.1

Exposure modelling

An OSI inspection area may range up to 1,000 square

kilometres in size. This correlates to a square with
sides of 31.6 km or a circle with a radius of 17.8 km.
If a height of five metres is used to define an air volume
(that is, at least twice the height of any human being),
this area becomes five cubic kilometres or five trillion
cubic meters. These values are too huge to be of use
in an assessment of chemical risk, so an alternative
approach, three illustrative air volumes, was developed
(see Table 1)
These were:
Low toxicity air volume, of 1,000 square kilometers
(square with sides of 31.6 km or a circle with a radius
of 17.8 km). This would be used for assessment of
residual or ambient risks;
Moderate toxicity air volume, of one square kilometer (square with sides of 1 km or a circle with a radius
Table 1.

of 0.56 km). This would be used for assessment of

potentially hazardous risks;
High toxicity air volume, to be used to assess point
source contamination, of 100 square metres (square
with sides of 10 m or a circle with a radius of
5.64 m). This would be used for assessment of
specific chemical based risks.

4.2

A simple risk assessment can be made by calculating the minimum amount of chemical that would be
required to be present in a specified air volume (low,
moderate, high) as estimated from its IDLH. While it
is accepted that there are many imponderables in estimating risks under operational OSI situations, if it is
assumed that all the available chemical is present in
the air volume fully mixed, this provides a crude estimate of risk. For example, if 100 g of a chemical is
required to be present to fill a specified air volume
to its IDLH, this will give an approximate measure

Volume modelling data.

Length
km

Width
m

km

Area
m

Height

km2

m2

100 m2
Square
0.01
10.00 0.01
10.00
0.0001
Circle
0.0056
5.64 0.0056
5.64
0.0001
1 km2
Square
1.00
1,000.00 1.00
1,000,00
1.00
Circle
0.56
564.19 0.56
564.19
1.00
1000 km2
Square
31.62
31,622.78 31.62
31,622.78 1,000.00
Circle
17.84
17,841.24 17.84
17,841.24 1,000.00

Table 2.

Exposure assessment

km
0.005

Volume
m

km2

m2

5.00

100.00
100.00

0.01
0.01

10.00 0.00
10.00 0.00

1,000,000
1,000,000

0.01
0.01

10.00 0.005 5,000,000

10.00 0.005 5,000,000

1,000,000,000 0.01
1,000,000,000 0.01

10.00 5.00
10.00 5.00

5000
5000

5,000,000,000
5,000,000,000

Chemical Identity, IDLH values and illustrative risk assessment calculations.

Substance name

CAS
Number

1994 IDLH
(mg/m3 )

High
Toxicity (g)

Moderate
Toxicity (kg)

Low
Toxicity (kg)

Paraquat (mist)
Hydrogen selenide
Acreolin
Chlorinated biphenyls (54% chlorine)
Mercury (elemental vapour or inorganic dusts)
Hydroquinone (vapour)
Lead and its compounds (dust)
Formaldehyde
Manganese (dust)
Phosphoric acid (mist)
Carbon Black

1910-42-5
7783-07-5
107-02-8
11097-69-1
7439-97-6
123-31-9
7439-92-1
50-00-0
7439-96-5
7664-38-2
1333-86-4

1.00
3.32
4.60
5.00
10.00
50.00
100.00
246.00
500.00
1,000.00
1,750.00

0.50
1.66
2.30
2.50
5.00
25.00
50.00
123.00
250.00
500.00
875.00

5.0
16.6
23.0
25.0
50.0
250.0
500.0
1,230.0
2,500.0
5,000.0
8,750.0

5,000
16,600
23,000
25,000
50,000
250,000
5,000,000
1,230,000
2,500,000
5,000,000
8,750,000

1087

of risk, which can be used for prioritisation of more

detailed assessments.
Eleven boundary chemicals representing the
categories mentioned above were selected for this
report (see above). Table 2 shows the identity and differing IDLH values for these chemicals, ranging from
1 mg/m3 to 1, 750 mg/m3 , providing a full range of
IDLH values.
When the three air volumes in Table 1 are used
to estimate the minimum amount of material that
would be present at the IDLH for the eleven boundary
chemicals, the data in Table 2 are obtained.
4.3

Risk characterisation

While a criterion for what constitutes an amount of

material that might be reasonably considered to be
present in an OSI is yet to be determined, a value of
10 kg would not seem implausible, and is used below.
It can be seen that gram amounts of chemicals
(0.5 g for Paraquat) to almost kilogram amounts (875 g
for Carbon black) are required to contaminate an air
volume of 100 m2 to a height of 5 m. Bearing in mind
these are minimum concentrations, these represent
significantly hazardous situations.
However, in moderate or even low toxicity air
volumes, kilogram amounts (for example, 5 kg of
Paraquat in a moderate air volume) are required to
be present, representing a situation of relatively low
hazard.
This being the case, it is possible to conclude that in
the absence of information or evidence of presence of
industrial chemicals, OSI team members are unlikely
to be at risk from exposure to toxic chemicals. This
may be sufficient to allow establishment of an operational principle. However, where such information or
evidence is available, then a more stringent risk assessment should be conducted and appropriate control and
preventive measures should be adopted.
5

DISCUSSION

Chemical exposures may be encountered in a range of

occupational and environmental situations. Some of
these exposures are definitionally beyond the scope
of traditional exposures encountered in workplaces.

If such situations arise, conventional risk acceptability

criteria, such as exposure standards, may not be
appropriate to estimate risk. Nevertheless, the responsibility of the employer to provide and maintain a
safe working environment still exists, regardless of
the unusualness of the exposure or the difficulty in
assessing its dangers.
One such exposure is of inspectors or inspection
teams confronted by a potentially dangerous environment of unknown chemical composition, for example,
teams conducting CTBT OSIs as a part of verification measures upon request by any State Party and
approval by the Executive Council of the CTBTO once
the Treaty enters into force.
A risk assessment process was developed based on
a toxicity assessment component based on recommendations for exposures that are immediately dangerous
to life and health, and an exposure assessment component based on plausible exposure circumstances.
When the risk assessment process was used to characterise risks, it was possible to show that for most
foreseeable risk situations, risks from exposure to
industrial chemicals were quite low.
The risk assessment process was sufficiently robust
to allow assessment of a number of exposure scenarios, and, combined with operational procedures and
recommended risk mitigation activities, allows a preliminary assessment of risks sufficient to conclude that
in the absence of information or evidence of presence
of industrial chemicals, Inspection Team members
are unlikely to be at risk from exposure to toxic
chemicals.

REFERENCES
Grandjean, E. 1988. Fitting the Task to the Man: A Textbook of
Occupational Ergonomics, fourth edition. London: Taylor
and Frances, pp. 8394.
NIOSH. 1994. Documentation for Considered Immediately Hazardous to Life or Health Concentrations. NTIS
Publication No PB-94-195047. Cincinnati: US National
Institute for Occupational Safety and Health. At: http://
www.cdc.gov/niosh/idlh/idlh-1.html
Winder, C. 2004. Toxicity of Gases, Vapours and Particulates.
Chapter 15 in: Occupational Toxicology, second edition,
Winder, C., Stacey, N.H., editors. Boca Raton: CRC Press,
pp. 399423.

1088

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Comparison of different methodologies to estimate the evacuation radius

in the case of a toxic release
M.I. Montoya & E. Planas
Center for Technological Risk Studies, Universitat Politcnica de Catalunya, Barcelona, Spain

ABSTRACT: A comparison of different methodologies used to estimate the evacuation radius, is presented.
The evaluation is accomplished doing some modifications in the way of computing indoor concentration (taking
into account the profile of outdoor concentration and the possible sorption into indoor surfaces) and toxic load.
The casualties probability estimation is made using the probit analysis, but a parallel analysis to assess the
effectiveness of shelter in place, based on the AEGL-3, is also made with the aim of comparison and to confirm
the results obtained with the probit analysis. Five substances of different toxicity (chlorine, methyl isocyanate,
acroleine, hydrogen fluoride and mustard gas) are used to study the performance of one or another methodology,
finding that for cumulative substances there is an overestimation with the method used by Catalonian government.
Concerning meteorological conditions, larger distances were obtained when increasing the atmospheric stability.

INTRODUCTION

Dispersion of toxic gas clouds can be originated by

diverse events, including accidental releases in the
industry or during transportation of hazardous materials, and biological or chemical warfare agents attacks.
One protection way against this phenomenon is to
shelter in a building and wait until de toxic plume
has passed. Since most people spend the majority
of their daily time inside buildings they could stay
indoors, where the concentration of the toxic substance
is supposed to be lower than outside and thus, the
toxic load (TL) to which people are exposed is lower.
Even though, there is not a unified or totally reliable
method to determine the effectiveness of this protection method, and thus, there are several methodologies
used by authorities and local emergency commissions
which underestimate or overestimate the evacuation
radius, leading to unexpected deaths, health affections
or unnecessary efforts, respectively.
As community protection measures, Catalonian
government has established three actuation zones:
alert, intervention and LC01, based on outside concentrations (PlaseQcat 2005). The first zone refers to
the region in which accidents consequences, although
perceptible by the community, does not justify intervention. The second is a zone in which those consequences generate such a level of damages that
immediately protection measures might be taken. The
last zone involves the area in which 1% casualties are
expected at the exterior. The first two zones are defined
using the threshold values given by the Spanish

government (Directriz bsica 2003), which are the

Acute Exposure Guidelines Levels (AEGL-1 and
AEGL-2, respectively), the Emergency Response
Planning Guidelines (ERPG-1, ERPG-2) or the
Temporary Emergency Exposure Limits (TEEL-1,
TEEL-2) if the first ones (AEGL) are not available
for the substance studied. The LC01 involves the estimation of a percentage of casualties which is usually
accomplished through a vulnerability study using the
probit analysis (Purple book 1999). In addition to this
zones, and taking into account that protection measures in the case of a toxic gas dispersion recommend
shelter, emergency managers should have established
the radius within shelter in place is effective and thus an
evacuation radius must be defined. The criteria used by
Catalonian government to calculate this radius is that
casualties inside the building, within this distance, are
0.1% (PlaseQcat 2005).
For the vulnerability study one of the most common techniques is the probit analysis which relates
the magnitude of the incident with the damage caused
(Casal et al. 1999a, El Harbawi et al. 2008, Geeta
et al. 1993). In the case of toxic dispersion, the magnitude of the incident is the TL received by people
exposed. This magnitude is used to calculate the probit variable and then, the casualties probability. The
TL depends on the time and the concentration to which
people are exposed, that is why the estimation of the
indoor concentration (Ci ) is so important.
Indoor concentration depends mainly on the air
exchange rate of the building -since it conditions
the entrance of outdoor air- and on the sorption

1089

phenomena that could take place in indoor surfaces

(Yuan 2000). In this work a comparison of four
methodologies to estimate the evacuation radius: the
method used by Catalonian government (Method 1),
two modifications of this method changing the way of
estimating Ci , without sorption (Method 2) and with
sorption (Method 3), and a fourth one changing the
air exchange rate used (Method 4); is outlined, and an
analysis of the response behaviour due to external conditions like substance type and atmospheric stability
is also presented.
2

Table 1.

Simplifications for indoor concentration models.

Model

Simplifications

Reference

Without
adsorption

ka = kd
= k1 = k2 = 0

Jetter & Whitfield

2005

Deposition

kd = k1 = k2 = 0

Yuan 2000,
Karlsson 1994,
Casal et al. 1999b

One-sink

k1 = k2 = 0

Karlsson & Huber

1996, Jrgensen &
Bjrseth 1999,
Jrgensen et al.
1999

Sink
diffusion

k1 = k2

Jrgensen et al.
2000, Singer et al.
2004, Zhang et al.
2002a

INDOOR CONCENTRATION MODELS

Mathematical models proposed for the estimation of

indoor concentration inside a residence, are based on a
mass balance assuming perfect mixing between fresh
and inside air. The most complex case considers pollutants entering with air due to infiltration and mechanical ventilation, these pollutants could be partly
retained by the corresponding filtration filters, and
inside the building they could be eliminated through
adsorption and diffusion into the materials presents
(Karlsson 1994, Karlsson & Huber 1996, Shair &
Heitner 1974). Even though, in the case of sheltering it is assumed that mechanical ventilation systems
are turned off and the only ventilation path is infiltration (Yuan 2000). With the above simplification, these
models can be classified into four categories:
1. Without adsorption (without adsorption model).
2. Adsorption rate constant, without desorption
(deposition model).
3. Variable adsorption/desorption (one-sink model).
4. Variable adsorption/desorption with internal diffusion (sink diffusion and two-sink models).
In the first case indoor concentration would be only
lower than outdoors by a lapse of time equivalent to
the inverse of the air exchange rate. In the others cases
a lower concentration is achieved since adsorption on
materials is taken into account (Yuan 2000, Karlsson
1994, Casal et al. 1999b, Liu & Nazaroff 2001).
However, desorption can causes a residual indoor gas
concentration over an extended time, increasing indoor
exposure and approaching to outdoors if the place is
not adequately ventilated after the toxic gas plume
has passed (Karlsson and Huber, 1996). Mathematical representations of these models could be derived
from the two-sink model (Singer et al. 2004, Singer
et al. 2005a), presented in Equations 13, applying the
simplifications shown in Table 1.

N 

Aj
dCi
= no Co no Ci
dt
V
j
(ka, j Ci kd, j m1, j k1, j m1, j + k2, j m2, j ) (1)

dm2, j
dm1, j
= ka, j Ci kd, j m1, j
,
dt
dt
dm2, j
= k1, j m1, j k2, j m2, j ,
dt

j = 1...N

j = 1...N

(2)
(3)

where Co and Ci (mg/m3 ) refer to the outdoor and

indoor concentration, respectively; no to the air
exchange rate (h1 ); A to indoor surfaces (m2 ); V
to indoor volume (m3 ); ka and kd are the adsorption
(m/h) and desorption (h1 ) constants, respectively; k1
and k2 are the mass transfer coefficients to and from
the embedded sink (h1 ), respectively; m1 and m2 to
the amount of gas deposited per unit area on the surface
and in the embedded sink (mg/m2 ); and j represents
the different material present.
Although all these models could be applied to
any gas, they have been specially used in assessing
indoor concentration of common pollutants that affect
indoor air quality, and therefore adsorption and desorption parameters (ki ) of substances like CO2 , O3
or VOCs have been more studied than toxic gases.
That is why there is a lack of adsorption and desorption parameters for toxic gases, which increase
with the complexity of the model, and thus limits its application. Since the main difference in the
indoor concentration behaviour applying the 5 models
appears between the model without adsorption and the
models that considers adsorption, as shown in figure 1
for the case of sarin (adsorption parameters taken from
Singer et al. (2005b)), the development of this work
would just consider the model without adsorption and
the deposition model (in representation of the sorption
models.

1090

to evaluate the TL reduction gained as being indoors.

In the first case, there are some indicators known as
safety factors, they represent the constant by which
the exposure concentrations can be multiplied without
exceeding the TLL (Chan et al. 2007a). In this work
the TLL was taken as AEGL-3 at 30 min, since it represents a concentration above which general population,
including susceptible individuals, could experience
life-threatening health effects or death, and could be
used as a reference to compare evacuation radius.
The indoor and outdoor safety factors (SFi and SFo ,
respectively) were determined trough Equations 7, 8.

SFi =

SFo =
Figure 1. Indoor concentration of sarin. Models: Without
adsorption Deposition One-sink - - - - - - Sink
diffusion Two-sink - .

METHODOLOGY

As mentioned in the first section, casualties percentage is estimated through the probit analysis. The probit
variable (Y ) is related to the probability (P) according
to the following expression:



1
Y 5
P=
1 + erf

2
2

(4)

(7)
 1n
(8)

For the second type of indicators, the toxic load

reduction factor (TLRF) and the safety factor multiplier (SFM) were used. The TLRF relates the TL that
is avoided by being indoors with the TL that is perceived outside (Eq. 9), a value near 1 means that a good
TL reduction is achieved while a value near 0 means
an ineffective shelter, like being outdoors. The SFM
(Eq. 10) represents the relation between the indoor and
outdoor safety factors; a high value of the SFM indicates an effective shelter while a value approaching 1
means an ineffective shelter.
TLRF (t) =


(5)

 1n

TLL
TLo (t)

 t2

The probit variable for the case of a toxic dispersion

is expressed as:
t

Y = A + B ln [C (t)]n dt

TLL
TLi (t)

SFM =

TLo
TLi

t
Con dt 0 1 Cin dt
TLi
=1
t
TLo
(1) 0 2 Con dt

(9)

 1n
(10)

where A, B and n are constants that depends on the

substance (Purple book 1999, Casal et al. 1999a),
while C(t) refers to the concentration and the integral
of this concentration in time, to the toxic load (TL)
received (Eq. 6).
t
TL =

[C (t)]n dt

The methodology currently used by the Catalonian government to determine the evacuation radius
given an external concentration is represented by the
following steps (PlaseQcat 2005):
1. Estimate indoor concentration at different distances
with Equation (11), using an air exchange rate (no )
of 2 h1 .

(6)

1 exp (no t)
Ci
=1
Co
no t

In this work, in addition to the probit analysis, shelter effectiveness was also estimated. To assess shelter
in place effectiveness there are two type of indicators:
ones that are used to evaluate if a toxic load limit (TLL)
is exceeded, it means a limit from which certain health
affections could take place; and others that are used

(11)

2. Calculate the TLi assuming a constant indoor

concentration for each distance (Eq. 12).

1091

TLi = Cin t

(12)

Table 2.

Parameters of the three methodologies.

Parameter

Method 1

Method 2

Method 3

Method 4

no

2
1 exp (no t)
Ci
=1
Co
no t

1
dCi
= no (Co Ci )
dt

1
dCi
= no Co no Ci
dt
A
ka Ci
V
t n
Co dt

1
1 exp (no t)
Ci
=1
Co
no t

Ci

TLo
TLi

Con t
Cin t

t
0
t

Con dt

0
t

Cin dt

Cin t

Con dt

3. Estimate Y using Equation 13 and P from Equation 4.

Y = A + B ln(Ci t)

Con t

(13)

4. The distance at which P equals 0.001 is considered

the evacuation radius.
Methods 2 and 3 differ from this one specially in the
way of estimating Ci . In method 2, the model without
adsorption is used, while the deposition model is utilized within method 3. In both cases, the air exchange
rate is assumed to be 1 h1 , since in a previous study
it was found that more of 92% Catalonian dwellings
exhibit air exchange rates lower than one. The TL, Y
and P estimation in these methods are computed using
Equations 46. In method 4 instead, the only difference with method 1 is the air exchange rate which is
assumed as 1 h1 . Table 2 summarized the parameters
used in the four methods.
Outdoor concentration was estimated with the
Gaussian dispersion model presented by Seinfeld &
Pandis (1998). The calculus was made downwind,
at a height of 2 m, neutral stability, wind speed
of 4 m/s, emission source at ground level, emission rate of 33.33 kg/min, ground roughness of 1 m
(urban, industrial areas) and emission duration of
30 min.
For method 3, where adsorption over indoor surfaces is taken into account, A/V relation was taken as
3.5 m1 and ka as 1 104 m/s. This last value corresponds to the average deposition velocity of gases in
residential buildings presented by Karlsson (1994).
Substances chosen for this analysis were chlorine (Cl), methyl isocyanate (Iso), acroleine (Acro),
hydrogen fluoride (HF) and mustard gas (MG). These
substances were selected due to its toxicity and because
they have different values of n, so the behaviour of this
parameter could be studied. To assess the influence
of this parameter on the estimation of the evacuation radius, the value of n was varied between 0.5

Table 3.

Parameters of the different substances.

Substance

A1

B1

n1

AEGL-3 30 min
mg/m3 (ppm)2

Cl
Iso
Acro
HF
MG

6.35
1.2
4.1
8.4
5.473

0.5
1
1
1
1.13

2.75
0.7
1
1.5
2

81.194 (28)
0.933 (0.4)
5.732 (2.5)
50.715 (62)
2.7 (0.41)

Constants taken from Serida Database 1.3 (1999), concentrations must be in mg/m3 and time in min.
Values taken from EPA (http://www.epa.gov/oppt/aegl/
pubs-/results56.htm).
3 Values estimated from lethal dosages Hartman (2002).
2

and 3 for isocyanate, acroleine and mustard gas, and

the corresponding evacuation distances were obtained.
Probit constants an AEGL-3 for these substances are
shown in Table 3. Probit constants for mustard gas
were estimated from lethal dosages (LC50 and LC01 )
reported by Hartman (2002) fixing n = 2, since the
constants included within the Serida Database for this
substance correspond to the default values (Espinar
2005), leading to an overestimation of the evacuation
radius.
All calculations were made for a lapse time of
30 min from the beginning of the release.

RESULTS AND DISCUSSION

Evacuation radiuses estimated with the methodologies

for each substance are shown in Table 4. Distances
obtained with method 1 are always higher than those
obtained with methods 2 and 3, increasing with the
toxicity of the substance. However, if Catalonian government methodology is applied using an air exchange
rate of 1 h1 (method 4), the difference decrease, specially with chlorine and hydrogen fluoride that are the

1092

Table 4.

Results obtained for the evacuation radius.

Substance

Method

Evacuation
radius (m)

Chlorine

Method 1
Method 4
Method 2
Method 3
Method 1
Method 4
Method 2
Method 3
Method 1
Method 4
Method 2
Method 3
Method 1
Method 4
Method 2
Method 3
Method 1
Method 4
Method 2
Method 3

203
149
163
144
4695
3300
1997
1891
716
518
481
437
223
163
167
150
1280
910
890
800

Methyl
isocyanate

Acroleine

Hydrogen
fluoride

Mustard
gas

TLL

SFo

SFi

SFM

TLRF

Po

Pi

5.35E + 06

0.25214
0.14718
0.17187
0.13985
0.52076
0.30177
0.21134
0.18845
0.1553
0.0895
0.0848
0.0715
0.1855
0.1075
0.1140
0.0944
0.1918
0.1112
0.1152
0.0939

0.68539
0.69079
0.6827
0.68389
1.4156
1.4164
1.4146
1.4145
0.4222
0.4202
0.4233
0.4211
0.5043
0.5044
0.5075
0.5079
0.5214
0.5221
0.5264
0.5136

2.7183
4.6935
3.9721
4.8903
2.7183
4.6935
6.6935
7.5062
2.7183
4.6935
4.9908
5.8881
2.7183
4.6935
4.4511
5.3826
2.7183
4.6935
4.6102
5.4701

0.93607
0.98576
0.97747
0.98728
0.50341
0.66119
0.73573
0.7561
0.63212
0.78694
0.79963
0.83016
0.77687
0.90165
0.89351
0.91992
0.86466
0.9546
0.9529
0.9665

4.37%
16.62%
11.85%
18.44%
0.84%
2.23%
3.93%
4.66%
1.83%
6.19%
6.87%
9.43%
5.68%
22.26%
19.70%
28.48%
18.12%
61.32%
58.29%
74.55%

0.10%

28.587

171.96

10835

218.7

less toxic substances, for which the distances are even

minor than the ones obtained with method 2, subestimating it. It could also be seen that in almost all
cases the SFo is lower than 1, as it was expected
since the TLL is supposed to be exceeded outside and
thus sheltering or evacuation might be advised to the
population. For each substance the SFi is the same
since the TLi must be the same to match the evacuation criteria (P = 0.001) as presented in Table 4.
This parameter was also expected to be under or
around one, since from AEGL-3 definition, no one
would be in risk of death if this threshold value is not
exceeded; even though SFi for methyl isocyanate is a
bit higher than one and thus evacuation distances are
overestimated.
Plots of Pi and SFi as a function of distance for
all methods were obtained for each substance, Po and
SFo were also plotted (e.g. Acroleine in Figure 2).
The behaviour of P is always decreasing with distance while the SF is always increasing. There is
also a big difference between Po and the Pi estimated
with all the methods, which assures that an important toxic load reduction, is achieved with shelter.
For all substances, the Pi and the SFi estimated with
method 3 were the lowest and the highest, respectively; while the results obtained with method 1 were
the opposite. In the case of methods 2 and 4, there
is a special behaviour related with the toxicity of the
substance, for the more toxic subtances (acroleine,

0.10%

0.10%

0.10%

0.10%

isocyanate, mustard gas) lower distances were

obtained with method 2, while for chlorine and
hydrogen fluoride with method 4.
The results of the variation of the evacuation radius
as a function of n for acroleine, methyl isocyanate and
mustard gas are presented in Figure 3. Two behaviours
could be identified, one in which the distance increase
with n and the other in which it decrease. The first
case takes place when the concentration needed to
reach the TLi that satisfies the probit function for
a defined probability is higher than 1 mg/m3 (e.g.
acroleine and mustard gas); in that case, the evacuation radius increase with n. The second situation
takes place when the above mention concentration is
less than 1 mg/m3 (e.g. Acroleine method 1), then,
the evacuation radius decrease with n. Particularly,
method 1 for methyl isocyanate (Iso Method 1) follows
the second behaviour, since it is necessary that indoor
concentration (assumed constant in this method) stay
below 1 mg/m3 to reach a TLi of 22.4 (mg/m3 )0.7 min,
which is the one required by the probit analysis to
have a probability of 0.1%. Such low concentrations
(<1 mg/m3 ) would only be needed when dealing with
a really high toxic substance.
Concerning the meteorology: atmospheric stability and wind speed, the evacuation radius obtained
for more stable atmospheres (class E, 2 m/s) were as
large as twice the distance obtained in a neutral atmosphere, while the radius obtained in a more instable

1093

Figure 2.

Probability and safety factor behavior with distance for acroleine.

Figure 3.

Variation of evacuation radius with n.

1094

to compare the values obtained and then to give more

reliability to the results found.
For methyl isocyanate, evacuation radius obtained
with method 1, even using an air exchange of 1 h1
(method 4), are really high compared with those
obtained with the others methodologies. The main
reason relies on the difference between the methods
analyzed to estimate the toxic load. If it is estimated
through the integration of indoor concentration profile, a closer appreciation of reality could be obtained
than with the integration of a constant. This is especially important in the case of cumulative substances
(n 1) like methyl isocyanate, were the toxic load
depends more on the behaviour of indoor concentration than on maximum indoor concentration, which is
the case of peak chemicals (2 < n < 3.5) (Mannan &
Lilpatrick 2000).

REFERENCES
Figure 4.
stability.

Variation of evacuation radius with atmospheric

atmosphere (class B, 2 m/s) is smaller, like half the

distance in a neutral atmosphere, as presented in
Figure 4 for acroleine. This behaviour is due to the
quick dispersion and turbulence present in an instable atmosphere which makes that the toxic substance
disperses better and faster.

CONCLUSIONS

One of the most important limitations when assessing the indoor concentration is the lack of knowledge
about the air exchange rate, which, as seen with
methods 1 and 4, is a very important parameter that
conditions the evacuations radius obtained and could
lead to under or overestimations.
For the substances studied, method 3 was the one
that predict the lowest evacuation radius and the highest safety factor for the same distance; the other
methods gave more conservative results. From a theoretically point of view, method 3 is also the most
rigorous, since it incorporates an indoor concentration
model that takes into account deposition over indoor
surfaces, and assesses the toxic load according to the
profile of the indoor concentration. Therefore, this
methodology should be the one that better represents
the reality. Even though, since there is not experimental data available to validate the results obtained with
the four methods, a final statement could not be given.
When estimating and evacuation radius, a parallel
analysis with another guideline level (like AEGL-3,
as shown here) is a good tool, since it could be used

Casal J., Montiel H., Planas E., Vilchez J.A. 1999a. Anlisis del Riesgo en Instalaciones Industriales. Barcelona:
Ediciones UPC.
Casal J., Planas-Cuchi E., Casal J. 1999b. Sheltering as a
protective measure against airborne virus spread. Journal
of Hazardous Materials., A68: 179189.
Chan W.R., Nazaroff W.W., Price P.N., Gadgil A.J. 2007a.
Effectiveness of urban shelter-in-placeI: Idealized conditions. Atmospheric Environment. 41, 49624976.
Directrz bsica 1196. 2003. Ministerio de trabajo y asuntos
sociales. Instituto nacional de seguridad e higiene en el
trabajo.
El Harbawi M., Mustapha S., Choong T.S.Y., Abdul Rashid
S., Kadir S.A.S.A., Abdul Rashid Z. 2008. Rapid analysis
of risk assessment using developed simulation of chemical
industrial accidents software package. Int. J. Environ. Sci.
Tech., 5 (1): 5364.
Espinar J. 2005. Estudi comparatiu dels parmetres de
toxicitat i vulnerabilitat de substncies perilloses. Undergraduate final work. Universitat Politcnica de Catalunya.
Geeta B., Tripathi A., Marasimhan S. 1993. Analytical
expressions for estimating damage area in toxic gas
releases.
Hartman H.M. 2002. Evaluation of risk assessment guideline
levels for the chemical warfare agents mustard, GB, and
VX. Regulatory Toxicology and Pharmacology, 35 (3):
347356.
Jetter J.J., Whitfield C. 2005. Effectiveness of expedient
sheltering in place in a residence. Journal of Hazardous
Materials, A119: 3140.
Jrgensen R.B., Bjrseth O. 1999. Sorption behaviour of
volatile organic compounds on material surfacesthe
influence of combinations of compounds and materials
compared to sorption of single compounds on single
materials. Environment International, 25 (1): 1727.
Jrgensen, Rikke B., Bjrseth, Olav., Malvik, Bjarne.
1999. Chamber testing of adsorption of volatile organic
compounds (VOCs) on material surfaces. Indoor Air,
9: 29.

1095

Jrgensen R.B., Dokka T.H., Bjrseth O. 2000. Introduction of a sink-diffusion model to describe the interaction
between volatile organic compounds (VOCs) and material
Surfaces. Indoor Air, 10: 2738.
Karlsson E. 1994. Indoor deposition reducing the effect
of toxic gas clouds in ordinary buildings. Journal of
Hazardous Materials, 38: 313327.
Karlsson E., Huber U. 1996. Influence of Desorption on
the Indoor Concentration of Toxic Gases. Journal of
Hazardous Materials, 49: 1527.
Liu D., Nazaroff W.W. 2001. Modeling pollutant penetration
across building envelopes. Atmospheric Environment, 35:
44514462.
Mannan M.S., Lilpatrick D.L. 2000. The pros and cons of
shelter in place. Process Safety Progress, 19 (4): 210218.
PlaseQcat. 2005. Pla demergncia exterior del sector qumic
de Catalunya. Generalitat de Catalunya. Secreteria de
Seguretat Pblica, Direcci General dEmergencies i
Seguretat Civil.
Purple book. TNO. 1999. Guideline for quantitative risk
assessment. (CPR18E). Committee for the prevention of
disasters. The Hauge.
Seinfeld J.H., Pandis S.N. 1998. Atmospheric Chemistry and
Physics: From Air Pollution to Climate Change. Wiley
Interscience Publication, New York, NY, 926950.

Shair F.H., Heitner K.L. 1974. Theoretical Model for Relating Indoor Pollutant Concentrations to Those Outside.
Environmental Science and Technology, 8 (5): 444451.
Singer B.C., Revzan K.L., Hotchi T., Hodgson A.T., Brown
N.J. 2004. Sorption of organic gases in a furnished room.
Atmospheric Environment, 38: 24832494.
Singer B.C., Hodgson A.T., Destaillats H., Hotchi T., Revzan
K.L., Sextro R.G. 2005a. Indoor sorption of surrogates for
sarin and related nerve agents. Environmental Science and
Technology, 39: 32033214.
Singer B.C., Hodgson A.T., Hotchi T., Ming K.Y.,
Sextro R. G., Wood E.E. 2005b. Sorption of organic
gases in residential bedrooms and bathrooms. Proceedings, Indoor Air: 10th International Conference on Indoor
Air Quality and Climate, Beijing, China.
Yuan L.L. 2000. Sheltering Effects of Buildings from Biological Weapons. Science and Global Security, 8: 329355.
Zhang J.S., Zhang J.S., Chen Q., Yang X. 2002a. A critical
review on studies of volatile organic compound (VOC)
sorption on building materials. ASHRAE Transactions,
108 (1): 162174.

1096

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Conceptualizing and managing risk networks. New insights for risk

management
R.W. Schrder
Witten/Herdecke University (UW/H), Witten, Germany

ABSTRACT: Risks are connected with one another by cause-related and/or by effect-related reasons. Thus,
two risks are interdependent as they are caused by interrelated events, or because the events inducing them impact
on interconnected targets. While the former relations can be represented by a cause-related risk network, the
latter ones form an effect-related network. Together both grids merge in a comprehensive risk network, requiring
changes and possibly new ways in risk management. However and at the same time, such a risk network entails
new approaches for the risk management process. This paper examines the nature of risk networks to provide
insights into the risks interrelations, their management, and their implications for the risk management process.
Consequently and contrary to the risk management performed so far, the paper shows that the information on
risk networks is desirable to effectively identify, evaluate and mitigate all risks encountered.

INTRODUCTION

In a context of an increasingly dynamic and uncertain

environment, risks have gained in importance. However, all risks are normally considered separately from
one another, disregarding any dependencies among
them. These relations may either stem from the riskinducing events or from the affected targets. Both
forms of dependencies among risks incorporate supplementary effects hampering or facilitating the risks
management.
This paper seizes the suggestions implied by the
risks dependencies, thereby providing insights into an
expanded risk management process. The paper is organized in four major sections. The second section characterizes networks and translates their particulars into
risk networks. Moreover, it addresses the constituents
of a risk network, i.e. the risks and alternative forms of
their dependencies, distinguishing between cause- and
effect-related relations. Following up this differentiation, the third section describes the risk management
process focusing on (a) the cause-related and (b)
effect-related dependencies among risks. To draw to a
close, the fourth section integrates the presented management processes and the risk management process
itself, leading to the management of risk networks.
Finally, the paper discusses the insights attained.
2

RISK NETWORKS

A network is a meshwork of relations, cause-andeffect chains between extensively autonomous partners,

e.g. individuals, items (Mller 2006). Transferring

this definition to risks, the risks represent the linked
partners. Each one of them may be understood as
a potential target deviation initiated by a possibly
realized event (Schrder, in prep.).
rij = pe  tj

(1)

where pe = occurrence probability of the event ei ;

rij = risk; tj = induced target deviation.
Hence every risk (rij ) itself constitutes a causeand-effect chain. These chains are interrelated, as
supposedly all targets (tj ) interact and as the riskinducing events (ei ) are connected. Consequently, four
cases of risk dependencies exist (see fig. 1):
1. The risks are independent, if neither the riskinducing events nor the affected targets are interrelated.
2. If two risks stem from either the same or from
connected events, they are cause-related.
3. Two risk-inducing events have an impact on the
same or on interrelated targets. In this case, the risks
considered may be referred to as effect-related.
4. This case merges the second and the third one,
assuming that the risk-inducing events and the
affected targets both are interrelated.
All in all, every constitutional element of a network
is at hand, which is why the risks may be summarized in
their own network. As shown in figure 2, such network
affiliates all relevant risks regarding their cause- and
their effect-related dependencies, supposing that any
one of these relations is existent.

1097

Effect-related
f
r
d ddependencies
p
ci s
((casee 3)

Two risk-inducing events are equivalent, if their

probabilities of occurrence correspond and if these
probabilities are concordant with the probability of
a jointly realization of both events. Yet, from the
equivalency of the events does not follow that they
are coherent.
Such coherence is however given, if one event forms
a part of the other event. If the first event is realized,
the occurrence of the latter event is given. However,
this sequence is irreversible.
Opposed to these forms of a positive event dependency, a jointly realization of two negatively related
events is impossible. If, for example, two events are
incompatible, the realization of one event prevents that
the second event also occurs. However it is possible
that none of the events happens. This conjoint nonrealization is ruled out, if one event represents the
opposite of the other event. Figure 2 gives a review
of possible cause-related risk dependencies.

Cause-- and eeffect-related

rel
dependencies
e
es

ddependent
e t

r11 :

e1

t1

r11 :

e1

t1

r22 :

e2

t2

r22 :

e2

t2

Independence
e

Cause-related dependencies

((casee 1))

independent
pe
t

Effect-related
f c
ted componentss

(casee 4))
(c

(case 2)

r11 :

e1

t1

r11 :

e1

t1

r22 :

e2

t2

r22 :

e2

t2

independent

dependent

Cause-related
C
u -r e components
co
nt

Figure 1.

Risk dependencies.

2.2

Figure 2.

2.1

Risk network.

Cause-related risk dependencies

Cause-oriented risk relations specify in which way two

risk-inducing events interact. Here we may distinguish
between (a) a positive and (b) a negative relationship.
In the former case, the occurrence probability of any
event positively connected with the event assumed as
realized rises. In the latter case a negative relationship between both events is assumed, the occurrence
probability of the non-realized event then declines.
Referring to the statistical theory of probabilities
various specific forms of positive event dependencies
exist (Keel 1995; Bamberg & Baur 2001):
The same event causes deviations from more than
one target. Regarding their causes, the resulting
risks then are identical.

Effect-related risk dependencies

Even those risks that are induced by independent

events may be interrelated. Therefor responsible are
the effects caused on different targets which are
interdependent.
Regarding the number of the objectives where the
risk-inducing events cause a deviation, we have to distinguish between two cases. Firstly, the events can refer
to the same target. Then an effect-related identity of the
induced risks is given. As long as this dependency does
not coincide with a cause-related identity, no difficulties arise. Otherwise the same risk is repeatedly and
thus, spuriously considered. In the second case, the
risk-inducing events cause deviations from the desired
values of various targets which are connected.
Considering the impact ratio, the second case can be
subdivided into (a) a dependency of the targets striven
for and (b) their interdependency. If the targets are
interdependent, they are interactive, i.e. the level of
achievement of one target influences the achievement
of the other objective. In the case of dependent risks
or more explicitly targets, the deviation induced at one
target affects the achievement of the second target.
However, this deviation does not retroact, i.e. it does
not reciprocate an influence on the first target.
Regardless of the impact ratio, the effect-related
dependencies can act in the same or different directions. While in the former case a complementarity of
the affected targets is implied, the latter case supposes
a goal conflict. Besides, the two relevant targets can
adopt a neutral position to one another.
Figure 4 summarizes the diversity of effect-related
dependencies among risks. Together with the causerelated dependencies these impart knowledge of the
multiplicity of relations in a risk network, entailing
new tasks for the risk management process.

1098

TRANSFERRING THE RISK MANAGEMENT

PROCESS TO RISK DEPENDENCIES

Based on a risk strategy presenting a guideline for the

risks management, the business objectives may be
derived. These targets provide a starting point for the
risk identification for which each company can revert
to the assistance of weak signals (Ansoff 1975).
As soon as all risks are identified, their evaluation
begins. Therefore the company has to determine the
occurrence probability of the risk-inducing event and
the extent to which the affected target will be missed
due to the events realization.
Subsequently to the evaluation, the company has to
decide on an appropriate measure to mitigate the risk.
The company can either avoid, reduce or transfer the
risks to someone else, e.g. its clients. Additionally it
is feasible to simply take and accept the risk together
with its potential consequences.
In the following step the outcome of each risk is
monitored. The findings the company hereby gets may
result in modifications of the risk strategy and/or the
objectives aimed at.
To identify deficiencies in the risk management
process it is attended by a monitoring or a risk
accounting system. Lck has developed a picture of
these steps of a risk management process (Lck 1998,
fig. 4), to which the following considerations refer.
3.1

Managing effect-related risk dependencies

The management of the effect-related risk dependencies starts with the formulation of the target system,
as the definition of the former relations is therein
included. Thus, a separate identification of these
dependencies is unnecessary and the second phase of
the process in figure 4 can be skipped. It is rather
examined if the formulated target system is complete
and consistent. Thereby, the company has to especially attend to the supposed dependencies among the
relevant targets. If any of the defined relations is inconsistent, an adjustment of the target system is obligatory.
As soon as the target system is accurately formulated, the evaluation of the incorporated dependencies
begins. This does not only mean that the direction of
the induced effects is established, but also that their
strength, e.g. a fairly small deviation from one target
may lead to a substantial missing of another aim.
Regarding a mitigation of the effect-related risk
dependencies, the company may either accept them
or rearrange its target system. When monitoring these
relations, the company must bear these alterations in
mind. The revision not only has to again revise the
consistency of the dependencies, but also to decide
whether the actions taken to mitigate them were appropriate. If any deficiencies become apparent, the target
system must again be reviewed and, if necessary,

assimilated. Then, the described management process

recommences.
3.2

Managing cause-related risk dependencies

According to the depicted risk management process the management of cause-related risk dependencies starts with an identification of potentially
risk-inducing events. For this purpose, the company
can rely on pattern recognition techniques. A pattern
represents the combination of an objects features and
particulars (Mertens 1977; Heno 1983). In the present
context, we can treat the risk dependencies as the
object to analyze.
For this analysis several techniques seem
appropriate:
The nearest-neighbor principle (Schrmann 1996),
The discriminant analysis (Backhaus et al. 2006),
A multinomial logistic regression (Anre et al.
1997; Backhaus et al. 2006) and
Artificial neural networks (Rojas 1996; Backhaus
et al. 2006).
None of these techniques clearly stands out from the
others as several comparisons by pairs show (Schrder
thereto provides an overview of studies, Schrder
2005). Thus, each company has to decide on its
own which technique to apply for identifying the
cause-related risk dependencies.
Having recorded all risk-inducing events, their
evaluation starts. Firstly the occurrence probability
of each event has to be verified. Than the effect of
one events supposed realisation on other so far nonrealised events has to be determined. Therefor the
company can apply conditional probabilities.
p(ei |eh ) =

p(ei eh )
p(eh )

(2)

where i  = h.
The value of these probabilities depends on the
relation the two events are in (fig. 3). If they are positively dependent, this relation causes a rise of the
non-realized events probability. For example, if the
two events are equivalent regarding their occurrence
probabilities, than the realization of one of the two
events induces an entrance of the second event.
In case of negative dependency, the realization of
one event diminishes the probability of the other event.
This reduction can go so far as a probability of zero for
the second event, i.e. its realization is no longer possible. Table 1 lists the conditional probabilities subject
to the existent dependency of the two relevant events.
The identification and evaluation of the causerelated risk dependencies bear indications helpful for
their mitigation. Firstly, an assurance taken to safeguard against the joint realization of two risks may be

1099

Table 1. Effects of the events dependency on the occurrence probability if one event is realized.

Cause-related risk dependencies

in general

in general

Events dependency

Effect of the realisation

of eh on the occurrence
probability of ei

specific forms:
- event identity
- equivalency
- one event is part
of the second

specific forms:
- incompatibility
- complementarity

Positive dependency
Equivalency
One event is part of the other
Negative dependency
Incompatibility
Complementarity

p(ei |eh ) > p(ei )

p(ei |eh ) = 1
p(ei |eh ) = 1
p(ei |eh ) < p(ei )
p(ei |eh ) = 0
p(ei |eh ) = 0

Positive dependency

Figure 3.

Negative dependency

Cause-related dependencies among risks.

apparent, the company has to take counteractive

actions to rectify the shortcomings.
Recapitulating the considerations made in sections 3.1 and 3.2, cause-related and effect-related risk
dependencies need a particular management. However, these processes closely resemble the risk management process (fig. 5), i.e. the required processing
steps are alike. To pinpoint the resemblances, figure 6
represents the thoughts given to the dependencies
management in a process flow.

Effect-related
t
risk dependencies

Identity
t of the
affected target

Figure 4.

Dependency of the
affected targets

Interdependency
r
of
the affected targets

Indifference

Indifference

Target complel
mentarity

Target comple-i
mentarity

Goal conflict

Goal conflict

Effect-related dependencies among risks.

rendered unnecessary, if these events are incompatible.

Moreover, any mitigation measure implemented to
reduce an events occurrence probability will be misdirected if the event is negatively related to all other
events whose probability of occurrence is higher.
In the following, the company must scrutinize all
detected cause-related risk dependencies to verify that
the relations have been fully and correctly captured.
It must also evaluate the effectiveness of the induced
mitigation measures. If any discrepancy becomes

RISK NETWORK MANAGEMENT

A risk network consists of the risks a company is confronted with, i.e. any risk-inducing event together with
its effects, and of their linkages (fig. 2). Accordingly,
the networks management has to allow for the risks
management (fig. 5) and the management of any of
their dependencies (fig. 6). Compared to the sole risk
management, the coordination of the three processes

1st Phase
Formulation/review of the risk
strategy and the target system

5th Phase
Risk monitoring

Figure 5.

Monitoring/
risk
accounting
system

2nd Phase
Risk identification and analysis (e.g.
referring to weak signals)

4th Phase

3rd Phase

Risk mitigation (e.g. acceptance,

avoidance, transfer)

Risk evaluation (occurrence probability?, extent of deviation?)

Risk management process.

1100

1st Phase
(Re-)formulation of relations between
targets, i.e. effect-related risk
dependencies
---

5th Phase
Controlling of the target systems
consistency and assessment of the
effectiveness of implemented risk
mitigation measures
Controlling and assessment of
identified event dependencies

Figure 6.

2nd Phase

Monitoring/
risk
accounting
system

Identification of cause-related risk

dependencies using pattern
recognition techniques

4th Phase

3rd Phase

Acceptance, rearrangement or use of

target dependencies to the companys
advantage

Assessment of the direction and

strength of target relations

Acceptance or manipulation of event

dependencies to the companys
advantage

Evaluation of event dependencies

using conditional probabilities

Managing risk dependencies.

presented costs the company a higher effort and additional time. In the following, we will oppose these
costs to the benefits of an integrated risk network
management, considering (a) the risk identification,
(b) the risk assessment, (c) the risk mitigation and (d)
the risk control.

4.1

Controlling of the target systems

completeness and consistency

Risk identification

A company set out to identify the dangers or target

deviations it may be confronted with in the future
analyzes its environment (Baum et al. 2007), and thus,
tries to attain knowledge about potential risk-inducing
events. Having identified upcoming events, the company correlates these and its target system to identify
potential risks.
In a risk network all risks are possibly related by
their causes and/or by their effects. These connections can serve as weak signals (Ansoff 1975), i.e.
one identified risk may hint at
Deviations the identified risk-inducing event provokes from the desired value of other targets due to
a direct influence,

Deviations the same event induces regarding other

targets because the target the event initially influences is related to these other targets,
Other risk-inducing events, as the events relation to
these other events are known from the past.
Thus, risk dependencies enlarge the sphere of influence of any measure for risk identification, facilitate
the identification of all and not just a limited part of
the relevant risks, and provide further insight into the
corporate risks and the necessary risk management
process.

4.2

Risk assessment

The risks evaluation can also benefit from the risks

relations, as these indicate if a risk is over-or underestimated.
If, for instance, two risk-inducing events act complementarily to one another, their occurrence probabilities must sum up to one. In case of a smaller or
larger amount, the difference indicates a mistake in the
assessment of at least one of the risk-inducing events
and their probabilities.

1101

4.3

Risk mitigation

Regarding the mitigation of its risks, a company can

choose either the avoidance of the risk, its reduction, its
retention or its transfer to a contractual partner. Normally, just one risk is considered while choosing a
mitigation measure, i.e. all of the risks dependencies
to other risks are abstracted away from. This implies
that the effects the measure chosen exerts on other
risks are not considered. Assuming two complementarily acting events, the realization of one of them,
renders the realization of the second event impossible, i.e. all mitigation measures applied to this second
event prove themselves unnecessary. The company has
wasted money.
Even if two risks are not related by their causes,
the avoidance of one of the risks, e.g. by transferring it to another party, may prove harmful. Then the
avoided risk can no longer be used to compensate for
the consequences of other risks.
4.4

Risk control

According to the nature of a risk just two achievement

levels for the affected target prove possible: Either the
risk is realized, i.e. the implied target deviation occurs
(t > 0), or the risk-inducing target does not enter,
and the target is fully achieved (t = 0). If at the end
of the period contemplated, the target deviation is
Larger than zero but smaller than the target level
to be achieved, if the risk-inducing event occurred
(te ), (0 < t < te ), or
Even higher as the proposed deviation, (0 <
te < t),
another explanation for the determined target deviation (t) has to be found. Here again risk dependencies render themselves useful. Suppose, there is a
second risk in addition to the one already contemplated
and this event affects a target behaving complementarily to the first one, than the resulting deviation of
the first target will be higher than the one announced,
as the second risk amplifies the deviation from the
desired achievement level of the first target.
Finally, the risks dependencies yield profits even
for the risk control and may thus initiate another run
of the risk (network) management process.
5

The reason for this interdependence is either a common

or interrelated risk-inducing event, or the same or
interconnected targets are affected. Thus, the risks a
company can be treated as a network.
As the risk management process proposed in literature and often applied in practice is not only applicable
for the risks management, but can be transferred to
any risk dependency, it can also be consulted for risk
networks.
The application of the risk management process to
risk networks proposes a number of benefits which
have to be challenged by comparing them to the
costs incurred. Moreover, the advantages and potential disadvantages of such a risk network management
process await the processs practical use.
REFERENCES
Andre, H.-J. et al. 1997. Analyse von Tabellen und kategorialen Daten. Berlin: Springer.
Ansoff, H.I. 1975. Managing Strategic Surprise by Response
to Weak Signals. California Management Review 18(2):
2133.
Backhaus, Kl. et al. 2006. Multivariate Analysemethoden.
Berlin: Springer.
Bamberg, G. & Baur, F. 2001. Statistik. Mnchen: Vahlen.
Baum, H.-G. et al. 2007. Strategisches Controlling. Stuttgart:
Schffer-Poeschel.
Heno, R. 1983. Kreditwrdigkeitsprfung mit Hilfe von
Verfahren der Mustererkennung. Bern: Paul Haupt.
Keel, A. 1995. Statistik II. St. Gallen: Wilhelm Surbir.
Lck, W. 1998. Der Umgang mit unternehmerischen Risiken
durch ein Risikomanagementsystem und durch ein
berwachungssystem. Der Betrieb 51(39): 19251930.
Mertens, P. 1977. Die Theorie der Mustererkennung in
den Wirtschaftswissenschaften. Zeitschrift fr betriebswirtschaftliche Forschung 29: 777794.
Mller, K. 2006. Wertschpfung in Netzwerken. Mnchen:
Vahlen.
Rojas, R. 1996. Neural Networks. Berlin: Springer.
Schrder, R.W. in prep. Ausdehnung der Risikopublizitt auf
Risikobeziehungen.
Schrder, R.W. 2005. Risikoaggregation unter Beachtung der
Abhngigkeiten zwischen Risiken. Baden-Baden: Nomos.
Schrmann, J. 1996. Pattern Classification. A Unified
View of Statistical and Neural Approaches. New York/
Chichester/Brisbane: John Wiley & Sons Inc.

FINAL REMARKS

Although most risks are so far considered separately

from one another, most of them are closely aligned.

1102

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Developments in fault tree techniques and importance measures

J.K. Vaurio
Lappeenranta University of Technology, Lappeenranta, and Prometh Solutions, Porvoo, Finland

ABSTRACT: This paper shows an efficient order for calculating importance measures and develops several
new measures related to fault diagnostics, system failure intensity and system failure count. Useful interpretations
and applications are pointed out, and many roles of the Birnbaum importance are highlighted. Another important
topic is the accuracy of various alternative methods used for quantification of accident sequence probabilities
when negations or success branches of event trees are involved. Thirdly, the role of truncation errors is described,
and criteria are developed for selecting truncation limits and cutoff errors so that importance measures can be
estimated reliably and risk-informed decision making is robust, without unreasonable conservatism and without
unwarranted optimism.

INTRODUCTION

Risk-informed decision making is becoming common

practice e.g. in nuclear, chemical and transportation industries. Typical subjects for decision making are ranking or prioritizing Systems, Structures,
Components or Human actions (SSCH) for enhanced
maintenance, quality assurance, safety categorization,
periodic testing programs or other such Preventive
Safety-Assurance Activities (PSAA), and optimizing these programs. Another area of applications is
control of plant state or system configurations when
failures occur in operation, or when planning plant outages or component maintenances. A necessary tool in
this decision making is a high quality plant-specific
probabilistic safety assessment (PSA) and the capability to obtain accurate quantitative risk results and
importance measures. Computerized fault-tree and
event-tree software tools have been used in this process extensively. However, there are signs that some
of the techniques used may not be sufficiently accurate, or that full potential of the tools may not be used
(Epstein & Rauzy 2005). This paper addresses several
of these issues.
The basic importance measures, their roles and
interpretations are introduced in Section 2. Although
close to the original definitions, the process is such
that only one of the measures needs to be solved from
a fault tree model for each basic event. The others can
then be obtained readily. Some relationships are different for events that are mutually exclusive rather than
statistically independent.
Section 3 is devoted to comparing various methods used in quantification of accident sequences and
importances, with special focus on handling success

branches of event trees. The roles of truncation errors

and other approximations in this process are evaluated.
Truncation or cut-off is a way to limit the number
of minimal cut sets (MCS) used in quantification.
New criteria are suggested for selecting a tolerable
upper limit for the truncation error, which then determines the truncation limit. Explicit use of truncation
errors and importances in these criteria is new and

extends earlier studies (e.g. Cepin

2005, Duflot 2006).
Numerical examples illustrate the techniques. Several
new importance measures are introduced in Section 4.
These are relevant to system failure diagnosis, intensities, failure counts and integral quantities. Birnbaum
importance plays a major role in the new developments, which are felt to be relevant improvements
to earlier measures. This paper is a continuation
and extension of earlier works by the author (Vaurio
2006a,b, Vaurio 2007).
2

IMPORTANCE MEASURES

Consider a TOP-event Y of a fault tree, a Boolean

expression of the failed state of a system or function in terms of the basic event failed states Zk and
possibly also success states Zk , k = 1, . . . , K. The
probability y = P(Y ) = y(z, z ) is formally linear in
terms of all K basic event probabilities zk = P(Zk )
and zk = P(Zk ) = 1 zk . The complements can be
preserved as individual events or by substitution we
have y = y(z1 , . . . , zK ). In coherent models negations
and complements Zk are not needed. The importance
measures of basic events were originally defined only
for coherent models when the basic events are mutually
statistically independent. The definitions are extended

1103

here to some non-coherent systems and mutually

exclusive events so that all important interpretations
and applications can be preserved. The order of computation is suggested such that a minimum number of
model runs are needed.
A key quantity for all importance measures of
basic event Zk with respect to Y is the low-end value
yk = y(z, z |zk = 0, zk  = 1). This is obtained as
an analytic expression or using the minimal cut sets
(MCS) of Y and some order Sylvester-Poincar development (S-P; the inclusion-exclusion principle). This
yk can be obtained also directly using fault-trees Y
and YZ k as yk = P(Y |Zk  ) = P(YZ k  )/P(Zk  ) =
[P(Y ) P(YZ k )]/(1 zk ). K + 1 quantifications are
needed to obtain y and all K values yk .
Many software tools calculate also K high-end
values yk + = y(z, z |zk = 1 zk  = 0) from the system model. This is unnecessary because yk + can be
obtained from y and yk directly, as will be shown.
Thus, only K + 1 system quantifications are needed.
One of the most important measures is the Criticality Importance (i.e. fractional contribution)

easy to assess the total impact of changes (Vaurio

2006a, b).
Another measure Risk Reduction Worth RRW k =
y/yk can be obtained as

y yk
CIk = CIk (Y ) = CI(Zk ; Y ) =
y

RIFk = RIFk (Y ) = RIF(Zk ; Y ) =

(1)

This definition is useful for any linear function no

matter what kinds of dependences couple the basic
events. Thus, all definitions from now on hold even
if y is not a probability of a single TOP event but a
linear combination of many TOP-event or sequence
probabilities. The criticality interpretation is:
CIk is the fractional risk reduction when event
k (failure) is eliminated (i.e. zk 0).
Criticality is utilized for assessing the impact of a
potential change of zk in the numerator of Equation 1
while the denominator y is the nominal risk yo with all
inputs zk at the nominal values. When changing zk < 1
within [0,1) does not change the other probabilities
(except the complement zk  when it is in the model),
y is a linear function of zk and
CIk is the fractional risk increase if the failure
probability is doubled (i.e. zk 2zk ).
Note that this can hold when Zk is independent but
also when it is mutually exclusive with some other
events like other failure modes of the same component. Linearity in both directions makes CI very
suitable for ranking systems, structures, components
and human actions for different levels of preventive
safety-assurance activities such as maintenance categories, safety classification and quality assurance etc.,
actions that influence the reliability within reasonable
limits, not like a complete failure or repair. When a
preventive activity influences many components, the
criticalities are practically additive, which makes it

RRWk = (1 CIk )1

(2)

and gives the same ranking as CIk .

The Birnbaum Importance is defined here as
BIk = yCIk /zk

(3)

Under mutual independence this is consistent with

the original definition and equal to y/zk (even if
negations would exist) but does not require separate
quantification of yk + . This nearly ignored measure
has plenty of interesting uses and interpretations as
will be shown in Section 4.
Finally, the Risk-Increase Factor, also called Risk
Achievement Worth, is defined as
yk +
,
y

(4)

measuring the impact of a real or predicted failure:

RIFk indicates a factor by which the risk is increased
if event Zk is in failed state or the component is taken
out of service.
RIF is widely useable in configuration management to determine allowed outage (repair) times and
planning maintenance activities. The denominator y
is again fixed at nominal value so that usually RIF is
insensitive to the reliability zk of the component that
it measures. Thus, RIF is unsuitable for ranking preventive activities (PSAA) that do not fail or repair a
component. Extensions of RIF to mutually exclusive
events and for controlling uncertain configurations of
multiple failures have been presented (Vaurio 2006a,
b). Notice that it is not necessary to quantify K values
yk + from the model to obtain all RIF k .
When a basic event Zk is statistically independent
of all other basic events, it follows from the linearity
and the definition of yk that
yk + = P(Y |Zk ) = P(YZ k )/zk
= [P(Y ) (1 zk )yk ]/zk

= (y yk )/zk + yk

(5)

and thus
RIFk = 1 +

1104

CIk
CIk
zk

(6)

This holds also for non-coherent trees such that both

Zk and Zk appear in the model but are independent of
the other basic events.
Next, consider a basic event Zk mutually exclusive with a subset Jk of basic events while statistically
independent of all other basic events. The mutual
exclusivity means Zk Zj = , and if zk = 1 then zj = 0
for all Zj Jk . When reducing the value of zk does not
influence the other values zj , then
RIFk = 1 +


CIk
CIk
CIj
Zj Jk
zk

(7)

This kind of dependency can exist between different

failure modes of a component (when only one mode
can be present at a time and otherwise they do not influence each other). Also common cause failures within
a group of components are sometimes assumed mutually exclusive in this way. Initiating events are also
usually considered mutually exclusive. For initiators
RIF is not meaningful but the Birnbaum importance is
relevant as the conditional probability of an accident
given the initiating event.
Finally, let us consider y a linear combination of
linear Boolean event probabilities P(Yi ), i.e. y =
i ci P[Yi (z)], where ci are scalar constants (positive or negative). Then the low-end value yk =
i ci P[Yi (z)|zk = 0] is a linear combination of the
corresponding values of each term, and all importance
measures with respect to the sum y are obtained via
Equations 1, 2, 3 6 and 7 without additional model
quantifications. If the importance measures have been
determined with respect to each Yi separately, then
they can be obtained with respect to y as weighted
sums.

ACCIDENT SEQUENCES

The risk function of a plant can generally be presented

as the accident frequency, a function
y = y(z1 , z2 , . . . , zI +J )
I
J (i)
=
fi
Pi,j (q1 , q2 , . . . , qJ )
i=1

all initiating events and sequences, system models

and basic events for all plant operating states (POS)
covering the normal full power operating period and
refueling/maintenance outages that can have several
stationary states and transitions while shutting down
and starting up again. The total risk is calculated for an
operating cycle of length L (typically a year), and the
initiating event rates are defined per phase (POS) or
cycle length. The number of sequences J(i) depends
on the initiator and accident consequence: in some
cases sequences can be combined through an OR-gate
to form a single TOP of a fault-tree with probability
Pi . In other cases even a single sequence can not be
modeled by a coherent TOP-event but by a difference
of two TOP-event probabilities. Even then y is a linear
function of the basic event probabilities with familiar
relationships for importance measures.
Fig. 1 represents an event tree with an initiating
event (I) and three headings, safety functions (D,H,L)
modeled by fault-trees (which are not visible here).
The TOP-events have the same names as the functions. A function may consist of several systems. The
upper branches represent successes of the corresponding safety functions, lower branches represent failures.
The Boolean formulae associated with the sequences
are given on the right.
Consider a sequence S defined by failed functions
(headings) Fl and success functions (headings) Gl 
as S = F1 . . . Fm G1  . . . Gn  . = FG  , where F =
F1 . . .Fm and G = G1 +. . .+Gn , based on de Morgans
rule (G1 + G2 ) = G1  G2  . This suggests an exact way
to calculate P(S) as a difference of two coherent models: P(S) = P(F) P(FG). Instead of this, several
approximate methods seem to be commonly in use.
In the following, an estimate of the probability
P(Y) of any coherent TOP-event Y can be exact or
a rare-event approximation REA(Y), or a min-cutupper-bound MCUB(Y). The approximations are conservative and REA(Y) MCUB(Y) P(Y) when
calculated using the complete set of minimal cut sets
of Y, set MCS[Y]. However, when truncation is used,
only a subset of MCS[Y] is identified and used in
quantification. Then it is not sure whether an upper or

(8)

j=1

of probabilities zk , either initiating event rates fi

(probabilities per unit time), or dimensionless unavailabilities qj , probabilities of failed states of components
or human actions. The dimension of y is per unit time
or operating cycle. Pi,j is the conditional probability
of accident sequence j for initiator i, given initiating
event i. An accident or consequence is different in each
application. In nuclear industry y is often a reactor
core damage frequency (CDF) or a large early radioactive release frequency (LERF). In principle, y includes

C1:IDHL
C2:IDHL
C3:IDHL
C4:IDHL
C5:IDHL
C6:IDHL
C7:IDH

Figure 1. An event tree and Boolean formulae associated

with its sequences.

1105

lower bound of P(Y) is actually obtained. Both terms

of the exact expression P(S) = P(F)P(FG) are subject to these approximations and truncation in handling
TOP-events F and FG. Other complications arise if a
system model contains negations or NOT-gates other
than success functions.
Let us consider several approximate methods, first
without truncation. The first approximation consists
of ignoring success branches, i.e.
AS1(S) = P(F1 . . . Fm )

(9)

The second and third approximations include correction factors,

AS2(S) = P(F1 . . . Fm )

n
j=1

[1 P(Gj )]

(10)

AS3(S) = P(F1 . . . Fm )[1 P(G1 + + Gn )]

(11)
If conservatism is required, P(G) should be as small
as possible, i.e. MCUB rather than REA, if exact
calculation is not possible.
The fourth approximation consists of deleting from
the MCS of F1 . . . Fm all MCS that satisfy the fault
tree function G1 + + Gn , i.e. make G = TRUE:
AS4(S) = P{Union of MCS[F1 . . . Fm ]not
satisfying (G1 + + Gn )}

This approach can be applied in principle to any

NOT-gates that appear in a fault tree model, not just for
event tree headings. A word of warning: unless computerized, this approach can be cumbersome for large
models. The method works best if negations appear
near the bottom of a fault tree, at the level of basic
events or modules. Otherwise there is a danger of
exploding the solution space.
Finally, there is the exact way to obtain P(S) without
all complications, P(S) = P(F)P(FG), a difference
of the probabilities of two TOP-events, F and FG. The
probabilities are linear in terms of basic event probabilities, even coherent unless there are NOT-gates
other than the success branches. Because there can be
other approximations in quantification, this approach
is defined AS7:
AS7(S) = P(F) P(FG)

(14)

(12)

This union of a subset of MCSs is not generally an

exact representation of the sequence S because S can
not be presented by MCSs when there are negations or
success branches. There is no need to solve the MCSs
of the union G1 + + Gn to make the deletion; thus
no additional truncation error is caused here.
The fifth method adds a reduction factor,
AS5(S) = AS4(S)[1 P(G1 + + Gn )]

Minimal cut sets of the rewritten formula are

computed.
Those terms that contain both an event and its
encoded complement are eliminated.
The remaining union is quantified. (If higherorder inclusion-exclusion development is used, also
higher-order terms that contain both an event and a
negation should be deleted.)

(13)

This is close to or exactly what can be done

with Risk Spectrum Professional, for example. A
conservative P(G) leads to some underestimation of
the reduction factor. However, another feature makes
the reduction factor conservative: P(G) should actually be quantified under the condition F = TRUE,
and this P(G|F) would generally be larger than the
unconditional P(G). Anyway, it is difficult to estimate
over- or underestimation, especially when combined
with truncation and REA or MCUB.
The next advanced method AS6(S) processes success branches as follows:
Negations caused by G1  . . . Gm  in FG  are pushed
down toward basic events by de Morgans Laws.
New variables (basic events) are introduced to
represent negated basic events.

Advantages: This approach needs only the very

basic fault-tree techniques to quantify TOP-events F
and FG. It avoids extra factors and ad-hoc manipulation of MCSs. It also allows estimation or bracketing
of the approximation error (by MCUB and 2nd order
Sylvester-Poincar formula). No bracketing has been
developed for approaches that use coefficients or
manipulate MCSs in a complex numerical way. The
computational burden to quantify F and FG is not
essentially worse than quantifying F and G, as needed
in versions AS2 AS5. It is a bit surprising that the
formally exact AS7 has not been recognized and used
in commercial fault-tree software tools.
In all versions additional approximation may be
caused by truncation or by the quantification using
REA, MCUB or a higher order S-P expression.
As an example, consider a sequence S = FG  ,
where after truncation the Boolean models are
F = (A1 + B1 )(A2 + B2 )(A3 + B3 )

(15)

G = (A1 + A2 + A3 ) + (D1 + D2 + D3 + D4 + D5 )
(16)
with numerical probabilities P(Aj ) = P(Bk ) =
P(Di ) = 0.1 and P(Cj ) = 0.001. This example is
adapted from Epstein & Rauzy (2005). Note that Ci

1106

Table 1.
methods.

Sequence probabilities P(S) with different

Table 2. Criticality importance CI(A1 ) with different

methods.

Method
formula

Rare-event
approximation

Min-cut
upper bound

Complete
S-P form

Method
formula

AS1
AS2
AS3
AS4
AS5
AS6
AS7

0.0080000
0.0028000
0.0016000
0.0010000
0.0002000
0.0004305
0.0005000

0.0079721
0.0034317
0.0034317
0.0010000
0.0004305
0.0004305
0.0004966

0.0068590
0.0029526
0.0029526
0.0010000
0.0004305
0.0004305
0.0004305

AS1
AS2
AS3
AS4
AS5
AS6
AS7

Rare-event
approximation
0.500000
0.428571
0.250000
0.0
0.500000
0.111111
0.0

Min-cut
upper bound
0.499000
0.443333
0.443333
0.0
0.111111
0.111111
0.004010

Complete
S-P form
0.473684
0.415205
0.415205
0.0
0.111111
0.111111
0.111111

is missing from Equations 15 and 16 because of truncation with truncation limit TL = 2 105 (to be
discussed later).
Numerical results for P(FG) obtained with different approximations (REA, MCUB) and with complete
S-P formula calculation are presented in Table 1. All
TOP-events in each formula have been quantified with
the approximations indicated. The results confirm that
methods AS1AS4 are inadequate, as was also shown
by Epstein & Rauzy (2005). They did not consider
methods AS5AS7.
Other conclusions and comments on this
example:

example it happens that setting P(A1 ) = 0 [when

calculating P(S|A1 = )] reduces P(FG) more than
P(F), leading to negative CI (A1 ) for many methods.
Even if this example is indeed somewhat artificial,
negative importance is a real possibility in success
sequence calculations. This example shows complete
inadequacy of AS1 through AS4, AS5 and AS7 with
REA, and AS7 with MCUB. AS6 is the best, otherwise the results call for efficient 2nd and 3rd order
S-P solutions to be implemented in software.

AS6 is most accurate with approximations. This is

partially due to the simplicity of the example which
yields only one single MCS with AS6. For realistic
large systems repeated use of de Morgans rules can
cause problems.
Methods AS5 and AS7 are reasonable with approximation MCUB. But one should notice that only AS7
is always exact with complete S-P-formula. AS5 is
accurate in this simple example.
With REA, AS5 is not acceptable and AS7 is on the
border line of acceptability.
REA does not lead to the most conservative result
for AS5 (and AS2 and AS3) because P(G) is also
conservatively quantified with REA. The ordering
P(Y ) MCUB(Y ) REA(Y ) holds generally
only when Y is a single coherent TOP-event.
These results indicate that higher order approximations or completely new methods may be needed
so that accuracy can be improved or at least proper
error assessments can be done.

Solving all minimal cut sets of a large risk model is

practically impossible because of computer time and
memory restrictions. A common practice is to set a
truncation limit TL and delete all terms (MCS) that
are known to have smaller probability than TL. The
remaining terms form the truncated model that is used
in all quantification of risk and importance measures.
This can obviously lead to some underestimation of
risk, but setting TL small enough one tries to cover the
essential portion of the risk. Important characteristics
of the truncation process are:

3.1 Importance measures

To see how different approximations behave in importance calculations, the criticality importance of the
basic event A1 is listed in Table 2. The same values
are valid for A2 and A3 due to symmetry. They are
of particular interest because A1 appears in both failure branches (F) and success branches (G). In this

3.2 Truncation error assessment and use

An upper limit for the total deleted risk (probability

of the union of the deleted MCS) can be estimated.
This is called the truncation error TE(Y).
The deleted terms are not identified as MCSs at the
basic event level: it is not fully known which basic
events contribute to the truncation error.
Consequently, it is not known which importance
measures are erroneous and how much, when they
are determined from the truncated model. For basic
events that are deleted by truncation the calculation
always yields CI k = 0 and RIF k = 1.
When the truncated model is quantified by a conservative method (REA, MCUB), it is not known
whether the result is above or below the real risk.
There are examples in the literature indicating that
a large portion of basic events of a realistic PSA
model can be left out because of truncation, 60%

1107

to 90% according to Epstein & Rauzy (2005), and

the truncation error can be as large as the TOPevent probability. This may be unavoidable because
iteration with decreasing TL could lead to a false
plateau (unchanging y), or explode the computer
time and memory requirements (Duflot et al 2006).
Consequently, the acceptability of risk should be
based on y + TE(Y ) rather than the truncated result
y alone.
Because it is not known which basic events are
contained in the deleted terms, one does not know
how much the deleted terms would contribute to
each low-end value yk (or high-end value yk + ) that
are needed in importance measures. To be conservative in estimating the criticalities, the total TE
is assumed to contribute to the TOP event in calculating y but not in the terms of yk . In principle then an upper bound (conservative) criticality is
obtained as
y yk + TE(y)
UCIk =
y + TE(y)

(17)

TE(y) depends on all truncation errors generated when

quantifying the TOP-events needed for y, in our example F, G and FG, depending on the formula. In the
difference P(S) = P(F) P(FG) in AS7 one can
add the positive TE of F. However, one can not subtract TE(FG) because the actual error in P(FG) can be
significantly smaller than the upper bound TE. The
impact of TE(y) on other importance measures follows directly from Equations 1, 2, 3, 6 and 7 with
UCIk replacing CIk . Especially RIFk (Equation 4) can
be very much influenced by TE(y) because the maximum impact on yk + could be TE(y)/zk with possibly
a very small zk .
In the previous example calculations were carried
out without considering any truncation error, i.e. using
the models F and G as given, as the truncated model.
Actually a truncation limit TL = 2 105 was used
and some terms in F were truncated (deleted) so that
Equations 15 and 16 form a truncated model. The basic
events C1, C2 and C3 are not even within the identified model. The upper limit of the total truncation
error TE(F) for P(F) was estimated to be 9 105 ,
no truncation error for G, and TE(FG) was 7 105 .
Using TE(y) = 9105 in Equation 17 yields UCI for
A1 equal to UCI(A1 ) = 0.0810, significantly different
from 0.111 as obtained from the truncated model.
This means a serious uncertainty about the importance
of A1 and could lead to rather incorrect ranking. The
way out is to reduce TL and identify additional MCSs
so that TE is reduced. In this example using a smaller
TL = 5 106 leads to discovery of additional terms
and reduction of truncation errors. Additional terms in
F are (B1 B2 +B1 B3 +B2 B3 )(C1 +C2 +C3 ), explaining
the earlier TE(F) and TE(FG). Because the additional

terms did not contain A1 , the final importance is

CI(A1 ) = 0.111.
3.3

Criteria for selecting the truncation limit

Truncation limit TL can only be guessed in advance but

the final value must be based on the truncation error
TE observed when carrying out calculations with some
TL. The first criterion to be considered is the accuracy
or degree of conservatism tolerated for the risk estimate y + TE(y). One does not wish the error term to
be a large fraction of y because it could lead to misguided plant modifications or other hasty decisions. A
reasonable limit is
TE(y) < 0.1 y.

(18)

For large models TL needs to be orders of magnitude smaller than TE. With y = 104 this criterion
easily leads to TL of the order of 108 or smaller.
Another criterion may be based on threshold values that are used for the criticality importances CI
of components. A criticality threshold CIT could be
a value used in ranking components for maintenance
or safety categories. Such values CIT usually range
from 106 to 103 . There are two types of potential
ranking-errors due to truncation:
For components that are not included in the truncated model or are essential in the deleted terms:
the criticality (y yk )/y underestimates and fails
to rank the components properly for PSAA.
On the other hand, with large TE(y), use of Equation 17 leads to many unimportant components
classified highly in maintenance programs or safety
categories, without serving real safety.
Thus, it makes sense to minimize TE(y) as much as
possible. The condition TE(y)/y < 0.1 CIT yields
TE(y) < 0.1 y CIT

(19)

With y = 104 and CIT = 104 this leads to the

requirement TE(y) < 109 , harder than Equation 18.
A third criterion is based on the sensitivity of RIF to
the truncation error. If the truncated terms include significant MCSs sensitive to a certain component, using
the truncated model with zk = 1 could be a serious
underestimation. It is generally possible to determine
a threshold limit DT such that if RIFk < 1 + DT,
no particular restrictions are set for plant operation
or shutdown, at least until the next regular maintenance. A reasonable value for DT is between 0.01
and 0.1. The basic events belong to two groups based

1108

on the truncated model, and two different criteria are

needed:
a. For basic events that already have RIFk 1 > DT,
we dont want TE to dominate decisions. The
condition TE(y)/(yz k ) << RIFk 1 leads to
TE(y) < 0.2 (RIFk 1) y zk

(20a)

b. For basic events with RIFk 1 < DT we dont want

TE to decide what actions to take. The requirement
TE(y)/(yz k ) << DT then yields
TE(y) < 0.2 DT y zk

(20b)

These ideas to tie the criteria of truncation error to

the risk-informed decision making (ranking and configuration control) and the uses of truncation errors

explicitly seem to be new. Earlier Cepin

(2005) proposed a convergence test for y, but Duflot et al (2006)
found several problems with it. They suggested an
approach not to cut off MCSs that could become large
when some zk is made equal to 1. That approach
potentially reduces the number of basic events that
are under-ranked by the truncated model (i.e. belong
to group b above), but it does not eliminate the need to
use both criteria 20a and 20b for the residual TE, and
Equation 17 for the upper bound criticality. For events
that can not satisfy those conditions one has to accept
the conservative importances (Equation 17 combined
with 4) and rankings that follow. Or one can take the
un-truncated model also for importance calculations,
set Zk = FALSE (or TRUE), and truncate later when
solving the MCSs.
When the assessment is carried out in additive parts
(e.g. separately for different initiating events), the conditions for truncation errors may be allocated so that
the sum satisfies Equations 18 through 20.
Some software suppliers are developing new
approaches such as BDD for solving the truncated
model exactly (instead of REA or MCUB). This helps
in reducing possible conservatism in the truncated
model, but it does not eliminate the errors caused
by truncation. Possibly 2nd or 3rd order S-P developments are needed with efficient algorithms to allow
small truncation limits.

ADDITIONAL IMPORTANCE MEASURES

In addition to risk assessment certain importance measures have roles to play in failure diagnostics. Assume
that Y is a fault-tree TOP-event, a Boolean function, and any basic event is completely statistically
independent or possibly mutually exclusive with some
subset of events. Failure of a component does not
always reveal itself but needs to be identified after

the system has ceased operating. Common examples

include electronic circuits and automobiles. The question arises: which of the possible components should
be inspected or replaced first? It is natural to base
the prioritization to the conditional probability of
a component (event) being failed when the system
is found to be failed. This leads to one diagnostic
importance
DIk = P(Zk |Y ) =

zk P(Y |Zk )
= zk RIFk
y

(21)

This is applicable when a system failed state is

observed but not the failure event occurrence itself,
and when it is not so important whether a single repair
returns the system immediately to service.
Another diagnostic importance is the conditional
probability that a component (event Zk ) is in a failed
state when the system is found failed, and repair of that
event also repairs the system. In this state the system
is not failed without Zk failed. Because the system
can be failed only with Zk or without Zk , we have the
diagnostic criticality importance
DCIk =

y yk
P(Y ) P(Y |Zk  )
=
= CIk
P(Y )
y

(22)

Thus, besides ranking components for potential risk

reduction and PSAA, the criticality importance can be
useful for prioritizing inspection programs.
Next we consider several roles of the Birnbaum
measure when all basic events are mutually statistically independent. First, Equation 3 means that BIk =
y/zk , and the linear differential y = BIk zk
implies that BIk gives the correct ranking if the same
absolute change is made in all basic event probabilities. The differential also offers a simple measure of
uncertainty with small variations or uncertainties in
basic event probabilities. The sum j BI2j j2 is an estimate of the variance of y when k 2 is the variance of zk .
A reasonable uncertainty importance measure of event
Zk is UIk = BIk 2 k 2 /j BIj 2 j 2 , the relative contribution of event Zk to the total variance of y. Based on
a general form Y = Zk Gk + Hk = Zk Gk Hk  + Hk ,
where Hk is the union of MCSs not containing Zk , it
can be concluded that BI k is the probability of a system state such that the state of event Zk dictates the
system state. This leads to interesting interpretations:
BIk = P(Gk Hk  ) is
the fraction of event k failures that lead to immediate
system failure, and
the fraction of event k repairs that change the system
from failed to operational.
Furthermore, BIk is a probability of a system state
such that

1109

 if the next change is failure of event k, the system

also fails, and
if the next change is repair completion of event k,
then the system also transfers from a failed state to
an operational state.

(t) = Ws,k (0) + ws,k (t  )dt  . A natural failure count

contribution (criticality) importance is

These facts lead to an economic way to calculate the

system failure (and repair) intensity. When all basic
events are independent and have failure intensities wk
and repair intensities vk , the system or function failure
intensity ws (t) and repair intensity vs (t) are

This deviates from the Barlow-Proschan importance

t
BIk (t  )wk (t  )dt 
(27)
BPIk = 0  t


0 ws (t )dt

ws =
vs =

K
k=1

K

k=1

BIk (t)wk (t)

and
(23)

BIk (t)vk (t).

Traditionally fault tree software tools first calculate

the failure intensities of individual MCSs and then
add these up. Equation 23 is much more economic
because the number of basic events (K) is generally
much smaller than the number of MCSs, and BI k or
CI k are automatically produced by fault tree codes.
Furthermore, Equation 23 is exact if BI k is exact, while
the traditional formula as a sum of MCS intensities is
only an approximation (REA).
Let us define the intensity contribution of Zk as
ws,k (t) = ws (t) ws (t|zk = 0, wk = 0),

ws,k (t)
.
ws (t)

Ws,k (t)
Ws (t)

(26)

which is an integral form of the last failure importance

because it includes in the numerator only the intensity
term caused by event Zk as the last failure when the system fails. It does not account for the full contribution
of component Zk to the failure intensity.
Let us consider simple series and parallel systems
that have K components with independent failure and
repair times, the mean times to failure k and the
mean times to repair k , k = 1, . . . , K. Asymptotically in the renewal process the unavailabilities are
zk = limt zk (t) = k /(k + k ) and the failure
intensities are wk = limt wk (t) = 1/(k + k ).
For a series system the unavailability is y = 1
j=1, ... ,K (1zj ) and the Birnbaum importances BI k =
(1 y)/(1 zk ) = j =k [j /(j + j )]. Then

(24)

the sum of all terms BI j (t)wj (t) that contain as a factor either wk (t), or zk (t). This is how much the system
failure intensity would be reduced if Zk were made
non-failable. Besides ranking components this can be
useful in a diagnostic problem. Assume that a system
failure occurrence at time t is observed, without knowing which components were failed and contributed
to the system failure. A question is: which component should be inspected or replaced first? A natural
ranking indicator for Zk is the intensity criticality
importance defined here as
ICIk =

CCIk =

(25)

This was earlier defined only within the rare-event

approximation. ICI k is useful if one wants to repair all
failed components before system re-start.
In case one wants to identify which component
was most likely the last one to fail the system, an
appropriate ranking index is the last failure importance
LIk = BIk (t)wk (t)/ws (t). LI is useful if one can take
the system back to service as soon as it is operational
even if some components are still un-repaired.
The total expected number of system failures is the
integral over time [0,t], Ws (t) = y(0) + ws (t  )dt  ,
and the contribution made by event Zk is Ws,k

BPIk,series =

K

j=1

(28)

(1/j )

This is independent of repair times because for the

system to fail all components needs to be up and then
any failure fails the system, i.e. then any failure is the
last failure causing the system to fail.
For a parallel system the unavailability is the product y = zj , and the Birnbaum importances BIk =
y/zk = j =k [j /(j + j )]. Then
BPIk,parallel =

K

j=1

(29)

(1/j )

is independent of times to failure. This is because first

all components have to fail for the system to fail, and
then any repair repairs the system.
For a series system the intensity criticality (and the
asymptotic failure count criticality) is
ICIk,series =

1110

wk
K

(1 zk )2

wj
j=1 1zj

zk
1 zk

BPIk zk
1 zk


k
k
= 1+
BPIk,series
k
k
=

(30)

This is generally smaller than BPIk,series because

eliminating a downtime k increases the failure effects
of other components in a series system.
In a parallel system making any component perfect eliminates all system failures. Thus, ws (t|zk = 0,
wk = 0) = 0 and ICIk,parallel = 1.
5

and unwarranted optimism. This work extends earlier

works done on analyzing the reliability of systems with
repairable components and defining importance measures for average (time-integral) quantities. Some of
these have been defined and extended to multi-phase
missions with repairable and non-repairable components (Vaurio 2007).

CONCLUSIONS
REFERENCES

An efficient process was presented for calculating

importance measures so that a minimum number
of model runs are needed, taking into account also
mutually exclusive events and negations. An efficient
way to calculate the exact system failure intensity
was pointed out, and several new measures were
developed related to fault diagnostics, system failure intensity and system failure count. Useful interpretations and potential applications were identified,
and several roles of the Birnbaum importance were
highlighted. Another topic was the accuracy of various methods used for quantification of accident
sequence probabilities and importances when negations or success branches of event trees are involved.
The results call for effective higher-order S-P formalism in fault tree codes, or completely new methods.
Thirdly, the roles of truncation errors were described,
and new criteria were developed for selecting truncation limits to control errors so that importance
measures can be relied on, and risk-informed decisions can be made without unreasonable conservatism

Cepin,
M. 2005. Analysis of truncation limit in probabilistic safety assessment. Reliability Engineering & System
Safety 87:395403.
Duflot, N. et al. 2006. How to build an adequate set of minimal cut sets for PSA importance measures calculation.
Proceedings of PSAM 8 conference, New Orleans, 1419
May 2006. New York: ASME.
Epstein, S. & Rauzy, A. 2005. Can we trust PRA? Reliability
Engineering & System Safety 88(3):195205.
Vaurio, J.K. 2006a. Developments in importance measures
for risk-informed ranking and other applications. Proceedings of PSAM 8 Conference, New Orleans, 1419
May 2006. New York: ASME.
Vaurio, J.K. 2006b: Configuration management and maintenance ranking by reliability importance measures. Proceedings of ESREL 2006 Conference, Estoril, Portugal,
1822 September 2006. Leiden: Balkema, Taylor &
Francis.
Vaurio, J.K. 2007. Definition and quantification of importance measures for multi-phase missions. Proc. ESREL
2007 Conference. Stavanger, Norway, June 2007. Leiden:
Balkema, Taylor & Francis.

1111

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Dutch registration of risk situations

J.P. van t Sant & H.J. Manuel
RIVM (National Institute for Public Health and the Environment), Bilthoven, The Netherlands

A. van den Berg

IPO (Association of Provincial Authorities), The Hague, The Netherlands

ABSTRACT: The legislation for registration of risk situations involving hazardous substances in The
Netherlands came into effect in the year 2007. Competent authorities are now required to submit data on
establishments and transport situations to the Register for risk situations involving hazardous substances. This
Register is an Internet application that has been developed in the past years, based on older concepts of the
legislation. With the final details of the legislation now being established, the application had to be adapted
considerably. After that the data from the original database have been converted to the final release. Other kinds
of risks (e.g. forest fire, collapse of high buildings) are registered by a second Internet application, called ISOR.
Also mandatory, competent authorities have to submit data on such risks. To comply with future European legislation registration of areas prone to flooding by sea or rivers have also recently been added. The data from both
applications are combined and shown on maps via Internet. Professionals (only civil servants of the government,
provinces and municipalities) have access to the maps and all data through a protected site. The general public
has access to the maps and a selection of the data through an open site. The general goal is that public and
authorities get insight and overview on risk situations. Municipalities can use the maps to fulfill their obligation
to inform their citizens on risks within their territories. Both spatial and emergency planning is expected to
improve with the combined data from Register and ISOR. That will be enhanced by the fact that the maps also
contain data on vulnerable objects (e.g. hospitals, schools, tall buildings).

INTRODUCTION

As one of the consequences of the major fireworks

accident in May 2000 in Enschede (The Netherlands)
(Ale 2002), the Dutch Government decided to establish a registration of activities involving hazardous
substances which can cause a major accident. For
this purpose an Internet application called Register
for risk situations involving hazardous substances
(further called RRGS) was built. In the RRGS
competent authorities can submit data on establishments handling, storing or processing dangerous materials as well as on transportation of such
materials.
For the registration of vulnerable objects (e.g.
schools, hospitals, tall buildings) and risk situations
not related to dangerous substances (e.g. floods, forest fire, riots) a second Internet application was built
called Information System for Other Disaster Types
(further called ISOR). Again competent authorities
can submit data on said objects and risk situations in
this application.

The registered data (except effect distances) are

shown to the public on Internet maps, called risk
maps. The effect distances are not shown for security
reasons. Professionals (i.e. authorized civil servants)
can access all registered data through a protected site.
The risk maps give an overview per province of vulnerable objects and risk situations with and without
dangerous materials. See Figure 1 for the relation
between applications and maps.
To comply with European legislation areas prone
to flooding by sea or rivers have also recently been
added to the maps.
The original RRGS had been built in 2002 based
on concept legislation from the Ministry of Housing,
Spatial Planning and the Environment (Vant Sant et al.
2003). Additionally a Regulation from the Ministry
of the Interior and Kingdom Relations was drafted.
It concerns risky situations in general. It extends the
number of situations with hazardous materials which
have to be placed in the Register. In 2007 the final
legislation came in to force. This necessitated us to
adapt the applications considerably.

1113

RRGS data

RRGS
database

Data
replication

ISOR data
input

National
database of
the
provinces

12 provinces

Professional maps

12 provinces

Public maps
Figure 1.

2
2.1

Relationship between RRGS and ISOR applications and the risk maps.

LEGISLATION
Environmental management Act
and Registration Order external safety

The registrations mentioned above are based on several new pieces of legislation. First an amendment
of the Environmental management Act was established, which came into force by the end of March
2007 (Anonymous 2005). The amendment formalises
the institution of a register for risk situations with
hazardous substances. It imposes the obligation of
competent authorities to submit data to this register on
risk situations for which they have a formal responsibility. Another obligation is the authorization of the
submitted data. Authorization signifies that the data
are correct and can be shown on the risk map (see
Section 4). RIVM (the Dutch National Institute for
Public Health and the Environment) has been charged
with the management of this register. In turn, RIVM
subcontracted the management to IPO (Association of
Provincial Authorities).
The details of the register are regulated in an Order
in Council, called Registration Order external safety
(Anonymous 2007a). In this Order the categories of
risk situations are defined. Examples are Seveso-II
establishments, LPG filling stations, establishments
storing or processing fireworks and transport of hazardous substances. The Order also prescribes the data

to be registered, such as name, address, geographical

location, hazardous substances involved, individual
risk contours of 106 per year (The individual risk
is the risk to an (unprotected) individual at a specific
location, expressed as the probability in one year of
dying as a direct result of an on-site accident involving dangerous substances. The location-based risk is
visualized by risk contours on a map.) and, if available,
societal risk graph.
In many cases it will not be realistic and practical to ask authorities to supply risk data based on full
Quantitative Risk Analyses. Therefore, the Regulation external safety of establishments (Anonymous
2007b) provides generic risk data for certain categories
of establishments.
Important in view of the possible uses that can
be made of the Register is the fact that the data to be
submitted for establishments should be based on the
public part of Environmental management Act licence.
This means that not the actual presence of hazardous
substances is registered, but rather the maximum
amounts of all substances that are laid down in the
license.

2.2 Regulation provincial risk map

Since the fireworks disaster the municipalities have
the duty to inform their citizens on all possible risks

1114

in their territories. For this purpose Internet maps

have been developed. The government provided the
functional design and the provincial authorities were
made responsible for embedding the maps in their
provincial IT environments.
The Regulation provincial risk map (Anonymous
2007c) prescribes that competent authorities submit
data to the RRGS on categorized situations with
hazardous substances and to ISOR on other risk situations (e.g. floods, collapse of tall buildings). The
situations with hazardous substances are less dangerous (main effect distances extending beyond the
establishments limits) than the situations designated
in the above described Registration Order external
safety (see Section 2.1) (individual risk contour of
106 per year extending beyond the establishments
limits).
The Regulation requires also that vulnerable objects
(e.g. schools, hospitals) are to be registered in ISOR
and shown on the maps.
It charges the provinces with the management of
their individual risk maps.

THE RRGS APPLICATION

3.1 Structure
RRGS has one database, containing three sets of data
on risk situations with hazardous substances. The first
set consists of the mandatory registration of situations based on the Registration Order external safety.
The second set contains the mandatory registration
of situations based on the Regulation provincial risk
map. The third set is relatively small and comprises
optionally registered situations (main effect distance
not extending beyond the establishments limits; see
Section 2.2). E.g. emergency response organisations
may use the data in the third set for occupational safety
measures.
The database is filled through an Internet application. Competent authorities can log in to a protected
site. In all cases sets of screens containing compulsory and optional fields are to be filled. There
are screens for general information on the establishment, including a drawing facility for the border of the premises of the establishment. Next the
type of establishment is specified through screens
which define for the installations present the category from the Registration Order or the Regulation.
For the installations, amongst others, the location,
substances present will be filled in. Finally there
is a set of screens for additional information on
the establishment, like the applicability of other
licences or the presence of a plan of attack for the
fire brigade. For transport routes comparable sets of
screens apply.

3.2 Adaptation of the RRGS to the new legislation

The final pieces of legislation necessitated the RRGS
to be reworked considerably. A few categories of
establishments were removed from the Registration
Order external safety. A large number of new situations had to be added based on the Regulation
provincial risk map. Some new functions were introduced and flaws were restored. At the start of the
project a thorough definition of the new screens
was achieved in 3 days. The work was done in 4 increments, with testing in between. Thus improvements
and solutions for flaws could be taken to the next increment, saving time in the final acceptation phase. On
behalf of the conversion (see below) the changes in
the data model and database were documented meticulously. This huge project was completed (almost)
within the agreed time (7 months).
Next, the data form the old database had to be
converted to the new database. It took a lot of effort
to prepare the scripts for the conversion programme.
The compulsory fields of each record were distributed
to the competent authorities in the form of Excel
spreadsheets. These fields from the old database were
complemented with new compulsory fields for the
new database. The competent authorities were asked
to fill in these new fields and to fill in empty fields
from the old database. Next, the spreadsheets were
returned and converted.
Only about 40% of the competent authorities
returned completed spreadsheets. The main argument
for non-cooperation was lack of time. The conversion
team had communicated that unprocessed spreadsheets would be converted, but still had to be processed
afterwards. Moreover, during conversion the status of
the record would be set back. That means that the
record would have to be re-authorized to publish the
data on the provincial risk map.

3.3 Filling
For establishments the authority granting the Environmental license is responsible for submission of the
registration for that establishment. In almost 300 cases
the authority is the central government (Ministry of
Housing, Spatial Planning and the Environment for the
larger nuclear installations and certain defence establishments; Ministry of Economic Affairs for onshore
mining industry establishments), in almost 700 cases
(in general large establishments) a province and in
most cases (about 8500) a municipality. For transportation of dangerous substances via roads, railways
or inland waterways, the authority that should submit the registration is the maintenance authority; for
pipelines the Ministry of Housing, Spatial Planning
and the Environment.

1115

Figure 2. Provincial risk map (province of South Holland near the town of Dordrecht) showing vulnerable objects and
establishments with individual risk contours. The small map at the top left-hand side is part of the search facilities. The map
legend at the right-hand side is adaptable to the users wishes with +/ buttons and check boxes.

Assuming optional registration (see Section 3.1) to

be nil, the number of registrations based on the Registration Order external safety is about 45% of the
total of 9500 establishments and based on the Regulation provincial risk map about 55%. All numbers
and percentages above are based on the RRGS filling
in February 2008.

4
4.1

RISK MAPS
General description

The contents of the RRGS database are synchronized

daily with a central provincial database. Data from
the ISOR application are sent directly to this central
database. Each province takes the data for its own territory from the central database plus data from the
neighbouring provinces up to a maximum of 15 km
from its border to fill the rectangle of the computer
screen.
On the public maps only authorized situations are
shown (see Figure 2).
All 12 provincial risk maps can be accessed through
a portal site. The portal provides the users with information and explanation on the types of risks, risk data,
legislation and 12 links to the provincial sites.
Originally the RRGS would have had its own map.
However, during the development of RRGS, ISOR
and risk maps it was decided to have only one kind

of risk map, viz. the provincial risk map with all risks
and vulnerable objects shown.

4.2

Use of the risk maps and underlying data

The aim of the register is to give insight and overview

of risk situations in The Netherlands. For the public
that will mean a survey of risk situations e.g. in ones
neighbourhood. The average number of hits for the
maps for the general public is about 4000 per day.
Commercial data brokers have been granted access
to the (publicly available) data. They use specific
sections of the risk maps e.g. for information folders
for clients of real-estate agents.
The use of the professional maps is restricted to
civil servants that have been granted access rights.
These rights range from right to look at the maps and
underlying data to right to submit or change data
to right to authorize data to superuser rights. In
general staff from environmental departments are submitting and authorising the data. Their colleagues from
the same department, planners from fire brigades or
ambulance services and employees from spatial planning use the maps for planning purposes. Note that the
use of the RRGS and maps is limited for e.g. firemen
as the registration comprises the worst case situation
as recorded in the environmental licence. It does not
show actual amounts of and categories of hazardous
substances that are present during an actual incident.

1116

Figure 3. Floodscenario shown on the provincial risk map (same as Figure 2). The legend shows under the headings
Natuurrampen (Natural disasters)and Overstromingsdiepte (Flood depth) water depth ranges (minder dan = less than;
meer dan = more than).

4.3

Floods

The European Directive on the assessment and management of flood risks (Eur-lex 2007) requires all state
members to have flood maps by 2013 and disaster
control programmes for floods by 2015. A layer with
potentially floodable area has been added to the provincial risk maps as a first anticipation on the legislation.
Per dike district the water depth has been calculated
for squares of 50 by 50 m. These are shown on the map
layer in different shades of blue per depth range (see
Figure 3).
In the future more map layers concerning floods will
be added, showing e.g. flooding risk or a simulation
of a certain flooding scenario (collapse of dike X at
location Y).

DISCUSSION

The RRGS is a useful instrument for the mandatory

registration of risk situations involving hazardous substances, since the recent modifications of the application. The situations to be registered are establishments
storing, handling or processing of hazardous materials
or their transport.
Conversion of a database after such a large modification of the application is difficult and requires

much preparation. In retrospect and seen from the

project of building a registration and a display system, it would have been wiser to integrate the pieces
of legislation involved and to develop from such legislation an integrated application for data submission,
database storage and display.
The display tool for the risk situations, the provincial risk map, gives insight and overview to the public
and to professionals for these situations, especially
because vulnerable objects are shown as well.
It is a useful planning tool. The RRGS registers
the worst-case situation from the license. It is in the
interest of licensees to apply only for relatively small
amounts of hazardous substances and few numbers of
different chemicals in their license-applications. Of
course the usefulness of the system depends on the
completeness of the mandatory data and the presence
of all risk situations and vulnerable objects. The legislation provides for terms in which the records should
be adapted due to changes in the license. Every 5 years
all records are to be reviewed.
The addition of the flooding layer in the risk map
shows that the system is versatile. The system can be
expanded with other risk situations.
The RRGS itself can be combined with other
applications: E.g. real estate information, crisis management, emergency planning.

1117

REFERENCES
Ale, B.J.M. 2002. The explosion of a fireworks storage
facility and its causes. In E.J. Bonano et al. (eds.)
PSAM6, 6th International Conference on Probabilistic
Safety Assessment and Management, San Juan, Puerto
Rico, USA, June 2328 2002: 8792. Oxford: Elsevier.
Anonymous 2005. Staatsblad 2005, 483. Wet milieubeheer,
titel 12.2 (Registratie gegevens externe veiligheid
inrichtingen, transportroutes en buisleidingen).
Bulletin of Acts, Orders and Decrees 2005, 483. Environmental management Act, title 12.2 (Registration of
external safety data on establishments, transport routes
and pipelines).
Anonymous 2007a. Staatsblad 2007, 102. Registratiebesluit
externe veiligheid.
Bulletin of Acts, Orders and Decrees 2007, 102. Registration Order external safety.

Anonymous 2007b. Staatscourant 2007, 249. Regeling

externe veiligheid inrichtingen.
Gouvernement Gazette 2007, 249. Regulation external
safety of establishments.
Anonymous 2007c. Staatscourant 2007, 75. Regeling provinciale risicokaart.
Gouvernement Gazette 2007, 75. Regulation provincial
risk map.
Eur-lex, (http://eur-lex.europa.eu) Directive 2007/60/EC of
the European Parliament and of the Council of 23 October
2007 on the assessment and management of flood risks
(Text with EEA relevance).
Van t Sant, J.P. et al. 2003. The new Dutch Register for risk
situations involving hazardous substances. In T. Bedford
and P.H.A.J.M. van Gelder (eds), Safety and Reliability:
13831388. Lisse: Swets and Zeitlinger.

1118

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Experimental study of jet fires

M. Gmez-Mares, A. Palacios, A. Peiretti, M. Muoz & J. Casal
Centre for Studies on Technological Risk (CERTEC), Department of Chemical Engineering,
Universitat Politcnica de Catalunya, Barcelona, Spain

ABSTRACT: Jet fires have been frequently the first step of severe accidents which have involveddue to the
domino effectfurther explosions or large fires. Nevertheless, the current knowledge of jet fires is still rather
poor. Several authors have published experimental data, generally from small scale jet fires or from flares. In
this communication, the experimental results obtained with relatively large jet fires (with flames length up to
10 m) are discussed. The fuel was propane, and both sonic and subsonic jet exit velocities were obtained from
different outlet diameters. The distribution of temperatures on the flame main axis was measured with a set of
thermocouples. The jet fires were filmed with a videocamera registering visible light (VHS) and a thermographic
camera (IR). The main flame geometrical features were analyzed as a function of the fuel velocity, mass flow
rate and jet outlet diameter: lift-off, flame shape, flame size.

INTRODUCTION

Fires are the most frequent accident that occurs in

industrial plants or in the transportation of hazardous
materials: 59% of these events are fires, followed by
35% for explosions and 6% for gas clouds (GmezMares & Casal, 2008). Often, fire accidents have a
shorter damage radius than those corresponding to
other major accidents (explosions, toxic clouds), their
dangerous effects being restricted to a relatively small
area. Nevertheless, these areas typically contain other
equipment (pipes, tanks, etc.) that can be damaged and,
thus, incorporated into the accident via the domino
effect. This is especially important in plants with a
compact layout such as offshore oil platforms and
some processing and storage plants.
Among the different fire accidents, jet fire direct
effects are the least severe, due to their relatively
reduced size as compared to a pool or a flash fire.
However, jet fires can affect equipment and, as they
can be locally very intenseespecially if there is flame
impingementthe equipment may subsequently fail
and ultimately amplify the scale of the accident. In fact,
jet fires have been in many cases the first step of severe
accidents which have involveddue to the domino
effectexplosions or large fires: a recent historical
analysis (Gmez-Mares & Casal, 2008) has shown
that among the accidents registered in the data bases,
in approximately 50% of the cases in which there was
a jet fire it caused another event with severe effects.
Nevertheless, the current knowledge on the main
features and behaviour of jet fires is still rather poor.
Several authors have published experimental data,

although many of them were obtained from small scale

jet fires or from flares, and most of them at subsonic jet exit velocity, i.e. at conditions quite different
from those found in a real accidental jet fire. Other
authors have published mathematical models to estimate the shape and size of the flames. However, most
of these models assume shapes (for example, a frustum of a cone) that do not correspond to those found
in accidental jet fires but to other phenomena such
as, for example, flares. Finally, rather scarce information has been published on the thermal features of jet
fires: flame temperature distribution, emissivity and
emissive power of the flames, which are especially
interesting to estimate the heat flux reaching a given
target.
An experimental study has been developed analyzing the main features of relatively large jet fires (flames
length up to 10 m). The fuel was propane, and both
sonic and subsonic jet exit velocities were obtained
with several outlet diameters. The distribution of temperatures on the flame main axis was measured with
a set of thermocouples. The jet fires were filmed with
a video-camera registering visible light (VHS) and a
thermographic camera (IR). The main geometrical features of the flames were analyzed as a function of the
fuel mass flow rate and the jet outlet diameter: lift-off,
flame shape, flame size length.
2

EXPERIMENTAL SET-UP

The experimental facility was built at the Security Formation Centre Can Padr. A schema of the field

1119

test apparatus is shown in Figure 1. The facility consisted of a set of pipes which allowed obtaining vertical
and horizontal jet fires. This paper concerns only to
vertical jet fires.
The gas pipe exit had an interchangeable cap which
allowed selecting different exit diameters (ranging
between 10 mm and 43.1 mm). In addition, pressure
measurements were taken at the gas outlet in order
to calculate the mass flow rate. The fuel (commercial propane) was contained in a tank located on an
upper site.
Jet flame geometric parameters were studied analyzing the images filmed by two video cameras located
orthogonally to the flame.
An AGEMA 570 Infrared Thermographic Camera
(IR), located next to one of the previously mentioned
video cameras, was used to know the temperature and
radiation distribution of the jet flame, and also to compare the geometric parameters obtained by the video
cameras. The IR camera vision field was 24 18 ,
the spectral range 7.5 to 13 micrometers.
Type B (Pt 30% Rh / Pt 6% Rh) and S (Pt 13%
Rh /Pt) thermocouples were used to measure the flame
axial temperature distribution. They were arranged in
a mast as shown in Figure 1. A meteorological station was used to measure the ambient temperature, the
relative humidity and the wind direction and velocity.
A Field Point module was employed as a collection data system. It is formed by a communication
module FP-1001(RS-485,115 kbps), three connection terminals FP-TB-1 and three input/output (I/O)
modules. The diverse measurement devices (thermocouples, pressure gage, etc.) were connected to this
system.
Two laptops were used to collect the data from the
different sensors. They controlled the measurement

Infrared and Video Cameras

Pressure gage entrance

Thread

Cap
Thermocouples Mast

Figure 1.

Experimental set-up.

Propane

devices operation. The IR camera, the meteorological

station and the Field Point were connected to them.
Furthermore, the two laptops were connected
through a network in order to synchronize the data
acquisition.
3

SHAPE AND SIZE OF A JET FIRE

3.1 Jet exit velocity

Jet fires are originated by the release and ignition of
a flammable gas or vapour through a hole, a broken
pipe, a flange, etc., or in process flares.
The jet exit velocity has a certain influence on some
of the features of the fire (flames length, lift-off). This
velocity will increase with the pressure inside the container or the pipe, finally reaching the sonic velocity,
i.e. the velocity of sound in that gas. This represents
the choked condition: further increase in pressure will
not modify the so-called critical velocity of the gas.
Thus, two regimes can be found: subsonic flow and
sonic flow. In the flow of a gas through an opening, the
sonic velocity is reached if the following relationship
is fulfilled:
 1

P0
2

Pin
+1

(1)

For most gases, the sonic velocity is reached if the

pressure at the source is greater than 1.9 bara. Therefore, accidental jet fires are often sonic. However, most
of the authors who have studied jet fires have worked
with subsonic flow.
3.2 Flame shape
The shape of jet fires is strongly influenced by the
direction of the jet (vertical, horizontal or inclined),
by the eventual existence of wind and by the jet exit
velocity. Kalghatgi (1983) and Chamberlain (1987)
assumed that the jet fire was a frustum of a cone,
more or less inclined according to the wind intensity.
They analyzed experimental data obtained from large
scale flares with rather low exit velocities. However,
in vertical jets with sonic gas exit velocity, the shape is
somewhat different and corresponds rather to that of a
spindle (Figure 2). In this case, a simplified approach
is to consider it as a cylinder with a given length and
an average diameter.
In the case of horizontal jet fires, two zones must
be considered. In the first one, dominated by the high
exit momentum, the jet is essentially horizontal with a
spindle/cylinder shape; however, when the velocity
in the jet decreases, the buoyancy has an increasing influence and the jet moves towards the vertical
direction.

1120

100
90
80
70
60
50
40
30
20
10
0

d (mm)

S /d

43.1
30
20
15
12.75

sonic

0.1

0.2

0.3

0.4

0.5

0.6

-1

m (kg s )
Figure 3. Variation of the lift-off as a function of the
fuel mass flow rate for different outlet diameters, for both
subsonic and sonic flow.

of the flame and the distance over which there can be

flame impingement.
The results obtained with different outlet diameters
and at different mass flow rates can be seen in Figure 3.
Lift-off is clearly a function of the orifice diameter
and the jet velocity (or the mass flow rate for a given
diameter). In this figure, data at both subsonic and
sonic conditions have been plotted. The trend is the
same for both regimes.
Diverse authors (Peter & Williams, 1983; Santos
and Costa, 2005; Kiran and Mishra, 2007) have proposed equations for predicting the lift-off as a function
of orifice diameter and gas exit velocity, according to
the following equation:
S
=c
d
Figure 2.

A vertical jet of propane (sonic flow).

In all cases, large jet fires show a significant turbulence which creates some difficulties to establish the
flame length.

3.3

Lift-off

The lift-off is the distance between the gas outlet and

the start of the flame. This phenomenon occurs due
to the fact that at the jet exit the average velocity
exceeds the burning velocity; at a farther downstream
position, equilibrium is reached between both velocities and the flame is stabilized. The existence of a
zone in which the fuel-air mixture is not between the
flammability limits can also have an influence.
The prediction of lift-off is interesting because,
together with the flame length, determines the position

V
d


(2)

In this expression, the constant c has the dimension of time. For the present study, the experimental
subsonic data could be correlated with relatively good
accuracy with c = 0.0027 s: the data corresponding
to the orifice diameters ranging between 12.75 mm
and 30 mm followed the same trend in a S/d vs. V /d
plot, but the data for d = 43.1 mm gave larger values
of S/d.
Eq. (2) can only be applied at subsonic conditions, as gas velocity does not increase any more once
the sonic flow has been reached. However, ifonce
the sonic velocity has been reachedthe pressure
upstream the jet outlet rises, the lift-off length becomes
larger. This is due to the increase in the gas density
upstream the orifice which finally leads to a larger
fuel mass flow rate.
Nevertheless, most accidental jet fires occur at exit
sonic velocity. Thus, an equation allowing the calculation of flame lift-off at sonic conditions would be of
outmost interest.

1121

3.4

Flame length

Diverse authors have studied the variation of jet fire

length with some operating variables (orifice diameter, jet velocity, etc.) (Suris et. al. 1977, Sonju &
Hustad 1984, Santos & Costa 2005, Kiran & Mishra
2007). These experimental works were performed
mostly with small scale jet fires, with flame lengths
up to 2 m. Only Sonju & Hustad (1984) worked with
larger flames (up to 8 m).
All these authors proposed the same expression to
predict the flame length as a function of Froude number and orifice diameter over the buoyancy-dominated
flow regime:
L
= A Fr n
d

(3a)

According to them, when applied to higher jet

velocities, i.e. over the momentum-dominated regime,
n would become 0 and L would be independent
from Fr. Table 1 summarizes the information about
Table 1. Values of A, n and Fr ranges (Eq. (3-a)) based on
experimental data.
Author

Fuel

Suris et. al.

(1977)

propane
methane
hydrogen

Sonju and
Hustad
(1984)
Santos and
Costa
(2005)
Kiran and
Mishra
(2007)

Froude
number range
up to 3 104

40
27 and 29 0.2
14 and 16

(3b)

Taking into account the agreement with the present

experimental data, this expression would thus be suitable for predicting the length of flames for Froude
number values up to approximately 1.75 105
(Figure 4).
However, in this work, the results for propane
flames were extended to the regime where the sonic
exit velocity is reached. And it should be noted that
for a given value of Fr (i.e. for a constant sonic velocity
of the gas at the orifice) and an orifice outlet diameter,
larger flame lengths can still be obtained (Figure 4).
This is due to the fact that for a given outlet diameter,
flame height increases with mass flow, as a result of
the increment in the gas density inside the pipeline.
Again, additional research would be required in the
sonic flow regime.
4

FLAME TEMPERATURE

up to 1 105

27
21
36
24

0.2

0.2

The temperature of the flame is important because it

is directly associated to the radiative characteristics of
the fire. The thermal radiation emitted from the fire is
usually estimated from the emissive power, which is a
function of the flame temperature and emissivity.

LPG

up to 4.5 104 30

0.2

4.1 Temperature heterogeneity

up to 2 104

d (mm)

L/d

L
= 53 Fr 0.12
d

propane
methane
propane
ethylene

1,000
43.1
30
20
12.75
10

53Fr

0.12

100

10
100

the type of fuel and the Froude number range of

applicability.
As for the present results, the subsonic propane data
obtained when plotting L/d versus Fr (Figure 4) were
also correlated according to the same expression.
In the buoyancy dominated regime the experimental
data L/d can be correlated with:

1,000

10,000

100,000

1,000,000

In large accidental fires, the temperature is not uniform over the flame due to the turbulence and to the
existence of fluctuating luminous, less-luminous and
non-luminous parts. This is especially significant in
pool fires, in which the combustion is rather poor
Muoz et al., 2004).
In accidental jet fires, due to the turbulence of the
phenomenon and the important entrainment of air by
the jetand to the fact that the fuel is often a gasthe
combustion is much more efficient and thus the variation of the temperature is not as significant as in a
pool fire.

Fr = V /g d

4.2 Axial temperature distribution

Figure 4. Variation of normalized flame length as a function

of Froude number: experimental results (subsonic and sonic
velocity) and Eq. (3-b) (subsonic data).

Information concerning the distribution of temperatures over the jet fire is rather scarce in the literature.
Frequently it is not well detailed and generally regards

1122

only to small or medium size flames (Santos and Costa,

2005; McCaffrey, 1989; Becker & Yamazaki, 1978).
To measure the variation of the temperature in the jet
fires, a set of four thermocouples were located along
the jet axis. The measurement was especially difficult
due to the turbulence and fluctuation of the jet fire, and
consequently the experimental data showed always a
certain scattering.
Figure 5 shows a typical temperature distribution
as a function of the jet fire axial position. It can be
observed that the temperature in the jet axis increases
with height, reaching a maximum at an intermediate height (at approximately 60% of the visible flame
length) and then decreasing again; the temperatures
over the top zone of the flame are significantly higher
than at its bottom.
4.3

Temperature distribution over the jet

Figure 6 shows an infrared image of a propane jet fire.

The highest temperatures are near the jet axis and at the
bottom and top of the flame the temperatures are lower.
1800

Temperature (C)

1600
1400
1200
1000
800
600
400
200
0
0

20

40

60

80

100

% Flame axial position

Figure 5. Flame temperature variation as function of jet fire

axial position.

The maximum axial temperatures measured were

around 1600 C for mass flow ranging between 0.06
and 0.28 kg s1 , and the lower ones corresponded to
0.29 and 0.43 kg s1 .
5

Although jet fires can play a significant role as a first

step in a domino effect, the literature survey has shown
that the existing knowledge on their main features is
still rather poor. The number of experimental studies published with large fires is reduced. Furthermore,
most of these works were performed only with subsonic flow, while accidental jet fires involve usually
sonic gas jets.
The results obtained with relatively large vertical
jet fires (flame length up to 10 m) have shown that
they can be considered to have a spindle/cylinder
shape, subjected however to certain turbulences and
intermittences.
For subsonic flow, the lift-off could be correlated
for most of the orifice diameters by Eq. (2), as proposed already by other authors, with c = 0.0027. This
expression is restricted, however, to subsonic flow and
can not be applied to sonic flow, at which the velocity V has a constant value but the mass flow and the
lift-off distance still increase.
As for the flame length, the data obtained at
subsonic conditions could be correlatedat the
buoyancy-dominated regimeas a function of Froude
number (Eq. (3-a)). Again, this expression could not
be applied to sonic flow. The results obtained show
clearly that the length of the jet fire continues increasing at sonic flow, when the mass flow rate is increased
(i.e. when more fuel is feeded to the fire). This indicates again that more research effort is required in the
sonic jet fire field.
Finally, the values obtained when measuring the
axial flame temperature distribution clearly show that
the flame temperature of a jet fire can not be assumed
to be uniform: it increases from the bottom of the
flame, reaching a maximum and then decreasing
again. The distribution can be represented by a polynomial equation, which reaches its maximum temperature at the middle of the jet flame. The maximum
temperatures reached were about 1600 C. This variation of temperature can be important when estimating
the emissive power of the jet.
6

Figure 6. Infrared image of a vertical jet fire of propane

(sonic velocity).

CONCLUSIONS

A
c
d
Fr
g

1123

NOTATION
constant in Eq. (3-a)
constant (s)
orifice or outlet diameter (m or mm)
Froude number (V 2 g 1 d 1 ) ()
acceleration of gravity (m s2 )

L
m
n
Pin
P0
S
T
V

length of the flame (m)

fuel mass flow rate (kg s1 )
constant in Eq. (3-a)
pressure inside the pipe or the tank (bara)
atmospheric pressure (bara)
lift-off distance (m)
temperature (K or C)
velocity in the jet at the gas outlet (m s1 )
ratio of heat capacities, cp /cv ()

ACKNOWLEDGEMENTS
The authors acknowledge the Spanish Ministerio de
Educacin y Ciencia (project no. CTQ2005-06231 and
a doctoral grant for M. G-M) and the Autonomous
Government of Catalonia (doctoral grant for A. P) for
funding.
REFERENCES
Becker, H.A. & Yamazaki, S. 1978. Entrainment, Momentum
Flux and Temperature in Vertical Free Turbulent Diffusion
Flames. Combustion and Flame 33: 123149.
Chamberlain, G.A. 1987. Developments in design methods
for predicting thermal radiation from flares. Chemical
Engineering Research and Development 65: 299309.

Gmez-Mares, M., Zrate, L. & Casal, J. 2008. Jet fires and

the domino effect. Fire Safety Journal (in press).
Kalghatgi, G.T. 1983. The visible shape and size of a turbulent hydrocarbon jet diffusion flame in a cross-wind.
Combustion and Flame 52(1): 91106.
Kiran, D.Y., Mishra, D.P. 2007. Experimental Studies of
flame stability and emission characteristics of simple LPG
jet diffusion flame. Fuel 86: 15451551.
McCaffrey, B.J. 1989. Momentum Diffusion Flame Characteristics and the Effects of Water Spray. Combust. Sci. and
Tech. 63: 315335.
Muoz, M., Arnaldos, J., Casal, J., Planas E. 2004. Analysis of the geometric and radiative characteristics of
hydrocarbon pool fires. Combustion and Flame 139,
263277.
Peters, N., Williams, F.A. 1983. Liftoff characteristics
of turbulent jet diffusion flames. AIAA Journal 21(3):
423429.
Santos, A., Costa, M. 2005. Reexamination of the scaling laws for NOx emissions from hydrocarbon turbulent jet diffusion flames. Combustion and Flame 142:
160169.
Sonju, O.K., Hustad, J. 1984. An experimental study of turbulent jet diffusion flames. Norwegian Maritime Research
4: 211.
Suris, A.L., Flankin, E.V., Shorin, S.N. 1977. Length of
free diffusion flames. Combustion, Explosion, and Shock
Waves 13 (4): 459462.

1124

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Failure mode and effect analysis algorithm for tunneling projects

K. Rezaie, V. Ebrahimipour & S. Shokravi
Department of Industrial Engineering, Faculty of Engineering, University of Tehran, Tehran, Iran

ABSTRACT: This paper is an approach for tunneling risk management to have safe and sustainable tunnel
for people, equipment and of course employees. Identifying hazard, categorizing the risks according to their
frequency and consequences and employing risk assessment methods to reduce, diminish or accept those risk,
help to reach this goal. For this purpose an upgraded algorithm for FMEA, which will be more common for all
phases of tunneling process and more both qualitative and quantitative than conventional FMEA, is used with an
eye on specific attributes of each road or metro tunnel. However different parts of a tunnel, existence or absence
of them in a tunneling project are concerned to make the options for each and every tunnel customized.

INTRODUCTION

There is one reason for people to be afraid underground spaces, which of course, is the lack of natural
light decreases the ability of orientation in time and
place. Another major reason for negative reactions
toward traffic tunnels and underground space is the
fear of accidents and fire in an enclosed area [1].
Therefore men reckon to decrease these incidents
and accidents in underground environments to defeat
the fear. Identifying hazards which lead to the accidents is first step of this process, and then is looking
for control or diminishing approaches to deal with
them which need various methods of identification,
assessment, analysis and at last management.
There are a numerous type of papers who address
different risk assessment and management methods
but only some of them were follow the same goal of
this paper.
In Korea for two tunnels (14-3, 13-3) diagnose and
note environmental risk such as ground, water leakage,
faults, etc from design to construction stages and with
appropriate risk analysis methods try to rank the identified risks and manage them to the acceptable level or
diminish them if it is possible [2].
Okazaki, et al, [3] in Japan assessed the convergence displacement and settlement, as contained in
the tunnel construction records, with lithofacies confirmed by boring cores. In addition, magnetic susceptibility of boring cores was measured, and a comparative
study was conducted regarding its correlation with
magnetic intensity distribution by helicopter magnetic
survey in helicopter borne survey.
Konstantinos, et al, [4] in Athens University
present a paper which object is to provide a generic

framework for developing and implementing a Tunnel

Safety Management System (TSMS). The development of the content and structure of the proposed
TSMS was based on: i) the plan, do, check, act
management cycle, ii) the tunnel lifecycle elements
(i.e. design, construction, operation, maintenance, and
refurbishment), and iii) the prevailing legal, institutional, organizational, operational and technological
environment. The implementation of the proposed
approach led to the identification of a set of guidelines
for developing a TSMS.
NG, et al, [5] in Hong Kong presents a paper that
describes how a risk-based maintenance management
model was formulated to systematically prioritize PM
activities. The model was based on the five core
elements of the risk management process (RMP): identification, measurement, assessment, valuation, and
control and monitoring. This model was applied to a
toll road/tunnel company in Hong Kong to enhance
the PM operations of its lighting system.
2

RISK MANAGEMENT

Risk management is a central part of any organizations strategic management. It is the process whereby
organizations methodically address the risks attaching
to their activities with the goal of achieving sustained
benefit within each activity and across the portfolio
of all activities. The focus of good risk management
is the identification and treatment of these risks. Its
objective is to add maximum sustainable value to
all the activities of the organization. It marshals the
understanding of the potential upside and downside
of all those factors which can affect the organization.

1125

It increases the probability of success, and reduces

both the probability of failure and the uncertainty of
achieving the organizations overall objectives.
Risk management should be a continuous and
developing process which runs throughout the organizations strategy and the implementation of that
strategy. It should address methodically all the risks
surrounding the organizations activities past, present
and in particular, future. It must be integrated into the
culture of the organization with an effective policy and
a programmed led by the most senior management.
It must translate the strategy into tactical and operational objectives, assigning responsibility throughout
the organization with each manager and employee
responsible for the management of risk as part of
their job description. It supports accountability, performance measurement and reward, thus promoting
operational efficiency at all levels.
In other word risk management means to take
rigorous decisions to enhance success, good effect
and decrease consequences. It is astute planning to
identify, quantify, manage and assess issues and
applications could end up with a financial loss.
Conventional risk management applied only on
insurable risks. But modern one concentrates on either
insurable or uninsurable risks which are affect costs
and benefits.
3

RISK ASSESSMENT

Risk assessment is employed to determine system

safety [6]. It is two-stage process [7]. In the first stage,
a risk analysis process defines the scope of the risk
assessment, identifies information resources (assets),
and determines and prioritizes risks to the assets. In the
second stage, a risk management process makes decisions to control unacceptable risks; a decision may be
to transfer a risk (for example via insurance), ignore
a risk, or reduce a risk via selection of appropriate
safeguards.
4

FAILURE MODE AND EFFECT ANALYSIS

Failure mode and effect analysis is yet another

powerful tool used by system safety and reliability
engineers/analysts to identify critical components/
parts/functions whose failure will lead to undesirable
outcomes such as production loss, injury or even an
accident. The tool was first proposed by NASA in
year 1963 for their obvious reliability requirements.
Since then, it has been extensively used as a powerful technique for system safety and reliability analysis
of products and processes in wide range of industriesparticularly aerospace, nuclear, automotive
and medical. The main objective of FMEA is to

discover and prioritize the potential failure modes (by

computing respective RPN), which pose a detrimental
effect on the system and its performance. The results
of the analysis help managers and engineers to identify the failure modes, their causes and correct them
during the stages of design and production.
The main disadvantage of RPN analysis is that various sets of failure occurrence probability [Of ], severity
[S] and detect ability [Od ] may produce an identical
value, however, the risk implication may be totally
different which may result in high-risk events may go
unnoticed. The other disadvantage of the RPN ranking method is that it neglects the relative importance
among Of , S and Od . The three factors are assumed to
have the same importance but in real practical applications the relative importance among the factors exists.
To address these disadvantages related to traditional
FMEA, an algorithm is revealed to simplify the steps
and use occurance-severity diagram to meet the rigorous assessment of tunneling projects that is one of the
crucial projects for human safety ever.
5

FACTORS IN THE SELECTION OF RISK

ASSESSMENT METHOD

According to numerous works carried out by several

authors after summarize interrelated factors there are
seven factors that are more important to organizations,
contractors and experts than other factors.
Factors are:

usability
credibility
complexity
completeness
adaptability
validity
and cost
RISKS OF UNDERGROUND ENVIRONMENT
CONSTRUCTION

In subject of risk management, at first it is important

to define and categorize all kind of risks and with this
approach clear their scope of management. According
to several types of risks, the management of them is
various as well. Project risks define in overall point of
view and related to cost and time issues although there
are several methods and tools for them after design
stage and during construction phase.
In 2004 Raley and Bown enumerate tunneling risk
as below:
Risk of loss or damage with a potential of death or
injury, economical and equipment failure and credit
loss for others involved in the project.

1126

12

1. Tunn el attributes:
Dimensions
Equipment
Human and machinery traffic
Tunnel plan
Geological and geotechnical information

occurrence

10

2. Define project objectives and ultimate goal

Figure 2.

10

Occurrence-severity diagram.

HAZARDS FOR ROAD, RAILWAY/SUBWAY

TUNNELS AND UNDERGROUND
PARKING SPACES

Seven countries- Australia, Canada, Germany, Italy,

Japan, Sweden, and the U.S.A- submitted information
for technical tables for road, railway/subway tunnels
and underground parking spaces [1]. The categories
addressed in the technical tables deal with seven
important aspects of safety:

End

Diminish

Yes
6. Minimize orinsurance,.

7. Appraisal
Is it Standard?
Is it safe according to international regulations?

Figure 1.

severity

No
No

Low

5. Select the method of risk management for all below mentioned

steps:
Purchasing
Work methods
Contractors
Before each shift
Reporting
Investigating incidents
Planning emergency

Vital?

Moderate

4. Risk matrix and RPN

To improve FMEA define a barriers to limit the detail scanning and
vast the point of view with an eye on cost and budgets.

Yes

High

3. Hazard Identification
3.1. Use workers and
Train them to get familiar to hazard and risk definition
Endorse their reports
Learn them standards and regulations
3.2. Categorization by tunnel users and groups affected by those risks
Owner
Contractor
Designer
Employees
Supplier
Workers
People

Acceptable?

Tunneling risk assessment approach.

Risk of not meeting defined standards and criteria

of design, operation, support and quality stages.
Risk of tardiness in due date of the project finish
which postpone profit making of it.
Risk of serious increase in project and support costs.
In other sources there are 3 different categories:
Financial loss of equipment, machineries, constructions.
Financial loss to third party assets
Injury of workers and third party

Fire and collapse-resistant structures;

Elimination of combustion materials;
Detection and alarms;
Compartmentalization;
Emergency systems(rescue, escape and refuge);
Fire and smoke control;
Maintenance of safety system.

A tunneling problem is simply a risk that is realized.

However, when risks are anticipated, their occurrence
and consequences can be reduced. This is an essence of
risk management, which to be successful must follow
from through risk assessment. Using risk assessment
and management techniques help tunnel owners and
constructors choose among alternative and complete
tunneling projects that have the lowest overall cost and
least effect on their community.
Two main categories of risk assessment methods
can be distinguished:
Deterministic safety assessment: the consequences
for loss of life of tunnel users and tunnel structure
are analyzed and assessed for possible accidents that
can be occur in a tunnel.
Probabilistic safety assessment: the consequences
for loss of life of tunnel users and tunnel structure and the frequency per year that these consequences will occur are analyzed. Consequences
and the frequency of the consequences are multiplied and presented in risk for the individual

1127

REFERENCES

Figure 3.

Risk matrix.

tunnel user, a societal risk and a risk for tunnel

damage.
8

CONCLUSION

Identifying hazard, categorizing the risks according to

their frequency and consequences is crucial for every
industry, hence an appropriate risk assessment method
should be selected to address those risks and be able
to introduce a suitable way to diminish them. An algorithm which is shown in figure 1, present 7 simple steps
to make a list of hazards in tunnels according to each
tunnel attributes, from early phase to the very end and
use RPN ranking analysis, occurrence-severity diagram (figure 2) and risk matrix (figure 3) to decide
the acceptable and unacceptable ones and based on this
categorizing, ignore the risk or employ an insurance
company to take care of it.

[1] Fire and Life Safety for Underground Facilities: Present

Status of Fire and Life Safety Principles Related to
Underground Facilities, ITA Working Group 4, Subsurface Planning, Arnica Nor mark, Animator.
[2] Korean Risk Management Practices: a Contractors
Perspective, Woong-Suk Yoo, President, SK E&C President, Korea, ITA-AITES World Tunnel Congress 2006
& 32nd General Assembly, Seoul, Korea, April 2006.
[3] Risk management for the new tunneling construction
using other tunnel records and helicopter borne survey
in accretionary complex, K. Okazaki, et al, Tunnelling
and Underground Space Technology 21(2006) 244.
[4] Development and Implementation of an Integrated Tunnel Safety Management System, Prof. Konstantinos,
et al, Transportation Systems and Logistics Laboratory,
Athens University of Economics and Business.
[5] A risk-based maintenance management model for toll
road/tunnel operations, M.F.NG, V.M. Rao Tummala
and Richard C.M. Yam, Engineering Department,
Rome 3 (CPS) Company Limited, NT, Hong Kong, Construction Management and Economics (July 2003) 21,
495510.
[6] System safety engineering and risk assessment, Nicolas
J. Baher, 2004.
[7] Factors in the selection of a risk assessment method,
Sharman Lichetenstein, 1996, Department of Inf. Sys.,
Monash Uni., Melbourne, Australia, Information Management and Computer Security, 2025.
[8] Risk Assessment Guidance for Tunnels, M. Molag,
I.J.M. Trijssenaar-Buhre TNO, Safe & Reliable Tunnels.
Innovative European Achievements.

1128

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Fuzzy FMEA: A study case on a discontinuous distillation plant

S.S. Rivera & J.E. Nez Mc Leod
CEDIAC Institute, Engineering Faculty, Cuyo National University, Argentina

ABSTRACT: Fuzzy FMEA provides a tool that can work in a better way with vague concepts and without
sufficient information but it is necessary to look for the suitable membership function to the study case. To
evaluate the revised matrix FMEA and fuzzy considerations, we present its application on a Discontinuous
Distillation Plant of biofuel. First, it is shown the matrix obtained from the ordinary method with severity
ranking and failure modes corresponding to the failure effects. Then, it is presented an FMEA in matrix form
and finally both of them are presented with fuzzy technique applied. The most important conclusions as a result
of the fuzzy techniques are shown and the most important result of the FMEA in a Discontinuous Distillation
Plan of biofuel is presented.

INTRODUCTION

Failure modes and effect analysis (FMEA) began in

army area and then it was applied in space, nuclear,
car manufacturers and nowadays in other areas.
The method is a procedure to analyze failure modes
and classified them by severity. It is a systematic process for identifying potential failures before they occur
with the intent to eliminate them or minimize the risk
associated with them. A group of experts make this
quantification gathering information from memory
and experience of the plant personnel.
The most known way to implement this analysis
is in an ordinary tabular form which is difficult to
trace. An FMEA worksheet is often arranged in a lot
of columns with inconvenient horizontal scrolling. In
order to eliminate this trouble a matrix method was
developed. The idea has already been explored for
different authors (Xu et al. 2002).
The matrix FMEA is a pictorial representation of
relationships between several FMEA elements. Traditionally, the numbers in the matrix are a prioritization of failures based on ranked numbers evaluating
concepts as severity, frequency of occurrence and
detectability of failure. These numbers are combined
in one number called Risk Priority Number.
The Risk Priority Number (RPN) methodology is a
technique for analyzing the risk associated with potential problems identified during a Failure Mode and
Effects Analysis (FMEA). The RPN for each issue is
calculated by multiplying Severity Occurrence
Detection.
Vague or ambiguous information and subjectivity in
the ranking scales adds inherent inconsistency. Some

authors eliminate this deficiency by introducing fuzzy

logic.
To evaluate the revised matrix FMEA and fuzzy
considerations, we present its application on a Discontinuous Distillation Plant of biofuel. First, it is shown
the matrix obtained from the ordinary method with
decreasing order of severity ranking and failure modes
corresponding to the failure effects. Then, causes of
failure are listed in decreasing order of magnitude of
criticality rank.
In fuzzy matrix application the analysis uses linguistic variables to describe the severity, frequency
of occurrence and detectability of failure. Different
types of membership functions and defuzzification
processes were used to evaluate the most appropriated
in this matter.
2

FAILURE MODE AND EFFECT ANALYSIS

Failure mode and effect analysis (FMEA) is a widely

used risk assessment tool to identify the potential
failure modes of a product or a process.
However, there could be several difficulties during
conducting conventional FMEA.
Thus, a new risk assessment system based on fuzzy
theory is proposed by some authors (Yeh & Hsieh
2007) to deal with these difficulties.
This technique is applied to a discontinuous distillation plant and it is presented in this work.
2.1

Traditional FMEA

There are several types of FMEAs; some are used

much more often than others. FMEAs should always

1129

be done whenever failures would mean potential harm

or injury to the user of the end item being designed.
The types of FMEA are:
Systemfocuses on global system functions
Designfocuses on components and subsystems
Processfocuses on manufacturing and assembly
processes
Servicefocuses on service functions
Softwarefocuses on software functions
This work is focused on components at design state
and follows the steps suggested by (Yeh & Hsieh 2007).
These steps are:
Collect the function of system and build a hierarchical structure, divide the system into several
subsystems, composing of components.
Determine the failure modes of each component and
its effects. Assign the severity rating (S) of each
failure mode according to the respective effects on
the system.
Determine the causes of failure modes and estimate the likelihood of each failure occurring, assign
the occurrence rating (O) of each failure mode
according to its likelihood of occurrence.
List the approaches to detect the failures and evaluate the ability of system to detect the failures prior
to the failures occurring. Assign the detection rating
(D) of each failure mode.
Calculate the risk priority number (RPN) and establish the priorities for attention.
Take recommended actions to enhance the performance of system.
Conduct FMEA report in a tabular form.
But with this ordinary tabular form, it is difficult to trace and/or cross-reference a particular failure
mode, cause or effect, because there are many-tomany relationships among failure modes, effects and
causes. The Matrix FMEA form overcomes this difficulty by storing all related qualitative and quantitative
information in a matrix (Sankar & Prabhu 2001).

Fuzzy sets have been introduced by Lotfi A. Zadeh

(Klir & Yuan 1995; Yager R. 1997) in 1965, as an
extension of the classical notion of set.
In classical set theory, the membership of elements
in a set is assessed in binary terms according to a bivalent condition; an element either belongs or does not
belong to the set.
By contrast, fuzzy set theory permits the gradual
assessment of the membership of elements in a set;
this is described with the aid of a membership function valued in the real unit interval [0, 1]. In fuzzy
logic the variables used are linguistic variables and are
described by membership functions. For that reason
fuzzy logic theory is adequate in this case.
Severity, occurrence and detection are the linguistic
variables (Devorak A. & Novk V. 2004) considered.
Each linguistic variable has four linguistic terms
to describe it. These linguistic terms are, Low (L),
Moderate (M), High (H) and Very High (VH).
Based in some authors (Xu et al. 2002; Yeh & Hsieh
2007) triangular fuzzy numbers were used in this work.
Membership functions (Anoop et al. 2006) are
shown in Figure 1.
Linguistic variable Risk (R) is obtained from fuzzy
rules (Dubois & Prade 1996) which are constructed
like is presented in Table 1.
Headings of this table are severity (S), occurrence
(O), detection (D) and riskness (R).
In the first column are numbers that represent fuzzy
rules used.
In this case of study 15 fuzzy rules were generated
to obtain fuzzy riskness.
These rules can be constructed by IF-THEN relationships. This approach will turn easy the future
analysis.
Table 1 was constructed from expert opinion in this
matter.

Membership function
1

2.2

Fuzzy FMEA

The conventional FMEA has been one of the wellaccepted reliability and safety analysis tool due to its
visibility and easiness. However, FMEA team usually suffers from several difficulties when conducting
FMEA in real industrial situation (Xu et al. 2002;
Yeh & Hsieh 2007).
A lot of difficulties are originated from the use of the
natural language due to the necessity to assign values
to those concepts.
Fuzzy sets are sets whose elements have degrees of
membership.

H
VH

Linguistic variable

10

Figure 1. Membership function of L, M, H and VH linguistic variables.

1130

Table 1.

Risk linguistic variable.

Rule #

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

L
L
L
M
M
M
M
H
H
H
H
VH
VH
VH
VH

L, M, H
VH
VH
L, M
L, M
H, VH
H, VH
L, M
L, M
H, VH
H, VH
L, M
L, M
H, VH
H, HV

L, M, H, VH
L, M
H, VH
L, M
H, VH
L, M
H, VH
L, M
H, VH
L, M
H, VH
L, M
H, VH
L, M
H, VH

L
L
M
M
M
M
M
H
M
H
M
VH
H
VH
H

Several defuzzification algorithms (Roventa E. &

Spircu T. 2003; Abbasbandya S. & Asadyb B. 2002)
have been developed. In this work, the center average
defuzzifier (Klir & Yuan 1995) was adopted due to its
simplicity.
In this approach the center of the fuzzy set is similar to the mean value of a random variable and it is
calculated by de following equation:

Value =

yi wi


(1)

DISCONTINUOUS DISTILLATION PLANT

OF BIOFUEL

Figure 2 shows the Discontinuous Distillation Plant

object of study in this work.
The Plant is located at the Engineering Faculty
Cuyo National UniversityMendozaArgentina.
The plant is composed by:
Boiler of 250 liters (10 KW)
Column of filling structured of 4 2.5 ms (1820
theoretical stages)
Condenser of helmet head and tubes 0.22 m2
Condensated buffer
Vacuum system (510 mmHg)
Distilled reception tanks
Account with a security system made up of:

Safty valve
Presostato
Pushbutton in plant for total cut
Fire protection system

Figure 2.

Discontinuous distillation plant.

This plant was built to be use in teaching, research

and services. It is a membership of the Bioenergy Program developed at Engineering Faculty, Cuyo National
University. The Bioenergy Program is concerned with
the generation of knowledge in connection with the
new sources of energy and focuses on how to develop
and exploit them. It also centers on human resources
and how to provide them with grade and post grade
levels of training so that they can access a new field
of study.
Each component was analyzed by traditional and
fuzzy FMEA.
4

REMARKS

In traditional FMEA the prioritization of failures is

performed based on the Risk Priority Number (RPN).
This number is mathematical product obtained like in
Equation 2.
RPN = S O D

(2)

As a result, some S-O-D scenarios produce RPNs

that are lower than other combinations, but potentially
more dangerous. (Sankar & Prabhu 2001) noticed as a
result, some S-O-D scenarios produce RPNs that are
lower than other combinations, but potentially more

1131

dangerous. This deficiency can be eliminated by introducing a new technique to calculate the Criticality
Rank using fuzzy logic. This concept was applied in
this work.
Other difficult in traditional FMEA is to trace
and/or cross-reference a particular failure mode, cause
or effect, because there are many-to-many relationships among failure modes, effects and causes. The
matrix form proposed by the same authors is a pictorial representation of relationships between several
FMEA elements.
Some authors introduce fuzzy logic for prioritizing
failures for corrective actions in a system FMECA and
FMEA.
The resulting fuzzy inputs are evaluated using a linguistic rule base and fuzzy logic operations to yield
a classification of the riskness of the failure and an
associated degree of membership in the risk class. This
fuzzy output is then defuzzified to give the Criticality
Rank for each failure.

RESULTS

It was considered 10 components called with upper

cases (A, B, C, D, E, F, G, H, I, J). Each failure mode
was called FX# , where the first subscript represents the
component and the second represents the number of
the failure mode.
With this notation traditional FMEA tables were
constructed. From these tables it was used fuzzy logic
to evaluate riskness and finally a matrix FMEA was
constructed.
In Table 2 the traditional FMEA is shown.
In this and next tables some columns without
interest for this discussion are excluded.

Table 2.

Example of traditional FMEA table.

Component

Failure mode

A
A
A
B
B
C
C
D
E
F
G
H
I
J
J

FA1
FA2
FA3
FB1
FB2
FC1
FC2
FD1
FE1
FF1
FG1
FH1
FI1
FJ1
FJ2

4
4
4
3
6
2
2
4
7
3
2
4
5
4
5

2
2
4
6
2
2
2
2
3
4
2
5
2
2
2

2
2
5
4
7
2
2
2
7
2
2
4
3
5
4

RPN
16
16
80
72
84
8
8
16
147
24
8
80
30
40
40

Table 3.

Fuzzy FMEA.

Component

Failure
mode

Fuzzy
S

Fuzzy
O

Fuzzy
D

Fuzzy
R

FA1
FA2
FA3
FB1
FB2
FC1
FC2
FD1
FE1
FF1
FG1
FH1
FI1
FJ1
FJ2

M
M
M
M
H
L
L
M
H
M
L
M
M
M
M

L
L
M
H
L
L
L
L
L
M
L
M
L
L
L

L
L
M
M
H
L
L
L
H
L
L
M
M
M
M

M
M
M
M
M
L
L
M
M
M
L
M
M
M
M

A
A
A
B
B
C
C
D
E
F
G
H
I
J
J

M
0
0
0
0
0
0
0
0
0

M
0
0
0
0
0
0
0
0
0

M
0
0
0
0
0
0
0
0
0

0
M
0
0
0
0
0
0
0
0

0
M
0
0
0
0
0
0
0
0

0
0
L
0
0
0
0
0
0
0

0
0
L
0
0
0
0
0
0
0

0
0
0
M
0
0
0
0
0
0

0
0
0
0
M
0
0
0
0
0

0
0
0
0
0
M
0
0
0
0

0
0
0
0
0
0
L
0
0
0

0
0
0
0
0
0
0
M
0
0

0
0
0
0
0
0
0
0
M
0

0
0
0
0
0
0
0
0
0
M

0
0
0
0
0
0
0
0
0
M

Applying fuzzy rules from Table 1 the following

table was obtained.
From Table 3 a fuzzy matrix was obtained and it is
shown in Table 4.
In Table 3 columns are different failure modes and
rows are each component.
Traditional matrix FMEA is shown in Table 4.
In Table 4 some highlighted elements looks more
important due to they are high values. Nevertheless in
Table 3 can be noticed that they have similar riskness
to others.
In particular, in the failure mode FE1 it is possible
to note that the riskness is not such important if it is
taking account that its detection is H.
One more step is recommended (Sankar & Prabhu
2001) in traditional matrix FMEA. It may be arrange in
decreasing order of severity ranking. By columns it is
used a critically rank in decreasing order too. With the
ordering used, the most critical areas are conveniently
found at the top left corner of the matrix where causes
have high rank for prioritization. The cell at the intersection of a failure effect and cause of failure shows
the Criticality Rank in both quantitative and qualitative
factors that affect the riskness of a design.

1132

Table 4.
16 16

It is no clear if it is necessary to construct membership function gathering expert knowledge. In this work
it is used simple triangular membership functions that
represent the whole rank used in a traditional FMEA.
Results of fuzzy techniques applied to this case
demonstrate consistency.
Risk priority number is used in prioritizing which
items require additional quality planning or action. By
fuzzy logic approach applied to matrix FMEA form
allows to detect easily the main contributor to the risk
in a better way.

Matrix FMEA.
80 0

0 72

84

16

0 147 0

24

80

30

ACKNOWLEDGEMENTS
To the Cuyo National University that support this
project. To Eng. Pablo Gusberti and to the staff
personnel of the Bioenergy Program.
REFERENCES

40 40

In this work fuzzy priority number are not arranged

in decreasing order.
Each row in Table 3 represents a component. Each
column represents a failure mode. This organization
shows quickly how much failure modes each component has, and which is the most important contributing
to the risk. At the same time it is so simple to extract
them.
From this case of study it is detected that it is
important take into account all measurements variables. All components are adequate for this plant but it
is necessary to evaluate the contribution of the human
error.
6

CONCLUSIONS

In this work it is gathered and combined a set of

techniques to improve the approach of FMEA studies.
A Continuous Distillation Plant was taken as case
of study.
Each technique alone or combined contribute to
explain a different point of view.
It is important to note the contribution of the human
error to this type of Plants and the spark production and
static electricity to the risk.

Abbasbandya S. & Asadyb B. 2002. Note on a newapproach

for defuzzyfication. Fuzzy Sets and Systems 128 (2002)
131132.
Anoop et al. 2006. Conversion of probabilistic information into fuzzy sets for engineering decision analysis.
Computers and Structures 84 (2006) 141155.
Devorak A. & Novk V. 2004. Formal theories and linguistic
descriptions. Fuzzy Sets and Systems 143 (2004) 169188.
Dubois D. & Prade H. 1996. What are fuzzy rules and how
to use them. Fuzzy Sets and Systems 84 (1996) 169185.
Dubois D. & Prade H. 1998. An introduction to fuzzy
systems. Clinica Chimica Acta 270 (1998) 329.
Klir G.J. & Yuan Bo 1995. Fuzzy Sets and Fuzzy
LogicTheory and Applications, Prentice Hall P T R.
Roventa E. & Spircu T. 2003. Averaging procedures in
defuzzification processes. Fuzzy sets and systems 136
(2003) 375385.
Sankar & Prabhu 2001. Application of fuzzy logic to matrix
FMECA. Review of Progress in Quantitative Nondestructive Evaluation. Thompson and D.E. Chimenti (ed.).
Vol. 20. American Institute of Physics.
Xu et al. 2002. Fuzzy assessment of FMEA for engine systems. Reliability Engineering and System Safety 75 (2002)
1729.
Yager R. 1997. Fuzzy logics and artificial intelligence. Fuzzy
sets and systems 90 (1997) 193198.
Yeh R.H. & Hsieh M. 2007. Fuzzy assessment of fmea for a
sewage plant. Journal of the Chinese Institute of Industrial
Engineers Vol. 24, No. 6, pp. 505512.

1133

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk analysis in extreme environmental conditions for Aconcagua

Mountain station
J.E. Nez Mc Leod & S.S. Rivera
CEDIAC Institute, Cuyo National University, Argentina

ABSTRACT: This work arises like a need to diminish the risk of data loss during the operation of a data
harvesting station. The station collects GPS and weather information on the top of the Aconcagua Mountain
located in Argentina. The access is only possible during summer season. Between April and November of each
year the access to the station is impossible due to the extreme climatic conditions. Under such conditions the
station should be capable of continuing its operation after to have a total loss of energy.
A failure mode and effect analysis was carried out by components. Critical failure modes were identified and
eliminated through the redesign.
The analysis carried out was oriented by the strong influence of the environment in the behavior of each
component. The correct characterization of the environment turned out to be fundamental. The design was
concluded with a series of modifications arisen of the analysis.

INTRODUCTION

Aconcagua Mountain, at 22,830 feet (6,959 meters) is

the highest point in the Western and Southern hemisphere, towering above the surrounding peaks in the
Argentine Andes. It has a very steep and massive face
on its south and a gentle slope on the north.
Aconcagua has become an increasingly popular
goal. This mountain can be a rigorous and rewarding
introduction to the demands of expedition climbing
to nearly 7,000 meters. On Aconcagua, the task of
load carrying to stock high camps, and the challenging
weather with its high winds, as well as the fortitude required to stay with one climbing objective for
three weeks, have shown many people what expedition
climbing is all about.
Until now geology it could discover that the Los
Andes grow, but not how much nor in what time.
For that reason, a system of measurement through
the GPS technology was installed to obtain data to
reveal the unknown of these mountains, highest of
America.
Although a priori it can seem superfluous to measure how many centimeters grow the Aconcagua and
the rest of hills, in decades, and even centuries,
the scientists assured that from the set of the investigations new and more precise knowledge for the
Geodynamics, Ionosphere, Geodesy and Glaciology

will arise, Cartography, and until for the territorial

ordering.
The system that settled in the summit of the
Aconcagua Mountain is of millimetric precision and
of small dimensions and it is connected with the
network of similar devices that exist to both sides of
the Mountain range.
A data harvesting station on the top of Aconcagua
Mountain was put it with the main to study its sliding.
Due to the need to diminish the risk of data loss
during the operation of the data harvesting station a
Failure Mode and Effect Analysis was done.
The station collects GPS (Global Position System) data and weather information on the top of the
Aconcagua Mountain. The access is only possible during summer season that is to say that would have access
once a year (between December and March). Between
April and November of each year the access to the
station is impossible due to the extreme climatic conditions like temperatures below zero, strong winds with
snow, snow-covered and persistent cloudiness. Under
such conditions the station should be capable of continuing its operation after to have a total loss of energy
(solar energy and batteries).
The design of the station should allow its disarmed
in smaller parts, such that could be transported for the
members of the mission up to the top of the mountain. Likewise the design for the armed one on the top

1135

should eliminate the possibilities of error. A failure

mode and effect analysis was carried out by components. Feedback from tests of reliability components
to low temperatures was obtained. It was analyzed and
subsequently experienced the thermal behavior of the
box that contained the electronic instrumental. Critical
failure modes were identified and eliminated through
the redesign.
The analysis carried out was oriented by the strong
influence of the environment in the behavior of each
component. The correct characterization of the environment turned out to be fundamental. Aconcagua
Mountain experts and personnel performing their tasks
in the Antarctica were consulted. The design was concluded with a series of modifications arisen of the
analysis.

Batteries

GPS Trimble

Photovoltaic
load
regulator

Figure 1.

Photovoltaic
load
regulator

GPS subsystem connection.

FAILURE MODE AND EFFECT ANALISYS

A Failure mode and effects analysis (FMEA) is a procedure for analysis of potential failure modes within a
system for the classification by severity or determination of the failures effect upon the system. It is widely
used in the manufacturing industries in various phases
of the product life cycle. Failure causes are any error or
defect in process or design, especially ones that affect
the customer, and can be potential or actual. Effects
analysis refers to studying the consequences of those
failures.
In FMEA, Failures are prioritized according to how
serious their consequences are, how frequently they
occur and how easily they can be detected. A FMEA
also documents current knowledge and actions about
the risks of failures, for use in continuous improvement. FMEA is used during the design stage with an
aim to avoid future failures. Later it is used for process control, before and during ongoing operation of
the process. Ideally, FMEA begins during the earliest
conceptual stage of design and continues throughout
the life of the product or service.
In this work two subsystems are studied: GPS and
Meteorological stations. The first gather GPS information measuring xyz coordinates of the point where
the box with both systems are anchored. The second
subsystem is a meteorological station for temperature,
atmospheric pressure, humidity and wind velocity and
direction measurement.

Freewave

SYSTEM DESCRIPTION

The station was built and was installed on the top of the
Aconcagua Mountain in the summer of 2006 (Blanco,
unpubl.).

3.1 GPS subsystem

This subsystem is shown in Figure 1.
Some connectors are located on the wall of the box
to allow the connection of the external devices.
The subsystem must like objective receive satellite
information by means of the device NetRS Trimble Receiver to calculate the position in which it
is on the globe and to transmit that information by
means of an Ethernet connection to the Freewave
transmitter that transmits it by radio frequency until
the position of Horcones (Nez Mc Leod & Rivera
2005).
The Trimble receiver counts with diverse power
sources. Its main power source comes from a set of
3 batteries in parallel of 12V, which are loaded as well,
through a load regulator, by the electrical energy produced by 2 photovoltaic panels (who do not appear
in the graph). These power sources are denominated
main in this document. Also it counts with a secondary source, which consists of 2 batteries in parallel
of 12V, which they are loaded, under the control of
load regulator (Solartec S.A. 2005a, b) by a solar
panel.
The Freewave transmitter (Ethernet bridge) is fed
from the bank of 2 batteries of 12V in parallel.
Both devices have antennas. First receive the satellite information and the second communicate bidirectionally by radio near to the mountain. Both antennas
have been equipped with an overvoltage protection
based on connected gas unloaders to earth.
The device is thought to be working in permanent
form the 365 days of the year. Nevertheless it is to hope
that long storm processes that prevent the generation
of electrical energy on the part of the solar panels take
to a state of exhaustion of the batteries after 3 days of
uninterrupted operation.

1136

The following components have not been

considered for the analysis: container, inner separating
plate of the container, anchor meteorological sensors,
subjection nails, inner cables and ground return.

Battery

ISS
Envoy

Photovoltaic
load
regulator

4.1 Justification of components exclusion

It is considered that:

Figure 2.

Harvesting subsystem of weather data.

The Trimble device can easily be acceded with a

protocol of transmission TCP/IP for its reconfiguration, unloads and erased of data by means of a friendly
interface HTML. The collected data are stored in a
Compact Flash Card of 1GB of capacity (FT 2005,
TNL 2003, TNL 2005).
3.2

Harvesting subsystem of weather data

This subsystem is shown in the following figure.

Figure 2 shows the components including within
the container specially designed to protect them of
the climatic inclemency of the top of the Aconcagua.
Externally it is had a support (approx 1 m of height)
in which a humidity anemometer, weathervane, thermometer and measurer are arranged.
The storage device of the information, the Envoy,
receives wireless information of the ISS. The Envoy
receives energy of a 6V Battery which is recharged,
under the control of a load regulator, by a solar panel.
This device cannot transmit the information that
stores. The frequency of taking data is set each 2 hours.
Thus the information is captured during 215 days and
later one unloading in situ. The access scheme is
quite rigid considering that the period to access to the
top of the Aconcagua extends from December to the
beginning of March.

METHODOLOGY

4.1.1 Cables
With the cold the cables improve their conductivity;
but they make worse its mechanical properties. Nevertheless the only cables under tensional requirements
are the external cables to the container. Cables adequates for extreme temperatures (below zero Celsius)
were acquired.
4.1.2 Anemometer and weathervane
There is not information for the analysis of the physical
fault of the anemometer and the vane due to the prevailing conditions on the top of the Aconcagua Mountain.
For this reason is considered that the selection fit the
requirements.
4.1.3 Interference
Radio frequency interference like root cause failure
has not been considered.
4.1.4 Other faults
Failures originated by inadequate move of the equipment have not been considered.
Human failures due to the installation of the station
have not considered either.
At the top of the Aconcagua Mountain the reflections and the cognitive activities are degraded by the
lack of oxygen.
This process can lead to the attainment of errors
which have been tried to diminish with a design that
reduces the decision making in the armed one, if it
does not eliminate them.

The standard methodology has been followed for

the accomplishment of the Failure Mode and Effects
Analysis.
The detail level has been by components, without evaluating the engineering of components such as
NetRS Trimble, Freewave, Load regulators, etc. These
considerations have the purpose of the study the availability of the station with the components that had been
acquired specially for the project.
The analysis considers the station in normal operation and/or with extreme temperatures. In this last case
always the considered temperature is clarified.

SUCCESSFUL STATES

Remembering that the most excellent point is the safeguard of the satellite data process, the following states
of success have been considered:
The device of harvesting and storage (NetRS Trimble) and the device of transmission (Freewave) in
operation.
The device of harvesting and storage (NetRS Trimble) and the device of transmission (Freewave) in
delay. This case would occur before the eventuality
of climatic conditions such that by more than 3 days
prevent the generation of electrical energy from the

1137

Table 1.

FMEA table example.

Components
#

ID

Description

Failure mode

Battery

Battery

Failure to function (normal

operation)
Failure to function (low
temperature)

solar panels, which would lead to the total unloading

of the batteries.
The device of harvesting and storage (NetRS Trimble) in operation.
The device of harvesting and storage (NetRS Trimble) in delay.

6
6.1

REMARKS
Batteries

It is considered that as much the load as the unloading of the batteries strongly is influenced by the
temperature variation (Enersys Inc. 2002, Enersystem 2005). A battery of the type used in the
Aconcagua station lets be able to be loaded when
the temperature descends to 15 C, point in which
with an inferior load to 50% it begins to produce the
freezing of the same one. Finally from 20 C the
possibility of unloading of the battery stops.
Before maintained conditions of low extreme temperatures, only the electrical generation of the solar
paddles will be able to put into operation the diverse
devices; but it will not obtain the load of the batteries
(Batteryworld 2005).
Conservatively it was not considered the heating of
the atmosphere in the container due to the lack of
suitable models to simulate its effects (Cirprotec
2005).

Effect on
the system

Cause failure

Detection

None

Open circuit

None

None

Frost

None

Connectors

The connectors standard begin to present problems exposed to inferior temperatures to 10 C

to 15 C, degrading their performance until being
necessary to replace them by special types to
25 C.
The system does not count with special connectors
and therefore a special emphasis has become in its
analysis. In these cases they considered the open
circuit failure mode and defective contact. Both of
them can be reached from several points, including
temperature factor.

Redundant
device
None. Solar
panels are
available.

FMEA TABLE

For this work FMEA tables were constructed with the

following headings:

Component ID
Component description or function
Failure mode
Effect on the system
Cause of failure
Detection way
Mitigation/restoration

Each component was analyzed resulting in 158

items and gathering in tables like it is shown in
Table 1.
Each row displays a component with its failure
mode. Describe the effect on the system and its
possible cause of failure. Finally it is evaluated if
there is a detection way and mitigation or restoration
possibility.
When a component has more than one failure mode
it is necessary to use so many rows as are necessary.
For this work the results obtained from the tables
are qualitative.
Suggestions and recommendations were obtained
and some of them contributed to improve the original
design.

8
6.2

Mitigation/
restoration

RESULTS

The results of the analysis can be grouped in suggestions of design changes and recommendations of
another nature.
8.1 Suggestions
Provision of primary energy to the Trimble device
(GPS) from the set of 3 redundant batteries. In the
original design energy was provided from this set to
the secondary connector exclusively.
Provision of secondary energy to the Trimble device
(GPS) from the set of 2 redundant batteries.

1138

 Simplification of connections in the terminals using

plates to make bridges.
8.2

Recommendations

To install diodes in sections that joint terminals and

Load Regulators, since these last ones do not make
the cut (open circuit) when the tension of the solar
panels is minor than the batteries.
To test photovoltaic load regulators to confirm the
correction of low and high set points.
To acquire of Ni-Cd batteries due to they are adapted
to work at extreme temperatures.
To begin connections with whole terminals with
weld points applied to each screw proving at the
end definitive conductivity.
To eliminate all sign of dirt and dust the cleaning
with dry cloth of the internal conductor, the contacts
of the gas capsule and the screw of fixation of the
capsule is recommended.
To place a plastic cover with an insulator at the
bottom part of the inner dividing plate of the box.
To test ground return on the top of the mountain to
ensure the correct working.
To use special connectors in external connections at
low temperature.
9

Future work is necessary about the construction of

a thermal model of the container and its main devices
to improve the study.

CONCLUSIONS

Broadcast of data was initiated successfully; nevertheless with the first storm the station to stopped
functioning. In the summer from 2007 the cause of
the failure was determined due to a change carried out
during the installation on the top of the mountain. The
solar panels were covered by a metallic fabric. The
fabric allowed to the snow to be accumulated and to
resist defrosting.
Results tested that the methodology of analysis was
correct. None of the ways of residual failure related to
the original design affected the station.
Meteorological station must be reexamined to
improve availability and to increase storage capability
and remote manage.

ACKNOWLEDGEMENTS
To the Engs. L. Euillades & M. Blanco (personnel of
the Satellital Image Division of the CEDIAC Institute)
for supply us the relevant information for accomplish
this study.
REFERENCES
Blanco, M. 2005. Circuito de Conexiones del Proyecto
Aconcagua, Documento Interno. Instituto CEDIAC, Facultad de Ingeniera, Universidad Nacional de Cuyo.
Mendoza, Argentina.
Batteryworld 2005. Understanding batteries. Battery World.
www.batteryworld.com.au.
CIRPROTEC 2005. Protecciones de seales de radiofrecuencia, www.cirprotec.com.
Enersys Inc 2002. Genesis NP 12-12, Sealed Rechargable
Lead-Acid Battery. Specification sheet. USA.
Enersystem, 2005. Catalogo de bateras YUASA. Series
NP/NPH/NPX. Selladas y Recargables, Bateras de
Plomo-cido.
FT 2005. Freewave, FGR Ethernet Series. Specification
sheet. Freewave Technologies, Bouder, Colorado, USA.
Nez Mc Leod, J. E. & Rivera, S. Estacin de Recoleccin de
Informacin GPS y Meteorolgica Aconcagua. Mmico de
Conexiones. CEDIAC-GD-IN002-06. Instituto CEDIAC,
Facultad de Ingeniera, Universidad Nacional de Cuyo.
Solartec S.A. 2005a. Regulador de Carga Solartec R5 y R8.
Manual del Usuario. Buenos Aires, Argentina.
Solartec S.A. 2005b. Generadores Elctricos KS5-KS7KS10-KS12-KS16-KS20. Manual del Usuario. Buenos
Aires, Argentina.
TNL 2003. NetRS GPS Receiver User Guide. Trimble.
Trimble Navigation Limited. USA.
TNL 2005. Estacin de referencia GPS Trimble NetRS.
Trimble Navigation Limited. Hoja de datos. USA.

1139

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Geographic information system for evaluation of technical condition

and residual life of pipelines
P. Yukhymets
Paton Welding Institute, Kiev, Ukraine

R. Spitsa
Institute of Geography, Kiev, Ukraine

S. Kobelsky
Strength Problem Institute, Kiev, Ukraine

ABSTRACT: GIS principal scheme includes pipeline sections technical state database and analytic block for
making corresponding decision. GIS distinctive feature is opportunity to estimate pipeline accident risk taking
into account: pipelines remaining life prediction using statistical treatment of pipeline thickness measurements
in pits or/and during in-line inspection; stress-strain state determination of pipeline typical elements with corrosive defects using FEA; estimation of hazardous geodynamical processes capable to upset balance between
nature and man-caused system in the pipeline territory. At present demo version of the system realized at one
pipeline in Ukrainian oil producing company.

INTRODUCTION

GIS technologies are important unit of modern

investigations at the design, construction and operation
stagesofindustrialobjectshavingdifferentassignment.
GIS opportunities in factorial and complex modeling of
space objects and phenomenon as well as investigations
intheirstaticanddynamiccharacteristicsallowconsider
them universal mean for data analysis and synthesis.
In contemporary understanding GIS is informational system providing collection, storing, processing,
access, imaging and spreading of space coordinated
data. It contains data about spatial objects in the
form of their digital presentation and includes set of
functions adjusted to being solved tasks.
Significant data volume used in pipelines systems
management has spatial component that characterize
objects in connection with their geographical location
(Koshkarev 1990). This certainly promotes pipeline
companies to use GIS technologies for automation and
integration of their activity.
2

(Maiorova 2005). Under certain conditions operational life growth in combination with impact of
thermal and mechanical loads, environment influence
could involve corrosion and flow accelerated processes and base metal saturation by new chemical
elements. As a result this may initiate development
of defects that at the beginning stage of operation
had allowable size. Beside that failure probability of
some pipeline sections increases in consequence of
bulging and corrugate forming, stress rising due to
landslide and mechanical damage. This way operation extend decision should be made differentially
and based on operation conditions singularity of the
pipeline section and its technical state. These circumstances and change from traditional pipeline schedule
technical service to service according to technical state
predetermined GIS principal scheme that includes
pipeline sections technical state database and analytic
block for making corresponding decision.
As a base program product for GIS development
was used MapInfo. Its main advantages are low cost,
hardware recourses tolerance and ease using.

GIS PRINCIPAL SCHEME

3

Efficient technical state monitoring is major condition of pipeline system safe operation. Continuing pipelines aging influences their actual reliability
level and is accompanied by accident rate rising

DATABASE DESCRIPTION

Pipeline company database stores all available at

the moment information about its pipelines including crossings, used pipes, soil types, electrochemical

1141

protection potential, rust-preventive cover state,

revealed defects, fittings, operational characteristics,
pipeline corridor topographic maps etc. The sources
of data are documentation and technical diagnosis
results.
As far as information content is in accordance with
actual standards the system could be integrated in
existing GIS.
Pipeline technical state diagnosis is fulfilled in next
order (Instruction 1998):
technical documentation analysis;
pipeline passage refinement, passage area properties study, rust preventive cover state investigations
by above-ground method (cover state is estimated
in accordance to generated high frequency current
dumping);
composition of technical state diagnosis program
considering:
available corrosion and flow formation zones;
rust preventive cover check results;
non-destructive testing;
pipeline strength and residual life calculation.
Database information is structured (STP 2003):

ANALYTICAL BLOCK

4.1 Risk assessment model

Comparative risk assessment model is a core of analytical block. According to it (GROWTH Project
2001):
Hazard risk = Consequence Probability

For data collection are used tables (see example in

Tab. 1).
Database interface provides possibility for user to
view and edit different data and form required query.
Taking into account pipeline geographical aspects
system includes geodesic objects positioning. Combined application of inspection results and technical
documentation data is based on referee GPS coordinates. It is provide convergence of investigation

Hazard consequences categories.

Pipeline section category

Predicted category
consequence of hazard

III
II
I
B

Low
Moderate
Severe
Very severe

Table 3.

Pipeline failure evaluation.

Probability

Description

Low
Moderate
High

Essentially impossible
Potentially possible
Might happened during operation
time
It is likely that the event has
occurred at the site if the facility
is more than a few years old

Very high

Object Stalk database input.

Characteristic
Point ID
Distance along pipeline, m
Height, m
Technological N o
Diameter wall thickness
Assembling date
Last inspection date
State
Drawing
Photo

(1)

Predicted consequence of hazard is set in direct

relation to its category shown in Table 2 according
to (Structural Norma and Rules 1985). Pipeline failure
evaluation due to some hazard effect is done according
to Table 3 (Sutton 2002).
Risk ranking is done using matrix 4 4 (Tab. 4).
Pipeline section technical state assessment is carried out using next main criteria on the base of data
accumulated in the course of periodical inspections:
Table 2.

pipeline general technical and operation information;

pipeline above-ground informational and measuring objects;
pipeline technological elements;
pipeline crossings;
dig pit or/and in-line inspection results.

Table 1.

methods which use diverse physical methods and

coordinate binding modes.

Value/Reference
Table 4.

Point risk ranking matrix.

Consequences category

Probability

Low

Moderate

Severe

Very severe

Low
Moderate
High
Very high

1
1
2
3

1
2
3
3

2
3
3
4

3
3
4
4

1142

Table 5.
Technical
documentation

Pipeline failure probability.

Residual life
Hazard
consequence
categories

Pipeline section
failure
probability point
estimation by
major criteria
( 1, 2 A10)

Separated
criterion risk
estimation

Total failure
point
estimation

years

Probability description

>30
20 30
10 20
<10

Essentially impossible
Potentially possible
Might happened during operation time
It is likely that the event has occurred at
the site if the facility is more than a few
years old

Inspection
results

Figure 1.

or/and during in-line inspections (Instruction 1998,

DSTU 2001).
Purpose of statistical treatment is defining parameters of wall thickness distribution function and estimation with required reliability maximal depth of
corrosion damage.
Pipeline linear part residual life prediction Tp is
based on wall thickness decrease dynamic taking
into consideration its thickness minimal allowable
value and statistical treatment results of thickness
measurement:

Pipeline section failure risk estimation.

pipeline tightness (A1 );

rust preventive cover state (A2 );
corrosion state of pipeline metal (A3 );
welds quality (A4 );
roaming current danger (A5 );
electrochemical protection state (A6 );
FAC (A7 );
soil corrosion aggressiveness (A8 );
geodynamical stability (A9 );
pipeline depth location (A10 ).

Tp =

Risk estimation is fulfilled for every of abovementioned criterion separately. After that pipeline section
failure total risk ST is defined by addition 10 separated
estimations obtained for every criterion:
ST = A1 + A2 + + A10

(2)

where A1 , A2 . . . A10 = pipeline section failure point

estimation by corresponding criterion.
Both estimationstotal and separated are taken
into consideration during analysis (Fig. 1). First of
all separated high point risk estimations should be
loweredthese measures simultaneously reduce total
estimation for given pipeline section. In case of point
risk estimation equality for several sections in the
first turn preventive measures should be taken toward
sections having grater total estimation. Anyway special attention should be paid for sections having
absolutely high total estimation ST .
Results of analytical block functioning provide possibility to develop pipeline maintenance plan, create
corrosion maps of its sections, determine allowable
pressure considering revealed defects.
4.2

Residual life assessment

Residual life is one of the main failure risk estimation

criterion. Its calculation is based on statistical treatment of pipeline periodical investigations in dig pits

Sav min Sclc

,
V
p

(3)

where Sav min = minimal probable average value of

pipeline wall thickness; V = corrosion rate; Sclc =
pipeline calculated wall thickness.
Pipeline failure probability is evaluated from the
base of calculated residual life (Table 5).
Such approach could be considered as additional
or alternative to in-line investigations because deliberate geometrical information (wall thickness data,
pipes out-of-roundness, presence of corrugations
etc.) should be supplemented with data about metal
actual state, local stresses, surface defects specified
information.
If surface volume damage of corrosion or flow
accelerated origin is revealed (for example in a pipeline
linear part, elbow or T-joint) than calculated estimation of its danger from position of static and cyclic
strength should be done.
At the base of static calculation lays hypothesis
according to which surface volume damage influence
is similar to influence of single through hole. Defect
reduction to single through hole is done by mean of its
cross area along pipeline axis because pipeline bearing
capacity is defined by hoop stresses (Fig. 2).
Along with operational loading and material cyclic
properties pipeline cyclic strength is based first of all
on its stress-strain state in stress concentration zones.
Stress-strain state of damaged zone under elastic
load is defined by FEA. Method is realized in the

1143

Figure 2. Defect limit dimensions curve: 1theoretical

curve; 2curve used in calculations; tn nominal thickness;
tmin effective thickness; ti current thickness; Lp defect
length along effective length; Li current defect length.

problem oriented program SPACE-3D environment

(Software 2004). Program performs multiple-choice
numerical calculations of pipelines elements. Parameterized library of pipeline typical elements allows to
build required model of the damaged element using its
base model. This way the geometrical model building
task is significantly simplified (Figs. 3, 4).
The halfellipsoid is used as geometrical model of
concerned defects. Its axes are correspondingly equal
to maximal sizes of defect in axial, hoop and radial
directions of pipe (Fig. 5).
k
Maximal local elastic-plastic strain e aj
determination is defined using FEA calculation results and
interpolating correlation for concentration factors:
k
= Kej
e aj

eaj
ey

(4)

where eaj = nominal strain amplitude in j load

cycle; ey = yield strain; Kej = elastic-plastic strain
concentration factor in j load cycle.
Static and cyclic calculations results are compared
with uniform corrosion data (3) and could influence predicted residual life assessment which do not
consider defects (Fig. 6).
Employed in calculations methods were approved
by full-scale specimens testing (Fig. 7).
4.3

Geodynamical stability

Numerous researches carried out in last years demonstrated negative influence of soil subsidence, landslides, karst and erosion processes, impoundment,

Figure 3.

T-joint FEA mesh.

extreme wind loadings (in case of overland parts) on

integrity and functioning of a pipeline (Govdiak 2002,
Zaharov 1988).
Most of the pipeline failures of East European
Platform take place at the contacts of different geomorphologic levels (terraces, alignment surfaces) characterized by environment corrosion activity rising. Also
technical state of linear structures could be negatively
influenced by, so called, power dangerous zones that
form on the tectonic structure borders having different
neo-tectonic and modern tectonic activity (Hudiakov
2000). It is also true for morphostructural knots that
are formed in intersections of tectonically active fractures having differing directions. In these locations are
observed rising of seismicity, rate and amplitude of
the earths crust vertical slow movements that finally
could influence pipeline strain-stress state (Kendzera
2003, Palienko 1992).
Taking into consideration aforementioned lithosphere main parameters are included in the analytical
block as separated submodule. This allows to evaluate morphology structural and neo-tectonic conditions
in pipeline passage zone. Averaged activities of areal
morphology are the base of morphology structural
zoning As a result the territory is divided into active,

1144

Uniform
corrosion

Inspection
results

Local
corrosion

Figure 6.

Figure 4.

Pipeline
residual life Tp
static
calculation

Damaged
section
residual life
Tss static
calculation

S-S state under

elastic load

Tp
min Tss
Tsc

Damaged
section
residual life
Tsc cyclic
calculation

General scheme of residual life (A3 ) calculation.

Figure 7. Full-scale specimen D325 after inner pressure

failure test.

Stress intensity in defect zone.

b/2
c

d
Figure 5. Volumetric surface defect modeled with halfellipsoid: clength, bwidth, ddepth; hpipe wall
thickness.

Figure 8. Pipeline failure risk estimation by rust preventive

cover state criterion.

1145

moderately active, low active and not active structures and territory geodynamic stability index is one
of the criterions of pipeline technical state integrated
estimate.
5

CONCLUSIONS

GIS distinctive feature is opportunity to estimate

pipeline accident risk taking into account:
pipelines remaining life prediction using statistical
treatment of pipeline thickness measurements in pits
or/and during in-line inspections;
stress-strain state determination of pipelines typical
elements with corrosive defects using FEA;
estimation of hazardous geodynamical processes
capable to upset balance between nature and mancaused system in the pipeline territory.
On the base of permanent collection, accumulation,
processing and examination of information related to
pipeline state proposed system determines possibility
to operate individual sections and the pipeline on the
whole, assign corresponding preventive and corrective
steps.
At present demo version of the system realized
at one pipeline in Ukrainian oil producing company
(Fig. 8).
Further project development directions in our opinion should be:
risk assessment module improvement;
economical aspects of inspection and accidents;
advanced study of all factors used in risk assessment.

Govdiak, R. 2002. Oil pipeline design improvement subject

to environment interaction. Ivano-Frankivsk: Candidate
of science thesis.
GROWTH Project G1RD-CT-2001-03008 RIMAP, 1-11F-2004-01-1.
Hudiakov, G. 2000. Energy dangerous zones and trans
regional pipelines/problems of ecological geomorphology
Materials of Intergovernmental XXV geomorphological
conference of Russian Academy of Science. Belgorod:
Publishing hose of Belgorod State University.
Instruction on inspection and residual life assessment of
underground pipeline I1-98. Kiev: Paton Welding Institute.
Kendzera, A. 2003. Modern movements of Earths crust in
Ukraine. Geoinformatic-No. 4: 6673.
Koshkarev, A. 1990. Cartography and geoinformatic: ways of
interaction. Reports of Acad. of Science of USSR, Geology
series, No. 1: 3238.
Maiorova, O. 2005. Failures and crashes on trunk gas
pipelines: causes and consequences. Actual problems of
contemporary science, No. 2:118.
Palienko, V. 1992. Latest geodynamic and its reflection in
Ukraine relief. Kiev: Naukova Dumka.
Software 2004. 3-D FEA modeling of thermal and thermotension state of machinery construction elements (SPACE)
Certification system UkrSEPRO. Certificate of conformity
No. UA1.017.0054634-04.
STP 3209.30019801.084-2003 Trunk pipelines. Requirements to passport system data. Kiev: PK Ukrtrsnsgaz.
Structural Norms and Rules 2.05.06-85. Trunk pipelines.
Moscow: Gosstroi USSR.
Sutton, I.S. 2002. Process Hazards Analysis. Houston, Texas:
Sutton Books.
Zaharov, J. 1988. Engineering and geologic monitoring of
nature-anthropogenic systems Moskow: Ph.D. thesis.

REFERENCES
DSTU 40462001. Technological equipment of oil-refining,
petrochemical and chemical manufacture. Kiev: Derjstandart of Ukraine.

1146

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Inherent safety indices for the design of layout plans

A. Tugnoli & V. Cozzani
Dipartimento di Ingegneria Chimica, Mineraria e delle tecnologie Ambientali,
Alma Mater Studiorum, Universit di Bologna, Bologna, Italy

F.I. Khan
Faculty of Engineering & Applied Science, Memorial University, St. Johns, NL, Canada

P.R. Amyotte
Department of Process Engineering & Applied Science, Dalhousie University, Halifax, NS, Canada

ABSTRACT: The design of the layout of chemical processes offers significant opportunities for improving
safety through the implementation of inherent safety principles. However, the application of inherent safety as
a guideline for layout design is seldom explored and practical support tools for the design activity are totally
missing. In the current contribution, a metric is introduced to assess the safety level in the analysis of layout
options. Several parameters are mapped by the indices in the metric: the hazard of the plant units, the domino
hazard of the layout plan, the application of inherent safety principles, the use of traditional risk control devices
and the safety economics. The framework is simple, auditable and easily applicable. Specific support rules are
defined in order to reduce subjectivity in the assignation of scores. The application to a case study points out as
the assessment yields effective indications by a swift procedure.

INTRODUCTION

The design of the layout of chemical processes deeply

affects the final safety performance of the plant. As
a matter of fact, design choices influence critical
elements such as the possibility of chain accidental
events, the maximum magnitude of the accident consequences, the possibility of exposure and fatalities in
populated targets (e.g. buildings), the accessibility of
the different areas in a plant (Khan & Abbasi 1999,
Cozzani et al. 2007).
Layout design is usually performed in successive
steps having an increasing detail. The early steps,
when relative positions of the units are defined,
present larger chances for effective introduction of
safety considerations. Safety issues are usually introduced in the early layout design by generic rulesof-thumb (Mecklenburg 1973, Lees 1996), usually
expressed through conventional safety distances that
do not account for the specific features of each
case. However, some attempts to include safety
constraints in the economic optimization of layout plan have been made (Penteado & Ciric 1996,
Patsiatzis et al. 2004). Nevertheless, the application of inherent safety in layout design is seldom
explored (Cozzani et al. 2007). Moreover reliable

assessment tools to support the design activities are

totally missing.
In the current contribution, an inherent safety metric is introduced to assess the safety level of layout
options. The assessment tool is based on a hierarchy
of indices and considers both strictly inherent as well
as passive safety measures in achieving layout safety.
Several parameters are scored by the indices: the hazard of the plant units, the application of inherent safety
principles, the requirement of traditional risk control
measures and the safety economics. The framework is
simple, auditable and easily applicable. The level of
information detail required by the assessment procedure is compatible with the early design phases. Moreover the metric presents a common framework with
inherent safety tools used in other phases of design
lifecycle (e.g. process design) (Khan & Amyotte 2004,
2005). In order to reduce the elements of subjectivity
in score assignation, reference sub-indices are defined
to support the assessment. Among them is the definition of a domino hazard index that scores the accident
propagation potential on the basis of the actual consequences of possible escalation scenarios. A case study
is presented in order to demonstrate the application of
the method. The results point out that the assessment
yields effective indications by a swift procedure.

1147

THE APPLICATION OF INHERENT SAFETY

IN LAYOUT DESIGN

Inherent safety aims for the elimination, or the reasonably practicable reduction, of the hazards in a
system (Kletz 1978). Moreover, inherently safer systems can reduce the high costs usually associated with
the full plant lifecyclefrom hazard management to
regulatory liabilities and safety system maintenance
(Kletz 1998, Khan & Amyotte 2005, Gupta et al.
2003). The basic principles of inherent safety can be
described trough an set of key guidewords (Hendershot
1999): minimization, substitution, attenuation, simplification, and limitation of effects. The guidewords
attenuation and limitation of effects can be generally
combined as moderation. Layout design is, however, an area where the distinction of the latter two
guidewords is deemed beneficial.
The inherent safety guidewords can be applied both
at different stages of the system design cycle and at
different levels of the safety strategies for control.
In the present work, the design stage of concern is
layout design, with particular reference to the early
phases (design of process items and utilities location,
building locations, on-site roads and access routes,
etc.). The safety strategies for control can be conventionally and hierarchically classified as inherent,
passive (engineered), active (engineered), and procedural. Application of the inherent safety guidewords to
the inherent safety strategies themselves is obviously
the most effective and straightforward approach, and
has received the majority of attention in prior development of assessment tools (Khan & Amyotte 2003).
However, the guidewords can also be applied at the
other levels of the hierarchy, for example leading to
add-on measures that are more reliable, effective and
thusin a broad senseinherently safer. In the current work, both strictly inherent measures as well as
passive measures have been investigated for their ability to improve the safety performance of the layout
plot. The perspective of layout design considered here,
therefore, is one in which the entire set of items placed
in the facility (no matter if they are pieces of equipment or blast walls) contributes in defining the global
hazard of the plant as a system (e.g. the potential for
maximum domino effect escalation). This shift in perspective justifies the choice to consider both safety
strategy levels (inherent and passive) in the current
analysis. Active and procedural safety strategies are
not considered here because, by their definition, they
do not generally belong to the first stages of layout
design.
The design of plant layout requires consideration
at once of several different issues: process requirements, cost, safety, services and utilities availability,
plant construction, regulations, etc. Furthermore, constraints related to choices in previous design steps

(e.g. chemical route choice, process design, equipment

selection, etc.) exist in layout design. This situation
limits the applicability of inherent safety principles in
layout design activities:
The minimization guideword is generally not applicable because equipment characteristics and material inventories have already been selected in previous design phases (conceptual and basic design).
The substitution guideword in layout design, both
for equipment and materials, is affected by limitations similar to minimization. Thus, substitution
applicability is generally limited.
The attenuation guideword, in its usual reference
to changes in unit operating conditions, has a limited applicability as these conditions will have been
fixed in previous design steps. However, this guideword may be applied to changes in the arrangement
of units. Changing unit arrangement and/or increasing unit segregation may reduce, if not eliminate,
the potential of domino effects and thus the hazard within the system. As a matter of fact, accident
escalation has been identified as the key hazard
source related to process layout design (Cozzani
et al. 2007). The effectiveness of layout in reducing
this hazard is thus the application of the guideword
attenuation for the plant considered as a system
(i.e. from the perspective of layout analysis). In
essence, the current work is proposing that just
as the materials being processed are the building blocks of process design, so then the process
units are the building blocks of layout design. So
processing hazardous materials under less hazardous conditions (Kletz 1998) becomes using
hazardous units in the least hazardous form.
The simplification guideword is readily applicable
to layout design. The choice of unit spatial organization has great potential to affect the simplicity of a
plant. Complexity can easily arise as the disposition
of units diverts from the logical process flow order,
or as further items (e.g. walls, equipment of other
production lines, and buildings) are added to the
plan. Therefore, it is quite likely that layout design
choices, even if oriented to satisfy the other inherent
safety guidewords, eventually result in a negative
feedback with respect to simplification.
Limitation of effects is a guideword that deals with
the reduction of the extent of negative consequences
arising from accidental events. In early layout
design, the potential adverse effects can be limited
by both inherent and passive strategies. Thus, the
limitation of effects guideword has been considered
in the safety analysis of both inherent and passive
measures. Three main applications of limitation
of . . . in layout design were identified:

1148

i. limitation of the effects of domino escalation:

reduction of the effects and consequences of

domino escalation events, considering the integrated action of inherent and passive strategies.
Note that this is a different aspect than the one
considered for the applicability of the attenuation guideword. With attenuation, the focus
was on reduction of the embedded hazard, such
reduction being attained only by inherent measures. With limitation of effects, the focus is on
the effects themselves that can be controlled by
both inherent and passive strategies.
ii. limitation of the damage potential to target
buildings: appropriate location of buildings
(workshops, administrative buildings, trailers,
etc.) and control or emergency structures (control room, medical centre, etc.) in the layout plan
so as to limit harm to people and impairment of
accident response.
iii. limitation of the affected area: limitation (generally by passive measures) of the spatial area
affected by the consequences of an accidental
event, regardless of the effects on other units,
buildings, etc.
The conclusion from the above analysis is that out of
the five guidewords, three (attenuation, simplification
and limitation of . . . ) are of particular interest for the
safety assessment of layout plans.

PROPOSED SAFETY ASSESSMENT

METHOD

The newly proposed layout assessment tool is meant

to extend the inherent safety perspective from the process design, where it is more frequently invoked, to
the layout design. A practical assessment tool should
share the same logical structure in all the phases of the
design, in order to give continuity in the inherent safety
analysis. Thus the framework of one of the more comprehensive assessment schemes proposed for process
design, the Integrated Inherent Safety Index (I2SI)
(Khan & Amyotte 2004, 2005), is adopted for layout assessment. Clearly enough, this new application
of the method required a deep revision of the indices
and sub-indices used in the analysis, in order to match
the unique issues of layout design: i) the inherent
safety indices were updated in order to address the
aforementioned role of inherent safety guidewords in
layout analysis; ii) reference indices were provided to
reduce the subjectivity in the evaluation of the extent
of the applicability of inherent safety guidewords;
iii) specific criteria for safety performance scoring
with respect to domino escalation were proposed;
iv) the safety cost indexing procedure was expanded to
account for accidental losses from chain effects. However, all the positive features of the I2SI approach for
assessment of options in early stages of the design are

conserved in the novel application (Khan & Amyotte

2004, 2005).
In the following sections, the I2SI for layout analysis is presented, devoting particular attention to the
new elements introduced. For a detailed description of
the I2SI framework, the original references (Khan &
Amyotte 2004, 2005) may be consulted.
3.1

The Integrated Inherent Safety Index (I2SI)

in Layout Analysis

The information necessary to perform the assessment

of a layout option are a preliminary definition of the
plant plot and of the parameters necessary to describe
the hazard of the pieces of equipment.
The first step of I2SI assessment for safety in layout
is the identification of the units in a given option.
For each unit, the Integrated Inherent Safety Index
(I2SI) is comprised of two main sub-indices (Equation 1): a Hazard Index (HI) and an Inherent Safety
Potential Index (ISPI).
I 2SI =

ISPI
HI

(1)

The Hazard Index is a measure of the damage

potential of a single unit after taking into account the
process and hazard control measures. The Inherent
Safety Potential Index, on the other hand, accounts for
the applicability of the inherent safety principles (or
guidewords) to the unit. The HI is calculated for the
units of an arbitrary reference layout optioncalled
the base caseand the values remain the same for
the corresponding units in all other possible options.
As evident, an I2SI value greater than unity denotes
a positive response to the inherent safety guideword
application (i.e. an inherently safer option). The higher
the value of I2SI, the more pronounced the inherent
safety impact.
To evaluate alternative layout options for the same
plant, the I2SI values for all the N single units
considered are combined according to Equation 2:

I 2SIsystem =

N


1/2
I 2SIi

(2)

i=1

The Inherent Safety Potential Index (ISPI) is comprised, of two sub-indices: an Inherent Safety Index
(ISI) and a Hazard Control Index (HCI). The ISPI for
single units is computed as shown in Equation 3:
ISPI =

1149

ISI
HCI

(3)

HI =

DI
PHCI

(4)

and limitation of effects these values are estimated by

conversion monographs (Figs. 1 and 2) that relate the
score of the extent of applicability of the guideword
in the assessment option to an ISI score. In the current approach, specific reference indices are proposed
to facilitate the quantification of the extent of applicability. For the guideword simplification, where the
authors experienced objective difficulty in quantifying
this subjective parameter, an arbitrary reference table
is proposed for the direct assessment of the index value
by linguistic guidelines (Table 1). The specific values
of ISI for the single guidewords are combined together
to yield the final ISI for the assessed unit, according
to Equation 5:
ISI = [Max(2 , ISIa 2 + ISIsi ISIsi  + ISIl 2 )]1/2 (5)
Extent of guideword applicability

10

10 0

Inherent Safety Index

The numerator, ISI, is calculated by using scores

based on the applicability of the inherent safety
guidewords, as reported in the next section.
The original version of ISPI (Khan & Amyotte
2004) used PHCI after the implementation of safety
measures as the denominator in Equation 3. The Process and Hazard Control Index (PHCI) is calculated for
various add-on process and hazard control measures
that are required or are present in the system. This
index is quantified on a scale mutually agreed upon by
process safety experts. The index ranges from 1 to 10
for any control arrangement (e.g. temperature control,
level control, blast wall, sprinkler system, etc.) and
is quantified based on the necessity of this control
arrangement in maintaining safe operation for the unit
(Khan & Amyotte 2005). However, for layout considerations, the denominator in Equation 3 is redefined
as HCI (Hazard Control Index) after the implementation of safety measures. In the assessment of HCI, the
requirement to install further add-on hazard control
measures after the previous analysis and implementation of safety measures in the layout option is assessed.
Process controls are not considered here, since they
are not effective in layout safety. The scores of HCI are
evaluated by the same rules as PHCI (Khan & Amyotte
2005).
The Hazard Index (HI) for layout assessment is evaluated for each unit following the same procedure as for
process assessment (Khan & Amyotte 2004). The HI
is comprised of two sub-indices: a Damage Index (DI)
and a Process and Hazard Control Index (PHCI). The
numerical value of HI for the unit being considered is
calculated by dividing the DI by the PHCI, as shown
in Equation 4:

80

60

40

20

0
1

Figure 1.
word.

0.8

0.6
0.4
Reference index

0.2

Inherent Safety Index for the attenuation guideExtent of guideword applicability

The Damage Index (DI) is a function of four hazard

parameters, namely: fire and explosion, acute toxicity,
chronic toxicity, and environmental damage. These are
estimated as a function of the expected damage radius
of each scenario and have values ranging up to 100.
Damage radii may be calculated using simple, validated approaches (see e.g. Khan et al. 2001, Tugnoli
et al. 2007).

10

10 0
(a)

Inherent Safety Index

3.2

The Inherent Safety Index (ISI)

80
(b)
60
(c)
40

20

The ISI calculation follows the same procedure as a

HAZOP study in which guidewords (in the present
case, inherent safety guidewords) are applied to the
assessed system. Based on the extent of applicability
and the ability to reduce the hazard, an index value
is computed for each guideword identified as the relevant guidewords for layout design. For each guideword
a specific value of ISI is estimated. For attenuation

0
1

0.8

0.6
0.4
Reference index

0.2

Figure 2. Inherent Safety Index for the limitation of effects

guideword: Limitation of effects of domino escalation (a),
limitation of the potential to target buildings (b), limitation
of the affected area (c).

1150

Table 1. Extended guidelines to decide the ISI value for

the guideword simplification in layout assessment. For the
positive values of ISI, refer to Khan & Amyotte (2004).
Description

ISI

Non-applicable
0
Minor complication & no substantial hazard increase 10
Minor complication & hazard increased moderately
20
Complicated moderately & moderate hazard increase 30
Complicated moderately & hazard increased
40
Complicated & hazard increased
50
Complicated & new hazards introduced
60
Complicated to large extent & moderate hazard incr.
70
Complicated to large extent & hazard increase
80
Complicated to large extent & significant hazard incr. 90
Complicated to large extent & hazard highly increase 100

where the subscripts refer to the considered guidewords (a for attenuation, si for simplification and
l for limitation of effects). Equation 5 allows negative values for the simplification parameter, although
limiting to 0 the minimum of the final ISI. In
the subsequent analysis, the minimum of the ISI range
(i.e. ) is set equal to the minimum of HCI ( = 5),
in order to balance the calculation of ISPI.
3.2.1 ISI for attenuation
Figure 1 reports the monograph proposed to convert
the extent of applicability of the guideword attenuation into an ISI value. The extent of applicability
of this guideword is assessed mainly as the ability of
the layout option to reduce the hazard potential from
domino effects. To overcome the subjectivity in the
assessment, a reference index based on the ratio of
the Domino Hazard Index (DHI) among the unit of
concern and the base option is adopted:
Reference Index =

DHIoption
DHIbase case

(6)

(a similar ratio of a key parameters will be the

approach used for all the following reference indices).
The DHI is specifically aimed at assessing the domino
effect hazards caused by a unit in a specific layout. The
DHI values for a unit range from 0 to 100. The maximum value means that the unit can affect multiple
other units, triggering severe domino consequences;
the zero-value indicates no domino possibility from
the unit (i.e. inherent safety). It is calculated from the
scores attributed to the potential maximum ability of
a unit to trigger domino escalation.
DHIi =


k

maxh (DHSi,k,h )

(7)

where DHSi,k,h is the score given to the hazard in

terms of an escalation from unit i to unit k by event
h. The criteria for the assignation of the scores are
described in detail by Tugnoli et al. (in press). In this
application, the protection provided by passive devices
is not accounted for in DHI as the focus is on the
domino escalation potential (i.e. hazard) that can be
reduced only by inherent measures.
3.2.2 ISI for simplification
As previously discussed, layout options aiming at safer
performance with respect to attenuation and limitation
of effects usually incur an increase in layout complexity. Some important factors that may result in the
complexity are:
Complication of pipe connection among units: the
displacing of units from the logical process-flow
arrangement makes the piping network, pumping
and control more complicated.
Complication of pipe connection among units: the
need for longer pipelines to connect units provides
additional sources of release. Thus, additional units
should be considered (transportation units).
Increase in the number of items in a plant: in addition to the process units, other elements (e.g. blast
walls, fire walls, etc.) contribute to the number
of items present on the site. As the number of
items increases, the requirements for management
and maintenance increase, thus complicating procedures and increasing the probability of errors.
Moreover, it creates obstacles (blast walls, dikes,
etc.) that limit the ease of access to the units, complicating regular operations (e.g. maintenance) as
well as emergency response operations (e.g. firefighting).
To assess the complexity introduced by both inherent and passive measures, the former ISI of simplification is extended to account for negative values
(Table 1). This is based on the idea that complexity
can be defined as negation of simplicity. Pursuing simplification indeed limits the increase in complexity up
to values that are overbalanced by the positive effects
from the application of other guidewords.
It is worth remembering that simplification is a
matter of interrelation among process units. It must
be judged not by focusing only on a single unit, but
with respect to the whole plant (or occasionally to a
plant section). The extent of the applicability of this
guideword should be assigned in terms of unit groups.
3.2.3 ISI for limitation of effects
Analysis of the applicability of limitation of effects
to layout design involves three different elements, as
identified in the discussion on the applicability of the
guidewords in layout design. These parameters are

1151

combined by Equation 8:

3.3 Cost indexing

ISIl = Min{100, [(ISIle )3 + (ISIlb )3 + (ISIla )3 ]1/3 }

(8)

The cost indexing procedure of I2SI accounts for and

evaluates the economic aspects of inherent safety. The
costing system is comprised of three sub-indices: the
Conventional Safety Cost Index (CSCI), the Inherent
Safety Cost Index (ISCI) and the Loss Saving Index
(LSI).
The CSCI and ISCI express, for the base case and
the options improved by inherent safety application
respectively, the costs of safety per unit of potential
monetary loss in case of accidents. The LSI accounts
for the savings from avoided potential loss due to a
reduction of domino escalation possibility. They are
calculated by Equations 1113:

Also in this case, suggestions are provided for

guidance in the evaluation of extent of applicability,
striving to reduce the degree of subjectivity in the
analysis:
i) Limitation of the effects of domino escalation
(ISIle ) can be estimated by resorting to a reference
index based on the ratio of Domino Hazard Indices
(DHIs). The approach is similar to that followed for
attenuation, but the focus, as discussed earlier, is
different in this case. In limitation of the effects of
domino escalation, the DHI is calculated considering
the synergistic effect of passive and inherent measure
protection on domino consequence limitation.
ii) Limitation of the damage potential to target
buildings (ISIlb ) aims to assess the location of the facilitys populated buildings (control rooms, offices, etc.)
in relation to the hazardous units of the process.
A proposal for guiding this assessment is based on
the grouping of the buildings into hazard-susceptible
areas (i.e. areas affected by fire, explosion and acute
toxic effects). The assessment takes into account
the combined effect of different primary units on the
same building. Thus, a reference index is defined as
the ratio of a parameter (A), that is calculated for each
target building (j) according to Equations 910:


Di,j
ai,j = Max 1
;0
Bi
A=

 
ai,j >0

ai,j

CConvSafety
CLoss

CSCI =
ISCI =

CInhSafety
CLoss

LSIoption =

(11)

(12)

CInhSafety,opt. + (CLoss,opt. CLoss,base opt. )

(13)
Closs,base option

Cost of loss, CLoss , represents the dollar value of

expected losses caused by accidental events in a unit.
It is comprised of five components, as shown in
Equation 14:

CLoss = CPL + CAL + CHHL + CECC + CDEC

(14)

(9)
(10)

where Di,j is the distance between the i-th unit and the
j-th building, and Bi is the maximum damage distance of the i-th unit for fire, explosion and acute toxic
effects.
iii) Limitation of the affected area (ISIla ) accounts
for the effects of passive measures to decrease the
area susceptible to dangerous consequences, no matter if particular structures are located there (e.g. units
or buildings), but simply because final targets (e.g.
people, environment) can potentially be present. The
suggested guideline for quantitative assessment of this
aspect is based on the percentage decrease of damage
area compared (i.e. ratio) to the same unit in the base
option. The parameter used in the ratio is thus the
affected area (AA) exposed to the consequence from
the unit considered (e.g. if no protective devices exist,
this is the area encompassed by the damage radius; if
protective devices exist, the upwind protected areas
are subtracted).

Production Loss (PL) is the economic consequence

of production shutdown. Direct Asset Loss (AL) represents the value of the physical unit itself which is
depleted by the accidental event. Human Health Loss
(HHL) is calculated in terms of the cost of fatalities/
injuries directly caused by the accident at the unit. The
current authors acknowledge that there can be a high
degree of subjectivity and discomfort associated with
assigning a dollar value to fatality and/or injury. However surrogate values can be defined (e.g. insurance
costs, worker compensation rates, etc.). Environmental Cleanup Cost (ECC) is associated with the mass or
volume of soil, water and air that were contaminated by
the accidental event. Domino Escalation Cost (DEC)
is a cost term explicitly introduced to account for the
loss consequences of the possible chain of accidents.
It represents the sum of the loss of secondary units
involved, weighted by a parameter that features the
probability of being involved, as for Equation 15:
CDEC =

1152


k

sk (CAL,k + CHHL,k + CECC,k )

(15)

1.0

0.8

The proposed index methodology is demonstrated by

a case study on the selection of the inherently safer
layout for an acrylic acid production plant. The acrylic
acid synthesis is based on propylene oxidation. Three
alternative layout options are proposed and compared:

sk

0.6
0.4
0.2
0.0

5
6
DHSi,k

10

Figure 3. Credit factor for domino escalation as a function

of maximum DHSi,k for the primary unit.

where CAL,k , CHHL,k and CECC,k are respectively the

additional direct asset loss, human health loss and
environmental cleanup costs for the failure of each
k-th secondary unit, as a result of escalation. The factor sk accounts for the credibility that the failure of the
considered unit affects the k-th secondary unit. It can
be evaluated as a function of the maximum Domino
Hazard Score, as in Figure 3.
The cost of conventional safety, CConvSafety , is the
sum of the costs of process control measures and addon (end-of-pipe) safety measures (i.e. CConvSafety =
CControl +CAdd-on ). It can be estimated by the number
of measures required and their representative reference
costs (see e.g. Khan & Amyotte 2005).
The cost of safety with inherent safety implementation, CInhSafety , is the sum of the costs of inherent safety
implementation, of process control measures, and of
add-on (end-of-pipe) safety measures still required
in the inherently safer layout option (i.e. CInhSafety =
CInherent + CControl + CAdd-on ). The costs for inherent
safety implementation is the marginal cost (i.e. capital
cost difference of the assessed option relative to the
base case), corrected by the extent of applicability of
each guideword (Equation 16). For example, the cost
of extra space required for increased unit segregation
is estimated and referred to as the cost of implementing
the guideword attenuation.

EXAMPLE OF APPLICATION

Option 1 (base option), Figure 4a, was designed in

accordance with typical safety rules used in industrial practice for separation distances (Mecklenburg
1973, Lees 1996 and references cited therein). No
passive protection devices were considered for this
option. The plant can be divided into four distinct
areas: the process area, the pressurized storage of
liquefied propylene, the atmospheric tank farm and
the loading facility.
Option 2 presents the same plot for the units as in
the base case (Fig. 4a). However, passive protection
measures (catch basin for liquid storage tanks bunds
and fire insulation of the vessels) are considered
in this case. This option is explicitly proposed to
demonstrate how safety performance is influenced
by the passive protections in place.
Option 3 incorporates segregation of units, a
modified spatial arrangement and passive protection devices, all aimed at enhancing layout safety
(Fig. 4b).
Application of the proposed index methodology to
the plot plan of the base case (Table 2) reveals that
the safety distances from the literature are effective
in preventing escalation from storage to process area
and vice-versa (the oxidation reactor is assumed as
reference for the process area). This is also due to the

2 3

2 3 5

7 4 6
5

CInherent = CA /EA + CSi /ESi + CL /EL

(16)

7
6

1
0

where the C variables are the costs and E the extents

of applicability of, respectively, attenuation (A), simplification (Si), and limitation of effects (L).

50 m

(a)

(b)

Figure 4. Layout plan considered in the case study: (a)

options 1 and 2; (b) option 3.

1153

Table 2. Summary of the principal indices evaluated in the assessment. AA: acrylic acid, AcA: acetic acid, Sol: solvent,
P: propylene.
#

Unit

ISI
Option 1

HCI
DI
(base case)

1
2
3
4
5
6
7
8
9

Reactor
AA storage
AA storage
AA storage
AcA storage
AcA storage
Sol storage
P storage
P tanker

5.0
5.0
5.0
5.0
5.0
5.0
5.0
5.0
5.0

42
24
24
24
24
24
24
36
36

47
36
36
36
38
38
14
40
27

PHCI

I2SI

CSCI

LSI

92
56
56
56
56
56
55
73
73
Total

0.23
0.32
0.32
0.32
0.31
0.31
0.82
0.25
0.38
0.0077

0.46
0.28
0.28
0.28
0.28
0.28
0.15
0.66
0.18

0.46
0.28
0.28
0.28
0.28
0.28
0.15
0.66
0.18

Option 2
1
2
3
4
5
6
7
8
9

Reactor
AA storage
AA storage
AA storage
AcA storage
AcA storage
Sol storage
P storage
P tanker

5.0
44.0
32.6
33.0
35.4
40.6
68.7
5.0
5.0

42
24
24
24
24
24
24
36
36

47
36
36
36
38
38
14
40
27

92
56
56
56
56
56
55
73
73
Total

Option 3
1
2
3
4
5
6
7
8
9

Reactor
AA storage
AA storage
AA storage
AcA storage
AcA storage
Sol storage
P storage
P tanker

5.0
50.8
50.8
50.8
68.5
68.5
94.9
5.0
60.5

42
24
24
24
24
24
24
36
36

47
36
36
36
38
38
14
40
27

choice, derived by safety experience and common to

all proposed layouts, of locating the propylene storage at the furthest feasible distance from the process
area. Another observation concerning the propylene
storage is that this area has the same vessel layout in
all options and therefore does not require assessment
of safety improvement possibilities. (This again highlights the fact that indices such as I2SI are intended
to be used in a relative, not absolute, manner to effect
risk reduction). The I2SI values for the storage and
loading area (Table 2) are below unity. The Inherent
Safety Index (ISI) has low values, as expected for the
base case. The values of Damage Index (DI) illustrate that all units have significant damage distances

92
56
56
56
56
56
55
73
73
Total

I2SI

ISCI

LSI

0.23
2.85
2.11
2.14
2.17
2.49
11.3
0.25
0.38
4.1

0.39
0.25
0.25
0.25
0.25
0.25
0.14
0.63
0.18

0.19
0.46
0.44
0.44
0.45
0.46
0.81
0.57
0.18

I2SI

ISCI

LSI

0.23
3.29
3.29
3.29
4.20
4.20
15.5
0.25
4.54
53.0

0.33
0.25
0.25
0.25
0.25
0.25
0.15
0.71
0.15

0.27
0.47
0.47
0.47
0.50
0.50
0.82
0.52
0.47

and hence significant potential to trigger escalation.

As expected, propylene storage appears as a critical
safety issue. The storage of the solvent, however, is a
relatively low hazard unit with respect to escalation.
From Table 2, it can be observed that storage units
usually have large costs of losses, mainly due to the
large extent of damage propagation by domino effect.
For storage, ISCI values are therefore lower than unity,
suggesting that further protective measures should be
applied.
In option 2 (Tab. 2), the spatial disposition of the
units is the same as in option 1. Thus, the I2SI values mainly reflect the effect of passive measures in the
limitation of escalation. The combined protection of

1154

bunding and fire insulation increases the index values

of the atmospheric storage units well above unity.
With respect to the cost indices, option 2 shows only a
minor decrease in ISCI values compared to those for
CSCI, meaning that the safety savings are usually an
order of magnitude lower than the total costs. This is
a direct consequence of the similar values of costs for
conventional safety and inherent safety.
In option 3 (Table 2), the improved unit spatial disposition has no effect in preventing escalation among
unprotected atmospheric storage units (low values of
ISIa ). However, if combined with passive protection,
an enhanced safety performance is obtained (high values of ISIl , mainly due to limitation of the effects of
domino escalation). On the other hand, the increased
segregation does prevent, from an inherent safety perspective, escalation triggered by the propylene tank
trucks at the loading station. The I2SI values are well
above unity for all units, with good performance of the
solvent storage unit made possible by coupling both
inherent and passive measures. With respect to cost
indices, the same conclusions of option 2 generally
hold for option 3. Here, though, the ISCI for propylene storage is higher than the CSCI because of the
high land cost for unit separation. The system I2SI for
options 3 clearly shows the better performance, due to
the positive contribution of every unit. The analysis of
the Loss Saving Index (Table 2) generally results in the
same outcomes already discussed for the I2SI results.
This is due to the predominant effect of the values of
DHI on both indices (LSI and I2SI), as is evident in
the case of the loading facility. The presence of several
negative values means that the cost of these measures
is fully compensated by the expected decrease in loss
in the event of an accident.
5

CONCLUSIONS

Layout plays a key role in defining the safety of a process plant. An attempt to bring inherent safety criteria
into layout design has been presented in the current
paper. A novel indexing approach was developed to
guide inherently safer choices in the early phases of
layout design. Application of the proposed safety

index methodology to a case study has demonstrated

the effectiveness of the approach. It allowed identifying the hazard related to the unit spatial arrangement
and the effect of the design measures adopted to
improve the layout safety.
REFERENCES
Cozzani, V., Tugnoli A. & Salzano E. 2007. Prevention
of domino effect: from active and passive strategies to
inherently safer design. J. Haz. Mat. A139: 209219.
Gupta, J.P., Hendershot, D.C. & Mannan, M.S. 2003. The real
cost of process safetya clear case for inherent safety,
Proc. Safety Env. Protect. 81: 406413.
Hendershot, D.C. 1999. Designing safety into a chemical process. In Proc. 5th Asia Pacific Responsible Care
Conference, 710 November 1999; Shanghai.
Khan, F.I. & Abbasi, S.A. 1999. The worlds worst industrial
accident of the 1990s. Proc. Safety Prog. 18: 135145.
Khan, F.I. & Amyotte P.R. 2003. How to make inherent safety
practice a reality. Can. J. Chem. Eng. 81: 216.
Khan, F.I. & Amyotte, P.R. 2004. Integrated inherent safety
index (I2SI): a tool for inherent safety evaluation. Proc.
Safety Prog. 23: 136148.
Khan, F.I. & Amyotte, P.R. 2005. I2SI: a comprehensive
quantitative tool for inherent safety and cost evaluation.
J. Loss Prev. Proc. Ind. 18: 310326.
Khan, F.I., Husain, T. & Abbasi, S.A. 2001. Safety weighted
hazard index (SWeHI), Proc. Safety Env. Protect.
79: 6580.
Kletz, T.A. 1978. What you dont have, cant leak. Chem. &
Ind. May 6: 287292.
Kletz, T.A. 1998. Process plants: a handbook for inherent
safer design, Bristol: Taylor & Francis.
Lees, F.P. 1996. Loss prevention in the process industries.
2nd ed. Oxford: Butterworth-Heinemann.
Mecklenburg, J.C. 1973. Plant layout. New York: John
Wiley & Sons.
Patsiatzis, D.I., Knight, G. & Papageorgiou L.G. 2004. An
MILP approach to safe process plant layout, Chem. Eng.
Res. Des. 82: 579586.
Penteado, F.D. & Ciric, A.R. 1996. An MILP approach
for safe process plant layout. Ind. Eng. Chem. Res.
4: 13541361.
Tugnoli A., Cozzani V. & Landucci G. 2007. A Consequence Based Approach to the Quantitative Assessment
of Inherent Safety. AIChE Journal 53 (12): 31713182.
Tugnoli, A., Khan, F.I., Amyotte P.R. & Cozzani, V.
in press. Safety Assessment in Plant Layout Design
using Indexing Approach: Implementing Inherent Safety
Perspective. Part 2Domino Hazard Index and Case
Study. J. Haz. Mat.

1155

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Minmax defense strategy for multi-state systems

G. Levitin
The Israel Electric Corporation Ltd., Haifa, Israel

K. Hausken
Faculty of Social Sciences, University of Stavanger, Norway

ABSTRACT: Protecting against intentional attacks is fundamentally different from protecting against accidents or natural cataclysms. Choosing the time, place, and means of attacks, the attacker has always an advantage
over the defender. Therefore, the optimal defense policy should take into account the attackers strategy. When
applied to multi-state systems, the damage caused by the destruction of elements with different performance rates
will be different. Therefore, the performance rates of system elements should be taken into account when the
damage caused by the attack is estimated. This paper presents a generalized model of damage caused to a complex
multi-state system by intentional attack. The model takes into account the defense strategy that presumes separation and protection of system elements. The minmax defense strategy optimization methodology is suggested,
in which the defender chooses a separation-protection strategy that minimizes the damage that the attacker can
cause by most harmful attack.
Keywords: multi-state system, survivability, defense, attack, minmax, universal generating function.

INTRODUCTION

Protecting against intentional attacks is fundamentally

different from protecting against accidents or natural
cataclysms (Bier et al. 2005, Azaiez & Bier 2007).
Adaptive strategy allows the attacker to target the most
sensitive parts of a system, which may be those that
are most essential for system functioning, or easiest to
attack, or most difficult to defend. Choosing the time,
place, and means of attacks, the attacker usually has
an advantage over the defender. The optimal policy for
allocating resources among possible defensive investments should take into account the attackers strategy.
This paper presents a general optimization methodology that merges game theory (Fudenberg & Tirole
1991) and multi-state system survivability theory. It
considers a situation where the system is composed
from components consisting of elements characterized
by different performance and reliability. The defender
can separate the elements and protect them using
different protection actions.
Different models of attackers behavior are considered. In probabilistic models the attacker uses only
single attack method, but can optimally chose subset of system elements to attack. The attack success
depends only on the type of protection. In combinatorial model the attacker can use different combination of
attack actions against separated and protected groups

of elements. For each protected group of elements the

attack success probability (probability of the group
destruction) is endogenously determined for every
combination of protection and attack actions.
The damage caused by the attack is associated with
partial system incapacitation (reduction of the overall
system performance) and with the losses of inherent
value of the destroyed system elements. In order to
determine the expected damage a procedure for evaluating system performance distribution based on a
universal generating function technique is suggested.
In Section 2, the basic definitions and the system
model are presented and the optimization problem
is formulated. Section 3 introduces different attack
models. The computational technique for evaluating
the system performance for arbitrary attacker and
defender strategies is described in Section 4. The
genetic algorithm based optimization approach is
discussed in Section 5.

2
2.1

THE MODEL
Basic definitions

Element. Lowest-level part of the system, which

is characterized by its inherent value, availability, and nominal performance rate; and can have

1157

the same successful attack. More than one PG cannot

be destroyed by a single attack.
Because system elements with the same functionality can have different performance rates, and different availability, the way the elements are distributed
among the PGs affects the system survivability. The
element separation problem for each component n can
be considered as a problem of partitioning a set n
of Kn items into a collection of Mn mutually disjoint
subsets nm , i.e. such that

two states: normal operation, and total failure.

Multi-state systems that can work with different
performance rate can usually be decomposed into
two-state elements.
Performance rate. Quantitative measure of task performing intensity of element or system (capacity,
productivity, processing speed, task completion
time, etc.)
Availability. Probability that the element is able to
perform its task (not in failure state) when it is
needed.
Component. Collection of elements with the same
functionality connected in parallel in the reliability
logic-diagram sense.
Protection. Technical or organizational measure
exerted by defenders aimed at the reduction of
the destruction probability of a group of system
elements in the case of attack.
Separation. Action exerted by defenders aimed at
preventing the simultaneous destruction of several
elements in the case of a single attack (can be
performed by spatial dispersion, by encapsulating
different elements into different protective casings,
by using different power sources, etc.)
Protection group. Group of system elements separated from other elements, and possibly protected,
so that a single external impact destroying elements belonging to a certain group cannot destroy
elements from other groups.
Attack. Action exerted by attackers to damage system
by destroying it or its parts.
Inherent value. Loss measured in some currency
incurred by defenders irrespective of the loss caused
by reduced system performance.
Vulnerability. Probability that a PG is destroyed causing zero performance of all its elements.
Attack action. Attackers method chosen to destroy a
PG (suicide bombing, missile attack etc.) characterized by its cost and efficiency (influence on PG
vulnerability).

The partition of the set n can be represented by

the vector {nk , 1 k Kn }, where nk is the
number of the subset to which element k belongs
(1 nk Mn ). The matrix of values nk for
1 k Kn , and 1 n N determines the elements
distribution among the protection groups for the entire
system (separation strategy of the defender).
For each protection group belonging to component
n, there are en + 1 available protection actions. Each
protection of type nm (0 nm en ) is characterized by its cost cn (nm ). The protection cost of any
PG m in component n can also depend on the number
of elements it comprises: cn (nm , |nm |). Protection
action nm = 0 corresponds to the absence of any protection. The cost of protection action 0 can be greater
than zero because it represents the cost of the common infrastructure of the PG (the separation usually
requires additional areas, constructions, communications, etc.) The matrix of the values of nm chosen
for any PG m, and component n, represents the entire
protection strategy of the defender.
The entire defenders defense strategy is defined
as = { , }. The total cost of the system defense
strategy (separation and protection) can be determined as

2.2

c() =

System description and problem formulation

Mn


nm = n ,ni

nj = ,

i  = j.

(1)

m=1

Mn
N 


cn (nm , |nm |).

(2)

n=1 m=1

The system consists of N statistically independent

components composing a series-parallel configuration. Each component n consists of Kn elements of the
same functionality connected in parallel. Each element
k in component n is characterized by its inherent value
wnk , nominal performance rate xnk , and availability
pnk . The states of the elements are independent.
The elements within any component can be separated (to avoid the entire component destruction by
a single attack). Parallel elements not separated from
one another are considered to belong to the same protection group (PG). The PGs can be protected. All the
elements belonging to the same PG are destroyed by

It is assumed that only a single attack is possible

against each PG since the attack leads to the attacker
being detected and disabled (Bier et al. 2005, Azaiez &
Bier 2007).
The strategy of the attacker can be represented by
matrix = {nm |1 n N , 1 m Mn }, where
nm characterizes attack against PG m in component n
(depending on problem formulation nm can represent either attack probability or attack type). For any
given defenders strategy , and attackers strategy ,
one can determine the probabilistic distribution of the
entire system performance (pmf of random value G)

1158

in the form gs , qs (, ) = Pr(G = gs )(1 s S)

using the algorithm presented in Section 3).
If the system totally fails when its performance
becomes lower than the demand (and the entire
demand F is not supplied), the expected cost of the
damage is

(, , F) = hF

S


qs (, ) 1(gs < F).

(3)

s=1

where h is the cost of damage caused by the system

performance reduction by 1 unit below the demand. If
the damage associated with the system performance
reduction below the demand F is proportional to the
unsupplied demand, the expected cost of the damage is

(, , F) = h

S


qs (, ) max(F gs , 0).

(4)

The total expected damage caused by the attack

should include the cost of losses associated with system performance reduction, and losses of inherent
values of the destroyed elements and the infrastructure.
The cost of the total expected damage is

d(, , F) =

3.1 Probabilistic models

These models presume that the attacker chooses a subset of PGs to attack and attacks different PGs the same
manner. The destruction probability of any attacked
PG depends only on its protection and nm defines
the probability of the attack on PG m in component n.
In this case the vulnerability of PG m in component
n takes the form vn (nm , nm ) = nm vn (nm ), where
vn (nm ) is the conditional destruction probability of
PG with protection nm given it is attacked.
3.1.1 Single attack
The assumption that only a single attack is possible
is realistic when the attacker has limited resources, or
when the attack leads to the attacker being detected &
disabled. In this case, the attacks on different PG are
mutually exclusive events, and
nm = 1.

(7)

n=1 m=1

In the case when the attacker has perfect knowledge about the system and its defenses, the attackers
strategy is
= (n, m), where
(n, m) =

vn (nm , nm )

n=1 m=1

ynm +

arg

{d((n, m), , ) max}

1nN ,1mMn

(8)

wnk + (, , F).

knm

(5)
where ynm is inherent value of the m-th PG infrastructure in component n, vn (nm , nm ) is expected
vulnerability (destruction probability) of PG m in
component n for attack nm and protection nm .
We consider the case when the defender has limited
budget b. The defender builds the system over time.
The attacker takes it as given when he chooses his
attack strategy. Therefore, we analyze a two period
minmax game where the defender moves in the first
period, and the attacker moves in the second period.
This means that the defender chooses a strategy in
the first period that minimizes the maximum damage
that the attacker can cause in the second period:

Here (n, m) is the matrix in which all the elements

are equal to zero, besides element nm which is equal
to one.
If the attacker has perfect knowledge about the system itself, but not about its defenses, the attacker tries
to maximize the expected damage assuming that different PG are equally protected (it can be assumed that
protections of type 0 are used for any PG). In this case,
the optimal attackers strategy is
= (n, m), where
(n, m) =

arg

{d((n, m), 0, ) max} (9)

1nN ,1mMn

If the attacker has no information about the system,

or cannot direct the attack precisely, the defender can
assume that the attacker chooses targets at random, and
N


= { , } = arg {d( (), ) min},

nm = 1

where ( ) = arg {d(, ) max} for any

subject to c( ) b.

ATTACK MODELS

Mn
N 


s=1

Mn
N 


Mn

n=1

(6)

for any component n and PG m.

1159

(10)

3.1.2 Multiple attacks

The attacks can take place sequentially, or simultaneously. However, following (Azaiez & Bier 2007), we
assume that the attacks are independent. Their probabilities cannot be changed in accordance with achieved
results; and successes, and failures of different attacks
are independent events.
Because several targets can be attacked, the assumption (7) on the attackers strategy does not hold. In the
worst case of unlimited attackers resources, any target can be attacked with probability 1: nm = 1 for
1 n N , 1 m Mn . If the attackers budget is
limited, and the attackers knowledge about the system
is perfect, the most effective attack strategy is
= arg{d(, , ) max} subject to
Mn
N 


nm Cnm B,

nm = {0, 1},

= arg{d(, , ) max} subject to C() B.

(13)

Having the system performance distribution in the

form gs , qs (, ) for 1 s S, one can obtain
the expected damage d(, ) using (3), (4) and (5).
The system performance distribution can be obtained
for arbitrary defenders and attackers strategies using
the universal generating function (u-function) technique, proven to be an effective tool for reliability
analysis, and optimization (Levitin 2005).
4.1

where Cnm is the cost of the attack on PG m in

component n, and B is the attackers budget.
When the attackers knowledge about the system
is imperfect or deceptive, the attack probabilities can
have positive or negative correlation with the expected
damage caused by the attacks.

3.2 Combinatorial model

This model presumes that the attacker has a set of En +
1 available attack actions against any component n.
For example, the same target (PG) can be attacked
by a suicide bomber, by missiles with different types
of warhead etc. The strategy of the attacker can be
represented by matrix = {nm |1 n N , 1
m Mn }. Each attack action nm (0 nm En ) is
characterized by its cost Cn (nm ). If nm = 0 the PG
m of component n is not attacked. The cost of attack
action 0 is Cn (0) = 0. The total attack cost for any
attack strategy is
Mn
N 


Cn (nm ).

EVALUATING THE SYSTEM

PERFORMANCE DISTRIBUTION

(11)

n=1 m=1

C() =

strategy is

(12)

n=1 m=1

Having the protection strategy and the attackers

strategy , one can obtain the probability of destruction for any PG m in component n as vn (nm , nm ).
This function can be elicited from previous experience or from the expert opinion and can be represented
in a table form for each component. It is obvious
that vn (nm , 0) = 0 for any nm . For any given
defense strategy = { , } the most effective attack

Universal generating function technique

The u-function representing the pmf of a discrete

random variable  is defined as a polynomial
u (z) =

J


j z j ,

(14)

j=0

where the variable  has J + 1 possible values, j is

the j-th realization of , and j = Pr( = j ).
To obtain the u-function representing the pmf
of a function of two independent random variables
(, ), the following composition operator is used:

J

j z j
U(,) (z) = u (z)u (z) =

j=0

I

i=0


i z i

I
J 


j i z (j ,i )

j=0 i=0

(15)
This polynomial represents all of the possible mutually exclusive combinations of realizations of the
variables , and  by relating the probabilities of
each combination to the value of the function (, )
for this combination.
In our case, the u-functions can represent performance distributions of individual system elements,
and their groups. Any element k of component n
can have two states: functioning with nominal performance xnk (with probability pnk ), and total failure
(with probability 1-pnk ). The performance of a failed
element is zero. The u-function representing this

1160

performance distribution takes the form

unk (z) = pnk z xnk + (1 pnk )z 0

(16)

If, for any pair of elements connected in series or

in parallel, their cumulative performance is defined
as a function of individual performances of the elements, the pmf of the entire system performance can
be obtained using the following recursive procedure
(Levitin 2005).
Procedure A.
1. Find any pair of system elements connected in
parallel, or in series.
2. Obtain the -function of this pair using the
corresponding composition operator over two

u-functions of the elements, where the function is

determined by the nature of the interaction between
elements performances.
3. Replace the pair with a single element having the
u-function obtained in step 2.
4. If the system contains more than one element,
return to step 1.
The choice of the composition functions depends
on the type of connection between the elements, and
on the type of the system. Different types of these functions are considered in (Levitin 2005). For example,
in systems with performance measure defined as productivity or capacity (continuous materials or energy
transmission systems, manufacturing systems, power
supply systems), the total performance of elements
connected in parallel is equal to the sum of the performances of its elements. Therefore, the composition
function for a pair of elements connected in parallel
takes the form
par (, ) =  + .

(17)

When the elements are connected in series, the

element with the lowest performance becomes the
bottleneck of the system. Therefore the composition
function for a pair of elements connected in series is
ser (, ) = min(, ).
4.2

(18)

Incorporating PG destruction probability

The u-function Unm (z) for any PG m in component n

can be obtained using Procedure 1 with composition
operator over all the elements belonging to the
par

set nm . This u-function represents the conditional

pmf of the PGs cumulative performance given the PG
is not destroyed by an attack. If the PG is protected
by the protection of type nm and the number of the
attack action undertaken against this PG is nm , the PG

be destroyed with probability vn (nm , nm ). To obtain

the unconditional pmf of the PGs performance, one
should multiply by 1vn (nm , nm ) the probabilities of
all the PGs states in which the group has nonzero performance rates. The u-function U nm (z) representing
the unconditional pmf can be obtained as follows.
U nm (z) = [1 vn (nm , nm )]nm (z) + vn (nm , nm )z 0
(19)
Having the operators (15) and (19), we can apply
the following procedure for obtaining the pmf of the
entire system performance for any given defenders
strategy , and attackers strategy .
Procedure B.
1. For any component n = 1, . . . , N :
1.1. Define Un (z) = z 0 .
1.2. For any nonempty PG (set nm ):
1.2.1. Define Unm (z) = z 0 .
1.2.2. For any element k belonging to nm ,
modify Unm (z) as follows:
Unm (z) = Unm (z) unk (z).
par

1.3. For 1 m Mn obtain the u-function U nm (z)

representing the unconditional pmf of PG m
using (19).
1.4. Modify the u-function Un (z) as follows:
Un (z) = Un (z) U nm (z).
par

2. Apply Procedure A over the u-functions of the

components in accordance with the series-parallel
system structure.

APPLYING A GENETIC ALGORITHM

TO DETERMINE THE MINMAX SOLUTION

When facing complexity, heuristic search procedures

are often called for. One way around this challenge is to
apply genetic algorithms which have proven successful in many areas including reliability optimization
(Levitin 2006a, b) and game theory (Marks 2002).
5.1 Structure of GA
The basic structure of the version of GA referred to
as GENITOR (Whitley 1989) is as follows. First, an
initial population of Ns randomly constructed solutions (strings) is generated. Within this population,
new solutions are obtained during the genetic cycle
by using crossover, and mutation operators. The

1161

crossover produces a new solution (offspring) from

a randomly selected pair of parent solutions, facilitating the inheritance of some basic properties from
the parents by the offspring. Mutation results in slight
changes to the offsprings structure, and maintains a
diversity of solutions. This procedure avoids premature convergence to a local optimum, and facilitates
jumps in the solution space.
Each new solution is decoded, and its objective
function (fitness) values are estimated. The fitness
values, which are a measure of quality, are used
to compare different solutions. The comparison is
accomplished by a selection procedure that determines
which solution is better: the newly obtained solution,
or the worst solution in the population. The better
solution joins the population, while the other is discarded. If the population contains equivalent solutions
following selection, redundancies are eliminated, and
the population size decreases as a result.
After new solutions are produced Nrep times, new
randomlyconstructedsolutionsaregeneratedtoreplenish
theshrunkenpopulation, andanewgeneticcyclebegins.
The GA is terminated after Nc genetic cycles. The
final population contains the best solution achieved.
It also contains different near-optimal solutions which
may be of interest in the decision-making process.
5.2

Solution encoding in the GA

The GA requires solution representation in the

form of strings. Any defense strategy = { , }
can be represented by concatenation of n integer
strings {nk , 1 k Kn }, and n integer strings
{nm , 1 m Mn } for 1 n N . The total length of

the solution representation string is 2 Nn=1 Kn . The
substring determines the distribution of elements
amongprotectiongroups, andthesubstringdetermines
types of protections chosen for the PGs. Because the
maximal possible number of protections is equal to the
total number of elements in the system (in the case
of total element separation), the length of substring
should be equal to the total number of the elements. If
the number of PG defined by substring is less than
the total number of system elements, the redundant
elements of substring are ignored.
For probabilistic attack model (11) any attack strategy can be represented by concatenation of n
Boolean strings {nm , 1 m Mn } for 1 n N ,
nm {0, 1}. For combinatorial attack model any
attack strategy can be represented by concatenation
of n integer strings {nm , 1 m Mn } for 1 n N .
5.3

Double-GA

The double-GA approach that guarantees obtaining

a minmax solutions uses two independent GA procedures for finding optimal defenders and attackers

strategies. An external GA seeks for minimizing

the expected damage d. For each generated by the
external GA during the genetic search an internal GA
is applied that seeks maximizing d while is fixed.
The maximal value of d obtained by the internal GA
is used as the value of the objective function in the
external GA.
Examples of minmax strategies obtained by the
GA can be found in (Levitin 2007) and (Levitin &
Hausken).

CONCLUSION

This paper considers a system which constitutes a

series-parallel combination of components consisting of functionally identical parallel elements. The
defender can separate the elements within each component into protection groups to prevent that the entire
component is destroyed by an attack. All the elements
belonging to the same protection group are destroyed
by the same successful attack. The defender has an
available set of protection strategies for all possible
separation strategies, which specifies how each protection group can be protected. A cost of separation
and protection is associated with each strategy. In the
probabilistic attack model the attacker chooses a subset of protection groups to attack. In the combinatorial
attack model the attacker has an available set of attack
actions and decides which action to apply against each
protection group.
The defenders goal is to minimize the maximal
possible expected damage caused by the attack. The
defender takes into account the multi-state nature of
the system, the unreliability of system elements, the
probabilistic distribution of the system performance,
the losses of inherent values of destroyed elements,
and the required demand. The damage can be proportional to the unsupplied demand, or to the probability
of not meeting the demand. The defender and attacker
can have limited or unlimited budgets.
The optimization methodology consists in determining the minmax solution for the two period noncooperative game between the defender moving first
and the attacker moving second. The defender chooses
the optimal separation and protection for each protection group from the available set of separation and
protection strategies, while the attacker chooses the
optimal attack strategy maximizing the expected damage. Both the defender and the attacker have perfect
knowledge about the structure of the game, the strategy
sets, and all parameters.
A universal generating function technique for evaluating the losses caused by system performance reduction is presented. A double-genetic algorithm which
determines the solution is discussed.

1162

REFERENCES
Azaiez, N. and Bier, V.M. (2007). Optimal Resource Allocation for Security in Reliability Systems. European Journal
of Operational Research, 181 (2), pp.773786.
Bier, V.M., Nagaraj, A. and Abhichandani, V. (2005). Protection of simple series and parallel systems with components
of different values, Reliability Engineering and System
Safety 87, pp. 313323.
Fudenberg, D. and Tirole, J. (1991). Game Theory. MIT
Press, Cambridge, MA.
Levitin, G. (2005). Universal generating function in reliability analysis and optimization, Springer-Verlag.
Levitin, G. (Ed.). (2006a). Computational Intelligence in
Reliability Engineering. Evolutionary Techniques in Reliability Analysis and Optimization, Series: Studies in
Computational Intelligence, Vol. 39, Springer-Verlag.

Levitin, G. (2006b). Genetic Algorithms in Reliability Engineering. Guest editorial. Reliability Engineering and
System Safety, 91(9), pp. 975976.
Levitin, G. (2007). Optimal Defense Strategy against Intentional Attacks. IEEE Transactions on Reliability, 56(1),
pp. 148157.
Levitin, G. and Hausken, K. (2008). Minmax defense strategy
for complex multi-state systems. Reliability Engineering
and System Safety. (submitted).
Marks, R.E. (2002). Playing games with Genetic Algorithms. In S.-H. Chen (editor), Evolutionary Computation
in Economics and Finance, Physica Verlag.
Whitley, D. (1989). The GENITOR Algorithm and Selective Pressure: Why Rank-Based Allocation of Reproductive Trials is Best. Proc. 3th International Conf. on
Genetic Algorithms. D. Schaffer, ed., Morgan Kaufmann,
pp. 116121.

1163

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Multicriteria risk assessment for risk ranking of natural gas pipelines

A.J. de M. Brito, C.A.V. Cavalcante, R.J.P. Ferreira & A.T. de Almeida
Federal University of Pernambuco, Recife, Pernambuco, Brazil

ABSTRACT: A multicriteria decision model for risk assessment and for risk ranking of sections of natural
gas pipelines is proposed. The model is based on Multi-Attribute Utility Theory (MAUT). The reasons for a
multicriteria approach for risk assessment are discussed. The model takes into account the assessment of human,
environmental and financial dimensions of the impacts of pipeline accidents. Therefore, three dimensions of
impact and the need to translate decision-makers preferences into risk management decisions are highlighted.
The proposed model approaches these factors by using a multi-attribute utility function which is combined with
probability distributions for consequences and with prior probabilities of accident scenarios in order to result
in a multicriteria risk measurement. Pipeline hazard scenarios and risk assessment on natural gas pipelines are
discussed. To help natural gas companies to prioritize critical sections of pipeline, the multi-attribute approach
also allows sections of pipeline to be ranked into a risk hierarchy. In order to illustrate the use of the model, a
numerical application based on a real case study is presented.

INTRODUCTION

Risk assessment is a widely used tool for identifying and estimating technological risks, such as in
pipeline facilities. It is traditionally based on the survey of probabilities related to human fatalities (Jo &
Ahn 2005). However, accidents recorded around the
world indicate that, due to the complexity of their
consequences, the environmental and financial dimensions of accident impacts must also be considered, in
order to avoid incomplete and inadequate approaches
to pipeline risk assessment. These arise from the
need to reconcile the concerns of society, the State
and gas companies regarding the safe operation of
pipelines.
The decision-making process involving risks in
pipelines is complex. It often involves analyzing conflicting aspects, usually produces long-term results,
and not rarely is performed against a background of
lack of information. Despite these characteristics, several methodologies used for risk assessment are still
vague, subject to doubtful interpretations, and are not
effective in supporting reliable decisions, such as prioritizing pipeline sections in order that, as appropriate,
they might receive supplementary safety investments
within a limited budget.
According to Papadakis (2000), the limitations
of traditional deterministic techniques deployed for
risk assessment may be overcome by probabilistic
approaches such as risk ranking, which combines
probabilities and consequences in a single measure
of risk. This has been shown to be more adequate for

setting up replacement policies and other priority level

problems (Papadakis 2000).
This paper presents a multicriteria decision model
for risk assessment and for ranking pipeline sections
into a risk priority hierarchy. Related research can
be found in Brito & Almeida (2008). This model
was conceived in order to support the planning of
actions for reducing the risks of accident to people,
to the environment and to the financial health of gas
companies and industries that consume gas. The risk
ranking of pipeline sections allows most critical segments to be prioritized under resource constraints.
This risk approach is developed within a decisionmaking context, and Multi-Attribute Utility Theory
(Keeney & Raiffa 1976) is used in order to aggregate
the human, environmental and financial dimensions of
consequences due to a natural gas pipeline accident,
and to incorporate the decision-makers preferences
and behavior regarding risk.
2

PROBLEM ANALYSIS

Among the several modes for transporting dangerous substances, such as natural gas, pipelines are one
of the safest options (Papadakis 2000). Their accident frequencies are less then those related to road
or rail haulage. However, even with low probabilities, the consequences of accidents arising from a
natural gas leakage cannot be neglected. These consequences are often associated with a wide-ranging set of
monetary and non-monetary impacts. By taking into

1165

consideration the human, ecological and economical

aspects involved when a pipeline accident happens,
measures are required to effectively quantify and
mitigate the risks.
In order to ensure an adequate approach to pipeline
risk assessment, uncertainties related to accident scenarios and the multiple dimensions of their consequences must be considered.

2.1.3 3 : CVCE
When gaseous mixtures are accumulated in confined
spaces, and undergo delayed ignition, a scenario called
Confined Vapor Cloud Explosion (CVCE) happens.
This is very dangerous to buildings within an area
where the gas is not out of its flammability limits, due
to the possibility of the confined vapor cloud exploding inside these buildings or in underground spaces
(Sklavounos & Rigas 2006).

2.1

2.1.4 4 : Flash fire

For very poor or very rich natural gas mixtures which
still remain within the flammability limits, a phenomenon known as flash fire may occur. This happens
when the gas-air mixture gas suffers a delayed unconfined ignition, and the resulting fires usually spread at
low speeds and with negligible increases of pressure.

Pipeline hazard scenarios

The evolution of events, from a natural gas release

to accident scenarios becoming concrete, is related to
the following factors (Jo & Ahn 2002, Sklavounos &
Rigas 2006): the failure mode of the pipeline, the time
gap between gas leakage and a possible ignition of
the resulting gas cloud, and the degree of confinement
of the surrounding space. The combination of these
factors may be schematized into five different accident scenarios (here represented by ), similarly for
ruptures or punctures in the pipeline (Sklavounos &
Rigas 2006):
2.1.1 1 : Deflagration/detonation
When a leakage of gas occurs, if the released gas is
sufficiently mixed with air, and it quickly finds an
ignition source within a confined space, combustion
starts quickly with a medium increase in pressure. This
is called deflagration, and it may occur when the mixture of air and gas is inside gas flammability limits.
However, if considerable confinement exists and the
proportion of oxygen inside the gas cloud is close to
the Zero Oxygen Balance, the propagation speed of the
fires increases swiftly, reaching sonic proportions and
developing a strong overpressure with shock waves.
This is called detonation.
2.1.2 2 : Fireball/Jet Fire
In the case of ruptures as well as of punctures, when
a large mass of gas is released quickly and finds
an ignition source immediately, the gas cloud that is
mixed with oxygen ignites, but still in a limited way.
While the core of the gas cloud remains with concentrations above the upper flammability limit (UFL),
a combustion process begins on the external layer of
the cloud. Under forces of buoyancy, the gas cloud
starts to become more spherical and to rise, and
assumes a fire ball form. As the fire ball rises, the
gas-oxygen mixture is increased and the mass of gas
within flammability limits is enlarged, thus sustaining the fires (Sklavounos & Rigas 2006). However,
according to Jo & Ahn (2002), fire balls are quickly
dispersed, and a sustained jet of flaming gas, called jet
fire, is established in an almost stationary state.

2.1.5 5 : Gas dispersion

Finally, when there is no immediate or delayed ignition, an accumulation or a dispersion of the vapor
cloud may happen, depending on the confinement conditions of the surrounding spaces. If the gas is trapped,
it may not ignite until it falls between its flammability
limits, in which case a CVCE scenario may occur. On
the other hand, if the gas is dispersed, it usually poses
no significant hazard, although it may injure people
close to the leakage point due to the heat of gas at high
temperatures.
2.2 The multiple dimensions of consequences
The impacts of natural gas pipeline accidents may have
several dimensions: damage is usually not limited to
fatalities and human injuries, but may involve environmental damage provoked by fires and huge sums
of money lost due to supply and revenue interruptions.
The process of pipeline risk assessment is based
fundamentally on the study of probabilities for different numbers of human fatalities (Jo & Ahn 2005), these
certainly being the most important aspects. However,
this traditional approach can obscure the identification of pipeline segments that are under considerable
risk regarding environmental or economic factors, and
which may also require substantial attention. Moreover, companies are influenced by competition, by
rigid environmental legislation being applied, and
by the general publics awareness of pipeline safety
and accidents to them. Therefore, gas companies
need to concomitantly pay regard to their competitive goals, and the requirements for human safety
and environmental preservation demanded by society
and the State. These factors demand a risk approach,
which includes the human, environmental and financial dimensions of the possible accident impacts, to be
observed.

1166

Since risk assessment is also a subjective process,

it is also necessary to consider, besides the multiple
dimensions of the impacts of accidents, the decision makers perceptions and behavior regarding risk.
This process can be undertaken within a multicriteria approach, by means of Multi-attribute Utility
TheoryMAUT(Keeney & Raiffa 1976).
3

PROBLEM MODELLING WITH MAUT

Multi-attribute Utility Theory is a consistent theory for

tacking multiple-objective problems within the Decision Analysis framework (Keeney & Raiffa 1976).
It has an axiomatic base that, through a structured
process of eliciting preferences, allows the possible
accident consequences to be assessed by endeavoring
to incorporate the decision-makers behavior regarding
risk (Keeney & Raiffa 1976).
As for the problem of risk assessment, MAUT
allows the decision-maker to aggregate several types
of consequences which may arise from a pipeline accident, which helps subsequently to minimize risks.
These consequences, here represented by a vector p,
are taken as the attributes of the problem. The decisionmakers preferences and behavior regarding risk are
modeled in order to obtain a multi-attribute utility
function U (p), which aggregates the utility function
for each attribute.
In order to assess the risk related to each pipeline
section si , the multi-attribute utility function U (p) is
combined with a consequence function P(p| , si ), represented by its probability density function f (p| , si ).
The loss for each pair of scenario and pipeline section
is obtained by applying Equation 1 below:

L(, si ) = P(p|, si )U (p)dp
(1)
p

According to Decision Analysis (Raiffa 1970, Berger

1985), the risk related to each pipeline section is calculated as the expected loss by taking into consideration
all accident scenarios, as shown in Equation 2 below:


r(si ) =
( ) P(p| , si ) u(p) dp
(2)

where i ( ) are the probabilities of the accident scenarios and the operational normality scenario for a
given pipeline section si .
4

THE MULTICRITERIA MODEL

This section presents a decision model which is put

forward to help gas companies undertake a pipeline

risk assessment within a multicriteria point-of-view.

This model also allows the ranking of sections of a
pipeline into a hierarchy of multidimensional risks.
So, if budget, time and labor restrictions come together
such that they do not allow measures for risk reduction to be undertaken on all sections of pipeline, then
a risk ranking of pipeline sections may help decisionmakers to prioritize a set of those sections that present
highest potential damage when the human, environmental and financial dimensions are simultaneously
taken into account.
In this problem, a decision-maker (DM) must be
identified. The DM is the person who influences the
risk assessment and the ranking process by establishing value judgments that represent stakeholders
preferences and perceptions, as seen by the company,
society and the Government. It is the decision-makers
preferences which will be elicited and which will
guide the construction of utility functions. The role
of decision-maker is generally performed by the manager of the operation, maintenance or safety of the
pipeline.
Next, a discrete set of sections S = {s1 , s2 , . . . , sn }
to be the object of risk assessment and risk ranking is
created. This set is formed for all the pipeline segments
in which a pipeline grid may be theoretically divided.
Such sections must present uniform features regarding
diameter, pressure, age of the pipeline section, soil
characteristics, land use, polymeric coating and other
factors taken as relevant to influence the risks related
to each section, such as the segment failure rate (Jo &
Ahn 2005). These sections are distinguished by their
characteristics, and they sum up to the total length of
the pipeline under study.
In pipeline risk assessment, uncertainties are related
to the hazard scenarios, which have already been discussed in Section 2.1. These scenarios form a set
= {} which represents, in this model, all the states
of nature when an accidental leak of natural gas occurs.
The uncertainties on the hazard scenarios are modeled by means of prior probability distributions, here
represented by (). These probabilities are obtained
based on historical data of accidents and leakages
(Papadakis 1999), through the elicitation of specialists prior knowledge (Raiffa 1970), or by exploiting
these two sources of information.
Given the different hazard scenarios discussed in
section 2.1, for rupture or puncture failure modes
(which sum up to ten scenarios), object exposure
analyses and vulnerability analyses are used for each
hazard scenario and pipeline section in order to obtain
a set of possible consequences or payoffs. These consequences are due to exposure to heat and energy,
to flames or explosions, and to supply interruptions
which may affect people, third party properties,
industries, the pipeline structure, areas of vegetation
etc. Thus, the payoff set is modeled as to how it

1167

ranges within three different dimensions or criteria for

analysis:
The human dimension, H : This is related to the
number of persons exposed, at least, to second
degree burns. This is a conservative but appropriate
measurement when dealing with impacts on human
beings, since any type of physical harm must be
avoided.
The environmental dimension, E: These consequences concern the vegetation area (in m2 )
destroyed due to the combustion of natural gas after
a leakage. Although natural gas usually quickly
disperses in unconfined spaces, if some mass of
released gas finds an ignition source, the environmental consequences may involve fires and vegetation being scorched. This also may affect the
integrity of animal species in a given area, and
like human consequences, it cannot be expressed
in monetary indices.
The financial dimension, F: This comprises the
sum of financial expenses due to operational losses
caused by a natural gas pipeline accident. These
expenses may involve: refunds to customers, losses
in revenues from supply interruptions, expenses on
equipment replacement etc.
Thus, the vector p = (h, e, f ) is concerned with the
trio of human, environmental and financial outcomes.
Next, the elicitation of the multi-attribute utility
function U (h, e, f ), is performed. This model assumes
that the utility function U (h, e, f ) is additive. This
is often found in practice (Vincke 1992), including in maintenance problems (Almeida & Bohoris
1996). Thus, in order to obtain U (h, e, f ), the onedimensional utility functions U (H ), U (E) and U (F)
are combined, as shown in Equation 3:
U (h, e, f ) = k1 U (h) + k2 U (e) + k3 U ( f )

values will be influenced by the physical and geographical particularities related to each section of
pipeline. These assumptions allow the probabilities
P(h| , si ), P(e| , si ), and P( f |, si ), to be separately
estimated by means of the probability density functions f (h|, si ), f (e|, si ) and f ( f |, si ), respectively.
The above probabilities can be obtained by means
of eliciting experts knowledge or using results of
object exposure analyses. These probability functions
may have several shapes, which depend on the mathematical models and on the simulation computer tools
adopted.
As presented in Equation 1, for the three-dimension
set of consequences, the utility function U (h, e, f )
is combined to the probability density function
f (h, e, f | , si ). Equation 1 becomes:
  
L( , si ) =

f (h, e, f |, si ) U (h, e, f )df de dh

H E F

(4)

As previously stated, additive multi-attribute utility function and the independence of P(h| , si ),
P(e| , si ), and P( f |, si ) have been assumed (Brito &
Almeida 2008). Thus, Equation 4 becomes:


L(, si ) = f (h|, si ) k1 U (h)dh
H


f (e|, si ) k2 U (e)de

+
E


+

f ( f |, si ) k3 U ( f )df

(5)

(3)

where k1 + k2 + k3 = 1. These are scaling constants estimated by means of elicitation procedures,

also applied for obtaining the one-dimensional utility
functions U (h), U (e) and U ( f ). In these procedures,
the DM answers questions to give information on his
preferences between deterministic consequences and
lotteries with specific probabilities between best case
and worst case values (Keeney & Raiffa 1976).
Uncertainties in this problem are not only related to
, but also to H , E and F, the values which the human,
environmental and financial sets of consequences will
assume. These uncertainties are considered by means
of a probability distribution over the deterministic
consequences, P(h, e, f |, si ), where si is any pipeline
section under analysis.
This model makes the assumptions that the
consequences H , E and F occur randomly and independently, and with no significant correlation. Their

According to Equation 2, the risk for each si is

then calculated as the expected loss under all accident
scenarios , as shown in Equation 6:




r(si ) =
i () f (h|, si ) k1 U (h)dh

f (e|, si ) k2 U (e)de

+
E


+

f ( f |, si ) k3 U ( f )df

+ (1) i (N )

(6)

Where 1 is the loss related to the scenario N

of pipeline operational normality, and (N ) is equal

1168

to 1 (failure ), the probability of not happening

a failure. The risk scale within this multicriteria
approach for risk assessment should not be misinterpreted (Raiifa 1970, Berger 1985). Since utility values
range from 0 (least preferred) to 1 (most preferred),
then loss values and risk values will be on a scale from
1 (lowest risk) to 0 (highest risk) due to the negative
signs in Equations 5 and 6.
After Equation 6 is applied for si , (i = 1, 2, . . . , n),
a multidimensional risk value is obtained for each
section of pipeline, by taking into consideration all
accident scenarios and the human, environmental and
financial consequences which may arise from a gas
release accident. The segments of pipeline can now
be placed in a decreasing ranking of risk, so that the
most critical pipeline sections can be prioritized, in a
context of financial or technical restrictions, in order
to reduce the risks related to this pipeline.
5

AN APPLICATION OF THIS MODEL

This section presents an illustration of the model

proposed for risk assessment and for risk ranking
of natural gas pipelines. This illustration concerns a
numerical application based on real data from a natural gas company. The numbers in this application are
not real, but they have been realistically changed.
This company used to apply the usual safety and
maintenance procedures to its pipeline grid, and also
used some qualitative techniques, such as HAZOP and
FMEA (Modarres et al. 1999) for safety management.
In parallel to the use of these techniques, the companys directorate decided to try more formal methods
in order to obtain a more confident decision-making
process regarding risk management.
The company wished to apply this risk assessment model to its pipeline and to arrange the pipeline
segments into a risk ranking in order to guide the allocation of resources for installing supplementary monitoring devices in the most critical pipeline sections.
This company was also interested in the multicriteria approach to prioritize sections of pipeline by
taking into consideration the human, environmental
and financial consequences that may arise from a
accidental gas leakage in a given pipeline section.
The safety manager was the member of staff
responsible for conducting this application along with
academic consultants. He was designated by this company as the decision-maker for the application of
this model. Some preliminary meetings for conceptual explanations and discussions were scheduled by
the consultants in order to explain this model and
the main concepts of MAUT. The manager was supported by technical subordinates that jointly divided
the pipeline grid into 10 segments, which differed from
each other regarding factors such as age, pressure,

third-party interference, soil type, external activities

and demographic concentration on the surface surrounding each pipeline segment. Polymeric coating
was a factor considered as reasonably uniform all over
the pipeline and was not taken into account for the
pipeline segmentation.
For the companys own pipeline grid, sufficient historical data on hazard scenarios were not available. The
study group decided to adopt the ten hazard scenarios
presented, although the impacts related to the flash
fire scenarios (for rupture and puncture) were taken as
negligible. In order to obtain i (), the accident scenario probabilities for each pipeline section, values
presented in a EGIG report (EGIG 2005) and in other
publications (Bartenev et al. 1996, Dziubnski et al.
2006) were used as references. EGIG data was used
because it provides conservative estimates when compared to other databases and also considers distinctions
for values among rupture and puncture failure modes.
A basic frequency for occurrence of a gas leakage of
0.00041/km.year (EGIG 2005) has been taken as one
of the references by the technical group performing
this application, so they could estimate prior probability () based on the frequency of occurrence and
taking into account the pipeline length. Once suitable
data were not available, this information was used
to drive the experts in order to estimate the probability taking in account some specific aspects from
the pipeline. The estimates have adjustments considering some particularities presented on each section
such as length, age, land use and soil type, third party
activities, etc. Thus, a set of hazard scenario probabilities was obtained for the rupture and puncture
failure modes, as presented in Table 1 and Table 2,
respectively.
The company has estimated that the probabilities
for the flash fire scenarios ((R4 ) or (P4 )) were
negligible for every section of pipeline as shown in

Table 1. Prior probabilities for hazard scenarios, from a

pipeline rupture.
Rupture failure mode (104 )
(failure )
Section (104 ) (R1 ) (R2 ) (R3 ) (R4 ) (R5 )
N1
N2
N3
N4
E1
E2
W1
S1
S2
S3

1169

3.50
4.50
3.90
3.50
2.50
5.00
3.60
4.30
7.80
6.20

0.01
0.00
0.01
0.00
0.01
0.01
0.00
0.00
0.04
0.03

0.17
0.04
0.22
0.28
0.28
0.14
0.05
0.09
0.67
0.53

0.10
0.04
0.09
0.11
0.07
0.06
0.02
0.05
0.28
0.22

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

0.95
0.36
0.84
1.01
0.63
0.54
0.29
0.49
2.53
2.01

Table 2. Prior probabilities for hazard scenarios, from a

pipeline puncture.
Puncture failure mode (104 )
Section
N1
N2
N3
N4
E1
E2
W1
S1
S2
S3

(P1 )

(P2 )

(P3 )

(P4 )

(P5 )

0.01
0.00
0.02
0.00
0.00
0.03
0.02
0.01
0.03
0.03

0.26
0.20
0.31
0.29
0.37
0.56
0.37
0.43
0.61
0.49

0.30
0.57
0.36
0.27
0.22
0.54
0.03
0.48
0.54
0.43

0.00
0.01
0.00
0.00
0.00
0.01
0.00
0.00
0.01
0.00

1.70
3.27
2.04
1.54
0.90
3.10
2.82
2.73
3.10
2.46

Tables 1 and 2. In order to perform an object exposure analysis, several simulations on computer tools
were performed for each scenario and section of
pipeline si .
The estimation of probability density function
demanded intensive and focused efforts. Some discussions were held with the company support group and
the consultants on how to model f (h|, si ), f (e| , si )
and f ( f | , si ). Based on the simulation results, these
consequence probabilities were adjusted to some wellknown families of probability density functions, such
as Lognormal, Gamma and Exponential. The parameters of these functions were changed from section to
section based on technical knowledge of the particularities related to each section.
In order to obtain a multi-attribute utility function
U (h, e, f ), consultants put several questions for the
DM to answer, based on a structured elicitation protocol (as presented in Keeney & Raiffa 1976). The
decision-maker was asked to make choices between
deterministic consequences and lotteries involving
these consequences. This is a process which allows
some utility values to be obtained for the human, environmental and financial dimensions of consequences.
These values were plotted and used for exponential
curve regressions, which have been shown to represent satisfactorily the decision-makers preferences.
From this elicitation procedure, the following shape
parameters were obtained for the exponential utility
functions:
For U (h) : h = 0.12(R 2 = 0.90);
For U (e) : e = 0.0017(R 2 = 0.88);
For U ( f ) : f = 3.5 107 (R 2 = 0.92).
The scaling constants k1 , k2 and k3 presented in
Equation 3 were also obtained based on similar elicitation procedures (Raiffa 1970, Keeney & Raiffa 1976).
These procedures were also formed by probabilistic

choices of lotteries involving several values for the trio

of human, environmental and financial consequences.
These trios were compared to combinations of best
and worst one-dimensional values, which allowed the
following scaling constants to be obtained:
For U (h): k1 = 0.60;
For U (e): k2 = 0.10;
For U ( f ): k2 = 0.30.
Thus, in order to obtain the set of losses according to
Equation 5, the study group integrated the probability
density functions f (h|, si ), f (e|, si ) and f ( f | , si )
with the one-dimension utility functions U (h), U (e)
and U ( f ). The results are presented in Table 3 and
Table 4, also for rupture and puncture failure modes,
respectively.
As previously stated and in order to simplify the calculations, the company assumed that the losses related
to the flash fire scenarios were reasonably negligible
for all dimensions of consequences (Jo & Ahn 2002).
This corresponds to a loss value of 1, according to
the explanations on this risk scale in Section 4.
On the results of Tables 3 and 4, the consultants applied a linear scale transformation L (, si ) =
1000 L( , si ) + 1000, in order to facilitate the visualization and the comparison of these values. This
procedure preserves the relative order of losses, and

Table 3.

Losses L( , si ) for the rupture failure mode.

Section L(R1, si ) L(R2, si ) L(R3, si ) L(R4, si ) L(R5, si )

N1
N2
N3
N4
E1
E2
W1
S1
S2
S3

Table 4.

0.455
0.498
0.277
0.280
0.687
0.469
0.099
0.336
0.301
0.274

0.394
0.349
0.198
0.216
0.567
0.420
0.077
0.298
0.247
0.191

0.605
0.409
0.420
0.377
0.786
0.638
0.253
0.488
0.590
0.346

1
1
1
1
1
1
1
1
1
1

0.950
0.857
0.902
0.840
0.967
0.873
0.829
0.849
0.912
0.848

Losses L( , si ) for the puncture failure mode.

Section L(P1, si ) L(P2, si ) L(P3, si ) L(P4, si ) L(P5, si )

N1
N2
N3
N4
E1
E2
W1
S1
S2
S3

1170

0.512
0.581
0.390
0.367
0.740
0.531
0.158
0.415
0.385
0.386

0.472
0.448
0.295
0.290
0.681
0.438
0.128
0.393
0.319
0.269

0.714
0.492
0.532
0.554
0.872
0.718
0.382
0.619
0.734
0.582

1
1
1
1
1
1
1
1
1
1

0.971
0.945
0.946
0.894
0.977
0.906
0.884
0.871
0.938
0.875

Table 5. Results of the multi-attribute risk

assessment and risk ranking of the pipeline
sections under study.
Section

r(si )

S3
S2
S1
E2
N4
N3
W1
N2
N1
E1

0,176
0,164
0,097
0,095
0,094
0,083
0,079
0,069
0,047
0,033

it may be applied to Neumann and Morgenstern utility

(or loss) functions in the context of Utility Theory, which are within an interval scale (Raiffa 1970,
Keeney & Raiffa 1976, Berger 1985).
Finally, in order to obtain a multi-attribute risk value
for each pipeline section, the study group multiplied
the probabilities from Tables 1 and 2 to the losses from
Tables 3 and 4 (after the linear transformation) respectively, and summed them up. Based on the risk value
obtained for each si , a ranking of pipeline sections was
obtained as presented in Table 5.
The results of Table 5 were presented to the
safety manager. Sections S3, S2 and S1 were recommended as the first three segments to be prioritized so they would receive the supplementary
monitoring devices available. These segments have
been shown to be the most critical ones within a
more complete and generalized perspective of risk,
that is related to the multiple dimensions of consequences which may arise from a natural gas pipeline
accident.
As the consultants stressed to the safety manager,
unless any condition related to the model parameters
changes, this priority ranking may still guide further allocations of resources to the pipeline sections.
Depending on the amount of resources to be invested,
the number of sections to be benefited may increase
until all the resources are employed to improve safety
on the companys pipeline grid.

CONCLUSIONS

This paper has approached a multicriteria perspective for risk assessment of natural gas pipelines. It
has discussed that risk management involves complex decision-making processes which must reconcile
concerns from different stakeholders: society, the
State and gas companies. Therefore, it tackles how to

involve human, environmental and financial factors

in risk assessment and risk management processes
undertaken by gas companies.
However, traditional methodologies for risk assessment present an incomplete framework regarding the
need to take into consideration the multi-dimensional
consequences that an accident in natural gas pipelines
may bring about. In order to overcome these limitations, Multi-attribute Utility Theory was employed
to model the pipeline risk assessment process and a
pipeline ranking process under a multi-dimensional
perspective, which not only sought to translate the
uncertainties of the scenarios, but also allowed aggregation of the decision makers aversion to risk regarding human, environmental and financial dimensions
of consequences which may arise from pipeline accidents. With MAUT, these consequences were regarded
as the attributes of the problem, and a multi-attribute
utility function U (h, e, f ) was obtained in order to
aggregate the utility, for the decision-maker, of the
consequences on each attribute.
The results of the application on this paper have
shown that this model was effective in assessing risks
and in ranking sections of pipeline under a multiattribute perspective. They have also shown that such
a formal model provides precise and quantitative support for a more confident decision-making process
regarding risk management, but it demands intensive
and focused efforts to assess the models parameters
in order to ensure that consistent results are obtained.
The application of elicitation protocols to obtain
utility functions and scaling constants has been shown
to be successful when the decision-maker makes himself available, is patient and has a solid background.
Such background and also experience must be present
in the consultants when applying this risk assessment
model in order that they can point out inconsistencies
and guide this process efficiently.

REFERENCES
Almeida A.T. de, Bohoris, G.A. 1996. Decision theory in
the maintenance strategy of a standby system with gamma
distribution repair time. IEEE Transactions on Reliability.
45 (2): 216219.
Bartenev, A.M., Gelfand, B.E., Makhviladze, G.M. &
Roberts, J.P. 1996. Statistical analysis of accidents on the
Middle Asia-Centre gas pipelines. Journal of Hazardous
Materials. 46: 5769.
Berger J.O. 1985. Statistical Decision Theory and Bayesian
Analysis. New York: Springer.
Brito, A.J. & Almeida, A.T. de. 2008. Multi-attribute
risk assessment for risk ranking of natural gas
pipelines. Reliability Engineering and System Safety. doi
10.1016/j.ress.2008.02.014.
Dziubnski, M., Fratczak, M. & Markowski, A.S. 2006.
Aspects of risk analysis associated with major failures of

1171

fuel pipelines. Journal of Loss Prevention in the Process

Industries. 19: 399408.
EGIG. 2005. 6th EGIG-report 19702004. Gas Pipeline Incidents, Report EGIG 05.R.0002, December 2005. Groningen: EGIG. http://www.egig.nl/downloads/6th ReportEGIG.pdf.
Jo, Y.D. & Ahn, B.J. 2002. Analysis of hazard areas associated
with high-pressure natural-gas pipelines. Journal of Loss
Prevention in the Process Industries. 15: 179188.
Jo, Y.D. & Ahn, B.J. 2005. A method of quantitative risk
assessment for transmission pipeline carrying natural gas.
Journal of Hazardous Materials. A123: 112.
Keeney, R.L. & Raiffa, H. 1976. Decision with multiple
objectives: preferences and value tradeoffs. New York:
Wiley.
Modarres, M., Kaminskiy, M. & Krivtsov, V. 1999. Reliability Engineering and Risk Analysis. New York: Marcel
Dekker Inc.

Papadakis, G.A. 2000. Assessment of requirements on safety

management systems in EU regulations for the control of
major hazard pipelines. Journal of Hazardous Materials.
78: 6389.
Papadakis, G.A., Porter, S. & Wettig, J. 1999. EU initiative on the control of major accident hazards arising
from pipelines. Journal of Loss Prevention in the Process
Industries. 12: 8590
Raiffa, H. Decision Analysis. 1970. Reading, MA: AddisonWesley.
Sklavounos, S. & Rigas, F. 2006. Estimation of safety distances in the vicinity of fuel gas pipelines. Journal of Loss
Prevention in the Process Industries. 19: 2431.
Vincke P. Multicriteria Decision-Aid. 1992. Bruxelles: John
Wiley & Sons.

1172

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

New insight into PFDavg and PFH

F. Innal & Y. Dutuit
IMS-LAPS/Universit Bordeaux 1, Talence, France

A. Rauzy
IML/CNRS, Marseille, France

J.-P. Signoret
Total SA, Pau, France

ABSTRACT: The aim of this paper is to give a new insight into some fundamental concepts of the IEC 61508
standard. In a first part, we examine low and high or continuous demand modes of operation. We study how
to determine the accident frequency in the case the system under study is made of one element under control
and its associated safety instrumented system. In a second part, we study the relationship between the average
probabilities of failure on demand and the risk reduction factor. Finally, we consider the probability of failure
per hour of a safety instrumented system. We propose different ways to compute it.

INTRODUCTION

Today IEC 61508 (IEC 61508 2000) is one of main

standards for specification, design and operation
for Safety Instrumented Systems (SIS, for short).
However, several of its fundamental concepts are
ambiguous. This article aims to clarify up the situation.
In a first part, we study the two modes of operation
of SIS. We introduce a new criterion to do so. Then, we
propose a generic model involving both low and high
or continuous demand modes of operation. We use this
model to illustrate the relationship between the accident
frequencyofaproductionsystemcontrolledbyaSISand
the average probability of failure on demand (PFDavg )
or probability of failure per hour (PFHavg ) of the SIS.
In the second part, we show that the whole risk
reduction factor (RFF) obtained with several layers of
protection is not equal to the product of the RFF for
individual layer.
The third part is devoted to the notion of PFH. We
define its nature in a precise way. Then, we describe
three methods to compute it.
A short conclusion ends this work.
2

Low demand mode: where the frequency of

demands for operation made on a safety-related system is no greater than one per year and no greater
than twice the proof-test frequency.
High demand mode: where the frequency of
demands for operation made on a safety-related system is greater than one per year or greater than twice
the proof-test frequency.
Both definitions refer to a threshold of the frequency of demand and to the ratio between demand
frequency and proof-test frequency. The values considered for these two quantities seem somewhat arbitrary and, at least, neither explained nor justified. We
have shown (Dutuit et al. 2008) that the factor 2 for the
ratio is valid, under some assumptions, in the particular
case of 1oo1 architecture. But it is not suitable for other
architectures. Morover, a SIS is usually made up of
three types of components (sensors / logic solver / actuators). In general, the proof-test frequencies of these
components are different. So the above definitions left
much too much room for interpretation.
To solve this problem, we propose hereafter a new
criterion to discriminate the two modes of operation.

THE TWO MODES OF OPERATION OF SIS

2.2 A discriminating criterion

2.1

Official definitions

Let us start by recalling the official definition of the

two operating modes of SIS:

Our demonstration is based on a generic model of a SIS

incorporating the demand of the controlled process.
This model is depicted Figure 1.

1173

If d >> , then we are in a continuous mode of

operation. In this case, Equation 2 reduces to:

OK

OK

KO

STOP

OK
1

OK
3

2
r

wacc = wSIS

The above equality indicates that an accident

occurs as soon as the SIS fails. Morover, by considering d as an approximation of the failure
frequency wd , the condition (d >> ) can be
rewritten as follows.

MDTSIS

KO
ACCIDENT
4

wd >> =

Figure 1. A generic Markovian model incorporating process demand.

Each state of the model is identified by three labels.

The upper label indicates the state of the SIS: OK
stands for working state and KO stands for failed
state. The median label indicates the state the Element Under Control (EUC). The nominal state is
denoted by OK. The shutdown state is denoted
by STOP. Finally, the accident state is denoted by
ACCIDENT. The lower label is just an index of the
state.
d denotes the demand rate. D denotes the dangerous failure rate (D = DD + DU ).
r is the inverse of the mean time required to restart
the EUC after its shutdown. = 1/MDTSIS .
The value allocated to R is very large to mask
the time spent after an accident, see (Dutuit
et al. 2008) for further explanation. It refers to
the so-called renewal transition (Sato & Black,
pers. comm.).
It is shown in (Dutuit et al. 2008) that the average
behaviour, over an important observation period, of
a given system can be approximated by its asymptotic behaviour. This makes it possible to compute the average value of the accident frequency
named wacc .
wacc = p3 () d = p1 ()

D
d
d +

(1)

p1 () D is the average value wSIS of the failure

frequency of the SIS. Therefore, the following equality
holds.
wacc =

d
wSIS
(d + )

(3)

(2)

Two extreme configurations can be deduced from

Equation (2):

1
MDTSIS

(4)

If d << , then we are in a low demand mode of

operation. In this case, Equation 1 gives:
wacc = p3 () d PFDavg d
= PFDavg wd

(5)

The above considerations show the expected criterion suitable to discriminate low demand mode of
operation from high or continuous demand mode is
obtained by comparing the product wd MDTSIS to
the unity as follows.
2.3 A generic model for the accident frequency
Low and high demand modes of operation are considered separately in IEC 61508 standard. This point
of view is a simplistic, because any SIS can be successively in both of them. As an illustration, consider
a shutdown valve, which has to close when an overpressure occurs in an upstream part of a pipe. The
considered SIS is working in accordance with low
demand mode until the overpressure arises but, after
that and until the overpressure is exhausted, the SIS
works under continuous mode of operation to preserve
the pipe against this overpressure.
This general configuration must be taken into
account when EUC is not put instantaneously into a
safe state after the demand. This corresponds to the
state 2 of the model depicted in Figure 2, which extends
the previous one (Fig. 1).
The meaning of all the states and parameters used
in this new model remains the same as in the previous
one. Here d stands for the inverse of the mean latency
time of the demand, and state 2 is the demand state.
Table 1.

A discriminatory criterion.

Modes of operation

Conditions

Low demand
High demand
Continuous demand

wd MDTSIS < 1
wd MDTSIS >= 1
wd MDTSIS >> 1

1174

OK
DEMAND
2

OK
STOP
4

wacc PFDSIS wdemand +PFHSIS p(demand)

OK

KO

OK
1

OK
3
MDTSIS

KO
ACCIDENT
5

Figure 2. An extended generic Markovian model incorporating process demand.

Let us compute the accident frequency by using the

critical working states approach.

wacc =

1
T

T

1
T

It is worth noticing 1) that the above expression

involves the contributions of both low and high or continuous demand modes of operation and, 2) that the
SISs failure modes concerned by these two different
demand modes are also different.
That is to say that, in Equation 10, PFDSIS refers
to failure modes of SIS components, which lead to
the non closure of the shutdown valve, when PFHSIS
refers to failure modes resulting in a spurious opening
of the same valve. In other words, the two products
of the right hand term of Equation 10 constitute two
mutually exclusive scenarii as shown in Figure 2.

PFDAVG VS RRF

w(t)dt

In this section, we consider low demand SIS. For such a

SIS, IEC 61508 and 61511 standards establish the relationship between the PDFavg and the RRF by means
of the following equality.

(p3 (t) d + p2 (t) D ) dt

wacc PFDavg wd

T
0

T
p3 (t)dt +

T
p2 (t)dt

(6)

As before, we can approximate the mean values of

the probabilities by their asymptotic values. Thus, we
get:
wacc d p3 () + D p2 ()

(7)

D
d
d +
D
(d + )
(D + d )

wacc

Protection layer 1

(8)

Protection layer 2
(1- p1) .wd
(safe outcome)

1 p1
wd

(9)

1 p2

p1 .(1 p2) .wd

(safe outcome)

Demand
p1 = PFD1

Morover D / and d /d express respectively the

average value of the PFD of the SIS and the average value of the probability of demand, i.e., PFDSIS
and p(demand), when D and d respectively approximate the PFH of the SIS, PFHSIS , and the demand
frequency, wdemand .
Under the above conditions, Equation 10 can be
rewritten as follows:

(11)

Independent protection layers

Because d << and 1/d << 1/D , we have:

D
d
d +

wd
1
=
= RRF
wacc PFDavg

This equality is correct only when a single layer of

protection is considered. But, what happens if several
layers can be sequentially activated? This latter case is
illustrated Figure 3.
For the sake of the simplicity, assume that the first
layer is a SIS with a 1oo1 architecture and the second one is a relief valve. Assume moreover, that the
accident occurs only if both layers fail. The parameters of interest for respectively the SIS and the relief

By considering that p1 () 1 and expliciting

p2 () and p3 (), it comes:
wacc

(10)

RRF1 = 1/ p1

p1 .wd

New demand p2 = PFD2 p1 .p2.wd= wacc

RRF2 = 1/ p2

Accident
frequency

Figure 3. Risk reduction performed by two independent

protection layers.

1175

valve are (DU 1 , DD1 , MTTR, T1 ) and (DU 2 , MTTR,

T2 = 2T1 ), where Ti (i = 1, 2) is the proof-test interval.
We shall apply first the approach of the standard to
compute the global RRF. Then, we shall propose the
right approach.
3.1

Standard approach

In accordance with the layout of Figure 3, we can write:

1
1
= PFD1 PFD2
=
RRFstd
(RRF1 RRF2 )




T1
+ MTTR + DD1 MTTR
= DU 1
2


T2
+ MTTR
(12)
DU 2
2
where the index std stands for standard.
A rather good approximation of Equation 12 is
given hereafter:

 

1
T1
T2
= DU 1
DU 2
RRFstd
2
2

3.3

Conclusion

The comparison between Equation 14 and Equation 16

shows the standard approach leads to an (slightly, in
our case) optimistic value for the risk reduction factor.
This can be not acceptable from a safety point of view.
The observed difference confirms the average value
of a product of functions differs from the product of
their own average values.
4
4.1

THE TRUE NATURE OF THE PFH

Preliminary remarks

If PFDavg of a SIS is now widely recognized as its

average unavailability, the nature of PFH is still clearly
defined neither in IEC 61508 standard, nor in the literature. To highlight this lack of definition, the reader can
check that the part 4 of the standard, entirely devoted
to definitions and abbreviations, gives no definition of
the probability of failure per hour! The note 4 of paragraph 7.6.2.9 of IEC 61508-1 is the only place where
a so-to-say definition is given:

(13)

The parameter in Table 3 for high demand or

continuous mode of operation, probability of a
dangerous failure per hour, is sometimes refered
as the frequency of dangerous failures, or dangerous failure rate, in units of dangerous failures
per hour.

The above approach is erroneous because the whole

PFDavg of the safety-related system made up of two
layers is obviously not equal to the simple product
of their individual PFDavg . The exact value must be
computed as follows:

Unfortunately, such a sentence is very ambiguous,

because a (failure) frequency differs from a (failure)
rate.
The note 5 of the same paragraph does not give a
definition of PFH, but a way to compute it:

=
3.2

DU 1 DU 2
2

T12

Exact approach

PFDavg =

1
T2

T2
PFD1 (t) PFD2 (t) dt

. . . Determine the required probability of failure of the safety function during the mission time
and divide this by the mission time, to give a
required probability of failure per hour.

(14)

By approximating PFD1 (t) and PFD2 (t) by respectively DU 1 .t and DU 1 .t over a proof-test period, we
have:
PFDavg =

DU 1 DU 2
2T1

T
1
2T1
t 2 dt + (t T1 ) tdt
0

4.2 Is PFH an average failure rate?

(15)

From note 5 mentioned above, it appears that PFH can

be expressed as follows.

T1

PFH =

By resolving Equation 15, we get:

PFDavg =

In the sequel, we examine these two pseudodefinitions and we propose further investigation to
converge onto an acceptable definition.

7
1
DU 1 DU 2 T12 =
12
RRFex

where the index ex stands for exact.

(16)

probability of failure of a SIF during T

(17)
T

What does mean the expression given in the numerator of this ratio? A reasonable interpretation is the
cumulative distribution function F(t), in other words
the unreliability, of the SIS.

1176

PFH =

F(T )
T

(18)

At first glance the latter expression is surprising,

because it tends to 0 as T goes to infinity, due to the
bounded nature of F(T ). Nevertheless, let us enter into
more details.


T
1 exp (t)dt

1 R(T )
=
T


1 exp avg .T
=
T

PFH =

T
(19)

If avg .T << 1, which is presumably the case in

practice, the following fact holds.

where A denotes the event component C was working

at time t = 0 and is working at time t.
w(t) = lim

dt

prob (C fails between t and t + dt/B)

dt
(23)

where B denotes the event component C was working

at time t = 0.
These definitions show that v (t) and w(t) cannot
be identified. Morover, it is well-known (Kumamoto &
Henley 1996) that the following equality holds.

wavg

1
=
T

T
w(t)dt =

W (0, T )
T

(24)

PFH

avg .T
= avg
T

(20)

Unfortunately this interpretation is not correct,

because the instantaneous failure rate (t) becomes
undefined as soon as the SIS fails. Therefore, the computation of avg is meaningless. Another interpretation
is then required.
4.3

Is PFH an average density function?

Equation 18 can be rewritten as follows.

1
F(T )
=
PFH =
T
T

T
f (t) dt = favg

(21)

As previously, the value of favg depends on T . But,

it is not constrained by the condition avg .T << 1. Its
limit is defined when T increases.
Unfortunately this limit is zero, which is quite unrealistic. Once again, another interpretation is needed.
4.4

Is PFH an average failure intensity?

The common problem with the two previous attempts

to define the PFH stands in the fact that they are only
related to the reliability of the SIS. It seems more
relevant to choose another indicator which is able to
take into account the successive failures and repairs
of this system. Two other parameters are suitable for
that: the conditional failure intensity v (t) and the
unconditional failure intensity w(t).
v (t) = lim

dt

prob (C fails between t and t + dt/A)

dt
(22)

where W (0, T ) is the expected number of failures

occuring over the period T . The ratio W (0, T )/T can
be assimilated to the mean time between two consecutive failures, i.e., the so-called MTBF. w(t) can be
identified to a failure frequency. Now,
The PFH is strongly related with the accident frequency into the continuous mode of operation.
Therefore, it should be a failure frequency.
The PFH is measured by favg over the period before
the first failure of the SIS. w(t) coincides with f (t)
on this period.
For all these reasons, we advocate that the right
measure of the PFH of a SIS is its average unconditional failure intensity wavg , or, in other words, its
average failure frequency.

4.5 How to compute wavg ?

The analytical formulae proposed in annex B of IEC
61508-6 make quite restrictive assumptions. They cannot be easily extended beyond these hypotheses. On
the other hand, Fault Tree, Markov Model and Stochastic Petri nets can be to compute the PFH in the general
case.
4.5.1 Fault tree model (FT-model)
From a theoritical point of view, the FT-model can
be used only if the components of the studied SIS
are as independent one another. In this case, fault
tree model makes it possible to compute directly the
unconditional failure intensity wS (t) of the SIS (Dutuit
& Rauzy 2005) on the basis of the so-called critical working states approach (Pages & Gondran 1979)
(Cocozza-Thivent 1997) as follows:

1177

wS (t) =

IB (S, ci ) wi (t)

(25)

AN ILLUSTRATIVE EXAMPLE

where wi (t) is the unconditional failure intensity of the

component i and IB (S, ei ) is its Birnbaum importance
factor (Birnbaum 1969).
PFH is then deduced by applying the first equality
of the Equation 25.
4.5.2 Multi-phases markov model (MPM-model)
In a previous paper (Innal et al. 2007), we showed
multi-phases Markov model is the right model to
handle systems with periodically tested components
(any SIS whose proof-tests are periodically performed
belongs to this category). The critical working states
approach can again be applied on this kind of models
through the following expression:
PFH (t) = wS (t) =

i pi (t)

5.1 Used models

To illustrate the ability of the three above models to
compute the PFH, we apply each of them to a simple 1oo2 architecture. FT, MPM and PN models are
respectively depicted in Figures 46.
FT-model deserves no particular comments.
The multi-phases Markov model requires some
explanations.

1oo2 dangerous
failure
G1

CCF failure

Independant
failures

G2

G3

(26)

ic

where c denotes the set of critical working states and

i is the sum of failure rates attached to all transitions
from the critical working state i to any failed state.
The average value of PFH (t) over the period T is
directly deduced from Equation 26:
1
PFH =
T

T

DU-CCF

Channel 1 failed

Channel 2 failed

e1

e2

G4

G5

DD_failure

DU_failure

DD_failure

DU_failure

e3

e4

e5

e6

Figure 4.

i pi (t) dt

FT-model related to 1oo2 architecture.

iMC

T
1 
i pi (t) dt
=
T iM
C

DD_CCF

4
2(1- D

1 
i CSTi [0, T ]
T iM

DD

DD

1KO (DD)
1OK

(27)

2OK

DU

DD

1KO (DD)
1KO (DU)

1
DD

and finally:
PFHmoy =

2KO
(DD)

DD

DD

i APSi [0, T ]

1KO (DU)
1OK

(28)

2(1-)

DU

iMC

where CSTi [0, T ] and APSi [0, T ] denote respectively

the cumulative sojourn time in the critical working
state i, over the periode T , and the average probability
of sojourn in this state over the same period.

4.5.3 Petri net model (PN-model)

When Petri net approach is used wavg , and then PFH, is
obtained by estimating the expected number of firing
W (0, T ) of transitions leading directly to any place
related to a system failed state, and by dividing it by
the mission time T , according to the second part of the
Equation 24.

Figure 5.

DD

2KO
(DU)

DU

DU

1
p1 (bi )
p2 (bi ) 0

p3 (bi ) 0
p (b ) = 0
4 i
p5 (bi ) 0
0
p6 (bi )

1178

0
1
0
0
0
0

0
1
0
0
0
0

0
0
0
1
0
0

0
0
0
1
0
0

p1 (ei1 )
0
0 p2 (ei1 )

0 p3 (ei1 )

1 p4 (ei1 )

0 p5 (ei1 )
0
p6 (ei1 )

MPM-model related to 1oo2 architecture.

Table 2.

Variables used in the PN-model.

Domain

Name

Definition/Initial Value

Real
Real

LDDIN
LDUIN

Real
Real
Real
Real
Real
Real
Boolean
Boolean
Boolean
Boolean
Boolean
Real

LDDCD
LDDCU
BETA
DC
LAMBDA
BETAD
A_KO
B_KO
CCF_DD
CCF_DU
TEST
1oo2_KO

(LAMBDA 0.5) (1- BETAD) DC

(LAMBDA 0.5) (1- BETAD)
(1-DC)
(LAMBDA 0.5) BETAD DC
(LAMBDA 0.5) BETA (1-DC)
------------------------BETA 0.5
false
false
false
false
false
#1 == 0&#11 == 0

Table 3. Numerical results obtained from holistic models

and from standard.
Failure rates (h1 )

5.0E-06

Approaches
Figure 6.

PN-model related to 1oo2 architecture.

IEC 61508

This model shows that common cause failures

(CCF) must be not considered independently from
individual failures.
The values of the state probabilities at the beginning (bi ) of the period i are computed from the ones
obtained at the end of the period i-1, by means of
the matrix pictured Figure 5.
The PN-model of Figure 6 is rather hard to understand without any explanation. However, an explicit
description of the main features of this kind of model
is out of the scope of this paper. Thus, the only information which can be given concerns the use of boolean
and real variables which greatly improves the ability
of PN to capture any aspect of the behaviour of studied
systems. In our case, the used variables are grouped in
Table 2.

60
90
0

FT-model

DC
(%)

60
90
0

MPM-model

DC
(%)

60
90
0

NP-model

5.2

DC
(%)

DC
(%)

60
90

Numerical results

Some results obtained from the three models and those

given in table B.13 of annex B of IEC 61508-6 are
gathered in Table 3, with the reliability parameters of
interest (T1 = 8760 h, , DC, , D , MTTR = 8 h).
We can observe a good agreement between the four
sets of results. We can also note that the numerical data
got from the standard are not systematically greater

= 2D
= 2D
= 2D
= 2D
= 2D
= 2D

= 10%
= 20%
= 10%
= 20%
= 10%
= 20%

2.90E-7
5.40E-7
1.90E-7
3.70E-7
1.40E-7
2.80E-7

= 2D
= 2D
= 2D
= 2D
= 2D
= 2D

= 10%
= 20%
= 10%
= 20%
= 10%
= 20%

2.93E-7
5.33E-7
1.93E-7
3.65E-7
1.42E-7
2.79E-7

= 2D
= 2D
= 2D
= 2D
= 2D
= 2D

= 10%
= 20%
= 10%
= 20%
= 10%
= 20%

2.93E-7
5.33E-7
1.93E-7
3.65E-7
1.42E-7
2.79E-7

= D
= D
= D
= D
= D
= D

= 10%
= 20%
= 10%
= 20%
= 10%
= 20%

2.97E-7
5.37E-7
1.91E-7
3.66E-7
1.46E-7
2.81E-7

than those obtained from the used holistic models,

conversely to what has been observed for the case of
PFDavg (Innal et al. 2007).

1179

CONCLUSION

This paper summarizes some qualitative and quantitative results from recent advanced work (Innal, in prep.)
on several topics related to IEC 61508 standard. New
insights have been given about some of the main definitions and concepts of this standard, including the
following.
The low demand and high demand or continuous
mode of operation.
The relationship between the whole risk reduction
factor obtained by associating of several layers of
protection and the combination of their individual
PFDavg .
The true nature of the PFH.
We hope our contribution will provide the reader
with a better understanding of part 6 of the standard.
Some important concepts however, such as safe failure
fraction and spurious failures, have not been discussed
here, due to the lack of space.
REFERENCES

Cocozza-Thivent, C. 1997. Processus stochastiques et fiabilit des systmes. Berlin: Springer.

Dutuit, Y. & Rauzy, A. 2005. Approximate estimation of
system reliability via fault trees. Reliab Engng Syst Safety
87 (2): 163172.
Dutuit, Y., Innal, F. & Rauzy, A. 2008. Safety instrumented
systems study according to the IEC 6150-6 standard and
TOTAL methodology. Arboost report.
IEC 61508 2000. Functional safety of electric/electronic/
programmable electronic safety-related systems. Parts
17.
Innal, F., Dutuit, Y. Rauzy, A. & Signoret, J.P. 2007. An
attempt to understand better and apply some recommendation of IEC 61508 standard. In Helge Langseth &
Giacomo Cojazzi (eds), Reliability of Safety-Critical Systems; Proc. Of the 30th ESREDA Seminar, Trondheim,
78 June 2007. Tspra: JRC European Commission.
Innal, F. in prep. Analyse critique de la norme internationale CEI 61508 rgissant les systmes instruments de
scurit. PhD. Bordeaux: Universit Bordeaux-1.
Kumamoto, H. & Henly, E.J. 1996. Probabilistic Risk Assessment and Management for Engineers and Scientists. N.Y:
IEE Press.
Pages, A. & Gondran, M. 1979. Fiabilit des systmes. Paris:
Eyrolles.
Sato, Y. & Black, B. pers. comm. Basics of Safety Integrity
Levels and Risk Frequency (i.e. Hazardous Event Rate).

Birnbaum, Z.W. 1969. On the importance of different components and a multicomponent system. In Korishnaiah P.R.
(ed.), Multivariable analysis II. N.Y: Academic Press.

1180

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

On causes and dependencies of errors in human and organizational

barriers against major accidents
J.E. Vinnem
Preventor AS/University of Stavanger, Norway

ABSTRACT: The BORA project has developed a model based on use of Fault Trees, Event Trees, Influence
Diagrams, Risk Influencing Factors and simplified modelling of dependencies between these factors. The work
has been further developed in the OTS project, which in 2007 has carried out a pilot project in order to test the
approach to assessment of performance of Risk Influencing Factors with effect on human and organisational
barriers. One of the challenges that so far has not been researched extensively, is the influence of common
cause failures in the human and organisational barriers. The topic is well known with respect to the technical
barriers, but is far from well developed for human and organisational barriers. Some case studies are reviewed
in order to illustrate the possible effect of common cause failures on human and organisational barriers, from
incidents on offshore petroleum installations. The dependencies that are thus created are important to integrate
into risk analysis studies relating to human and organisational barriers. The common cause failures in human and
organisational barriers is one of the main topics in a new research project at UiS, with emphasis on modelling of
risk of major hazards relating to organisational, human and technical barriers. Some of the challenges that are
focused on in the project are discussed.

1
1.1

INTRODUCTION
Background

Offshore risk assessment of major hazards is to an

increasing extent focusing on modelling of hazards
which are strongly influenced by operational errors
and where operational barriers are important for risk
minimization. The same applies to analyzing major
hazards for onshore petroleum facilities.
Almost 2/3 of all leaks on offshore installations
in the period 20012005 according to the Risk level
project by Petroleum Safety Authority (PSA; see
Vinnem, 2006), resulted from manual operations and
interventions, as well as shut-down and start-up, confirming what is considered common knowledge; that
incidents and accidents often are caused by failure of
operational barriers.
The data from the PSAs Risk level project apply
in particular to hazards associated with hydrocarbon
leaks. Other hazards, in particular relating to marine
systems, are also strongly dependent on human and
operational aspects. This applies to ballast systems,
dynamic positioning systems and anchor systems, see
for instance Vinnem et al (2003a), Chen & Moan
(2005).
The focus in the future for the Norwegian offshore
industry is on operation of existing installations,

extending the operational life of some of these

installations as well as tying in new subsea installations. This implies that operational safety is receiving
more and more attention, in contrast to design safety.
Authorities are also more focused on further risk
reduction in the operations phase.
At the same time, the trends in the last ten years have
been either constant or slightly increasing, with respect
to several aspects of major hazard risk, for existing production installations in the Norwegian sector,
according to the Risk Level Project by PSA, Vinnem
et al., (2006). The industry may need a new initiative
in order to turn the trends in the right direction. The
industry has formulated ambitious goals for reduction
(50%) of the number of hydrocarbon leaks in the period
20062008. Reduction of the contribution from operational causes will be required, in order to achieve this
reduction.
1.2

BORA project

In a paper presented at ESREL 2003 (Vinnem et al.,

2003b), operational risk assessments were discussed.
It was concluded that there is a clear need for improvement of the analysis of barriers. These aspects form
the outset for an extensive research activity called
the BORA (Barrier and Operational Risk Analysis)
project (Vinnem et al., 2003b).

1181

3. Establish an overview of the factors that influence

the initiating events and the operational barrier performance and define Performance standards (PS).
4. Define performance requirements (PR) for these
PS that are suitable for measurement of the
condition and establish methods for how to measure the status.
5. Establish a system for aggregating the score of each
PR into an overall indicator for operational safety
condition.

OTS/TTS
Performance
standards
(PS)

9 RIFs

22 Safety
barriers

Performance requirements
(PR)
- Main work
processes
- PUFF (Plan,
Do, Check,
Improve)

Functionality
Integrity
Vulnerability
Management

The OTS system is an extension of the TTS verification system, as outlined above. The TTS and OTS
systems may be integrated when fully developed.

Check lists

1.4 Risk OMT project

Assessment activities

Figure 1. Relationship between performance standards,

performance requirements, check lists and assessment
activities.

One of the main aspects of the project is to address

the barrier situation in detail when operational activities are carried out. A list of ten suitably defined
activities and conditions that are associated with
hydrocarbon leak risk was established during the work
with activity indicators (Vinnem et al., 2004b), which
is also used in the BORA project.
The BORA project has been carried out in the
period 2003 through 2006, and was concluded in 2006
with a generalised methodology, based on the initial
methodology formulation as well as the experience
from the case studies. This paper presents the generalised methodology, which is not presented in any
previous papers.
1.3

OTS project

The objective of OTS is to have a system for assessment of the operational safety condition on an offshore
installation/onshore plant, with particular emphasis on
how operational barriers contribute to prevention of
major hazard risk and the effect of human and organisational factors (HOFs) on the barrier performance.
This objective is reached through a development
process with the following targets:
1. Identify and describe human and operational barriers for selected accident scenarios, with major
hazard risk potential.
2. Identify those tasks that are most critical from a risk
point of view, either through initiation of failures
(initiating events) or failures of important barrier
functions.

The objective of the Risk OMT programme is to

provide new knowledge and tools for major hazard risk management for installations and plants,
based on improved understanding of the influence of
organizational, human and technical factors.
These challenges will be addressed and the main
objectives will be achieved through meeting the following sub-goals of the project:
Identify and describe organizational and operational
barriers for risk control.
Provide new knowledge about the effectiveness of
organizational, human and technical factors for the
performance of operational barriers.
Define indicators for these factors that are suitable
for measurement of barrier performance and establish methods for how to measure the status of these
factors.
Develop new models for barrier performance
reflecting organizational and operational management factors.
Demonstrate use through case studies and proposed
risk reduction measures.
Analyze experience data in order to identify those
risk management regimes that are most effective.

1.5 Abbreviations
BBN
Bayesian Belief Network
BORA
Barrier and Operational Risk Analysis
FPSO
Floating Production, Storage and
Offloading
MTO
Man, Technology and Organization
OMT
Organization, huMan and Technology
PSA
Petroleum Safety Authority
PSF
Performance Shaping Factor
RIF
Risk Influencing factor
RNNS
Risk level project
SMS
Safety Management System

1182

Table 1. Distribution of leak scenarios per 100 installation

years for all UK and Norwegian companies.
Circumstance of leak
Technical faults
Latent errors from
intervention
Immediate errors from
intervention
Process control
Design error
Total

UK
installations

Norwegian
installations

11.8

11.2

4.1

15.3

1.7
0.2
0.5
18.2

4.1
4.8
2.7
38.1

3.1

Table 2. Distribution of leak scenarios per 100 installation

years for two groups of companies.
Circumstance of leak
Technical faults
Latent errors from
intervention
Immediate errors from
intervention
Process control
Design error
Total

2 Norwegian
companies

Other
companies

14.0

6.9

27.3

3.4

6.7
6.7
6.7
61.3

1.4
1.4
0.7
13.8

HYDROCARBON LEAKS OVERVIEW

Hydrocarbon leak statistics was presented in a paper to

the ESREL2007 conference, Vinnem et al., (2007b).
It was shown that leak frequencies on Norwegian
installations are considerably higher when compared to UK installations. This is illustrated in
Table 1.
It was further shown in that paper that there is a considerable difference between Norwegian companies,
as illustrated in Table 2. Two Norwegian companies
have much higher leak frequencies than the other companies, the second group may be compared to the
average of UK companies, with respect to average leak
frequencies.
The main differences between the companies with
high average leak frequencies and those with lower
frequencies are the leaks caused by manual intervention in the process systems, which is shown
clearly by Table 2, when comparing values for the
two rows marked Latent errors. . . and Immediate
errors. . .. Especially the latent errors give a very significant contribution. This paper focuses mainly on
the latent errors, when leaks caused during manual
interventions are analyzed.

COMMON CAUSE FAILURES

About the terms

Common cause and common mode failures are well

known in relation to engineered systems. The more
general expression is Dependent failures. Common
causes exist when e.g. two different components fail
due to influence of the same (common) factor, such
as extreme environment or external impact. Common mode failures occur when two components fail
in the same mode due to being influenced in the same
manner.
This term is perhaps not so widely used in relation to Human and Organizational Factors (HOF), we
may instead talk about dependencies. So how could
we apply a parallel logic to human and organizational
errors?
A basic approach might imply to consider the
human element as a common cause why multiple
operational barriers may fail. This appears to be
an oversimplification, because a systems perspective
on operational failures may look for root causes in
the physical systems, organizations, management and
culture in order to establish causal relationships.
In the OTS development project, we have tried to
establish verification tools in order to identify weaknesses in organizations, management and cultural
aspects in order to look as the vulnerability of an operating organization with respect to occurrence of major
accidents.
The factors are often referred to as Performance
Shaping Factors (PSFs), but are in our project referred
to as Risk Influencing Factors (RIFs).
When we assess an organizations performance with
respect to vulnerability to major hazards, we may
determine that several of these factors are substandard.
This is important in itself, as input to identification
of actions to reduce the vulnerability by severalto
some extent overlappingactions. If we are trying to
quantify the proneness for major accidents to occur,
we need to make sure that we avoid double-counting
as much as possible. Thus we have to identify what
are common cause effects, or dependencies between
factors that we consider.
One could ask; is dependency a problem that we
have to tackle or should we try to avoid such dependencies? But dependency is highly wanted: it may
be stated that the objective of safety management
is to introduce a common mode influence (a Safety
Management System, SMS) to bring all base event
probabilities to the lower bound of their intrinsic
range, Hale (2007). The objective in an organization
should be to have as much dependency (i.e. influence) as possible, although it makes the modelling
of dependencies a real challenge. The OTS methodology report, Steen et al. (2007) has illustrated the

1183

dynamic elements. To what extent this is effective, will

also depend on how active and efficient the supervision is. There are indications that learning loops and
continuous improvement and feedback are far from
common on many installations.
It may also be argued that when the OTS pilot verification was performed in 2007, the findings of root
causes in investigations for the plant in question over
a period of 56 years concluded that the root causes
had not changed significantly during this period. There
are therefore reasons to claim that the models may be
developed without a large focus on learning loops and
other aspects of dynamics.

Operational error
resulting in leak

Work practice in
maintenance
operations

Communication

Procedures &
documentation

Work practice in
.....

Physical
working
environment

Competence

Work practice in
.....

Working time
factors

Change
management

Supervision

Management

3.3
Figure 2. Relationship between Risk Influencing Factors
and operational leaks, based on Steen et al., (2007).

dependencies between the RIFs with the following

illustration, see Figure 2.
3.2

Modelling of dependencies

Dependency is a particular aspect of quantification, which needs to be addressed thoroughly.

A simplified approach to dependencies was used in
the BORA methodology, Vinnem et al., (2004a).
A further consideration is to what extent a dynamic
modelling should be the ultimate goal. Hale (2007)
asks the challenging question: Can you hide the
dynamic inside a static black box?delivery systems,
learning loops, . . . . This is a fundamental question
which needs to be considered carefully.
Ideally, one would want to capture the dynamics of
the organizations and operations in a realistic model,
in a transparent and objective manner. On the other
hand, how dynamic are these conditions?
It should be noted first of all that the focus is on
major hazard aspects in these projects. This implies
that even precursors (hopefully) will be quite rare.
A major part of the data basis for these projects comes
from leaks of hydrocarbon gas and liquid on offshore
and onshore petroleum facilities. It has been claimed
that for an average employee, critical leaks (i.e. such
leaks that could escalate into a major accident) may
typically be experienced about once every ten years
(not necessarily by the same person, but at least by
persons on the same shift). In the view of such a low
occurrence frequency, it may be questionable if there
are learning loops that are effective.
The maintenance operations that have been found
to be the major source of leaks is on the other hand routine maintenance operations, such operations that are
carried out regularly many times each year. From this
perspective, there could be learning loops and other

Dependencies in BORA & OTS projects

The BORA project adopted a rather coarse modelling

of dependencies (Sklet et al., 2006), based on the
thinking that the appropriateness of the modelling
should be demonstrated first, before a lot of efforts
was invested in advanced modelling of dependencies between Risk Influencing Factors. The modelling
assumed somewhat arbitrarily that if one RIF is dependent on two or more RIFs, and at least two of those
RIFs are given either a very high or a very low score,
then the score of that first RIF is adjusted one score
category up of down.
The OTS project has considered each performance
shaping or risk influencing factor separately, and has
therefore not needed to consider dependencies. The
methodology report has a brief discussion of dependencies, but has not special focus on this aspect, due to
consideration of each factor separately. Relationships
have been shown however. Figure 2 is based on the
relationships presented by Steen et al., (2007), and has
developed these relationships somewhat further. OTS
has combined management and supervision into one
common factor, but these are kept separate in Figure 2.
The hierarchical structure of Figure 2 suggests that
a BBN may be a possible modelling technique, considerably more suitable than the single level influence
diagram used in the BORA project. The data problem on the other hand grows considerably, and has to
be solved in an efficient manner in order to have an
approach which is feasible in risk assessment studies
performed in a commercial context.
3.4

Leak frequencies and risk influencing factors

The Risk Level Project, (Vinnem et al., 2005) has collected considerable amounts of incident data (lagging
indicators) and barrier elements test data (leading indicators). The data has not been analyzed extensively
with respect to identify differences in risk management practices and approaches, to the extent that the
effectiveness of the various alternatives has been established. A limited study has on the other hand been

1184

completed in 2007 (Vinnem et al., 2007a), whereby

incidence rate of leak frequencies and questionnaire
data were analyzed in order to identify correlations.
No correlations could be found however, pointing to a
possible explanation that the questionnaire data in the
PSA Risk Level Project are too general and dominated
by generic working environment issues.
A recent paper (Vinnem, 2008) used an alternative questionnaire survey in order to look for potential
correlations between risk influencing factors and incidence rate of leaks. The survey used is an internal
survey which has somewhat more pointed questions.
This is probably the main reason why it was possible
for find some correlations. The following correlation
coefficients were determined for indexes expressing
some selected risk influencing factors against a leak
incidence rate expressed as number of leaks per 10,000
work orders:
Work practice:
Management:
Supervision:

0.50
0.46
0.40

5. Maloperation
of valve(s)
during manual
operation; 6

6. Maloperation
of temporary
hoses; 1
1. Incorrect
blinding/isolation;
11

4.Erroneous
choice or
installations
of sealing
device; 2

3. Valve(s) in
incorrect
position after
maintenance; 6
2. Incorrect
fitting of
flanges or bolts
during
maintenance; 10

Figure 3.

Distribution of latent failures on sub categories.

These correlation coefficients are much higher

than those found in the study which used the RNNS
(Risk Level project) questionnaire, as discussed above
(Vinnem et al., 2007a). This is noteworthy, as the
questionnaire used was not developed with particular
emphasis on major hazards.

20012006, which have occurred during, or as a consequence of, maintenance activities in the hydrocarbon
processing areas. Previous studies have also considered the same data, see for instance Vinnem et al.,
(2007b), but the present study is the first to discuss
the detailed failure scenarios and the risk influencing
factors involved.

3.5 Other projects

4.1 Data basis

According to Hale (2007) there are few studies of

empirical nature that have attempted to consider the
efficiency of the various approaches to management
of major accident risks. A similar observation is also
made by Robson et al., (2007).
Hurst et al., (1996) analyzed data from accidents
and incidents and safety attitude survey questionnaires
from six major hazard sites in four European countries. One of the conclusions made by Hurst et al., is
that replies to the survey do not predict neither the
occurrence of injuries nor loss of containment events.
This implies that one of the main tasks of the Risk
OMT project may be particularly useful and contribute
to fill a gap in the knowledge of effectiveness of risk
management practices, especially with emphasis on
major hazards.

The data basis consists of 38 hydrocarbon leaks on

offshore installations, where the leaks have occurred
as a consequence of maintenance in the process area.
The overall distribution of leak scenarios is as follows:

ANALYSIS OF HC LEAKS DURING

MAINTENANCE

This chapter reports some results of a limited data

analysis of precursor data, i.e. hydrocarbon leaks
on Norwegian offshore installations in the period

Immediate leak during intervention: 1 case

Leak due to latent failure introduced
during intervention:
36 cases
Leak due to inadequate design:
1 case
The cases with latent failures introduced during
intervention in the process systems are split on subcategories in Figure 3, in relation to more precise
definitions of what error that was performed.
Incorrect blinding/isolation, incorrect fitting of
flanges or bolts, valves in incorrect position and
maloperation of valves account for 33 of 36 cases.
Four installations account for 19 of the 38 leak
cases, with respectively 7; 5; 4 and 3 cases during
the period. No installation has had more than one leak
in average per year. But one installation actually had
five leaks during one year, implying that the same
shift(s) could have been involved in more than one
leak that particular year. It is unlikely that the same
shift is involved in more than one leak per year in all

1185

Resetting; 4

Planner; 5

Planning &
preparation; 7

Execution &
resetting; 1

Area
responsible; 4

CCR personnel;
6

Planning &
resetting; 2
Area
technician; 20

Execution; 12

Maintenance
personnel; 15

Preparation; 11
Prod
technician; 2

Preparation &
execution; 1

Figure 5.
errors.
Figure 4.

Overview of work phases when errors are made.

other cases, and there will probably be several years

between leaks that the same shift is involved in, for the
majority of the installations.
4.2

Work phases and personnel involved

Typical maintenance activities were analyzed using

a Task analysis approach (Steen et al., 2007). The
following are the four main work phases of any intervention in the process systems on these installations:

Overview of personnel who have performed

Planning
Preparation
Execution
Resetting & restarting

The distribution of errors in the leaks on the different work phases, including the cases where errors have
occurred in two work phases, is shown in Figure 4.
The errors that occur during the planning phase are
all requiring at least one additional error in order to
cause a leak, most of the additional errors are during
preparation. In virtually all of the cases with second
failure in the preparation phase, these errors appear
to be strongly correlated with the error made during
the planning phase. As an illustration, if the planning
phase failed to complete a checklist for which valves
to operate and which blindings to install, then the failure to close all relevant valves or omission to install
blindings are examples of strongly correlated errors.
In the remaining phases, preparation, execution and
resetting, the majority of the errors are single errors
that lead to latent failures that have caused leaks when
the systems are being or have been started.
It is also worthwhile to consider what personnel
that have been involved in the errors made during the

maintenance activities. Figure 5 gives an overview of

the personnel who have been involved in making the
errors. When the error is made during planning, the
split between planners and Area responsible personnel
is about even.
When the errors are made in the preparation phase,
the Area technicians are the mainly involved personnel, with some few additional cases with control room
(CCR) personnel and the mechanics. The only group
involved in the errors during execution is the group
of mechanics carrying out the work, whereas the Area
technician are the only personnel involved in the errors
during resetting.
This implies that all groups of personnel who are
involved in the interventions are involved in making
the errors. It could also be noted that in the case where
five leaks occurred on one installation during one calendar year, one of these started with planning, three
with the preparation and one with the execution. Area
technician, area responsible and mechanics were the
personnel groups involved in these five cases.

4.3 Root causes according to investigations

It was required that investigation reports were available
in order to determine root causes. This was available in
30 of the 38 cases considered. The distribution of root
causes according to the investigation reports is shown
in Figure 6. The average number of root causes given
per case is just above 2.0.
It should be noted that the root cause analysis is
often the weakest element of an investigation, especially the investigations that are carried out without
the involvement of onshore specialists. This implies
that there is some uncertainty associated with these
distributions.

1186

100 %
90 %

Work time
Documentation;
factors; 1
5

Communication;
6

80 %
70 %
Communication

60 %

Competence;
12

Competence
Management/ supervision

50 %

Work practice
Documentation

40 %

Work time factors

30 %
20 %

Work practice;
28

10 %
0%
Planning

Management/
supervision; 16

Figure 8.
Figure 6.

Fraction of leaks where each RIF is contributing (N=30)

Execution

Resetting

Root cause contributions in the work phases.

Distribution of root causes involved.

100 %
90 %
80 %
70 %
60 %
50 %
40 %
30 %
20 %
10 %
0%
Communication

Figure 7.

Preparation

Competence

Management/
supervision

Work practice

Documentation

Work time factors

Fraction of leaks where factors have contributed.

Work practice is given as one of the root causes in

over 90% (28 out of 30) of the cases. Management/
supervision is given as one of the root causes in just
over half (16) of the cases.
If we consider work practice or management/
supervision, all cases have at least one of these as the
root causes. Competence is the third most important
root cause. The fractions of the leaks where each of the
risk influencing factors have been mentioned amongst
the root causes is shown in Figure 7.
It may also be worthwhile to consider differences
between the contributions of root causes in the different stages of an intervention. This is presented in
Figure 8.
Work practice is the highest factor in all phases.
Management is the second highest contribution in all
phases but one (execution), where competence is the
second highest contribution. Competence is also a high
contribution in the resetting phase. It should however,
be noted that the number of cases due to resetting
where root causes are known, is as low as 3 cases.
Communication appears to be a significant contribution mainly during planning. Competence is a

high contribution during execution and resetting. Management is a lower contribution during the execution. These differences are most likely not significant
differences in a statistical sense.
It could be mentioned that the root causes identified were determined by the research team, based on
interpretation of the descriptive texts in the investigation reports. This should imply that the root causes
are identified in a systematic manner, and thus more
reliable than what is often found as root causes
in investigation reports. But the reliability in each
case is obviously dependent on the correctness of the
description of circumstances in the investigations.

DISCUSSION

The previous chapter has illustrated a substantial

amount of details relating to human and organizational factors for 3040 major hazard precursor events,
i.e. hydrocarbon leaks of a magnitude that may cause
escalating fire and/or explosion. The analysis of the
majority of these cases may be based on relatively
detailed investigation reports. As a data set, this is
a fairly good set of data of a reasonable size. The
next question is whether it is suitable in order to reveal
dependencies or common causes of an organizational
nature.
We have seen from the previous chapter that more
than 90% of the leaks are caused by factors which
include incorrect or inadequate work practice. Management is also a contributing factor in about half of
the cases.
We have further demonstrated that inadequate communication appears to be a significant contribution
to human errors during planning. Insufficient competence appears to be a significant cause of human
errors during the actual execution of maintenance

1187

work. Inadequate management appears to be one of the

main causes of planning errors. Inadequate work practice is on the other hand the most common contributor
in all phases of the work.
The fact that work practice and management are the
two dominating root causes may be taken to support
the structure suggested in Figure 2, at least as far as the
central role that work practice appears to have. The
assumption implied in Figure 2 is that management
is to some extent a common cause which influences
on all other risk influencing factors.
Work practice and management were also in the
OTS pilot project for a land based petroleum facility
among the main aspects where improvements were
recommended.
But the difficult question is as follows: Are the
findings given here sufficient in order to state that
management is the common cause of leaks due to
operational factors?
The response to such a question is regrettably that
one is far from being able to conclude on which are the
common causes to human and organizational errors,
even with the detailed information that is available in
the present study.
It was already noted in Section 3.5 that there are few
studies that have attempted to assess the effectiveness
of risk management practices based on empirical data,
especially in the case of major hazard risk. This is
probably a reflection of the same challenge as we face
with the data analyzed in this paper. Even with quite
detailed data, it is very limited what we are able to
conclude about common causes and dependencies.
What would be required? We have already briefly
mentioned the possibility to use a more detailed and
more relevant questionnaire survey, with particular
emphasis on major hazards. This is a possibility that
we hope to be able to put into effect in the present
project. But will it help towards the ultimate goal?
It may help in relation to being able to measure differences between how the management practices are
experienced on the different installations. It may also
help in relation to seeing more distinctively what the
characteristics are for those installations with low incidence rate of hydrocarbon leaks in relation to those
installations with high rate. But questionnaire surveys
will probably never be the solution to provide the final
answer about dependencies.
Based on the results of a detailed and pointed questionnaire survey, one will probably have to conduct
field studies, in order to reveal the basic differences
between the risk management practices, and how these
have effects on the incidence rate of hydrocarbon
leaks.

CONCLUSIONS

Dependencies amongst root causes of human and organizational errors that may cause major hazards are very
challenging. It has been shown that there is little scientific work done in this field, especially based on
empirical data. The considerable challenge involved
in such work is probably one of the factors.
This paper has reported some results from an analysis of circumstances and root causes of major accident
precursors on offshore installations. 38 cases where
leaks have occurred as a result of the maintenance work
in the process area have been considered in detail.
The results may illustrate some of the dependencies, but are not capable of showing what these
dependencies are. The study has demonstrated that
inadequate work practice is a contributing factor to
more than 90% of the hydrocarbon leaks that occurred
on Norwegian installations in the period 200106.
It has further been shown that inadequate management practices have contributed in about 50% of the
cases.
A hierarchical structure of risk influencing factors
is suggested, whereby work practice is the immediate
cause of operational leaks, with a number of influencing factors, amongst which management system may
be the ultimate common cause. This structure is at
present a hypothesis, we will attempt to confirm or
modify this through the programme.
The Risk OMT programme will continue until the
end of 2010, and will make an attempt to study common causes and dependencies amongst human and
organizational errors one of the key subjects during
the second half of the programme period.
The work in order to identify possible risk reducing measures and the most effective risk management
approaches may be considered to be the most valuable
input to enhancement of safe operation. This will be
thoroughly researched in the programme, and we hope
to be able to devote papers to this topic in the future.

ACKNOWLEDGEMENT
The Research Council of Norway and StatoilHydro are
acknowledged for the funding of the programme. Also
the team members of the BORA and OTS projects
are gratefully acknowledged for their contributions,
in addition to Professor Andrew Hale for valuable
discussions around this topic.

1188

REFERENCES
Chen, H. & Moan, T. 2005. DP incidents on mobile offshore
drilling units on the Norwegian Continental Shelf. European Safety and Reliability Conference, Gdansk, Poland,
2630 June 2005.
Hale, A. 2007. Modelling safety management & culture,
presentation to Risk OMT workshop, 1112 December
2007.
Hurst, N.W., Young, S., Donald, I., Gibson, H. & Muyselaar,
A. 1996. Measures of safety management performance
and attitudes to safety at major hazard sites, J. Loss Prev.
Process Ind., 9, 161172.
Robson, L.S. et al., 2007. The effectiveness of occupational
health and safety management system interventions:
A systematic review, Safety Science, 45, 329353.
Sklet, S., Vinnem, J.E., and Aven, T. 2006. Barrier and
operational risk analysis of hydrocarbon releases (BORARelease); Part II Results from a case study, Journal of
Hazardous Materials, A137, 692708.
Steen et al., 2007. OTSMethodology for execution of OTS
verification (in Norwegian only), StatoilHydro report
December 2007.
Vinnem, J.E. et al., 2003a. Operational safety analysis
of FPSO Shuttle tankercollision risk reveals areas of
improvement, presented at Offshore Technology Conference, Houston, 58 May, 2003.
Vinnem, J.E. et al., 2003b. Risk assessment for offshore
installations in the operational phase, presented at the
ESREL 2003 conference, Maastrict, 1618 June, 2003.
Vinnem, J.E., Aven, T., Hauge, S., Seljelid, J. & Veire, G.
2004a. Integrated Barrier Analysis in Operational Risk
Assessment in Offshore Petroleum Operations, presented
at PSAM7, Berlin, 1418 June, 2004.

Vinnem, J.E., Veire, G., Heide Knudsen, B. & Aven, T. 2004b.

A method for developing and structuring risk activity indicators for major accidents, presented at PSAM7, Berlin,
1418 June, 2004.
Vinnem, J.E., Aven, T., Huseb, T., Seljelid, J. & Tveit, O.
2005. Major hazard risk indicators for monitoring of
trends in the Norwegian offshore petroleum sector. Reliability Engineering & Systems Safety, Volume 91, Issue 7,
778791.
Vinnem, J.E. et al., 2007a. Leak frequencies and risk influencing factors (in Norwegian only), Preventor Report
P200793-02, 21.12.2007.
Vinnem, J.E., Seljelid, J., Haugen, S. & Huseb, T. 2007b.
Analysis of Hydrocarbon Leaks on Offshore Installations, presented at ESREL 2007, Stavanger, 2427 June,
2007.
Vinnem, J.E. 2008. Challenges in analyzing offshore hazards of an Operational nature, presented at PSAM9, Hong
Kong, 1823 May, 2008.

1189

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Quantitative risk analysis method for warehouses

with packaged hazardous materials
D. Riedstra, G.M.H. Laheij & A.A.C. van Vliet
National Institute of Public Health and the Environment (RIVM), Bilthoven, The Netherlands

ABSTRACT: In The Netherlands, Quantitative Risk Assessment (QRA) is applied in land use planning around
establishments handling, processing or storing dangerous substances. The QRA method for warehouses with
packaged hazardous materials has recently been evaluated and redrafted. This study discusses the revised QRA
method proposed to the Dutch Ministry of Housing, Spatial Planning and the Environment. Important riskdetermining parameters, such as the conversion factor for nitrogenous substances to the toxic combustion
product NO2 and the survival fraction of non-combusted extremely toxic substances, have been adapted. The
consequences for the calculated individual risk 106 /year distances are presented.

INTRODUCTION

In the Netherlands there are about 900 establishments

where more than 10,000 kg of packaged hazardous
materials are stored. The risk these establishments
pose for human health is mainly determined by a fire
that causes a release of toxic combustion products such
as nitrogen dioxide (NO2 ), hydrogen chloride (HCl)
and sulphur dioxide (SO2 ). If extremely toxic substances like pesticides/herbicides are stored the release
of non-combusted substances may determine the risk
as well.
During the first stage of a fire, in particular, the
toxic fumes are relatively cold and will disperse into
the surrounding area of the warehouse. At a later
stage when the building structure collapses, plume
rise is presumed (by the generated heat) and the possible lethal consequences of the fire are assumed to be
negligible.
Warehouses with packaged hazardous materials
have to comply with stringent requirements in order to
prevent (major) accidents to happen. For the remaining
risk a quantitative risk assessment (QRA) is applied. In
the Netherlands this approach is followed in land use
planning around establishments handling, processing
or storing dangerous substances. Recently, the criteria
for the acceptability of risk were laid down explicitly
in Dutch legislation. In 2010, housing and vulnerable objects like schools and hospitals will no longer
be allowed within the individual risk contour 106 per
year.
Individual risk is defined as the probability that a
person permanently residing at a fixed location (outside the establishment) will be killed as a result of an

accident with hazardous substances. The individual

risk is presented in the form of iso-risk contours.
The Dutch Guideline for Quantitative Risk Assessment, the Purple Book (CPR 1999), has been redrafted
in order to structure the method more in agreement
with the new legislation and to update the risk calculation method where appropriate (RIVM 2008). Part of
the Dutch guideline is the QRA method for warehouses
with a maximum storage area of 2500 m2 in which
more than 10,000 kg of packaged hazardous materials, chemical waste materials or pesticides/herbicides
are stored. The current method originates from 1997
(VROM 1997).
In 2005, the guideline for warehouses with packaged hazardous materials was actualised to the socalled PGS-15 (VROM 2005). The guideline has been
extended to the storage of cylinders, aerosols, solid
flammable materials and limited quantities of (less
dangerous) organic peroxides. Further, the classification of hazardous substances has been changed from
the EU Directive 67/548/ EEC to the European classification for transportation of hazardous materials
(ADR). Due to these changes, the QRA method had
to be redrafted. During this revision process, other
specific issues were clarified as well.

2
2.1

CURRENT QRA METHOD

General

The current QRA method defines fire scenarios with

different sizes and probabilities. Table 1 details this
for all firefighting systems used in the Netherlands.

1191

Table 1.

Probability per fire area scenario and applied ventilation rate per firefighting systems.
Probability of different fire sizes
Ventilation
rate [hr 1 ]

20 m2

50 m2

100 m2

300 m2

900 m2

4&
4&
4&
4&

45%
63%
63%
99%

44%
26%
26%
n.a.

10%
10%
10%
n.a.

0.5%
0.5%
0.5%
n.a.

0.5%
0.5%
0.5%
1%

4&
4&

89%
89%
35%
n.a.

9%
9%
45%
20%

1%
1%
10%
30%

0.5%
0.5%
5%
28%

0.5%
0.5%
5%
22%

Protection level 2: Fire detection & fire brigade

<15 min
Flammable liquids (ADR class 3; max 100 tonnes)
stored in synthetic packaging
NOT stored in synthetic packaging
No flammable liquids (ADR class 3)

n.a.
n.a.
n.a.

n.a.
n.a.
n.a.

n.a.
n.a.
n.a.

55%
78%
78%

45%
22%
22%

Protection level 3: Only preventive measurements

n.a.

n.a.

n.a.

78%

22%

Firefighting system
Protection level 1: Firefighting system or company
fire brigade
Automatic sprinkler installation
Automatic in rack sprinkler installation
Automatic deluge installation
Automatic inerting gas installation
Automatic high expansion foam installation
with OUTSIDE air
with INSIDE air
Company fire brigade with manual deluge system
Company fire brigade only

A fire scenario is characterised by the fire duration,

size of the fire (maximum 900 m2 ) and burn rate. The
latter depends on the composition of the substances
stored and the quantity of oxygen available during the
fire. The composition of the materials stored codetermines which toxic combustion products, such as
NO2 , HCl/HF/HBr and SO2 , are formed and in what
quantities.
2.2

Frequency and size of the warehouse fire

The initial probability of fire in a warehouse is equal

to the probability of fire in a pesticide/herbicide storage area (Molag 1991). This has been set at 8.8
104 /year, based on 4 major fires in warehouses with
pesticides/herbicides among 350 warehouses between
1975 and 1987 in the Netherlands.
Hazardous substances can be stored in a warehouse
under three different levels of protection:
Protection level 1: Automatic firefighting system
or firefighting by a company fire brigade;
Protection level 2: Fire detection system and firefighting by the local fire brigade;
Protection level 3: Storage facility where only
preventive measurements are taken.
For protection levels 1 and 2, the same probability is
assumed as for a fire in a warehouse where pesticides/
herbicides are stored. This is because substances
stored under protection levels 1 and 2 cause a risk comparable to pesticides/herbicides. A five times lower

probability of fire has been prescribed for substances

stored in a protection level 3 warehouse, namely:
1.8 104 /year, since flammable substances are not
allowed under these storage conditions.
It is assumed that the duration of a fire is equal to the
time needed to extinguish the fire, which depends on
the firefighting system used. However, a limit is set for
the duration of the fire, which is equal to the assumed
maximum exposure duration of the population near the
warehouse, namely: 30 minutes. After this period it is
assumed that the fire has been mainly extinguished,
the people outside the warehouse have been warned
to shelter or the roof of the warehouse collapses and
plume rise occurs.
For several protection level 1 firefighting systems,
two ventilation regimes are distinguished (4 hr 1 or
; see table 1). If the automatic self-closing doors
function well and close after detection of an initial
fire, the fire will be ventilation controlled (or oxygen restricted; probability 0.98) and if not the fire
will be fuel restricted (probability 0.02). Other firefighting systems always cause a fuel-restricted fire
only.

2.3 Release rate

The release rate of the toxic combustion products (and
of non-combusted products if extremely toxic substances are present) is determined by the composition
of the hazardous materials stored, the burning rate
(determined by the quantity of oxygen present) and

1192

the size of the fire. The maximum burn rate for most
substances is 0, 025 kg/m2 s (if the fire is not ventilation controlled). For flammable liquids a burn rate of
0, 100 kg/m2 s is applied.
2.3.1 Toxic combustion products for ventilation
controlled fires
For a warehouse with an average composition of
Ca Hb Oc Cld Ne Sf X the release rate  for NO2 , SO2
and HCl (HF and HBr are considered as HCl) are:
NO2 = min(Bvc , Bfr ) e 46 /Mw

(1)

SO2 = min(Bvc , Bfr ) f 64 /Mw

(2)

HCl = min(Bvc , Bfr ) d 36.5 /Mw

(3)

with
Bfr = Bmax A

(4)

Bvc = O2 Mw /OD

(5)

where min (Bvc , Bfr ) is the minimum value of the

ventilation-controlled or fuel-restricted burn rate,
Bmax the maximum burn rate, the conversion factor (35% for NO2 and 100% for both SO2 and HCl;
Laheij 1995), Mw the molecular weight and A the
fire area. The quantity of oxygen available O2 and
oxygen demand OD for the combustion of 1 mole of
Ca Hb Oc Cld Ne Sf X are defined as:
O2 = 0.2 (1 + 1/2F)V/(24 t)

(6)

OD = a + (b d)/4 + 0.1e + f c/2

(7)

where 0.2 is the oxygen content in air, F the ventilation

rate, V the volume of the warehouse, 24 the molar volume and t the maximum exposure time (1800s). The
letters a, b, c, d, e and f refer to the average composition
of the materials stored.
2.3.2 Toxic combustion products for fuel
restricted fires
For fuel restricted fires the release rate  is only
determined by the burn rate Bfr in the formulas (1)(3).
2.3.3 Non-combusted toxic substances
The release rate of non-combusted extremely toxic
substances unburnt is calculated by multiplying the
resulting burn rate (ventilation controlled or fuel
restricted with the quantity of these substances
(mass%) and the survival fraction (sf):
unburnt = min (Bvc , Bfr ) mass% sf

(8)

For fuel restricted fires the release rate unburnt is only

determined by the burn rate Bfr in formula (8).

For the survival fraction it is assumed that 10%

of the extremely toxic materials is released in a
non-combusted form during a fire if the flashpoint
100 C and that 2% of the remaining extremely toxic
substances is released in the non-combusted form.
The height of materials being released from the
warehouse is equal to the height of the storage facility.
It is assumed that the products being released in the
recirculation zone of the building will disperse like a
neutral gas into the surrounding area. Probit functions
are used to calculate the probability of death due to
toxic substances at a given exposure.
3

REVISED QRA METHOD

The most important elements in the QRA method that

have been evaluated and, where necessary, revised are
detailed in this section.
3.1 Extension to new ADR classes
The actualised PGS-15 guideline for warehouses with
packaged hazardous materials has been extended to the
storage of cylinders, aerosols, solid flammable materials and limited quantities of (less dangerous) organic
peroxides in 2005.
Cylinders, defined as a pressure vessel with a
water volume less than 150 litres, are normally stored
separately from other hazardous materials outside a
warehouse. Therefore the QRA scenarios deviate from
the warehouse fire scenarios. Two failure scenarios are
identified: instantaneous failure and leakage from a
3.3 mm hole, both with a 5 107 /year frequency per
cylinder. These frequencies are based on an analyses
of 74 accidents with cylinders between 1977 and 1984
(Marchal, 1985).
Storage of aerosols with a flammable (gas as a) propellant can only result in a normal violent fire. If the
aerosols contain nitrogen, sulphur or chlorine (and fluorine or bromine) then the release of toxic combustion
products is possible and the warehouse should be considered in a QRA. If the firefighting system is not
capable of extinguishing an initial fire with a limited
quantity of aerosols, the fire will rapidly expand to
the entire storage area due to the rocketing effects of
burning aerosols. Therefore only two fire area scenarios are considered: 20 m2 and the maximum storage
area. The probability of the latter scenario is one minus
the change of occurrence of the 20 m2 fire scenario
according to Table 1. Further, a maximum burning
rate of 0, 100 kg/m2 s is applied.
In the Netherlands, only in a limited number
of warehouses flammable solids are stored (ADR
class 4). The proportion of flammable solids in these
storage facilities is limited and usually restricted to
ADR class 4.1. For this subcategory (flammable

1193

solids) as well as spontaneously combustible substances (4.2) and substances which emit flammable gas
(in contact with water; 4.3) the rule is that they must
be stored separately or separately from flammable substances (ADR class 3). If the risk of these flammable
solids stored is not determined solely by the release of
toxic combustion products, the QRA method implies
that additional safety measurements are being taken
to prevent these scenarios from happening. Therefore no additional QRA scenarios are considered for
this heterogeneous category (stored in relatively small
quantities).
According to the PGS-15 guideline, storage of
organic peroxides (ADR class 5.2) in warehouses is
only allowed for a small quantity of the less dangerous
peroxides (<1000 kg per establishment) packaged in
so-called limited quantities (LQ). Therefore organic
peroxides are not considered separately from other
hazardous materials in the method as well.
3.2

Flammable/non-flammable materials

Since according to the current QRA method it is

not completely clear under which conditions a fire
with lethal effects outside the establishment is possible, QRAs are currently also made for the storage
of non-flammable substances in packaging (like certain types of fertilisers and cement). Consequently the
revised method pays greater attention to the actual
fire hazard of the dangerous substances stored. The
method discriminates between flammable (including
non-classified substances) and non-flammable substances. The latter are divided into materials that
contribute to the fire (by decomposition or evaporation) and those that do not. If only non-flammable
substances are stored a fire with lethal effects outside
the establishment is not a credible scenario (and a QRA
is not relevant).
If flammable and non-flammable substances are
stored within the same warehouse, fire is regarded as a
credible scenario and should be considered in a QRA,
irrespective of the quantity of flammable materials
stored.
3.3

Failure frequency

Although the current value of 8.8 104 per year for

warehouses has given rise to a lot of discussion, the
value has been maintained, because warehouse fires
are never systematically registered in the Netherlands.
Companies in particular argue that this failure frequency, based on outdated and limited information,
is too high. Yet in this discussion it is all too easily
forgotten that the value is based on four major fires
and small fires (like 20 m2 ) are included in this frequency as well. According to the QRA method, a fire
of 300 m2 in a warehouse with automatic sprinklers,

for example, has a probability 100 times lower than

the initial fire frequency (see Table 1).
Further in Belgium much higher fire frequencies are used, bases on research of Hymes [Hymes
1992]. In this research for the HSE (United Kingdom)
351 warehouse fires among a population of 36,000
warehouse.years between 19761987 have been studied. For severe warehouse fires, resulting in a
collapse of the building structure a frequency of 6.9
104 and 2.5103 per year is applied for warehouses
with and without an automatic sprinkler system.
3.4 Conversion factor
An important parameter for determining the risks of
warehouses is the quantity of nitrogenous substances
stored. During a fire these compounds are converted
to various hazardous and non-hazardous combustion
products, like nitrogen (N2 ), nitrogen oxides (NO,
N2 O, NO2 ,), ammonia (NH3 ) and hydrogen cyanide
(HCN). The current method only considers NO2 and
assumes a conversion factor of 35% (Laheij 1995).
The conversion factor has been evaluated by a literature search and by consulting a group of experts from
RIS (DK), HSL (UK) and TNO (NL). The research
was not restricted to the conversion to NOx , HCN was
considered as well. The experts recommended evaluating both conversions with the experiments from two
large international research projects, carried out by the
above-mentioned institutes in the 1990s (COMBUSTION and TOXFIRE). These combustion experiments
varied from laboratory scale (<1 gram) to relatively
large-scale experiments (50 kg). The average of the
highest reported conversion factor for NOx and HCN
for each experiment (and within some experiments for
each scale) reported has been calculated: 6.2% for both
NOx (32 experiments) and HCN (21 experiments) with
a standard deviation of 5.6% and 4.5% respectively
(Matthijsen, 2007).
Since NOx is mainly formed under well-ventilated
conditions and HCN during ventilation-controlled
fires, the average conversion factors for both compounds will not be achieved at the same time and the
addition of the two values would be a very conservative
approach. Therefore a conversion factor of 10% was
proposed. This concurs with factors used in several
other European countries, such as the UK, Denmark
and Belgium (proposed).
The conversion factors for chlorine and sulphurcontaining substances into hydrogen chloride (HCl)
and sulphur dioxide (SO2 ) have not been adapted and
both remain 100%.
3.5 Survival fraction
If extremely toxic substances like pesticides/herbicides
are stored in a warehouse then a certain quantity of

1194

Table 2. Applied survival fraction for (extremely) toxic

substances stored at different storage levels.
Phase of the (extremely) toxic
substances for the different
protection levels
Toxic liquids and powders
Protection level 1
automatic sprinkler, deluge,
gas extinguish installation,
hi-ex foam system with
inside air and company
fire brigade with
manual deluge system
automatic hi-ex foam system
with outside air and company
fire brigade only
Protection level 2
Protection level 3
Other toxic solids (granulate)
Protection level 1, 2 or 3

Storage height
B

Bvc
Bfr

1.8 m

>1.8 m

10%
1%

30%
10%

Bfr

1%

10%

Bfr
Bfr

1%
1%

10%
10%

Bfr

1%

1%

substances with a default probit value (modelled as

nitrogen):
I: Pr = 5.9 + ln(C2 t/60)
II: Pr = 10.1 + ln(C2 t/60)
where C is the concentration and t the exposure time.
The probit values are based on LC50 (rat, 4h) values
for the inhalation of dust and mist:
I: LC50 (rat, 4h) 0.2 mg/l
II: 0.2 < LC50 (rat, 4h) 2 mg/l
The LC50 (rat, 4h) for packaging group I is based on
the average value (0.2) of these substances available
in the Dutch Safety and Environmental Risk Database
(SERIDA, withdrawn in 2008): 0,06 mg/l. For packaging group II, a conservative value of 0.5 mg/l has
been proposed.
4

B = ventilation controlled (Bvc ) or fuel restricted burn

rate (Bfr ).

these compounds will be released during a fire in a

non-combusted form. Experts from RIS and HSL
were consulted about the current QRA modelling of
this survival fraction, based on its flashpoints (2% or
10%). Experts from both institutes are of the opinion that the storage height rather than the flashpoint
is crucial for determining the survival fraction. This
was confirmed in several studies, in which high survival fractions were reported if liquid and powdered
organophosphorous pesticides were spilled from relatively high pallet levels into the flames on the ground
level of a warehouse (Atkinson, 1992, 1994a, 1994b).
A significant quantity of these droplets and particles
were driven upwards along with the fumes before the
ground level was reached. For granulates this effect is
less pronounced. Another parameter stated in the literature is the ventilation rate. A higher survival fraction
was found if a fire is ventilation controlled.
Therefore two storage levels are distinguished in
the revised QRA method: below and above the test
fall height of UN approved packages (1.80 m). Further is discriminated between ventilation controlled
and fuel restricted fires (See table 2). In compliance
with the values described by the Safety Report Assessment Guide of the HSE (HSE 2002), three different
survival fractions are proposed in the revised QRA
method, 1%, 10% and 30%.
Probit functions are used to calculate the probability
of death due to toxic substances at a given exposure. In
the revised QRA method, the toxicological effects of
the non-combusted toxic substances are modelled for
each ADR class 6.1 packaging group using imaginary

RESULTS

This section describes the results of the risk calculations according to the revised QRA method proposed
to the Ministry of Housing, Spatial Planning and the
Environment.
4.1

Individual risk 106 /year distances

(toxic combustion products)

For several types of establishments in the Netherlands, tables with individual risk 106 /year distances
are available for land use planning. The new safety
distances for warehouses with packaged hazardous
materials will be published this year and will be based
on the results given in Table 3. In this table individual risk 106 /year distances are given for warehouses
with different firefighting systems and storage sizes
(100 m2 , 300 m2 , 600 m2 and 9002500 m2 ). For protection level 1 warehouses, safety distances are only
given for the hazardous materials stored which contain
less than 15% nitrogen. For protection level 2 and 3
warehouses, the table discriminates between three different nitrogen contents (5%, 510% and 1015%).
The individual risk 106 /year distances for protection
level 1 warehouses are relatively small (except the firefighting system company fire brigade only), since
the automatic firefighting systems or firefighting by a
company fire brigade make an immediate or fast interference possible in the event of a warehouse fire. In
protection level 2 and 3 storage facilities, an initial fire
will expand to larger sizes before firefighting starts.
This results in greater safety distances.
Although the conversion factor for nitrogenous substances has been reduced from 35% to 10%, the
nitrogen content is still the most dominant parameter
determining the risk of warehouse fires. The contribution of HCl and SO2 may even be neglected if the

1195

Table 3.

Individual risk 106 /year distances for warehouses with different storage sizes
Individual risk 106 /year distance [m]

Firefighting system
Protection level 1: Firefighting system or company
fire brigade
Automatic sprinkler installation
Automatic in rack sprinkler installation
Automatic deluge installation
Automatic inerting gas installation
Automatic high expansion foam installation
with OUTSIDE air
with INSIDE air
Company fire brigade with manual deluge system
Company fire brigade only
Protection level 2: Fire detection & local fire
brigade < 15 min
Flammable liquids (ADR class 3; max 100 tonnes)
stored in synthetic packaging
Flammable liquids (ADR class 3; max 100 tonnes)
NOT stored in synthetic packaging
No flammable liquids (ADR class 3)

Protection level 3 Only preventive measurements

Nitrogen
content

100 m2

300 m2

600 m2

9002500 m2

15%
15%
15%
15%

30
30
30
20

30
30
30
20

40
40
40
20

50
50
50
35

15%
15%
15%
15%

50
20
20
310

50
20
30
550

60
20
40
660

60
20
85
720

5%
10%
15%
5%
10%
15%
5%
10%
15%

290
290
340
290
290
340
50
130
210

360
500
620
360
500
620
150
360
530

230
470
660
170
380
560
170
380
560

290
550
750
220
400
570
220
400
570

5%
10%
15%

30
65
90

75
150
210

80
170
240

85
180
270

chlorine or sulphur content in the substances stored is

less than 10%.
The next part of this section describes how table 3
was compiled.
The effect of the toxic combustion compounds
released from a warehouse fire will be significantly
reduced by the new proposed conversion factor for
nitrogenous substances. Since the frequency has not
been changed, the risk (frequency effect) of these
fires has inevitably become lower. The result is smaller
individual risk 106 /year contours.
The current safety distances only apply to warehouses with hazardous substances containing less than
1.5% nitrogen.
4.1.1 Protection level 1
The objective of the Ministry of Housing, Spatial Planning and the Environment has been to maintain the
current individual risk 106 distances for protection
level 1 warehouses. The relatively small safety distances for these storage facilities hardly ever give rise
to problems in land use planning. For many protection
level 1 firefighting systems, the current applied safety
distances are reached at a nitrogen content of 15%.

Since average nitrogen contents of more than 15% are

not likely, the chemical composition of the hazardous
substances stored has become irrelevant. Therefore,
for risk analysis, the administration of the composition of the materials stored, by companies as well as the
competent authorities considered as inconvenient, has
become unnecessary for almost all protection level 1
warehouses.
4.1.2 Protection level 2 and 3
For protection level 2 and 3 firefighting systems,
the current applied safety distances are reached at a
nitrogen content of 510% and 1015%. For most
protection level 2 and 3 warehouses the individual
risk 106 distances are decreased significantly by the
reduced conversion factor for nitrogenous substances,
with regard to the current distances for 1.5%. This
gives rise to less problems in land use planning.
The applied safety distances depend on the nitrogen content of the substances stored as well as the
storage area. The table discriminates between three
levels, 5%, 510% and 1015%. Contrary to protection level 1 warehouses, most protection level 2
or 3 warehouses still need to report their maximum

1196

nitrogen content of the substances stored in a permit

request. For companies with many different hazardous
substances in packaging and where the average composition changes daily (e.g. the transport sector),
a fixed nitrogen content of 10% is proposed. This
value can be used to determine the applied safety distance in Table 3 or for a QRA. Nevertheless, for all
types of chemical warehouses a QRA should always
be performed if desirable.
Finally the individual risk 106 /year distances for
protection level 2 warehouses with flammable liquids show a discontinuity between 300 and 600 m2 .
Since 100 tonnes flammable liquids are allowed to
be stored under these conditions, a higher burn rate
than 0, 025 kg/m2 s has been used for 100 and 300 m2
warehouses, which results in higher individual risk
106 /year distances.
4.2

Non-combusted products

If certain quantities of extremely toxic substances like

pesticides/herbicides are present in a warehouse the
risk is determined by the release of non-combusted
materials rather than the release of toxic combustion
substances. Therefore table 3 is applicable for storage facilities with less than 5000 kg of ADR class
6.1 packaging group I stored or less than 50,000 kg
of packaging group II. For SEVESO-II establishments, for which the table is not applicable, a QRA
is obligatory.
Figure 1 shows an example of the influence of
non-combusted extremely toxic substances (group I)
on the calculated individual risk 106 /year distances
for a 2500 m2 warehouse with an automatic sprinkler installation. Above a certain threshold quantity
this contribution increases rapidly to large safety distances. Since the threshold contribution for ADR class
500

Individual risk 10-6/year distances [m]

450
400
350
300

sf 1%
sf 10%
sf 30%

250
200
150
100
50
0
1%

10%

100%

Amount of extremely toxic substances stored [mass%]

Figure 1. Individual risk 106 /year distances for various

amounts of non-combusted extremely toxic substances (ADR
class 6.1, packaging group I) stored in a 2500 m2 warehouse
with automatic sprinkler installation and a survival fraction
1%, 10% and 30%.

6.1 group II substances is relatively high, the noncombusted release of these substances hardly ever
dominates the risk.
5

CONCLUSIONS

Following an evaluation, several major changes were

proposed to the current QRA method for warehouses
with packaged hazardous materials. In particular, the
adaptation of the conversion factor for nitrogenous
substances to NOx from 35% to 10% leads to smaller
consequences of a warehouse fire and, therefore,
smaller individual risk 106 distances for almost all
warehouses. A possible increase in the safety distances
under certain storage situations with respect to the current calculated distances only occurs if extremely toxic
substances like pesticides/herbicides (ADR class 6.1
packaging group I) are stored.
The revised QRA method was proposed to the Ministry of Housing, Spatial Planning and the Environment in April 2008. The Ministry need to authorise the
method before it can be incorporated in the redrafted
Dutch guideline for Quantitative Risk Assessment.
REFERENCES
Atkinson G.T. & Jagger S.F. 1992. Exposure of organophosphorus pesticides to turbulent diffusion flames, J. Loss
Prev. Process Ind. Vol 5, No 5 (1992) p. 271277.
Atkinson G.T. & Jagger S.F. 1994a. Assessment of hazards
from warehouse fires involving toxic materials, Fire Safety
Journal Vol 22 (1994) p. 107123.
Atkinson G.T. & Jagger S.F. et al 1994b. Plume Seeding Mechanisms in Warehouse Fires, EC Industrial Fires
workhop, Cadarche May 1718th 1994.
CPR 1999. Committtee for the Prevention of Disasters.
Guidelines for Quantitative Risk Assessment, CPR 18E.
HSE 2002. Safety Report Assessment Guide: Chemical
WarehousesHazards, HSE version 6.
Hymes I. and Flynn J.F. 1992. The Probability of Fires in
Warehouses and Storage Premises, SRD/HSE R578.
Laheij 1995. Studie naar de gevormde hoeveelheid stikstofoxiden bij brandi nopslagen van chemicalin of bestrijdingsmiddelen, RIVM report 610066003 (in Dutch).
Marchal 1985. Bouteilles gazAnalyse Statistique des
accidents, IGC Document 31/85/F.
Matthijsen A.J.C.M. & Riedstra D, Omzettingspercentage
van stikstofhoudende verbindingen naar NOx en HCN bij
brand, RIVM letter 191/07, 2007 (in Dutch).
Molag M. & Blom-Bruggeman J.M. 1991. Onderzoek
naar de gevaren van de opslag vanbestrijdingsmiddelenriscoanalysemethodiek, TNO report 90424 (in
Dutch).
RIVM 2008. Handleiding Risicoberekeningen BEVI, versie
3.0 (in Dutch).
VROM 1997. Risico-analyse methodiek CPR-15 bedrijven,
VROM (in Dutch).
VROM 2005. Publicatiereeks Gevaarlijke Stoffen 15, Opslag
van verpakte gevaarlijke stoffen, VROM (in Dutch).

1197

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Ranking the attractiveness of industrial plants to external acts of interference

M. Sabatini & S. Zanelli
Dipartimento di Ingegneria Chimica, Chimica Industriale e Scienza dei Materiali, Pisa, Italy

S. Ganapini, S. Bonvicini & V. Cozzani

Dipartimento di Ingegneria Chimica, Mineraria e delle Tecnologie Ambientali, Alma Mater Studiorum, Bologna, Italy

ABSTRACT: The concern about the possibility of malicious acts of interference involving process plants
greatly increased in the last years, since the industrial sites may represent a very attractive target. The methods
proposed for the evaluation of the vulnerability of a plant to external attacks generally lack of quantitative criteria,
dealing mainly with security-related issues, and not affording in depth the technical aspects of the problem. The
aim of the present work was the development of a method for the quantitative evaluation of the attractiveness
of a plant with respect to the possible consequences of external attacks, based on the characteristics both of
the industrial site and of the surroundings area. A set of indices was defined to quantify these hazard factors.
The final aim of the procedure was to provide criteria and priorities for emergency planning and for protection
actions.
1

INTRODUCTION

Consequences of attacks or acts of interference involving industrial sites where relevant quantities of dangerous substances are processed or stored may result in
very severe consequences. Several Security Vulnerability Assessment (SVA) methods were developed to
analyze the problem (SFK, 2002; API-NPRA, 2003;
Uth, 2005; Moore, 2006). However, all the SVA
procedures deal mainly with security and target
vulnerability, but not much with safety and do not
take into account the evaluation of the possible consequences of acts of interference. This is a crucial point,
because the aim of an external attack is usually to
obtain the most severe consequences as possible: in
order to rank the attractiveness of a plant the vulnerability of the surrounding areas can not be neglected,
because higher is the number of people affected, higher
is the attractiveness of the target.
In the following, a method to evaluate the hazard
posed by an intentional act of interference is presented. A procedure based on a two-step assessment
was developed. The first part of the procedure was
based on the assessment of the attractiveness of the
target. This was based on the screening of the hazard of the site, the distribution of the population and
of aggregation points for population in the surroundings of the plant were taken into account by the use
of indexes. The combination of the indexes allowed a
ranking of the plant attractiveness.

The second part of the procedure afford a more

detailed assessment of the potential damage distances
of an attack, based on the characteristics of the facility
and the likeliness of various types of attack.

2
2.1

EVALUATING THE ATTRACTIVENESS

OF A PROCESS PLANT
Data required for the assessment

As stated above, the aim of an attack is to cause

an important damage. This is possible if i) Accidental scenarios triggered by the act of interference are severe, and if ii) The damage areas of
the triggered scenarios impact on densely populated
zones, thus potentially involving a relevant number of
persons.
Starting from this consideration, in the present
study a method to quantitatively assess the attractiveness of a plant was developed. A set of indexes was
defined to characterize both the vulnerability of territory in the surroundings of the plant, and the plant
hazard. The combination of the index values allowed
the ranking of the attractiveness of the targets. Three
levels were defined: high, medium and low.
The indexes may be calculated using information on
population distribution in the surroundings of the site.
A GIS tool may be used to implement the data and to
support their elaboration.

1199

A large number of plants were analyzed, in order

to test the use of these indexes and to better calibrate
their values.

score, named ITP in the following is obtained by this

procedure.
2.3 Vulnerability of the potential impact area

2.2

Site hazard indexes

The starting point to evaluate plant attractiveness is to

know the types and the quantities of the hazardous
materials present in the site. These data are available for the plants that fall under the obligations of
the EC Seveso Directive (Directive 96/82/EC and
following amendments).
In order to evaluate the hazard posed by the plant, it
may be convenient to schematize it using a normalized
substance or substance category hazard index, defined
as follows:
Ii =

Wi
Ti

(1)

where Wi is the quantity of material i (or of all substances in the i-th category, as defined by Annex 1
of the Directive), and Ti is the corresponding threshold reported in the Directive (the thresholds may be
provided for categories of substances or for named
substances in the Annex I of the Directive). Obviously,
for at least one substance or substance category this
index will be higher than 1, if the site falls under the
obligations of the Directive.
This index has to be calculated for each substance
or substance category present in the plant.
A simple sum of the indexes for all the substances
or substance categories present provides an overall
site value: the so-called site hazard index. It is
convenient to sum separately toxics and flammable
substances, thus defining flammable hazard and a
toxic hazard site indexes:


Isub =
Iifl +
Ijtox
(2)
i

The site hazard index, Isub , may be used to rank

plant attractiveness. Table 1 shows the 6 classes of
attractiveness defined in the present study (from 1 to
6 decreasing the attractiveness). A site attractiveness
Table 1. Categories of attractiveness based
on the site hazard index, Isub .
Isub

ITP

<10
11 50
51 150
151 300
301 650
>650

6
5
4
3
2
1

Clearly enough, beside the hazard posed by the site, the

attractiveness also depends on the vulnerability of the
potential impact area, that thus needs to be determined.
The first step of the procedure is the identification
of the potential impact area. The site characteristics
may be also useful to identify the extension of the
area surrounding the site that should be considered for
vulnerability analysis.
If only a very limited number of accidental scenarios are possible for the site, a detailed analysis may
be carried out to quantify the potential impact area
of the scenarios triggered by acts of interference (see
section 3). However, in complex sites a preliminary
screening usually is needed to understand the potential hazard deriving from the scenarios. In the stage
of hazard screening, the potential impact area may
be approximated. If only flammable substances are
present on the site, an impact distance of 1 km may
be assumed. If also toxic and volatile substances are
present, an impact area of 7 km may be considered.
The impact radii assumed in the analysis are conservative damage distances derived from the experience
and from the analysis of past accidents. However, it
must be considered that these reference distances are
adopted only to limit the extension of the area that
needs to be analyzed for vulnerability characterization.
These assumptions thus do not affect the calculation
of damage distances or the consequence analysis of
the scenarios (see section 3).
The vulnerability of the surrounding area may be
related to the number of people present. This parameter is correlated to the population density, that may be
derived from census data, but also to the presence of
the so-called vulnerability centres. These are sites in
which the population density is supposed to be higher
than in normal residential areas: hospitals, commercial
centres, schools are examples of these vulnerability
centres. These sites are usually mostly critical in the
case of an emergency, due to the high number of people
that might need to be evacuated. However, these sites
become important in zones in which the population
density is low: when the population density is high,
the presence of aggregation sites has a lower influence
on the vulnerability. For this reason, in the present
framework the vulnerability centres were considered
only when the population density was lower than 2000
inhabitants per square kilometre.
In order to speed-up the screening, it is mostly useful to have these data available on form of GIS layers,
in order to represent and analyze them with a GIS
tool (the commercial software ArcGis was used in the
present framework).

1200

Table 2.

Table 4. Ranking of attractiveness basing on IA values.

Calculation of IRes .

Population in the potential

impact area (inhabitants)

IRes

IA

Attractiveness level

<1000
1000 10000
10000 500000
>500000

4
3
2
1

36
7 10
11 14

High
Medium
Low

Table 3.

Calculation of Icv .

Number of vulnerability centers

in the potential impact area

ICV

<2
2 10
11 50
>50

4
3
2
1

A vulnerability index of the potential impact area

may be defined considering two parameters:
The number of persons present in the impact area
(calculated on the basis of the census data in the
potential impact area, e.g. using a proper GIS
extension), expressed by the index Ires
The number of the vulnerability centres in the
impact areas with low population density, expressed
by Icv
Tables 2 and 3 show the criteria used for the calculation of Ires and Icv . The sum of these indexes
yields a global index, Itv , expressing the territorial
vulnerability for the considered area:
ITV = Icv + IRes

(3)

2.4 Assessment of overall site attractiveness

Once having calculated the site hazard index, ITP , and
the impact area vulnerability index, ITV , it is possible
to combine them in order to express the overall attractiveness of the site. An overall site attractiveness index,
named IA in the following, is obtained summing the
site hazard and area vulnerability indexes:
IA = ITV + ITP

Figure 1. Scheme of the procedure for calculate plant

attractiveness.

EVALUATING THE CONSEQUENCES

OF AN EXTERNAL ACT OF INTERFERENCE

3.1 Definition of the type of interference

First of all, in order to estimate the consequences of
an external act of interference it is important to analyze the types of interferences that may be performed.
The literature analysis lead to the identification of
10 kinds of attacks (listed in Table 5), each characterized by three parameters useful to estimate the
likeness and the severity of the damages: the potential impact vector (i.e. the physical effect or the action
that is responsible of the damage to the process or
storage equipment units: e.g. radiation, overpressure,
projectile penetration or other types of effects), the
expected release rate. Four categories of release rate
were considered, defined on the basis of the approach
provided by the Purple Book (Uijt de Haag 1999),
ranging from R1 to R4 with the increasing severity of
the release and the required level of knowledge necessary to conceive, plan and perform the malicious act
of interference (3 levels ranging from A to C as the
necessary information required increases).

(4)

In order to understand more clearly the level of

attractiveness, three categories may be defined on the
basis of the overall index, IA . Table 4 shows the criteria used for the definition of three attractiveness
categories: low, medium and high.
Figure 1 summarizes the procedure used to calculate the overall attractiveness index.

3.2 Target attractiveness

As it was said previously, the aim of an intentional
act of interference is to obtain the most severe consequences. That is, the devices giving the higher
damage distances are the most attractive and critical
with respect to the risk of an intentional attack. In the
present work a classification of the various process

1201

Table 5. Characterization of the ten categories of possible

acts of interference identified, using the parameters identified
in the present study.

Type of
interference
Deliberate
misoperation
Interference by
simple means
Interference by
major aids
Arson by simple
means
Arson by incendiary devices
Shooting
(minor)
Shooting
(major)
Explosives
Vehicle impact
Plane impact

Required
level of
information

Expected
release rate
(atmosp.
Eq.)

Expected
release
rate (pressur. Eq.)

R2

R1

R2

R1

R3

R2

R3

R2

R4

R3

R1

R1

R4

R4

B
B
A

R4
R3
R4

R4
R3
R4

Figure 2. Layers of protection in case of shooting minor

interference (see Table 1).

3.3 Credible acts of interference

Table 6. Ranking of the attractiveness of process

equipment.
Equipment type
Large
Column Reactors/
Physical conditions
diameter type Heat exchof inventory
Tanks pipelines eq.
angers
Liquefied gas
stored under
pressure
Fluids with low
vapor pressure
stored in liquid
phase
Gas/liquid stored in
gas phase
Cryogenic storage
Liquid phase

2
1

2
1

2
1

1
1

A successful act of interference is to be intended as

an act that actually causes a release of hazardous substance. The probability of success thus depends on the
effective protections present in the plant. A protection
can be effective or not with respect to specific modes of
interference, depending on the damage vectors associated to the interference: e.g. for Arson, radiation
is the damage vector at which the protections must
be effective. Analyzing several kinds of possible acts
of interference from this point of view allowed the
achievement of two main results:
protections actually used in a site may be organized
in layers of protection, from which the estimation
of the probability of success of the attack may be
assessed
other protection devices may be identified, that may
be useful in resisting to the attacks.
In Figure 2 a representation of the layers of protection for a given interference mode is shown, while
Table 7 reports some proposed correlations for the
evaluation of damage probabilities.
3.4 Identification of accidental scenarios
and consequence assessment

equipment was produced, based on several characteristics: the type of hazard associated to the substances
present (flammable, toxic or both), the physical conditions (influencing the behaviour of the substance after
the release: formation of a toxic cloud, flash, vaporization, liquid spread, and so on) and the amount of
substance present in the unit (which may be considered
a function of the type of device: given the equipment
volume, e.g. a tank may have a higher hold-up than a
column). Table 6 reports this attractiveness index.

On the basis of the previous considerations, for each

device and each kind of interference an expected
release rate may be estimated. The evaluation of
accidental scenarios and consequences can be carried on with conventional methods and procedures as
event trees and physical effects correlations. Attention
must be dedicated to the identification of accidental
scenarios for some peculiar attacks, as Arson or
Explosives interference (see Table 5): the characteristics of the act of interference may affect the likeliness
of some scenarios, since, for example, ignition may be
more probable than in other situations.
Domino effect may also be evaluated by this
method, using proper threshold values for physical

1202

Table 7. damage vectors and proposed correlations for

evaluating success probability of an attack.
Act of
interference

Damage
vector

Damage
correlation

Deliberate
misoperation
Interference by
simple means
Interference by
major aids

n.d.

Not available in
the form of
mathematic
expression

Arson by simple
means
Arson by
incendiary
devices
Shooting (minor)
Shooting (major)
Explosives

radiation

Vehicle

impact

Plane impact

impact

n.d.
n.d.

radiation

Table 8.

Plants characteristics for case study.

Name

Impact radius
to be considered (m)

Plant A
Plant B
Plant C
Plant D
Plant E
Plant F

7000
7000
1000
1000
7000
1000

640
599
0
0
30
27

Ifl


50
60
50
30
0
0

Itox

Isub

ITP

690
659
50
30
30
27

1
1
5
5
5
5

Equipment
vulnerability
model (Uijt de
Haag et al. 1999)

impact
impact
overpressure
Specific studies or
simplified vulnerability models
(Susini et al. 2006)

effects, since it would increase significantly damage

distances.

4
4.1

Figure 3. Collocation of plants (crosses) in the case study

area. Shaded areas are high populated zones; white areas have
low population density (less than 2000 inhabitants/square
km). Small circles correspond to vulnerability centers.

CASE STUDY
Part 1: Ranking the attractiveness

In order to validate the procedure and to better understand the results obtained by its application, the
method was applied to a case study.
In the case study, real data were used for the population. These refer to an industrial area of an Italian
region and are supplied by Italian government. Also
the plants considered in the area actually exist, but their
position is completely different from that considered
in the case study (data from competent authorities of
a different region of Italy were used).
In Table 8 the characteristics of the industrial sites
considered are reported, while Figure 5 shows the
position considered in the case-study.
As shown by Table 8, the site hazard index is quite
high for two of the plants considered. This means that
the quantities of hazardous materials present in these
sites are far above the thresholds for the application of
the Seveso-II Directive.
On the other hand, plants A, E and C are quite close
to populated zones (shaded areas in Figure 3), while
the other plants are farer from these areas.
In order to calculate the vulnerability index of the
potential impact areas, a specific GIS extension was

developed. This allowed the calculation of the number

of inhabitants in a given circular area, on the basis of
the implemented density data.
Figure 4 shows the potential impact areas of some of
the sites, while Table 9 reports the potential impact area
vulnerability indexes calculated for all the six plants
considered in the case-study.
The results show that, even if site E has a less
critical ITP , due to a limited inventory of hazardous
substances, it has a quite high IVT , because its impact
area involves highly populated zones.
Table 10 reports the overall attractiveness index IA
values and the corresponding attractiveness levels for
the six plants.
The results of the analysis show that while plants A
and B have high overall attractiveness due mainly to
the large inventories and not to the area vulnerability,
plant E, even if its inventory is not very large, has a
high value of the vulnerability index of the potential
impact area, since it is near to populated zones, and its
global attractiveness is higher than the ones of other
similar plants (as D or F).

1203

of consequences of external attack was made for this

site, that consisted in a section of a tank farm of a
petrochemical industry. The lay-out considered for
the case-study is reported in Figure 5. As shown in
the lay-out, the presence of several atmospheric tanks
(AT_Xs), pressurized tanks (PVA_Xs and PVB_Xs),
a spherical LPG tank (SF_F) and a buried LPG tank
(TUM_F) were considered.
The procedure defined above was applied to all the
tanks present in the lay-out. The attractiveness rank
of each tank was determined, and expected release
rates were estimated. In a following step, conventional
methods and commercial simulation software were
used to assess worst case scenarios and to calculate

Figure 4. Impact areas of plants E, A and C. It is evident that impact area of plant E involves highly populated
areas and several vulnerability centers, while, for example,
consequences of an accident in plant C impact on scarcely
populated areas.
Table 9.
values.

Area vulnerability data and corresponding index

Plant

Inhabitants

Number of
vulnerability
centers

A
B
C
D
E
F

225832
23916
1640
0
386538
0

4
12
0
0
210
0

IRes

ICV

IVT

2
2
3
4
2
4

3
2
4
4
1
4

5
4
7
8
3
8

Table 10. Overall site attractiveness index calculated for the

six plants considered.
Plant

ITP

IVT

IA

Attractiveness level

A
B
C
D
E
F

1
1
5
5
5
5

5
4
7
8
3
8

6
5
12
13
8
13

High
High
Low
Low
Medium
Low

4.2

Part 2: Detailed analysis of possible

consequences

In order to test the procedure developed, the case study

was further developed, examining in detail one of
the sites considered in the preliminary screening and
resulting in a high attractiveness level. The evaluation

Figure 5.

Layout of the considered case study.

Table 11.

Worst-case scenarios and distances.

Act of interference

Device

Exp.
release
rate

Deliberate
misoperation
Interference by
simple means
Interference by
major aids
Arson by simple
means
Arson by incendiary
devices
Shooting (minor)
Shooting (major)
Explosives
Vehicle impact
Plane impact

TUM

R1

448

TUM

R1

448

TUM

R2

448

AT

R2

260

PV

R3

1048

AT
PV
SF
SF
SF

R1
R4
R3
R3
R4

92
1048
6118

6118

1204

Damage
distance
(m)

Figure 6. Worst case damage distances for the considered

plant, including surrounding vulnerability (increasing from
F to A). Dot line refers to spherical tank, continuous lines
regard pressurized tanks, wavy line refers to buried tank.

the damage distances for each scenarios. The results

obtained are reported in Table 11.
As expected, the pressurized spherical LPG storage
tank comes out as the most critical device, since it has
the higher damage distances, in particular for the more
severe acts of interferences (Explosives and Plane
Impact, see Table 5).
Due to the large impact areas, the consequences
on the surrounding areas are of particular importance.
The territory around the site was classified on the
basis of Italian criteria for land use planning. These
criteria allowed the classification of the area around
the plant on the basis of population density and type
and number of aggregation centers for the population.
A classification ranging from A to F, decreasing in
vulnerability, is thus used. Figure 6 shown the results
of this activity.
The large impact areas due to the spherical tanks
involved also high vulnerability zones: from this consideration arises the concept of attractiveness as a
function both of the inherent hazard of the plant
itself and of the vulnerability of surroundings.

CONCLUSIONS

In the present study a method to estimate plant vulnerability towards external acts of interference is presented.
The method consists in a two-step assessment. In the
first step, a preliminary screening is carried out, by
the use of indexes quantifying the inherent hazard
due to the site and the vulnerability of the potential
impact area. In a second step, a detailed analysis may
be carried out, in order to refine the analysis and to
calculate more in detail the impact areas of the potential scenarios that may be triggered by external acts of
interference.
The methodology was applied to the analysis of a
case-study, based on realistic data. The results allowed
a preliminary ranking of the attractiveness of the targets, and a detailed analysis of the potential areas of
impact of scenarios that may be triggered by acts of
interference. The methodology thus provides criteria
to orient more detailed analysis and provides criteria to prioritize emergency planning, protection and
prevention actions.
REFERENCES
American Petroleum Institute, National Petrochemical &
Refinery Association 2003. Security Vulnerability
Assessment Methodology for the Petroleum and Petrochemical Industries.
Moore D.A. 2005. Application of the API/NPRA SVA
methodology to transportation security issues. Journal of
Hazardous Materials 130: 107121.
Str-fall Kommission (SFK) 2002. SFKGS38, Report
of the German Hazardous Incident Commission.
Susini A. 2006. Analisi del rischio connesso alla proiezione di
frammenti in strutture off-shore. M. D. Thesis, University
of Pisa, etd-06232006-101740.
Uth H.-S. 2005. Combating interference by unauthorised persons. Journal of Loss Prevention in the process industries,
18: 293300.
Uijt de Haag PAM, Ale BJM, 1999. Guidelines for Quantitative Risk Assessment (Purple Book). The Hague, The
Netherlands: Committee for the Prevention of Disasters.

1205

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Review and discussion of uncertainty taxonomies used in risk analysis

T.E. Nkland
University of Stavanger, Stavanger, Norway
IRIS, International Research Institute of Stavanger, Norway

T. Aven
University of Stavanger, Norway

ABSTRACT: Uncertainty is a key concept in risk assessment, but its meaning varies a lot between different
risk assessments schools and perspectives. This causes confusion. Various attempts have been made to structure the uncertainties and in this paper we review and discuss some of the existing taxonomies. Some restrict
uncertainty to lack of knowledge about unknown quantities. Others distinguish between stochastic (aleatory)
uncertainty and epistemic (lack of knowledge), and distinction is being made between knowledge uncertainty,
modelling uncertainty and limited predictability/ unpredictability. More comprehensive structures also exist.
For example, in the well-known Morgan and Henrion classification system it is distinguished between uncertain
types of quantitiessuch as empirical quantities, constants, decision variables, value parameters and model
parametersand sources of uncertainty, such as random error and statistical variation, systematic error and
subjective judgment, linguistic imprecision, variability, randomness and unpredictability, disagreement and
approximation. In the paper we introduce a structure to group the taxonomies and the uncertainty categories.
We argue that it is essential to clarify ones perspective to risk and probability to prevent confusion.

INTRODUCTION

Risk analyses are performed in many areas of our

society. They relate to accident risk, economic risk,
environmental risk, health risk, project risk, societal
risk etc. The risk analyses are performed in numerous
ways with methods varying from simple checklists to
complex computer assisted calculations. The intention of performing a risk analysis is to provide insight
about risks related to an activity or a given system in
order to support decisions about possible solutions and
measures (Aven 2008).
Uncertainty is an important concept in risk and risk
analysis. The word uncertainty has many synonyms
such as e.g. ambiguity, anxiety, confusion, guesswork,
unpredictability, vagueness to mention a few (Rogets
New Millennium Thesaurus, 2007). Uncertainty is
the opposite of certainty, sureness and definiteness,
and it refers to something that we cannot know for sure.
In a risk analysis context uncertainty relates to occurrence of events, physical quantities, models, model
parameters, risk estimates, assumptions etc. We have
uncertainties in both input and output quantities. To
structure the uncertainties a number of taxonomies
have been suggested. These taxonomies can be seen
as an attempt to categorize and describe the nature of
uncertainty in order to help us interpret and discuss

the results from risk analysis. Different perspectives

exist and a variety of terminology is used to describe
the uncertainties. The key questions raised are:
What type of uncertainty exists?
What are the sources of uncertainty?
How is uncertainty expressed?
The number of uncertainty categories introduced
and discussed in the literature is long and consists
of e.g. aleatory uncertainty (stochastic uncertainty,
randomness, variability, irreducible uncertainty) epistemic uncertainty (uncertainty due to lack of knowledge, reducible uncertainty, subjective uncertainty),
ignorance and ambiguity. Other categories suggested
are metrical, structural, temporal and translational
uncertainty, as well as uncertainty in effect, cause,
and uncertainty in the relationship between cause and
effect.
The list of sources to uncertainty is equally versatile and examples are statistical variation, subjective
judgement, linguistic imprecision, variability, inherent randomness, disagreement, approximation, complexity, measurement error, bias error and systematic
error.
Probability is the most common measure to express
and quantify uncertainty, but there are different views
on how it should be interpreted, applied, and to what

1207

types of uncertainties is should be applied. Other tools

also exist, for example fuzzy theory, Dempster-Shafer
theory of evidence, possibility theory and entropy
measures; however, such tools will not be discussed
in this paper.
The remaining part of the paper is organised as
follows. In section 2 we review some common taxonomies suggested in the literature. In Section 3 we
discuss these taxonomies, specifically looking at differences and similarities. Based on this review and
discussion we suggest in Section 4 two structures to
group the taxonomies and the uncertainty categories.
Section 5 provides some concluding remarks.
2

UNCERTAINTY TAXONOMIES

A number of uncertainty taxonomies have been suggested in the literature, and in this section we give a
brief review of some of these:
A. Epistemic uncertainty vs. aleatory uncertainty
B. Knowledge uncertainty, modelling uncertainty and
limited predictability
C. Sources of uncertainty in empirical quantities
D. Metrical, structural, temporal and translatoric
uncertainty
E. Uncertainty due to lack of knowledge
Other taxonomies also exist, see e.g. NRC (1994),
Helton (1994), Ferson & Ginzburg (1996), and Levin
(2005), but these will not be presented in this paper.
2.1

Epistemic vs. aleatory uncertainty (A)

A common taxonomy of uncertainty in risk analysis is

the division of uncertainty due to epistemic uncertainty
and aleatory uncertainty (e.g. Pat-Cornell 1996,
Winkler 1996, Hora 1996 et.al.). Aleatory uncertainty stems from variability in known (or observable)
populations, and represents randomness in samples.
Other words that have been used to describe aleatory
uncertainty are variability, stochastic variability and
interindividual variability. Aleatory uncertainty is
considered not reducible through further measurement
or more research, but may be reduced through changing the physical system. A classic example to illustrate
variability is the inherent randomness in outcomes of
tossing a coin. The proportion of heads to the total
number of tosses represents the aleatory uncertainty.
The aleatory uncertainty is represented by probability
distributions, where the interpretation of probability
is relative frequency based. Say that we observe the
lifetimes of a type of machines. Then the aleatory
uncertainty is represented by the lifetime distribution,
for example the exponential distribution.
Epistemic uncertainty stems from lack of knowledge and is typically expressed as the assessors lack of

knowledge about the parameters that characterise the

physical system that is being modelled, for example the
parameter of the exponential distribution. Uncertainty
is considered reducible through further measurement
or studies, or through consulting more experts. It is
by definition subjective since it is dependent on the
assessor.
The combination of aleatory and epistemic uncertainty act together and form what Vose (2000) describe
as total uncertainty which erode our ability to predict
what the future hold. Risk analysis based on the distinction between aleatory and epistemic uncertainty
is often referred to as the probability of frequency
approach (Kaplan and Garrick 1981, Aven 2003) and
the classical approach with uncertainty analysis (Aven
2003).
2.2

Knowledge uncertainty, modelling uncertainty

and limited predictability (B)

HSE (2001) has published the document Reducing

Risk, Protecting People describing the official process for decision making of the UK safety authorities.
The document includes an uncertainty taxonomy and
advice on how the different manifestation (categories)
of uncertainty should be treated.
First uncertainty is separated from ignorance
defined as being a lack of awareness of factors influencing the issue. A result of ignorance in risk assessment is the fact that the identification of hazards may
be incomplete.
HSE defines uncertainty as a state of knowledge
in which, although the factors influencing the issue
are identified, the likelihood of any adverse effects or
the effects themselves cannot be precisely described.
Three manifestations of uncertainty is described and
explained how they affect the approach for handling.
The three manifestations are (HSE 2001);
Knowledge uncertaintyThis arises when knowledge is represented by data based on sparse statistics
or subject to random errors in experiments. There
are established techniques for representing this kind
of uncertainty, for example confidence limits. The
effect on risk assessment can be estimated by sensitivity analysis. This provides information relating to the
importance of different sources of uncertainty which
can then be used to prioritise further research and
action, which is the only feasible way to address the
uncertainty, though in some cases research may not
be technically possible or cost-effective.
Modelling uncertaintyThis concerns the validity
of the way chosen to represent in mathematical terms,
or in an analogue fashion, the process giving rise to
the risks. An example is the growth of a crack in the
wall of a pressure vessel. The model would postulate
the way the growth rate is affected by factors such as

1208

the material properties and the stress history to which

the vessel is exposed in service. The model will provide
prediction of failure in terms of time and the nature
of the failure. It will inform intervention strategies
such as the material specification, in-service monitoring and mitigation measures. All these factors may be
modelled in many ways with the assumptions for each
one open to question. The rigour of the peer review
process and openness to alternative hypotheses are
the main safeguards. However, the most intractable
problems arise when it is not practical or physically
possible to subject the alternative hypothesis to rigorous testing. In such cases, the exercise of expert
judgement is paramount and confidence depends on
the procedures adopted for selection of the experts and
the management of bias (or appearance of bias).
Limited predictability or unpredictability
There are limits to the predictability of phenomena when the outcomes are very sensitive to the
assumed initial conditions. Systems that begin in the
same nominal state do not end up in the same final
state. Any inaccuracy in determining the actual initial state will limit our ability to predict the future
and in some cases the system behaviour will become
unpredictable.

Table 1. Summary of types of quantity (Morgan & Henrion

1990).

2.3

sources and present seven sources in their taxonomy.

The appropriate method to characterize the uncertainty, and the appropriate method for trying to reduce
it, generally depends on the particular kind of source.

Sources of uncertainty in empirical

quantities (C)

Morgan & Henrion (1990) state that subjective probability distributions often are a good way to express
uncertainty, but emphasize that this holds for only a
certain type of quantities used in risk analysis. They
refer to such quantities as empirical or chance quantities, representing properties or states of the world.
However there are other types of quantities that play a
role in a risk analysis model, including decision variables, value parameters and others. In case these other
quantities are unknown (or not determined) at the time
of the analysis they should be treated as parameter or
switchover, i.e. defined. See Table 1.
The only uncertain quantity found appropriate to be
treated probabilistic is empirical quantities. Empirical
quantities represent measurable properties of the realworld systems being modelled, i.e. they will, at least
in principle, be measurable now or at some time in the
future. The reason why empirical quantities are the
only type of quantity whose uncertainty may appropriately be represented in probabilistic terms is that
they are the only that is both uncertain and could be
said to have a true, as opposed to an appropriate or
good value. The list of quantities is context dependent, which means that e.g. what is decision variable
or value parameter in one analysis can be empirical
quantity in another.
Morgan and Henrion state that uncertainty in empirical quantities can arise from a variety of different

Type of quantity

Example

Empirical
Thermal efficiency,
parameter or
oxidation rate, fuel
chance variable
price
Defined
Atomic weight, ,
constants
joules per
kilowatt-hr
Decision
Plant size (utility),
variables
emission cap
Value parameters Discount rate,
value of life,
risk tolerance
Index parameters Longitude and
latitude, height,
time period
Model domain
Geographic region,
parameters
time horizon,
time increment
Outcome criteria Net present value,
utility

Treatment of
uncertainty
Probabilistic,
parametric or
switchover
Certain by
definition
Parametric or
switchover
Parametric or
switchover
Certain by
definition
Parametric or
switchover
Determined by
treatment of
its inputs

2.4 Temporal, structural, metrical,

and translational uncertainty (D)
According to Rowe (1994) there is more information
that we do not know than we do know for making most
critical decisions involving risks, thus our focus should
be on understanding and effectively deal with what we
do not know. Rowe presents a classification of four
types (dimensions) of uncertainties and discusses their
respective sources. See Tables 3 & 4. All four classes
occur in any situation, but one or more dominate.
Parameters of unknown information show what
information is unknown and uncertain for each uncertainty class. Discriminator parameters are used to
discriminate between different levels of uncertainty.
Valuation parameters indicates to what degree the
uncertainty has been successfully addressed, while the
last parameter set give primary methods used to handle the uncertainties. Temporal uncertainty (future)
Future uncertainty is values by luck. We consider
desirable outcomes as lucky and undesirable ones
as unlucky when constructing probability models of
possible future events. Temporal uncertainty (past)
Past information is subject to different interpretations
as requirements change or political correctness is

1209

Table 2. A list of sources from which uncertainty in empirical quantities can arise (Morgan & Henrion 1990).
Source of uncertainty

Description

Random error and statistical

variation

Imperfections in the measurement instrument will inevitably give rise to

variations from one observation to the next. The resulting uncertainty
depends on the size of variations between observations and the number
of observations taken. Techniques for quantifying this uncertainty:
standard deviation, confidence intervals, and others.
Systematic error is defined as the difference between the true value of the
quantity of interest and the value to which the mean converges as more
measurements are taken. It arises from biases in the measuring apparatus
and experimental procedure. It is not reduced by additional observations.
Subjective judgement of errors that are essentially unknown at the time is
difficult.
In everyday conversation and even professional writing, we refer to
events or quantities with imprecise language that would not pass the
clarity test. The clarity test is a conceptual way to sharpen up the
notion of well-specifiedness. Imagine a clairvoyant who could know
all facts about the universe, past, present and future. Given the description
of the event or quantity, could she say unambiguously whether the event
will occur (or had occurred), or could she give the exact numerical value
of the quantity? If so, it is well specified. There is no successful techniques
to eliminate it, but clear specification of events and quantities
reduce linguistic imprecision.
Many quantities are variable over time or space. Populations of such
quantities can be described with frequency distributions. It is possible to
have a high degree of certainty about frequency distributions. A common
mistake is failure to distinguish between variability due to sampling from
a frequency distribution and empirical uncertainty that arises from
incomplete scientific or technical knowledge.
Inherent randomness is sometimes distinguished from other kinds of
uncertainty in being irreducible even in principle, ref Heisenberg
uncertainty principle. Even if there were actually a well-defined pattern
that could enable us to predict values in principle, to do so in practice
requires knowledge of the pattern.
Disagreements between scientific experts are often an important source
of uncertainty in risk and policy analysis. Disagreement arises from
different technical interpretations of the same scientific evidence, and
is often complicated by the fact that the people view the problem from
very different perspectives. Common approach is to combine opinions
using weights.
Approximation uncertainty arises because the model is only a simplified
version of the real-world system being modelled. Each approximation
represents a trade-off between the precision of the representation and
the computational costs of the model. Often a feasible approach in
estimating the approximation uncertainty is to increase the resolution
of scope, for example reducing grid size or time steps to examine the
effects on the results.

Systematic error and subjective

judgement

Linguistic imprecision

Variability

Inherent randomness

Disagreement

Approximations

Table 3.

Parameters of the classes of uncertainty (Rowe, 1994).

Uncertainty class

Unknown information

Discriminator parameter

Valuation parameter

Method

Temporal
Temporal
Structural
Metrical
Translational

Future
Past
Complexity
Measurement
Perspective

Probability
Historical data
Usefulness
Precision
Goals/values

Luck
Correctness
Confidence
Accuracy
Understanding

Prediction
Retrodiction
Models
Statistics
Communication

1210

Table 4.

Sources of uncertainty (Rowe, 1994).

Type of uncertainty

Sources of uncertainty

Temporal (future)
Prediction

Uncertainty in future states

Apparent inherent randomness of nature (still an open metaphysical question)
Luck in the short run (unusual combination of outcomes)
Inconsistent human behaviour (we often do the unexpected)
Nonlinear dynamic (chaotic) systems behaviour (highly sensitive to
initial conditions
Individual versus expected values behaviour
Sparse rare events (e.g. low-probability/high consequence events)
Rare events imbedded in noise (e.g. new cancers in a population with
competing causes)
Short versus long time frames for observation (e.g. miss long term-trends)
Changing internal system parameters (e.g. genetic drift)
Uncertainty in past states
Incomplete historical data (measurement error)
Biased history (bias error)
Changing system parameters preventing identical conditions to be revisited
(systematic error)
Hindsight vs foresight
Lack of external data references (validation limited)
Imposition of political correctness (systematic bias)
Conflicting reports (from both the viewing position at events and the
viewpoint of the observer)
Measured parameters inappropriate for the purpose used (lack of correctness)
Uncertainty due to complexity, including models and their validation
Inherent random processes
Inconsistent human behaviour
Number of systematic degrees of freedom (complexity)
Completeness of parameter identification
Interdependence of parameters
Initial conditions in chaotic systems
Differences in judgements and interpretation by experts
Biases, beliefs, and dogma of experts and proponents
Oversimplified models
Inability to verify the validity of alternative models empirically
Selection of models that support pre-established hypotheses
Inappropriate analogies
Uncertainty and variability in measurement
Apparent randomness of nature
Precision of measurements (resolution of instruments)
Accuracy of measurements (quality of measurement)
Measurement interaction (Heisenberg uncertainty principle)
Systematic measurement errors (measurement bias)
Data inclusion/exclusion judgement
Sample size adequacy
Objective vs subjective methods for sampling
Objective vs subjective methods for statistical analysis and reduction
Differences in judgment and interpretation by experts
Biases, beliefs, and dogma of experts (bias error in making observation or
slants in interpretation of results)
Uncertainty in explaining uncertain result
Different approach/perspectives to risk (scientific, regulatory,
design, management)
Value diversity

Measurement

Temporal (past)
Retrodiction

Interpretation of data

Measurement uncertainty
Structural
Systematic fluctuations
Parameter interaction

Interpretation of models
Model choice uncertainty

Metrical
Empirical observations

Interpretation of observations

Interpretation of measurements
Translational

1211

imposed. Structural uncertaintyIs a result of complexity, and depends on both the numbers of parameters used to describe a situation and their interaction as
well as the degree to which models of the complex situation are useful. Metrical uncertaintyMeasurement
uncertainty lies in our inability to discriminate among
values within a parameter (i.e. imprecision). Translational uncertaintyThis differs from the other three
and occurs after the three have been considered. People
have different perspectives, goals, and values as well
as different capabilities and levels of training. These
affect the interpretation of results and, thereby, interact with these interpretations, often contributing to the
uncertainty.
2.5

Uncertainty is due to lack of knowledge (E)

Watson (1994) and Aven (2003) refers to only one type

of uncertainty in a risk analysis context, the lack of
knowledge about unknown quantities, typical physical
quantities such as cost and the number of fatalities.
Hence all uncertainties are epistemic uncertainties.
Variation is not uncertainty for the assessor. Uncertainty is expressed through probabilities (subjective
probabilities), where the value of the probability or the
shape of the probability distribution reflect our uncertainty about the unknown quantity. If the assessor is
to assess the uncertainty of a lifetime of a unit, he/she
may use the variation in populations as background
information to assess the uncertainty about the specific lifetime. However, this variation is not referred
to as uncertainty.
3

SIMILARITIES AND DIFFERENCES

BETWEEN UNCERTAINTY TAXONOMIES

lack of knowledge, while taxonomy B uses the term

knowledge uncertainty to express sources of sparse
statistics and random errors in experiments.
Modelling uncertainty is used in taxonomy B
referring to uncertainty that rise due to concerns about
validity of models. A similar concept is found in
taxonomy D in the structural uncertainty dimension
where several sources of uncertainty are presented,
for example inherent random processes, interdependence of parameters, oversimplified models and
inability to verify the validity of alternative models
empirically. Taxonomy C does not explicitly categorise modelling uncertainty as a type of uncertainty
but refers to approximation and simplification of
models as a source of uncertainty. The taxonomy
E, as described by Aven (2003), considers model
uncertainty (accuracy) as a source of uncertainty in
predictions. A model is a part of the background
information used to assign probabilities and express
risk.
Probabilities and likelihoods are subject to uncertainties in taxonomies A and B. Such uncertainties do
not exist in taxonomy E, as probability is a subjective
measure of uncertainty.
Taxonomy C categorize a type of uncertainty due
to disagreement, which is close to the description
of translational uncertainty in taxonomy D. Linguistic imprecision is only present as a source/type of
uncertainty in taxonomy C.
In Table 5 we have suggested affinity between
the uncertainties terminologies used in the five
taxonomies.

The five taxonomies are different in many respects,

which make comparison a puzzling exercise. A main
challenge relates to the distinction between types/
classes/categories of uncertainties and sources of
uncertainty. For example, taxonomy C uses sources of
uncertainty as labels of types of uncertainty, taxonomy
D introduce dimensions of uncertainty which all have
overlapping sources of uncertainty, while the others
operate with distinctions between type and source.
In taxonomy A aleatory uncertainty is identified
as a type of uncertainty, whereas in taxonomy
C and E variability is identified as a source of uncertainty. Taxonomy D considers variability as a source of
uncertainty in all uncertainty classes except for translational uncertainty. Taxonomy B does not refer to
variability/aleatory uncertainty at all.
The terminology used to describe types (and
sources) of uncertainty varies among the different taxonomies. Epistemic uncertainty in taxonomy A and E
are similar as this type is used for uncertainty due to

TWO STRUCTURES TO GROUP

THE UNCERTAINTY CATEGORIES

To structure the above uncertainty categories, we need

to frame the problem we are concerned about. We will
distinguish between two reference situations:
i. The aim of the analysis is to assess unknown
statistical quantities (parameters) , for example
expressing the expected value EX or a quantile
P(X > x0 ). Probability has a relative frequency
interpretation.
ii. The aim of the risk analysis is to assess an unknown
physical quantity X, for example the number of failures of a system, or the number of fatalities in a
process plant, or the cost of a project. To express
likelihoods, subjective probabilities are used.
In the following two subsections we will clarify
the meaning of key uncertainty categories for these
situations.

1212

Table 5.

Suggested affinity between terminologies used to describe types of uncertainty.

AEpistemic vs.
aleatory uncertainty
Epistemic

BKnowledge
uncertainty, modelling
uncertainty and
limited predictability/
unpredictability
Knowledge uncertainty:
Variation in data in a
classical statistical
setting, for example
expressed by a
confidence interval
for a statistical
parameter

CSources of
uncertainty
in empirical
quantities
Uncertainty about
unknown empirical
quantities

Sources of uncertainty:
Expert disagreement
Systematic error
Random error and
statistical variation
Aleatory

Probability model
uncertainty: A
way of describing
aleatory uncertainty.
A probability model,
such as the exponential
is a model of the
variation in lifetimes
for a population
of similar units

DTemporal,
structural,
metrical, and
translational
uncertainty
Temporal
uncertainty in
future and
past states

Inherent randomness
Approximations
(the model is a
simplified version
of the real world)

Lack of
knowledge

Metrical
(measurement
errors)

Variability

Modelling uncertainty,
this relates to
the accuracy of the
model to represent
the real world.

EUncertainty
is due to lack
of knowledge

Variability,
a source of
uncertainty
Structural
uncertainty, relates
to both accuracy
of models and
confidence in
models

Physical models:
Included in the
background
information
to assess
uncertainties
(assign
probabilities)

Limited predictability/
unpredictability
Linguistic
imprecision
Translational

4.1 Assessing the statistical parameter

As an example we interpret as the expected number of failures for a specific system and a specific
period of time. The expected number we interpret as
the average number if we consider (real or fictional)
an infinite number of similar systems. The variation
in the number of failures in this population of systems
constitutes the aleatory uncertaintythe stochastic
uncertainty. A source for this variation we may refer
to as randomness. The parameter is unknown and
the risk analysis provides an estimate of . This estimate can be based on relevant observed data for this
system or similar systems, and/or expert judgments,
as well as modelling. There is uncertainty related to
this estimate, in the sense that we do not know how
close it is compared to the underlying true value of

. Expressed differently, there are uncertainties about

epistemic uncertaintiesas we do not know the
value of . These uncertainties are a result of lack of
knowledge, and if we had had access to data from thousands of similar systems we would be able to produce
very accurate estimates and reduce the uncertainties
to a negligible level. However, in practice we seldom
have access to such data and we face significant uncertainties in the estimate and about . Different methods
are used to express these epistemic uncertainties. In
a traditional classical statistical setting confidence
intervals are used to express uncertainties due to variation in the data observed. Bayesian methods are also
used. Then also expert judgment can be taken into
account. A posterior probability distribution, given
the observed data, is then established expressing the

1213

epistemic uncertainties about . This leads to the probability of frequency approach to risk analysis. In this
approach frequency refers to EX and probability to
subjective probabilities.
In this context variation in the observed data and
disagreement among experts are sources of epistemic
uncertainties.
Different types of models are introduced to assess
. Consider the following two examples:
a. The number of failures in year i, Xi , is given by a
Poisson distribution with parameter c + d, where
c and d are constants.
b. = p q, where is the number of failure
exposures, p equals the event that barrier 1 fails
(given exposure) and q equals the probability that
barrier 2 fails given that barrier 1 fails.
Model a) assume a linear relation between the
expected number of failures from one year to the next.
Model b) assume the expected number of failures to be
a function of the parameters on a lower level. Model
uncertainty is here an issue. The issue is to what degree
the model is close to the correct values. In case a),
the model uncertainty relates to the deviation between
the model c + d and the true value of EXi , and the
probability distribution given by the Poisson distribution and the underlying true distribution. In case b) the
model uncertainty relates to deviation between pq,
when these values are true, and the true value of .
Both models are simplified descriptions of the system or activity being studied. And approximations are
sources of uncertainty.
If the assumption of the linear form c + d is
wrong, we may refer to this as a systematic error. It
would systematically lead to poor estimates of the EXi ,
even if we have accurate estimates of .
4.2 Assessing the physical quantity X
The physical quantity X, for example the number of
failures of a system, is unknown at the point of assessment, and hence we have uncertainties about its true
value.
To be able to accurately predict X, we introduce
models, denoted g, such that
X = g(Y),
where Y = (Y1 , Y2 , . . . , Yn ) is a vector of quantities on a lower system level, for example representing
lifetimes of specific components of the system.
To describe uncertainties, we use subjective probabilities P, which are conditional on the background
knowledge (referred to as K). This uncertainty is
epistemic, i.e. a result of lack of knowledge.
Let H(y), where y = (y1 , y2 , . . . , yn ) denote the
assigned probability distribution for Y, i.e. H(y) =

P(Y y). Then uncertainty about X is described

through the distribution:

P(X x) =
dP(Y y)
{y:g(y)x}

Consider an event tree model of a process plant, and

let X denote the number of leakages resulting in accident propagation in a specific area, specified by to the
model
X = i Y2i Y3i ,

(4.1)

where the sum is over the number of leakages Y1 , Y2i

equals 1 if the leakage i ignites and 0 otherwise, and
Y3i equals 1 if the ignited leakage results in explosion
and 0 otherwise.
This is a model, which is considered sufficient accurate for its purpose. The models accuracy relative to
the true physical quantities we may refer to as model
uncertainty. When using the model to assign probabilities, model uncertainty is not relevant. The model
is a part of the background knowledge used to assign
the probabilities. Returning to the event tree example,
assume that the probabilistic analysis based on formula
(4.1) produces P(X > 0|K) = 0.02, i.e. a probability of 2% for a leakage occurring resulting in accident
propagation. This probability is conditional on the use
of the model g determined by formula (4.1). Model
uncertainty does not exist with respect to P, as P is a
way of expressing uncertainties given the background
knowledge (including models used).
Variability in data is a source for the uncertainties.
Say that we have observations expressing the number of leakages zj in different years for similar types
of plants, j = 1, 2, . . . , n. Based on these data we
may predict say 1.5 leakages per year, and express a
distribution showing the assessors uncertainties about
Y1 . The data zj are not uncertainthey are observed,
but they provide a basis for assessing the uncertainties
about the future unknown quantity Y1 .
5

CONCLUDING REMARKS

As stated in the introduction, the intention of performing a risk analysis is to provide insight about risks
related to an activity or a given system in order to support decisions about possible solutions and measures.
Various types and sources of uncertainty are associated with the risk analysis and its results. We have
seen that the terminology varies a lot depending on
the taxonomy adopted. For risk analyst and decision
makers, the situation is confusing.
A structure is required. We believe that such a structure best can be obtained by distinguishing between

1214

the two references situations discussed in Section 4.

Some analysts and researchers base their work on a traditional analysis as described in Section 4.1 whereas
others prefer the use of subjective probabilities as discussed in Section 4.2. In order to obtain clarity in
analysis and communication of uncertainty we believe
that it is essential to clarify ones perspective to risk
and probability, as done in Section 4. Mixing these
two perspectives, means chaos.
Using the structures presented in Section 4, we
are able to give substance to the various categories
of uncertainty introduced in different taxonomies.
However, not all type of uncertainties are covered.
Ambiguity, linguistic impression etc. have been left
out of the discussion. In our view they should not be
considered in risk assessments. A quantity addressed
in a professional risk assessment should be welldefined, such that ambiguity and linguistic impression
are removed.
REFERENCES
Aven, T. 2008. Risk analysis. John Wiley & Sons, Ltd.
Chichester.
Aven, T. 2003. Foundations of Risk Analysis: A knowledge
and decision-oriented perspective. John Wiley & Sons,
Ltd. Chichester.
Ferson, S. & Ginzburg, L.R. 1996. Different methods are
needed to propagate ignorance and variability. Reliability
Engineering and System Safety 54 (23): 133144.
Helton, J.C. 1994. Treatment of uncertainty in performance assessments for complex systems. Risk Analysis
14: 483511.

Hora, S.C. 1996. Aleatory and epistemic uncertainty in

probability elicitation with an example from hazardous
waste management. Reliability Engineering and System
Safety 54: 217223.
HSE. 2001. Reducing Risks, Protecting People, HSEs decision making process [R2P2]. Part 3, Pages 2829. HSE
books: http://www.hse.gov.uk/dst/r2p2.pdf.
Kaplan, S. & B.J. Garrick. 1981. On the quantitative
definition of risk. Risk Analysis 1: 1127.
Levin, R. 2005. Uncertainty in risk assessment contents and
modes of communication. Thesis in Philosophy from the
Royal Institute of Technology 8.71 + viii pp. Stockholm.
ISBN 91-7178-175-7.
Morgan, M.G. & Henrion, M. 1990. Uncertainty: A Guide to
Dealing with Uncertainty in Quantitative Risk and Policy
Analysis. Cambridge University Press, Cambridge.
NRC (National Research Council) 1994. Science and Judgement in Risk Assessment. Washington, D.C.: National
Academy Press.
Pat-Cornell, M.E. 1996. Uncertainties in risk analysis: Six
levels of treatment. Reliability Engineering and System
Safety 54: 95111.
Rowe, W.D. 1994. Understanding Uncertainty. Risk Analysis
14(5): 743750.
Rogets New Millennium Thesaurus, First Edition
(v 1.3.1). Retrieved November 18, 2007, from:
http://thesaurus.reference.com/browse/uncertainty
Vose, D. 2000. Risk Analysis: A Quantitative Guide. 2nd
edition, John Wiley & Sons, Ltd, Chichester.
Watson, S.R. 1994. The meaning of probability in probabilistic safety analysis. Reliability Engineering and System
Safety 45: 261269.
Winkler, R.L. 1996. Uncertainty in probabilistic risk
assessment. Reliability Engineering & System Safety
54: 127132.

1215

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk analysis in the frame of the ATEX Directive and the preparation
of an Explosion Protection Document
Alexis Pey, Georg Suter, Martin Glor, Pablo Lerena & Jordi Campos
Swiss Institute for the Promotion of Safety and Security, Basel, Switzerland

ABSTRACT: Directive 1999/92/EC, known as ATEX Directive, lays down minimum requirements for the
safety and health protection of workers potentially at risk from explosive atmospheres. Specifically, its article
8 enforces the employer to ensure that a document, known as the explosion protection document, is drawn
up and kept up to date. In the explosion protection document it has to be demonstrated that the explosion risks
have been determined and assessed thus consequently adequate measures will be taken to attain the aims of the
Directive. Notwithstanding this fact, no further reference to the way in which those risks should be analysed
can be found on the Directive. Therefore, quite a high number of different ways are used to assess explosion
risks, showing, some of them, clear difficulties to unveil certain types of ignition sources. The aim of this work
is to make a contribution in the field of risk analysis related to the potential presence of explosive atmospheres,
specially taking into account the RASE Project (EU Project No: SMT4-CT97-2169) which has been already
published on this subject.

1
1.1

1.2

ATMOSPHERES
Source of risk

It is well known that during handling flammable liquids and gases or combustible solids, many situations
may be given in which a mixture of those substances
with air is generated in concentrations between the
explosion limits.
In the frame of the so called ATEX 137 Directive (DIRECTIVE 1999/92/EC), these situations are
defined as follows:
For the purposes of this Directive, explosive
atmosphere means a mixture with air, under atmospheric conditions, of flammable substances in the
form of gases, vapours, mists or dusts in which, after
ignition has occurred, combustion spreads to the entire
unburned mixture.
From this definition, it is clear that the physical
space where this mixture is present is the so-called
explosive atmosphere. As well, it is obvious, that
once this atmosphere is generated, the presence of
an ignition source may cause it to ignite; therefore causing certain effects in the space occupied
by the atmosphere and, to a certain extend, in its
surroundings.
These situations are not exclusively linked to a
certain type of industry, activity or equipment. Therefore, the risk of generating dangerous atmospheres
may be widely present on many different industrial
fields.

Explosive atmosphere?

As can be seen, any atmosphere which can be ignited

is commonly named after explosive atmosphere, but,
is it really so? Will any ignitable atmosphere cause an
explosion if ignited?
The answer is clearly no. Many different factors
have an influence on the effects caused when such
type of flammable atmospheres are ignited.
Amount of substance, concentration, degree of confinement, turbulence, capacity to extend the ignition
to a bigger amount of substance than the one which
initially caused the atmosphere, etc. are some of the
factors that will have an influence while defining the
consequences of an ignition.
In one hand a relatively high amount of flammable
atmosphere is required to cause an explosion if ignited
without confinement, this phenomenon is known as
Unconfined Vapour Cloud Explosion (UVCE). On the
other hand a small amount of flammable substance
may cause an explosion if ignited under confinement.
In any case the ignition of a flammable atmosphere will cause radiation and overpressure effects,
and depending on the ignition conditions, as those
listed above, those effects will have a higher or lower
magnitude.
1.3

What do we intend to protect?

The ATEX 137 Directive has the aim to protect workers from the potential consequences that the ignition

1217

of such flammable atmospheres would generate. Of

course, in most industrial installations if measures
achieve the aim of protecting people, equipment will
also be protected.
But stressing the fact that the aim is to protect people is necessary to bring into consideration not only
factors related to the physical properties of the atmospheres when ignited but also factors related to the
way those atmospheres are generated. For instance,
the generation of the flammable atmosphere may be
not independent from the presence of a worker. Such
would be the case if we think on manually feeding a
mill, transferring a flammable liquid from a drum to
a small can, discharging a hopper, taking a sample or
loading solids in a reactor through the manhole.
Many other situations under normal operation
conditions could be described in which the actions
performed by an operator will cause a flammable
atmosphere to be present. Furthermore, in most of
those cases, the amount of substance which will constitute the flammable, maybe explosive, atmosphere
will be rather small; a few grams of solid or a few
grams of liquid evaporated.

Table 1.

1.4

2.2 Risk reduction principles in ATEX

Ignition effects

Thinking on the situations described and on the properties of flammable atmospheres generated the risk is
caused by rather small amounts of substances, in an
open system and with the presence of operators in or
next to the atmosphere.
From then on it easy to assess the fact that, in case
of ignition, mainly the radiation effects may easily
overcome the operator, causing serious burnings.
Obviously bigger amounts of substances can play a
role if they are dispersed due to the primary explosion
and get ignited as well.
But even if only the initial, rather small, flammable
atmosphere is generating low radiation and/or overpressure effects, the key fact to consider while assessing this risk is that a flash fire of vapour, gas or
dust atmospheres could generate grave consequences
on the operators and almost have not effect on the
equipment or the installation.

2
2.1

RISK FACTORS
Classical risk factors

Many risk analysis methodologies assess somehow

the frequency and severity of the dangerous events
identified.
Then the risk is defined as a function which combines the frequency at which the scenario will occur
and the severity of the effects that would generate.

Example of a classical risk matrix.

Remote

Occasional

Moderate

Frequent

Catastrophic
Critical
Marginal
Negligible

Most of the times, this combination of frequency

and severity is expressed through a risk matrix (see
Table 1), where it may be established the risk acceptance criteria by balancing the frequency and the
severity. Usually, the higher the frequency the lower
the severity required to have an acceptable risk and the
other way round.
In the table above, the zone above the double line
would represent all those scenarios where the risk is
not acceptable, while the zone below the double line
represent the scenarios where the risk is acceptable.
Typically, when a non acceptable risk is identified,
it has to be then decided whether it is more feasible
to get into the acceptable risk area by reducing the
severity or by reducing the frequency.

The Directive 1999/92/EC on its article 3 defines

the basic principles to provide protection against
explosions to the operators.
Next those basic principles are presented, as in the
Directive, in order of priority:
I. Prevention of the formation of explosive atmospheres.
II. Avoidance of the ignition of explosive atmospheres.
III. Mitigation of the detrimental effects of an explosion so as to ensure the health and safety of
workers.
Observing the first and second principles, it is clear
that the intention is to avoid the completion of the fire
triangle, that is, the coincidence in the same point of
a flammable substance, a comburent and an ignition
source.
In most of the situations the comburent is the
Oxygen present in the Earth atmosphere. Therefore,
principle I can be achieved by reducing the Oxygen
concentration, that is inerting, or by reducing the concentration of flammable substance to a value below de
lower explosive limit.
Inerting is mostly a feasible principle in closed systems, while diluting or preventing the generation of
the atmosphere by reducing the power of the emission
source is most feasible on open systems.
The second principle aims to avoid the presence of
an ignition source on those cases where the formation

1218

of the explosive atmosphere can not be avoided. Many

different ignition sources can be identified which
could effectively ignite a flammable atmosphere. Nevertheless, most common ignition sources are hot
surfaces, flames or hot gases, mechanical sparks,
electrical equipment and electrostatic discharges.
To be rigorous, the assessment should assess the
presence but also the effectiveness of ignition sources.
Some ignition sources can only be avoided to a certain
degree of magnitude, not totally, therefore, it is possible that some ignition sources are present but will
not be able to effectively ignite the explosive atmospheres with which they are in contact. Of course, this
assessment depends as well on the properties of the
substances causing the atmosphere.
Finally if the formation of the explosive atmosphere
and the presence of ignition sources can not be avoided,
then it is necessary to implement mitigation measures for the effects of an explosion. Typically, those
measures are based on venting, suppression or containment of the explosion together with isolation of the
equipment or area where explosion may take place.

Zone 0

Zone 20

Zone 1

Zone 21

Zone 2

Zone 22

Figure 1. Symbols used to represent classified Ex-Zones

on a plant layout.

3
3.1

EX-ZONES
Definitions

Before presenting the methodology to assess the risk

linked to the presence of explosive atmospheres. It is
necessary to briefly present how the zones where an
explosive atmosphere are identified in an installation.
These so called Ex-Zones and its definition can be
found in the Annex I of Directive 1999/92/EC, but also
in other norms which have been used for a long time
before the implementation of the ATEX Directive.
Definition of zones is next presented:
Zone 0
A place in which an explosive atmosphere consisting of a mixture with air of flammable substances in
the form of gas, vapour or mist is present continuously
or for long periods or frequently.
Zone 1
A place in which an explosive atmosphere consisting of a mixture with air or flammable substances in
the form of gas, vapour or mist is likely to occur in
normal operation occasionally.
Zone 2
A place in which an explosive atmosphere consisting of a mixture with air of flammable substances in
the form of gas, vapour or mist is not likely to occur
in normal operation but, if it does occur, will persist
for a short period only.
Zone 20
A place in which an explosive atmosphere in the
form of a cloud of combustable dust in air is present
continuously, or for long periods or frequently.

Zone 21
A place in which an explosive atmosphere in the
form of a cloud of combustible dust in air is likely to
occur in normal operation occasionally.
Zone 22
A place in which an explosive atmosphere in the
form of a cloud of combustible dust in air is not likely
to occur in normal operation but, if it does occur, will
persist for a short period only.
After reading the definitions used to rate the ExZones it can be said that they are defined based on a
qualitative assessment of the probability of presence
of an explosive atmosphere.
The definition of a Zone 0 or 20, 1 or 21 and 2 or 22
is equivalent, the only difference comes from the fact
that Zones 0, 1 and 2 are defined due to the presence
of gases, vapours or mists and Zones 20, 21 and 22 are
defined due to the presence of dust.

3.2 Zone identification

Once zones have been defined, they are usually represented on plant layouts in order to easily identify them
on an installation.
Next, the symbology used to represent zones is
presented:
These so called Ex-Zones and its definition can be
found in the Annex I of Directive 1999/92/EC, but also
in other norms which have been used for a long time
before the implementation of the ATEX Directive.

1219

4.1

PROPOSAL OF A METHODOLOGY
TO ASSESS THE RISK LINKED TO ATEX

Table 2.

Initial statement regarding risk assessment

Up to now, all the necessary elements to present a

methodology to assess the risk linked to the presence
of an explosive atmosphere in the frame of the ATEX
137 Directive have been introduced. Next the basic
assumptions on which the methodology is defined are
presented.
In first place, it has been considered that many situations occur in which the presence of operators is
required to generate an explosive atmosphere. Additionally it could be reasonable to argue that even if
an explosive atmosphere is generated due to a unintended operative condition without the presence of an
operator, for instance, leaking of a flange, rupture of
a drum, etc. the presence of an operator may be quite
frequent since when this conditions is noticed somehow an action will be taken to solve it, most of the
times requiring the intervention of an operator.
Secondly it is clear that even if small amount of
substances generate a flammable atmosphere and it is
ignited, the consequences may cause severe injuries to
people present and low or no damage to equipment. As
well, the aim of the ATEX 137 Directive is to protect
operators.
In third place, classical risk assessment methodologies rate the consequences of the identified scenarios.
In this case the scenarios deal with the ignition of
flammable atmospheres and the consequences, as has
been explained, can be quite commonly severe.
Therefore, as an initial statement it is proposed to
leave the assessment of the consequences out of the
risk assessment, and assess only the risk of generate
and ignite a flammable atmosphere.
4.2

ATEX risk analysis matrix.

No Zone

Ignition Source
probability

Ex-Zone Classification
Zone
Zone
0/20
1/21

Zone
2 / 22

A
B
C
D

two probabilities will define the acceptable and non

acceptable risk areas.
As the Ex-Zones definition has been already presented, next the criteria to rate the probability of
presence of ignition sources is presented.
4.3 Probability of presence of ignition sources
The rating of the probability of presence of ignition
sources has been defined in a similar way to that used
while assessing the safety of equipment in front of the
presence of an ignition source, which is defining under
which conditions an ignition source will be generated
(DIRECTIVE 94/9/EC).
Four possible ratings A, B, C and D are considered;
next its definition is presented:
A: occurs permanently or frequently during normal
operation
B: does not occur during normal operation but only
as a result of rare malfunctions
C: does not occur during normal operation but only
as a result of very rare malfunctions
D: does not occur, or if it occurs, can not ignite the
explosive atmosphere.

Risk assessment factors

Once the reasons why the rating of consequences has

been left out of the risk assessment, factors that will be
considered while assessing the risk will be presented.
In first place it is quite easy to consider as a risk
factor the Ex-Zone definition, as has been explained
already this definition is based on a qualitative assessment of the probability of generating an explosive
atmosphere.
Then it is necessary to define another risk factor
in order to be able to set a risk matrix. This factor
has been defined as the probability of presence of an
ignition source.
By defining the risk as a function between the presence of an explosive atmosphere and the presence of an
ignition source the main principles to reduce the risk
according the ATEX 137 Directive can be assessed
(see point 2.2). That is, the relation between those

4.4 Risk matrix

Based on those two factors, the Zone classification
and the probability of presence of ignition sources, it
is possible to define a risk matrix.
This risk matrix has the aim to define the acceptable and non-acceptable risk areas based on a balance
between how probable is the presence of an explosive
atmosphere in an area and how probable is the presence
of an ignition source in the same area.
Next the risk matrix is presented:
In the matrix above, the zone above the double line
would represent an area where the risk is not acceptable, while the zone below the double line represents
the area of acceptable risk.
As can bee seen ignition sources occurring permanently or frequently during normal operation can only
be acceptable in a non-classified area.

1220

On the other hand, when a Zone 0 or 20 is defined

it has to be excluded the occurrence of any effective
ignition source.
4.5

Factor independence

After having presented the risk matrix it is important

to make a consideration regarding the independence
of those two factors.
Some ignition sources are completely independent
from the presence of an explosive atmosphere. As
an example one could consider an electric spark in
a switch to turn on and turn off the lights of an installation. The spark will be generated each time the
switch is pressed, independently from the presence of
an explosive atmosphere.
On the other hand, there are some ignition sources
which will be generated due to the same action that
creates the explosive atmosphere. That would be the
case of cone discharges while filling a silo. Therefore
the probability of presence of some ignitions sources
is not independent from the probability of presence of
an explosive atmosphere.
This fact has to be taken into account while using
the risk matrix defined above to analyse the risk of an
installation.
Then, what happens when the generation of an
explosive atmosphere can not be avoided as well as
the presence of an ignition source?
4.6

III Risk reduction principle

Then the third risk reduction principle defined in the

ATEX 137 Directive has to be applied.
If it is not possible to reduce the zone classification
or reduce the probability of ignition sources, then the
risk of a certain ignition source and Ex-Zone will be
placed in a non-acceptable risk area. Therefore the
only solution is to take protective and measures to
mitigate the effects of an explosion.
It can be seen as a second layer of the risk matrix
defined.
5

This methodology considers that the severity of any

explosion may be too high; therefore the assessment
follows the aim of defining the risk based on two
probabilities.
The first probability which is considered, takes into
account the definition of Ex-Zones. Thus the probability of presence of an explosive atmosphere in a certain
physical area is included in the risk analysis.
The second probability which is considered, takes
into account the presence of ignition sources. Thus the
probability that equipment or operations performed
generate an ignition source is included in the risk
analysis, even considering the conditions under which
this ignition source will be generated, that is normal
operation, failure, etc.
When the probability of presence of an explosive
atmosphere and an ignition sources are too high, the
risk is considered to be non-acceptable. Then, the
strategy to end in an acceptable risk area can be to
reduce the zone classification or reduce the presence
of ignition sources.
When none of those actions can be implemented,
the solution is to adopt protective measures in order to
mitigate the effects of an explosion.
REFERENCES
DIRECTIVE 1999/92/EC OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 16 December 1999
on minimum requirements for improving the safety and
health protection of workers potentially at risk from explosive atmospheres (15th individual Directive within the
meaning of Article 16(1) of Directive 89/391/EEC).
DIRECTIVE 94/9/EC OF THE EUROPEAN PARLIAMENT AND THE COUNCIL of 23 March 1994 on the
approximation of the laws of the Member States concerning equipment and protective systems intended for use in
potentially explosive atmospheres.
EU Project No: SMT4-CT97-2169. 2000. The RASE Project
Explosive Atmosphere: Risk Assessment of Unit Operations and Equipment.

CONCLUSIONS

A risk analysis methodology has been presented which

allows following in a systematic way the risk reduction
principles defined in the ATEX 137 Directive.

1221

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk reduction by use of a buffer zone

S.I. Wijnant-Timmerman & T. Wiersma
TNO, Netherlands Organisation for Applied Scientific Research, Department of Industrial and External Safety,
Utrecht, The Netherlands

ABSTRACT: In situations where chemical industries and residential areas are situated close to each other,
the population runs a safety risk associated with the accidental release of toxic gases. TNO has investigated the
possibilities to reduce this risk by integrating safety measures in the area between the chemical industry and the
residential area, the so called buffer zone. Two measures were selected and have been studied in more detail,
being; the effect of physical obstacles such as buildings, walls and trees and the effect of water curtains. The
effect of different configurations of physical obstacles was measured with wind tunnel experiments. The project
pointed out that physical obstacles and water curtains lead to a reduction of concentration of the gas cloud and
reduces the number of victims in the residential area behind the buffer zone.
1

INTRODUCTION

The Netherlands is a small country, with a high population density. This leads to an intensive use of the
available land. In specific situations it happens that
chemical industries and residential areas are situated
close to each other, causing a safety risk for the
population.
In a town in the south of the Netherlands a chemical plant and a residential area are separated only a
few hundreds of meters (Wiersma 2007). The area in
between, the so called buffer zone, will be developed
in the future. This provides the opportunity to take
measures to prevent the dispersion of a gas cloud in
the direction of the residential area, or to dilute such a
gas cloud.
The first step in the investigation was the inventory of measures that influence the dispersion of a gas
cloud.
As a next step two measures were selected and have
been studied in more detail, being (a) the effect of
physical obstacles such as buildings, walls and trees
and (b) the effect of water curtains.
The investigation took place under the authority of
the administration of this town. Representatives of the
administration, ofthechemicalindustryandtheregional
fire department participated in the project group.

INVENTORY OF MEASURES

A brainstorm session was organized in order to obtain

an extensive list of measures, which prohibit the
dispersion or reduce the concentration of a gas cloud.

Passive measures (e.g. walls and buildings), as well

as active measures (e.g. fans and water curtains) were
taken into account.
The measures were evaluated in a qualitative way
and ranked, based on the following criteria:

Technical feasibility
Effectiveness
Suitability in the area
Costs
Delay time of activation of the system
Side effects (e.g. noise or aesthetic value)

Two measures were selected and have been studied

in more detail, being; the effect of physical obstacles
such as buildings, walls and trees and the effect of
water curtains.

WATER CURTAINS

The first measure investigated, is the effect of water

curtains. This was evaluated, based on literature study.
Water curtains in industry are used for different
purposes:
To prevent the occurrence of a flammable gas
mixture
As a shield between a flammable gas mixture and
an ignition source
To extinguish fire
To prevent an object to heat up by heat radiation
To dilute or neutralize a toxic cloud
This last application of water curtains is studied as
measure to prohibit the dispersion of a gas cloud.

1223

Water curtains are most effective when positioned

near the source. Typical situations where water curtains are used are near loading and unloading points
and near process installations. In literature no situations were found where water curtains are situated
outside the chemical site, so this will be a new
application.
Field experiments and literature show that concentration reduction of water curtains varies from 40% to
90%. The efficiency strongly depends upon the design
of the water curtain, the nozzle parameters, the release
conditions and the meteorological conditions.
The characteristics of an efficient water curtain
are:

Large water flow

Positioned near to sources
Near to the source downward or horizontal flow
Far from source (in field) upward flow, in order to
create more turbulence in the air.

The disadvantage of water curtains is that they

are quite expensive, compared to the use of physical
obstacles. Thats why it is recommended to use water
curtains only for limited lengths, for example where a
wall is interrupted by a road.

4
4.1

4.2 Modeling of the configurations

The dimensions of the total buffer zone are too large
to test in the wind tunnel, that is why only a part of
the area is modelled. The experiments are performed
on a scale of 1:350. Table 1 shows the dimensions of
the buffer zone and of the modelled buffer zone (not
scaled).
The model is positioned on a rotating disc, in order
to study the effect for winds angular to the buffer zone.
The emission source is modelled as an impulse free
source and emits a tracer gas like a continuous release.
The emission source is situated up stream of the buffer
zone.
The gas samplers are situated at various distances
behind the buffer zone, varying from 50 meters to 600
meters, so they can measure a concentration profile.
Figure 2 shows the receptor points situated on lines
perpendicular to the buffer zone.

Table 1. Dimension of the buffer zone, actual and modelled

(not scaled).

Width buffer zone

Length buffer zone
Dimension of the
residential area starting
from the buffer zone
Distance of source to
buffer zone

PHYSICAL OBSTACLES: EXPERIMENTAL

SETUP
Wind tunnel

Actual

Modelled

140280 m
1.7 km
1 km

200 m
470 m
600 m

4001150 m

400 m

The effect of physical obstacles on the concentration

of a gas cloud was measured in a wind tunnel of TNO
in Apeldoorn (Wiersma 2007). This wind tunnel is has
been designed to be representative of an atmospheric
boundary layer. The inside dimensions of the tunnel
are 2 meters (height) by 3 meters (width) (figure 1).
The wind tunnel is suitable to measure neutral gas
releases in a neutrally stable atmosphere.

Figure 1.

Picture of the wind tunnel, inside.

Figure 2.
zone.

1224

Receptor points, perpendicular to the buffer

The residential area is modelled as a flat plane, covered with masonry sand, to prevent the build up of an
internal boundary layer.
The nature of the buffer zone itself is grass land,
modelled as material with the same roughness as grass
land.
4.3

Configurations with maximum height of 20

meters:
8: massive wall of 20 m height
9: rows of trees 20 m height

Description of configurations

Nine configurations were tested in the wind tunnel.

A configuration can consist of the following elements:
(a) (rows of) buildings, (b) (rows of) trees and (c)
walls. To get insight into the optimal possible configuration, it is important to measure the effect on the
concentration profiles of different issues:

6: 1 row of buildings 30 m height, distance between

buildings 25 m
7: 1 row of buildings 30 m height, distance between
buildings 50 m

Height of the building

Trees versus buildings
Increasing versus decreasing height of the buildings
The effect of spacing between the buildings

Figure 3 shows an example of one of the configurations in the wind tunnel.

Measurements are performed on the following
configurations:

All buildings have a length of 100 meters. The distance between the rows, in wind direction is 40 meters.
The distance between the buildings in the direction
perpendicular to the wind is 25 meters.
4.4 Measurements
The following effects were evaluated:
the concentrations for perpendicular wind flow on
the buffer
the concentration at the edges of the cloud, at a
distance of 140 m and 280 meters
the concentrations for winds angular to the buffer
zone

0: no buffer zone (reference measurement)

Configuration with maximum height of 50
meters:

RESULTS OF WIND TUNNEL

MEASUREMENTS

1: massive wall of 50 m height

2: rows of buildings of increasing height 10, 30, 50 m
height
3: rows of buildings of decreasing height 50, 30, 10 m
height
4: 1 row of trees 10 m height, row buildings 30, 50 m
height

5.1 Results for perpendicular flow

Configurations with maximum height of 30

meters:

Configuration 0: no obstacles
Configuration 1: wall of 50 meters

Point of interest was the concentrations of the gas

clouds for flows perpendicular to the buffer zone, for
receptor points situated on the axis of the gas cloud.
Figure 4 shows the result of the measurements for three
of the configurations for flows perpendicular to the
buffer zone:

5: 1 row of trees 10 m height, row buildings 30, 10 m

height
C on centration profile, p erpend icula r flow
0: no obst acle

k value

2,50E-04

2,00E-04

6: row of
buildings 30
meters

1,50E-04

2: rows of
buildings: 10,
30, 50 meters
8: wall 20
meters

1,00E-04

1: wall 50
meters

5,00E-05

0,0 0E+00
0

20 0

4 00

6 00

m eters

Figure 3. Scale model of one of the configurations, with

modeled trees on the left and buildings in the middle.

Figure 4. Concentration profiles for flows perpendicular to

the buffer zone.

1225

Table 2. Concentration reduction with respect to configurations without obstacles.

Configuration
1: wall of 50 m height
2: rows of buildings of 10 m , 30 m,
50 m height
3: 3 rows of buildings of 50 m, 30 m,
10 m height
4: 1 row of trees 10 m, 2 rows of
buildings 30 m, 50 m height
5: 1 row of trees 10 m, 2 rows of
buildings 30 m, 10 meter
6: 1 row of buildings 30 meter height
(distance between buildings
perpendicular to the wind: 25 m)
7: wall of 20 m height
8: 1 row of buildings 30 meter height
(distance between buildings
perpendicular to the wind: 50 m)
9: row of trees 20 m height

Average
concentration
reduction
76%
55%
62%
59%

An unbroken wall is more effective than an unbroken row of trees (of 20 meters height)
Increasing the distance between the buildings (perpendicular to the wind direction) from 25 to 50
meters did not influence the concentration reduction
for this specific configuration. It can be expected
that the concentration will finally increase, when
increasing the distance further
In case of rows of buildings with different heights,
it is more effective to situate the highest building
the nearest to the source

36%

5.2
38%
58%
36%
11%

Configuration 2: rows of buildings of 10, 30 and

50 m height.
Configuration 6: 1 row of buildings 30 m height.
Configuration 8: wall of 20 meters
The x-axis of the graph represents the distance to
the buffer zone, the y-axis of the graph represents a
dimensionless concentration.
The concentration profile with the highest (dimensionless) concentration is the one without obstacles.
The graph shows that configuration 1, the wall of
50 meters height, results in the highest concentrations
reduction. The concentration reduction of configuration 2, with maximum height of 50 meters, is
comparable to that of configuration 8, a wall of
20 meters.
These concentration profiles are measured for all
studied configurations. In order to easily compare the
different configurations, an average reduction factor
is calculated, being the average of the receptors on the
axis of the cloud.
Table 2 shows the average concentration reduction
of the configurations which were studied.
This table shows that the presence of physical
obstacles leads to a considerable reduction of the
concentration in the gas cloud. This can vary from
11% up to 76%, depending on the configuration.
The following conclusions were drawn, regarding the
effectiveness of the configuration:
The higher the obstacle, the more effective
An unbroken wall is more effective than an interrupted row of buildings

Results for angular flow

The effect on concentration reduction of winds, angular to the buffer zone has also been studied. Results
showed that when the wind direction was at an angle
of 30 degrees to the buffer zone, a concentration
reduction occurred (compared to the situation without
obstacle).
Experiments on angular winds described in literature (Gandemer 1975) show examples where gas
clouds role over an obstacle like a screw driver causing
a higher concentration behind an obstacle.
Wind tunnel experiments on the studied configurations pointed out that for the configurations 1, 6, 7
and 9 the angle of the flow has no influence on the
concentration reduction. For the configurations 2, 4
and 8 it did have a considerable influence. For these
configurations the concentration for flows angular to
the buffer zone showed a higher concentration compared to the concentrations for flows perpendicular to
the buffer zone.
The effect of an increase of concentration behind
an obstacles has not been observed when the first row
of buildings is high enough (50 meters) and with a row
of trees.
The results of angular flow for the remaining configurations (3, 5) are not presented, because there were
insufficient receptor points available on the axes of the
plume behind the configuration.

5.3

Results on the edges of the cloud

Another point of interest is the concentration at the

edges of the gas cloud, parallel to the axes. Under influence of an obstacle the gas cloud spreads. Compared
to the situation without obstacles, the concentrations
on the edges of the cloud increases slightly. Nevertheless the total amount of concentration reduces in the
residential area behind the buffer zone. These leads
to positive effects with respect to external safety (as
explained in paragraph 6).
An illustration of this effects is shown in
figure 5.

1226

Without buffer zone

Table 3.

Hazard source and distance to buffer zone.

Hazard source

Distance to buffer zone

NH3 release
HCN release
NO2 release

1150 m
400 m
600 m

Withbuffer zone

Figure 5. Buffer zone spreads the cloud, it becomes shorter

and wider.

are the potential sources of toxic gases which are

situated nearest to the buffer zone. Table 3 shows
the nature of the source and their distances to the
buffer zone.
The effects of the scenarios are studied for configuration 7 (wall of 20 meters height).
6.1 Assumptions
The concentration reduction, measured with the wind
tunnel, is used to determine new (reduced) concentration profiles. Therefore the following assumptions are
made:

Figure 6. Example of how to integrate physical obstacles

in the landscape.

5.4

Conclusion

The configurations showed a concentration reduction

of 11% up to 76%, depending on the configuration.
In this project the wall of 20 meters was regarded
as best solution for this specific situation. The configuration resulted in a relative high concentration
reduction of nearly 60%. Furthermore the wall fits
in the requirements of the zoning plan of the area.
When the wall is interrupted by a road, a water curtain can fill the space. An example of how to integrate
a wall in the area is shown in figure 6.
This configuration is selected by the project group
to study the effect in more detail with respect to
reduced effect distances and the reduction of numbers
of casualties.

The concentration reduction of neutral gas is applied

to heavy gas dispersion scenarios.
The wind tunnel is not equipped to measure on
heavy gas dispersions. The effect of obstacles is
expected to be bigger for dense gas clouds than for
neutral gas clouds because a heavy cloud is lower
to the ground, so obstacles have more effect.
In a further stage of the study this assumption
and the consequence of this assumption should be
validated to CFD calculations or field experiments.
The concentration reduction is based on concentration on the symmetry axis of the plume, perpendicular to the buffer zone
The reduction is measured for a fixed distance of
the hazard source to buffer zone
The experiments are performed for a source distance of 400 m from the buffer zone. More detailed
measurements on one of the configurations showed
however that the concentration reduction reduces,
with decreasing distance (of 400 to 850 meters).
This effect should be studied in more detail.
Effect of weather conditions
The wind tunnel operates under weather conditions corresponding to Pasquill class D. The concentration reduction is assumed to be indepen-dent
of this. This should be checked with experiments or
CFD calculations.
6.2 Reduction of effect distances

EFFECT ON DISTANCES AND VICTIMS

In order to quantify the effects of the concentration

reduction, three possible releases are defined. These

The reduction of effect distances is determined using

the calculations of quantitative risk analysis model
Safeti.nl. The adapted, reduced concentration profiles

1227

Table 4.

D5
m

F2
m

625
400

775
400

<400
<400

2300
1000

ammoniac scenario, where the distance between the

source and the buffer zone is large compared to
the other two scenarios, the concentration near to the
buffer zone is low. In that case the buffer zone has
little effect on the lethal percentage. Besides the effect
on lethality, the buffer zone also affects the number of
injuries in the area. This reduces for all toxic scenarios.
To calculate the reduction in number of victims, the
inhabitants of the residential area should be taken into
account. This was no part of the study.

900
650

900
700

Effect distance for 1% lethality.

Weather class
NH3 scenario
Without buffer zone
Wall of 20 m height
HCN scenario
Without buffer zone
Wall of 20 m height
NO2 scenario
Without buffer zone
Wall of 20 m height

Table 5. Lethality percentage for 3 scenarios with and

without buffer zone.

Distance from source

to area just behind
buffer zone [m]
Lethality without
buffer zone
Lethality with
wall 20 meter

NH3

Scenarios
HCN

NO2

1350

600

800

<1%

88%

19%

<1%

19%

<1%

are calculated, using the concentration reduction factor

and the calculated concentration profiles. To illustrate
the effect of the concentration reduction, table 4 shows
the distances for 1% lethality for Pasquill class F (wind
velocity 2 m/s) and Pasquill class D (wind velocity
5 m/s). The buffer zone is assumed to be situated at
400 meters from the source, for all scenarios.
The 1% lethal distance is derived based on the
method described in the purple book (Committee for
the Prevention of Disasters, 1999).
The distances of the 1% lethality shows a large
reduction, especially for the stable weather class F2.
The distances for the LBW value and the AGW value
were calculated as well (Wiersma 2007). The LBW
(live threatening threshold value) and AGW value
(alert threshold value) are threshold values used in the
Netherlands. These distances were reduced by 20%
to 50%.
6.3

Reduction of lethality percentages

The buffer zone leads to a decrease in lethality. Table 5

shows the results for 20 meter wall in the buffer zone,
for Pasquill class F wind speed 2 m/s, based on the
distances of source to buffer zone as in table 3.
The table shows that the buffer zone is effective
especially for the NO2 and the HCN scenario. For the

CONCLUSION

Active water curtains reduce the concentration of a

gas cloud for roughly 40% to 90%, depending upon
the configuration. Because they are expensive, they
should be applied for limited lengths.
Physical obstacles (walls and buildings) can lead to
a considerable reduction of the concentration of the gas
cloud. Depending on the configurations, reductions
of 11% up to 76 % were measured, for gas clouds
perpendicular to the buffer zone, on the symmetry axis
of the cloud.
The concentration reduction leads to a considerable
reduction of victims and injuries.
Because obstacles have a large influence on the
concentration of the gas cloud and number of victims, physical obstacles should be taken into account
for detailed risk analyses. Furthermore it should be
considered as a measure to increase the safety.
8

RECOMMENDATIONS

More research is needed to check the assumptions or

study the effect of concentration reduction on:
The differences between neutral and dense gas
scenarios
The effect of distance of the hazard source to buffer
zone
Effect of weather conditions
This can be studied by CFD calculations, literature
study of field experiments, more detailed wind tunnel
experiments or field experiments.
When a definitive configuration is chosen, the
effect of the configuration on concentration reduction should be studied in more detail on the following
points:
Effects of distance to source
More detailed insight in the effect of angular flow
Calculation of reduction of number of victims by
taking into account the population of the residential
area.

1228

REFERENCES
Committee for the Prevention of Disasters, 1999, Purple Book, Guidelines for quantitative risk assessment,
CPR-18E, The Hague, The Netherlands.
Dandrieux A., Dusserre G., Ollivier J. & Fournet, H. 2001,
Effectiveness of water curtains to protect firemen in case
of an accidental release of ammonia: comparison of the
effectiveness for two different release rates of ammonia, Journal of Loss Prevention in the Process Industries,
volume 14: 349355.
Dandrieux, A., Dusserre G., Ollivier J., 2002, Small scale
field experiments of chlorine dispersion, Journal of Loss
Prevention in the Process Industries, volume 15, 510.
Dutch Association of Cost Engineers, Prijzenboekje, 24ste
editie, november 2006.
EPA-report to Congress, 1992, Hydrogen fluoride study.

Gandemer, 1975. Wind environment around buildings:

Aerodynamic concepts. Proc. 4th Int. Conf. Wind Effects
on Buildings and Structures, Heathrow.
Hald K., Buchlin Dandrieux A. & Dusserre, G., 2005,
Heavy gas dispersion by water spray curtainsa research
methodology, Journal of Loss Prevention in the Process
Industries, volume 18, 506511.
Gant, S.E., 2006, CFD modelling of water spray barriers, HSL-report, HSL/2006/79. http://www.hse.gov.uk/
RESEARCH/hsl_pdf/2006/hsl0679.pdf
Rivm.safeti.nl
Wiersma, T., Wijnant-Timmerman, S.I. & Kamperveen, J.
& Boonstra B., 2007, Bufferzone Kampstraat, Op zoek
naar een veiligheidsverhogende inrichting van het gebied
tussen chemische industrie en woonwijk, TNO report Ref
No 2007-A-R0303/B.

1229

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Safety in engineering practice

Z. Smalko
Air Force Institute of Technology, Warszawa, Poland

J. Szpytko
AGH University of Science and Technology, Cracow, Poland

ABSTRACT: The paper is focusing on safety engineering, as a fast developing discipline branch. The paper
is describing major objectives of safety engineering, based on the system approach, understanding the structure
of safety management system, including the human factor, and all interrelationships between the system components. Major issues on safety engineering of any business include: selection of suitable tolerance zones of
selected key operation parameters of the system, taking the human factor as a part of the system into account,
possible risk under the system operation identification with taken under consideration the time and cost factor.
Designing of complex business systems is always associated with the risk that the system fails to reach the
desired safety level and losses and damages shall be higher than the assumed acceptable level and the surrounding conditions may alter so much that the presumed respond measures and objects of the system shall no longer
be adequate.

INTRODUCTION

The basic quality measures of the objective (device)

are its usefulness, safety, reliability, accuracy and others (e.g. esthetic). With increasing the automation
level of technical systems and devices, more and more
attention has been focuses on problems of safety and
reliability (Smalko & Szpytko, 2003; Szpytko, 2004a;
Smalko, 2005; Smalko, 2008; Szpytko, 2007a). More
information properly related to the safety it is possible to achieve from publications (Aven & Vinnem,
2007; Jazwinski & Wazynska-Fiok, 1993; Sklet,
2006), and to reliability from the example references
(Tobias & Trindade, 1986; Mourbray, 1991; Borgon et al, 1992; Elsayed, 1996; Saleh & Marais,
2006) and standards. The safety can be a subject of
control, and its required level can be obtain by inbuild security systems, which enables restraining work
in critical point. The critical state of any technical
device could arise from the loss of operation abilities
after exceeding the allowable values. Reliability of the
object/device includes the proper features of a device
and the system of detecting and removing the failures, which occurred during utilization. The failures
are the results of operation and exceeding critical values of the device, and their changes are mostly random.
The above results with new today problems in engineering practice (Szpytko, 2007a), so-called safety
engineering, waiting for description and solving, for
example:

investigation of results of lack memory of the

technical systems including objects/devices,
causeresults type approach to investigation of
damages in arrangement of possible states of the
system,
correlation of system safety and reliability,
integrated perception the life phases of objects/
devices from regard the safety and reliability
aspects,
follow-up systems and objects/devices structures
integration.
Safety engineering is a fast developing discipline
branch. The today modern technical and organizational systems have become so large and sophisticated
that the safe management is nearly infeasible and
exceeds capability of organizations that are established
to be in charge of economy, industry and transport
and others. The major objective of safety engineering,
based on the system approach, is to understand the
structure of safety management system, including
the human factor, and all interrelationships between
the system components. Major issues on safety engineering of any business include: selection of suitable
tolerance zones of selected key operation parameters
of the system, taking the human factor as a part of the
system into account, possible risk under the system
operation identification with taken under consideration the time and cost factor. Designing of complex
business systems is always associated with the risk that

1231

the system fails to reach the desired safety level and

losses and damages shall be higher than the assumed
acceptable level and the surrounding conditions may
alter so much that the presumed respond measures and
objectives of the system shall no longer be adequate.
Safety engineering approach is especially important in designing large complex systems. The first
task of safety engineering is to establish the concept of comprehensive management of the system.
The safety management of the system is the first and
crucial step to find out interrelationships between various distinguishes components of the system. The
direct objective of safety engineering is the reasonable
deployment of a tangible product, derived by systematic ideas in accordance to the predefined rules. The
approach bases on transient dynamic type characteristics and analysis of potential damages and losses
caused by individual components of the system.

SAFETY ENGINEERING SYSTEMATIC

FORMULATION

Safety engineering, from the systematic point of view

(Bertalanfy, 1973; Cempel, 2006; Klir, 1972), predominantly deals with artificial systems that exhibit
the following features:
they are integrated, which means that all the system
components are dedicated to one specific objective, which must be precisely established prior to
the designing process starts,
the system is large: system is made up of a large
number of various components, performs a number
of functions and is expensive,
mutual interdependence between system components is significant (changes of one system variable
affect a number of other system variables),
system inputs are subject to stochastic stimulation,
the system is partly automated: some of the system
functions are still performed by human operators,
but other ones are a subject of programming using
dedicated tools.
Safety engineering deals with preparation of decisions dedicated to technical and/or organizational
matters that include:
analysis of the problem with consideration of all the
essential internal and external interrelationships,
definition of desired targets for the problem solution
with respect to the required system safety level,
disclosing of all the possible options for the possible
final solution,
comparison between specific solutions and rating
them in terms of detrimental effect against the
required expenditures,
selection of the most advantageous option.

Definition of the system on the subject of research

comprises two basic operations: quantization and simplification of the system. Quantization consists of subdividing of the system into its elementary components
(parts) and establishing of quantitative interrelationships of such elementary units. On the other hand,
simplification of any system remains the art of engineering as it bases, to the substantial degree, on the
engineers intuition.
Major issues on system safety engineering systematic formulation stage include selection of suitable
criteria (criterion), seeking for feasible solutions, consideration to intangible matters, taking the human
factor into account, introduction of risk and timeinduced deterioration. Designing of complex systems
is always associated with the risk that the system fails
to reach the desired safety level and both losses and
damages shall be higher than the assumed acceptable level, as well as the surrounding conditions may
alter so much that the presumed respond measures and
objectives of the system shall no longer be adequate.
During the systematic system analysis process,
three different objectives of systematic analysis are
possible to distinguish: for the primary real system
under consideration, for the virtual system developed
based on the primary system and system carried out
analysis. The next stage include definition of the
system on the subject of research with further specification of such system, supported by analysis of the
system structure and behaviour and ideas leading for
improvements and corrections of the existing system
(e.g. integration of subsystems and components) with
above implementation as the last step. Definition of
the system on the subject of research includes the
following tasks:
checking whether specification of the system objectives is correct,
selection of the system components,
selection of the system surrounding components,
selection of essential links between system components,
selection of essential links between the system and
its surrounding,
deployment of a graphic image of the system,
development of matrix-based representation of the
system,
other forms of system representation.
Moreover, it is highly desirable to make the decision
whether system components will deemed as actions or
organization units, select the abstraction degree and
determine the system components where the decisionmaking process takes place, as well as those where
only simple information processing occurs, and finally
find out how make description of the system fully or
partly automatic at least. The large system analysis

1232

from safety attention needs support computer-based

techniques.
The important step of any operation system analysis focusing especially on safety problems are system
structure and behaviour as well as interaction between
specified elements. Analysis of the safety system
structure and behaviour should provide answers to the
following questions:
what is the system nature and objectives?
what is the system structure and how it corresponds
to its objective and what is the system behaviour?
why does the system behaviour fail to be in line with
the initial assumptions?
is it possible to improve the system operation and
how to achieve this?
how shall the suggested improvements affect the
system behaviour?

DESIGNING OF SAFETY SYSTEMS

Development of safety systems, based on real circumstances (on an existing system), is started from
a systematic search for real equivalents to the applied
terminology, which is already known from systematic
analysis. Then the system layout should be present
in a perceivable form (verbal definition, tabularized
scheme, graph, decision diagram and/or computer
presentation, etc.). The system development process
incorporates definition of components and interconnections among them. Some selected factors are
significant, whereas other are neglecting. Possible
interferences between system components must be
taking into account, as well as sequences and frequencies of system processes and events are a subject
of analyzing. Then categorization of the engineering
problem must be carrying out, but unity of terminology
must permanently being keeping in mind.
Systematic approaches have a direct input to the
mental activities of researchers in the field of safety
engineering in relatively fixed sequences that comprises the following steps:
identification and diagnosis of a research subject
which have been extracted from the entire system,
system identification, definition of the system on
real and objectively existing circumstances and
environment,
representation of a real system by means of its model
(meta-system designing),
quantization of characteristics attributable to the
designed system,
algorithmization and experimental modelling (solutions, calculations),
analysis and interpretation of result obtained from
systematic studies,

analysis of implementation opportunities associated with result obtained from systematic studies.
The above classification is only an arbitrary assumption.
So-called safety systems should met the following
requirements:
be capable to correctly perform their functions with
no conflicts with the surrounding environment,
provide sufficient controllability and maximum
possible degree of automation and computerization
of operators activities,
guarantee that technical durability, economic lifetime and innovative qualities of he facilities shall
be employed in reasonable degree,
make sure that environmental requirements shall
be met,
be secured against unauthorized and undesired
access to the facilities,
be furnished with ergonomic technical solutions,
including easy control and operation facilities,
guarantee that the facilities are ready to perform
their function in fault-free manner wit due regard
to technical condition of the facilities as well as
physical and mental status of operators,
enable current and permanent insensitivity to exposure to internal and external harmful factors, tolerance to inherent faults and operators errors,
be equipped with technical and medical neutralizing
and counteracting means as well as measures for
elimination of effects of critical disasters,
meet requirements to insure the facilities against
damages and operators against third-party responsibility as a consequence of failures and injures to
operators and damages to the facilities and goods.
Methodological development of safety systems
should give answers to the following questions: Who
claims needs to initiate safe execution of the assignment with system use? Whose needs related to system
safety assurance must be fulfilling? Whose point of
view is taking into account when the needs for system
safety assurance are considered? How many viewpoints can be respect when the need for system safety
assurance is considered? What is the position of system safety assurance in the hierarchy of organizations
needs? When fulfilling of the system safety assurance
needs should start and how long it should last? Which
techniques for system safety assurance are applied,
and. which methods and expenditures are anticipated?
What external circumstances are predictable during
system assurance process? Which disturbances and
which conductive factors are expected? What are
constraints related to safety assurance means and
methods? How make the system assurance process
effective? What methodological and subject matter
reasonable approaches are to be applied? How to

1233

balance capabilities and demands for safety assurance

means?
Designing process of safety systems requires
expenditures (what kind) and material means (how
much) which are involved into the safety assurance
area, as well as the approaches to fulfil the overlooked
need with use the system (what) and items details (what
kind: material securing means, technical or medical
services or methods) which are included into system
safe completionFigure 1.
Other actions must concentrated on clients that
are responsible to specify the system safety assurance applied and decides upon functions of safety
means and the method in which these means enable
safety assurance. Moreover, on unit/man which has
right to sets up system operation and acceptance criteria for safety means and verifies whether the selected
safety means conform to the regulations and requirements (checking usability, functionality, ergonomic
features and others). The operation process of the system is connecting with maintenance. Also essential
is knowledge related to the manner of further operation and maintenance of the system safety (how much
is needed, how to measure the demand), as well as
requirements which are necessary to achieve top efficiency and inherent uniformity of the system safety
(are reusable or disposable safety means to be used
and what is the anticipated lifetime thereof).
Maintenance phase is dominated by the preventive
type approach and decision-making based on system technical state or expressed operational potential.
The key matter is proper activities control of special
services. The above implementation needs to be continuously monitoring the technical state of the system
and its degradation. To analyze the system degradation processes the causeeffect methods are useful:
FMECA (Failure Modes and Effect Analysis), HAZOP

Potential hazard analysis:

range, identification of
potential sources,
consequences estimation
Decision - making: activity
acceptable, system capacity
(operation potential)

Protection techniques,
protection devices

Figure 1.

Scheme of safety systems designing

(Hazard and Operability studies), FTA (Fault-Tree

Analysis methods).
During the design process, it is necessary to take
under consideration that the system never reaches system safety assumption and losses will be higher then
planned, as well as environmental state of the system
will change so farso founded previously activities
and targets of the system will be no more suitable.

TRANSPORT SAFETY SUPERVISION

SYSTEM BASED ON ACTIVE
KNOWLEDGE MODEL

Safety of transportation devices is a priority problem in exploitation, particularly for inspected one.
Requirements placed to the device maker, as well
as technological progress, make possible to improve
constructions taking into consideration users (clients)
needs and expectation. The device safety concept is
drifting from a passive approach to mixed taking into
account active and adaptation type of solutions. The
above required system approach to device safety shaping, with taken into account distinguished devices
subsystems, integration of all actors engaged into
the operating process (people and loads movements,
impact with surrounding), continuous and evolutionary character. The device safety is reached in all
devices life phases and is verified in operation phase
(exploitation is including operation and maintenance).
Obtained user needs in a field of safety is in practise difficult to obtain, because the impacts have a
complex character, the device unit costs is very high,
and mostly affects the unit production. It is the reason why in practise special attention is focusing on
integrated undertaken which rationalise exploitation
safety of the device in a particular environmental conditions, as well as modernisation type (reengineering
approach) based on such fields of activity like: technology, data mining, management. To keep overlooked
safety level the complex approach is necessary.
The market globalization places new challenges in
the management of activities in particular in transportation activities. There are searching techniques
make possible effective management of transportation systems, with taken safety and reliability into
consideration. Such chances produce telematics technique (Szpytko, Jazwinski, Kocerba, 2007). As the
result of telematics used in different fields of man
activities it is possible to reduce exploitation costs of
technical devices. Techniques, which support telematics (e.g. software, hardware and tools), are rapidly
improved. It is growing requirement on so-called intelligent transport service ITS type (Intelligent Transport Services) and dynamical type management of

1234

devices (e.g. airbags, safety belt, both acting force

level sensitive),
automatic undertaken emergency procedures under
hazardous situations,
signals devices addressed to the other user operating
at the same place (e.g. sound signals, visual signals,
added light).
Each agent is an object of supply and controls.
Man-operators are equipment with modules of knowledge and skills I (with use of own in-build sensors),
which make possible auto-correction of done controls as the results of undertaken activities. Moreover,
the device, depending on automation level, may be
equipped in auto-corrective module SA (self-acting).
The output products of activities undertaken by individual agents are shaping for decision-making needs
in quality module so-called telematics.
Telematics TE is a quality module, which help
management:

Figure 2. The transport active knowledge model: Iknowledge and skills module, INinputs, SAauto-corrective
module, TEtelematic unit.

transportation devices DTM (Dynamic Traffic Management), both on large distances and at integrated
automated manufactures.
The transport system is mostly composed from
three categories of agents/actors: device A, manoperator B (device B1, service/maintenance B2, general coordinator/management B3), safety subsystem
C (in complex systems the total number of actors
is equal to N, e.g. D-surrounding). Between each
actor exist specified relation/controls (Figure 2).
For example between operator and device attributes
exists several correlations, for example: perception
information visualization, knowledgemonitoring,
skillsoperation realization ability, decision making
abilitycorrective auto-activity, reaction on external
stimulussafety device and strength.
Example expectations from the safety subsystem
are as follows:
recognition and signal of the device technical state,
recognition and signal of the psychical and psychological operator predisposition,
recognition and signal of the possible hazards coming from surroundings (e.g. anticollision sensors),
passive safety (over-dimensioning of the subsystems),
active safety: overloads limiter, active and oriented
energy dissipation during undesirable events, intelligent based safety systems used sensors estimated
acting forces and hand-worked special safety

to get exit products generated by agents which have

been taken part in directed undertakings,
base on quantity and qualitative assessment of exit
products to generate integrate products (including
safety indicators),
to extract conclusions which help to compose dedicated input controls and implement them in practice.
A significant purpose to use telematics in practice is to optimize activities (or combined activities)
as the result of more effective getting and utilization of possible information (activities accompanying
data). Telematics is based on techniques in the field of
informatics (Information Technology), vision sources
(optoelectronics), automation and telecommunication
and decision making approach.
Activities D implemented by i-th agent are possible
to present by dependence:
D(i) = f [D(i, k, j), T, T, M, M, SK, SK]
for i = 1 . . . N, k = 1 . . . p, j = 1 . . . r
where:
D
D(i)

(1)

D(i, k, j)

activity,
activity D implemented by i-th
agent,
j-th category of information accompanying of k-th category of activity
D realized by i-th agent including
safety indicators (which are describing system safety feature),

time, time delay T,

1235

M = f (x, y, z)
SK

operation space 3D type and her

parameters, change of space parameters M,
internal type connections concerning the object activity, change
of connection SK,

i-th agent which is implemented definite activities,

k-th category of activity realized by agent,
j-th category of information as the result of activities
put into practice by agents.
The formula D(i) determine that i-th agent (which
possess specified knowledge and skills) has at owns
disposal information accompanying realization of k-th
category activities D oriented on load and/or man with
consequences expressed with j-th information, taken
into consideration safety indicators. The system activity is started and finished by agent B (operator) based
on external information IN or/and own knowledge and
skills. Information generated, as the result of activity
of particular agent, can be available to him and others
agents with the time delay T or are their lack. Information lack or receiving them with no acceptable time
delay T be effective with no accurate (or without)
decision making by the agent.

FINAL CONCLUSION

Scientific disciplines related to the systematic

approach to safety issues have one common subjectthe investigated system. They also use similar
methods that are employs for research activities: interdisciplinary approach, broad metal horizon, ability to
overcome formal borders between specialized disciplines and individual features of business and administrative organizations. Resulting the known references
on safety engineering, the systematic approach to
safety assurance can be still considered as the most
suitable method and should be subject to further
development and investigation.
Analysis of today progress in any complex system
including safety problems and needs makes possible
the following conclusions:
1. further evolution into more complex system where
human is playing essential part,
2. are overlooking new solutions together increasing
safety and reliability of system composite as manmachine set,
3. solution evolution of remote warning before possible systems threats and dangers,
4. more stronger integration of systems structures and
activities, systems follow-up,
5. more stronger improvement of preventive approach
to maintenance process,

6. better system set process understanding (not only

analysis of system with use the Fault Tree Analysis method) in result of investigation of system
exploitation potential changes and his invariability
(the intelligence, adaptation),
7. developing improvements in methods and tools
focusing on decision making process with use of
man-machine interfaces,
8. much more stronger integration of manmachineenvironment set,
9. progressing in shaping system safety and reliability taking into account life-phases of the manmachine set.
The paper describes major objectives of safety engineering, based on the system approach, understanding
the structure of safety management system, and interrelationships between the system components. The
attention have had taken on safety engineering systematic formulation, designing of safety systems, as
well as on practice related transport safety supervision
system example based on active knowledge model.
Major issues on safety engineering in practice
include selection of suitable tolerance zones of
selected key operation parameters of the system, taking the human factor as a part of the system into
account, and possible risk under the system operation identification with taken under consideration the
time and cost factor. Designing of complex business
systems is always associated with the risk that the system fails to reach the desired safety level and losses
and damages shall be higher than the assumed acceptable level and the surrounding conditions may alter so
much that the presumed respond measures and objects
of the system shall no longer be adequate.

ACKNOWLEDGEMENT
The research project has been financed from the Polish
Science budget.

REFERENCES
Aven, T. & Vinnem, J.E., Eds. (2007). Risk, Reliability and
Societal Safety. Taylor & Francis Group, London.
Bertalanfy von, L. (1973). General Systems Theory.
Brazilier, NY.
Borgon, J., Jazwinski, J., Sikorski, M. & Wazynska-Fiok, K.
(1992). Niezawodnosc statkw powietrznych. ITWL,
Warszawa.
Cempel, C. (2006). Teoria i inzynieria systemw (Theory and
Engineering of Systems). ITE, Radom.
Elsayed, A.E. (1996). Reliability engineering. Addison
Wesley Longman Inc., Reading.
Jazwinski, J. & Wazynska-Fiok, K. (1993). Bezpieczenstwo
systemw. PWN, Warszawa.

1236

Klir, G.J. (1972). Trends in General Systems Theory. Prentice

Hall, NY.
Mourbray, J. (1991). Reliability Centered Maintenance.
Butterworth-Heinemann Ltd, Oxford.
Saleh, J.H. & K. Marais (2006). Highlights from the early
(and pre-) history of reliability engineering. Reliability
Engineering & System Safety, 91 (2), 249256.
Sklet, S. (2006). Safety barriers: Definition, classification, and performance. Journal of Loss Prevention in the
Process Industries, 19 (5), 494506.
Smalko, Z. & Szpytko, J. (2003). Kierunki i perspektywy rozwoju transportu w aspekcie globalizacji rynku. Inzynieria
Maszyn, s. 14 20, v.8, z.1.
Smalko, Z. (2005). Podstawy projektowania niezawodnych
maszyn i urza dzen mechanicznych. PWN, Warszawa.
Smalko, Z. (2008). Wybrane problemy inzynierii bezpieczenstwa. Materiay 36 Szkoy Zimowej Niezawodnosci, s. 298297, PAN, Szczyrk.

Szpytko, J., Jazwinski, J. & Kocerba, A. (2007). System

koordynacji dziaania rozproszonych srodkw transportu
(Management system of distributed transport devices).
Prace Naukowe, Transport, z. 63, s.249256, Oficyna
Wydawnicza PW, Politechnika Warszawska, Warszawa.
Szpytko, J. (2007a). To increase systems safety and reliability by better man machine interface. Proceedings on 8th
IFAC International Symposium on cost oriented automationaffordable automation systems, XII Convencion y
Expo InternacionalInformatica 2007, La Habana.
Szpytko, J. (2004a). Ksztaltowanie procesu eksploatacji
srodkw transportu bliskiego. Monografia, Biblioteka
Problemw Eksploatacji, ITE, KrakwRadom.
Tobias, P.A., Trindade, D. (1986). Applied reliability.
New York, Van Nostrand Reinhold.

1237

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Why ISO 13702 and NFPA 15 standards may lead to unsafe design
S. Medonos
Petrellus Engineering Ltd, UK

R. Raman
Lloyds Register, Australia

ABSTRACT: Many international standards specify firewater deluge rates intended to protect personnel from
thermal radiation from the fires during escape, and to cool equipment and structures affected by thermal radiation
or direct flame impingement. ISO 13702 and NFPA 15 are popular international Standards, frequently used for
the design of firewater systems in the oil & gas industry both offshore and onshore, and the chemical process
industry. Process fires in these industries are often pool fires and jet fires, where jet fires are high momentum
gas or multi-phase fires. Recent attempts for practical applications in plant design of the deluge rates specified
in these Standards have shown that these deluge rates are not fire specific, but generic. Since fires vary widely
in intensity and mechanism of escalation to structures, use of generic deluge rates may lead to inadequate fire
protection of personnel and plant. This Paper:
examines potential fire situations, and escalation mechanisms
provides a critique of existing fire protection standards with respect to deluge systems
identifies and suggests further development of technical information and Standards to provide design guidance,
and
proposes deluge rates for interim practical applications in the hydrocarbon process industries until new
information becomes available.

FIRE SCENARIOS TO PROTECT AGAINST

1.1

Fire types and definitions

The following fires may occur on an offshore oil &

gas installation or onshore refinery or petrochemical
plant:

Pool fires of burning flammable liquid;

Jet fires of burning gas;
Jet fires of multi-phase flammable fluids;
Jet fires of flammable liquid and its droplets.

Figures 1 and 2 show examples of offshore installations on fire. The following definitions are normally
used:
Pool fire: A turbulent diffusion fire burning on a
pool of liquid fuel spilt onto an open surface, the burning rate of which is determined by fuel type and pool
size.
Jet fire: A turbulent diffusion flame resulting from
the combustion of a steady release of pressurised
gaseous, multi-phase or liquid fuel: Significantly

higher rates of burning are likely due to turbulent fuel/

air mixing.
Deluge system: A sprinkler system where the water
distribution piping is equipped with open spray nozzles for discharging over an area. A sprinkler system
is a fire protection system consisting of at least one
automatic water supply and a distribution piping system equipped with sprinklers for discharge over an
area to be protected.
Depending on confinement and size, the surface
heat flux of hydrocarbon pool fires may be in the range
from 50 kW/m2 to 250 kW/m2 and from 100 kW/m2
to 400 kW/m2 for jet fires (Ref. 1). Not all fires are
large as those shown when they start. Only if the fire is
not controlled, it escalates to involve major structures
and grows to a stage where no protection is possible.
The surface heat flux in the bright light yellow areas
in the jet flame in Figure 2 is close to 400 kW/m2 , i.e.
approximately 1400 C.
The objective is to prevent such escalation and deluge is a major safety critical element in the control of
fires. The mechanisms involved in the transfer of heat
from a fire (flame) onto an object are thermal radiation

1239

Figure 1.

Pool fire on the sea surface.

Figure 2.

Jet fire from a riser on an offshore platform.

1240

and convection. The convective together with radiative

heat transfer only applies for the transfer of heat from
flame (or hot gases) onto the object that is directly
impinged or engulfed by the fire.
Heat transfer by thermal radiation is from a flame
which is at some distance from the threatened object.
In the case of pool fires the heat transfer to an
impinged object occurs largely by thermal radiation
(approximately 70%), and the rest by convection.
Hence escalation from these fires may be controlled
by the generic deluge density. However, in the case of
jet fires, the reverse is true, i.e. nearly 70% of the heat
input is by convection. This makes such rapid temperature increase on the receiving body that generic
deluge is often insufficient to prevent escalation
(Ref. 2, 3).
1.2

Fire protection by deluge

Depending on the rate, deluge can,

reduce fire intensity,
cool the surface of the objects that the flame
impinges on,
reduce fire intensity and cool the surface of the
objects that the flame impinges on,
reduce the thermal radiation from non-impinging
fires onto personnel and plant, and
extinguish the fire.
Whether or not deluge is effective for the reduction of fire intensity and/or cooling of objects depends
on kinetic energy of the deluge droplets with respect
to the kinetic energy of the flame. For a pool fire,
because the kinetic energy of the flame is given
by the buoyancy of the burning gas, it is relatively small, and even low rate deluge may have a
chance to,
penetrate the flame to reduce the fire intensity, and
establish a continuous water film over the surface
area of the protected object to provide cooling of it.
The thermal radiation from a non-impinging fire
onto personnel will also be reduced by deluge, but the
deluge is unlikely to extinguish the fire unless the fire
is very small and the deluge rate is relatively high.
For jet fires, due to the high kinetic energy of jet
flame, the jet flame will push deluge droplets out of
its path and away from the surface that the deluge is
supposed to cool, unless the deluge is applied at a
relatively high rate.
When a fire occurs in a process facility, there are
two competing processes that occur concurrently as
time progresses. One is the potential for the fire to
escalate beyond the initial source, making control of
fire near impossible, resulting in significant structural
damage. The other is the escape and evacuation of personnel to a safe area, and, in the case of onshore plants,

protection of fire fighting personnel from potential

escalation of fire. The faster of the processes wins. The
objective of deluge should be either to prevent or significantly slow down the escalation, so that personnel
protection can be ensured.
The information on what deluge rates should be
applied for which fire types and under which circumstances, does not seem to be readily available. Some
limited data is scattered through the literature.
2

A CRITIQUE OF DELUGE RATES

FROM STANDARDS AND GUIDANCE

Table 1 gives an overview of the equipment to be protected and the guidance that can be obtained from
current Standards and Guidelines NFPA 15, FABIG,
OTI, NORSOK and ISO 13702 (Refs. 4, 5, 6, 7, 8
respectively). In Table 1, the information missing in
these Standards and Guidelines is written in the form
of questions in italics.
3

WORK THAT PROVIDES CRITIQUE

OF THE STANDARD DELUGE RATES

For a jet fire a model for a water spray system described

in Ref. 10 was developed by Lev (1991) and takes into
account, among other things, the high thermal radiation within the jet flame. Lev recommends a design
cooling rate of 60 to 120 l/min/m2 . Droplet size is also
noted to be importanta 0.5 mm diameter droplet is
considered to be vulnerable and it is suggested that a
1 mm diameter should be aimed for. This applies to jet
flames of liquefied petroleum gas (LPG) with stream
temperatures ranging from 5001300 C and jet mean
velocities of 40 m/s.
The information on effectiveness of deluge for
engulfing jet fires is detailed in Ref. 5. It states that
the due to the high kinetic energy of the jet fire, deluge
will be penetrated by the flame. It refers to the work
carried out by Lev mentioned above and concludes
that deluge rates far in excess of current requirements
may be required. It states that Lev concluded that high
velocity nozzles or water monitors would be best to
tackle the impinging jet fires.
Identical information is given in Ref. 6, which states
that for . . . impinging jet fire: deluge likely to be
ineffective but this requires special attention.
The Paper in Ref. 11 describes full scale tests of the
effectiveness of deluge for the protection of LPG tanks.
The referenced Paper states that the recommended deluge rate of 10.2 l/min/m2 for such systems is designed
to provide adequate protection against hydrocarbon
pool fires. However, it is now recognised that the incident scenario most likely to threaten the integrity of
such vessels is impingement by a hydrocarbon jet fire.

1241

Table 1.

Deluge rates from standards and guidelines.

Objects to be protected and

against which fire types

Deluge rates from standards and guidelines (the missing information required
for correct practical application is described in the form of questions in italics)

Wellheads
To be protected mostly against
jet fires, including jet fires
from well blowouts
Jet fires are directly from
very high well pressure

NFPA 15 (2007 version) does not specifically address fire protection of wellheads.
FABIG Technical Note 8 (2005) gives information on page 128 for design guidance.
It gives a figure of 400 litres/min/well for the protection of a wellhead. Is this for
pool or jet fires, impinging/engulfing, non-impinging/non-engulfing?
The rates given in the FABIG Technical Note are based on the document Interim
Guidance Notes for the Design and Protection of Topside Structures against Fires
and Explosions, The Steel Construction Institute, Publication Number SCI-P-112,
1992, Ref. 9.
OTI 92 607 gives 400 litres/min/well, which is taken from Availability and properties
of passive and active fire protection systems, Steel Construction Institute for the HSE
OTI 92 607. Is this for impinging/engulfing or non-impinging/non-engulfing flame?
NORSOK does not specifically address fire protection of wellheads.
EN ISO 13702 gives typical minimum 10 litres/min/m2 or 400 litres/min/wellfor
wellhead/manifold areas.
What type of fire is this for? These rates do not seem to be equivalent to each other, as
it would implicitly mean that a wellhead may occupy an area of 40 m2 .

Wellhead areas
To be protected against
jet fires and pool fires
Jet fires may be directly
from very high well pressure
but their jet flame momentum
may be reduced by the flame
impacting on equipment and
structures
Processing areas and auxiliary
equipment, and pressure
vessels
To be protected against pool
and jet fires

NFPA 15 (2007 version) does not specifically address fire protection of wellhead areas.
FABIG and OTI do not specifically address wellhead areas.
NORSOK S001 (Ref. 7) gives 20 litres/min/m2 .
What type of fire is this for?
EN ISO 13702 gives typical minimum 10 litres/min/m2 or 400 litres/min/well
for wellhead/manifold area, and 400 litres/min/m2 for blowout preventer area.
What type of fire is this for? The 10 litres/min/m2 or 400 litres/min/well do not
seem to be equivalent to each other.

Jet fires are from inventories

of pressure which is generally
lower than that of wellheads

NFPA 15 (2007 version) specifies a minimum of 10.2 l/min/m2 for transformers

regardless of type of fire. Can auxiliary equipment be interpreted in this way?
It mentions a rate of 6.1 l/min/m2 for the expected non-absorbing ground surface
area of exposure. Is it possible to use this figure for general coverage of an area and
for which types of fire (pool or jet fire, impinging/engulfing, non-impinging /
non-engulfing?
This is for exposure protection. What types of fire is this for?
Information is also provided on the control of burning for pumps/compressors and
related equipmentthe rate should be 20.4 l/min/m2 . What types of fire is this for?
FABIG Technical Note 8 (2005) gives information on page 128 for design guidance.
To protect against pool fires it suggests 10 litres/min/m2 . Which type of equipment is
this for? Is this for impinging/engulfing or non-impinging/non-engulfing flame?
It gives a general water application rate of 12.2 litres/min/m2 . Which type of
equipment is this for? Is this for impinging/engulfing or non-impinging/nonengulfing flame?
To protect against high pressure jet fires it suggests 20 litres/min/m2 . Which type of
equipment is this for? Is this for impinging/engulfing or non-impinging/nonengulfing flame? It gives a general water application rate of 12.2 litres/min/m2 .
Which type of equipment is this for? Is this for impinging/engulfing or nonimpinging/non-engulfing flame?
It gives 400 litres/min/m2 to protect against impinging/engulfing jet flame.
It also states that the offshore industry often uses the rates recommended by NFPA 15.
The rates given in the FABIG Technical Note are based on the document Interim
Guidance Notes for the Design and Protection of Topside Structures against Fires
and Explosions, The Steel Construction Institute, Publication Number SCI-P-112,
1992, (Ref. 9).
OTI 92 607 gives 20 litres/min/m2 for jet fire and 10 litres/min/m2 for pool fire
which is taken from Availability and properties of passive and active fire protection
systems, Steel Construction Institute for the HSE OTI 92 607. Is this for impinging /
engulfing or non-impinging/non-engulfing flame?
(continued)

1242

Table 1.

(continued)

Objects to be Protected and

against which Fire Types

Deluge Rates from Standards and Guidelines (the missing information required
for correct practical application is described in the form of questions in italics)
NORSOK S001 (Ref. 7) gives 10 litres/min/m2 for pool fire and jet fire. Is this for
impinging/engulfing or non-impinging/non-engulfing flame?
Deluge Rates from Standards and Guidelines (the missing information
EN ISO 13702 gives typical minimum 10 litres/min/m2 with the exception of pumps/
compressors where it gives 20 litres/min/m2 . What type of fire is this for?
EN ISO 13702 does not specifically address fire protection of pressure vessels, but gives
typical minimum10 litres/min/m2 with the exception of pumps/compressors where
it gives typical minimum 20 litres/min/m2 . What type of fire is this for?

Horizontal structural steel

To be protected against
pool and jet fires

Vertical structural steel

To be protected against
pool and jet fires

NFPA 15 (2007 version) specifies a minimum of 4.1 l/min/m2 over the wetted area
regardless of type of fire. Is this for pool fire, jet fire, impinging/engulfing,
non-impinging/non-engulfing?
FABIG Technical Note 8 (2005) Same as for other equipment above.
OTI 92 607 Same as for other equipment above.
NORSOK does not specifically address fire protection of horizontal structural steel.
EN ISO 13702 gives 4 litres/min/m2 .
What type of fire is this for?
NFPA 15 (2007 version) states that vertical structural steel should be protected at a rate
of 10.2 l/min/m2 regardless of type of fire. Is this for pool fire, jet fire, impinging/
engulfing, non-impinging/non-engulfing?
FABIG Technical Note 8 (2005) Same as for other equipment above.
OTI 92 607 Same for other equipment above.
NORSOK does not specifically address fire protection of vertical structural steel.
EN ISO 13702 gives 10 litres/min/m2 .
What type of fire is this for?

The tests concluded that the rate of 10.2 l/min/m2 did

not maintain a water film over the whole tank surface
in an approximately 2 kg/s flashing liquid propane jet
fire. At the highest delivery rate possible in the tests of
approximately 23 l/min/m2 the water deluge appeared
close to giving protection against the size of the jet fire
used.
The information given in Ref. 12 is principally concerned with LPG vessels. It states a deluge rate in
the range of 250 to 500 gallons/min (946 litres/min
to 1893 litres/min (assumed American gallons, 1 US
gallon = 3.785 litres)) for an impinging high velocity
jet flame.
For pool fires, work described in Ref. 13 refers to
crude oil fire deluge tests in which the test rig was
engulfed in flame. It was shown that the deluge provided significant cooling to the rig and roughly to
halves the temperature after 80 seconds. The tests were
completed with deluge rates ranging from 12 l/min/m2
up to 28.6 l/min/m2 . The use of area deluge was
concluded to be the best.
A study of the transient processes where fire engulfment of process equipment occurs in a refinery found
that, for conventional deluge systems, escalation could
occur for jet flame impingement cases at the very
moment fire fighters are setting up a large water battery for high density deluge to be applied manually,
putting the fire fighters lives at risk (Ref. 14). While

it is known that a fire-fighting water monitor could

be effective if aimed directly at the impinged area
(Ref. 15), setting up the monitor needs to be manually (unless done remotely) and thus there is potential
exposure of personnel.
It is clear that the various generic deluge rates specified in the existing Standards and Guidelines may be
inadequate for some common fire situations. While
it is known that deluge can delay the escalation of
fire events, it is not clear if the delay is sufficient
for escape and evacuation of personnel, and for the
protection of fire fighters. This may result in unsafe
design. Improved protection for people and plant can
be achieved if adequate information is generated for
design. This will require time and money so that full
scale tests and/or CFD analyses and dynamic simulations on the response of structures to fires can be
carried out, and prediction models developed for practical design applications. An approach that can be used
in the interim is suggested in the next Section.

SUGGESTED INTERIM APPROACH

FOR THE DESIGN OF DELUGE SYSTEMS

As mentioned above, due to the high kinetic energy of

jet flame, the jet flame will push deluge droplets out

1243

1244

Process areas
and auxiliary
equipment (incl.
pressure vessels,
pumps and
compressors)

Wellhead
areas

Wellheads

Objects to be
protected

Design deluge rate

Jet fires from small and medium releases

Yes
Yes
Yes
Jet fires from large releases
Yes
Yes
Yes
Jet fires from small and medium releases
No
Yes
No
Jet fires from large releases
No
No
No
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No

400 litre/min/m 2

10

20

20

10

Jet fires from small and medium releases

Yes
Yes
Yes
Jet fires from large releases
Yes
Yes
Yes
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No

First
impingement

400 litre/min/m2

Subsequent
impingements

Jet fires from small and medium releases

Yes
Yes
Yes
Jet fires from large releases
Yes
Yes
Yes
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No

First
impingement

No
No

No

No

No
No

No

No

Yes
Yes

Yes

No

No
Yes

No

No

Yes
No

Yes

No

No
Yes

No

No

Yes
No

Yes

Extinguishment
of fire

Yes

Subsequent
impingements

Cooling of the object to be protected

400 litre/min/well
or
400 litre/min/m2

Deluge rates in
stnds/guids
(litres/min/m2 )
(unless stated
otherwise)

Reduction of fire intensity

Jet fire

Table 2. Summary of suggested deluge protection against impinging/engulfing fires.

Cooling
of the
objects
to be
protected

Small and medium fires

Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes

Small and medium fires

Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes

Small and medium fires

Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes

Reduction
of fire
intensity

Pool fire

No

No

No

No

No

Yes

No

No

No

Yes

No

No

No

Yes

Extinguishment
of fire

1245

Vertical
structural
steel

Horizontal
structural
steel

Jet fires from small and medium releases

Yes
Yes
Yes
Jet fires from large releases
Yes
Yes
Yes
Jet fires from small and medium releases
No
Yes
No
Jet fires from large releases
No
No
No
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No

10

20

No
No

No

No

No
No

No

No

Yes
Yes

Yes

No

No
Yes

No

No

No
No

No

No

No
No

No

Yes

No

Yes

400 litre/min/m2

10

20

No
Yes

Jet fires from small and medium releases

Yes
Yes
Yes
Jet fires from large releases
Yes
Yes
Yes
Jet fires from small and medium releases
No
Yes
No
Jet fires from large releases
No
No
No
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No
Jet fires from small and medium releases
No
No
No
Jet fires from large releases
No
No
No

400 litre/min/m2

No

Yes

Jet fires from small and medium releases

No
No
No
No
Jet fires from large releases
No
No
No
No

Small and medium fires

Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes

Small and medium fires

Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes
Small and medium fires
Yes
Yes
Large fires
Yes
Yes
Small and medium fires
No
No
Large fires
No
No

Small and medium fires

No
No
Large fires
No
No

No

No

No

No

No

Yes

No

No

No

No

No

No

No

Yes

No

No

Figure 3.

Computer Plot Representation of Impingement of Gas Jet Flame onto an Object as Obtained from Brilliant.

of its path and away from the surface that the deluge
is supposed to cool. The flame kinetic energy will be
lower for small releases than for large releases. The
kinetic energy will be higher upstream of the jet flame
than downstream of the flame. Also, most of the jet
fires will be obstructed, where:
The part of the jet flame before it impacts on a first
obstacle will have a high momentum, but a large
part of the momentum may be lost by the impact,
depending on the size and shape of the obstacle; and
The part of the jet flame downstream of the impact
will have lower momentum than the part upstream
of the impact.
It is therefore possible that deluge of a rate that will
be pushed away by the flame upstream of the impact
will be effective (i.e. not pushed away) downstream of
the impact.
Taking this into consideration and due to lack of
other data, Table 2 (Ref. 16) has been developed qualitatively for considerations in risk assessments before
any quantitative information and/or calculation models will become available. Table 2 takes deluge rates
from the Standards and Guidelines in Table 1 and provides information on what fire protection can or cannot
be practically achieved for various fire scenarios.
Table 2 covers scenarios where the deluge is applied
into the flame to reduce the flame intensity and onto

surfaces of equipment and structures to provide cooling. The Table does not address the situation whereby
the deluge is applied into the area between the flame
and the object or person to be protected from thermal
radiation. The radiative heat flux received by an object
or person reduces exponentially with the distance from
the heat source. The effectiveness of protection against
thermal radiation therefore strongly depends on how
far the threatened object or person is from the fire.
In a review on fire research in offshore installations,
Dalzell (Ref. 17) has stated that the design of deluge
systems requires much more focus with the assignment
of specific roles, the cooling of exposed critical equipment, the cooling and control of pool fires through the
suppression of fuel vaporisation, and the interaction
with combustion in the high level hot zone. These
require totally different design philosophies and one
catch all design cannot achieve them all.

POSSIBLE WAY FORWARD

The deluge rates for adequate fire protection should be

obtained from full scale tests, but the fire scenarios and
adequate deluge rates to mitigate them represent many
various permutations, which can be time-consuming
and costly to represent by the tests. An efficient

1246

approach for the determination of these relationships

is the application of Computational Fluid Dynamics
(CFD) method such as that developed in the multiphysics computer program Brilliant (Ref. 18). Figure 3
shows impingement of a large gas jet flame onto an
object, which was obtained from Brilliant (Ref. 19).
The arrows in Figure 3 represent gas velocity of
the burning gas; unburnt gas near the gas nozzle and
burning/burnt gas at a distance away from the nozzle.
As there is a large variation of the gas velocity, the
arrows were scaled down to fit the picture.
The computer program Brilliant has already been
verified against full scale tests. Additional verification
can be carried out for the specific application of deluge
droplets entrained into the burning gas or other fluid
stream. CFD-based analyses can then be performed
with confidence and data on which deluge rates are
suitable for which fire scenarios can be obtained at a
fraction of costs of full scale tests.
Guidance on design of pressure vessels subject to
severe fires has been suggested in a publication by
Institute of Petroleum (Ref. 1), which covers higher
deluge densities, higher wall thickness of vessels
and faster blowdown as possible options for optimal
design.

CONCLUSIONS

The following conclusions have been made based on

the above information:
1. The Standards or Guidelines for the application of
deluge in the hydrocarbon production and processing industries are generic, unclear on the nature of
fire it is designed to protect against, and not fireor equipment-specific.
2. The current information available may lead to
under- or over-protection, where the underprotection may lead to unsafe design. Overprotection will result in design of deluge systems
and firewater pumps that are larger than necessary,
with higher capital and maintenance costs.
3. The gaps in information on fire protection by deluge need to be addressed by full scale tests or
verified CFD-based multi-physics computer analyses, where the results of the tests or analyses can
be used for the development of fire protection standards and guidance applicable to protecting people
and plant for various types of fires.

[2] Bennet, J.F., Shirvill, L.C., and Pritchard, M.J., Efficiency of Water Spray Protection Against Jet Fires
Impinging on LPG Storage Tanks, HSE Contract
Research Report, 137, 1997.
[3] The Effectiveness of Area Water Deluge in Mitigating
Offshore Fires; Safety in Offshore Installations 1999;
Gosse A.J., and Evans, J.A.
[4] NFPA 15, Standard for Water Spray Fixed Systems
for Fire Protection, 2007 Edition.
[5] Availability and Properties of Passive and Active Fire
Protection Systems, Prepared by the Steel Construction Institute for the Health and Safety Executive, OTI
92 607.
[6] Protection of Piping Systems Subject to Fires and
Explosions, FABIG, Technical Note 8.
[7] Technical Safety, NORSOK S001, Revision 3, January 2000.
[8] Petroleum and Natural Gas Industries. Control and
Mitigation of Fires and Explosions on Offshore Production Installations. Requirements and Guidelines,
EN ISO 13702, June 1999.
[9] Interim Guidance Notes for the Design and Protection
of Topside Structures against Fires and Explosions,
The Steel Construction Institute, Publication Number
SCI-P-112, 1992.
[10] Loss Prevention in the Process Industries, Hazard
Identification, Assessment and Control, Second Edition, Frank P. Lees.
[11] Directed Deluge System Designs and Determination
of the Effectiveness of the Currently Recommended
Minimum Deluge Rate for the Protection of LPG
Tanks, T.A. Roberts, Health and Safety Laboratory,
Journal of Loss Prevention in the Process Industries
17 (2004) 103109.
[12] Fire Protection Considerations for the Design and
Operation of Liquefied Petroleum Gas (LPG) Storage
Facilities.
[13] FABIG Newsletter 11, December 1994.
[14] Raman, R., Accounting for Dynamic Processes in
Process Emergency Response Using Event Tree Modelling, Paper presented at the 19th CCPS International
Conference, Orlando, Florida, 2004.
[15] Roberts, T.A., Medonos, S. and Shirvill, L.C., Review
of the Response of Pressurised Process Vessels and
Equipment to Fire Attack, Offshore Technology
ReportOTO 2000 051, Health & Safety Executive,
2000.
[16] Consideration of Deluge Systems in Risk Assessment,
an internal research project, Petrellus Engineering Ltd,
2007.
[17] Dalzell, G., Offshore Fire ResearchA Summary,
Hazards VXIIIInherent safety, 2006.
[18] Berge, G., Theory and Description of the Structure of
the Multi-Physics Simulation System Brilliant, Petrell
as, 2006.
[19] Berge, G., Multi-Physics Simulation of Big Scale Jet
Fire Experiments, Petrell as, 2004.

REFERENCES
[1] Guidelines for the Design and Protection of Pressure
Systems to Withstand Severe Fires, The Institute of
Petroleum, March 2003, ISBN 0 85293 279 0.

1247

Risk control in complex environments

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Is there an optimal type for high reliability organization?

A study of the UK offshore industry
J.S. Busby & A. Collins
Lancaster University Management School, Lancaster, UK

R. Miles
Health and Safety Executive, Bootle, UK

ABSTRACT: Work on so-called high reliability organization suggests that it is highly generalized qualities of
behaviour like collective mindfulness, not particular choices of organizational structure, that confer very high
levels of reliability on inherently vulnerable and hazardous organizational activities. Similarly, actors in such
organizations seem to stress the capacity to refine and perfect a chosen way of organizing, rather than choosing a
specific way of organizing in the first place. Our aim in this study, which is part of a larger programme of work,
was to examine what this means in the context of the offshore hydrocarbons industry. In particular, we propose
the idea of rigour as a general concept for encapsulating the capacity to make whatever model of organization is
chosen a safe and reliable one. The paper describes an attempt to characterize what rigour means in this context
by analyzing a dataset of offshore incident reports, in which the diagnosis of the report writers were interpreted
in terms of shortfalls in rigour. This led to a characterization of rigour in terms of multiple aspects including,
for example reflexiveness (the capacity to reflect on the limitations and consequences of the activity itself) and
scepticism (the capacity to be aware of and inquisitive about potential discrepancy and problem). Some of these
aspectsfor example situatedness and uniformityoften appear contradictory in particular situations, however.
So rigour has to be seen recursively: its various aspects are brought to bear in particular situations in a way that
is itself rigorous.

INTRODUCTION

Recent years have seen the entry of small, low-cost

operators without large technical staffs, long historical experience and extensive standards to intrinsically
hazardous industries, such as aviation and offshore
hydrocarbons extraction. Perhaps unexpectedly there
appears not to have been a marked increase in failureseither as near-misses or fully-developed breakdowns. This raises basic questions about whether high
reliability organization (Roberts, 1990; La Porte and
Consolini, 1991; La Porte, 1996) requires the adherence to one particular model, or whether there are
multiple models that achieve similar levels of reliability. It also raises questions about how regulators
deal with hazardous operations. How far, for instance,
should they insist on codified practices and standards,
how much technical expertise should they expect
within the operating organization, and how much can
they rely on formal processes of risk assessment? Past
studies point to the importance of a regulatory contribution to high reliability organization that goes well

beyond merely enforcing compliance (La Porte and

Thomas, 1995).
The purpose of the study we describe here has been
to study the problem of new entrants in the UK offshore
industry, and to determine how significantly different
forms of organization perform in maintaining reliability. Our interest is particularly in installations which
were formerly operated by large, well-established
operating companies, but are now being bought by
much newer and smaller entrants to the industry, often
at a time that is close to the end of the installations
design life.
However, it is probably not objectively determinable whether new entrants are worse, better or the
same because there are few of them and the events of
interestlarge scale catastrophic failuresare rare.
There are also many confounding factors, since the
quality of being a new entrant tends to be associated
with other qualities that might influence reliability but
which are not inherent in the newness of the operator. Even in domains such as aviation, where there is
a good deal more data on the relative performance of

1251

established firms and new operators following deregulation, analysis tends to be inconclusive (for example
Bier et al, 2003; Oster et al, 1992). Moreover, even if it
could somehow be determined that being a new entrant
made an operating firm more reliable or less reliable,
this knowledge could not easily be acted on. Firms cannot simply be made into new entrants or established
operators, and cannot simply have the qualities of new
or established operators imposed on them. From a regulatory standpoint, if a company is operating within
the law the regulator has no basis for telling it to change
its nature or practice.
In fact, when we look at the understandings of
experienced, knowledgeable and reflective observers,
these seem to indicate that the type of organization
including the question of whether it is a new entrant
or established firmis not central to explanations of
its reliability. In particular, in a small number of interviews that preceded this work, regulatory inspectors
pointed to a somewhat vaguer, but deeper, explanation
of reliability. This essentially concerned the capacity
of organizations to recognize the implications of whatever ways of organizing they had chosen, to refine
whatever practices they had chosen, and to follow
through as it were on their choices. It was less the particular choices they made (for example about whether
to subcontract certain activities or operate them internally), and more the thoroughness with which those
choices were then organized and managed. This is
not to say that the choices were arbitrary. But the
basic feeling was that they were less instrumental in
achieving reliability than the capacity to make those
choices work. For example, in some cases the organization owning an installation was quite different from
the organization that operated it. What mattered to
safety, in the experience of some actors at least, was
not whether ownership and operation were separated,
but whether the consequences of this separation were
fully understood and managed.
Part of our study of this problem has involved
the analysis of historical incidents and accidents, as
explained in more detail in section 3. The idea was
to analyse the way in which official sense was made
of these incidents in terms of what it said about this
notion of making choices work.
2

LITERATURE

The relevant literature is that on high reliability organizations (or HROs). This literature points to a number
of ways in which social organization solves the problem of making inherently hazardous systems acceptably reliable. It refers to redundancy and slack of
various kinds (for example Roberts, 1990; Rochlin
et al, 1987; Schulman 1993). It places considerable
stress on learning from simulated rather than concrete

experience (Weick, 1987), on the continuous search

for improvement (La Porte, 1996), and on the avoidance of blame (La Porte, 1996; Grabowski and
Roberts, 1997). It refers to notions of heedful interrelating (Weick and Roberts, 1993), collective mindfulness, (Weick et al, 1999) and extraordinarily
dense patterns of cooperative behavior (La Porte,
1996). And it draws on the idea of cultures of reliability
(Weick, 1987) and flexible, expert-centred authority
structures (Roberts et al, 1994; La Porte and Consolini,
1991; La Porte, 1996; Bigley and Roberts, 2001).
But in this literature there is also a recurring theme
of strength in adversity. What gets an organization
from having merely ordinary levels of reliability to
high reliability is, counter-intuitively, an environment
that is particularly testing. Or it is the apparently
counter-productive and self-confounding behavior of
the organization itself. Thus, for example, Weick
(1987) points out that reliability tends to increase when
performance pressures are high, rather than low, as
can be seen by the fall in error rates when air traffic
control centers are handling heavy workloads. It turns
out that it is in conditions of heavy, rather than light,
load that air traffic controllers sweep the entire radar
screen. Rochlin et al (1987) essentially say that the
instability and turbulence produced by regular crew
turnover and rotation on naval vessels is a source of,
not impediment to, reliability. The continual retraining
that this requires seems to avoid the loss of vigilance
that Rochlin et al associate with more stable civilian organizations. It also leads to a regular scrutiny
of operating procedures that makes them more robust.
And the movement of people about the naval organization rapidly diffuses lessons learned from practice.
Early ideas about HROs rather contradicted the notion
of strength in adversity, referring for instance to the
buffering of environmental surprises and protecting the stability of a technical core (La Porte and
Consolini, 1991). But subsequent work emphasized
accepting and dealing with fluctuations instead of trying to remove them, seeking resilience rather than
invariance, and tolerating and even protecting organizational ambiguities in order to cope with unforeseen
events (Schulman, 1993; Roe et al, 2005).
Although HRO theorists do not particularly make
the connection, this idea of strength through adversity seems explicable when we look at the quality of
collective mindfulness that Weick et al (1999) particularly stress. Adversity begets reliability because it
provokes a compensating, or over-compensating reaction, provided there is mindfulness of this adversity.
This notion of reflecting on what makes a setting
adverse is particularly striking when the adversity is
of your own makingwhich is seen most clearly in
Rochlin et als (1987) examples of what the US navy
does to make life difficult for itself, yet seems to profit
from.

1252

The idea of being mindful also reflects what Weick

wrote earlier (Weick, 1979): Most managers get into
trouble because they forget to think in circles. Managerial problems persist because managers continue
to believe that there are such things as unilateral causation, independent and dependent variables, origins,
and terminations. Reliability is not simply a function
of a series of identifiable factors, whether internal to
an organization or external: it is equally a function of
how those factors are interpreted and responded to; and
even a function of how the responses are responded to.
Thus forgetting to think of organization in this circular
pattern seems likely to lead to the view that the problem of reliability is simply making the right choices of
organizational structure or processwhereas a much
better view would be that reliability is obtained by
continually acting and reflecting in order to act again.
This takes us very close to the point we reached in
the Introduction: the thesis that what seemed to matter was making your choices work, rather than making
the optimal choices in the first place. The particular
point about strength from adversity is that continuing
adverse conditions are a stimulus to take this more
circular approach to achieving reliability.
However, it seems likely to us that if you asked
managers how their organizations work they would
probably not be so simplistic as to go around claiming the world consists of independent and dependent
variables. They are quite likely to acknowledge that
every solution is equally a source of problems, and
that reasoning circularly is an important aspect of how
organizations should be managed. So rather than this
circularity being unknown, it is probably truer to say
that the conditions of organization tend to make actors
behave as though they did not believe in circularity.
Bounded rationality, institutionalisation, negotiation
and so on tend to bring about situations in which
people do not think in circles. People do not think
about howwhen they introduce a procedure to avoid
some particular kind of accidentthis same procedure could produce a different kind of accident: and
they do not do so because inter alia they need to be
seen to be doing something, they need to be seen to
be decisive not equivocal, their role means that they
have a higher authority than those who have to follow the new procedure, and so on. In other words,
organization creates the conditions in which actors end
up behaving as though reliability problems were linear, whether or not those actors really believe them
to be so.
The main challenge in achieving organizational
reliability becomes the overcoming of these influences, and the consciousness of the need to continually
reason about the consequences of your commitments.
We have described this overcoming as rigour. The
reason for not simply reusing the term mindfulness
as found in Weick et al (1999) and Weick and Sutcliffe

(2006) is that it is not just about being mindful, it

is also about acting on this mindfulness. We can see
this in Weick et als (1999) formulation of mindfulnesswhich is strongly cognitiveset beside Sagans
strongly political thinking (1994). Organizations that
are political do not see an automatic translation of
mindfulness into action, as in fact Rijpma (1997)
points out. Learning starts out as cognition, but then
has to become politics in order to have any effect.
Therefore we have stressed not so much a mindful
way of thinking as a rigorous way of acting a way
of acting that, in particular, involves attending to the
consequences and entailments of whatever ways of
organizing have been chosen by the actors in question. The rigour label is the closest we could get to
this idea. Rigour also has the advantage of not privileging any particular source of reliability-enhancing
factors. It does not say that the source has to be culture, or psychology, for example, because in principle
either could be a source of rigorous notions of acting,
as could rule systems, or physical devices. There may
be nothing to stop us particularly building into physical devices the capacity to detect the results of their
action and review those results, and in some sense all
closed loop devices do this.
Again then we are led to a working thesis that what
matters in the achievement of reliability is less the
choice of ways of organizing, and more the attention to and (over)compensation of the consequences
and entailments of this choice. This seems to fly
in the face of some HRO work which points to the
need for the other qualities like redundancy. But the
extent to which redundancy really confers reliability is disputed anyway (for example Sagan, 2004).
A better thesis on redundancy in fact would be a
synthesis of the redundancy-is-good thesis and the
redundancy-is-bad antithesis. This would say that reliability comes from either 1) having high redundancy
and coping with the consequences or 2) having little redundancy and coping with the consequences
of that. What matters is acting on the consequences
of having more redundancy or less redundancy, just
as much as the choice of how much redundancy
to have.
One implication of the idea that rigour is what
matters is that best practice is a misleading basis
for judging organizations trying to achieve reliability. Instead of reliability and safety being attained by
adopting some optimal kind of practice, the concept
of rigour gives us a view that that different actors can
select different ways of organizing, perhaps even for
spurious reasons, but make their chosen way work
equally well. We should not measure an organization
according to whether its practices are the best practices we have seen in the same industry, but according
to how fully it has refined its practices in the light of
analysis and experience.

1253

METHOD

Nonetheless rigour remains a vague notion and needs

elaborating on if it is not to be an empty, or circular
concept. The aim in this study was to do so by analyzing a set of secondary data: the reports of incidents and
accidents. What we have done is to construe what are
labelled as failures, or contributory factors, as being
deficits in rigour. This kind of construal means asking
of the data in what way are the consequences or entailments of the chosen way of organizing not attended
to? The answers to this can be built into a taxonomy
of rigour.
What we do not get from this is any kind of test of
whether different ways of organizing really are equifinal with respect to reliability, however. And this is
the main limitation of the analysis. It seems plausible
to say that we should look for ways of being rigorous
rather than ways of organizing, and that being rigorous
means different ways of organizing can lead to comparable levels of reliability. But this is an assumption,
not a clear truth.
We used two datasets: 1) a set of 40 incident reports
published by the StepChange organization, which are
mostly of minor incidents on offshore installations in
UK waters, analysed informally by local individuals,
and 2) a set of 34 MMS (Minerals Management Service) investigatory reports on incidents in the Gulf
of Mexico, which are much more extensive accounts
of more formally investigated cases. The StepChange
reports are typically a page or so of narrative, whereas
the MMS reports are typically 40 pages of narrative
with diagrams and photographs. There are limitations
in both datasets, and in the whole process of analyzing such reports. They may give a misleading view in
hindsight of what was really knowable at the time, and
they are the interpretations of particular individuals.
As public documents they are also very likely to be
subjected to self-censorship. Nonetheless our aim is
not to diagnose particular incidents definitively, but to
understand the texture of rigour.
The procedure was to work through each report and
look for causal or contributory factors that involved in
some way 1) choosing a way of organizing and 2) failing to manage the consequences of this choice. Each
factor of this kind was generalized upon as a kind
of rigour. For example, in one case a permit-to-work
system was not fully understood and as a result not
properly used. This was inferred to be a case of rigour
as giving substance to a nominal system. In other
words, there was a system in name but, in the absence
of an understanding how to use it and an understanding of why to use it, it remained nominal or token and
not substantive.
An attempt was then made to produce general
aspects of rigour from these specific deficits. For
example, the idea of rigour as making a nominal

system into something that was properly implemented

was seen as an example of a general idea of substantiveness. In other words, one aspect of what we mean
by rigour in an activity is that it is substantive, dealing with substance not mere form or appearance. This
process of going from particular observations to
general aspects of rigour was inductive rather than
deductiveand involved stripping away aspects of the
context in particular cases that seemed uninteresting.
This is subjective, as a particular analyst makes his
or her own judgment about what is interesting, and
it assumes that there is no objectively correct way of
categorizing some phenomenon. In many cases the
same incident could be categorized in more than one
way, so that sometimes several aspects of rigour were
implicated in one case.

RESULTS

The results of this procedure are shown in Table 1.

Each entry shows a particular aspect of rigour, in
terms of a particular quality of activity, with a brief
explanation.
Space precludes giving examples for all these qualities, but Table 2 provides some examples for what
are perhaps the most distinctive of them. Each entry
consists of a row giving the aspect of rigour (taken
from Table 1), the nature of the rigour implied in the
incident and the shortfall that was described in the
report writers diagnosis, followed by a row giving an
abbreviated description of the incident.

DISCUSSION

Some of the aspects of rigour identified in the results

are hard to specify precisely. For example, one aspect
was the idea of rigour as giving substance to a system
that is otherwise only nominal or token. A characteristic example involved the last entry in Table 2, in
which there was a permit-to-work system in operation.
Its operation was inadequately understood by a group
of workers that included specialist sub-contractors as
well as employees, it lacked specificity and seems
not to have been checked by a competent authority, nor counter-signed by the actors undertaking the
work. After the fact it is therefore possible to say in
detail what a shortfall in rigour amounted to. But
the deficits in this particular case do not represent
all possible deficits, and it seems unlikely that you
could ever spell out just what it means to turn a nominal system into a substantive one. What points to
a lack of rigour is that the situation was correctible
or improvable, not a specific prescription for exactly
what it should have been improved to. This shows

1254

Table 1.

Aspects of rigour.

Aspect of rigour Explanation

Authoritativeness

Being a source of authority, expertise

and direction, and being willing
to exercise power
Boundedness
Being constrained in your ambitions
and objectives, not extending
capabilities too far
Collectiveness Being concerned to involve people
widely, to consult, to engage in
joint decision making
Compliance
Being willing to follow agreed
norms and standards
Currency
Being up-to-date, being adapted to the
existing environment
Embodiedness Being put into a concrete, specific
form in contracts, agreements and
so on that create obligations
Engagement
Being actively involved with others, often
subordinates, within a larger system
Facilitation
Being supported by devices and
artefacts that amplify capabilities or
compensate for short-comings
Formality
Being thoughtful, formal, penetrating
and pro-found rather than shallow
and intuitive
Historicism
Being aware of historical events
and taking them into account in
current activity
Informativeness Being a source of information and
under-standing for those who have
dealings with the activity
Planfulness
Being systematic in making prior
arrangements and plans
Reflexivity
Being able to reflect on the limitations
and consequences of the activity itself
Scepticism
Being aware of, and inquisitive about,
potential problems, errors and
differences of interpretation
Seriousness
Being sufficiently concerned,
conscientious and energetic, not
being casual, passive or superficial
Situatedness
Being tailored to specific circumstances
and not just lazily inherited from
a general procedure or practice
Substantiveness Being concerned with substance,
action and deeper implications, not
being merely token or nominal
Unambiguity
Being clear and specific and leaving no
problematic or unproductive doubt
Uniformity
Being similar or consistent in different
places or on different occasions,
without unconsidered variation

things even when finding these ways might take some

experimentation. It is the commitment to acting to
give systems substance that is important to a notion of
rigour.
Some of the aspects seem to contradict one another.
For example, one is concerned with being situatedthat is, being adapted to particular circumstanceswhile another is about being uniformthat
is, being consistent or similar from one time to the
next. It is not hard to imagine situations in which
these can be construed as being contradictory. Similarly, a further aspect of rigour is being compliant
while another is being sceptical. Again these are easily construed as being contradictory. Thus it would not
be true simply to say that rigour involves an adherence
to all these qualities mindlessly or inappropriately.
They are facets of rigour that emerge as being relevant to particular situations, not constant attributes
that have a permanent, ubiquitous importance. There
is a sense in which they are selectable according to the
circumstances surrounding the activity in question.
This raises the danger of defining rigour in a circular
way. Rigour is having qualities like being sceptical, but
you need to know whether to be skeptical in a particular instance by then being reflexive (another quality).
And being reflexive could get you into trouble unless
you were then being collective (another quality) to help
determine whether this would upset other people and
cause a failure as a result. This looks like circularity:
being rigorous is doing things in a way that is ultimately rigorous. But it is better to see rigour as being
recursive rather than circular in meaning, because at
each stage of the explication you take a further step
back from concrete action. It starts with an activity
like following a procedureand says there are certain qualities of that activity which confer rigour, like
adapting the way a procedure is followed to particular
circumstances. But how you adapt to circumstances
can be done with differing degrees of rigour. Some,
for instance, are more reflexivethat is, conscious
of their own limitations, than others. And then there
are different ways of being reflexive, some again more
rigorous than others. In practice it is obviously not possible to recurse more than once or twice, and knowing
when to stop is again a particular quality of behaving.
This leaves rigour as a quality that cannot be defined in
a closed form: we have to see it as being about entering into this recursive process, and not being about
reaching a particular state.
6

that rigour is not necessarily a specifiable condition, but can also be a process that may not have
reached a particular end-point. It is important to find
ways of making systems achieve proper, substantive

CONCLUSION

The implication is that what we need to look for in an

organization, in order to judge its capacity to be reliable and stay safe, is this kind of rigour. Its various
aspects include the qualities we have listed. And in

1255

Table 2.

Examples of some of the aspects of rigour.

Quality of
activity implying
rigour

Nature of rigour
implied

Shortfall in rigour

Taking account of histories

of by-passing normal controls

Failure to review task and

permit-to-work prior to further action

Incident description
Historicism

During attempts to open a valve on a mothballed test header, a plug shot out of the valve body; personnel evacuated;
then assumed that test header isolation valve closed and decision made to repressurise line; this resulted in a second
larger hydrocarbon release from bleed port as personnel fitting a replacement plug. Review conducted to bring valve
back into service inadequate and did not identify need to replace bleed plug; previous incident involving corrosion
of bleed plug prompted replacement of several plugs on operating valves but as this valve was mothballed at the time
the plug was not replaced; following initial release, PTW [permit to work] not reviewed to ensure new valve plug
fitted, gas trapped in system isolated and vented, new or changed hazards identified, adequate controls in place,
clearly defined and communicated process steps; lack of communication between control room and personnel in
field such that people not made aware of what actions taken or about to be taken
Planfulness

Planning in detailed, collaborative

way, especially in non-routine
activity

Failure to plan unusually-configured

operation in detail in collaboration
with contractor conducting it

Tubing supporting BOP [blow-out preventor] stack assembly bent suddenly under the load causing the BOP to break
free and fall overboard. A CT technician was attached to the BOP Assembly by his safety line and was pulled
overboard. There were no engineering calculations, the temporary work platform precluded attachment of
safety lines, there was a lack of detailed planning and no discussion of potential anomaly, the contractors were not
involved in the planning who did not know the unusual nature of the set-up until reaching site, there was no JSA
meeting or shift hand-off meeting, and no oversight of the contractor by the operator
Reflexivity

Ensuring reporting activity is

informative about its own limitations

Inspection reports do not draw attention

to lack of inspection opportunity

Gas release from corroded section of riser above water-line and not accessible to internal or external inspection;
no inspection hatches on the caisson, and diameter of this section of gas riser too small to allow intelligent pigging.
Should highlight any latent defects in inspection and maintenance regimes in inspection reports; should ensure
equipment designed to provide adequate access for inspection & maintenance; should recognise that when
responding to a potential gas release the response to the initiating event can also contribute to harm to people
Scepticism

Ensuring requests are met, making

physical observations

Failure of several managers to check their

assumptions, follow up unmet requests, visit the
site and take it on themselves to be sceptical

Attempt to free stuck tubing led to parting and ejection of slips that struck an individual and led to loss of well control;
failure by the deceased to calculate adequate safety factor when determining maximum pull force; deceased acted
beyond his authority; ineffective managerial oversight at all levels; operator entirely reliant on consultants; operation
should in any case have been shut down and regulator approval sought; neither of the operators supervisors shut
down the job in spite of the safety violations and unsafe work practicesnor report these to their managers; the
operations manager approved the activity with assumptions he failed to verify and was disengaged from day-to-day
operation; he also failed to follow up an unmet request for details of the procedure; an office consultant was
misinformed about the operation but failed to make any visit and failed to recognise potential conflict of interest
with the deceased representing two firms
Situatedness

Adapting generic instructions to

specific circumstances

Failure to tailor setting to

specific well

Loss of drilling fluids to highly permeable, geologically anomalous, thick sand deposit encountered while drilling;
sand zone not identified in 4 wells previously drilled from platform, and encountered less than 100 ft from nearest
location of those wells; failure or inability to forecast presence of thick anomalous sand through use of sparker and
shallow gas hazard surveys; lack of forewarning of odd morphology contributed to failure to plan for difficult
circulation problems; generic setting depth of drive-pipe prevented emergency isolation of thief zone, possibly
set by construction department at generic depth, rather than at a depth tailored to actual well-specific
requirements determined by drilling operations
(Continued)

1256

Table 2.

(continued).

Quality of
activity implying
rigour

Nature of rigour
implied

Shortfall in rigour

Incident description
Substantiveness

Giving substance to a nominal

system

PTW system was not fully understood and not

properly usedneeds to be explicit, specific,
checked, counter-signed

Offshore supply vessel having mud and base oil tanks cleaned; oxygen levels checked and three entered the
tank with suction hose; when all base oil removed, team started to remove hose but one connection jammed; to
facilitate removal hose end section in engine room removed; pumping operation ashore stopped and storage
tanker valves closed; misunderstanding about correct sequence of operations allowed residual vacuum in
suction hose to be lost, allowing base oil remaining in hose to flow over running engine, which ignited; permit
to work system not fully understood and permit issued inadequate for the activity.

principle there does not have to be one way of organizing that achieves it. Reflexiveness, for instance, can
be institutionalized and built into activity through the
formal means of controlling the activityand this is
conspicuous in the idea of doing risk assessments. But
this is not the only way to achieve reflexiveness, since
people can choose to reflect on the consequences of
what they are about to do without institutional constraints. Of course in certain circumstances one way
of achieving it might look better than another, and
sometimes we might come to a completely consensual
view about this. But in general there is nothing in the
aspects of rigour we have identified that indicates there
is only one way, or always an optimal way, of achieving
them.
ACKNOWLEDGEMENT
Many thanks are due to staff at the UK Health and
Safety Executive for their contributions to this work.
The programme of which this was a part has been
partly funded by the HSE under contract JN3369.

REFERENCES
Bier, V., Joosten, J. and Glyer, D. 2003. Effects of Deregulation on Safety. Kluwer Academic Publishers.
Bigley, G.A. and Roberts, K.H. 2001. The incident command system: high reliability organizing for complex
and volatile task environments. Academy of Management
Journal, 44: 12811300.
Grabowski, M. and Roberts, K. 1997. Risk Mitigation
in Large-Scale Systems: Lessons from High Reliability Organizations. California Management Review, 39:
152161.

La Porte, T.R. 1996. High Reliability Organizations:

Unlikely, Demanding and At Risk. Journal of Contingencies and Crisis Management, 4: 6071.
La Porte, T.R. and Consolini, P.M. 1991. Working in Practice
But Not in Theory: Theoretical Challenges of High Reliability Organizations . Journal of Public Administration
Research and Theory, 1: 1947.
La Porte, T.R. and Thomas, C.W. 1995. Regulatory
compliance and the ethos of quality enhancement:
surprises in nuclear power plant operations. Journal of Public Administration Research and Theory, 5:
109137.
Oster, C.V., Strong, J.S. and Zorn, C.K. 1992. Why Airplanes
Crash: Aviation Safety in a Changing World. Oxford
University Press.
Rijpma, J.A. 1997. Complexity, tight-coupling and reliability: connecting normal accidents theory and high
reliability theory. Journal of Contingencies and Crisis
Management, 5: 1523.
Roberts, K.H. 1990. Some Characteristics of One Type of
High Reliability Organization. Organization Science, 1:
160176.
Roberts, K.H., Stout, S.K. and Halpern, J.J. 1994. Decision
dynamics in two high reliability organizations. Management Science, 40: 614624.
Rochlin, G.I., La Porte, T.R. and Roberts, K.H. 1987. The
Self-Designing High Reliability Organization: Aircraft
Carrier Operations at Sea. Navel War College Review,
40: 7690.
Roe, E., Schulman, P., van Eeten, M. and de Bruijne, M.
2005. High reliability bandwidth management in large
technical systems: findings and implications of two case
studies. Journal of Public Administration Research and
Theory, 15: 263280.
Sagan, S.D. 1994. Toward a political theory of organizational reliability. Journal of Contingencies and Crisis
Management, 2: 228240.
Sagan, S.D. 2004. The problem of redundancy problem [sic]:
why more nuclear security forces may produce less nuclear
security. Risk Analysis, 24: 935946.

1257

Schulman, P.R. 1993. Negotiated order of organizational

reliability. Administration and Society, 25: 356372.
Weick, K.E. 1987. Organizational Culture as a Source of
High Reliability. California Management Review, XXIX:
112127.
Weick, K.E. and Roberts, K.H. 1993. Collective Mind in
Organizations: Heedful Interrelating on Flight Decks.
Administrative Science Quarterly, 38: 357381.

Weick, K.E. and Sutcliffe, K.M. 2006. Mindfulness and the

quality of organizational attention. Organization Science,
17: 514524.
Weick, K.E., Sutcliffe, K.M. and Obstfield, D. 1999.
Organizing for high reliability: processes of collective
mindfulness. Research in Organizational Behavior, 21:
81123.

1258

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The optimization of system safety: Rationality, insurance,

and optimal protection
R.B. Jongejan & J.K. Vrijling
Delft University of Technology, Delft, The Netherlands

ABSTRACT: The optimization of system safety involves a tradeoff between risk reduction and investment
cost. The expected utility model provides a rational basis for risk decision making within the context of a
single decision maker. Yet the costs of risk reduction and the consequences of failure are generally distributed
over a number of individuals. How to define optimality in terms of (different) individual preferences? This
paper applies insights from welfare and financial economics to the optimization of system safety. Failures of
technological systems are typically uncorrelated with market returns, so that technological risks couldat least
theoreticallybe financed against expected loss. It will be shown that mechanisms for risk transfer such as
insurance sometimes allow us relax assumptions that would otherwise have to be made concerning the exact
shape of individual utility functions and the interpersonal comparability of utility. Moreover, they allow us
to lower risk premiums as insurance premiums are typically lower than certainty equivalents. The practical
significance of these results is illustrated by a case study: the optimal design of a flood defense system.
Keywords:
1

system safety, optimization, expected utility, insurance.

INTRODUCTION

Quantitative risk analysis is frequently used as a

basis for balancing investments against risk reduction, a practice that is often linked to the expected
utility framework. Von Neumann and Morgenstern
(1944) and Savage (1974) construct a utility function
that describes the preference ordering of a rational individual, and show that the individual, faced
with uncertainty, ranks actions on the basis of the
expected utility of their consequences. The idea that
people maximize the expected utility of wealth rather
than expected wealth itself was first introduced by
Bernoulli (1738). Amongst other, the Joint Committee on Structural Safety has embraced the expected
utility framework for the comparison of decision alternatives (Faber et al., 2007), it underlies the Life Quality
Index methodology that can be used to optimize measures that increase life expectancy (Rackwitz, 2004),
and it serves as the basis for the optimization of
flood defenses in the Netherlands (Van Dantzig, 1956;
Vrijling et al., 1998).
Unfortunately, the expected utility model is not as
rational as is often suggested when it is applied in
a context in which a single decision affects multiple
individuals. It will be shown however that the complexities of collective choice can sometimes be avoided
when mechanisms for risk transfer operate efficiently.
Moreover, it will be shown that mechanisms for risk

transfer, such as insurance, can have a profound impact

on the optimization of system safety.
The text is organized as follows. First, the expected
utility framework is introduced to discuss the interplay
between insurance and prevention for the simple one
person case. Insurance theory typically deals with this
subject out of concern for moral hazard (e.g. Arrow,
1971; Kunreuther, 1996), but the underlying logic
applies equally to the optimal design of technological
systems such as flood defenses. The expected utility framework is then extended to a setting in which
a single decision affects multiple individuals. A case
study, the optimal design of a flood defense system,
will finally illustrate the practical implications of our
results.

2
2.1

THE ONE PERSON CASE

The rational utility maximizer

Let us first consider a single, rational decision maker

that lives in a world of perfect certainty. Denote
the admissible set of prospects by a. The individuals utility function U would rank these prospects
such that:
ai aj if and only if U (ai ) = U (aj )

1259

(1)

But how to rank prospects when outcomes are

uncertain? Under the expected utility theorem, a rational decision maker ranks prospects on the basis of
their respective expected utilities (Von Neumann &
Morgenstern, 1944). Let a now be a set of risky
prospects. The decision makers utility function would
now be such that:

by maximizing the expected utility E[U (x)] of total

income x, which is a function of coverage c:
E [U ] = p0 U (w q + c zc)
+ (1 p0 )U (w zc)

(4)

It should be noted that this utility function is not

unique: we could multiply it by positive constant, or
change its base point, without changing the ranking
of alternatives. This has important implications when
we are to consider the case of decisions that affect
multiple individuals. But before considering decisions
in a multi-actor setting, let us first consider the relations between insurance and optimal prevention for the
relatively simple one person case.

When the individuals utility function is concave

(risk averse behavior), the optimal amount of insurance coverage can be found by differentiating E[U (x)]
with respect to c. The optimum c is bound between
zero and maximum loss (after Kunreuther & Pauly,
2006). When c = 0, the rational utility maximizer would not purchase insurance coverage. And
when c > q, he or she would purchase full coverage (c = q) as it is impossible for claims to exceed
total loss q. Note that the outcome c > q can only be
obtained when insurance is available at a subsidized
rate as it implies that coverage is offered at a price
below expected loss.

2.2

2.4 Optimal prevention with insurance



ai aj if and only if E [U (ai )] E U (aj )

(2)

Optimal prevention without insurance

Consider an individual that is confronted with a probability of losing q dollars due to the failure of a
technological system. He or she can reduce this probability from unity to pm by investing m dollars in risk
reduction, where the probability pm of losing q dollars is a strictly decreasing function in m. Denote
initial wealth by w. The optimal investment m can
now be determined by maximizing the expected utility E[U (x)] of total income x, which is a function of
investment m.
E [U (x)] = pm U (w m q) + (1 pm )U (w m)
(3)
Differentiating expected utility with respect to m
then yields the optimal investment m . This optimal
investment depends, amongst other, on the shape of the
utility function. A risk averse individual that considers
a risk worse than expected loss should for instance
invest more than a risk neutral person. Conversely, the
certainty equivalent of a risk seeking individual would
be lower than expected loss, and such an individual
should invest less.
2.3

Optimal insurance

Let us now consider a rational utility maximizer that

contemplates an insurance purchase. The individual is
faced with a fixed probability p0 of suffering a loss q.
Denote the amount of damage that can be reclaimed
under the insurance contract by c, the annual cost per
unit coverage by z, and initial wealth by w. The optimal amount of coverage c can now be determined

The loss probability was treated as a constant in the

previous sub-paragraph. Let us now assume that it
depends on the investment in risk reduction (m), as
in subparagraph 2.2. Note that the cost per unit coverage (zm ) will depend on pm , and hence on m. It will
after all be less costly to purchase insurance coverage when the probability of suffering a loss is 1/10000
rather than 1/10. The cost per unit coverage will thus
be a strictly decreasing function in m of the investment in risk mitigation (dzm /dm < 0). To determine
the optimal amount of insurance coverage (c ) and the
optimal investment in risk reduction (m ), we again
have to maximize the expected utility of total income
E[U (x)], which is now a function of m and c:
E [U ] = pm U (w q + c zm c m)
+ (1 pm )U (w zm c m)

(5)

The outcome of the optimization depends on the

cost of insurance coverage. But what insurance premium could be expected? The failure of technological
systems is generally uncorrelated with market returns.
When mechanisms for risk transfer would operate efficiently, technological risks would cost expected loss.
With full insurance costing pm q, equation (5) could be
rewritten to:
E [U ] = U (w pm q m)

(6)

We would now be able to determine an optimal

investment in risk mitigation without having to worry
about the exact shape of the decision makers utility
function (as long as it is not convex). In fact, the decision problem could be formulated as if our decision

1260

maker was risk neutral (note also the irrelevance of

initial wealth w):
Max [E [U ]] = Max [m pm q]

(7)

Besides the advantage of no longer having to specify

a utility function, another important result is that we
would overinvest in risk reduction if were to ignore the
opportunities for risk transfer. Capital markets allow
us to diversify exposures and thereby reduce the cost
of risk bearing (provided markets operate efficiently
and people are risk averse). An investment in system
safety thus cannot be evaluated without concern for
the way potential losses are borne.

3
3.1

THE BENEVOLESCENT DESPOT:

THE MULTI-ACTOR CASE
The optimization of total output
without insurance

Many decisions concerning the safety of engineered

systems affect a public rather than a single individual.
Infrastructures such as tunnels, flood defenses and
high speed rail links are generally paid for through
general taxation, requiring all taxpayers to contribute.
And failures of technological systems often affect multiple individuals as well. How to decide on an optimal
investment in risk reduction if this decision is to be
based on individual preferences?
The single decision maker framework could be
extended to this multi-actor setting by considering
what a benevolescent decision maker would decide
given the publics appreciation of every decision alternative. But the public is merely a collection of individuals. When every individual would prefer prospect
a1 over a2 , we could say without much difficulty that
the public prefers a1 . But what if the public is divided
and ten percent of the population preferred a2 ? The
approach that is commonly taken to overcome this
dilemma is to consider the aggregate or populationaveraged costs and gains of different prospects. This
practice is based on the Kaldor-Hicks potential compensation criterion that allows us to consider total
output and distribution apart (Pearce & Nash, 1982).
But despite its widespread use, Arrow cautions us
that there is no meaning to total output independent
of distribution (Arrow, 1963: 40). Let us evaluate
this statement in some greater detail and see what its
implications are for the optimization of system safety.
As discussed in paragraph two, a persons utility
function is defined up to a positive affine transformation under the expected utility theorem. Without
a single scale and the possibility to meaningfully
compare the intensities of peoples likes and dislikes, it would be impossible to construct a ranking of

alternatives on the basis of total public pleasure. But

how to meaningfully add peoples pleasures and pains?
The aggregation of individual preferences presupposes
cardinal utility and the interpersonal comparison of
utility, two assumptions that are not uncontested. If,
however, these assumptions are accepted, a benevolescent despot who would be solely interested in
the aggregate utility of the population would rank all
alternatives such that:
ai aj if

n

k=1

E [Uk (ai )]

n



E Uk (aj )

(8)

k=1

Where n denotes the number of individuals that

would be affected by a decision. Unfortunately, additional assumptions are unavoidable if we are to evaluate total output using a single utility function that is
based on individual preferences. To see this, let us consider two individuals that are exposed to a hazard that
could cause them to lose q1 and q2 dollars respectively
(Q = q1 + q2 ). They can reduce the probability of a
loss from unity to pM by investing m1 and m2 dollars
in risk reduction (M = m1 + m2 ), with dpM /dM < 0.
Denote their initial wealth by w1 and w2 respectively
(W = w1 + w2 ). The individual investments that maximize the aggregate expected utility of income follow
from the maximization of:
E [U1 (x1 ) + U2 (x2 )] = E [U1 (x1 )] + E [U2 (x2 )]

(9)

Or
E [U1 (x1 ) + U2 (x2 )] = pM U1 (w1 m1 q1 )
+ (1 pM )U1 (w1 m1 )
+ pM U2 (w2 m2 q2 )
+ (1 pM )U (w2 m2 )
(10)
E [U1 (x1 ) + U2 (x2 )] = pM (U1 (w1 m1 q1 )
+ U2 (w2 m2 q2 ))
+ (1 pM )(U1 (w1 m1 )
+ U (w2 m2 ))

(11)

We could construct a single utility function for

aggregate consequences (U1+2 ) when the individual
utility functions are linear and the same for both individuals (or when m1 = m2 and q1 = q2 ). The above
expression could then be rewritten to:
E [U1+2 (x1 + x2 )] = pM U1+2 (W M Q)

1261

+ (1 pM )U1+2 (W M )
(12)

Or (note the absence of wealth effects):

Max [pM (M Q) + (1 pM )(M )]

(13)

Though convenient, the assumption that all individuals have linear and equal utility functions is rather
unrealistic. After all, it implies risk neutral behavior,
and the assumption that the marginal utility of income
is constant across individuals. But it seems unlikely
that an extra dollar in the hands of a millionaire brings
the same amount of pleasure as an extra dollar in the
hands of a poor man.
Optimizing the safety of technological systems
is clearly not without considerable difficulty. The
expected utility framework can be extended to a multiactor setting, but only when additional (disputable)
assumptions are made. But in defense of the optimization of total output, one might argue that its
ramifications remain limited when the unfair distributive consequences of a large number of decisions
cancel out. This, however, need not be the case.
3.2

Insurance, investment costs and optimal

prevention

The presence of efficient mechanisms for risk transfer

could strongly reduce the theoretical and normative
dilemmas that surround the use of the expected utility
framework in a multi-actor setting. When insurance
markets operate efficiently, every risk averse individual would purchase full insurance coverage. The
certainty equivalent of every risk averse (or risk neutral) individual could then be replaced by expected
loss, irrespective of peoples exact risk preferences.
But to allow for further simplification, we also have
to consider the distribution of investment cost. These
will often be concentrated in the hands of a few, but
when all individuals would be charged a contribution
proportional to their expected loss, the complexities
of the optimization of public safety could be strongly
reduced. This could for instance be the case when
exposures and the taxes to pay for risk reduction would
both be proportionate to wealth. And for publicly
traded firms, the proportionality between expenditures
and exposures is rather straightforward: corporate
investments in risk reduction and the costs of unexpected events both affect shareholders in proportion
their holdings (when default risk is negligible).
When perfect insurance is available, i.e. when full
coverage is available against the actuarially fair premium (expected loss), and when expenditures are
distributed in proportion to individual exposures, the
optimization of system safety could be reduced to an
expected value optimization. To see this, let us again
consider two persons that are faced with an event with
probability p that could cause them to lose q1 and q2
respectively (Q = q1 + q2 ). They can reduce this

probability by investing m1 and m2 in risk reduction

(M = m1 + m2 ), with dpM /dM < 0. When the individuals are risk averse, they would both purchase full
insurance coverage, costing pM q1 and pM q1 respectively. When each individuals contribution to system
safety is set in proportion to his or her exposure
(insurance premium), the interests of both would be
perfectly aligned. None of them could be made better
off by making the other worse off when resources are
allocated according to:
Min [M + pM Q]

(14)

With
m1 = M

q1
q1 + q2

m2 = M

q2
q1 + q2

(15)

When mechanisms for risk transfer operate efficiently, and when costs are borne in proportion to
individual exposures, the practical and normative
dilemmas associated with the aggregation of individual preferences could be completely overcome.
We could then consider gains and losses under the
condition of risk neutrality.
It seems unlikely that any of the two conditions
that allow us to strongly simplify the optimization
of system safety will ever be completely met. While
investment cost and exposures might for instance
be distributed in a fairly proportional way, it seems
unlikely that they will ever be distributed in an exactly
proportional way. And perfect insurance presupposes that all consequences of failure can be compensated for, that insurance is offered against the
actuarially fair premium (expected loss), and that
full compensation is so frictionless that people do
not even notice that they have suffered a loss. The
optimization of risks to the public thereby typically
involves a trade-off between practicality and scientific
rigor.

CASE STUDY: FLOOD PROTECTION

IN THE NETHERLANDS

4.1 The economic decision problem

The optimal investment in a flood defense system
could be determined by minimizing the sum of the
discounted investments in flood defense and the discounted expected value of future losses (after Van
Dantzig, 1956). A simplified one-period model without economic growth or sea level rise is presented here
for illustrative purposes (it is assumed that the NPV of

1262

the costs of delayed programs exceed the NPV of the

costs of direct investment):

Min [NPV ] = Min M +

T
pQe

rt

dt

(16)

Where NPV = net present value of total cost (euro);

M = total investment in flood defense (euro); p =
probability of flood (per year); Q = total damage in
case of flood (euro); T = planning horizon (year);
r = discount rate (per year).
When overtopping is the only failure mode of
a flood defense, and when exceedance frequencies
of water levels can be described by an exponential
distribution, the failure probability p can be defined as:
p = e

ha
b

(17)

Where p = probability of failure; h = dike height

(m); a, b = constants (m). When it is assumed that
the cost of dike heightening consist of a fixed and a
variable part, equation (17) can be expanded to:

Min [NPV ] = Min M0 + M h +



T
e

ha
b

Qe

rt

dt

(18)
Where M0 = fixed cost of dike heightening (euro);
M  = variable cost of dike heightening (euro); h =
dike height (m); a, b = constants (m).
The optimal standard of protection (h ) can now
be found by differentiating NPV with respect to h.
Figure 1 shows how the net present value of total
12

10

11

NPV in euro

10

10

10

10

10

6
7
height (h) in meter

10

Figure 1. Optimal investment in flood protection under risk

neutrality; M0 = 40 106 euro; M  = 25 106 euro/meter;
a = 2.3 m; b = 0.3 m; r = 0.015/yr; Q = 10 109 euro.

cost changes when flood defenses are raised (fictitious

parameter values). The optimal height of the flood
defenses would in this case be 5.5 m.
Equation (18) presupposes risk neutrality. But
why should the benevolescent despot only consider
expected loss when citizens are risk averse? Interestingly, equation (18) shows remarkable resemblance to
equation (14) which was based on the availability of
perfect insurance and contributions to system safety in
proportion to individual exposures. Although the latter
condition seems broadly reasonable, the first deserves
a closer look.
4.2 Relaxing the assumption of perfect insurance
Flood risks in the Netherlands are highly concentrated:
a single flood could give rise to billions of euros
of damages. Such low-probability, high-consequence
events are notoriously difficult to insure and private insurance against large-scale floods is currently
unavailable in the Netherlands. The assumption of
perfect insurance might therefore seem highly unrealistic. But the assumption of perfect insurance does
not presuppose the provision of insurance by private
enterprise.
When a government is trusted to provide swift and
complete compensation for losses, the assumption of a
negligible risk premium (or: efficient insurance) could
be fully justifiable (human tragedy apart). After all,
the Dutch government has an AAA-rating so that it is
able to borrow at the risk-free rate and sell coverage
at a price that corresponds to expected loss (see also
Cummins et al., 1999 for a more elaborate treatment).
To see this, consider an unexpected loss C. If the government were to finance this loss by issuing a perpetual
at the risk free rate r (a perpetual is a bond without a
maturity), it would pay rC annually to the bondholders. The present value of the interest payments would
be rC/r = C. The loss would have been spread over
an infinite time horizon without a surcharge for doing
so. When a government is able to spread losses over
time without extra cost, the NPV of its irregular loss
pattern is equivalent to the NPV of a constant stream
of annual losses with magnitude expected loss. The
government would thus be able to sell coverage at
expected loss. Indeed, the Delta Committee assumed
that funds would be set aside to cover future losses
and it treated the economic decision problem as an
insurance problem (Van Dantzig, 1956: 280).
So far, it has been assumed that all losses could
be compensated for. But as shown recently by the
flooding of New Orleans, the human tragedy of a
large-scale flood extends well beyond the loss of property and income. Floods can disrupt communities and
families, and leave considerable trauma behind. It
would be unreasonable if not immoral to argue that
decision makers and engineers should ignore these

1263

aspects and only consider the economic impacts of the

failure of technological systems. The approach that
was followed by the Delta Committee was to multiply
potential economic damage with a constant to account
for intangible losses. Distress and loss of life were thus
assumed to be proportionate to economic loss. This
practical solution allowed decision makers to evaluate
the influence of intangible losses on their optimizations. It also showed what minimum investment in
flood protection would be appropriate, as rational individuals would at least consider their economic losses.
The decision problem was thus reformulated to:

Min M0 + M  h + K

T
e

ha
b

Qert dt

(19)

Where K = multiplication factor to account for

intangible losses.
The optimal height of a flood defense, and hence the
optimal investment in flood protection, goes up when
the multiple of expected loss increases. As shown in
Figure 2, the optimal height of the flood defense is
about 6 meters when the certainty equivalent equals
five times expected loss.
The above formulation would also be appropriate
when full insurance coverage would be available, but
only at a price that exceeds expected loss. Reinsurance premiums for low-probability, large-scale losses
are typically a multiple of expected loss (Froot, 2001).
This multiple of expected loss should not be confused with a pricing rule for (re)insurance contracts.
It is merely a descriptive metric that illustrates how
costly insurance coverage can be relative to the actuarially fair premium. It offers a relatively simple

(20)

with
T

E [U ] = M0 M h

14

K=100
K=10
K=5
K=1

13

10

12

11

10

10

10

10

6
7
height (h) in meter

Cp,Q ert dt

(21)

10
NPV in euro

Max [E [U ]]

10

10

way to account for costly (re)insurance coverage in

the optimization of technological systems: we could
replace the certainty equivalent by an insurance premium that equals a multiple of expected loss (see also
Kok et al., 2002).
The formulation of the decision problem would
have to be adjusted considerably when the availability of flood insurance (or government compensation)
would be unrealistic. We would then have to consider
each individuals utility function, and the distribution
of initial wealth, investment costs, and risks to calculate certainty equivalents and maximize total public
pleasure. After all, the use of a single utility function
for aggregate gains and losses might now yield grossly
invalid results, even when everyones risk preferences
would be exactly similar. Consider for instance the
case in which an investment to reduce expected loss
would concentrate this risk in the hands of a few. When
every individual would be risk averse, the latter situation would be associated with a greater total certainty
equivalent than the first. But a risk-averse utility function for aggregate gains and losses would suggest the
opposite.
To reduce the complexities of the decision problem
to manageable proportions, let us assume that wealth
effects are absent, that exposures and investment costs
are uniformly distributed over the population, and that
everyone has the exact same utility function. These
assumptions allow us to consider a single utility function for aggregate gains and losses (no wealth effects).
The decision problem can now be reformulated to:

10

Figure 2. Optimal investment in flood protection under K

times expected loss; M0 = 40 106 euro; M  = 25 106
euro/meter; a = 2.3 m; b = 0.3 m; r = 0.015/yr; Q =
10 109 euro.

Where Cp,Q = certainty equivalent for a loss Q with

probability p (euro).
By assuming the absence of wealth effects, we
have assumed that individuals display constant absolute risk aversion (CARA). Despite the practicality
of this assumption, it places constraints on individual risk preferences that hardly seem realistic. After
all, it seems unlikely that peoples pains would depend
on the absolute severity of losses, rather than the
severity of losses relative to initial wealth. Relaxing the assumption that wealth effects are absent thus
seems reasonable. Iso-elastic utility functions describe
behavior that corresponds to constant relative risk
aversion (CRRA) (Chavas, 2004):
U (x) = x(1)

1264

for < 1

(22)

Where denotes the Arrow-Pratt coefficient of

relative risk aversion ( = xU  (x)/U  (x)). This
coefficient can be interpreted as an elasticity: it
expresses the percentage decrease in marginal utility due to a percentage increase in wealth (Chavas,
2004). Unfortunately, the use of a CRRA utility function requires us to drastically reformulate the objective
function. Let us assume that wealth is uniformly distributed over the exposed population. Denote annual
per capita income by w. Denote every individuals contribution to flood defense by m (with m0 = M0 /n and
m = M  /n). The maximization of aggregate expected
utility now boils down to:

1
T


rt
Max np w m0 m h qe dt
0

+ n (1 p) w m0 m h

1

(23)

Solving the above formulation of the economic

decision problem would require considerable computational effort. It seems questionable whether this
effort would be worthwhile given the disputable
assumptions that again have to be made (most notably
the assumptions of cardinal utility and the interpersonal comparability of utility). The Delta Committees
relatively simple model can assist decision makers in
balancing their views on the cost of risk bearing against
the cost of flood protection. If cost-benefit analysis is
to be used to support decision makers, such an insightful model seems preferable over a more refined yet
equally disputable model.
5

CONCLUSIONS

The expected utility framework provides a rational

basis for the evaluation of risks and the optimization
of system safety within the context of a single decision maker. Yet most decisions about system safety
concern risks to a public. When decisions are to be
based on the preferences of all potentially affected
individuals, the expected utility framework has to
be supplemented by several disputable assumptions.
When analysts for instance consider population-wide
or population-averaged effects, they implicitly make
assumptions concerning the comparability of individual preferences. While the consistent optimization of
total output might make everyone better off in the long
run, this need not be the case.
When risks are unsystematic and markets operate efficiently (or: when full insurance coverage is
available against expected loss), and when individual

contributions to system safety are set in accordance

to individual exposures, it would be appropriate to
simply balance total investment cost against reductions in total expected loss. In other cases, analysts
have to make assumptions concerning amongst other
the interpersonal comparability of utility, and the
shape of individual utility functions when attempting
to optimize system safety.
Efficient mechanisms for risk transfer reduce optimal investments in system safety. This has been
illustrated for the design of flood defenses in the lowlying Netherlands: the availability of an efficient flood
insurance program would reduce optimal investments
in flood defense to a minimum. Current cost-benefit
studies for the Dutch primary flood defenses are risk
neutral, which implies that they assume that full coverage is available against the actuarially fair premium.
Since that assumption hardly seems reasonable at
present, the results of such risk neutral studies could
be considered unduly optimistic.

REFERENCES
Arrow, J.K. 1971. Insurance, risk and resource allocation. In
K.J. Arrow (ed.) Essays in the Theory of Risk-Bearing:
134143. Amsterdam and London: North-Holland Publishing Company.
Arrow, J.K. 1963. Social Choice and Individual Values.
New Haven and London: Yale University Press.
Bernoulli, D. 1738. Exposition of a new theory on the
measurement of risk (original publication: Specimen
Theoriae Novae de Mensura Sortis, Commentarii Academicae Imperialis Petropolitanas, Tomus V). Econometrica 1954, 22: 2236.
Chavas, J.P. 2004. Risk Analysis in Theory and Practice.
San Diego: Elsevier Academic Press.
Cummins, J.D., Lewis, C.M. & Phillips, R.D. 1999. Pricing
Excess-of-Loss Reinsurance Contracts against Catastrophic Loss. In K.A. Froot (ed.) The Financing of
Catastrophe Risk: 93141. Chicago and London: The
University of Chicago Press.
Faber, M.H., Maes, M.A., Baker, J.W., Vrouwenvelder, T. &
Takada, T. 2007. Principles of risk assessment of engineered systems. In 10th International Conference on
Application of Statistic and Probability in Civil Engineering (ICASP10). Tokyo, Japan.
Froot, K.A. 2001. The market for catastrophic risk: a clinical examination. Journal of Financial Economics 60:
529571.
Kok, M., Vrijling, J.K., Van Gelder, P.H. A.J.M. & Vogelsang, M.P. 2002. Risk of flooding and insurance in The
Netherlands. In Wu (ed.) Flood Defence, New York:
Science Press Ltd.
Kunreuther, H. 1996. Mitigating disaster losses through
insurance. Journal of Risk and Uncertainty 12: 171187.
Kunreuther, H. & Pauly, M. 2006. Rules rather than discretion: Lessons from Hurricane Katrina. Journal of Risk
and Uncertainty 33: 101116.

1265

Pearce, D.W. & Nash, C.A. 1982. The Social Appraisal of

Projects: A Text in cost-Benefit Analysis. London and
Basingstoke: The Macmillan Press Ltd.
Rackwitz, R. 2004. Optimal and acceptable technical facilities involving risks. Risk Analysis 24: 675695.
Savage, L.J. 1974. The Foundations of Statistics. New York:
Dover Publications Inc.
Van Dantzig, D. 1956. Economic decision problems for flood
prevention. Econometrica 24: 276287.

Von Neumann, J. & Morgenstern, O. 1944. Theory of Games

and Economic Behavior. Princeton: Princeton University
Press.
Vrijling, J.K., Hengel, W.V. & Houben, R.J. 1998. Acceptable
risk as a basis for design. Reliability Engineering and
System Safety 59: 141150.

1266

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Thermal characteristic analysis of Y type zeolite by differential

scanning calorimetry
S.H. Wu & W.P. Weng
Graduate School of Engineering Science and Technology,
National Yunlin University of Science and Technology, Douliou, Yunlin, Taiwan, ROC

C.C. Hsieh & C.M. Shu

Department of Safety, Health, and Environmental Engineering,
National Yunlin University of Science and Technology, Douliou, Yunlin, Taiwan, ROC

ABSTRACT: Volatile Organic Compounds (VOCs) are the main factors involved in environmental pollution
and global warming in industrialized nations. Various treatment methods involving incineration, absorption,
etc., were employed to reduce VOCs concentration. Activated carbon, zeolite, silica gel or alumina, and so on
were broadly used to adsorb VOCs in many applications. Zeolite has been widely applied in various industrial
fields and it has generally been used in the form of beads or pellets. Differential Scanning Calorimetry (DSC)
was utilized to analyze the thermal characteristics of our self-made Y type zeolite. This investigation was used
to assess the adsorptivity of Y type zeolite under high temperature situations. In view of pollution control and
loss prevention, versatility and analysis of recycled zeolite were necessary and useful for various industrial and
household applications.
1

INTRODUCTION

Volatile organic compounds (VOCs) are one of the

major groups of atmospheric pollutants. It has been
well known that VOCs have adverse effects on human
health, cause odorous nuisances, and participate in
photochemical reactions. Surveys of VOCs have
shown that vehicle exhausts are the dominant sources
of emissions, especially in urban areas, accounting
for one half to two thirds of total emitted VOCs.
The purification of industrial waste gases containing
VOCs plays an important role in chemical engineering (Chen et al. 2006a,b; Chen et al. 2005; Salden and
Eigenberger, 2001).
Activated carbon, zeolite, silica gel or alumina,
and so on were broadly used to adsorb VOCs in various industrial and household applications. Practically
speaking, the ideal adsorbent characteristics for VOCs
removal process are as follows:
a. a large amount of reversible adsorption capacity
(vast accessible pore volume);
b. no catalytic activity;
c. a hydrophobic property;
d. high thermal and hydrothermal stability;
e. an easy and flexible regeneration property.
Zeolite has been widely applied in various industrial fields and it has generally been used in the form of

beads or pellets. Many studies have been applied various types of zeolite to analyze the adsorption effect for
VOCs chemicals. Many investigations were employed
to analyze the adsorption capability with various formats of zeolite (Ahmed et al., 1997; Breck, 1974;
Dyer, 1988; Ichiura, et al., 2003; Yang and Lee,
1998). This study was used to appraise the thermal
characteristic of self-made Y-type zeolite in the chemical industries. Figure 1 demonstrates the self-made
Y-type zeolite manufacturing flowchart in Taiwan.
VOCs were assimilated in zeolite rotor-wheel system.
The zeolite rotor-wheel system was composed of three
parts that include the adsorption area, desorption zone,
and the recuperative thermal oxidizer (RTO). Figure 2
exhibits the cross-sectional area of the zeolite rotorwheel. The zeolite rotor-wheel was minced with two
zones that involve the adsorption area and desorption
zone. The zeolite rotor-wheel system is continuous
equipment for the VOCs handling. Here, the green
zone is an adsorption area for process gases. Red zone
is a desorbtion area that was penetrated for hot air
under the boiling point of wastes. The thermal stability
of the zeolite is an important factor under the best the
VOCs handling. The differential scanning calorimetry
(DSC) was employed to evaluate and to assess the thermal stability of zeolite and activated carbons under 200
to 600 C (after chamber furnace tests). Results showed
that the zeolite is a quite stable material in the VOCs

1267

0.1

SiO

Y- Type
Zeolite

0.0

-0.1
-0.2
-0.3

Admixture

Alkaline
solution

-0.4
-0.5
-0.6

1st Hydrothermal
synthesis

-0.7
-0.8

Zeolite (sample mass: 13.82 mg)

Zeolite (sample mass: 10.40 mg)
Zeolite (sample mass: 10.00 mg)
o
Initial temperature: 30 C
o
Final temperature: 640 C
o
-1
Heating rate: 4 C min

2nd Hydrothermal
synthesis

100

200

300

400

500

600

700

Figure 3. Sample reconstruction tests of Y-type zeolite

under at 4 C min1 by DSC.

Drying

Table 1. Heat of endothermic reaction for three tests of selfmade Y-type zeolite under at 4 C min1 by DSC.
Test no.

Sample mass
(mg)

H
(J g1 )

1
2
3

13.82
10.40
10.00

209
190
148

Calcination
Figure 1. The self-made Y-type zeolite manufacturing
flowchart from our laboratory.

and the DSC thermal scanning tests for 30 to 640 C,

and the thermal stable tests for chamber furnace.

2.1 Sample

Figure 2. The cross-sectional of self-made Y-type zeolite

rotor-wheel.

adsorption, compared with other adsorbents, such as

activated carbon.

EXPERIMENTAL DESIGN AND METHODS

This study attempted to evaluate the thermal stability

of self-made Y-type zeolite and activated carbon. The
experimental processes involves sample preparation,

Zeolite was made in five steps: mixing, gelling, forming, drying, and calcination, in the manufacturing
process. First of all, we used the SiO2 mixed with
sodium silicoaluminate powder (mixed ratio was 6)
and sodium hydroxide (NaOH) solution in the well
stirring machine. Secondly, zeolites were put into the
extruding forming machine. Finally, zeolite was dried
at 100 C for 8 hrs and calcined at 450 C in oven
(10 hrs). Figure 3 delineates the sample reconstruction
tests of zeolite under heating rate () at 4 C min1 by
the DSC. Initial reaction behavior of Y-type zeolite was
an endothermic reaction at 30 to 200 C. The final reaction has an exothermic reaction at 550 to 640 C. Heat
of endothermic reaction was calculated as displayed in
Table 1. Sample reconstruction tests of zeolite under
at 4 C min1 by the DSC were calculated about
150200 J g1 . Adsorbent (such as activated carbon
and zeolite) has adsorbed a few wastes in air. To adsorb
a great deal of wastes, adsorbent must be desorbed at
high temperature.

1268

2.2

DSC

Scanning experiments were carried out on a Mettler

TA8000 System coupled with a DSC821e measuring cell that can withstand pressure up to about 100
bars. STAR e software was applied for acquiring curve
traces. An aluminum standard pan was employed to
avoid evaporation of the zeolites and activated carbons
during the scanning experiment. For better accuracy
in testing, the scanning rates chosen for the temperature range were from 30640 C with 4 C min1 under
atmospheric air.
Reactions can be investigated within 30640 C,
for the working temperature range of this thermostat.
Under high temperature, DSC is a highly sensitive
calorimeter for stability testing of various types of
materials. DSC is regarded as a useful tool for the
evaluation of thermal hazards and for the investigation
of decomposition mechanisms of reactive chemicals
if the experiments are carried out carefully. The
experimental conditions were set as follows:
a. Heating rates: 4 C min1 .
b. Materials mass: 120 mg.
Test cell: An aluminum standard pan was employed
to avoid evaporation of the Y-type zeolites and activated carbons during scanning experiment.

RESULTS AND DISCUSSION

3.1

Analysis of zeolite

This study was used to analyze the thermal capacity of

zeolite and activated carbon for chemical industries.
Figure 4 shows the thermal capacity of zeolite under
various Isothermal temperatures by DSC. Zeolite has
two reaction behaviors. One was an endothermic
reaction at low temperature, the other was an exothermic reaction under high temperature. From Figure 4,
0.1

0.1
0.0

zeolite(isothermaltimeis10hrsat200 C)

0.0

-1

Heat flow (W g )

zeolite(isothermaltimeis0hratroomtemperature)

-0.1
-0.2
-0.3

-0.1

-0.2

-0.3

-0.4
-0.4
-0.5
0

100

200

300

400

500

600

700

100

200

300

400

500

600

700

Temperature( C)

Temperature( C)
0.1

0.1

Heat flow (W g )

-1

zeolite(isothermaltimeis10hrsat300 C)

0.0

zeolite(isothermaltimeis10hrsat400 C)

0.0

-0.1

-0.2

-0.3

-0.1
-0.2
-0.3
-0.4

0
-1
-2
-3
-4
-5
-6

zeolite (isothermal time was 10 hrs at room temperature)

o
zeolite (isothermal time was 10 hrs at 200 C)
o
zeolite (isothermal time was 10 hrs at 300 C)
o
zeolite (isothermal time was 10 hrs at 400 C)
o
zeolite (isothermal time was 10 hrs at 500 C)
o
zeolite (isothermal time was 10 hrs at 600 C)

-7
-8
-9
-10
0

100

200

300

400

500

600

700

Figure 5. The best adsorption temperature test of Y-type

zeolite under various isothermal temperatures by DSC.
Table 2. Heat of endothermic reaction for Y-type zeolite
under various isothermal temperatures (holding time was
10 hrs) by DSC with 4 C min1 of heating rate.
Isothermal
temperature
( C)
Room
temperature
200
300
400
500
600

Sample mass
(mg)

Holding time
(hr)

H
(J g1 )

13.82

10

209

11.00
9.50
8.80
10.00
14.30

10
10
10
10
10

128
146
21
104
99

zeolite owned more stable situation after 500 C in

oven (isothermal time was 10 hrs). The operating
index, such as desorption temperature, adsorption
time, absorption surface area, etc., was used to adsorb
the VOC and waste as important parameters in the
chemical industries. The best adsorbent was tested
as disclosed in Figure 5. Here, heat of endothermic
reaction for Y-type zeolite was determined as given
in Table 2. Heat of endothermic reaction for zeolite that tested under room temperature (holding time
was 10 hrs) was about 209 J g1 , indicating a relative
unstable material.

-0.5
-0.6

-0.4
0

100

200

300

400

500

600

700

100

200

300

400

500

600

700

Temperature( C)

Temperature( C)

0.1

0.1

3.2

0.0

Heat flow (W g )

zeolite(isothermaltimeis10hrsat500 C)
-1

-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7

zeolite(isothermaltimeis10hrsat600 C)

0.0

-0.2

-0.3

-0.8
-0.4

-0.9
0

100

200

300

400

500

600

700

Analysis of activated carbon

-0.1

100

200

300

400

500

600

700

Figure 4. Thermal capacity of Y-type zeolite under various

isothermal temperatures by DSC.

In general, activated carbon is a traditional adsorbent

that is widely used to adsorb the waste in water or
air, but it has a flame problem in chemical industries. The price of activated carbon is cheaper than
the zeolite. To prevent the fire accident, this study was
employed to discover the exothermic behavior under

1269

3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5

Activated carbon (isothermal time was 10 hrs at room temperature)

4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
-1.0

100

200

300

400

500

600

700

500

600

700

500

600

700

500

600

700

500

600

700

500

600

700

Activated carbon (isothermal time was 10 hrs at 200 C)

4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5

100

200

300

400
o

Activated carbon (isothermal time was 10 hrs at 300 C)

5.0
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
-1.0

100

200

300

400
o

Activated carbon (isothermal time was 10 hrs at 400 C)

4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
-1.0

100

200

300

400
o

Activated carbon (isothermal time was 10 hrs at 500 C)

4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
-1.0

100

200

300

400

Figure 9. (a), (b). The thermal decomposition situation of

activated carbon under 200 to 600 C by chamber furnace.

Activated carbon (isothermal time was 10 hrs at 600 C)

100

200

300

400

Figure 6. The exothermic onset temperature of activated

carbon under 200 to 600 C (isothermal temperature) by DSC.

60
Activated carbon (isothermal time was 10 hrs at room temperature)
o
Activated carbon (isothermal time was 10 hrs at 200 C)
o
Activated carbon (isothermal time was 10 hrs at 300 C)
o
Activated carbon (isothermal time was 10 hrs at 400 C)
o
Activated carbon (isothermal time was 10 hrs at 500 C)
o
Activated carbon (isothermal time was 10 hrs at 600 C)

50

40

30

20

10

-10
0

100

200

300

400

500

600

700

Figure 7. The endothermic behavior of activated carbon

holding at 200 to 600 C by DSC.

various isothermal temperatures by DSC. Figure 6

shows the exothermic onset temperature of activated
carbon under 200 to 600 C (isothermal temperature)
by DSC. The endothermic behavior of activity carbon was determined by DSC as illustrated in Figure 7.
Results exhibited the ideal adsorption temperature at
500 C that has a low endothermic profile. Figures 6
and 7 show unfinished reactions of activated carbon
under heating rate at 4 C min1 by DSC. Figure 8
reveals the peak of activity carbon under heating rate
at 1 C min1 by DSC. The exothermic onset temperature of activated carbon was produced at 300 C
by DSC ( is 1 C min1 ). Thus, the peak (heat of
decomposition, Hd ) of activated carbon was integrated about 19, 200 J g1 by DSC. Figure 9 shows
the thermal decomposition situation of activated carbon under 200 to 600 C by chamber furnace. Activated
carbon was used to assess the thermal decomposition
under 200 to 600 C by chamber furnace. Activated
carbon was flam under 500 to 600 C by chamber furnace. Activated carbon was burst as shown in Figures 9
(a) and (b).
4

3.0

VOCs included various kinds of wastes and chemicals, such as trichloroethylene (TCE), acetaldehyde,
phenol, 1, 2-dichloroethane (DCE), etc., in the chemical industries. Adsorption material for VOCs must
control the thermal stability and safety characteristics for boiling point of wastes and chemicals. Zeolite
is a quite stable substance from room temperature
to 650 C. Activated carbon was a dangerous adsorbent that showed flame characteristic over 500 C.
In essence, the best desorption temperature of zeolite
is 500 C.

2.5

2.0

1.5

1.0

CONCLUSIONS

Activated carbon
o
Initial temperature = 30 C
o
End of temperature = 640 C
o
-1
Heating rate = 1 C min
-1
Heat of decomposition = 19,200 J g
o
Exothermic onset temperature = 300 C

0.5

0.0

-0.5
0

100

200

300

400

500

600

ACKNOWLEDGMENT

700

Figure 8. The peak and curve of activated carbon under

heating rate at 1 C min1 by DSC.

The authors would to thank Dr. Kuo-Ming Luo for

valuable suggestions on experiments and the measurements of a runaway reaction. In addition, the authors

1270

are grateful for professional operating techniques and

information.
REFERENCES
Ahmed, S., El-Faer, M.Z., Abdillahi, M.M., Shirokoff, J.,
Siddiqui, M.A.B. & Barri, S.A.B. 1997. Production of
methyl tert-butyl ether (MTBE) over MFI-type zeolites
synthesized by the rapid crystallization method and modified by varying Si/Al ratio and steaming. Appl. Catal. A.
161: 4758.
Breck, D.W., 1974. Zeolite Molecular SievesStructure,
Chemistry and Use. John Wiley & Sons, New York.
Chen, C.L., Fang, H.Y., & Shu, C.M. 2006a. J. Air & Waste
Manage. Assoc. 56: 82433.

Chen, C.L., Shu, C.M. & Fang, H.Y. 2006b. Environmental

Forensics 7: 19.
Chen, C.L., Fang, H.Y. & Shu, C.M. 2005. J. Air & Waste
Manage. Assoc. 55: 148797.
Dyer, A. 1988. An Introduction to Zeolite Molecular Sieves.
John Wiley & Sons, New York, USA.
Ichiura, H., Nozaki, M., Kitaoka, T. & Tanaka, H. 2003.
Influence of uniformity of zeolite sheets prepared using
a papermaking technique on VOC adsorptivity. Advances
in Environmental Research 7: 9759.
Salden, A. and Eigenberger, G. 2001. Multifunctional adsorber/reactor concept for waste-air purification. Chemical
Engineering Science 56: 160511.
Yang, J. & Lee, C.H. 1998. Adsorption dynamics of a layered bed PSA H2 recovery from coke oven gas. AIChE J.
44: 13251334.

1271

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Using network methodology to define emergency response

team location: The Brazilian refinery case study
E. Calixto & Emilio Lbre La Rovere
UFRJ-COPPE, Rio de Janeiro, Brazil

Joo Eustquio Beraldo

Petrobras, Minas Gerais, Brazil

ABSTRACT: The main objective of this study is to define Emergency Response Team Location in a specific
area based on the risk of plants and facilities. The Center of Gravity and Haikini Network methodologies are the
two different approaches used to define the location in a network based on index values and distance between
locations. These methodologies are different in regard to one basic concept concerning the possibility of defining
critical locations in the network in the first case and their boundaries in the second case. The index in this case
will be the frequency of hazardous events in each facility or plant located in the network.
The two methodologies will be implemented and the results will be assessed. Therefore, a sensitivity analysis
will be carried out looking at specific elements such as alternative routes and population dislocation in the case
of accidents. Furthermore, the real historical data and the usual data used in Brazil in hazardous event will be
assessed to check the influence on final results.
The refinery case study will be carried out to define the Emergency Response Team location in a Brazilian
refinery.

INTRODUCTION

Network methodologies involve different approaches

to solving different logistical problems, such as minimizing product flow, defining the best routes to
different customer locations, etc.
The Center of Gravity methodology has been long
used in physics and mathematics for resolving problems and is now used to provide logistical decisions
such as supplier location.
In risk management an emergency team has to be
prepared to deal with critical and catastrophic events.
In huge areas such as industrial plants, refineries and
urban centers one of the most difficult issues is defining where the emergency team will be based in order
to minimize its dislocation to deal with catastrophic
and critical events. A number of assumptions have to
be taken into account, such as probability of the occurrence of the event and its criticality in terms of harm to
employees, the population and the environment, etc.
Identifying and quantifying the vulnerabilities of such
complex systems is crucial for designing adequate protection, mitigation and emergency actions in the case
of failure (E. Zio et al 2007).
The network methodology using the Center of
Gravity and Hakini theory will be proposed to solve the

emergency team location problem in the refinery taking into account risks, distances and other important
related issues.
2

ANALYSIS METHODOLOGY

For this analysis a number of important steps have to

be followed to achieve consistent results which can
provide support for decision making concerning the
position of the emergency team in the refinery.
According to the Center of Gravity Methodology
the best location depends on the distance from one
point to others and the importance index which is usually represented in terms of cost or profit. Therefore,
the first step is to define the study boundaries as well
as the point locations in specific areas. Moreover, its
representation in terms of importance and the distance
from one point to the others is essential. It is thus very
important to know in advance in which area it will be
possible to locate the emergency team.
After defining the importance index and the distance points the methodology will be applied to define
the emergency team position. In addition, it is also necessary to check if in reality it is feasible to locate the
emergency team at the indicated point. Otherwise, it

1273

feasible for the best location. According to the Hakini

Methodology, however, the best location will be one
of the network points. In both cases sensitivity analyses will have to be carried out for the best location
because in real life the network space has many limitations and assumptions that have to be taken into
account.
These methodologies require a severity map of the
area being studied (B. Mazzarotta et al 2006). Severity maps of the study area can be obtained through
specific GIS (Geographic Information System) applications. In fact, a GIS permits the association of layers
of information with specific locations (points, lines,
polygons) identified on the basis of their geographical coordinates on a map. A number of GIS databases
are available, including information such as land use,
district boundaries, transportation networks, seismic
areas, flooded areas, location of landslides, etc. In
order to obtain risk maps the study area has to be
selected beforehand.

1 Boundaries Limits
2 Critical Points Positions
3 Importance Index
4 Gravity Center Methodology
5 Hakini Methodology
6 Critical Analysis
7 Conclusion
Figure 1.

Network methodology.

will be necessary to look for the most feasible area

near this point.
The second approach is Hakini Methodology, which
is used to check the consistency of the first methodology. Since both methodologies are similar the same
steps will be followed, but in the second methodology
the best position of all boundary points is taken into
account, because it is not feasible to locate the team
within the area.
The final step is to analyze the position in terms
of the feasibility of particular assumptions and limitations of real cases, such as impossible emergencies
routes, traffic in the operational area and the flow of
people and emergency resources. The methodological
steps are shown in Figure 1 below.

3.1

The center of gravity methodology

In a network with many points on the regional plane,

where n points are represented as (Xi , Yi ) and with the
importance index Fi , i = 1, 2, 3, . . . , n. The Center of
Gravity is found by solving the equation below:

MIN Z =

Network Methodologies are used nowadays in many

logistic problems to solve the maximization or minimization of product throughput and transportation
flows. It is a new approach to resolve this type of
problem whose use has been intensified through software simulation evaluations in the last two decades.
There are a group of network methodologies dedicated
to finding the best locations in the network model.
In this particular case study two of these approaches
will be used. The first one is the Center of Gravity and the other one is Hakini Methodology. The
basic concept is to discover the best location in a
specific network in order to maximize flow or minimize time spent to provide service in the network.
The difference between the two methodologies is that
the first one takes any position in the network to be

(Gx Xi )2 + (Gy Yi )2

The Gravity coordinate in Y will be solved by

equation below

Gy =

NETWORK METHODOLOGIES

Fi

i=1

n


n


i=1
n

i=1

Fi Yi
DEi
Fi
DEi

The Gravity coordinate in X will be solved by

equation below
n


Gx =

i=1
n

i=1

Fi Xi
DEi
Fi
DEi

And
DEi =

1274


(Gx Xi )2 + (Gy Yi )2

To stop the interactions an error limit will have to

be assumed, which will be represented as

 k
G G k1 
x
x

 k

G G k1 
y
y

And

This is the number of interactions necessary until

the acceptable error is achieved. The first step will take
into account:
DE 1 = DE 2 = DE 3 = DE n = DE
Therefore, in the first step the equation coordinates
will not take into account DE, so the equations will be
as follows:
n


Gx1

n


Fi Xi
n


Gy1

Fi

Fi Yi

n


Fi

The second interaction will use the results from the

first interaction and will be the same until k interactions achieve the error limits. The general equation
used from the second interaction onwards is shown
below:
n


Gxk

1
n

1

3.2

Fi Xi
DEik1
Fi
DEik1

n


and

Gyk

1
n

1

Fi Yi
DEik1
Fi
DEik1

this condition in order to locate the center point in

the correct position over time it is necessary to know
how importance varies over time, in other words, its
lifetime behavior has to be simulated.
To solve this problem the probability density function (PDF) of the importance index over time has to
be established to discover its frequency. Moreover, it
is also necessary to simulate its behavior taken as a
whole and considering all points at the same time.
Logistic problems usually have to define specific
criteria before responding to a specific network point,
which may be represented in the importance index.
However, some cases, such as emergency situations,
specific point requirements are not met, which can
have catastrophic consequences. Therefore, is advisable in some cases to check response capacity in order
to prepare a capacity that can meet one hundred percent
of demands.
In this particular case, c it is not enough to change
the center point based on its importance index frequency. It is also necessary to overload response
capacity in order to meet demand. In this way the team
which will fulfill the demand must be represented and
simulation will be required to check the behavior of
that team over time.
Direct Simulation will be proposed to check if team
response capacity is sufficient over time.

The Hakini methodology

The Hakini Methodology looks at the network with

a much lower number of points, taking an extreme
point or vertex as the center point. This is a different
approach and in some cases it is more realistic because
it does not consider everywhere in the space in question
to be suitable. The methodology follows three simple
steps, namely:
In the first step the matrix with the shortest path
between all the network points is created.
Next, each matrix element is multiplied by index
points that represent importance in terms of cost,
profit, risk and so on.
In the third and final step, all matrix elements in
same line are added. The best point will be the line
with the lowest value, which is represented on the
vertex in the space in question.
These two methodologies above are able to define
a single reference point in a group of points in space
but they are not able to test whether response can be
maximized over time. The importance index is determinant in most cases, but not in others, therefore the
center point probably will change over time. To check

THE DYNAMIC CENTER OF GRAVITY

METHODOLOGY

One important limitation of the traditional Center of

Gravity Methodology is that the importance index is
constant over time. Consequently, the center point will
be constant over time because both distance between
points and its importance index are constant. In real
life, despite the distance between points remaining
constant, in most of cases the importance index does
not. Even in relation to distance, in some networks new
points are added in the network and some interference
is required for the center point decision. Whenever
this happens, it is advisable to carry out a new analysis in order to determine the new center point. When
doing so the importance index changes, so a new analysis may be proposed, but it is not usual to consider
its behavior over time. Therefore, in some cases the
center point will not fit well with a specific period of
time. To solve this problem it is necessary to establish the probability density function of the importance
index and define a specific period of time to reanalyze
the network. In fact, whenever an importance index
changes a new analysis is required.
The figure below shows an importance index which
represents accident frequency resources in an operational area.

1275

capacity if necessary. In the simulation some results

like reliability, availability, expected number of events
and team capacity will be estimated. The results represent the probability of the events not happening in a
specific time, the percentage of time that the events do
not have an impact on the system, the expected number of events over time and response team capacity
related to requirements. It is actually a good methodology for analyzing center point location over time,
taking into account more than just distance and the
static importance index.

Figure 2.

PDF corrosion.

The above figure represents PDF corrosion for different types of risk resources in a refinery. It means
that corrosion is the basic cause of pipe linkage which
in turn causes catastrophic accidents. Despite different
PDF results, all resources have similar characteristics
for this event. Therefore, the center point in this
network will not change considerably because the
importance index varies similarly over time.
If new events, such as critical and non-critical situations, are taken into account the importance index for
pipe linkage will vary over time, forcing center point
change over time.
Regarding its PDF behavior over time, a new interaction will be required in both network methodologies
in order to analyze the index changes over time in center point decisions. It must be analyzed whenever the
event frequency index will not be constant anymore.

In order to clarify the theoretical method a case study

will be carried out. The case study stimulation is used
to evaluate the methodologies in question and can also
be used in other cases.
In the refinery area one of the most critical problems
is the location of facilities, suppliers, warehouses and
other types of logistics support for industrial plants due
to the many limitations such as space, environment,
risks to safety, and so forth.
The problem to be solved is at which point of space
will the emergency team be located in order to respond
to emergency requirements as fast as possible with
lowest level of vulnerability. This means the best location may be near to an industrial plant but with a
consequent high vulnerability. In order to solve those
problems three models were proposed:
Center of Gravity Methodology;
Hakini Methodology;
Monte Carlo Methodology.
6.1

RESPONSE CAPACITY SIMULATION

Network methodology usually considers that the

response team will meet all requirements on time
and that these requirements will occur one by one.
However, this is not the truth and in some cases
non-response to a requirement can have catastrophic
consequences, such as in the case of accidents. To
solve this problem response team behavior must be
simulated over time. Simulation is a powerful tool
to check emergency procedures which are expensive and difficult to check in reality (David & Isavit
2007). Direct simulation will be carried out to test
response team performance. For this analysis block
diagram methodology is used for PDF events and
response team capacity. The main objective is to discover response team capacity in order to improve their

CASE STUDY

Center of gravity methodology

In the first case the network will be built up based

on the positions of industrial plants in the area of the
refinery, with each point representing a catastrophic
risk resource. The importance index will be the probability of the catastrophic event happening. Therefore,
the model will be located at the center point nearest to
the most critical point, meeting the requirements of all
points regarding distance and the importance index.
The refinery area is show in Figure 3 below.
It is very important to state that the probability decision related to risk resource points was quantitative
and the consequence risk analysis took into account
catastrophic events as fire, explosion and toxic gas
dispersion. The table below represents the fourteen
points which represent the most critical events in the
refinery area.
Using the equations given in section 3 and subsection 3.1 the result of the problem is the center

1276

very difficult for an emergency team located in this

area to respond to catastrophic accidents. The second
is that the emergency route makes people go through
the area from different points, which makes it difficult
for the emergency team to move around freely. Using
the Centre of Gravity methodology the circumferences
around the center point define the nearest area to locate
the emergency team. Therefore, the Hakini methodology will be analyzed in order to define the best point
position and check the second and third best points.
Figure 3.
Table 1.

6.2

Brazilian refinery area.

Center of gravity table.
CG model

Points

U-1
U-3
U-4
U-5
U-2
Off site
Gate 1
Substation

1282,78
1291,35
1203,69
1376,08
1558,06
497,697
0
300

716,15
480,57
500,45
672,97
451,4
939,599
300
800

0,001024
0,001286
0,001548
0,003317
0,002447
0,013255
0
0,013255

Table 2.

Center of gravity table.

Gravity center
GX1

GY1

Erro X

Erro Y

815,7277
GX2
752,4623

729,2447
GY2
585,2872

0
Erro X
63,2654

0
Erro Y
143,9575

point which is found through the second interaction,

as shown in Table 2 below.
The results of this methodology notwithstanding, it
is necessary to analyze the feasibility of the location.
Furthermore, in the case study two points are proposed for the location of the emergency team support,
therefore sensitivity analysis is required.
The first proposed location is near the entrance gate
of the refinery, represented by X(0,0) in the refinery map. The second is located around X(800,800)
due to the available area. Based on the results of the
model, the second one is the better position but there
are two important events that undermine this solution.
The first is that when it rains a lot the area becomes
flooded, so in the case of emergencies it would be

Hakini methodology

As stated in sub-section 3.2, the methodology defines

the best point based on distance and the importance index using a matrix and adds values to final
index which take into account distance between points
(meters) and the importance index (accident frequency). The Hakini results are shown in Table 3
below.
The best point is offsite because it contains the
lowest value in the last right column in the matrix.
The assumptions taken into account in the previous
methodology in regard to flooding and the emergency
route also applies. This assumption, the delay of team
movement by more than 3 minutes, represents 1500
meters of movement, so this value will be added
to substation and offsite to represent the delay. The
new matrix with the new results is shown in Table 4
below. U-1,U-3,U-4 and U-5 are Unit plants around
refinery area represented for point coordinates in map
(figure 3).
In relation to these assumptions the substation and
offsite is the worse location point for the emergency
team. The U-613 location is better, but due to possibility of accident in the area, another point must be
chosen. To locate the emergency team near the plant
is to assume a high vulnerability because in the case
of a catastrophic accident the emergency team will be
unavailable. Therefore the best location is Gate 1.
Nevertheless, the two methodologies provide enough support to decide in which position the emergency
team will be located. What is important is that in this
position the emergency team will be available to deal
with catastrophic events. The Monte Carlo methodology will be used to answer this question and will be
looked at in the next item.
6.3

Monte Carlo methodology

The main objective of this methodology is to compare the impact of emergency localization in terms of
response capacity. Its takes into account the probability of catastrophic events occurring over time and
response duration time.
The first step is create the Safety Block Diagram
in order to represent catastrophic events and simulate

1277

Table 3.

Hakini matrix results.

Hakini matriz 1
Distance

U-1

U-3

U-4

U-5

U-2

Off-site

Gate 1

Substation

Total

U-1
Gate 1
Substation
U-3
U-4
U-5
U-2
Off-site

0
1,6384
1,0752
0,256
0,3072
0,1024
0,6144
0,6144

0,3215
1,929
1,6718
0
0,1286
0,2572
0,3858
1,8004

0,4644
2,1672
1,8576
0,1548
0
0,4644
0,6192
1,7028

0,3317
5,6389
3,6487
0,6634
0,9951
0
1,82435
3,6487

1,4682
4,4046
3,9152
0,7341
0,9788
1,3459
0
2,9364

7,953
13,255
2,651
18,557
14,5805
14,5805
15,906
0

0
0
0
0
0
0
0
0

11,9295
10,604
0
17,2315
15,906
14,5805
21,208
2,651

22,4683
39,6371
14,8195
37,5968
32,8962
31,33085
40,55775
13,3537

Table 4.

New Hakini matrix results.

Hakini matriz 2
Distance

U-1

U-3

U-4

U-5

U-2

Off-site

Gate 1

Substation

Total

U-1
Gate 1
Substation
U-3
U-4
U-5
U-2
Off-site

0
1,6384
2,6112
0,256
0,3072
0,1024
0,6144
2,1504

0,3215
1,929
3,6008
0
0,1286
0,2572
0,3858
3,7294

0,4644
2,1672
4,1796
0,1548
0
0,4644
0,6192
4,0248

0,3317
5,6389
7,2974
0,6634
0,9951
0
1,82435
8,6242

1,4682
4,4046
7,5857
0,7341
0,9788
1,3459
0
6,6069

7,953
13,255
22,5335
18,557
14,5805
14,5805
15,906
0

0
0
0
0
0
0
0
0

11,9295
10,604
0
17,2315
15,906
14,5805
21,208
22,5335

22,4683
39,6371
47,8082
37,5968
32,8962
31,33085
40,55775
47,6692

Table 5.

Emergency team availability.

Detailed crew information

Figure 4.

Safety block diagram.

them over time. Figure 4 below shows the Safety Block

Diagram.
For the location of the emergency team at both proposed sites, namely the entrance gate X(0,0) and the
substation X(800,800) there will be 100% availability
and 75.6% reliability in twenty years. This means that
100% of time there will be no catastrophic accidents
and there is a 75.6 % chance that none will happen
over the next 20 years. The emergency team will have
100% of availability because they will not be required,
as shown in Table 5 below.
The table above shows the emergency team availability (100%), the number of expected calls received
and accepted and their average duration (one hour).
The scenarios can take place and require emergency
support in many different events happening at almost
the same time, in which case because the emergency
team is occupied responding to one event it will not be

Crew Policy: Portaria 1

Call Summary
Calls Received:
Accepted Calls:
Percent Accepted:
Percent Rejected:
Utilization
Total Utilization:
Average Call Duration:
Total Wait Time:

0,288
0,288
100,00%
0,00%
0,2953
1,0254
0

available to respond to others. These scenarios must

happen if critical and non-critical accidents are considered in the safety block diagram. In this case, in order
to deal with the most critical events, a prior assumption
would be proposed.
7

CRITICAL ANALYSIS

The case study involves three types of different

approaches used to define emergency team location

1278

and to discover whether this position has an influence

on emergency team availability. Although catastrophic
events occur rarely in reality, it has to be borne in mind
that the event frequency was considered to be constant,
which does not actually occur in real life.
For this reason, PDF events related to different risk
resources were taken into account. As a result the
dynamic network methodology was applied in order
to verify center point modification over time, after
which it was concluded that even PDF events did
not have much influence on emergency team location, because most events occur around same time.
Despite this, PDF events provide significant information allowing preventive action, such as pipeline maintenance to prevent corrosion which causes catastrophic
events.
The other important point is that when a particular
industrial point is taken into account it is necessary
to consider all catastrophic risk resources in order to
verify the results.
PDF events will be analyzed using Monte Carlo
Methodology in order to check their influence on the
final results.
8

SENSITIVITY ANALYSIS

In order to discover the influence on the final result

a sensitivity analysis was carried out on the three
methodologies in relation to PDF events. In the case
of network methodologies, there were no modifications but there were in relation to the Monte Carlos
methodology.
For the Monte Carlo methodology, in both cases
of emergency team location, the Entrance Gate and
the Substation, safety availability and safety reliability
was 99.98% and 0% in twenty years. This means that
99.98% of the time there will be no catastrophic events
but there is a 100% chance that they will occur during
this period. Furthermore, looking at frequency index
events in Figure 5 and comparing static and non-static
frequency events, it is clear how important it is to take
into account historical data. In the static frequency it
is clear that the event is rare over fifty years but in the
non-static frequency from fifteen years onwards the
frequency starts to increase. This means that catastrophic events have a greater chance of occurring over
time.
Despite the increase in the intensity of catastrophic
events, as a result of which it is expected that there will
be around thirty events in twenty years, the emergency
team will have 100% acceptance of emergency call
requirements in both location, namely the Substation
and the Entrance Gate.
Even though they do not have an important effect
on the resolution of problems, it is nonetheless very
important to take historical data into account to

Figure 5.

Safety block diagram.

establish the PDF in order to carry out preventive

action whenever this is possible.

CONCLUSION

The network methodology applied to the case study

shows how important it is to consider other assumptions, such as PDF events and response capacity over
time. In this specific case study they did not have
much influence on the final result but it is nonetheless reasonable to take into account their occurrence
depending on the different events considered.
The final decision was to locate the emergency team
at the entrance gate because the events are rare over
time and in both cases the response team capacity is
100%. Although the substation region is the best one
according to the network methodology solution, the
possibility of the area being flooded is critical and
if this happens when a emergency team is required
the consequences of not having the emergency team
available are terrible.
The refinery case study must to be evaluated in relation to critical and non critical events and new plants
to check the influences of these on the final results.
That methodology may be applied in offshore case
study regarding one specific area with some platforms
and drill in order to state the emergency team in the
better position in order to attend catastrophic accidents. In Some cases, the offshore area change its
configuration in terms of facilities having ships frequently surrounded or gathering those platform Unit.
Therefore, the dynamic approach is required.

REFERENCES
A.M. Cassula, Evaluation of Distribution System Reliability Considering Generation and Transmission Impacts,
Masters Dissertation, UNIFEI, Nov. 1998.
Boaventura, Paulo Oswaldo Netto, Grafos: Teoria, Modelos,
Algortimos. So Paulo, E. Blucher, 2003.

1279

B. Mazzarotta & B. Silvetti, A severity index to assess

technological and natural risks in a study area, Dipartimento di Ingegneria Chimica, Universit di Roma La
Sapienza, Rome, Italy. ESREL 2006.
E. Zio & G. Sansavini, Service reliability analysis of a
tramway network. Department of Nuclear Engineering,
Polytechnic of Milan, Milan, Italy, ESREL 2007.
IEEE Recommended Practice for the Design of Reliable
Industrial and Commercial Power SystemsIEEE Std.
493-1997.
Kececioglu, Dimitri, and Sun, Feng-Bin, Environmental
Stress ScreeningIts Quantification, Optimization and
Management, Prentice Hall PTR, New Jersey, 1995.
Lafraia, Joo R. Barusso, Manual de Confiabilidade,
Mantenabilidade e Disponibilidade, Qualimark, Rio de
Janeiro, Petrobras, 2001.

Moraes, Giovanni de Araujo. Elementos do Sistema de

Gesto de segurana meio ambiente e sade ocupacional.
Gerenciamento Verde Consultoria Rio de Janeiro: 2004
R. Billinton, and R.N. Allan, Reliability Evaluation of Engineering Systems: Concepts and Techniques, 1st Edition,
Plenum Press, New York, 1983.
ReliaSoft Corporation, Weibull++ 6.0 Software Package,
Tucson, AZ, www.Weibull.com.
V. Idasiak and David, Accident simulation: Design and
results, Laboratory Vision et Robotique, ENSI de
Bourges, ESREL 2007.
W.F. Schmitt, Distribution System Reliability: Chronological and Analytical Methodologies, Masters Dissertation,
UNIFEI, Oct./2002.

1280

Risk perception and communication

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

(Mis-)conceptions of safety principles

J.-T. Gayen
TU Braunschweig, Germany

H. Schbe
TV Rheinland InterTraffic, Cologne, Germany

ABSTRACT: Imprecise definitions and use of safety terms and principles result in misunderstanding and
misconception of safety principles. In this article we present a systematic overview and explanations of safety
principles as fail-safe, safe-life, fail silent, redundancy etc. For these safety principles we give a systematic
approach and show their advantages, shortcomings and pitfalls. All principles are presented in detail, compared
and their premises, use and possible problems are discussed. We present an overview of possible problems that
can occur when applying these safety principles erroneously. In addition, some examples for the application of
the safety principles are given.
1

MOTIVATION

Frequently terms describing safety principles are used,

as e.g. fail-safe. There exists a large variety of such
terms. It is not always clear, what is meant with a
safety principle, which limitations exist, and which
assumptions are used. In this connection, there are
many misunderstandings and misconceptions. They
are caused by different factors. Among them are:

These examples clearly show the open questions.

In our paper we will start with terminology in
section 2. In the third section we describe the known
safety principles. The fourth section is dedicated
to possibilities and limitations of safety principles.
Section five presents an example. In the last section,
we draw conclusions.

2
unclear terminology
unclear assumptions
unclear descriptions

2.1

In addition, there are differences between the versions in different languages of the same standard. This
situation has motivated the authors to consider the
following questions:
Which safety principles do exist at all?
How are they defined?
When can they be applied and under which conditions and assumptions?
Safety principles are widely applied in safety techniques. The most known is the fail-safe principle in
railway technology.
A safety related railway system must enter a safe
state upon occurrence of a failure. This implies the
following questions:
What is a failure?
Has the state the system was in before the failure
occurred been an unsafe state?
How is a safe state characterised?

BASIC TERMINOLOGY
Process, system boundaries and environment

In order to develop a system, especially for the

safety related sub-systems, the process that has to
be safely controlled has to be completely understood and described. System boundaries must be
defined unambiguously and the system environment (physical, chemical etc.) must be completely
described.
2.2

Safe state

Any safety related system must have at least one

safe state. A state is called safe, if the risk, which
arises from the system being in this specific state,
is smaller than the acceptable risk. For railway systems the railway authority must define the acceptable
risk.
Every transportation system has at least two safe
states:
Normal operational state,
Stop (system switched off).

1283

Specified requirements

Correct
development and
manufacturing

Incorrect development
and/ or
manufacturing mistake

System is correct

Safe operating state

Deficiency
System has a
physical failure

System has
wrong
information

System has
inherent faults

fault

failure
System is
defective

System fails

Figure 1.

Safe failure
state

malfunction

Causes for system malfunction.

Figure 2.

It is assumed that the system is free of (dangerous)

failures in the normal operating state. The stop state is
usually a state in which the system has less energy than
in the normal operating state. The safe stop state must
be reached in sufficiently short time by deactivation
of the system functions. The deactivation can be an
active process (additional function of the system).
There can exist situations, in which the previously
existing (safe) state cannot be reached immediately,
e.g. when a train is on fire in a tunnel. In this case, the
train is preferably stopped outside the tunnel.
2.3

Safety terminology

Figure 1 shows the main safety terms, inspired by DIN

40041. Equivalent standards exist in other countries
and languages.
The terms are defined as follows:
Failure: a physical functional unit seizes to function
within specified load and environmental conditions
Remark: A failure leads to malfunction if the
desired system function is demanded.
Mistake: A requirement (for planning, realisation,
design or manufacturing) has not been fulfilled.
Fault: Missing, erroneous or incomplete provision
of the desired function by the unit.
Malfunction: A fault emerges during specified use
of the unit due to an interior cause.
Remark: Malfunction of a system can be aroused
by different causes. Malfunctions occur if a failure of
the unit arises during specified use.

failure

Demand to system
coincides with the defect

SAFETY PRINCIPLES

Safety principles are used in order to assure system

safety in cases of failures, mistakes, deficiencies and
other deviations.

Strict view of the fail-safe principle.

To facilitate this, efficacy and scope of safety principles must be clear. Safety principles can be used to
control e.g.
a.
b.
c.
d.

random failures,
systematic failures,
hardware failures,
software failures.

Control of failures means that after a failure has

occurred or a fault has been activated, the reaction of
the system ensures that failure consequences are only
on the safe side. E.g. this can be a safe stop of the
system. Alternatively, safe function of the system can
be ensured further on for a limited time interval.
In the sequel, we will describe the safety principles
one by one.
3.1 The fail-safe principle
According to VDI/VDE 3542, the fail-safe principle is
dedicated to control of failures. It describes the ability
of a system to remain in a safe state or to go over into
a safe state, if a failure occurs. The following figure
illustrates this
According to EN 50129, a concept is called fail
safe (in German terminology funktionstechnisch
sicher), which is incorporated into the design of a
product such that, in the event of a failure, it enters or
remains in a safe state.
At another place (B3.1), the standard EN 50129
says:
It is necessary to ensure that the system/subsystem/equipment meets its tolerable hazard rate
(THR) in the event of single random fault. It is necessary to ensure that SIL 3 and SIL 4 systems remain
safe in the event of any kind of single random hardware
failure, which is recognized as possible.

1284

The principle can be implemented in three different

ways:

Feeding pipe

inherent fail-safety
reactive fail-safety
composite fail-safety

max.level

Note that, composition means that two channels are

present in the system that are used for crosschecks.
This has nothing to do with redundancy. The German
translation (Fail safety with Redundanz) is ambiguous. This example shows that careful definition and
description are crucial for correct understanding of
standards.
3.1.1 Inherent fail-safety
A system is designed according to the inherent failsafe principle, if all possible failure modes do not lead
to dangerous consequences. Physical, chemical and
other processes that are inherent to the system ensure
this. The system is always kept in safe states, even if a
failure occurs. If the assumptions are fulfilled, a safe
system can be constructed. The system would consist
of a single sub-system, e.g. a train detection system
using shunting of track circuits. If such a system is constructed, all dangerous failure modes must be known,
analyzed and for each such failure mode there must be
a proven physical, chemical or other inherent process
that brings the system into a safe stopping state. Since
inherent fail-safety is based on properties that cannot
vanish, such systems cannot fulfill complicated functions. This is caused by the fact that the safety functions
are simple and no complicated decisions can be made
by them.
Therefore:
No inherent fail-safety can exist without remaining
in or transition to (another) safe states.
No fail-safety without existence of a safe stopping
state.
Examples for this kind of systems are gravity dropout relays. With this kind of relays with carbon-silver
contact pairs the gravity prevents the relays from sticking in the closed position. If the coil is de-energized,
the closed relay contacts open, caused by the weight
of the contact bar.
For train detection with shunting devices in track
circuits, an alternating current is fed into the track circuit. If there is a train in the track circuit, both rails
are shunted so that the voltage cannot be detected at
the other end of the track circuit. All possible failures
(short circuit, open contacts) lead to a loss of the alternating current. This would imply a faulty detection of
a not existing train, which is a safe event. It must be
made sure, however, that no voltage can be induced by
EMC effects from other sources. This can be realized
by choosing an appropriate frequency for the alternating current. In addition, there exist requirements for

Over flow
duct

Outlet

servovalve

Figure 3.

Fluid level control.

railway equipment, that prevent interference of this

equipment with track circuits.
3.1.2 Reactive fail-safety
Reactive fail-safety is characterized by fast detection
of any failing function and an active reaction to bring
the system in to a safe state. Although there is usually
only one sub-system for the control function, there
might be a second sub-system present for detection,
diagnosis and testing. An example of this type of
systems is a system with a safe supervising function.
It becomes clear that this kind of systems need a
detection mechanism that is able to detect with sufficient coverage all dangerous failures. Moreover, a
reaction mechanism is needed that brings the system
into a safe stopping state by an active reaction.
We will now present a simple example in order to
illustrate the inherent fail safe principle as well as the
reactive fail-safe principle. Imagine an open reservoir
for a fluid. The reservoir has a feeding pipe on the
top and an outlet with a servovalve at the bottom.
To prevent the reservoir from being overfilled, there is
an overflow duct just below the upper edge. This is
an inherent fail-safe protection against overflow. It
is based on the physical principle of communicating
ducts. If the fluid level of the reservoir becomes to
high, the fluid is vented through the overflow duct.
Thus, the fluid cannot leave the reservoir over the
upper edge.
In addition, the system has a sensor that measures
the fluid level. If this reaches a critical value (near the
maximal level), the servovalve at the outlet is opened
to decrease the fluid level. This second protection
function is an example of a reactive fail-safe system.
3.1.3 Composite fail-safety
Systems designed according to the composite failsafety priciple must fulfill the following conditions:
Several (at least two) sub-systems must be present
which provide the same safety function
These sub-systems must be independent of each
other.

1285

Table 1.

3.2

Detection and toleration of failures and faults.

Safety principles

Failures

Systematic
faults

Inherent fail-safety
Reactive fail-safety
Composite fail-safety

X
X
X

X1)

The results of all these sub-systems must be identical (composite fail-safety with comparison) or a
majority of the sub-systems must come to the same
result (majority vote)
Detection of failure of sub-system and subsequent
safety reaction must be fast enough that probability
of a failure of a second sub-system is negligible.
A safe stopping state must exist, as with the other
two fail-safe principles
Note that, also the comparison or majority vote must
be carried out fail-safe and fast enough.
Composite fail-safe systems can also be adaptive.
E.g. a 2-out of-3 system with majority vote can be
transformed into a 2-out-of-2 system with comparison, after malfunction of one sub-system has been
detected and the sub-system has been disabled. Only
if a second sub-system fails too, the system has to enter
the safe stopping state. This is necessary, since with
two remaining sub-systems, a difference of the subsystems can be detected, but no decision can be made,
which of the sub-systems has failed.
Composite fail-safe systems with majority vote
need an uneven number of sub-systems, to prevent
deadlock situations (k-out-of-n with k (n + 1)/2).
With an even number of systems, the deadlock with
n/2 against n/2 votes would leave stopping the systems
as the only solution.
3.1.4 Detection of failures and faults
The following table gives a comparison how different principles contribute to detection and toleration of
failures and faults. It is evident that only composite
fail-safety is able to detect and tolerate failures. This
is caused by the fact that these systems are built of several sub-systems. Now it becomes clear, why for higher
safety integrity levels (SIL 3 and SIL 4) a system is
advised to be constructed from several subsystems,
since failure detection is very important.
Note that, for composite fail-safe systems, fault
detection does not come automatically. Fault detection is implemented only in the same degree as the
sub-systems are diverse. That means, that composite
systems with diverse sub-systems detect only specific
faults due to the kind of diversity.

The safe life principle

A system is designed according to the safe life principle if the safety relevant function is available with
high probability. Usually this is achieved using redundancy of the relevant sub-systems so that the function
is ensured even in case of failures. In addition, the
system has a failure detection function in order to prevent accumulation of failures. So, the mission can be
finalized in case of occurrence of a failure. There exist
also systems, which can be repaired during their mission without switching them off. This holds, e.g. for
many chemical plant control systems. Here, the failed
sub-system is repaired and restored to a functioning
state, while the other intact sub-systems carry out the
safety-relevant control function.
The safe-life principle is applied for systems that
must not fail during a certain mission time. Engines of
airplanes serve as an example. The safe life principle
is also applied to magnetic levitation trains. The principle of safe levitation and guidance is applied here.
This principle considers magnetic levitation and guidance as safe-life functions. That means, that the rate of
dangerous failures of the levitation and the guidance
function must not exceed a rate of 106 per year.
There exist several possibilities for the implementation of the principle:
Dimensioning (no damage)
Such an approach is usually applied for buildings,
in particular to bridges. These systems have only
one safe state (correct function) and there is no
transition into another safe (stopping) state. Therefore, their function must be warranted with a very
high probability even under extreme environmental
conditions.
Redundancy
Several components or sub-systems carry out the
same function, however, not all of them are necessary. An example is an airplane having more
than one engine. Magnetic levitation trains have
three operation control systems, four power supply
networks etc.

3.3 Fail-safe versus Safe-life

When shall we apply the fail-safe principle and when
shall we apply the safe-life principle?
The choice of the safety principle depends on the
process. First, it has to investigated, whether there
exists a safe stopping state, which can always be
reached.
The table shows that the fail-safe principle cannot
be applied in some cases, since the safe stopping state
does not exist. Then, the safe-life principle has to be
applied, since it does not require the safe stopping
state. In all cases, where the fail-safe principle can be

1286

Table 2.

Safety principles and their use in different systems.

System

Is the
safe stopping
state always
reachable?

Applicable
safety
principle

Railway
Airplane
Nuclear power station

yes
no
yes

Fail-safe
Safe-life
Fail-safe

Note

that, the emergency cooling system for nuclear power

stations has to be safe-life.

applied, of course the safe-life principle is also applicable. Such a choice does not necessarily bring a gain
in safety, but usually improves availability. Using one
safety principle in place of another must not corrupt
the safety level.
For the example of the railway and other guided
transport systems we refer to cases, where the safe
stopping state does not always exist. Then, specific
questions arise whether the propulsion must be always
available [5].
3.4

Redundancy

Redundancy is the presence of more means than necessary for the desired function (VDI/VDE 3542). For
safety related functions it has to be distinguished,
which means are used to ensure the desired level of
safety and which of them are present additionally. Only
the last can be declared as redundant. For example,
a composite fail-safe system with two sub-systems
and comparison is sufficient for safe applications.
However, in order to increase the availability, a third
sub-system can be added and the system can be redesigned as a 2-out-of-3 system with majority vote.
Hence, the third sub-system does not improve the
safety level at all, but it increases system availability.
Redundancy is introduced to improve reliability
and availability, but safety is not directly improved.
When considering an entire system, it might come out
that improvement of reliability and availability also
might improve safety. The following railway example
demonstrates this.
After the railway system has been brought into the
safe stopping state, railway operations need to be continued in a fall-back mode. Operation in the fall-back
mode is usually not on the same safety level as normal
operation. This is caused by the fact, that the systems
used in fall-back mode are much simpler and responsibility for safety relies much more on personnel. This
is acceptable, if the fall-back solution is not frequently
used. If now the availability of the safety system,
which is used for normal operations is improved,
the fall-back solution is used less frequently and,
therefore, overall system safety is improved.

Table 3. Comparison of composite fail-safety and

redundancy.
Composite fail-safety

Redundancy

The safety function is realised

with several sub-systems that
have the same function
Safety is ensured through the
function of all sub-systems.
Failure (and fault) detection
are based on the existence of
several (diverse) sub-systems
Safe comparison
Existence of more than one
sub-systems improve safety

Several sub-systems
realise the same
function.
There are more subsystems present than
necessary to carry
out the function
No comparison
Existence of more
than one sub-systems
improve availability
and reliability

The following types of redundancies need to be

distingished.
Hot standby:
A redundant (or several) sub-system is working in
parallel with the active sub-system. The redundant
sub-system is able to take over the functions of
the active system in very short time. Usually, the
redundant sub-system reads the same input data as
the active system and also performs the same processing operations, as the active one. However, the
redundant system doe not provide outputs to any
actors or other control elements. In such a situation, the hot standby-system is subject to aging to
the same extent, as the active one.
Warm standby:
The redundant system is partially active and can
take over the functions of the active system in short
time. Mostly, the controller of the standby system
is active, however attached auxiliary systems as
pumps etc. must be activated, before the function
can be taken over.
Cold standby:
The redundant sub-system is completely deactivated. In case of failure of the active sub-system,
the redundant one first needs to be activated. It
then needs a start-up phase, in which the control
unit is waked up, reads the necessary data, pumps
etc. are switched on and other preparing actions
are initiated. In several case, manual intervention
is necessary.
Composite fail-safety using several sub-systems
and redundancy must be strictly distinguished,
although redundancy is sometimes used in combination with fail-safe systems. The following table
shows the differences. Using several sub-systems for
composite fail-safety is a measure to improve safety,
redundancy to improve availability.

1287

Note that, in cases where the safety of a system is

achieved via availability of the safety system (safe-life
systems), this availability is ensured using redundancy. However, no majority vote or comparison is
carried out.
Frequently, a combination of composite fail-safety
and redundancy is used. In such systems, k subsystems are necessary for safe function and n-k
sub-systems are redundant ones. These systems are
characterized by the following properties:
The safety function is carried out by several subsystems with the same task.
Safety is ensured only if the majority of these subsystems functions.
The number of such subsystems is larger than
necessary for the safety function.
There exists a majority vote.
The system has a higher safety as well as a higher
availability than a single sub-system.
3.5

Diversity

Diversity characterises a situation, where several subsystems carry out the same function in a different
manner. Diversity can be in relation to the development
process, the algorithm, software, hardware etc. The
diversity principle can be applied for composite failsafe systems and redundant systems. In most cases,
comparisons and majority votes can be present only
on the functional level, caused by differences between
the sub-systems.
The following kinds of diversity are distinguished:

Diverse development teams

Design diversity
Hardware diversity
Software diversity
Diverse operating systems
Diverse programming languages
Diverse algorithms
...

The main intention of the diversity principle is to

exclude common-mode-failures as far as possible.
Before the diversity principle is applied, it must be
made sure, with respect to which kind of faults the
diversity principle shall be effective. Faults that do not
arise from the requirements specification, but arise
later in the development and manufacturing process,
can be controlled by choosing the right methods of
diversity.
E.g. faults in control units can be coped with by
using different processor types. A hardware fault in
a single processor can be revealed by using diverse
software, carrying out the same or equivalent tasks.

3.6

Fail silent

The term fail-silent describes a situation, in which

a system with a supporting function fails, perhaps
without detection of the failure. The principle can
only be applied to systems with supporting functions, e.g. driver assistance systems. Failure of an
assistance system is not dangerous as far as complete loss of function is considered. A malfunction
of the assistance function can, however, be dangerous, if the assistance function interferes with another
safety relevant function as e.g. steering. Therefore,
dangerous failures must be detected and the system
must be safely switched off. Switching off the system
is analogous as for a reactive fail-safe system, however, the system self can participate in supervision and
switching-off. Sensors and actors must be sufficiently
reliable.
As an example an active front steering system is
considered. Such systems interfere actively in the
steering. When a car is parked, the number of necessary revelations of the steering wheel is reduced.
For high velocity, this function is not provided. So,
the system provides a velocity dependent gear ratio
for the steering, by imputing additional revelations
of the steering columns with the help of an electromotor. Obviously, the steering column has to be
split into two parts for such a system. The system
must not introduce unintentional steering movements
that destabilize the vehicle. Therefore, dangerous failures must be revealed, the system switched off and
the splitted steering column must be reconnected.
Detecting dangerous failures and reacting must be
carried out in a very short time. After the system
has been switched of, the driver still has the normal
steering.
There exists a problem with fail-silent driver assistance systems. The problem has to do with properties
of human beings, getting used to situations, equipment etc. In case of loss of the assistance system, the
driver might not be trained or used to such a situation. An active brake assistance system that measures
the distance to a vehicle running ahead and braking,
might serve as an example. A driver might get accustomed to such a system so that he does not notice a
failure and will drive with very short distance. The
driving behaviour of some upper class car driver seems
to support such a hypothesis.
The fail-silent principle is frequently used in automobile technology, but rarely in railway technology.

POSSIBILITIES AND LIMITATIONS

OF SAFETY PRINCIPLES

The consideration of the different safety principles has

discovered the following.

1288

4.2

Failure
dangerous

Actuator works
in time

Figure 4.

4.1

Especially for the inherent fail-safe principle identification of all failure modes is important. The failures
have to be divided into safe and dangerous ones,
depending on the function of the component in the
system context. It must be proven, that dangerous failures are controlled. As a prerequisite, a safe stopping
state must exist.

safe

und etected

detected

detected

Safe side failures

undetected

Actuator fails

4.3

Actuators

The reliability of actuators is important. This holds

also for undesired activation or reaction of the actuators, since such a reaction can cause dangerous
situations.

Failure detection and reaction.

Failure detection

Identification of failures is an important feature: For

systems with reactive fail-safety, fail-silent systems
and redundant systems, failures must be detected to
activate the reaction that switches the system off,
switches sub-systems on or activates redundant subsystems. The degree of failure detection (coverage)
strongly influences the safety integrity level that can
be achieved with a given system. That is one of the
reasons that IEC 61508 requires a minimal failure
coverage for each safety integrity level.
In order to determine the failure coverage, it must be
analysed which failures can occur and by which means
and to which degree they can be detected. In addition,
failures of the actor must be taken into account. Consequently, the division of failures as shown in figure 4
has to be carried out. A failure occurs with a certain
probability. This probability has to be split into a part
for dangerous failures and those, which are not dangerous, according to the different failure modes. This has
to be substantiated by an analysis, e.g. a failure mode
and effects analysis. Dangerous failures are shaded in
grey in figure 4. These are either undetected dangerous failures or detected dangerous failures with a failed
actuator. Undetected safe failures have to be considered in addition. They are not dangerous as such, but
if they accumulate, necessary system functions can be
inhibited or degraded. Several undetected safe failures
together might reduce the failure detection ability of a
systems so that a larger fraction of dangerous failures
occurs.
Dangerous failures need to be excluded either by
design or the probability that they lead to an accident
must be sufficiently small. For the last alternative there
exist a tolerable hazard rate (THR) as a target value
for dangerous failures. Based on the THR, a safety
integrity level is determined for the relevant system
safety function.
When talking about detection of failures, also the
detection and reaction times have to be compared with
the values necessary for a safe reaction of the system.

4.4

Comparison and majority vote

When using several sub-systems with comparison or

majority vote, a safe voter must be used. The voter
must be subject of the safety analysis and satisfy the
same safety requirements as the system. The separate
sub-systems do not have to fulfill the same set of safety
requirements, as the entire system. However, the probability of simultaneous failures of several sub-systems
must be sufficiently small. Sometimes, safe comparators are composite systems themselves or built up as
combinations of safety relays. Comparison or majority vote can also be implemented in form of a software
solution and can be carried out in each of the subsystems. Faulty sub-systems can disable themselves
or be disabled by a majority of the other sub-systems.
4.5

Sub-systems and common mode failures

When using several (identical) sub-systems within a

system, common mode failures need to be considered.
They can ruin all advantages of composite systems.
Common mode failures are systematic failures, where
the deviations have been caused by a common cause
(see DIN/VDE 0801). This means, that one cause
induces several failures. For instance, a wrong algorithm can cause all sub-systems to fail with the same
failure mode.
4.6

Common cause failures

Common cause failures have to be distinguished from

common mode failures. This is not always done, sometimes both terms are used synonymously. A common
cause failure is a common failure in several subsystems, which should be independent (EN 50129).
EN 61511-1 defines a common cause failure that is
caused by one or more events that lead to failures
in two or more separated sub-systems in a compound
system, resulting in a system failure. A common cause
can be a short circuit that leads to failure of the power
supply of all sub-systems. A reason can be that the

1289

Table 4.

Sources for CCF and CMF.

Software (operating system, libraries,

compiler, linker)
Algorithm
Hardware (Processors, layouts, ..)
Power supply
Exterior influence (EMC, temperature,
pressure, radiation, ..)
Prozesses (Concept, design, manufacturing)

CCF

CMF

X()

X
X
X

X()
X()
X()

() It

is supposed that the same software etc. is used in all

sub-systems.

power supplies of the sub-systems are insufficiently

separated. A common cause failure is a systematic
failure, which however can be triggered by a random
failure serving as an initiator.
The following table shows the main differences
between common cause (CCF) and common mode
failures (CMF).
Table 4 shows the difference between common
mode and common cause failures. Processes and exterior influences cannot be a common mode. Algorithms
can cause common mode failures (if the same algorithm is used), but not common cause failures. Other
cause can lead to common mode failures as well as to
common cause failures.
4.7

Diversity

When diversity is used, the degree of diversity is

important. It must be analyzed, which common mode
failures and which common cause failures are prevented. It is recommended to carry out a common
mode failure analysis according to ARP 4761.
4.8

The fail-safe principle and risk considerations

In the past, many systems have been developed according to the fail-safe principle. The applied method is
a deterministic one, i.e. in fail-safe systems dangerous failures are excluded and thus do not occurat
least according to the applied philosophy. For systems
with reactive fail-safety or composite fail-safety probabilistic arguments are used. That means, there exists
a probability of unsafe system failures. When moving
forward from inherent fail safe systems to more complex safety principles, deterministic considerations
must be replaced by probabilistic ones.
5

systematic application of safety principles. These

guidelines have been applied to a particular project.
In this example, the guiding function will be discussed. In most cases guided transport systems, as
e.g. a railway, usually have two safe states, one is
standstill of the train. However, when discussing the
guidance function, it cannot be concluded that there
is a second safe state if the guidance function fails.
Therefore, the guidance function cannot be designed
as a fail-safe function but must be designed as a safe
life function, i.e. a function that is maintained. For a
classical train this is done by the rail-wheel pair, which
fails with a sufficient low probability. If this is not
the case, derailment protection measures are applied
(redundancy).
If one discusses guided transport that is kept on the
track not by mechanical systems but by immaterial
means, the guidance system has to be a safe-life system. Such a system is e.g. the Phileas, described by
Bouwman (2007). During the collaboration between
APTS, the company developping and manufacturing
the bus and TV Rheinland InterTraffic, the safety
principles have been applied.
First, the guidance system has been defined as a
safe-life system. Then, a k-out-of-n:system has been
used in order to ensure the guidance function technically on the axles. The Phileas is kept on the track
by different electronic sub-systems. In order to avoid
common-cause failures, the principle of diversity has
been applied rigorously, i.e. using different hardware,
software etc.

CONCLUSION

The analysis of safety principles has shown that a

clear distinction between safety principles and a clear
use of terms is necessary. Determination of safe system states, including the existence of a safe stopping
state is a basic pre-condition for correct application
of safety principles. System functions must be analyzed, to identify the appropriate safety principles.
Identification of all possible failure modes is necessary to define the scope of safety measures. Then, an
appropriate safety principle can be chosen.
Failures, mistakes and faults must be controlled to
arrive at a safe system. This can be ensured by the different safety principles. For a proper choice of a safety
principle a detailed analysis of system properties and
control processes is necessary.

REFERENCES

APPLICATION EXAMPLE

In this section, design principles for guided transport are derived in order to demonstrate the use of

ARP 4761 Aerospace Recommended Practice, Guidelines

and Methods for Conducting the Safety Assessment Process on Civil Airborne Systems and Equipment, 199612.

1290

Bouwman, R. Phileas, Tram on Virtual Rail, Increasing Standards, Shrinking BudgetsA Bridgeable Gap?,
2nd International Rail Symposium, TV Rheinland,
67.11.2007, Cologne, Proceedings, 22 p.
DIN 40041, Zuverlssigkeit; Begriffe, (Reliability, terms)
199012.
EN 50129, Railway applicationsCommunication, signalling and processing systemsSafety related electronic
systems for signalling, february 2003.
EN 61511-1, Functional safetySafety instrumented systems for the process industry sectorPart 1: Framework,

definitions, system, hardware and software requirements

200510.
Glker, J., Schbe, H., Is the propulsion vital?, 11th International Conference on Automated People Movers, Vienna
2225 April 2007, Conference Proceedings, 10 pages.
IEC 61508 Functional safety of electrical/electronic/programmable electronic safety-related systems, 199812.
VDI/VDE 3542, Blatt 4, Sicherheitstechnische Begriffe fr
Automatisierungssysteme, Zuverlssigkeit und Sicherheit
komplexer Systeme, 200010.

1291

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Climate change in the British press: The role of the visual

N.W. Smith & H. Joffe
Department of Psychology, University College London, UK

ABSTRACT: The present investigation identifies the key images that British newspapers use to represent
climate change risks. In doing so, it widens the scope of the burgeoning literature analysing textual content
of climate change media information. This is particularly important given the visual informations ability to
arouse emotion and the risk literatures increasing focus on the importance of affect in shaping risk perception.
From a thematic analysis of newspaper images, three broad themes emerged: the impact of climate change,
personification of climate change and representation of climate change in graphs. In particular, the depiction
of climate change as an issue affecting domestic populations rather than just other areas of the world brings the
threat closer to home. Challenging the perception that climate change is still a long term and future orientated
threat, visual images concretise the risk by providing viewers with tangible examples of climate changes impact.

INTRODUCTION

Global climate change is currently being constructed

as one of the biggest threats facing the world. Although
gradual warming and cooling of the Earths climate
is a natural phenomenon, unprecedented temperature increases in recent years are causing widespread
damage. The common consensus amongst climate
change scientists states that continued production of
greenhouse gas emissions at present levels is likely
to have irreversible consequences (IPCC, 2007). For
individuals who are not engaged in climate science
research, access to climate change information is typically through media reportage and communication
with people in their environments. As mass media
communication is hypothesised to play a key role in
the public understanding of risk (Wahlberg & Sjoberg,
2000), detailed analysis of this coverage is important.
The nature of this relationship has been the focus of
considerable debate in recent years. Although a direct
effects model of information uptake has increasingly
been rejected by researchers since the 1950s owing to
individuals bringing an already known component
to their understanding of media information (Corner,
1991), there has been little resolution in the field in part
due to inherent complexities of discerning media influence. Audiences of media information are active processors rather than passive receptacles who socially
construct their understanding of risk information
(Boholm, 1998). Furthermore, the messages involved
are multi-dimensional and many different channels of
media disseminate risks. Those with a more visual
bent may convey a set of messages that are absent
from those with a more textual dominance, though of

course, textual and visual portrayals are often combined. The visual, in particular, has the ability to
arouse emotion making it an effective medium for the
social construction of risk messages (Joffe, 2007).
Perlmutter (1998) argues that images, amongst
other factors, arouse and stimulate affective responses
from the viewer. In his typology outlining effects of
visual images, manifest content is thought to foster an emotional connection between the viewer and
what is viewed. Investigating responses to photographs
of the Kenneth Bigley kidnapping in 2004, Iyer and
Oldmeadow (2006) corroborate this theory identifying that individuals respond with greater feelings of
fear when presented with visual images of the kidnapping compared to individuals presented with textual
information only. Moving the viewer along affective
pathways, visual images illicit particular emotional
responses due to the vividness of their content.
In recent years there has been a considerable
increase in the quantity of both textual and visual
climate change information in the British press. At
a textual level, British newspapers are representing
climate change as an issue falling in and out of public consciousness (Carvalho & Burgess, 2005). There
is an under reporting, however, of images accompanying newspaper articles. Given the potential visual
information has to arouse emotion and concretise risk
messages for members of the public (Joffe, 2007), the
current investigation will explore the content of this
more recent coverage with particular focus on the types
of images being depicted and the impact this is likely
to have for public engagement.
Due to pragmatic and logistical challenges of
obtaining visual material from online newspaper

1293

databases, few investigations have explicitly included

images in media analyses. Those that do, illustrate
the potential power visual information has adding new
levels of meaning to the constructions media coverage is providing, above and beyond that obtained at
the textual level. Exploring media visualisations of the
Gulf War, Griffin and Lee (1995) for example, identify
mixed messages publication of pictures are providing
the public. A lack of images depicting war casualties
and other human costs contradicts claims that media
organisations are providing first hand coverage of the
conflict.
Boholms (1998) examination of visual images of
the Chernobyl disaster also illustrates the power of
visual information. Content analysing over 200 photographs from five European newspapers commemorating the tenth anniversary of the event, the most
prominent category of images concentrated on the fate
of human victims abroad. Presenting the suffering
and dying victims of the disaster as persons rather
than personnel helped to personalise the messages
reported and thereby concretise the risks communicated. Although desolated landscapes are apparent in
the images sampled, it is the personification of the
Chernobyl disaster that the newspapers chose to focus
on. Depicted as individuals, vivid images of suffering
and dying victims portray the devastating effects of
nuclear radiation (Boholm, 1998).
What these studies illustrate is the potential power
visual information has to construct meaning, above
and beyond that obtained at the textual level. A key
question to address, therefore, is the extent to which
media coverage of climate change information can
be appropriately represented in a visual form and the
potential impact this might have on individuals construction of the risks presented. In order to ascertain
such a relationship it is important to first identify
the content of visual images contained in newspaper articles. The present investigation will address the
following research questions:
What is the visual content of recent climate change
coverage in British newspaper press?
What role are these images likely to play in public
engagement with climate change risk information?
2
2.1

METHOD
Sample

Three Sunday broadsheet newspapers (Sunday Times,

Sunday Telegraph and Observer) and three Sunday
tabloid newspapers (News of the World, Mail on
Sunday and Sunday Mirror) were chosen based on
being the most read according to the Newspaper
Marketing Agency (NMAUK, 2006). The broadsheet and tabloid newspapers chosen not only enable

comparison between highbrow and lowbrow viewpoints, they also represent a broad spectrum of political editorial style (Washer, 2004). Sunday newspapers
are considered to contain longer and more analytical
articles than those in daily newspapers and also offer
a feel for how coverage has shifted over the preceding
week (Washer, 2004).
As there are no online databases cataloguing visual
images in newspapers, Lexis-Nexis was used to
initially identify articles about climate change. As
an online newspaper archive, Lexis-Nexis provides
researchers with an invaluable resource for accessing the text of newspaper articles. Using key words
global warming OR climate change OR greenhouse effect appearing in the headline or with a
major mention, all articles from the six Sunday newspapers sampled were selected between 1st January
2000 and 31st December 2006. Duplicates, letters and
other articles not specifically about global warming,
climate change or the greenhouse effect were removed
leaving a workable sample of 300 Sunday newspapers included in the media analysis (208 broadsheet,
92 tabloid). As the focus of this investigation is to
examine content of visual images accompanying these
articles, microfilm copies of all newspapers sampled
were obtained. Photocopies were taken of any image
accompanying an article. Of the 300 newspaper articles sampled, 188 (approximately 60%) had one or
more accompanying image.
2.2

Coding frame

The coding framework chosen contained predefined

high level categories and subcategories for deductive
coding, as well as inductive codes based on distinctions identified in the data (Joffe & Yardley, 2004).
The predefined categories mapped onto the categories
in which the United Nations Inter-Governmental Panel
on Climate Change (IPCC) has categorised the climate
change agenda: the scientific basis (investigating the
causes of climate change), impacts, adaptation and
vulnerability and mitigation. As a starting point
therefore, information contained in newspaper images
was coded Cause, Impact or Solution.
Upon closer inspection it became apparent that
numerous images could be subdivided into more specific categories. A more inductive coding scheme was
therefore devised staying truer to the data and allowing counter intuitive themes to develop. An image of
a flooded village in Britain, for example, rather than
simply coded as impact, was also coded as local
reflecting the location of the impact being depicted.
Further inductive distinctions for this and other themes
were made as the coding process developed. This more
fine grained level of analysis stays truer to the data
obtained whilst encouraging identification of unexpected and intuitive themes to emerge. Less restrictive

1294

than deductive coding, patterns and links operating

between codes are derived from the raw data itself,
making it easier to identify how images are related.
2.3

Data analysis

A content analysis of images was initially conducted

concentrating on salient higher level codes using chi
square statistics to assess important trends. Often considered trite (Silverman, 1993) establishing code frequencies gives the data quantitative rigor but ignores
the context within which information is constructed.
Each category is counted without reference to its relationship to other content. Thus the paths of stories
or linkages between ideas in images are lost. A thematic analysis was therefore also conducted to explore
patterns found within the data. As a more sophisticated level of qualitative analysis, a deeper account
of image content is obtained by emphasising symbolic meanings embedded within the articles sampled.
In the following section, both a content analysis and
exploration of themes will be presented, observable
from visual images accompanying British newspaper
articles on climate change.
3
3.1

RESULTS
Impacts of climate change: Bringing the threat
closer to home

The tangible and immediate effect climate change has

around the world is a prominent theme, with over
50% of images illustrating one or more climate change
impacts. Significant differences were evident between
newspaper types with broadsheet images more frequently depicting impacts than tabloid images (X2 ,
(1), = 4.17, p < .05). Exploring the content of these
images further, impacts relating to ice melt were the
most dominant. Photographs of retreating glaciers and
melting polar ice caps capture both the scale and speed
of the climate change problem. Two recent reports in
the Observer (30/01/05, 26/06/05), for example, use
images of melting ice caps to capture the ferocity of
these impacts. In both images clumps of jagged ice are
depicted breaking away from the ice shelves dramatically crashing into the seas below. These images are
often wide angled, panoramic shots capturing not only
the crumbling ice sheets but also the encroachment
of the ice by surrounding seas.
The retreating glacier provides a potent symbol of
the scale of the problem, made salient by before and
after photographs being provided, giving the viewer
a baseline upon which to form their representation.
An article in The News of the World (03/09/06) for
example discussing climate change impacts, presents
two photographs, one taken in 1980 showing a glacier
in full health, and one of the same landscape taken

26 years later showing the glacier in retreat. Although

glaciers are known to advance and retreat naturally,
such before and after type images help emphasise
the disastrous consequences climate change is having
on the world.
Associated with images of ice melt, newspapers
also include numerous photographs of polar bears in
articles about climate change. As inhabitants of Arctic
environments, their plight is made increasingly salient
as a result of melting polar ice caps. Depicted in several images as struggling to swim to safety, a sense
of futility and desperation is apparent (Mail on Sunday, 24/09/06). As symbols of power and fearlessness
polar bears struggle for survival is made ever more
poignant in the images provided.
Beyond ice melt, flooding also dominates the
content of photographs newspapers use to illustrate
climate change. With an increase of extreme weather
events reported in recent years, flooding is a particularly salient and disastrous consequence newspapers
are keen to visually represent. On the front cover
of a special environmental report in the Observer
(07/11/04), a flooded landscape is the first image seen.
It could easily be a lake were it not for the top of a
pedestrian walkway sign poking precariously out of
the water. In another image cars are shown stranded
on flooded streets (Observer, 11/08/02). A solitary
figure, half submerged walks ironically with umbrella
in hand. Such out of the ordinary visual landscapes
are being used to emphasise the severity of the climate
change challenge.
Although the array of impacts dominating newspaper images illustrates the immediacy of the threat,
the climate change problem is brought closer to home
in more recent coverage. Dividing the six years sampled into two groups (20002003 and 20042006), the
percentage of local impact images in more recent
coverage has increased from 30% to 40% whereas
the percentage of other impact images has decreased
from 48% to 42% in this same period. Despite neither
of these differences being statistically significant, the
direction of the trend is important. Elaborating on what
these labels refer to, local and other impacts often
share common ground depicting flooded landscapes,
biodiversity loss or summer heat waves, but the key
distinguishing feature is location. Both images might
therefore show a flooded landscape but the local
location would be British whereas the other location
would be abroad. Visually bringing the threat closer to
home climate change is being represented as a problem
no longer affecting just populations abroad.
3.2

Personification of the climate change problem

A considerable quantity of images collected did not

feature landscapes as the central focus but people. This
trend is particularly noticeable in tabloid newspapers

1295

reporting significantly more images of people than

broadsheet newspapers (X2 , (1) = 4.68, p < .05).
Members of an affected public, key political figures
and celebrities feature prominently in many of the
images sampled and help personalise the messages
conveyed.
3.2.1 An affected public
Members of an affected public are a key group of people that relay the immediacy of the climate change
problem through the situations in which they are
depicted. Often shown in a variety of dangerous and
life threatening scenarios, these photographs seem to
aim to evoke a heightened emotional response. This is
neatly illustrated from a photograph in the Observer
(05/01/03a) in which two fire-fighters rescue a grandfather and his six year old granddaughter from an
episode of flooding. This content, disambiguated by an
accompanying caption, conveys an it-could-be-you
sentiment enabling individuals to vicariously experience the first hand effects of climate change through
the lives of others. Although the picture also depicts
a flooded landscape, it is a focus on the individuals
being carried to safety that dominates its content.
Another set of images that use an affected public
to represent consequences of climate change focuses
on the inconvenience and disruption warming is having upon lifestyles. The scenario faced by two holiday
makers in the Swiss Alps, for example, is depicted
in one photograph by the lack of snow confronted
upon arrival (Sunday Telegraph, 17/12/06). Standing
on barren ground with snowboards clutched underarm, readers seem to be called to empathise with the
sense of frustration etched on both their faces. In
another image, a cyclist is pictured carrying his bike
along a road closed due to flooding (Sunday Telegraph,
26/09/04). There appears to be an intention to convey irony in this. Such examples illustrate the direct
impact climate change has on the general public but in
a manner that makes the message more personal.
Given the negative connotations mention of climate
change evokes, an interesting set of photographs depict
members of an affected public (children in particular),
enjoying unseasonal weather changes. In one photograph, for example, two small children enjoy being
pulled on a sledge after waking up to a blanket of
snowfall (Observer, 05/01/03b). In another, children
at Kew Gardens play in unseasonal Autumnal sunshine
(Observer, 15/10/06). If visual information plays an
important role in helping individuals foster representations of climate change, the contradictory impact of
such images may require future investigation.
3.2.2 Political leaders
An abundance of images depicting global political leaders illustrate the visual politicisation of the
climate change problem. As an issue dominating the

international stage, climate change has become an

inherently political topic. It is unsurprising therefore
that photographs of key political leaders are routinely
included in newspaper articles on climate change. Of
these, American President George Bush receives the
most attention. To help dramatise the climate change
threat, newspapers are attracted by political leaders
they can vilify. Given the lack of international cooperation America has offered in relation to climate
change treaties in recent years, President Bush is an
obvious choice and images reflect this.
In one image, for example, a picture of the sun is
targeted by a series of heat seeking missiles. President
Bush is depicted in the bottom right hand corner smiling and giving thumbs up to the situation (Observer,
05/11/06). Via this photomontage the viewer is provided with the insinuation that climate change is not
only caused by the sun but is also controllable through
the use of military force. This links to Bushs reputation as a war-mongering President as well as focussing
on the non-anthropogenic argument for the cause of
climate change.
Other images depicting President Bush are not as
easily interpreted owing to the lack of relevant context. Whereas it is often easy to discern the messages
contained in an image of a melting polar ice cap or
retreating glacier, there is often an inherent ambiguity surrounding the interpretation of images depicting
people. Disambiguation of these images is therefore
reliant on captions. The link connecting President
Bush and the climate change issue is made through
the visual, but the negative connotation attached is
only made salient though the caption.
Following popularisation of his climate change documentary An Inconvenient Truth, former American
Vice-President Al Gore is also regularly depicted in
photographs accompanying newspaper articles (for
example Sunday Telegraph, 19/11/06). Rather than
vilified by press coverage however, Al Gore is presented as a climate change ambassador. Whether it
is his stance or the facial signs expressed, a sense of
authority is gained from viewing these images often
depicting Al Gore delivering his lecture. Al Gore also
personifies the international effort being made to solve
the climate change problem.
Although Bush and Gore dominate visual images
of international political figures accompanying climate change articles, former British Prime Minister
Tony Blair and former Deputy Prime Minister John
Prescott attract their fair share of domestic coverage. In
a notable example, three real time photographs depict
John Prescott dramatically storming out of a recent
climate change convention (Observer, 26/11/00). Providing three successive images, taken just seconds
apart, gives the viewer a sense of action only usually gained viewing video footage. After attempting
to broker a proposal that would subsidise American

1296

greenhouse gas emissions, the photographs depict

Prescott leaving at the precise moment that talks collapsed. Not only do these images capture the volatility
of international climate change negotiations, but also
the real time footage gives the viewer a vicarious sense
of being there, amidst the turmoil.
3.2.3 Celebrity icons
Impassioned by the climate change cause, celebrities
are increasingly represented as environmental spokespeople in newspaper images. Pictures of famous
people are used to persuade individuals to take up
the green agenda and change their unsustainable
behaviours, reflecting the Hollywood-isation of environmental issues more generally. Celebrity protesters
Rory Bremner and Bianca Jagger, for example, are
depicted in one image promoting the boycott of Esso
petrol stations (Mail on Sunday, 06/05/01). Depicting
celebrities endorsing environmental consciousness is
a persuasive mechanism encouraging the public to do
the same.
Celebrities also capture visual newspaper attention
by fostering climate change activism. Headlining at
a recent Stop Climate Chaos rally in London, for
example, British pop singers Johnny Borrell and KT
Tunstall are depicted in two images aimed at raising
public awareness (Sunday Mirror, 05/11/06). Pictured singing to crowds of fans, they are keen to
show their commitment to environmental causes, as
accompanying captions suggest. In another image,
American pop group Scissor Sisters are shown
performing to fans in Londons Hyde Park (Sunday
Telegraph, 10/12/06). Although taken during the
recent Live8 concerts, the caption accompanying the
image emphasises their involvement with Global
Cool, a similar event encouraging people to reduce
carbon emissions. Although such images are not as
prominent in newspapers as an affected public or
political leaders, growing numbers of celebrities are
keen to have their image associated with the climate
change issue.
3.3

Graphical representation of climate change

A final group of images that feature prominently in

newspaper articles represent the climate change problem in graphs. Broadsheet newspapers, in particular,
adopt the conventional format of steadily rising line
graphs more common in scientific reports on climate
change (IPCC, 2007). Broadsheet readers are also provided with complex statistical information in readily
digestible formats more frequently than tabloid readers
(X2 , (1) = 8.16, p < .05).
Three noteworthy sets of images illustrate this trend.
A series of line graphs and bar charts are used to
represent the climate change threat in graphical form
(Observer, 27/02/00, 01/04/01 & 11/06/06). Most

notably, all three images have prominent bar charts

depicting a world carbon league table. Measuring the
relative greenhouse gas emissions different countries
contribute, America is in pole position. Such figures
symbolise the impact carbon emissions have on the
planet. The viewer is provided with a visual representation of America as the worst polluter. Furthermore,
this trend is contrasted to the relatively small contributions made by other counties. also use a variety
of line graphs to graphically represent salient climate
change impacts. In line with findings from the most
recent round of IPCC reports (IPCC, 2007), levels of
atmospheric carbon dioxide and global temperatures
in particular are predicted to rise steadily over coming
years. By using line graphs to represent this information, viewers are provided with a conceptually clear
format through which to appreciate these rises. Most
frequently, a single line is depicted, plotted on a graph
where the x axis is time (calculated in years) and the y
axis is impact measured (atmospheric carbon dioxide
calculated in parts per million, temperature calculated
in degrees Celsius and so on). What is most notable
from these graphs, however, is not the axes used or
even the impact depicted but rather the rate at which
they are rising. Indeed, to keep formats simple, minimal information is provided in the graphs presented.
In one image (Observer, 11/06/06), carbon dioxide is
shown rising sharply over recent centuries (from 280
parts per million in 1800 to over 360 parts per million
in 2000) but no elaboration is given regarding what
this means. Although the relative magnitude of this
200 year change is unknown, it is simply a dramatic
increase that is being depicted.

DISCUSSION

Visual information is clearly a dominant feature in

British broadsheet and tabloid newspaper coverage of
climate change. Accompanying nearly two thirds of
all articles sampled, the content of visual images can
be classified within three broad themes: the impact
of climate change: bringing the threat closer to home,
personification of climate change and graphical representation of climate change. The following section will
highlight likely reasons why newspapers have focused
on these three areas in their selection of climate change
imagery and in particular, focus on the role imagery
is likely to play in directing public perception of the
climate change issue.
Over half of all articles had one or more image
depicting a climate change impact. Ranging from
melting ice caps to retreating glaciers, newspapers
provide viewers with a visual catalogue of the influence climate change is having on the world. Bringing the problem closer to home, newspapers are
increasingly choosing to represent the climate change

1297

threat more locally, with newspaper images depicting

UK landscapes. At one level, this trend mirrors the
textual content of the newspaper articles accompanying these images. In reviews of this coverage, Carvalho and Burgess (2005) and more recently Boykoff
(2007) highlight a variety of extreme events that have
helped drive media interest in climate change over
recent years. Ranging from the devastation caused by
Hurricane Katrina to more recent summer flooding in
British towns and cities, extreme weather events provide the mass media with an attractive hook with which
to document tangible examples of alleged climate
change impacts.
The relationship current events have with newsworthiness also helps account for the depiction of the
climate change issue. Investigating the rise and fall
of risk reporting, Kitzinger and Reilly (1997) argue
that future orientated threats characterised by distant
and unknown impacts do not attract widespread media
attention. As newsworthiness of a story is ultimately
dependent on how tangible its impact is, presenting climate change as an unknown quantity hinders
the likelihood of its continued media interest. Visual
images are therefore used to represent the climate
change threat as real and no longer something potential
and future orientated. Although articles about retreating glaciers convey to the reader the realities of the
climate change threat, an accompanying image visually solidifies the tangibility of this threat by providing
the viewer with a concrete example of the impacts
climate change can have.
This need for media to have concrete material to
represent is neatly illustrated by an investigation of
media attention to BSE. Whilst conducting their interviews, Kitzinger and Reilly (1997) mention that one
news journalist, before being allowed to cover the
impending crisis, commented we needed dead people,
well, weve got them now (p. 344). Climate change
has been a topic of concern for the past two decades
but only recently have material events such as retreating glaciers, melting ice caps and summer flooding
served as news hooks permitting a change in degree of
newsworthiness. Although not icons as Perlmutter and
Wagner (2004) would describe, this group of images
can be considered symbols representing the climate
change threat, with concrete examples exemplifying
the immediacy of the impacts.
The second theme attracting visual media attention
in British newspapers focuses on the personalisation of
climate change. Particularly salient in tabloid newspapers, images of an affected public, key political figures
and celebrities account for the majority of these visual
images. In other media analyses of visual information, pictures of people are often found to accompany
textual news articles. Boholm (1998) for example, in
her investigation of media coverage of the Chernobyl
disaster, identified that the vast majority of newspaper

images depicted people rather than landscapes. Griffin

and Lee (1995) in their exploration of Gulf War photography in popular American newsmagazines, also
identify images of people depicted in numerous dominant picture genres. Attracted to the human face of
science and risk reporting (Kitzinger, 1999), journalists actively seek out stories in which a human interest
factor can bring the content alive. Including a photograph depicting a dramatic rescue in a newspaper
article about flooding, for example, is more likely
to generate a sense of empathy with the viewer than
simply presenting the text.
The power images of people can have has been reliably identified in a series of investigations exploring
charitable giving. Investigating the identified victim
effect, for example, Small, Lowenstein and Slovic
(2007) discovered that people are significantly more
willing to donate money when presented with pictures
of identified victims as opposed to statistical information alone. The strength of this effect diminishes,
however, with inclusion of extra people in pictures
presented. Exploring the role emotions are hypothesised to play in decision making processes, Kogut
and Ritov (2005) identify a strong positive relationship
between reported distress and willingness to contribute
to charities. Evoking vicarious feelings of both distress and empathy, images of identified victims seem
to exemplify the vividness of the contexts depicted.
Whether or not personification of climate change
risks in British newspaper imagery evokes a comparable emotional tone in readers will be important to
investigate.
The final group of images made salient in British
newspaper coverage depict graphical representations
of climate change risks with line graphs conceptualising the threats. Given the inherent conceptual limitations individuals have interpreting complex statistical
information (Kahneman & Tversky, 1979), there are
a surprising number of visual images depicting quantitative presentation of data. In particular, there is a
general trend for newspapers to use line graphs to
depict numerical information. Corroborating results
from Zacks, Levy, Tversky and Schiano (2002), in
which graphs from four US newspapers were content
analysed, line graphs accounted for half of all presentation formats. In particular, simple line graphs
depicting straightforward relationships between two
variables were the most common format identified.
As contemporary societies rely heavily on graphics to
communicate statistical data, simple graphical representation that has the power to make quantitative information perceptually salient and memorable, is an
important component of print media communication
(Zacks et al, 2002, p. 187).
Although the types of images newspapers choose
to accompany articles have been identified in the
present analysis, it is important to consider how these

1298

themes might influence public engagement with climate change risk. Given the complex empirical link
between mass media presentation and public uptake
of risk information, it is important to explore the role
visual imagery might play in mediating this relationship. The vividness of visual images, in particular,
has been highlighted as an important component in
understanding emotional reactions to graphic images
(Iyer & Oldmeadow, 2006). Photographs appear
to direct emotionally fearful responses due to the
graphic nature of the content presented. This has
enabled Iyer and Oldmeadow (2006) to speculate
about the role visual imagery can play in influencing individual responses to threatening events. At a
minimum, such images should at least attract attention to media reports through which information
about unknown and threatening phenomena is often
obtained.
This line of enquiry is particularly important to
consider given the strong affective component
researchers attribute to risk perception processes.
Whereas traditional risk perception theories focus
on the rational and probabilistic reasoning practices
individuals are hypothesised to employ when making judgements about unfamiliar information (Slovic,
1987), more recent research has emphasised the
important role positive and negative affect plays in
guiding judgement and decision making (Finucane,
Alhakami, Sloivc & Johnson, 2000). In their investigations of individual associations with climate change,
Leiserowitz (2005) and Lorenzoni, Leiserowitz, de
Franca Doria, Poortinga and Pigeon (2006) highlight
the importance affective images can play in public
engagement with risk information. More specifically, this identification helps develop a conceptual
understanding of the important role emotion plays in
engagement with risk. Furthermore, given the emotive
power of visual information, it will be important to
explore the suggestion that newspaper imagery helps
shape public engagement with climate change risk.
This is particularly pertinent given identification of
visual themes does not automatically uncover how the
public engages with the issue.
5

CONCLUDING COMMENTS

The present study has identified salient images newspapers choose to visually represent the climate change
threat. In doing so, it builds upon previous media
analysis research focused on the textual content of
newspaper articles. Notwithstanding the importance
findings from these studies have uncovered, visual
imagery offers researchers a novel and exciting opportunity to investigate media coverage of climate change
information. Furthermore, given the power the visual
has to evoke strong affective responses, newspaper

imagery is expected to play an important role fostering

public engagement with risk issues.
REFERENCES
Boholm, A. 1998. Visual images and risk messages: commemorating Chernobyl. Risk Decision and Policy, 3:
125143.
Boykoff, M.T. 2007. Flogging a dead norm? Newspaper
coverage of anthropogenic climate change in the United
States and United Kingdom from 2003 to 2006. Area, 39:
470481.
Carvalho, A. & Burgess, J. 2005. Cultural circuits of climate
change in UK broadsheet newspapers, 19852003. Risk
Analysis, 25: 14571469.
Corner, J. 1991. Meaning, genre and context. In Curren, J. &
Gurevitch, M. (Eds), Mass Media and Society: 267284.
London: Edward Arnold.
Finucane, M.L., Alhakami, A., Slovic, P. & Johnson, S.M.
2000. The affect heuristic in judgments of risks and
benefits. Journal of Behavioral Decision Making, 13:
117.
Griffin, M. & Lee, J. 1995. Picturing the Gulf War: Constructing an image of war in Time, Newsweek, and US News
and World Report. Journalism & Mass Communication
Quarterly, 72: 813825.
Intergovernmental Panel on Climate Change (IPCC). 2001.
Climate Change 2001: The Scientific Basis. Cambridge:
Cambridge University Press.
Iyer, A. & Oldmeadow, J. 2006. Picture this: Emotional and
political responses to photographs of the Kenneth Bigley
kidnapping. European Journal of Social Psychology, 36:
635647.
Joffe, H. 2007. The power of visual material: persuasion,
emotion and identification. Diogenes, 55: 8493.
Joffe, H. & Yardley, L. 2004. Content and thematic
analysis. In Marks, D.F. & Yardley, L. (Eds.), Research
Methods for Clinical and Health Psychology: 3955.
London: Sage.
Kahneman, D. & Tversky, A. 1979. Prospect Theory
Analysis of Decision Under Risk. Econometrica, 47:
263291.
Kitzinger, J. & Reilly, J. 1997. The rise and fall of risk reportingMedia coverage of human genetics research, False
Memory Syndrome and Mad Cow Disease. European
Journal of Communication, 12: 319350.
Kitzinger, J. 1999. Researching risk and the media. Health
Risk & Society, 1: 5569.
Kogut, T. & Ritov, I. 2005. The Identified victim effect:
An identified group, or just a single individual? Journal
of Behavioral Decision Making, 18: 157167.
Leiserowitz, A.A. 2005. American risk perceptions: Is
climate change dangerous? Risk Analysis, 25: 14331442.
Lorenzoni, I., Leiserowitz, A., Doria, M.D., Poortinga, W., &
Pidgeon, N.F. 2006. Cross-national comparisons of
image associations with global warming and climate change among laypeople in the United States of
America and Great Britain. Journal of Risk Research, 9:
265281.
Newspaper Marketing Agency (NMAUK) 2006. Newspaper readership figures. NMAUK [On-line]. Available:
http://www.nmauk.co.uk/nma/do/live/homePopulate

1299

Perlmutter, D. 1998. Photojournalism and Foreign Policy:

Icons of Outrage in International Crises. Westport:
Praegar.
Perlmutter, D. & Wagner, G.L. 2004. The anatomy of a
photojournalistic icon: marginalization of dissent in the
selection and framing of a death in Genoa. Visual
Communication, 3: 91108.
Silverman, D. 1993. Interpreting Qualitative Data: Methods
for Analysing Talk, Text and Interaction. London: Sage.
Slovic, P. 1987. Perception of Risk. Science, 236: 280285.
Small, D.A., Loewenstein, G., & Slovic, P. 2007. Sympathy
and callousness: The impact of deliberative thought on
donations to identifiable and statistical victims. Organizational Behavior and Human Decision Processes, 102:
143153.
Wahlberg, A.F. & Sjoberg, L. 2000. Risk perception and the
media. Journal of Risk Research, 3: 3150.
Washer, P. 2004. Representations of SARS in the British
newspapers. Social Science & Medicine, 59: 25612571.
Zacks, J., Levy, E., Tversky, B. & Schiano, D. 2002. Graphs
in Print. In Anderson, M. Meyer, B. & Oliver, P. (Eds.),
Diagrammatic Representation and Reasoning: 187206.
London: Springer.

NEWSPAPER REFERENCES
Mail on Sunday. May 6, 2001. Stars urge a boycott of all Esso
garages.
Mail on Sunday. September 24, 2006. A future too hot to
handle.
News of the World. September 3, 2006. Weve got ten years
to save the planet.

Observer. February, 27, 2000. Focus: Climate in crisis: Its

apocalypse now as world overheats: Scientists fear that
global warming has gone beyond the point of no return.
Observer. November 26, 2000. Climate talks end in disarray.
Observer. April 1, 2001. Global warming: The President who
bought power and sold the world.
Observer. August 11, 2002. Monsoon Britain.
Observer. January 5, 2003a. Whatevers happened to the
weather?
Observer. January 5, 2003b. Builders face huge flood tax
to meet cost of climate change.
Observer. November 7, 2004. Flood, sweat and tears.
Observer. January 30, 2005. How we put the heat on nature.
Observer. June 26, 2005. Turning the tide: Planet in Peril.
Observer. June 11, 2006. Energy: Its too late for the planet:
or can we pull it from the fire?
Observer. October 15, 2006. Climate change: Seasons of
mists?
Observer. November 5, 2006. George Bushs war on
warming.
Sunday Mirror. November 5, 2006. Ten years to save the
world.
Sunday Telegraph. September 26, 2004. Counting the cost of
global warming.
Sunday Telegraph. November 19, 2006. At state is nothing
less than the survival of human civilisation.
Sunday Telegraph. December 10, 2006. Rock stars to fight
greenhouse effect with cool aid concerts.
Sunday Times. December 17, 2006. Climate change may kill
European skiing.

1300

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Do the people exposed to a technological risk always want more

information about it? Some observations on cases of rejection
J. Espluga
Universitat Autnoma de Barcelona, Spain

J. Farr & J. Gonzalo

Universitat Rovira i Virgili, Spain

T. Horlick-Jones
Cardiff School of Social Sciences, Cardiff University, UK

A. Prades, C. Oltra & J. Navajas

CIEMAT, Barcelona, Spain

ABSTRACT: There exist situations in which people who perceive themselves exposed to a technological
risk feel uninformed but, at the same time, refuse to absorb new information about the said risk. With the
objective of understanding this paradoxical situation, in this text, we are going to compare some empirical
data obtained in two different studies carried out in Spain about the perception of technological risks, the first
one, very specific and visible, related to petrochemical plants; and the second one, more abstract and opaque,
related to nuclear fusion. According to the results some of the reasons that could explain this situation can
be found in the structure of the social and institutional context in which people experience and perceive risks;
a context in which acknowledging new information implies assuming new responsibilities on factors out of
peoples control.

INTRODUCTION

The relation between information about risks and the

social responses to them are influenced by various
social, contextual and situational factors (Irwin et al.
1999; Burger et al. 2007; Visschers et al. 2007). When,
during the last third of the 20th century, there was a
proliferation of social rejection to certain technologies,
it was considered a priority in governmental, industrial and scholarly circles to study the perception of
risk held by the individual, so as to take measures to
reduce the distance between lay people and experts.
This gave way to the concept of acceptable risk in
which, by weighing up all the factors and benefits the
citizen would be persuaded to stop opposing supposedly dangerous plants and technologies (Starr 1969).
With this, importance was placed on the investigation
into the communication of risks as it was evident to the
experts that for the public to give up their irrational
beliefs and behaviour it would be necessary to inform

them correctly about the risks. In fact it was not

until the end of the 1980s and during the 1990s when
emphasis began to be placed upon perspectives of
integration which took into account the social, economic, political and cultural framework in which the
process of perception and communication of risks
takes place (Frewer et al. 1996, Frewer 1999; HorlickJones 2005; Pidgeon et al. 2003; Wynne 1991,
1996).
The majority of classical studies about perception
and communication of risks have centered on the
analysis of situations in which, while experts and
decision makers assure that the risk is under control, the public are distrustful and express explicitly
their opposition to the technology (via individual or
collective protests, demonstrations etc.). However,
we are going to focus on another type of situation:
that in which the public do not publicly reject a risk
but resist being informed about it. We consider that
these types of situations have received less study and

1301

permit a clearer analysis of the weight of the social

and institutional context in relation to the citizens
experiencing the risks.
There exist situations in which people who perceive themselves exposed to a technological risk
feel uninformed but, at the same time, refuse to
absorb new information about the said risk. With the
objective of understanding this paradoxical situation,
in this text, we are going to compare some empirical data obtained in two different studies about the
perception of technological risks. The first study, titled
El proceso de comunicacin de riesgo en Tarragona
[The risk communication process in Tarragona], was
financed by the Spanish Government Plan General de
Investigacin del Ministerio de Educacin y Ciencia
(SEJ2004-00892). This project analyses the processes
of communication and perception of technological
risks in relation to the numerous chemical and petrochemical industries in the area of Tarragona, Spain,
occupying an overall area of 1,200 hectares and made
up of 27 companies (including such multi nationals
as Bayer, BASF, Dow Chemical and Repsol-YPF).
The second study, Investigating lay understanding
and reasoning about fusion technology, is a project
financed by European Fusion Development Agreement (EFDA 06/1427) and UK Economic and Social
Research Council (RES-000-22-2664). An important
objective of this latter research project is to find
practical ways of improving communication about
fusion, and to enhance the quality of lay engagement with sources of information about fusion. It
should be noted that the data analysis of this project
is still on-going, and that here we will be dealing
with preliminary insights from the data. Clearly additional hypothesis may still emerge as the analysis
progress.
Both research projects have been based on a qualitative methodology directed, among other things, at
obtaining discourses from citizens about their perceptions, experiences and expectations with respect to the
technologies in question, and their associated risks:
petrochemical and nuclear fusion. We consider it interesting to compare these two projects as they relate
to two very different technologies: the petrochemical industries are very visible and concrete elements
with well established risk evaluations and preventative
obligations, and the local population has been living
with them for decades (Farr & Fernndez, 2007);
meanwhile nuclear fusion is still a relatively abstract
area around which there are very few estimations of
risk or control norms (preventative measures, emergency plans etc.), and of which nobody has experience
(Prades et al 2007, 2008). This radically different nature of the technologies we will be analyzing
should be carefully considered when interpreting our
hypothesis.

2
2.1

METHODOLOGY
Research project on communication and social
perception of risk in the petrochemical area of
Tarragona

Eight focus groups were created with the objective of capturing the principal discourses about the
experiences, perceptions and expectations around the
chemical and petrochemical industries of Tarragona
and their potential risks. Two types of municipalities and neighbourhoods of Tarragona were chosen
according to their territorial proximity to the petrochemical plants: a) close (Bonavista, La Canonja,
Constant, La Pobla de Mafumet); and b) distant
(TarragonaCentro, Reus, Vila-seca). While in those
areas in the first group the predominant socioeconomic activities are closely related to the chemical
industry sector, those in the second group have a
concentration of administrative and economic activities based on the service industries such as mass
tourism. Field work was carried out between April
and June 2006. A total of 50 people participated (60%
women and 40% men) with an average age of 51 and
a range of ages from 1692, of whom 8% were uneducated, 36% had completed primary school, 42% had
completed secondary education and 14% were university educated; and had diverse labour situations.
For the analysis of the information we used a constant
comparative method according to grounded theory
procedures.
2.2

Research project on the social perception

of risk in nuclear fusion

A hybrid group-based methodology, using reconvened

focus groups (RFG), was developed in order to allow
groups of lay citizens to engage with issues concerned with fusion. In this way, respondents would
be taken through a learning and discussion process;
one that would allow them to assimilate sometimes
quite complex information, and examine this information in a considered way. RFG have shown their
suitability in attempts to generate deliberative process, where there is careful consideration of, and
discussion about, evidence for and against a measure,
with a view of making a decision (Horlick-Jones et al,
2007). Field work was carried out in Barcelona (Spain)
between October and December 2007, and six RFG
were run (each meeting on two occasions), composed
of suitably-selected lay citizens. When recruited, participants will be told that they are being invited to
take part in discussions about energy issues, and
that the work is being sponsored by the European
Union. Those likely to have specialist knowledge or
strong views about fusion were excluded. The groups

1302

were segmented along age (18 to 25; 26 to 40; over

40) and socio-economic lines (AB: High & Medium
High status; CD: Low & Medium Low status). All
groups were of mixed gender. As mentioned above,
we draw upon preliminary findings as the analysis is
on-going.
3
3.1

RESULTS
Evaluation of available information
about technologies

more reliable information. In general, they consider

the available information too technical and difficult
to understand, so more accessible information would
be needed for the general public. Our participants
even propose that the information they have should be
passed on to experts in the field to see what they will
say; i.e., to check if the experts are able to produce
an integrated and balanced version of the radically
different available informative materials.
3.2

Two simultaneous tendencies can be observed in the

case of the Tarragona petrochemical area. On one
hand, the majority of the participants in the focus
groups express certain ignorance about what the local
chemical plants produce as well as the processes and
materials used. Some of them try to name some of
these processes and products without much success
and with a high level of confusion. They also admit
to having little information about the incidents and
accidents which happen in these industries, which contribute to the opaque image. On the other hand, the
residents of the neighbourhoods and municipalities
closest to the plants tend to say that the public are well
informed about what happens in the industries as the
companies periodically hold conferences, especially
with the neighbours associations. Even so, they also
say that important topics about risks are not usually
discussed in these meetings and chats.
In the case of nuclear fusion (here we draw upon
preliminary findings from ongoing data analysis), during the first sessions of the group meetingswhen
knowledge about fusion was very limited indeed if
existent at all participants find it difficult to value
the technology, and tend to agree on their lack of
information about what it is, the context in which
these technologies can be developed, possible alternatives etc. They lack more public information about
the project to construct a big nuclear fusion laboratory in France, and especially about what this could
represent. At this stage of the process, and in general terms, participants tend to agree that making
information publicly available will be essential for
the project to be accepted socially, especially taking
into account the size of the project, the huge economic costs for the participating countries and the
future repercussions. In the second meetings of the
groups, after consulting the information provided by
the research team, most participants consider that there
is a wide difference between the materials elaborated
by the project promoters (fusion community) and the
ones from the ecological associations, which contributes to creating confusion, perplexity, and lack
of trust amongst the readers. This partial and differing information increases the sensation of needing

Evaluation of available information

about technological risks

In the case of the petrochemical area of Tarragona,

the focus groups participants barely refer to the necessity of information about accident risks. They merely
mention that they usually receive little or delayed information about incidents or accidents produced in the
petrochemical industry. On the other hand, a large
number of the participants say that they know that data
is collected about atmospheric contamination derived
from these industries, as many of them have seen the
measuring apparatus, but in general they do not seem
to know very well how to consult the data or who is
administers the data gathering.
It was also observed that none of the participants
have tried to find the information, which leads to the
deduction of a tendency to trust the administrators and
delegate control to them. In some focus groups it was
said that the apparatus dispersed around the territory
are from the environmental department of the regional
government (Generalitat de Catalunya) and some participants said that they knew how to find information
on atmospheric contamination from the departments
webpage. However, they think that the information
on the webpage is not very detailed and of little use to
know the real situation, as it only reports monthly rates
of contamination for various contaminants, but it is not
possible to know if there have been episodes of acute
contamination. In general there is agreement in the
focus groups that the population rarely have information from the chemical and petrochemical companies
about risks, but they assume that the local councils
must have this information although it is seldom made
public. In any case, the perception exists between the
participants that more information is available about
the industries now than in the past, which indicates a
positive evolution.
In the case of nuclear fusion, the group participants
perceive a significant lack of information about the
possible risks of a fusion power plant. More precisely,
participants lack details about possible risks in the
materials produced by the fusion community thaton
the contrarywere one of the key issues in the materials by the environmental groups, and thus, lead once
more to confusion and perplexity. Even though, as

1303

could be expected for a nuclear complex technological facilitythey would prefer the plant to be far from
their homes. And it seems that they do not need more
information to reach this conclusion. Furthermore they
are afraid that the developers are not a reliable source
of information about risks (as it happens in other siting contexts with which our participants are familiar)
and consequently do not have clear ideas on how to
get more information on the subject. On the other
hand they feel that the decision over the nuclear fusion
project has already been taken and it is unlikely to
be rescinded which also stops them from demanding
more information.
3.3

Information about risk control methods

There are no comments in this section from the

study on nuclear fusion as, being a technology still
to be developed, there are no preventative or public
emergency measures.
In the case of the petrochemical area of Tarragona,
some people (especially residents of neighbourhoods
and villages furthest from the plants), claim that they
do not know of the existence of emergency or evacuation plans in case of a grave accident, or at least say that
they have not received information about such plans.
However, in the closest neighbourhoods and municipalities the people know of the existence of such a plan
(called PLASEQTA) which establishes some public
obligations. They also feel that the companies do not
inform the public sufficiently about what to do in case
of emergency. For all these reasons there are people
who say that if there was an emergency they would not
know how to act.
Curiously, many of the focus group participants
affirm that they do not feel well informed about prevention and emergency measures, but say that, even if
they were, they would have no intention of doing what
is expected of them. This reasoning indicates that they
do have some idea about they should do in these cases.
In this sense, the people demonstrate their confusion
with respect to PLASEQTA, a plan created by the
governmental institutions which establishes a series
of norms to be followed in case of a serious chemical
accident (to run to home, close the windows, switch
on the radio, etc.). Here there is a paradox: at the same
time as they say that they do not have information,
they know what they should do. This demonstrates
that the demand for more information should be treated
carefullythey probably need more information but
of a different type to that which they are getting from
PLASEQTA. We can observe that these people will not
easily accept more information for various motives: in
the first place, because they generally have no intention of following the emergency procedures; they are
norms which collide with certain values which the people consider important (putting the family before your

own safety etc.). Secondly, by clinging to the attitude

that they do not want to think about the risks, following
the argument that it is useless as nothing can be done,
means that possible emergency norms seem irrelevant.
Furthermore people perceive that there are no means
of warning nor of evacuation in case of necessity. For
all these reasons the following of procedures in the
event of an emergency is doubtful for a large part of
the population.
3.4

Do the people really want to be better

informed?

In the case of petrochemical risks, in general, it does

not seem very clear that people desire more information about risks or preventative measures. It can be
deduced that there is an apparent rejection of receiving more information on this subject. On the one
hand, some people maintain that sometimes it is preferable not to know too much about contamination and
risks of accidents. On the other hand people recognise
thatirrespectively of the quantity quality of the available information those who want to be informed will
actively search and, thus, will be somehow informed,
and those who do not want to be informed, will remain
un-informed (even if they have plenty of information
just by their door). In this generalised attitude, the
way in which people perceive and experience the risks,
characterised by rigid interrelations between agents,
in which the public have very little power to change
things, plays an important role. For these reasons they
argue that it is the same to be informed or not as nothing can be done to avoid the risks. The only action
contemplated as a possible alternative is a change of
residence, something which the majority of people are
not considering. All together, it is a context where
information about risks and their prevention can be
understood as an obstacle to everyday life.
Some of the focus group participants suggest that
the information should be publicly available so that
those who want to be informed can be. However, it
seems that not everyone wants to inform themselves
about the risks or the preventative measures. In fact
a generalised perception exists that the public are not
used to demanding more information about the risks,
and, in this sense, there are those who find it significant that people who have lived their whole lives in the
Tarragona area have barely any knowledge about the
subject. Even though participants find this situation
lamentable, no-one proposes any solution to change
the situation and they tend to blame the local politicians, but without much conviction. Finally they end
up by relating the low level of information that the
people have with the scant social mobilisation against
the petrochemical risks, although they express doubts
about what could be the reason behind this link: is
there a lack of mobilisation because people have so

1304

little information or do people say that they have little

information to justify their lack of action?
In the case of nuclear fusion, in spite of emphasising the lack of informationeven in the second group
meetings after having consulted various documents on
the topicparticipants do not seem to be interested in
receiving more information as far as it is similar to the
available one they have just had access to but to some
extent they demand a new, balanced and integrated
version of what is known about fusion. Resistance to
the available information on fusion seem to rely, on
the perception that it is quite technical, and abstract,
so supposedly very few people would be able to understand, and that it is not sufficiently close to their lives
to dedicate their (often scarce) time to. More over, the
available information it is perceived to be biased and all
of this creates certain resistance to acquiring new information of this type, as it seems to generate even more
confusion about what the technology means. Additionally, participants consider that the general public would
not be interested in receiving more information, and
we observe that this lack of interest in getting more
information may be also explained by the fact that
the possible consequences (both positive and negative) of the nuclear fusion project are seen as long term,
even leading to the idea that another generation will be
affected (and not this one). Therefore they understand
that it will be up to others to accept or reject this given
technology. In a certain way, the participants express
their resistance to taking responsibility for something
they will probably never see, and about which (they
perceive) they can do nothing. For this reason, they
allege, beside their lack of time and maybe interest,
powerlessness to do anything to change the future,
resistance to take responsibility for effects that could
appear in a distant future (when they are no longer
here) etc.
In line with the idea of not looking for new information, our participants consider that it should be
the responsibility of the State to draw up reliable and
balanced information to the public. Even then they
acknowledge that they probably would not pay much
attention to it, but they feel that it is important that the
information should be available to the few people (in
their opinion) who would use it.
4

CONCLUSIONS

The results obtained offer a possible way to understand

the apparently paradoxical situation in which a population apparently prefers not to receive more information
about the risks to which it is exposed . To begin with,
it has been confirmed that the people studied recognise that they feel uninformed about the technologies
in question and their possible risks, from those who
feel exposed to a concrete and visible risk such as

the petrochemical plants, to those who reflect on an

abstract and still inexistent risk like nuclear fusion.
Therefore, the chemical groups say they are not
sure about what the petrochemical companies exactly
do, even though it can be seen that the people someway
connected to the plants do have a certain knowledge
about how they function, and, in general, feel exposed
to health risks. However, in spite of everything, they
do not dare to publicly express their evaluation of the
situation in terms of a balance of risks/benefits and
prefer to say that they are unaware of any risks. Meanwhile, the fusion groups recognize that they do not
have sufficient information about this technology nor
about its potential risks, and after reading the material
provided by the investigating group, most participants
seem to still be confused and with a greater sensation of a lack of reliable information than before. This
could, of course, reflect both the inherent difficulty of
assimilating information about fusion (as those able to
handle the technical details of fusion were willing to
learn more about it); and the perceived vested interests of the institutions producing such informative
materials (How can I know where is the truth?). In
both cases, however, we can see that the people are
prepared to show their ignorance and lack of adequate
information about technologies and their associated
risks. This does not necessarily imply that they demand
more explicit information. This factor is particular significant for the petrochemicals groups. It is perhaps
less surprising for the fusion groups, as they encounter
the technology largely as an abstraction, rather than a
tangible entity in their local environment.
Secondly, in the case of exposure to petrochemical
risks the people admit that they have a deficit of knowledge about the preventative norms and how to act in
case of emergency, while in their discourses there are
sufficient indications to think that they know what they
are expected to do. However it is apparent that they
prefer to give an image of relative ignorance. Probably the negation of the population to speak explicitly
about the risks and preventative measures is related
to the refusal to accept certain responsibilities which
the authorities (both public and industrial) try to foster
upon them (obligations in case of emergency). These
responsibilities are imposed in a framework in which
the citizen can decide nothing due to the high level of
dependancy and the lack of autonomy in which they
are exposed to risks. It could be argued that a similar
lack of capacity to act is also present for the participants thinking about the risks of nuclear fusion. They
tend to relate this technology with large projects in
which decision making is out of the hands of the citizen. Which is to say done deals against which little
can be done.
Here, we suggest that the concept of responsibility
could provide a key to interpreting the behaviour of
these people. It is a concept which implies an agent

1305

with a certain degree of autonomy given that someone

who cannot act or take decisions will find it difficult to
be responsible for something. Therefore, both in the
case of the population exposed to the petrochemical
risks in the area of Tarragona and the citizens that
reflect on the nature of nuclear fusion, we encounter
a population who (perceive that they) can not act nor
decide anything about the risks or their own degree of
exposure to the risks.
Possibly for this reason, people refuse to play the
game, and, in justification, allege ignorance and the
uselessness of doing anything (cf. Michael, 1996 on
the construction of ignorance). This can lead to individual and collective demoralisation and no demand
for more information. In a certain way it could be said
that this generalised strategy to maintain a discourse
of fate and inertia (needed on an individual level to
be able to live normally) facilitates the absence of an
explicit social conflict. This does not mean that the
people do not want information. What happens is they
find themselves subordinated by a social and institutional context in which it is preferable to pretend that
they do not know and do not want to know. For the
people to publicly express their desire for more information about risks it will be necessary to change the
relation of forces so that the there is an increase in
citizen autonomy in relation to the other actors.
These results have several practical implications.
First, the management of these technologies (both
petrochemical and nuclear fusion) must be accompanied by relaying information to people although
they did not ask for it. In other words, information
on risks and preventive measures should always be
available and accessible to everyone. In addition, this
information should be presented in different levels of
complexity, from the more technical type of information to the most informative, and there should be
mechanisms which lead from one level to another.
Secondly, information about the risks should be communicated through sources that are trusted by the
population, or that provide sufficient transparency in
order to reduce suspicions. Therefore, the involved
institutions themselves rarely may be the only or main
sources of information. Transparency in this area
seems more feasible in the case of concrete and visible risks (such as petrochemicals) than in the abstract
and invisible risks (such as nuclear fusion). Finally, we
must be aware that people are often reluctant to accept
theirs obligations in case of petrochemical risk emergency, and that this kind of prescriptive information
(about how people should behave) often proves to be
useless if it is not accompanied by structural reforms
designed to give greater decision-making power to the
people. People can hardly accept their responsibilities to protect themselves without acquiring a certain
capacity to make decisions about the risks to which
they are exposed. In this regard, the acceptance of

emergency procedures requires a process of empowerment of the population. Otherwise, communication

will become a simple defensive measure to absolve the
institutions of the potential harm to humans that could
be produced by a technological risk emergency.

ACKNOWLEDGMENTS
The fusion related work reported in this paper
was supported by the European Fusion Development
Agreement (EFDA) (contract TW6-TRE-FESS-B1)
and UKs Economic & Social Research Council (award
RES-000-22-2664). Parts of this work, supported by
the European Communities under the contract of Association between EURATOM/CIEMAT, were carried
out within the framework of the European Fusion
Development Agreement. The views and opinions
expressed herein do not necessarily reflect those of the
European Commission. The research project on perception of petrochemical risks was supported by the
Spanish Ministry of Education and Science (SEJ200400892) and was carried out by the Universitat Rovira
i Virgili, the Universitat Autnoma de Barcelona and
the Universitat Pompeu Fabra.

REFERENCES
Bauer, M. 1995. Resistance to New Technology. Cambridge
University Press, Cambridge, UK.
Burger, J., Greenberg, M., Gochfeld, M., Shukla, S.,
Lowrie, K. and Keren, R. (2007) Factors influencing
acquisition of ecological and exposure information about
hazards and risks from contaminated sites. Environmental
Monitoring and Assessment, 137 (13), 413425.
Farr, J. and Fernndez Cavia, J. 2007. Comunicaci i
risk petroqumic a Tarragona. De les definicions a les
prctiques institucionals. Tarragona: Universitat Rovira i
Virgili.
Flynn, J. 2003. Nuclear stigma, in: Pidgeon, N. Kasperson, R.
and Slovic, P. (eds.): The Social Amplification of Risk.
Cambridge University Press, Cambridge, 326352.
Frewer, L.J. 1999. Risk perception, social trust, and public
participation into strategic decision-making: implications
for emerging technologies. Ambio, 28, 569574.
Frewer, L.J., Howard, C., Hedderley, D. and Shepherd, R.
1996. What determines trust in information about foodrelated risks? Underlying psychological constructs. Risk
Analysis, 16, 473486.
Horlick-Jones, T. 2005. Informal logics of risk: contingency and modes of practical reasoning. Journal of Risk
Research, 8 (3), 253272.
Horlick-Jones, T., Walls, J. and Kitzinger, J. (2007) Bricolage in action: learning about, making sense of, and
discussing issues about genetically modified crops and
food, Health, Risk & Society, 9(1), pp. 83103.
Irwin, A., Simmons, P. and Walker, G. 1999. Faulty environments and risk reasoning: the local understanding of

1306

industrial hazards. Environment and Planning A. Vol. 31

(7), 13111326.
Michael, M. (1996) Constructing Identities: The Social, The
Nonhuman and Change, Sage, London.
Pidgeon, N., Kasperson, R.E. and Slovic, P. 2003. The
Social Amplification of Risk. Cambridge University
Press, UK.
Prades, A., Horlick-Jones, T., Oltra, C. and Sol, R. 2007.
Social Perception of Large Research and Development
Programs. Madrid: Ciemat.
Prades, A. Horlick-Jones, T. Oltra, C. and Sol, R. 2008.
Lay perception of nuclear fusion: multiple modes of
understanding. Science and Public Policy (in press).
Starr, C. 1969. Social Benefit versus Techonological Risck.
Science, nm. 165: 1.2321.238.

Turner, G. and Wynne, B. 1992. Risk communication: a

literature review and some implications for biotechnology. A: Durant, J. (ed.) Biotechnology in Public. A Review
of Recent Research. London: Science Museum for the
European Federation of Biotechnology. pp. 109141.
Visschers, V.H.M., Meertens, R.M., Passchier, W.F. and
DeVries, N.K. 2007. How does the general public evaluate
risk information? The impact of associations with other
risks. Risk Analysis, 27 (3), 715727.
Wynne, B. 1991. Knowledges in Context, Science,
Technology & Human Values, 16 (1), 111121.
Wynne, B. 1996. May the Sheep Safely Graze? A Reflexive View of the Expert-Lay Knowledge Divide. In:
Lash, S., Szerszynski, B., Wynne, B. Risk, Environment
and Modernity. London: Sage. 4483.

1307

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Media coverage, imaginary of risks and technological organizations

Ferenc Fodor & Gilles Deleuze
Electricit de France R&D, Av. du Gnral de Gandle, Clamart, France

ABSTRACT: We propose to analyze risks inside and outside technological organizations, through their
representations in the social imaginary by the media and fictions. We try to show how they can be interesting
and useful objects of research, what they tell us in complement to risk assessment and expert communication.
After presenting our assumptions justifying how the study of fictions is interesting to analyse risks, we briefly
present the disciplines and the methods used for our analyses: semiotics, semantics and storytelling. We finish
with two examples. First, the imaginary representation of nuclear power plant: how it is perceived, what are its
representations, questions directly related with social acceptability of technological organizations; second, the
narrative interpretation of the collapse of Air Libert, a commercial air transportation company. We conclude
with benefits of the approaches.

INTRODUCTION

We present an exploratory research combining risk

analysis and sciences of languages to analyze risks
regarding large organizations, through their representations in the social imaginary of insiders and outsiders
of the organization, influenced by the media and
fictions, especially movies.

2
2.1

ASSUMPTIONS
Imaginary of risks outside of the organization

2.1.1 Media discourses

In the modern western societies, it is before all the
media that convey information and knowledge. Thus,
they become indicators of the social life. There is a
close connection between the news diffused by the
media and the formation and the structuring of the
public opinion. However, even if the media influences
the public opinions, they also try to refer to it, creating
complex situations to analyze.
There is also the question of the credibility of the
media, and particularly television. It is necessary to
understand that an image reflects often more a point
of view than the reality itself. We were accustomed for
a long time to think that it is only the verb that is able
to lie. However, it is understood better today that what
we see is not always true; consequently it is a semantic
illusion to say, I believe it because I see it.
On the discursive level, we observe also that the
language of the media supports dramatization. Catastrophic information multiplies there as soon as an event

gives the opportunity of it. Let us remember of the

medical and food crises or the media treatment of the
aviary influenza recently. However, analyzing media
discourses is important but not enough if we want to
understand the imaginary, the social representations
of risks.
2.1.2 Importance of fictions in the imaginary
of risks
We propose to take also into account fictions (movies,
novels) for risk analysis, because they can reveal elements that the discourses of experts, of victims or the
representation by the media, do not reveal.
More over, nowadays, the knowledge, the expertises become increasingly compartmentalized. The
specialist cannot detect everything, his knowledge is
related to only one specific field; the media then often
have a quite particular point of view with a large
tendency to dramatize; storytelling from victims are
sources of bias and excess due to the special situation
of the victim.
At the opposite, fiction presents the world in an
original and total way. It does not suffer from the hyperseparation of knowledge; it includes and exceeds the
points of view of the expert, of the media or of the
victim. When separated approaches can lead to dead
ends, fictions show events in another way, they show
reality differently, in a way more inductive than the
speech of expert, victims or the media.
Fiction is an imaginary construction of the world,
essential for human beings. First, fiction fulfils a therapeutic function. It is one of the reasons for which
people tell stories. Second, fiction learns to us essential things about the human being and the society but at

1309

the same time it ensures us on the imaginary character

of the story. Thus, fiction fulfils a substitute function: it replaces pleasant or terrifying events which we
should not inevitably live ourselves in the real world.
Another function of the fiction, particularly of the
fiction of catastrophe, is cathartic because it makes
possible to exteriorize, through a novel or a movie, the
anguishes, fears or the memory of traumatizing events.
At the same time, fiction learns to us much about ourselves and about the possible unfolding of an event.
The fiction makes possible to refer to reality but it is
not reality. Moreover, it is not limited to the intrigue
but it extends to the field of characters, values, and
ideologies. The evoked imaginary world overflows, in
this way, the story itself and it questions human reality
as a whole. Even if any story is only a fragment of the
world, it returns, in an implicit way, the world in its
totality. Thus any story reporting a particular industrial
accident, for example, poses in general the problem of
this type of accident. In other words, fiction, which
is the construction of an imaginary universe, exceeds
what it shows us. It enables us to represent and to interpret the world in an unusual, even surprising, way, to
think from incredible scenarios.

2.2

Imaginary of risks within the organisation

The observation of recent examples of failures, loss

of reputation and disparition of large organizations in
technological domains like energy, telecom, commercial air transportation, lead to a question [31],[27]:
How experimented and skilled managers, committed to act for the benefit of their organization and their
own reputation, make wrong decisions and/or commit
actions obviously suicidal?
There is an abundant litterature about managerial
crisis. It provides usually interesting experiences and
lists of causes of crisis and simple lines of action, like
for example [26]. However, prevention through interpretation of weak signals may be missing. We try to
prove here that the signals could be detectable from
the narrative production of the management teams.
We propose the study of fiction to analyse and
describe so called managerial crisis, major failures
of organization due to management failure or lack of
control. The assumption is that many crisis come from
the collision between a real situation, and a group of
decision makers living in a story built over time.
By the way, some managers build a story from their
cultural and social context. With time, the belief in the
story grows as success arise, and it becomes a success story. Crisis occurs when the story collides with
reality, with often the consequence that the only but
ultimate failure overcomes all previous successes. We
can find a comparable assumption in [28], and the serie
of factors.

2.2.1 No proceduralisation
There is an important difference in term of decision making between the operator of a technological
system, driven by procedures; and a manager of a technological organizations, acting on strategy, investment
decisions, risk management. . . . As the proceduralization of such managerial decisions is substancially
impossible; role playing, visions of the world, personal
stakes have a larger importance.
2.2.2 Heroisation
The hero is presented in [12], as an individual mandated to solve a problem or transform a situation and
acting at the same time from is own will, motivated
by its personal desires of power, wealth or accomplishment. That is exactly the way many management
manuals describe the leading decision manager, highlighting individualistic values, individualistic decision
making, concentration of power, charisma, instinctive fast decision made in action. Being an executive
manager is like to drive a race car. If you like speed and
sensations, no problem, otherwise you are lost! [29].
Some managers like to present themselves as wearers
of a heroic mission that monopolize their thoughts:
Increase profit? It is continuously in my mind! Even
during the weekend, when I go to theatre. But it does
not bother me, I am a do-er. That is my mission.
2.2.3 Mass production of fictions and imaginary
frameworks
There may be also the emergence of a generation of
managers grown up in an environment and a culture
feed by movies, video games, telling standardized
stories produced in large quantities by mass culture
production centers, supported by the standardization
of semantic techniques. For example, the studies
of Campbell [5],[7], are influential on many action
movies.
Furthermore, managerial techniques may be also
influential on the way organizations build their own
stories [27], as storytelling is often used to make sense
and consolidate organizations.
The whole lay contribute to a standardization of
imaginary framework used by managers to make
decisions in an uncertain world.
2.2.4 Virtualization of reality
The quote of A.C. Clarke1 : Any sufficiently advanced
technology is indistinguishable from magic, illustrate the power of information technologies, creating
cognitive bubbles capable to produce a so called
virtual reality, immerging the decision maker in a

1 Author

1310

of 2001 Space Odyssey.

reconstruction of its context, the more deceptive that

it is realistic.
Furthermore, the globalisation and the size of many
organizations, including geographical and cultural dispersion between decision centers, production plants,
markets may increase the risk to lose contact with the
real organization.
At last, the information and the terms of a decision, are often at highest levels of decisions of large
organizations, translated in financial terms, that may
be misleading or produce self predicting indicators,
in the sense that In technological systems, the confidence is built on measures, in financial systems, the
measure is based on confidence [30].
As an illustration, of loss of contact with reality, the
movie Grounding tells the collapse of the Swissair as
a thriller and reminds a movie on the last days of the
DDR Politburo that did not perceive reality anymore
and did not see that the state was collapsing under
him.2
2.2.5 Importance of heuristics for decision makers
The overflow of instant information, the practice,
sometimes addictive, of quick decision making under
multiple sollicitations, the positive value put on the
capability to make instant decisions require often to
rely on standardized heuristics (framework helping
analysis and decision making). For example, Henry
Mintzberg, a management teacher [30], observes that,
as the majority of managers activities take less than
five consecutive minutes and that most of their activities hare unscheduled, they make thus their decision
mostly on intuition.
Every day starts with disturbances in my planning,
determined by the emails received during the night3
Since two years, I spend an average of one extra
hour per day to manage my electronic mails
This factor combines with the loneliness due to
heroisation presented above
Eventually, there is not so much people with whom
I can have discussions
They are often in a great loneliness. We imagine
they are strong, but I see them as flimsy. Many live in
fear and are surrounded by consultants and coaches
to make them feel more confident4
At the end, there is a need for heuristics, that are
built from the experience, and also the culture, the
imaginary of the individual, in other words, stories.

2 Tages

Anzeiger, traduction by Courrier International,

no 803, Mars 2006.
3 Extraits dinterviews publis sous forme anonyme dans
Haute sphre, ton univers impitoyable, Enjeux Les
Echos, Janvier 2006, pp. 8891.
4 Jacques Bonnet, professeur de management lENSAD.

2.2.6 Necessity of storytelling

for some organizations
At last, we also observe that globalized companies,
despite their technological and financial power, may
suffer from a lack of imaginary roots in a population,
a tradition, a nation, a territory, a culture, especially in
the case of conglomerates, built by successive fusion
and acquisitions, or complex supply chains and network of activities spreaded in many countries, often
subcontracted. Their survival relies on volatile financial resources brought by investors as long as they feel
confident. Thus, there is a need to built an identity,
shared values to create confidence inside and outside
the organization, often from scratch.
Some management schools advise [27] to use storytelling techniques to build sense in such organizations.
The risk is that, as a result of telling stories
to improve the robustness of the organization, the
management may become the subject of its own stories, acting according to standardized and improper
imaginary frameworks.
3

DISCIPLINES AND METHODS USED

FOR OUR ANALYSES

We briefly present the disciplines and the methods used

for our analyses: semiotics, semantics and storytelling.
3.1 Semiotic and semantic approaches for studying
the imaginary of risks outside the organization
Semiology (or semiotics) is one of the scientific disciplines that make possible this type of analyses. We
distinguish two main analytic phases: the systemic
phase (how function the recurrent and the peripheral
elements in the corpus) and the interpretative phase.
Analyzing the differences between the manifest content and the latent content is one of the semioticians
most important tasks; it is interesting to analyze the
differences between these two levels in order to find
the links between explicit forms and implicit content.
The semiotician tries to discover how the meaning is
built, what effects of meaning can be found in the
implicit content, in the latent significances or in the
cultural unknowns [15].
It is essential to do an in-depth analysis of the convergences in order to highlight the recurring elements
in a given corpus, paying special attention to scenic,
textual and semantic devices and to their referents and
values. It is also vital to analyze the role of the reader
and the text within the interpretation process. One then
needs to know how many readings can exist, how the
meaning comes to the image, to what extent one must
consider the role of subject and subjectivity throughout this process. Images are perceived (and not simply
received) by the receiver.

1311

During the analysis, the semiotician does not just

read the documents; he frames and hierarchically
organises the traits of the images and words, whilst
at the same time relating them to one another and
interpreting them. Every visual perception, even when
examined in an inattentive or indifferent manner, is
interpreted in relation to other previous visual perceptions. This intertextuality falls within the domain
of both subjectivity and of a cultural community.
Analysis of the different strata is a vital stage in throwing light upon the universes of values relating to a
given corpus. Semiotics borrows linguistic methodology that derives analytical structures from language
(drawing on phonology, lexicology, semantics, syntax, etc.). The separate analysis of the strata (firstly
the iconicscenic, colouristic devices, etc.and then
the linguisticlexical, syntaxical, rhetorical, etc.) is
completed by taking the denotative and connotative
aspects of the question into account. The semiotic
approach allows one to bring out the way in which the
object in question works, by demonstrating the dominant practices, the convergences and the divergences
[17]. Semantics studies the significances of the words
and is the science of the linguistic significances only,
without taking into account the iconic aspects.
3.2

Study of the imaginary of risks inside

the organization

Under our assumption of an improper storytelling

leading to wrong decisions, two approaches may be
combined to identify such situations. The first is the
description of crises with the structural semantics
of Greimas [12] and Propp [24], useful to define
the actors and the oppositions and alliances, the second is the description of the sequence of the crisis,
including psychological dimensions, with scenario
building techniques from Campbell, supported by an
ancient dramaturgic typology from Gozzi et Polti
[33]. The whole idea came from a seminal paper of
Palmlund [25].

4
4.1

AN IMAGINARY OF RISK OUTSIDE

AN ORGANISATION (CIVIL NUCLEAR)
The systemic phase of analysis

In the representations of the nuclear power plants,

the first structuring element relates to the disproportion of these worrying, threatening constructions of
large size in a natural environment, often the edge
of the sea or rivers. The disproportionate aspect of
these buildings is reinforced by a second convergent element which relates to the incomprehensible
character of the nuclear power plant, the idea that it
escapes from the understanding: it is a quasi occult

power which can be used at the same time for the

wellbeing of the men, but also for their destruction.
The nuclear power plant must contain, control and
channel an enormous power, at the same time good
and bad. This force can flee or explode because of
a certain supposed brittleness of the nuclear power
plants structures that is the third convergent element in the corpus. The fourth recurring element
concerns the representation of the specialists, engineers and technicians, who work inside this mysterious installation and who have real capacities but
who seem also to be playing the sorcerers apprentices. The iconic representation of humans in tight
white suit reinforces the hazardous character of work
in a nuclear power plant (this clothing icon often
appearing at the time of an action of decontamination) and consequently it creates an effect of
hazard, of threat. Specialists operate a mysterious,
very powerful object. The eternal question of the
transgression and the punishment appears then: playing with fire, wanting to go too far, involves a
punishment.
4.2

The interpretative phase

Results of the interpretative phase of semiotic analysis

led us to the formulation of the idea that the nuclear
power plant makes the semiosis (unite) specifically
the anguish through the theme of the madness, found
especially in the corpus of the nuclear bomb5 . The
theme of nuclear awakes a subjective fear which would
be the fear of imaginary which exists in each of us and
that finds its expression in an indirect, involuntary way.
We fear the uncontrolled emergence of this imaginary
and what it contains, i.e. impulses and desires. Fears
of a cataclysm, end of the world appear very clearly
around the phantasm of the explosion, in the majority
of the analyzed fictions, which they are the military
or civil applications of the nuclear power. The first
many times repeated image of destroying mushroom
cloud remains engraved in the collective memory. This
image was magnificently exploited by Kubrick in the
Dr. Strangelove movie and it seems to represent the
unconscious original definition of the nuclear power:
in last analysis, it would like to mean that all that is
nuclear explodes.
The semiotisation of the anguish is also shown in
the fictions about the nuclear power plant that awake a
certain anxiety since they point out the existence of the
nuclear weapon, which we do not see, and the feelings
are thus repressed. In the case of the nuclear in general,

5 we dont present results of analyses of the nuclear bomb

but links between the bomb and the nuclear power plant
are very frequent.

1312

and the nuclear power plant in particular, the knowledge and the techniques are represented like complex
and unintelligible; the lifespan of the elements evokes
the infinite; the radioactive waste, a quasi eternal dirt,
is bad substances and we fear for their invasion; the
radiations are invisible, the dangers are conceived in
terms of collective. Moreover, the nuclear power plant
is secret and what appears secret is renowned seldom
innocent. All these elements contribute to create the
effect of meaning of the anguish that characterizes the
nuclear6 .
We can also affirm that it is not possible to separate the discourses on the nuclear power plant and the
atomic bomb. We did not find any element in our corpus of fictions of the nuclear power plant in which a
bond, even weak, with the bomb is not identifiable.
In imaginary, nuclear civil and atomic bomb form an
inseparable whole. From the elements highlighted, we
can make the assumption that the nuclear would be
represented, in the unconscious one, like a non acceptable creation, in the same way that certain desires are
prohibited because punishable. If these desires find
any expression, they generate a guilt feeling and thus
anguish. The nuclear power plants (and also the atomic
bomb) are connected unconsciously with the prohibited desire (to become very powerful, to be able to
control everything); they have a concrete existence and
thus distressing.
Consequently, we make the proposition that the
anguish of the nuclear, as well as the anguish of
the catastrophe and the end of planet, reflects the
mechanism that make people feeling guilty of their
desires.
4.3

The narrative structure of an accident

To complement the semiotic analysis, using the

Greimas framework [12], we can illustrate the typical
narrative structure of an accident in a nuclear power
plant, as it is told in fictions (Figure 1). The hero (the
subject), is the person who the first understood that
there is a real risk of accident. He wants to avoid the
catastrophe (the object of his quest). Opponents are
the managers for example of the nuclear power plants
who do not want to reveal problems. The supporters
(Helpers) are the journalists, some colleagues. This
is, of course, a schematic presentation of the narration
but we can evoke The China Syndrome movie (1979)
as an illustration.

Axis of the communication

Sender
Scenario Writers

meet the same phenomena in the media (newspapers, Internet . . .) with a strong tendency to dramatization
moderated for a certain time by the arguments related to
the climate change (non emission by the nuclear of gases
contributing to the greenhouse effect).

Receiver
Spectators
International Public

Axis of the quest

(of the desire)
Helper
A specialist, an engineer,
authorities (nuclear safety,
government), journalists

Objectof the quest

Objectof desire of the hero
To prevent a nuclear
catastrophe

Opponent
The company which
manages the nuclear power
plant
The management who
minimizes or conceals the
risk
ri

Axis of the fight

Figure 1. The narrative structure of a catastrophe in a

nuclear power plant in fictions.

AN EXAMPLE OF IMAGINARY OF RISK

INSIDE AN ORGANISATION

5.1 A brief history of Air Libert

Air Libert was a passenger air transport company,
funded in 1987 as a charter airline, transporting
tourists to holidays spots overseas. It became in 1994
a regular airline, in the follow up of the liberalization of the air transport in Europe. The company CEO
decides to take market shares to the former monopoly,
Air Inter, on its domestic lines, proceeding in 6 months
to the opening of 11 lines, investing around 500 MC
=,
starting a price war and aggressive advertising. In
1996, the former monopoly counterattacked. The price
war expanded. Some domestic lines became at that
time the cheapest in the world per kilometer. However
Air Libert was very quickly financially exhausted and
lost around 40 MC
= for a revenue of 114 MC
=. The company was eventually purchased by British Airways in
1997.
The question is then why did the CEO of Air Libert,
recognized as a smart and successful entrepreneur,
who achieved to build a beneficial and sustainable
company, launch a very fast and suicidal competition
against a powerful monopoly?
5.2 Structural Analysis of the managerial crisis7
Using the same Greimas framework than Figure 1, we
can present the story, as the CEO dreamed of it (see

7 We
6 We

Subject
The hero of the drama,
Person who announces the risk

do not pretend to provide here the only interpretation of an history. There are others, like for example the
ambivalent strategy of on of its main investors, the possible role of the government, objective managerial failures
on account reporting. . . . However they do not explain the
strategic failure.

1313

Sender

Receiver

Main Shareholders

Customers, media

Subject Company CEO

The company

Competitor
(the former
monopoly)

Employees, assets
(material and
immaterial)

Helpers

Object

Deus ex machina

Opponents

Analysts, air traffic authorities,

regulators

Figure 2.

The semantic structure, as seen by the CEO.

Figure 1). The hero, the subject, is the Air Libert

CEO. Its mission, or object of his quest, is the success
of Air Libert company. He has been requested by
its main investors (the Sender). The people benefiting
of the mission (Receivers) are customers, lay public,
small shareholders.
Opponents are groups of people (Competitors, trade
unions . . .) or events, physical obstacles (kerosene
costs, economic growth . . .) that will oppose to the
quest.
The hero will find support (Helpers) with the employees of Air Libert, material (aircrafts, fly allocations)
and immaterial assets (reputation), the media (supposed to support the outsiders against the monopoly)
At last a deus ex machina8 supposed to be neutral will eventually solve the litigation, distribute the
awards and sanctions: the regulators, and financial
analysts.
The crisis can than be explained according to this
structure: the CEO, shifts the object of its quest (the
success of the company) to an other (the benefit of
the customer and the victory against the monopoly),
maybe mislead by its own storytelling (see Figure 2),
focuses on his fight against competition, and neglects
its receiver, the investor. The axis of the communication and the axis of the fight eventually permuted. The
CEO will later justify his failure by the inappropriate
role played by some media that not acted as helpers
and a deus ex machina actor, playing as an opponent
(the regulation authority, accused to secretly support
the former monopoly) [23].
5.2.1 Sequential analysis of the managerial crisis
It complements the structural analysis, introducing a
sequential frame and a dramaturgic sequence.

8 We have introduced that notion in addition to the original

concept of Greimas, in accordance with Palmlund [25].

Figure 3. The story that Air Libert told: Picture from a TV

advertisement presenting the former monopoly in contempt
of its customers.

We can built it from the Campbell theoretical works

[32], and the dramatic situation no 24 from Carlo
Gozzi [33], fighting in asymmetric position, that
we could call The outsider against the (evil) empire,
a framework common in fictions (Star Wars, Robin
Wood, Lords of the Rings . . .).
At the origins is a successful outsider. In our case,
the CEO is a successful self made man, successful in
tourism, building a charter company, and searching
recognition in media.
The initiating event of the story. The 1st January
1995, the deregulation of aerial transportation offers
the opportunity for new markets.
The enemy appears. The opponent the former
monopoly, presented in the Air Libert communication as sad, arrogant, showing contempt for its
customers, fighting against the positive values of freedom and entrepreneurship. The company has to fight
against the opponent on its own ground, the French
domestic sky.
First fights, first successes. Air Libert is beneficial and gain passengers against the former monopoly.
In august 1995, a survey present Air Libert as the
favoured company by French customer in air business.
Passengers are presented as being fed up with the former monopoly, strikes, and a bad sense of business that
do not consider them as customers, just as users.
Doubts and difficulties. At the moment of such fictions, the hero faces first difficulties, and doubts. He
overcomes them and finds the allies and knowledge
necessary to win the final fight against the empire [24].
In this real story, the CEO has also to face difficulties,
but could not overcome them.
For example, a legal decision requests Air Libert to contract as full employees interim personal,
the monopoly manages to keep close many flight

1314

schedules. He thinks seriously to change its strategy

or resign, but goes on in the fight. He wrote later:
[My children] are engaged into the battle. The walls
of their rooms are covered by drawings representing
Air Libert airplanes fighting adverse dragons, Air
France and Air Inter. To resign would mean to flee in
full battle . . . I decide to continue the fight. I have
no regrets[23]. At that moment, all happens as if
the CEO believes in its storytelling. He is no more
engaged in an economic competition, but in a story of
fight against a dragon. Forgetting the rationale of
business, the CEO starts its a final fight for market
shares, suicidal because of a lack of support.
Empire strikes back. The competitors eventually
manage to adapt and win the war of prices. At the difference of the fictional story, the hero suffers at that
moment losses and treason; he has to resign in 1996
on the commitment of one of its main investors, who
is negotiating an alliance with British Airways.
The empire wins. That is the end of the real story,
the rebel company has to merge in a larger group to
survive. The CEO publishes a revenge book explaining
the treason from deus ex machina and helpers acting
as opponents, with a meaningful title: The forfeited
sky, as a sad testimony of a stolen dream.

CONCLUSION: BENEFITS
OF THE APPROACHES

Previous works to this article show the importance

of the imaginary dimension in the development of
some risks [13][14]. We propose and illustrate in this
article the benefits of language sciences to analyse
and understand it. The approaches presented could be
experimented to analyse discourses and media produced by movies and organizations to detect weak
signals and trends to prevent crises, internal or
external to an organization [34].
They also permit to analyse crisis and risk by
storytelling techniques, opening the opportunity to
build counter stories, including explicit and implicit
knowledge, eventually able to help decision makers
and stakeholders to have a reflexive attitude towards
their own stories.

REFERENCES
[1] Aumont, J., Marie, Michel, Lanalyse des films, Paris,
Nathan, 1988.
[2] Aumont, J., Bergala, Alain, Esthtique du film, Paris,
Nathan, 1988.
[3] Barthes, R., Mythologies, Paris, Le Seuil, 1957.
[4] Barthes, R., Elments de smiologie, Communications, 1964, 4, pp. 91134.

[5] Barthes, R., Rhtorique de limage, Communications, 1964, 4, pp. 4051.

[6] Barthes, R., Le Systme de la mode, Paris, Le Seuil,
1967.
[7] Eco, U., La structure absente, Milan, Paris, Mercure
de France, 1968.
[8] Eco, U., Le signe, Bruxelles, Labor, 1973.
[9] Eco, U., Lector in fabula, Paris, Grasset, 1979.
[10] Eco, Umberto, La production des signes, Paris, Le
Livre de PocheEssais, 1992.
[11] Floch J.-M., Identits visuelles, Paris, PUF, 1999.
[12] Greimas A.J., Smantique structurale, Paris,
Larousse, 1966.
[13] Fodor, F., Langoisse de lavenir de la plante et
sa smiologisation dans les fictions du XXe sicle, Actes du Colloque international Smiologie
2003De la diversit lunit du domaine, publi
sur CD, Paris, Sorbonne, 2004, pp. 93100.
[14] Fodor, F., La menace vient-elle de lEst? Analyse smiologique de lvolution des reprsentations du nuclaire militaire russe dans des
films occidentaux (19621997), dans Langue
CommunautSignification (d. Harald WEYDT),
Berne, Peter Lang, 2002, pp. 355361.
[15] Fodor, F., Sur la mthodologie et les applications
de la smiologie, dans La linguistique fonctionnelle au tournant du sicle, (d. Claude TATILON et
Alain BAUDOT), Toronto, Editions du GREF, 2002,
pp. 270275.
[16] Guedeney, C., Mendel, G., Langoisse atomique et les
centrales nuclaires. Contribution psychanalytique
et sociopsychanalytique dun phnomne collectif,
Paris, Payot, 1973.
[17] Houdebine, A.M., Actualit de la smiologie, dans
Entreprise et smiologie (dir. B. FRAENKEL, C.
LEGRIS-DESPORTES), Paris, Dunod, 1999.
[18] Labbe, Marie-Hlne, La grande peur du nuclaire,
Paris, Presses de Sciences Po, 2000.
[19] Le Guelte, G., Histoire de la menace nuclaire, Paris,
Hachette, 1997.
[20] Metz, C., Langage et cinma, Paris, Editions Albatros,
1977.
[21] Odin, R., Cinma et Production du sens, Paris, A.
Colin, 1991.
[22] Paris, H., Latome rouge, Le nuclaire sovitique et
russe, Paris, LHarmattan, 1996.
[23] Balhassine, L. Le ciel confisqules relations incestueuses dAir France et de lEtat- 1997, Albin Michel.
[24] Propp, V. Morphologie du conte populaire russe,
Seuil, 1965.
[25] palmlund, I., Social Drama and Risk Evaluation
in Social Theories of Risk Sheldon Krimsky and
Dominic Golding, Editors Praeger Publishers, 1992.
[26] Finkelstein, S., Quand les grands patrons se plantent,
Editions dOrganisation, 2004.
[27] Morel, C., Les dcisions absurdes, Livre de poche,
2004.
[28] Salmon, C., Storytelling, la machine fabriquer des
histoires et formater les esprits, La Dcouverte,
2007.
[29] Vauhan, D., Rosenthal, U., Turner, B., Pidgeon, N.,
Man-made disasters, Butterworth-Heinemann Ltd;
2Rev Ed edition (7 April 1997).

1315

[30] Extraits dinterviews publis sous forme anonyme

dans Haute sphre, ton univers impitoyable, Enjeux
Les Echos, Janvier 2006, pp. 8891.
[31] Nicolet JL., Enron, un Tchernobyl financier?
congrs Lambda Mu, Bourges, 2004.
[32] Campbell J., Les hros sont ternels, Paris, Seghers,
1987.

[33] Rey F., Les 36 axes dramatiques de Carlo Gozzi,

Synopsis, 2003.
[34] Deleuze G., Un cadre conceptuel pour la comparaison sret et scurit de filires industrielles, WISG
08, Universit de Technologie de Troyes, 29 January
2008.

1316

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Media disaster coverage over time: Methodological issues and results

M. Kuttschreuter & J.M. Gutteling
Center for Risk & Safety Perception (CRiSP), University of Twente, Enschede, The Netherlands

ABSTRACT: In 2000 disaster struck Enschede in The Netherlands. Due to explosions at a fireworks factory,
22 people were killed. This study aims to describe the developments in the media coverage of this disaster.
Content analysis was performed on 4928 articles, derived from four distinct newspapers. After a period of
intense coverage, media attention for the disaster declined. In this first month 772 articles were run (local paper
457, national papers 315). After three years this number is approximately 30 articles per month. This decline can
be best described by an inverse function. This paper focuses on the changes in the amount of news coverage, and
concentrates on some methodological issues researchers encounter when they try to analyse the news coverage
of a disaster over time.

INTRODUCTION

On May 13, 2000 at 15:30 in the afternoon several

explosions destroyed a fireworks plant located in a suburban area of Enschede, a city in the eastern part of the
Netherlands with 150.000 inhabitants. The explosions
and the fire that was started by the explosions killed
22 citizens, and wounded nearly 1.000 others. Over
4.000 people were made homeless, and the damages
were estimated at several hundred millions of Euros.
Within hours the magnitude of this event became clear
and hundreds of journalists and news crews from all
over the world poured into Enschede. The first footage
aired on television showed the severe consequences of
an industrial calamity that the country had never witnessed in recent history; the previously lively suburban
area closely resembled a war zone.
Industrial disasters with citizen fatalities are rare
events in countries as the Netherlands. Events like this
have a high level of newsworthiness, and thus typically
attract much media attention. Where the first accounts
in newspapers and television are typically event driven
and describe what has happened, the news framing
soon changes its focus as journalists start to ask questions like how could this happen?, Why was this
disaster not prevented?, and Who is to blame?. The
tone-of voice of the news accounts changes as well, at
first we primarily see sympathy and compassion for
the victims and the inhabitants of Enschede, but soon
both local and national authorities are questioned for
not being able to prevent this disaster. Such questions
were fed by the report of an Independent Investigation
Committee, which was installed to find out which factors contributed to the disaster (Committee Oosting).
After a while, media coverage fades away, as other

issues arise and win the battle for the medias and
publics attention.
In the coverage of a disaster one may expect differences between local and national newspapers. Newspapers have to write stories which are interesting to
their readers. Local newspaper can be expected to pay
more attention to the disaster than the national newspapers. Also, local newspapers can be expected to focus
on the interests of and the personal relevance for the
local readership, which may be different from those
of the readership of national newspapers. One might
expect more stories with a human interest frame or
economical consequences frame in the local press, and
more stories with an responsibility or conflict frame
in the national press (Anderson & Marhadour, 2007).
In this paper we analyze the media coverage of the
Enschede disaster in Dutch local and national newspapers for a three-year period after the disaster. We
focus on the changes in the amount of news coverage, and concentrate on some methodological issues
researchers encounter when they try to describe these
changes in a mathematical function. We are not aware
of any attempts to do so before. These issues are:
Does the coverage in the local newspaper resemble
that in the national newspapers and does a similar
mathematical function apply to both? Or does one
have to analyse the developments in media coverage
for local and national newspapers separately?
Which mathematical function fits the observed
number of articles per period the best?
To analyse the developments in coverage over
time, the total period under investigation should be
divided into smaller time segments. Does the length
of these smaller periods affect the results?

1317

 How does one distinguish periods in which coverage

temporarily intensifies? How can peaks in media
coverage be related to events in the aftermath of the
disaster?
Does the mathematical function apply equally well
to specific subsets of the articles, which make use
of a particular thematic frame?
The consequences of the various methodological
choices will be illustrated by applying them to the news
coverage of the Enschede fireworks disaster. This will
result in a description of the developments in the media
coverage of the Enschede disaster.

a way that implies that responsibility for the issue

is in the hand of some actor, e.g. government. The
economic-consequences frame is present in an article
when it discloses information or opinions about the
financial consequences of an event, issue or problem
for an individual, group, organization, region or country. Interrater reliability for the coding of the frames
(Cohens Kappa) is moderate.
2.3 Sample characteristics

A content analysis of four Dutch newspapers was performed over a three year period after the disaster.
Sampled were the local newspaper in the disaster area
(TcTubantia), two national newspapers (Volkskrant,
Telegraaf) and one national/regional newspaper (AD)
from one of the major industrial regions of the country. These newspapers were selected on the basis of
their region of delivery, size of readership, and political and economic point of view. For these newspapers,
all articles were selected from electronic databases
that were published in the three years after the disaster. Selection criteria were the occurrence of keywords
as vuurwerkramp (fireworks disaster) or Fireworks
(SE Fireworks is the name of the company that owned
the plant). Analysis indicated that these keywords
returned the largest number of hits in the electronic
databases.

In the three year period after the disaster, a total

of 4.928 articles were found. After exclusion of
announcements of radio and television programs, and
articles in which the disaster was mentioned but was
not part of the storyline, 3.927 articles remained as our
corpus of analysis (that is 986 articles or 20% were
excluded). The local newspaper from the disaster area
published considerably more articles than the other
newspapers (local n = 2.679, national n = 1.248).
In these articles, the responsibility frame was used
most frequently (64% of all articles uses this frame),
followed by the conflict frame (52%), the humaninterest frame (36%), and the economic consequences
frames (31%). There were significant differences in
frame use between the local and the national newspapers (multivariate F(4, 3937) = 110, p .0005): the
local newspaper more often made use of the humaninterest frame (F() = 230, p .0005), and the
economic consequences frame (F() = 222, p
.0005) and slightly more often of the responsibility
frame (F() = 10, p .001), whereas the national
papers more often made use of the conflict frame
(F() = 12, p .001).

2.2

2.4 Analysis

2
2.1

METHOD
Design and sample

Coding

The articles were identified by newspaper, publication data and title. The 3.927 articles were coded by
several students based on a trained coding instruction.
The articles were read and the coders coded whether
one or more of the following frames were present:
conflict frame, human-interest frame, responsibility
frame, and economic-consequencesframe (yes, no).
The choice of frames was based on previous work
by Semetko & Valkenburg (2000) and De Kort &
dHaenens (2005). The conflict frame focuses on conflicts between individuals, groups or organizations.
This may reflect in the description of disclosure of
conflicts of opinions, judicial procedures, criticizing
the other party or defending oneself against the critique of others. The human-interest frame presents the
human emotional aspects of an event, issue or problem. Typical are stories describing individual victims
and disaster relief workers, identified with their full
name. The responsibility frame presents an issue in

The analysed focused on the developments in the

media coverage over time. The analysed period of three
years has been divided into periods of equal length; in
particular 39 periods of 4 weeks. In this way, each
period consisted of an (almost) identical number of
similar days of the week, 4 Mondays, 4 Tuesdays etc.
Occasionally, newspapers did not publish issues, such
as on Christmas Day A timeline was created. Characteristics of the media coverage of the disaster were
added, as were events that took place in the aftermath
of the disaster. Six types of events were identified on
the timeline: (1) the aftermath of the disaster and crisis
intervention, (2) recovery and rebuilding of the disaster stricken area, (3) prevention of similar events in
the future, (4) judicial research and procedures against
individuals or organizations that could be held responsible for their role in the disaster, (5) discussion about
consequences for active politicians and authorities,
and (6) external events.

1318

3
3.1

RESULTS

Table 1. Fit between the observed number of articles and

various mathematical functions.

The news coverage over time

In the three years following the fireworks disaster,

3.927 in-focus articles appeared in the investigated
newspapers. At first, coverage was high. In the 4-week
period directly after the disaster 772 articles were run
(local paper n = 457, national newspapers n = 315).
A sharp decline followed, which later levelled off to
a gradual decline. Even after three years, the media
still continued to cover the disaster. Three years after
the disaster, coverage amounted to approximately 30
articles per month, which is an article per day on
average.
The local paper (n = 2.679) published more articles on the fireworks disaster than the three national
papers combined (n = 1.248). This difference holds
for each 4-week period in the 2,5 years following the
disaster.
There also was a significant difference between the
local newspaper and the national newspapers in the
distribution of articles over time (chi-square = 215,
df = 38, p < .0005). The national newspapers published more articles in the period directly following
the disaster (25% of the articles appeared in period 1)
and in the period in which the results of the Independent Investigation Committee were published (7%
appeared in period 11) than the local paper, that covered the disaster more smoothly (17% and 5% of the
articles appeared in periods 1 and 11 respectively).
This means the decline in media coverage of the disaster after the first period was steeper for the national
newspapers than for the local newspapers, and that
the coverage of the national newspapers may be more
affected by events in the aftermath of the disaster, such
as the publication of the report of the independent
investigation Committee.
This leads to the conclusion that there are probably
substantial differences between the local and national
newspapers in the mathematical functions describing
the amount of media coverage over time. The coverage
in the local and national newspapers should therefore
be analysed separately.

3.2

Mathematical functions describing

the developments in media coverage

Data analysis has taken place by means of the SPSS12. This package allows one to investigate to what
extent observed data fit with a large number of hypothetical, mathematical function (R 2 ). The higher R 2 ,
the better the fit. Estimates for the regression weights
and a constant term are also given.
Table 1 shows the fits between the observed number
of articles and all available mathematical functions.

R square (R 2 )
Function

Local (n = 2.679)

National (n = 1.248)

Inverse
Logistic
Logarithmic
Cubic
Exponential
Growth
Compound
Quadratic
Power
S

.90
.76
.76
.71
.66
.66
.66
.62
.58
.32

.78
.47
.47
.43
.41
.41
.41
.37
.50
.39

Based

on 39 periods of 4 weeks.

The best fit between the observed number of articles and the hypothetical, mathematical function was
found for the inverse function. This function provided a very good fit for both the local newspaper
(R 2 = .90, standardized beta .95), and the national
newspapers (R 2 = .78, standardized beta .88). For
the local newspaper, the full mathematical function
describing the expected number of articles is E(Nt ) is
E(Nt ) = 22 + 426/t, where t is the number of the
four-week-period after the disaster. Similarly, for the
national newspapers, the full mathematical function
is E(Nt ) = 4 + 255/t. These formulas predict that
the local newspaper started out paying much more
attention to the disaster than the national newspapers
together (local E(N1 ) = 444, national E(N1 ) = 259)
and that the coverage in the local newspaper eventually (E(Nevent. ) = 22) remains higher than that in the
national newspapers (E(Nevent. ) = 4).
The inverse function described the developments
in the disaster coverage in the local newspaper (.90)
better than it did the coverage in the national newspaper
(.78). The question arose whether the higher R 2 for the
local newspaper could be attributed to the different
sample sizes or to other factors such as differences in
publication policy. To answer this question, a random
sample of n = 1.248 was drawn from the articles in
the local newspaper. The fit with the inverse function
for this reduced sample was R 2 = .88, which is hardly
any lower than for the full sample. The lower fit for
the national newspapers can thus not be explained by
a lower number of published articles.
3.3 Duration of the time periods
The analyses up till now have been based on the number of articles published in 39 periods of four weeks.
A relevant question was, whether the results were
affected by the chosen duration of the time-periods.

1319

Table 2. Fit between the number of articles and the inverse

function, for periods of various duration.
R square (R 2 )

1 week
2 weeks
4 weeks

Local (n = 2.679)

National (n = 1.248)

.75
.88
.90

.78
.76
.78

500
400
300
200
7

100
0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

Local newspaper

Four-weeks-periods

National newspapers

Figure 1. Number of articles in the local and national

newspapers over time.

To investigate this, the analyses were executed again

for the number of articles published in periods of two
weeks and one week respectively (see Table 2). The
results show that, for the national papers, there is not
much difference between the fits for periods based on
four, two or one week; R 2 varies from .76 to .78. For
the local newspaper, all fits were high, but the fit for
the one-week period (R 2 ) seemed to be less than the
fits for the two- and four-week periods. The reasons
for this are not clear. May be the local paper reacted
promptly to developments in the aftermath of the disaster, and compensated this extra attention by paying
less attention to the disaster aftermath in the following week, while the national newspapers reacted with
more restraint.
3.4 Identification of significant peaks in coverage
Figure 1 showed that the decline in media coverage was
not monotonous and that there were fluctuations. It is
tempting to attribute these fluctuations to events in the
aftermath of the disaster (see Figure 2). However, some
of the fluctuations might be random. This raised the
question if it were possible to determine whether fluctuations are random, or whether they could be labelled
significant peaks in media coverage.
A manner to decide whether the number of published articles deviates significantly from the expected

number based on the hypothesized function, would be

to inspect the confidence intervals. If the observed
number in a specific period falls within the confidence limits, one concludes the peak is a random
fluctuation. If the observed number falls outside the
confidence interval, one concludes there is a genuine
and significant peak in media coverage (see Figure 3).
There is a problem with this procedure. The estimation of the parameters in the inverse function
(regression coefficients and constants) depends on the
coverage in all periods, including the period in which
coverage may be excessively high. This means that
the expected value for the period suspected to show a
peak in coverage, is too high. As a result, there seems
to be a tendency for significant peaks to be considered
insignificant.
It is not clear how one should deal with this problem.
Theoretically, one might argue that the inverse function should be estimated based on all points, minus
the ones suspected of peaking. This should lead to
an corrected estimate, which can be used to decide
whether the observed number of articles is higher than
the upper-limit of the corrected estimate. In order to do
so, however, one first has to identify which points in
time may peak. This strategy thus seems to run in circles. This best correction we came up with, was using
the 90%-confidence interval instead of the usual 95%
interval, thus slightly lowering the critical upper-limit.
Employing the 90%-confidence interval, in the coverage of the local newspaper three four-week-periods
can be distinguished in which coverage peaked. These
were the three periods following the publication of
the report of the Independent Investigation Committee (periods 11, 12 and 13). Coverage in the national
newspapers peaked in two four-week-periods, namely
the period directly following the disaster and the period
following the presentation of the report by the Independent Investigation Committee (period 11). The results
thus showed that only a few of the peaks in media
coverage, suggested by Figure 2, were statistically
significant.
3.5

Mathematical functions describing the

developments in the use of specific frames

The question arose whether the inverse function would

also provide the best fit for the number of articles in
which a specific thematic frame was used. This proved
to be the case, for both the local and the national
newspapers. However, the fit between the number of
articles with a specific frame and the inverse function
is less than that for the number of articles in general
(see Table 3). The fit was best for the human-interest
frame, followed by the economic consequences frame.
The fits for the conflict and responsibility frame were
found to be lower. A major contributor to this fact
seemed to be the extensive disaster coverage with a

1320

500
Explosion destructs residential
area, 22 death, 1000 injured.
Start independent investigation

450

Board set free. Council of

Ministers promises 100
million euro's.

400
350

Suspect arrested.
Fireworks decision in
Council of Ministers
Report independent
investigation published

Attacks September 11,

Explosion Toulouse

Debate municipal local

board Enschede.
Alderman resigns.

Internal criticism
investigation suspect.

300
Cafe-fire
Volendam

250
200

Debate parliament.
Authorities not
prosecuted.

Start trial board members

Fireworks. Enschede decides to
move high risk companies.

Digital risk map

announced. Heavy
fireworks banned

150

Board
convicted.

Suspect
convicted.

Suspect released
after appeal.

100
50
0
1 2

3 4 5 6 7
2000

8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
2001
2002
2003

Local newspaper

Figure 2.

four-week-periods

National newspapers

Events in the aftermath of the Fireworks disaster and number of articles in the local and national newspapers.

Table 3. Fit between the number of articles and the inverse

function, for various thematic frames.

900
800

R square (R 2 )

700
600

Local

National

.90 n = 2.679
.94 n = 1.163

.78 n = 1.248
.70 n = 238

.91 n = 1.131

.62 n = 195

.75 n = 1.617
.71 n = 1.329

.53 n = 747
.47 n = 688

500

All in-focus articles

Human interest
frame
Economic consequences
frame
Responsibility frame
Conflict frame

400
300
200
100
0

9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

Based

-100
Observed number
Lower limit

Expected number Four-weeks-periods

Upper limit

Figure 3. Expected number of articles per period, with 90%confidence levels.

conflict and responsibility frame in the period following the publication of the report of the Independent
Investigation Committee. This suggests the disaster
coverage with a conflict and responsibility frame to
be more event driven than the disaster coverage with a
human interest and an economic consequences frame.
Again, the fit was better for the local newspaper
than for the national newspapers. It was observed that
there was a larger increase in articles with the conflict and responsibility frame in the period following

on 39 periods of 4 weeks.

the report of the Independent Investigation Committee

for the national papers than for the local paper. This
suggests the disaster coverage of the national newspaper was more event driven than the coverage in the
local newspaper.

DISCUSSION

Disasters typically attract much media coverage. After

a while, other issues with a high newsworthiness
emerge, and the media attention for the disaster fades.
No attempts have been made as yet to describe this
decline in media coverage by a mathematical function. This paper discusses some of the methodological

1321

issues we encountered, when we to do this for the

media coverage of the Enschede Fireworks disaster.
Other researchers who focus on the developments in
the media coverage of some other incident or disaster,
will to face similar issues.
The media coverage in the local and three national
newspaper was studied. The first question relates to
the duration of the total period to be analysed. While
it is a lot of work to code and analyse a long period
of time, our research showed that major changes in
news coverage were observed following the publication of the report of the Independent Investigation
Committee, nine months after the disaster; three years
after the disaster still about one article per day was
published. Researchers should thus guard themselves
against analysing too short a period.
The investigation showed that it might be important
to distinguish between local and national newspapers.
For both the local and the national newspaper, at first,
coverage was high. A sharp decline followed, which
later levelled off to a gradual decline. However, the
decline in media coverage of the disaster after the
first period was steeper for the national newspapers
than for the local newspapers. This suggested that
the coverage in the national newspapers may be more
affected by events in the aftermath of the disaster, such
as the publication of the report of the Independent
Investigation Committee. There were thus similarities and differences in coverage between the local
and the national newspapers. Researchers should thus
ask themselves whether the coverage in the local and
national newspapers should be studied separately.
An inverse function provided the best fit with the
observed number of articles, for both the local and
the national newspaper. However, the formulas for
the local and national newspapers were different. The
local newspaper started at a far higher level of coverage than the national newspapers, and three years
after the disaster, the local newspaper still showed a
higher level of coverage. The local disaster news coverage over time was thus similar, but not identical in
magnitude to the national disaster news coverage.
Additionally, it was found that the inverse function provided a better fit for the number of articles
published over the three years in the local newspaper (R 2 = .90), than in the national newspaper (R 2 =
.78). This difference could not be accounted for by differences in sample size. Perhaps this difference in fit
may be explained by differences in publication policy.
Support for this interpretation is found in the observation that there were large differences between the local
and national newspapers in the use of specific frames.
Newspapers have to publish articles that interest their
readers. The local paper focused its coverage on the
relevance of the disaster for the local residents; it made
use of an human-interest frame and an economic consequences frame more often than the national paper.

The national paper, on the other hand, emphasised the

national relevance of the disaster, and published more
articles with a conflict frame.
The events in the aftermath of the disaster, such as
the publication of the report of the Independent Investigation Committee, relate to the role of governmental
agencies, the Board of the Fireworks factory and a
potential arsonist in bringing about the disaster. Such
events stimulate stories with a conflict or responsibility frame, rather than a human-interest frame. The
national papers thus seemed to have reacted more
strongly to the events in the aftermath of the disaster, while the local newspaper covered the disaster
aftermath in a less volatile manner. Such differences
in coverage policy would result in the observed finding that the number of articles in the local newspaper
showed a closer fit to the inverse function than the
number of articles in the national newspapers.
Another question which relates to the fit with the
inverse function, concerns the thematic framing of the
articles. It was found that the inverse function also provide the best fit for the number of articles in which a
specific thematic frames was used, for both the local
and the national newspapers. However, the fit between
the number of articles with a specific frame and the
inverse function is less than that for the number of
articles in general. This may be due to the fact that
events in the aftermath of the disaster evoked articles
with specific frames, such as the conflict or responsibility frame. Researchers should realise, that different
theoretical models may apply to the news coverage of
an event than two news coverage of issues related to
that event. Whereas the model for the former might
be a monotonous decrease in attention over time, the
developments in framed news coverage might be better
explained by models such as the issue-attention model
(Downs, 1991).
An important question in studying media overage
over time relates to the duration of the periods. One
argument could be the fit with the inverse function. For
our investigation, the results showed that all fits were
high and that there was not much difference between
the fits for periods with a duration of four, two or one
week; for the national papers R 2 varied between .76
to .78, for the local paper between .75 and .90. For the
local newspaper, the fit for the one-week period (R2)
seemed to be less than the fits for the two- en fourweek periods. The reasons for this are not clear. May
be the local paper reacted promptly to developments
in the aftermath of the disaster and compensated this
extra attention by paying less attention to the disaster
aftermath in the following week, while the national
newspapers react less instantaneous. A two- or four
week period thus seems to be a proper choice. Choosing for a two-week period has the advantage that the
regression coefficient can be based upon two periods
with high coverage directly following the disaster, in

1322

stead of one, as was the case in our study. At the same

time, however, the global picture might become too
detailed to comprehend.
It was found that the decline in media coverage was
not monotonous and that there were fluctuations. It
is tempting to attribute such fluctuations to events in
the aftermath of the disaster. However, some fluctuations might be random, and researchers should be
aware. Inspection of the confidence intervals for the
estimated number of articles might give an indication. If the observed number of articles falls within
the confidence limits, one concludes the peak is a
random fluctuation. If the observed number falls outside the confidence interval, one concludes there is a
genuine and significant peak in media coverage. One
should bear in mind, however, that, though the procedure is theoretically correct, in practice it has some
methodological caveats. For the Enschede disaster,
many peaks were visually identified, but only the peak
in media coverage in the periods following the publication of the report of the Independent Investigation
Committee was found to be significant. So, things are
not always what they look.
Insight into media coverage processes is important,
because news coverage influences public opinion and
risk perception. Studying media coverage will help

to predict which influence and how much influence

is being exerted on the media consumers. This paper
has focused on the Enschede fireworks disaster and
has discussed some methodological issues and their
solution. Many questions still remain, in particular
related to the generalisability of the results to other
disasters. More research is clearly needed
REFERENCES
Anderson, A. & Marhadour, A. 2007. Slick PR? the media
politics of the prestige oil spill. Science Communication,
29(1): 96115.
Downs, A. 1991. Up and down with ecology: the issueattention cycle. Agenda setting: readings on media,
public opinion, and policymaking. In D.L.M. Protess &
M.E. McCombs (eds.). Hillsdale, New Jersey, Lawrence
Erlbaum Associates: 2733.
Kort, K. de & DHaenens, L. 2005. Berichtgeving over
Pim Fortuyn en extreem-rechts gedachtegoed in twee
Nederlandse kranten vergeleken. Tijdschrift voor Communicatiewetenschap, 33(2): 127145.
Semetko, H.A. & Valkenburg, P.M. 2000. Framing European
politics: A content analysis of press and television news.
Journal of Communication, 50(2): 93109.

1323

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk amplification and zoonosis

D.G. Duckett & J.S. Busby
Department of Management Science, Lancaster University, UK

ABSTRACT: This qualitative study presents a novel perspective on risk amplification, applying it to zoonosis.
We introduce the concept of folk epidemiology to trace a narrative dimension to understandings about zoonotic
disease. We see this narrative influence as an element that is bound-up with action around actual cases of zoonosis
drawing on sensemaking approaches. We adopt a wider perspective on communication theory drawing
on communion based readings beyond the narrow transmission metaphors that limit existing approaches.
The intention is to reanimate the actor from passive receiver of transmitted risk messages into someone who
concurrently shapes and is shaped by risk. This social action is described in terms of actors using accounts
from a reservoir of dominant narratives and selecting from a repertoire of available actions. Actors perform
these operations in the context of a specific case of zoonosis. Focus groups of lay publics and expert groups
provide our empirical data. Instances of narrative and action combination centered on cases are presented. This
empirical work aims to demonstrate the underlying structure of this folk epidemiology; a three element structure
of narrative, action and case that allows novel insights into how we humans understand the risks from animal
diseases that infect us.

INTRODUCTION

Zoonosesdiseases that cross species barriers from

animals to humanspresent pressing issues to our
understanding of risk perception. In the case of Avian
Influenza or bird flu, 2007 witnessed a high degree
of public concern in Britain immediately following
an outbreak of the H5N1 strain at a farm in Suffolk. This high degree of concern rapidly subsided,
at least from the headlines and main news items. The
Social Amplification if Risk Framework or SARF
(Kasperson, Renn et al. 1988) aims to account for
these phenomena of amplification and, attenuation.
(The mechanics of SARF are detailed in section 1.1).
In addition, zoonoses may be something about which
expert and lay judgements of risk diverge. Our study
focuses on these two areas, namely amplification
and lay/expert differences in perception and presents
novel concepts of folk understanding to extend the
social amplification framework.
1.1

The Social Amplification of Risk Framework

Probabilistic accounts do not capture the full complexity of risk. Human judgements about risk often
have profoundly consequential outcomes that seem
unrelated to the magnitude of the possible harm multiplied by the estimated frequency of the event. The
Social Amplification of Risk Framework (SARF),

(Kasperson, Renn et al. 1988) represents an attempt

to map some of this complexity.
The central insight from (SARF) can be summarized as the conceptualization of risk perceptions as
either being amplified or attenuated by social factors. This modulation is situated in social stations of
amplification. These stations are variously described
as large units comparable to, what mainstream sociology would call institutions (Renn, Burns et al. 1992),
or at the level of the individual, for example, scientists
and opinion formers (Kasperson, Renn et al. 1988).
We find this understanding both plausible and useful
and attempt to build upon it by proposing reservoirs
of narratives and repertoires of available actions that
coincide during particular cases of risk amplification.
We present this layer as a folk theory of risk in which
narratives, stories, and metaphors, seen collectively as
shared understandings, are bound up with risk related
behaviours.
SARF also addresses the theme of differences
between lay and expert judgements about risk. Risk
managers may be confounded by lay reactions to risks
that appear disproportionate to risk levels derived
from technical or expert assessments of risk. Risk
researchers have a long tradition of investigation into
this expert versus lay phenomenon, for example,
myriad studies into the siting of hazardous waste and
the proposed nuclear waste repository at Yucca Valley (Pijawka and Mushkatel 1991; Slovic, Layman

1325

et al. 1991; Slovic 1992; Slovic 1993; Easterling 1997;

Flynn 2003). This study tackles the lay/expert gap by
exploring the narratives deployed by each group and
the actions available to them.
1.2

Zoonosis

It is estimated that 60% of all known human diseases

are zoonotic (Bennet 2007). They are diverse in form
and effect:
They can be transmitted directly by contact with
an animal (e.g., rabies, through a bite), via a
contaminated environment (e.g., anthrax) and
via food (e.g., campylobacteriosis) or indirectly
via vectors, such as mosquitoes or ticks (e.g.,
West Nile fever and Lyme disease, respectively).
The organisms causing zoonoses include viruses,
bacteria, fungi, protozoa and other parasites,
with both domestic and wild animals acting as
reservoirs for these pathogens. The diseases
they cause in humans range from mild and selflimiting (e.g., most cases of toxoplasmosis) to
fatal (e.g., Ebola haemorrhagic fever).
(http://www.nwph.net/NWZG/importance.aspx)
Highly pathogenic avian influenza H5N1 belongs to
the severe variety of zoonotic diseases. It is not easily
transmissible from bird to human and only rarely from
human to human, however, when contracted, it has
killed up to 50% of human victims (Farndon 2005).
This high mortality has raised fears that if the virus
mutates to form a strain that has greater human to
human virulence, it will cause a human pandemic
on a scale last witnessed in 1918 when the Great
Influenza killed millions of people across the world
(Kolata 2000). Significantly this 20th century avian flu
(H1N1), devastating as it was, was not as pathogenic
as the current threat claiming a mere 2.5% of those
infected.
Other zoonotic viruses have infected human populations including SARS and while none have been as
devastating as 1918, the potential for a pandemic to
threaten millions of lives remains a serious concern to
epidemiologists.
Pandemic is not the only perceived risk connected
with bird flu or other zoonotic diseases and we are
equally interested in economic issues such as loss of
sales and potential collapse of markets. BSE typifies
this phenomenon.
The BSE story has been one of the most dramatic
risk crises in Britain in the last twenty years. It
has been an extremely long running and complex
saga with a definite risk narrative (Eldridge and
Reilly 2003:141).

The wide ranging social and economic effects

that a zoonosis can trigger have spurred considerable
literature which we draw on where relevant (Carey
1989; Eldridge and Reilly 2003; Campbell and Currie
2006; Horlick-Jones, Walls et al. 2007).1 We contribute to this literature with some contemporary cases
of zoonoses.
1.3 Objectives of this research
The objective of the study is to extend SARF in the
context of zoonoses and to reinterpret SARF where
necessary, through an empirical study. This empirical
work is centred on understandings of zoonosis and its
risks among both expert and lay groups.
We further address a long recognised structural
weakness in SARF: That it is rests on a narrow
communications theory metaphor based of signals
and receivers (Rayner 1988). This narrow view of
communication as transmission ignores the view of
communication as ritual, which draws on the etymological associations of communication with communion, community and commonality (Carey 1989).
We aim to broaden this view thereby overcoming the
limitations of the existing framework.

LIMITS OF SARF

2.1 Communications theory provenance

The structure of the framework depicting risk information communication is based on a particular conception of communication. Kasperson describes this
intellectual provenance as derived from classical
communications theory and he lists a number of texts
in his bibliography on communication (Kasperson,
Renn et al. 1988). Amongst these few citations is the
work of Hovland (1948), which takes us back to the
early days of studies into mass communication. Drawing a transmission orientated definition the author
outlines the origin of theory in WWII propaganda and
the burgeoning field of psychologically driven product
advertisement.
Amongst this sparse literature the approach that
holds most striking similarities to the Kaspersons
amplification is Lasswells communication address
(1964). Lasswell develops a model around the notion
of conductance.
Whoever performs a relay function can be
examined in relation to input and output. What

1 Horlick-Jones

(2007) concerns GM crops not zoonoses

however, its treatment of risk communication to the lay
public is highly relevant to our study.

1326

statements are brought to the attention of the relay

link? What does he pass on verbatim? What does
he drop out? What does he rework? What does he
add? How do differences in input and output correlate with culture and personality? By answering such questions it is possible to weigh the
various factors in conductance, no conductance
and modified conductance. (ibid:48)
Present above are several of the SARF elements
including recognition of both culture and personality
integrated into a mechanistic account of a signal being
modulated by social structure.
While this classical tradition allows us certain
insights into the formulation of SARF, we must recognize limitations bound up with adopting this heritage.
These limitations become apparent when one contrasts the transmission view with another prominent
interpretation of communication that exists in the
literature, namely communication as communion
(Carey 1989).
2.2 Communication: Transmission versus
communion
Firstly, the transmission view, bound up in religious
origins but increasingly secularized with scientific,
economic and technological contexts, relates to the
movement or transportation of ideasideas moving in
space, from the word of God through western expansionism to the telegraph and mass communication.
Following the Second World War, that witnessed the
scientification of propaganda and saw the emergence
of modern marketing and the blossoming of the electronic age, this definition has achieved considerable
prevalence.
Transmission is a goal directed perspective. The
goal of the exercise is defined by the party initiating
the process. This might be to convert the heathen, to
sell the product, or in our case, to manage the perception of risk. The carefully crafted message should
be preserved intact throughout the process (Lasswell
1964). The objective is for the receiver to receive the
message as the originating transmitter intends it to be
received. It is a top-down hierarchical view of communication. Kaspersons signal/receiver metaphor allies
strongly with this definition.
The second definition of communication is revealed
in an etymological kinship with communion and
community. It concerns ritual communication and
emphasizes sharing, participation, association, fellowship, or the possession of a common faith (one can
think of excommunication here as an exclusion from
a community). Commonness is another relevant
root. In this perspective understanding is an emergent social process. The linearity of transmission
is replaced by reflexivity. This latter understanding

lends itself favourably to ideas about folk understanding and ethnoscience (Atran and Medin 1999;
Coley, Medin et al. 1999; Rip 2006) that are presented here as complimentary to SARF. Communion
encompasses more fully cultural perspectives of risk
perception (Douglas and Wildavsky 1982; Douglas
1992). We see this second perspective as a fertile area
for our research given that is not well represented
currently in SARF. This analytical distinction offer
some prospects for the overarching framework that
SARF aspires to, avoiding the dangers of making a
monopoly out of transmission to the exclusion of the
communal perspectives outlined above. As Morgan
notes; Favored metaphors tend to trap us in specific modes of action. (Morgan 1997: 350). Here, the
trap is that the signal/receiver metaphor excludes the
alternative, etymologically based understandings of
communication as communal exercises, shared experiences, rituals, and, what we want to propose, folk
understandings. None of these concepts are contradictory to the cross disciplinary content of SARF
that overtly embraces Cultural Theory and wants to
develop social structures as key factors, but they are
outside the problematic central metaphor.
The electronic imagery is too passive to cope
with the complexity of human risk behaviour.
. . . . The metaphor reduces the richness of human
risk behaviour addressed by the subjectivist
paradigms to the mere noise it represents to the
technical risk assessor. (Rayner 1988:202)

AN ORGANIZING FRAMEWORK

We aim to develop an organizing framework that

addresses the limitations of SARF but is broadly meant
to achieve the same endthat is, not to provide a
predictive theory but to organize thinking about the
growth and dissipation of risk issues. The first aim in
developing this new framework is to represent individuals and organizations as actors in risk issues, not
merely perceivers of risk. Therefore, the development
of their intentions is an explicit part of the framework.
The second aim is to capture the idea that folk theorising is strongly influenced by the actions available
to folk actors (Rip 2006). This also reflects ideas of
sensemaking (Weick 2001) in which actors reasoning
follows action as much as it precedes it, and is heavily oriented toward rationalising rather than planning.
And it reflects ideas of situated cognition (Suchman
1987) in which peoples plans are not so much control
structures that precede actions as discursive resources
produced and used within activity.
The third aim of the framework is to posit the role of
narratives as being central to the way in which actors

1327

engage with issues. These narratives are not necessarily well structured, integrated or strongly worked out.
But they pre-date actors encounters with an issue, and
may well be modified as a consequence. The important
quality of narratives in this context is that they make
less commitment to a particular source of influences
on risk understandings than is found in psychometric risk (for example Slovic 1987) and work on the
cultural selection of risk, for example, Schwarz and
Thompson (1990). They do not have to be psychological resources within the individual, nor interpretive
resources within a particular culture, nor the explanations provided by some authority, but can be any of
these, or some mixture. Their source is less important than their availability at a given time to a given
actor.
The final aim is to represent the idea that the
particular case or situation is often instrumental in
explaining peoples interpretations. It is not just an
actors available choices, nor the available fragments
of narrative, but also the structure of the case that
matters. For example, in the UK aviation influenza
outbreak at the Bernard Matthews plant the outbreak
happened to occur in an intensive production facility.
Avian influenza was thus confounded with intensive food production. It need not have been. Other

outbreaks occurred in the wild bird population, or in

less intensively farmed flocks that had been infected
by wild birds. Thus the available narratives and available choices become activated and bound in the context
of a specific situation.
The resulting framework is represented in figure 1
showing that actors have a reservoir of narrative fragments available to them, and a repertoire of potential
actions.
Particularly relevant combinations are activated
in complex interactions and become bound together
in the context of a specific case. This leads people to enact particular choices. Actors are connected with each other in this process, particularly
in the way in which they can observe each others
intentions.
Figure 2 represents this for a member of the lay
public in the Bernard Matthews case.
We further propose a higher level structure reflecting the grouping of actors in particular cases. Certain
groups, through their homogeneity as groups, will
strongly share common reservoirs of narrative fragment and repertoires of available actions. And they
will have influences on other groups intentions that
are stereotypical for that group. This is indicated in
Figure 3.

specific actor s response processes

specific actors response processes

reservoir of
dominant
narrative
fragments

dominant
narrative
fragments

unsustainable mass
production, ecological
paradigm, industrygovernment conspiracy, happy animals,
sensational media, rational science

repertoire
of available
enactable
choices

available
enactable
choices

factory
farming is
unnatural

case binding narrative and

action

Boycott processed
turkey meat

state has to compensate producers loss

others boycotting turkey meat

observations of other actors

risk-related intentions
General framework for an actors response.

meat product
boycott

Bernard Matthews avian flu

outbreak

risk-related intentions

Figure 1.

consumption
choices, travel decisions, affiliation
decisions, dietary
choices, protest actions, biosecurity
precautions

Figure 2.

1328

Example of the framework.

Exotics include avian influenza and rabies; endemics

include Salmonella and E Coli. The expert group interviews were topical with current events influencing the
agenda.

regulatory agencies
broadcast media

4.2

lay publics
advocacy groups

industrial firms

Figure 3.

4
4.1

Higher level structure.

METHODS
Data collection

Since we are working within a loose theoretical framework, not a predictive theory, the empirical investigation was naturally an exercise in qualitative, inductive
analysis. Our aim is to elaborate on the framework we
have proposed, developing it for the case of zoonotic
risk both generally, and for specific cases of zoonosis.
The research design involved collecting rich textual
data about people making sense of zoonotic cases.
This data was to consist of lay focus groups and expert
groups.
The lay focus groups were intended to involve individuals who had no particular reason to have expert,
analytical or deep knowledge of zoonosis. The expert
groups were intended to involve individuals who were
closely involved in farming, food processing and regulation, and who had privileged practical or analytical
knowledge of zoonoses.
In both cases the data was collected in connection
with particular outbreaks of a zoonosis. The focus
groups were based on both exotic zoonosis and
endemic ones. These distinctions are deployed by
Defra (the UK Department for Environment, Food and
Rural Affairs) and denote whether a disease is established inside the UK or is originating from abroad.

Focus groups

Focus groups are extremely well suited to the exploration of shared understandings (Kreuger 1994).
Within a focus group participants can engage with
one another to retrieve narratives through sharing partial accounts and through developing collective, emergent accounts including elements of narrative, story
and metaphor.
Focus groups were designed to incorporate a strong
element of homogeneity in order to promote sharing
of opinions. The ideal size was determined at between
5 and 8 to discourage fragmentation, or groups within
groups and to increase opportunities to share insights.
The first group was constituted of PhD students
from management related disciplines. They were
shown a brief presentation based around headlines
from the BBC website related to the Suffolk outbreak of bird flu and subsequent stories dealing with
bird flu from February 2007 to February 2008. The
presentation was intended to focus thoughts on the
groups understanding of this particular zoonosis but
was carefully crafted to convey little detail about specific actors, and to avoid judgments about the risk.
It was designed to include dates, locations, numbers of birds involved and consequences in terms of
movement restrictions and control measures.
A second focus group comprised of veterinarian PhD students all with a different animal disease
specialism.
A third group was made up of retired lay people
with no particular knowledge of zoonosis.
Further groups are still in development.
4.3

Expert groups

Expert groups were necessary to this research design

to investigate the proposed expert/lay gap in understanding (Kraus, Malmfors et al. 1991). We were also
sensitive to the idea of problematizing expert understandings in their own right (Wynne 1996) rather than
simply accepting them as privileged forms of objective
knowledge.
The main expert groups relevant to the zoonosis
cases we used were regulators, farmers and processors,
academic experts, special interest groups (for example
the Royal Society for the Protection of Birds). At the
time of writing the groups covered have been regulators
both at Defra and the Food Standards Agency (FSA),
and one group of academic experts. Some groups
were assembled at the offices of Defra in Whitehall
where discussions were steered into areas of interest

1329

by the researchers and notes were taken. The attendees

had a range of responsibilities in relation to zoonotic
diseases including epidemiologists and veterinarians.
Other groups of experts were assembled at the National
Centre for Zoonosis in the North West of England
also including epidemiologists and vets. These latter
subjects were more familiar with research participation
and for this reason audio recording was made.
4.4

Analysis procedure

The analysis broadly followed a grounded approach

(for example Glaser and Strauss 1967), but had to
use convenience sampling in places, rather than theoretical sampling (ibid:59). Based on the framework
shown earlier it involved looking for narrative fragments and potential choices revealed as informants
discussed zoonotic cases. The nature of the narratives
and choices were then examined, following the idea
that the fundamental operation of analysis is to discover classes of things and the properties of these
classes (Schatzman and Strauss 1973: 110), and the
specification of differences among and within classes,
as well as similarities (Strauss and Corbin 1990).
The informants utterances were also examined in an
attempt to work out how narratives and choices came
together to produce intentions. The rationale for qualitative analysis of this kind can be found in various
texts (for example Walker 1985; Berg 1989; Bryman
and Burgess 1994; Mason 1996).
At this formative stage in the analysis we wish to
present some provisional findings.
4.5

A reservoir of narratives

The lay narratives that emerged were not particularly

about zoonosis as a concept or as a general category.
Notably, they did reason about the actors who they
think are responsible for specific cases and activities
related to those cases. What we found therefore, was
that reflections about zoonotic diseases appeared to be
case specific. The context of particular cases appeared
to coincide with certain widely shared narratives. A
clear example was the coincidence of the outbreak in
Suffolk at an intensive factory farm with generally
shared narratives that were critical towards this means
of production. In this case the focus group associated intensive production and high density stocking
of animals indoors with outbreaks of bird flu.
[Extract from focus group 1]it is a naturally
occurring thing between birds but the impact of
intensive farming due to mass consumption and
the need for this intensive farmingis that it has
causedand as I say the regulations between
birds being brought in from Asia or from Europe
where there is a higher proportion of it . . . and

the restrictions that we have brokenor not particularly been in placeand add that to mass
farming and turkeysyou then get a copy ofthe
consequence is then that you get bird flu in this
country. . .
Importantly we acknowledge that this was an inconsistent association. On the one hand there was recognition that free range outdoor rearing brought poultry
into contact with wild birds. This was accepted as a
risk factor. At the same time there was a alternative
argument made by the same respondents, that intensive stocking was a breeding ground for disease or
(as above) actually the cause of the disease. This type
of inconsistency, here that the disease is caused by
intensive farming and that it is at the same time naturally occurring in the wild, is a typical feature of
folk theory.
They [Folk Theories] are a form of expectations,
based in some experience, but not necessarily systematically checked. Their robustness depends from their
being generally accepted, and thus part of a repertoire
current in a group or in our culture more generally.
(Rip 2006:349)
The Defra expert group recognized the tension
between risk factors of intensive production and risk
factors from outdoor rearing yet were quite reticent about stating which practice had a higher risk
profile. With avian influenza, they argued, overhead netting and location away from wetland habitats
could offer good protection to outdoor reared birds.
With indoor rearing biosecurity should concentrate on
rodent control and facility cleaning between stocking.
This even handed assessment was perhaps in part
due to the experts responsibility to supervise all parts
of the industry. However, one expert respondent did
acknowledge that a gleaming, sterile indoor facility
would be the best precaution against the spread of
MRSA in pigs but that the public would judge such
a facility as the unacceptable face of factory farming preferring a naturalistic and consequently dirty
environment for pigs.
A clear conspiracy narrative was also voiced by
focus groups. When considering official response in
terms of both control measures and communications,
the focus groups again expressed a somewhat contradictory view that on the one hand the authorities had
done what they had to do in regards to the culling of
birds and the imposition of restrictions. On the other
hand, there was a shared feeling that the authorities
may well conceal information from the public and
expose the public to risks and that we would never
know.

1330

[Extract from focus group 1]Its a bit like

natural disasterI supposemaybe potentially

bird flu is something in our controlbut say for

example the Tsunamipeople had been monitoring thatthey knew or had an idea that there was
a potential problem that was going to happen yet
they didnt evacuate anybody.
[Extract from focus group 1]Maybe its just
all related to whether theres money involved. . . if
theres an outbreak of avian flu and we all stopped
buying chicken then thats a knock on effect to the
economy so maybe until they do know the full
things theyre not going to tell us. . .
If the public believe that real risks may well be
concealed, it is sensible for them to adopt extreme
precautionary behaviour and avoid products rather
than adopting more modest food preparation precautions. This conspiracy theory emerging from focus
groups may reinforce the strength of stigmatization
and avoidance behaviour discussed below.
The veterinarians also voiced this conspiracy narrative:
[Extract from focus group 2] Its quite scary how
much we rely on the media I think to tell us about
whats happening. . . Especially if its Defra (that)
has some sort of control over it or the government
has some sort of control over it to some extent you
know, surely that filters out certain things that
we should hear about. . . its all a bit conspiracy
theory isnt it but Im sure it goes on.
In summary, narratives containing a concern for
naturalistichusbandryasopposedtointensiveproduction,
and narratives suggesting conspiracy at institutional
levels, were both present in the empirical data.
4.6 A repertoire of available action
Avoidance was the primary course of action that participants suggested as responses to risk perceptions of
zoonoses. This was often bound up with stigmatization. Stigmatization of products is a key process of
amplification in SARF (Kasperson, Kasperson et al.
2003) and in the supporting literature (Slovic, Layman
et al. 1991). The lay groups were quick to stigmatize
diseased animals and when they did so there was a clear
association with the action strategy of avoidance. They
expressed disgust or a yuck factor (Rip 2006) towards
sick animals and associated intentions connected with
avoidancewouldnt go near it wouldnt touch it
wouldnt buy it. With poultry there was a more proactive strategy of avoidance that linked into the free
range versus intensive debate. Contributors expressed
preferences for happy meat choosing to disregard
risks surrounding free range rearing.
With beef and BSE there was a feeling that, by the
time the risk message was clear, it was already too late

to effectively avoid consumption if one had already

eaten the tainted product over a number of years.
Expert opinion was more circumspect. In the case
of bird flu from poultry, Salmonella, and most food
borne zoonoses, risk was considered negligible if
proper cooking temperatures were observed therefore
stigmatization and corresponding product avoidance
was thought to be an over reaction.
Avoidance could comprise simple brand switching.
Focus Group 1 advocated switching brands of chocolate in the face of the Cadburys salmonella incident.
However, avoidance might be evidenced in a more
blanket move away from a generic product such as
beef (CJD) or, eggs (salmonella).
5

DISCUSSION AND CONCLUSION

A number of anthropological studies into folk biology consider primitive taxonomic categories as cases
of folk understanding (for example Diamond and
Bishop 1999). We considered to what extent zoonosis was a meaningful category by encouraging the
focus groups to identify zoonotic diseases. If there
were a general understanding of zoonosis it may
be significant. However, the lay groups studied did
not appear to have a particular category of animal
diseases that cross the species barrier and could identify only rabies, mad cow disease and bird flu
readily. Other zoonotic diseases were recalled only
after further prompting indicating that the category
was not salient. Our contention is that lay groups
have no sophisticated taxonomic understanding about
zoonoses as a single category. By this we do not simply
mean that the word zoonosis is not generally recognized. We mean that the category is not meaningful.
The expert groups corroborated the story that zoonosis
is not generally understood as a category by the public. On several occasions experts lamented the fact that
whilst diseases crossing the species barrier are highly
significant in epidemiological terms, the public did
not recognize them as a single category.
This is, we propose, at least partly indicative of
a complex social landscape in which modern, urban
populations are distanced from livestock farming in
many instances. Experts regularly raised concerns
about this disjuncture between farm and fork. Understandings arise when the public is confronted with an
outbreak. At the point of confrontation they draw on
a reservoir of narratives and a repertoire of available
courses of action to construct their understanding. Initial findings suggest that wider narratives that already
have social currency, for example critiques of intensive production as unnatural or breeding grounds
for a variety of ills, may be drawn upon when needed in
order to understand a current event or case. As Kasperson expresses it Reference to a highly appreciated

1331

social value may increase the receivers tolerance for

weak evidence (Kasperson, Renn et al. 1988:181). In
the case of bird flu the evidence does not clearly
attribute higher risk to intensive production facilities yet weak evidence against intensive production
facilities was, at times, readily accepted by lay focus
groups. This pull from the narrative side is symmetrically matched by a pull from the action side.
Avoidance strategies of brand or product switching
are often readily available. Rip recognizes this when
he asserts, projections linked to an action perspective are more important than empirical support for the
claims (2006, p. 351).
Findings about avoidance strategies from the focus
groups supported Rips (2006) idea of folk understanding in which action is bound up with understanding.
Weick refers to this boundedness as The placement
of enactment near the beginning of sense making . . .
(2001:176).
We follow Atran and others (Atran and Medin
1999) but look at a new subset of folkology2 , namely
folk epidemiology. This study attempts to explain
behaviours by uncovering the understandings through
which people form their views of zoonoses.
What the findings indicate is that the general framework we have proposed can be used to organize
sensemaking about zoonosis risk in a way that gives
new insight.
Furthermore, the passivity of the SARF actors as
receivers of goal directed messages is rejected in
favour of dynamic social actors participating in the
selection and sharing of narratives and programmes of
action in response to cases. We propose this reanimation of the actor in a communion based view of communication as an improvement to the transmission
view present in SARF. The actors we spoke to did not
receive their understandings about zoonoses from any
one source be it science or the media. They constructed
it together, as members of groups, out of diverse components. These components can be modeled in terms
of a reservoir of narratives and a repertoire of available actions that have an essential relationship to
their particular, concrete circumstances (Suchman
1987:70).
The next stages of the work are to 1) continue elaborating on the narratives and available actions that actors
locate and deploy, and 2) get from the actor-by-actor
model to a population model, i.e. a more epidemiological one. It is at this level that we can talk about the

2 We coin the neologism folkology to conveniently refer

to the myriad folk science studies from folk neurology to

folk biology.

growth and decline of an issue, in a similar way to that

in which SARF refers to the amplification of risk.

REFERENCES
Atran, S. and D.L. Medin, Eds. (1999). Folkbiology. Cambridge, Mass., MIT Press.
Atran, S. and D.L. Medin. Cambridge, Mass., MIT Press.
Bennet, M. (2007). Personal communication at a meeting
of the project team at the National Zoonosis Centre.
D. Duckett. Liverpool.
Berg, B.L. (1989). Qualitative research methods for the social
sciences. Boston, Allyn and Bacon.
Bryman, A. and R.G. Burgess (1994). Analyzing qualitative
data. London; New York, Routledge.
Campbell, S. and G. Currie (2006). Against Beck: In
Defence of Risk Analysis. Philosophy of the Social
Sciences 36(2): 149172.
Carey, J.W. (1989). Communication as culture: essays on
media and society. Boston, Unwin Hyman.
Coley, J.D., D.L. Medin, et al. (1999). Inductive Reasoning
in Folkbiological Thought. Folkbiology.
Diamond, J. and K.D. Bishop (1999). Ethno-ornithology of
the ketengban People, Indonesian New Guinea. Folkbiology. S. Atran and D.L. Medin. Cambridge, Mass., MIT
Press.
Douglas, M. (1992). Risk and blame: essays in cultural
theory. London; New York, Routledge.
Douglas, M. and A. Wildavsky (1982). Risk and culture:
an essay on the selection of technological and environmental dangers. Berkeley, Calif., University of California
Press.
Easterling, D. (1997). The vulnerability of the Nevada visitor economy to a repository at Yucca Mounta. Risk
Analysis 17(5): 635647.
Eldridge, J. and J. Reilly (2003). Risk and relativity: BSE
and the British media. The Social Amplification of Risk.
N. Pidgeon, R.E. Kasaperson and P. Slovic. Cambridge,
Cambridge University.
Farndon, J. (2005). Bird flu: everything you need to know.
Thriplow, Icon Books.
Flynn, J. (2003). Nuclear stigma. The Social Amplification of Risk. N. Pidgeon, R.E. Kasaperson and P. Slovic.
Cambridge, Cambridge University: 326352.
Glaser, B.G. and A.L. Strauss (1967). The Discovery of
Grounded Theory. New York, Aldine.
Horlick-Jones, T., J. Walls, et al. (2007). The GM Debate:
Risk, Politics and Public Engagement. London, Routledge.
Hovland, C.I. (1948). Social Communication. Proceedings
of the American Philosophical Society 92(5): 371375.
Kasperson, J.X., R.E. Kasperson, et al. (2003). The social
amplification of risk: assessing fifteen years of research
and theory. The Social Amplification of Risk.
Kasperson, R.E., O. Renn, et al. (1988). The Social Amplification of Risk A Conceptual Framework. Risk Analysis
8(2): 177187.
Kolata, G.B. (2000). Flu: the story of the great influenza
pandemic of 1918 and the search for the virus that caused
it, Macmillan.

1332

Kraus, N., T. Malmfors, et al. (1991). Intuitive Toxicology:

Expert and Lay Judgments of Risks. Risk Analysis 12(2):
215232.
Kreuger, R.A. (1994). Focus Groups: A Practical Guide for
Applied Research. London, Sage.
Lasswell, H.D. (1964). The Structure and Function of Communication in Society. The Communication of Ideas.
L. Bryson. New York, Cooper Square.
Mason, J. (1996). Qualitative researching. London, Sage.
Morgan, G. (1997). Images of organization. Thousand Oaks,
Calif., Sage Publications.
Pidgeon, N. R.E. Kasaperson and P. Slovic. Cambridge,
Cambridge University.
Pijawka, K.D. and A.H. Mushkatel (1991). Public Opposition To The Siting Of The High-Level Nuclear Waste
Repository: The Importance Of Trust. Policy Studies
Review 10(4): 180194.
Rayner, S. (1988). Muddling Through Metaphors to Maturity: A Commentary on Kasperson et al. The Social
Amplification of Risk1. Risk Analysis 8(2): 201204.
Renn, O., W.J. Burns, et al. (1992). The Social Amplification of RiskTheoretical Foundations and Empirical
Applications. Journal of Social Issues 48(4): 137160.
Rip, A. (2006). Folk Theories of Nanotechnologists.
Science as Culture 15(4): 349365.
Schatzman, L. and A.L. Strauss (1973). Field research;
strategies for a natural sociology. Englewood Cliffs, N.J.,
Prentice-Hall.
Schwarz, M. and M. Thompson (1990). Divided we stand: redefining politics, technology and social choice. Philadelphia, Pa., Univ. of Pennsylvania Press.

Slovic, P. (1987). Perception of Risk. Science 236(4799):

280285.
Slovic, P. (1992). Perception of Risk: Reflections on the
Psychometric Paradigm. Social theories of risk. S. Krimsky and D. Golding. Westport, Conn., Praeger.
Slovic, P. (1993). Perceived Risk, Trust, and Democracy.
Risk Analysis 13(6): 675682.
Slovic, P., M. Layman, et al. (1991). Perceived Risk,
Stigma, and Potential Economic Impacts of a High-Level
Nuclear Waste Repository in Nevada. Risk Analysis
11(4): 683696.
Strauss, A.L. and J. Corbin (1990). Basics of qualitative
research: grounded theory procedures and techniques.
Newbury Park, CA, Sage.
Suchman, L.A. (1987). Human-Machine Reconfigurations:
Plans and situated actions: the problem of human-machine
communication. Cambridge Cambridgeshire; New York,
Cambridge University Press.
Walker, R. (1985). Applied qualitative research. Aldershot,
Hants, England; Brookfield, Vt., Gower.
Weick, K.E. (2001). Making Sense of the Organisation.
Oxford, Blackwell.
Wynne, B. (1996). May the Sheep Safely Graze? A Reflexive View of the Expert-Lay Knowledge Divide. Risk,
Environment & Modernity: Towards a New Ecology.
S. Lash, B. Szerszynski and B. Wynne. London, Sage
Publications.

1333

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk communication and addressing uncertainties in risk

assessmentsPresentation of a framework
John Stian strem
Risk Management Solutions, DNV Energy, Det Norske Veritas, Norway

Hvard Thevik
Risk Management Solutions, DNV Energy, Det Norske Veritas, Norway
If P&C, Norway

Roger Flage & Terje Aven

University of Stavanger, Norway

ABSTRACT: The purpose of this paper is to present a framework for simple and effective treatment of
uncertainties in risk assessments. In the framework, it is emphasized how uncertainties can be accounted for
when using the risk assessment results, whether it be to identify and select risk reducing measures or to express
a companys overall risk picture. The use of the framework is illustrated by applying it in a typical offshore risk
assessment case. The new approach is compared with the approach presently employed within the oil & gas
industry and benefits are highlighted. Uncertainties are only associated with physical, observable quantities.
A key element of the framework is the use of sensitivity assessment. This is a well-proven, intuitive, and simple
way of illustrating how the results of a risk assessment depend on the input quantities and parameters. Here,
sensitivity assessment is revisited in order to show how the decision-making process can be enhanced when
accounting for uncertainties. The sensitivity, combined with an evaluation of the background material, forms
the foundation of the framework for presentation of uncertainty presented in this paper. Finally, it is discussed
how the uncertainties should be handled in the decision-making process. Often, uncertainties can be large, and
comparing (point) estimates from the traditional risk assessments with numerical risk acceptance criteria is
not recommended. Rather, possible risk reducing measures should be assessed based on the ALARP principle.
A measure should be implemented as long as the cost is not in gross disproportion to the benefits gained.

INTRODUCTION

The purpose of the quantitative risk assessment (QRA)

is to give a comprehensive, informative and balanced
picture of the risk in order to support decisions; for
example on acceptability of risk and implementation
of risk reducing measures according to the As Low As
Reasonably Practicable (ALARP) principle. A traditional offshore QRA typically presents personnel risk
and impairment frequencies for a selection of safety
functions; for instance escape ways, main support
structure, temporary refuge etc. (Vinnem 2007).
Traditionally, indices like the Fatal Accident Rate
(FAR), Individual Risk (IR) and Potential Loss of Life
(PLL) are used to describe and communicate personnel risk. The Fatal Accident Rate is defined as the
expected number of fatalities per 108 exposed hours.
Individual Risk and Potential Loss of Life is defined
as the probability of death for a randomly selected person and the expected number of fatalities, respectively,

during a specified period of time (typically a year).

For safety functions, an impairment frequency (i.e.
an expected value) is calculated. An impairment frequency is the annual frequency for the event that a
safety function is unable to uphold its function as a
result of an accidental load.
Offshore operators on the Norwegian continental shelf are required to establish acceptance criteria
(tolerability limits) for major accident risk and environmental risk (The Management Regulations, 2002).
Some typical risk acceptance criteria are:
The average FAR value should not exceed 10.
Accidental loads resulting in loss of a safety function should have an annual frequency less than
1.0E-4 (The Facilities Regulations, 2002).
It is a common belief that there are large uncertainties inherent in risk assessments. In structural
engineering it is normal practice to allow for uncertainties through the use of safety factors. This has not

1335

been the normal practice in the context of quantitative

risk assessmentsinstead a best estimate approach
partly based on conservative assumptions is used,
where uncertainties are discussed, often only in a general and qualitative way, with no specific relations to
the results of the assessment. Sometimes uncertainty
is not mentioned at all.
There is a tendency to place a strong focus on the
calculated results and compare these to the defined
risk acceptance criteria. If the results are lower than the
defined acceptance criteria, the conclusion is often that
the risk is acceptablethe solution is OK, and there
is no need for further actions. No idea of the robustness of the calculated results has been conveyed. If
the acceptance criterion is a frequency of 1.0E-4 per
year, and the risk assessment presents a frequency of
9.0E-5 per year, does this mean that everything is all
right? The number is close to the acceptance criterion, and according to the ALARP principle measures
should be considered that reduce the risk to a level that
is as low as reasonably practicable. Nevertheless, it is
commonly understood that as long as the acceptance
criterion is not exceeded, everything is OK, even if the
calculated risk results are close to the defined acceptance criteria. This is also a strong indication of how
disproportionate the weight is on meeting the acceptance criteria, rather than implementing an ALARP
process.
Would this practise be the case if the user of the
QRA had a better understanding of the factors contributing to the risk results, as well as the analysts
judgement of the uncertainty factors? It is believed
that a better communication of the risk results and
uncertainties gives the user of the assessment an
improved understanding of the risk. This would in turn
strengthen the decision processes and ensure a proper
focus in the treatment of risk.
The presentation of risk should be simple and effective. It should be easy to understand for the user of the
assessment. In addition it should not be too time consuming to perform for the analyst. This applies to the
existing approach for presentation of the risk results,
as well as for any new approach. In the former case,
one should not have to expand or modify the existing
risk methods and procedures too much. The approach
should be based on the already existing assessment
and tools.
This paper presents a suggestion for how to communicate risk results and uncertainty, based on a new
assessment framework. This framework holds uncertainty, not probability, as the main component of
risk. Probabilities are regarded as epistemic-based
expressions of uncertainty. This means that probabilities are regarded as expressions of uncertainty,
based on some background knowledge. As they are
expressions of uncertainty, the probabilities are not
themselves uncertain. They must however always be

seen in relation to their basis, i.e. the background

knowledge. To evaluate the background knowledge we
have to question to what extent for instance established
databases and methodology are used, or if the assessments are based solely on engineering judgement.
Another key issue is the competence/background of
the analysts.
Taking into account that the background knowledge can hide uncertainties, a description of risk in
QRA should cover more than the traditional probabilities and expected values. As uncertainty is the main
component of risk, and probabilities are only tools
for expressing uncertainty, the uncertainty factors in
the background knowledge need to be highlighted and
their effect on risk and vulnerability assessed. We refer
to Aven (2008a, 2008b) for the theoretical foundation
of this framework.
It is believed that this way of presenting risk can
contribute to a better understanding of the uncertainties relating to the driving risk factors. By gaining
such knowledge, the user gets a better understanding
of the results of the risk assessment. Hopefully, such
an understanding will provide better decision support.
A case study has been used in the development of
the presented framework. The case is a QRA of an
offshore wellhead platform.

ASSESSMENT FRAMEWORK

Risk in this framework is characterised by the following components:

A: Events (e.g. initiating events, barrier failures)
C: Consequences
U: Uncertainties
P: Probabilities and expected values
S: Sensitivity
K: Background knowledge
See Figure 1. A complete risk picture covers all
these components. Traditionally the P component has
been dominating.
In the framework, uncertainty factors are assessed
and presented qualitatively. A qualitative approach

K
U

Figure 1.

1336

Risk description.

A, C

is believed to be more informative and accessible to

the user of the QRA. In addition, quantitative results
such as probability distributions are not in practice
possible to establish for all the uncertainty factors of
interest. As mentioned earlier, we have to see beyond
the probabilities and expected values.
The assessment of uncertainty factors should
preferably be presented together with the summarised
results of the QRA.
During the risk assessment the analyst should identify uncertainty factors. This can potentially be many
factors, and the analyst should concentrate on factors
that, based on experience, contribute the most to the
overall risk picture. Examples include the occurrence
of a process leak, ignition following a process leak etc.
The importance can be verified by running a series of
sensitivity assessment within the QRA tool, to determine the influence of each factor on the results. This
would typically be increasing a parameter by a factor
and see the influence on the overall risk picture. Focus
should also be on factors that are judged to have considerable degree of uncertainty, these factors should
be targeted for a vulnerability evaluation.
The next step would involve the analysts judgement of the level of uncertainty of the factor. What
is the confidence in the background knowledge? The
background knowledge includes the data used, the
tools used to analyse physical phenomena etc. The
combined judgement of these two factors, sensitivity
and uncertainty, will constitute the analysts judgement of the effect of the uncertainty factor on the
risk.
The uncertainty factors could be presented in a format like Table 1. As described above, the uncertainty
factors are first ranked according to their sensitivity on
the risk results. Then the confidence in the background
knowledge (level of uncertainty) is ranked. These two
rankings will determine the effect of the uncertainty
factor on the risk.
We need to define what we mean by Minor, Moderate and Significant. For instance, the definition
of Minor in relation to Degree of Uncertainty
could be:
Strong support in much relevant data (information/knowledge). Strong agreement/consensus on
values/methodology.
Definition of Moderate in relation to Degree of
Uncertainty:
Some data (information/knowledge) to support
chosen values. Some agreement/consensus on values/methodology.
Definition of Significant in relation to Degree of
Uncertainty:
Little or no data (information/knowledge) to
support the chosen values (engineering judgement).
There is no agreement/consensus on the values and/or
methodology.

For factors that are ranked high in the effect on

risk there should preferably be a more in-depth discussion on why it has received this ranking. For instance,
what is the sensitivity of the risk result, and why is
the confidence in the background low for this particular factor? This can be done relatively easily on all
existing QRAs without having to modify the already
established assessments.
An aspect of risk that is not normally communicated
is the vulnerability, i.e. consequences and associated
uncertainties given an initiating unwanted event. This
is considered to be useful information, since the risk
results can be hidden by a low frequency of the
initiating event. Even if the initiating event frequency
is low, the event could happen the following week. It
would therefore be of relevance to get insights about
the vulnerability of different accidental scenarios as
judged by the analyst.
By presenting personnel vulnerability, for instance
the number of fatalities given that an unwanted event
has occurred, one gets an additional dimension of
risk normally not presented. This dimension gets
lost when the risk is presented as a function of a
scenario frequency. Vulnerability becomes especially
important if the underlying uncertainties are large concerning the number of events occurring. In such cases,
ALARP evaluations could be appropriate, even if the
calculated risk figures are not necessarily high.
Table 2 presents an example of a vulnerability
assessment from the case studya wellhead platform.
The figures present the expected number of fatalities
given an ignited process leakage. The text and figures
have been slightly edited for reasons of anonymity.
We predict eight and six fatalities in the maintenance and well intervention groups, respectively, given
an ignited process leak. This is done by considering
the fraction of leaks that are small, medium and large.
For instance a small leak in area 1 will constitute 42%
of the total leak frequency.
The information needed to establish this vulnerability is normally available in all risk assessment results.
This is because it is information used in calculating
traditional risk indices such as FAR and IR. The information is however normally not available as output of
the assessment.
Ideally, we should report more than the expected
value, as this value could produce a poor prediction
of the actual outcome. A 90% prediction interval [a,
b] could be derived, meaning that that there is a 90%
probability that the outcome of C will be in the interval [a, b] given the background knowledge and the
event A. However, in many cases the traditional assessment would not easily provide the information required
to support such calculations. One of the reasons for
this is that the traditional QRA often consists of best
estimates of the end event consequences in event
trees. To establish prediction intervals, distributions

1337

Table 1.

Example of uncertainty factors and their effect on personnel risk.

Uncertainty factor

Sensitivity of risk results

Degree of uncertainty

Effect on overall personnel risk

Minor

Minor

Minor

1. No. Personnel
in each area
2. Number of well
operations
3. Ignition following
riser leak
4. Occurrence of
process leak
5. Duration of
process leak

Moderate

Significant

Leak
size

Process area 1

Small
Medium
Large

Process area 2

Small
Medium
Large

Group

E[C|A]

Maintenance
Well intervention
Maintenance
Well intervention
Maintenance
Well intervention
Maintenance
Well intervention
Maintenance
Well intervention
Maintenance
Well intervention

6,6
4,2
17,6
16,0
17,0
14,2
3,4
1,4
7,2
5,2
7,2
6,0

DISCUSSION

When applied to the case, which was an update of an

existing risk assessment, the described methodology
was to some extent restricted by the established structure of previous analyses. For instance, the uncertainty
factors became partly limited to those established in
the current risk assessment. However, the generation
and assessment of uncertainty factors extends beyond
the current assessment. It is an aim of the approach
to look for factors that could cause surprises relative
to the assigned probabilities and expected values. A
specific model used in the assessment may constitute
such a factor.

Significant

X
X

Moderate
X

rather than best estimates are required. But for an

already established QRA, this could mean extensive
modifications.

Significant

Table 2. Personnel vulnerability. E[C|A] expresses the

expected number of fatalities given an ignited process gas
leak.
Area

Moderate

Obviously, this framework is dependant on both

the judgement and integrity of the analysts, but no
more so than any other part of the risk assessment.
In fact, every part of the assessment is dependent on
these factors. However, standardisation is required,
both with respect to concepts and methodology. An
example is the definition of Minor in relations to
sensitivity and degree of uncertainty. In this way variation in interpretation from one analysis to another is
avoided.
The background knowledge has to be seen in relation to the analysts performing the study. If the analysts
have little competence in the phenomena studied, this
could affect the uncertainty assessment. But the competence of the analysts is, or at least should be, one
of the criteria of the user of the QRA in selecting the
personnel to perform the assessment. If the analysts
do not have a strong competence in the phenomena
and processes studied, they should not perform the
assessment. However, a QRA covers a broad number of disciplines. Ideally multi-disciplined teams are
preferable, but such teams are in practice not always
possible to implement. Often analysts will have to perform assessment of processes where their background
is limited. If this is the case, it should be reflected
through the Degree of uncertainty assessments.
Uncertainties are only associated with physical,
observable quantities, for example gas volumes,
explosion overpressure and the number of fatalities
the next ten-year period. Uncertainties about probabilities, for example related to ignition of a hydrocarbon
leakage, have no meaning in this framework. The
probabilities are regarded as the assessment groups
assignment based on the knowledge available. When
a probability is associated with a possible event, this is
in itself an expression of the uncertainty of the event
occurring (Aven, 2008a, 2008b).
This can be a somewhat challenging view when the
risk assessment in majority consists of probabilities

1338

and frequencies. In addition, common risk acceptance

criteria like those for individual risk and annual impairment frequencies are not physical, observable quantities. In the presented framework the risk figures are
not presented with any type of uncertainty/probability
distributions. They are presented with a sensitivity
assessment, a set of uncertainty factors and the background for these factors. For instance the analyst
has used a probability for ignition of a process leak.
According to the framework, this probability is not
uncertain. But the background information for this
probability will have to be evaluated. In the background information there will be uncertainty factors
related to physically observable quantities, such as the
gas concentration at a certain distance from the leak
point, the leak rate, the duration of the leak, ignition
following a leak etc. If there are differing expert opinions or little data to support the assessments, this will
be reflected in the framework through the uncertainty
factors, refer to Table 1. The total risk picture covering
(A, C, U, P, S, and K) provides the basis for concluding
about risk acceptance and the need for risk reducing
measures.
As mentioned, the traditional role of the QRA
results is control against predetermined risk acceptance criteria. This practice weakens the use of the
ALARP principle, which states that risk should be
reduced to a level that is as low as reasonably practicable. If the risk indices are below the acceptance
criteria, it is easy to think that no further evaluation
of risk reduction is necessary. With a presentation of
risk that is more comprehensive and nuanced one may
expect that the focus on the risk acceptance criteria
is reduced and focus on the ALARP principle gains a
stronger position.
But what about Dimensioning Accidental Loads
(DAL)? These are loads that an area should be dimensioned against: For instance, an explosion overpressure of 0.3 bar or a fire heat load of 250 kW/m2 with
duration of five minutes. These are based on a frequency criterion of 1.0E-4 per year. This means that
the area should be dimensioned against loads with a
frequency higher than 1.0E-4 per year. An area has to
be designed to withstand one specific load, but how
should this be handled with the presentation of a more
nuanced risk picture?
One can imagine a situation where the DAL calculations have associated uncertainty factors that have
been ranked as having a significant influence on the
risk. First of all, these are factors that are also present
in current QRAs, but they are hidden and thus not
communicated. This means that any confusion regarding what to design for is ignored, as long as one is
given one specific result to relate to. A more nuanced
risk picture may to some extent complicate the process but improve the basis for making an adequate
decision.

With the presented framework, the calculated

results will have to be seen in relation to the sensitivity analyses and the uncertainty factors. If there are
uncertainty factors giving a significant contribution,
this will have to be addressed. One solution could be to
increase the design criteria compared to the calculated
figures. Another approach is to focus on risk and
uncertainty reduction according to the ALARP principle. The QRA does not give the answer, just a support
basis for the decision maker to conclude.

CONCLUDING REMARKS

The proposed framework in this article is not set in

stone. It will need to be developed further; it has so
far only been briefly tested on one offshore QRA.
The work will have to be extended and the presented
approach for identifying and assessing uncertainty factors will have to be further refined and developed.
It still remains to be seen how uncertainty will be
addressed in future updates of relevant regulations
and standards, such as NORSOK Z-013 (2001). This
could also potentially influence the framework. Rather
than focus on the established acceptance criteria, it
is desirable to have a stronger focus on the ALARP
principle.
The decision support role of QRA is emphasized
by Falck et al. (2000) and discussed by e.g. Apostolakis (2004). The decision processes today are closely
linked to the present way of presenting risk results, in
the sense that focus is on compliance with acceptance
criteria. It is believed that the decision processes will
need to change in order for the presented framework
to be fully effective. Such changes have to be implemented also in regulations and relevant standards. It
is however outside the scope of the present paper to
assess in detail how this should be done. A main shift
in focus is required: more weight on ALARP processes
compared to acceptance criteria. If acceptance criteria are to be used, they should be defined in a more
nuanced way than what they are today. The uncertainty
factors should be taken into account. For instance, if
there are uncertainty factors with a large influence on
the risk result, a risk figure close to the acceptance
criteria should not automatically lead to acceptance
of the relevant arrangements or measures. The way
to deal with the uncertainty dimension needs to be
specified.

ACKNOWLEDGEMENTS
The authors would like to thank Andreas Falck, Brre
Johan Paaske and Jan Erik sland for their many useful
comments to an earlier version of this paper.

1339

REFERENCES
Apostolakis, G.E. (2004). How useful is quantitative risk
assessment? Risk Assessment, 24(3), 515520.
Aven, T. (2008a). A semi-quantitative approach to risk assessment, as an alternative to QRAs. Reliability Engineering
and System Safety, to appear.
Aven, T. (2008b). Risk AssessmentAssessing Uncertainties beyond Expected Values and Probabilities. New York:
Wiley.
Falck, A., Skramstad, E. & Berg, M. (2000). Use of QRA
for decision support in the design of an offshore oil production installation. Journal of Hazardous Materials, 71,
179192.

NORSOK Z-013 (2001). Risk and emergency pre paredness

analysis. Norwegian Technology Centre, Oslo.
The Facilities Regulations (2002). Regulations Relating to
Design and Outfitting of Facilities etc. in the Petroleum
Activities. http://www.ptil.no/regelverk/r2002/Innretn
ingsforskriften_e.htm.
The Management Regulations (2002). Regulations Relating to Management in the Petroleum Activities. http://
www.ptil.no/regelverk/r2002/Styringsforskriften_e.htm
Vinnem, J.E. (2007). Offshore Risk Assessment: Principles,
Modelling and Applications of QRA Studies. 2nd ed.
London: Springer.

1340

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk communication for industrial plants and radioactive waste repositories

F. Gugliermetti
Sapienza University of Rome, Rome, Italy

G. Guidi
ENEA Italian National Agency for New Technologies, Energy and the Environment, Rome, Italy

ABSTRACT: Past industrial major accidents (Bhopal, Seveso, Chernobyl, etc.) highlighted that information
and prevention constitute an inseparable pair. International laws give more rights to the public in terms of access
to information as well as in terms of consultation. A correct communication entails advantages for the company
both in normal conditions and in case of accidents. The success of risk communication programs depends upon
their sensitivity to the perceptions of risk and the behaviour of people, competent public authorities, plants
operators and stakeholders. Good risk communication results in an outcome where there is a high level of
agreement between all the interested parties.
For many reasons, the permanent disposal of radioactive waste is one of the most difficult environmental
problem. In this paper two case studies will be discussed: Scanzano Jonico (Italy) as an example of bad risk
communication and Gyeongju (South Corea) as an example of good risk communication.

THE DUTY OF INFORMATION ABOUT RISK

Public consent and reactions cant be ignored or dismissed, as irrational or illogical, when industrial hazardous activities are operated. Access to information,
participation in decision-making and planning, communication about emergency action and plans, regular
repetition of up to date news more and more strongly
have been asserted and supported by the Seveso Directives, 82/501/EEC (Seveso I), 96/82/EC (Seveso II)
and 2003/105/EC (Seveso III), that have been published in these last years. Right to know and need
to know of people are the correct base of the risk communication program. The community had to receive
at minimum information, on request or periodically,
including: the explanation, in simple terms, of activities undertaken at the establishment; the dangerous
substances used, with an indication of the main hazardous characteristics; general information relating to
the nature of major-accident hazards, including their
potential effects on population; adequate information
on how the public will be warned and kept informed
during an emergency, as well as on the behaviour they
should follow in this event.
The requirement of Article 8 of the Seveso I Directive for members of the public to be informed of
safety measures and how they should behave in the
event of an accident was not initially welcomed by
industry in European countries that were frightened

by the unforeseeable public reaction; but the great

damage of the chemical industries public image produced by the post-Bhopal events, produced a greater
willingness to communicate with the public and to
become actively involved with local communities
(Walker 1999).
The Seveso II Directive (art. 13) gives more rights
to the public in terms of access to information as well
as in terms of consultation. Operators as well as public authorities have well defined obligations to inform
the public, by means both of passive information
and active information. Whereas passive information means permanent availability of information, i.e.
that this information can be requested by the public,
active information means that operators or competent
authorities themselves need to be pro-active, for example through the distribution of leaflets or brochures
informing the public about the behaviour in the case of
an accident. The Italian Legislative Decree n. 334/99,
implementing the Seveso II Directive, states in its
Article n.22 that the Region should provide for the
access of interested people to the safety report. The
local government, where the establishment is located,
promptly acquaint the population with information
supplied by the establishments operator, if necessary
made more understandable. This information should
always be updated and permanently available to the
public. Article 23 suggests a greater involvement of
population in decision procedures in the event of new

1341

establishments build up, significant modifications of

the existing ones and settlement and infrastructures
creation around them. The information to be communicated to the public, as provided for in Article 13,
should not be considered as the only means mayors
may use to communicate risk, but it is necessary to
pursue the aim of an effective information.
The Directive 2003/105/EC (Seveso III) extends the
duty of information, about safety measures and on the
requisite behaviour in the event of an accident, to all
persons and to all institutions serving the public (such
as schools and hospitals) liable to be affected by a
major accident originating in an establishment covered
by article 9.
2

INDUSTRIAL RISK COMMUNICATION:

SOME BASIC PRINCIPLES

The National Research Council defined risk communication as an interactive process of exchange of
information and opinion among individuals, groups
and institutions. It involves multiple messages about
the nature of risk and other messages, not strictly about
risk, that express concerns, opinions or reactions to
risk messages or to legal and institutional arrangements for risk management (National Research
Council 1989).
An important concept outlined in this definition is
that risk communication is an exchange of information, or an interactive process requiring the establishment of two-way communication channels.
Building trust in the communicator, raising awareness, educating, reaching agreement, motivating action are some of the possible goals of risk communication (Rowan 1991). Because of this multiplicity
of aims, different strategies of risk communication
may be suitable for different objectives. For example, stakeholder participation methods are likely to
be more appropriate for reaching agreement on particulate matter, while simple risk communication
messages are best for raising awareness.
Any organization planning a risk communication
effort needs to clarify whether the goal is to gain consensus on a course of action, or only to educate people.
In the first case if the organization involves the stakeholders in the decision process, this will facilitate the
process of building trust and gaining agreement on the
course of action (Bier 2001). One of the basic principles of risk communication is establishing trust and
credibility. A broad-based loss of trust in the leaders of major social institutions and in the institutions
themselves has occurred over the past three decades
(Kasperson et al. 1992). This general loss of trust may
also be exacerbates in particular contexts (i.e. Three
Mile Island accident). In cases where trust is particularly low due to past history or the seriousness of an

event, restoring trust and building relationships might

be a primary objective. In order to build credibility and
trust it is indispensable:
to take the time to listen to people and try to
understand their perspective;
to be open and honest, communicating early and
often;
to encourage questions in any areas where there may
be concerns or interest;
to coordinate and collaborate with other credible
sources by building alliances with credible third
parties to raise own credibility.
Unfortunately, trust is difficult to gain but easy to
lose. Recent empirical research suggest that negative
(trust-destroying) events carry much greater weight
than positive events (Slovic 1993). Some actions,
within the risk communicators control, that can harm
credibility are: becoming defensive, hiding information, ignoring the public, disregarding suggestions and
concerns by stakeholders, appearing to act solely in
licensees interests, releasing risk information that
risk analyst dont support, not fulfilling commitments. Other factors, such as i.e. media coverage,
incidents at other facilities, national crises or emergencies (i.e. September 11, 2001) are instead outside
risk communicators control (U.S. Nuclear Regulatory
Commission 2004).
Risk communication processes for situations of distrust can involve several strategies. First, it is important
to recognize that when an audience has serious concerns or negative impressions, one must begin by
listening to them before conveying new information.
After acknowledging that peoples concerns are important and based on valid or understandable motivations,
they may be more willing to listen to new information
and other points of view. In situations of distrust, openness is the surest policy. An industrial plant adopted
an open-door policy, providing public access to all
of the permits and performance reports the company
files with regulatory agencies, in a separate building
located near the plant entrance. When the building was
first opened, it attracted a lot of concerned citizens but
after a while it rarely had a visitor, suggesting that the
companys open-door policy acted over time to create trust (Heath 1995). Very important factors are the
institutions and companies credibility and responsiveness in building trust. In situations of low trust it
is crucial to share information and to involve the public in developing situations in order to face the risk
problem.
Since the early 1990s it became clear that risk had
different meanings to different groups of people and
that all risks had to be understood within the larger
social, cultural and economic context. Although there
was considerable agreement between the public and
scientists on many risk assessments, there were also

1342

some, such as nuclear power, where there were large

differences of opinion. Public perception of radiation
risk differs from the experts one and is determined
by the contest in which the radiation is used (Slovic
1996). Research into perceived risk has highlighted
the lack of confidence the public has in the technical experts (Gough & Hooper 2003). General public
mostly tends not to accept the technical approach, a
quantitative type one, towards the risk concept, typical
of risk analysis experts. Consequently it is necessary to
pass to a new approach, a qualitative one. The mathematical risk expression in terms of probability (i.e.
1 105 ) is meaningless for the general public, while
for example the sentence it is dangerous to smoke is
immediately understandable. It could be helpful to use
familiar units of measure and transforming notation
into concrete examples based on numbers that can help
people understand the size of a risk. Instead of using a
risk that is expressed like the risk of additional cancers is 2.1106 it is possible to present the following
situation: imagine 10 cities of 100,000 persons each,
all with the same exposure to contaminant A. In eight
of these, probably no one would be affected. In each
of the other two cities, there would probably be one
additional cancer, on average (Covello, Sandman &
Slovic 1988).
The process of informing a civil population of the
risks it is exposed to is extremely delicate. The aim
of the information is essentially to prepare the population to act accordingly in an emergency. The security
that this implies should, in turn, reassure the population. Nevertheless an information campaign carried
out incorrectly may simply serve to spread alarm and
fear amongst the population. Information must try to
familiarize people with the risk and give each individual a certain sense of control. This can be achieved by
explaining the nature of the products and installations
to be found, for example, in chemical plants. Other
important factors to be clarified are the real dimension
of the danger, including the normally low probability
of an accident, and the possibility that the individual,
by means of a few simple actions, can notably increase
his or her safety and that of his or her family in the
event of an emergency. When part of the local population works in the industry at issue and, to a greater
or lesser extent, the local economy and development
will depend on that industry, this helps the dissemination of the idea of the industry as a necessary thing
which, despite its negative aspects also benefits the
local community (Casal et al. 1997).
Good risk communication and community involvement will enable government and industry to better
understand public perceptions, to more readily anticipate community responses and to reduce unwarranted
tensions.
Some key principles of effective risk communication are (Covello, Sandman & Slovic 1988):

1. accepting and involving the public as a partner and

stakeholder;
2. carefully planning and evaluating the nature and
content of the risk communication undertaken so
that it is relevant and understandable;
3. listening to publics specific concerns;
4. being honest, realistic and open;
5. speaking clearly and with compassion;
6. ensuring that information is accurate, consistent
between agencies and not speculative;
7. coordinating and collaborating with other credible
sources;
8. effectively communicating with the media;
9. appreciating that intentional communication is
often only a minor part of the message actually
conveyed; the manner of delivery and its tone may
be more important than its content;
10. acknowledging the public concerns and the effects
on the community;
11. focusing on issues and processes rather than
people and behaviours.
It is very important the involvement of stakeholders. If they are aware of risk issues and involved in
the early stage of the decision-making process, the
solutions will inevitably become more sensible and
legitimate. The quality of stakeholders participation
in risk decision-making is often determined by how
well informed they are about the nature of risk issues,
and how responsible institutions deals with them.
A basic reason for the emergency of risk communication research derives from the highly visible
dilemmas that have risen as a result of particular social
conflict over risks (for example, over the sitting or
expansion of hazardous facilities). Fostering appropriate forms of communication between the parties to
such disputes might contribute in some way to better
mutual understanding and hence to conflict resolution.

INFLUENCE OF RISK COMMUNICATION

ON PEOPLES PERCEPTION OF RISK

Risk communication can be aimed at influencing peoples behaviour and perceptions about risk. The goal
might be to place a risk in the right context or to
encourage a change to less risky behaviours.
The idea of acceptable risk or safety may change
quite suddenly after a single stunning accident. Examples of a sudden loss of a safe feeling are the catastrophe at Bhopal, Chernobyl or terrorist attack on the
World Trade Centre. Public opinion is influenced not
only by the accident itself, but maybe even more by
the attention which is paid to it by the media.
Research into risk perception has revealed that,
in addition to rationally considering the probabilities
about a risk, human beings rely on intuitive faculties

1343

and emotions when subconsciously deciding what to

fear, and how fearful to be. By understanding the
factors influencing peoples perception of risk, communicators can more effectively encourage them to
make the right choices.
Several factors influence peoples perception of
risks; the most important ones are:
trust: the more citizens trust people informing them
about a risk, or the institution or company exposing
them to the risk, or the government agencies that are
supposed to protect them, the less afraid they will
be. The less the trust, the greater the fear;
control: the more control population have over a
risk, the less threatening it seems. This explain why
it feels safer to drive than to fly, though the risk of
death from motor vehicle crashes is much higher;
dread: the more dreadful the nature of the harm from
a risk, the more worried people will be;
risk versus benefit: the more citizens get a benefit
from a choice or behaviour, the less concerned are
about any associated risk;
human-made versus natural: natural risks seems less
scary;
new or familiar: new threats generate concern. After
resident have lived with the risk for a while, familiarity lowers their fear;
children: any risk to a child seems more threatening
in the eyes of adults than the same risk does to them;
uncertainty: the less we know, or understand, about
a risk, the scarier it seems.
Perceiving risk through these emotional and intuitive lenses is natural human behaviour, but it can
lead to make dangerous personal choices. Driving
may have felt safer than flying after September 11th,
2001, but those who opted to drive rather than to fly
were actually raising their risk. Risk misperception
can threaten health by making people too afraid, or not
afraid enough. Failing to keep risk in perspective lead
persons to pressure government for protection against
relatively small risks, which diverts resources from
bigger ones. Only by understanding and respecting the
way people relate to risk, risk communicators can play
a vital role in improving the publics health.
4

RISK COMMUNICATION IN RADIOACTIVE

WASTE MANAGEMENT

Over the last decade the awareness of the necessity for the nuclear waste programmes to become
more communicative has increased worldwide. This
reflects advances in national programmes to the phase
of site selection and the necessary involvement of
regional and local bodies as well as concerned citizens, but it is also due to the introduction of legal
requirements for public consultation, often under the

umbrella of the Environmental Impact Assessment

(EIA). Requirements for information to or consultations with the affected or concerned public are often
part of EIA legislation. The European Union Directive 85/337/EEC as amended by Directive 97/11/EC
requires the concerned public to be informed as part
of the EIA process for certain projects, including disposal facilities and facilities for long-term storage of
radioactive waste. This information must be provided
in good time, thus allowing the public to express their
opinion before a decision is taken. Through legislation, the principles in EIA such as openness, broad
involvement of stakeholders, early participation, and
description of alternatives including the zero (do
nothing) solution, become part of the decision making
process.
A license for a repository for radioactive waste
needs not only a technical or regulatory decision but
also a political decision, which in turn requires broad
public consent. In most countries, members of the
public have the opportunity to object to a proposed
nuclear installation at certain phases in the decision
process. Gaining public acceptance is, therefore, a
prerequisite for implementing the final disposal solution. In Sweden, for example, the siting process is
based on voluntary participation and the feasibility
studies take place only in municipalities that have
given their consent. In France, a 1999 decree authorizes the implementation of a Local Information and
Monitoring Committee to be chaired by the Prefect
of Department where an underground research laboratory project is implemented. That committee shall
be responsible for ensuring that all information concerning the evolution of the URL project is freely
accessible. Switzerland has a long tradition of public
involvement in decision making at all political levels.
The licensing procedure for each license includes two
public consultations, which give opportunity for the
public to comment on or object to the project applied
for. In Canada, the Nuclear Safety and Control Act that
established the Canadian Nuclear Safety Commission
(CNSC) imposes requirements for public notification
and participation.
An interesting form of public participation takes
place in the Czech Republic, where representative
of the public sit on the Radioactive Waste Repository Authority (RAWRA) Board. The latter supervises
RAWRA activities and approves plans, budget, etc.
In this way, representative of the public participate
directly in the decision making process.
The regulatory and licensing bodies play a very
important role during the entire decision process for
new nuclear facilities. The independence and public
accountability of the regulators are vital to public confidence in the national high-level waste programme.
The success of public outreach programmes hinges
on the extent to which regulators effectively make

1344

their presence and role known and communicate their

independence.
Especially when the nuclear waste programmes
enter a phase of site selection, the issue arises of
whether the regulators can take an active role with
involvement in the community processes while maintaining independence for later licensing actions.
The traditional policy worldwide has been that the regulators should not be too intensely involved, if at all,
since that might put their independence into question.
Experiences have shown that active regulator
involvement is needed and can be achieved without
endangering independence and integrity as licensing
body. An active role is important to gain trust and
confidence in the communities that the regulator fulfils its protective function in the management process
(OECD 2003).
A basic condition for public dialogue is information
(early, understandable, credible, consistent, related to
all issues of public interest). To provide easy access
to information, the organization issues booklets, provides internet sites, and sets up information centres
near existing facilities or exhibitions in regions of
interest, organizes meetings and discussions (OECD
2003). Policies also reflect the increasing awareness
that public dialogue must be based on two-way communication rather than one-way information. Effective
two-way communication will ensure that the risk information is technically accurate, understandable and
relevant to publics concerns. In order to do this it is
essential first of all to enhance the capability of lay
people to deal with risk information. Still now many
people dont know the difference between hazard
and risk and use the terms interchangeably. Other
useful actions are:
to use qualitative explanations, understandable language and familiar examples or comparisons;
to explain the impact of the decision on them and
the licensee;
to encourage questions to clarify understanding.
An important task allocated to meetings, information centres and internet sites is, indeed, to obtain
feedback from the public in order to identify and later
address issues of public and social interest.
The communication process can only be effective
if all parties of the process are interested in communicating and if the dignity and the roles of all parties are
acknowledged. One pitfall of earlier communication
approaches was the emphasis by waste management
experts on technical issues and issues of natural sciences to the detriment of the wider social issues.
Effective communication requires ability to discuss all
issues of interest to the public.
According to CNSC (Canada), lack of credibility of
waste management organization and regulators seems
to reflect a lack of credibility in governments and

big business as a whole. For waste management

implementers and regulators, this translate not to
a lack of confidence in their competence, but to
scepticism about their integrity and intentions. CNSC
concludes that little can be done directly to change this
attitude, other than maintain a high decree of integrity
in dealing with the public. Relations with the mass
media are important in modern society. Openness and
availability are always required. It is also useful to
make acquaintance with journalists and media editors
so that they can identify the face and the personal
characteristics of the company spokesman as well as
of the media interface officer.
5

SITE SELECTION EXPERIENCES

In France and the UK, earlier experiences in site

selection have led to re-evaluation and redirection of
the overall national programmes. In France a period
of successive crises resulted, in 1991, in a law that
instituted a new approach to waste management in
general, and site selection in particular, with responsibility, transparency and democracy as lead principles.
The new approach to site selection looks for consensus with, and involves actively, responsible territorial
communities. The Law institutes a local information
and monitoring committee on each underground laboratory site. Important for the French approach is the
responsibility taken by Parliament. A key element in
the French programme is that other options (near surface storage and transmutation) are kept open until the
Parliament decides otherwise (OECD 2003).
In the UK, the refusal of the Nirex Rock Characterization Facility at Sellafield in 1997 led to a complete
reappraisal of radioactive waste management policy.
A Parliamentary enquiry recommended in 1999 that
the Government go ahead with underground disposal
but that its policy must be comprehensive and must
have public support. The Government replied that it
would seek public views but that it would look at all
waste management options before endorsing a particular one. After consultation, Ministers announced the
creation in 2003 of a new independent body to oversee
the review and to recommend the best option, or combination. The events of 1997 also led Nirex to adopt
a new Transparency Policy (1999) with a dialogue on
the future long-term management of wastes. A number of dialogue processes are now being tested and
used. The new approach has received initial positive
response. In Spain a delay in the programme allowed
for a strengthened educational programme, which met
with a favourable response.
5.1

Positive experiences

One example is Finland, where a site has been proposed recently by Posiva for detailed investigation in

1345

view of underground disposal of spent fuel, and the

municipality has given its consent. This is the result of
a long-term programme consistent with the decisionin-principle on the national political level and with
Environmental Impact Assessment (EIA) as the key
tool for communication with concerned municipalities and the general public. The site has now been
approved.
In Sweden, the EIA also has been a tool for the
dialogue between SKB (Swedish Nuclear Fuel and
Waste Management Co), the authorities, municipalities, counties and the public. Here the regulators
SKI (Swedish Nuclear Inspectorate) and SSI (Swedish
Radiation Protection Authority) have taken an active
role in the process, supported by a series of research
project. Also the municipalities themselves have taken
new initiative. The municipality of Oskarshamn has
further developed the EIA methodology within the
framework of representative democracy (the Oskarshamn model). The SKB proposed in 2000 three sites
for site investigations. The proposal has been subject
for review and approval by the authorities and the government. Two of three proposed municipalities have
approved and one has denied site investigation.
The progress made in site selection in Finland,
Sweden and France seems to be the result of a proactive programme for public involvement. Other positive
experiences include the site selection for a LLW/ILW
repository in Hungary where a majority of the local
public is supporting the site selection process, the
Vandellos decommissioning project and Andujar uranium mill decommissioning project in Spain, where
a firm proactive attitude was adopted towards all
range of authorities, political parties and the media.
Another example is Belgium where the earlier site
selection approach based on purely technical criteria
has been replaced by local partnership. The new principle is that any party that could be directly affected
by a collective decision, should have a say in it
(OECD 2003).
5.2

Negative experiences

One negative experience is the Sellafield case in UK

for an underground laboratory, where technical discussions did not address the real concerns of the local
community. Other cases are the Gorleben case in
Germany where distrust in officials grew and there
was no real participation and the negative cantonal referenda on Wellenberg in Switzerland in 1995
and 2002.
In Spain since September 1995 an antinuclear organization spread through the press maps of Spain
showing 35 zones that had been investigated according to the HLW site selection programme conducted by
ENRESA. This produced great concern in the vicinity of those areas and paved the way to protests and

to the creation of local anti-repository organizations.

After the successful change in the French approach
to site selection and the establishment of a laboratory
site in clay, a new siting project followed for a granite
site. The response of anti-nuclear movements brought
about refusal reactions from the local populations and
civil society organizations. The local political representatives preferred not to engage themselves in favour
of the project (OECD 2003).
6

SCANZANO JONICO: AN EXAMPLE

OF BAD RISK COMMUNICATION

After the Chernobyl accident, a referendum in 1987

has banned the use of nuclear power in Italy. Currently
Italy has to face the problem of dismantling nuclear
power stations and has to ensure safety solution for
nuclear waste. Some of the waste presently produced,
comes from research, industrial or medical activities.
The radioactive waste are as of today kept in more
than one places (generally close to old nuclear power
plants) in not really safe conditions.
On November 14th, 2003 the Italian Government
signed an emergency decree naming the small southern town of Scanzano Jonico as the site for the national
waste depository. The choice was said to be technical
not political based on the presence of deep salt rock
deposits beneath 300 metres of clay. The proposed site
was a flat area at sea level, 200 metres from the beach
and 100 metres from the Cavone River. The decree
was signed without a preliminary impact assessment
and without public participation. The identification
of the site has been neither concerted nor communicated in advance to the concerned local authorities.
The grievance of the institutions added to the population one. Furthermore the Regional administration
complained the lack of involvement and consultation
in the decision making process.
The issue of the Law Decree created alarm not only
among population of the area but also among institutions, stakeholders, political parties, trade union
organizations, environmental movements. Immediately began a pacific occupation of the proposed site
and blockading railway lines and major highways.
Information and public awareness activities have also
been launched. From the first moment was clear, to
lay people and local authorities, that the implementation of such a repository could represent a huge hazard
for an area larger than the municipal one, involving a
great part of Southern Italy. It was also clear that, apart
from the health hazard, the construction of a permanent nuclear waste storage would seriously damage
the tourism and the agriculture, that are the main economical resources for the area. By the time it was all
over, 100,000 protesters of all political persuasions had
forced the Italian Government to withdraw Scanzano

1346

as the intended site. The emendation is only a first

achievement; currently the conversion Draft Law is
object of a debate in the Parliament and the opposition
to it is strong and widespread.
All these events gives evidence of as the awareness of population on health, environment, selfdetermination, land use planning and institutional
relationship is mature enough so that such an impacting decision cannot be taken against the will of the
people and without agreement with local administration. Mostly, it was a sense of outrage at the militarystyle imposition of the decree, named Scanzano
decree, issued without public consultation or environmental impact studies. The allocation and storage of
nuclear waste cannot be faced as a mere question of
science or discussed as technical concern: it is above
all a political matter.
Since then the word Scanzano has become synonymous of successful anti-nuclear activism. Inspired
by Scanzano, strong popular movements have arisen to
support and ensure that an open, democratic, scientific
process and not a authoritarian decree will decide the
future of Italys radioactive waste.
The incident of Scanzano has been something more
than an advice for all those people who handle the
relationship between science and society: the theory
that problems can be solved by experts behind closed
doors has been proven to be false. Whatever the nature
of the problem (to minimize a risk, to grasp an opportunity) the solution can be searched for only through a
close relationship among experts, institutions and citizens. In other words, only through a process of real
communication.
In conclusion, in democratic societies, the solution of social problems generated by technoscience
must encompass the active participation of the general public. An efficient decisional process requires
the participation of all subjects and implies a transparent flow of communication among the subjects
themselves. The audiences for public involvement
activities should include representatives from local
communities, administrative units (i.e. national,
regional and local), government officials, regulatory
agencies, community and public interest groups, environmental organisations, industry and trade groups,
the scientific community and the news media.
To sum up Scanzano was a failure because there
were neither stakeholders participation nor public
consultation, and above all because the site was selected by means of a military-style decree, that caused
loss of trust in the institutions.
Fortunately in Italy things are changing. A survey
has been accomplished by Sogin (SOciet Gestione
Impianti Nucleari) about risk perception in Italian
nuclear sites on a representative sample of 4000 persons living in 7 areas where are nuclear facilities or
nuclear plants in decommissioning phase. It revealed

that 47% of people could accept a radioactive waste

repository in the place where they live, even in lack
of information. This percentage could notably rise
owing to an effective risk communication. People only
required safety and security measures and employment and development opportunities as compensative
measures.
7

GYEONGJU: A VERY RECENT EXAMPLE

OF GOOD RISK COMMUNICATION

For the past 20 years South Korea has been working

to site a facility capable of storing low and intermediate level radioactive waste materials. The project of
selecting a site for radioactive waste began in 1986.
Nine possible sites for permanent waste dump were
selected but every time residents were against the facility. Over the course of 21 years, two Science Ministers
and one Industry Minister had to resign over problems
related to the selection of a site. The baton was handed
to the Ministry of Commerce, Industry and Energy
in 1997. Buan in North Jeolla was selected, but residents fought pitched battles with police to prevent the
dump from coming to town. The Government changed
its strategy in 2005 and accepted applications from
areas that wanted the facility. There has been a certain
amount of interest in the employment and development opportunities that come with a willingness to
perform a necessary, but potentially unpopular, task.
The South Korean government decided to capitalize
on the expressed interest in development opportunities
and set up a first of a kind referendum that allowed
various regions, already preselected for having the
required space and appropriate geology for a low level
waste disposal facility, to hold a vote. Of the four
participating regions, the one with the largest portion of favourable votes would be awarded the right to
host the facility. In order to be considered the winner,
the region had to also achieve a voter turnout above
a specified minimum level (one-third of all eligible
voters).
On 2nd November 2005, the South Korean government announced that Gyeongju had won. Of the
eligible voters in the Gyeongju area, 70,8% turned
out to vote and 89,5% of the voters were in favour of
the plan.
In the three regions that lost the bid, severe criticism was targeted at environmental and anti-nuclear
energy activist groups that encouraged residents to
vote against the project. At Yeongdeok, Gunsan and
Pohang, some residents scuffled with activist group
officials. The residents said the activists had hindered
the development of regional economies and asked
them to leave the areas.
The waste products that the Gyeongju facility will
store are created in nuclear power plants, nuclear

1347

medical facilities and industrial facilities using nuclear

materials in measurement and processing. A circular building will be excavated from rock 80 metres
beneath the surface at the repository site in Gyeongju
and it will be large enough to hold 100,000 drums of
waste. Construction of the first phase of the repository is expected to be finished by the end of 2009.
A temporary holding facility will also be built on the
surface.
Gyeongju will receive a special state subsidy of
around 200 million. The disposal fees for the site are
expected to be approximately 5.5 million per year. In
addition, the area will have an increased opportunity
to compete for other nuclear related projects (Nuclear
Decommissioning Authority 2007).
This is the beginning of a new era of local autonomy
since it was decided through the direct participation of
local residents, instead of having the government unilaterally decide what happens with such state projects.
It sets a positive precedent for resolving social conflict.
Contrary to Scanzano, Gyeongju was a successful
project because the selection was made with the overwhelming support of local residents and in a fair and
transparent manner.
The Gyeongju nuclear waste repository is a significant success story as it stands as the first state project
determined in a plebiscite. Unfortunately there were
aspects of the process that left something to be desired.
That would especially be the case with the allegations
that Gyeongjus local government engaged in illegal
behaviour in the voting process. That is regrettable,
since it leaves a negative precedent for what is the
first application of the new local referendum law.
In addition, too much compensation is being used to
make the selection a success: Gyeongju will receive a
high subsidy, a proton accelerator centre, and it gets to
be the new base of the Korean Hydro & Nuclear Power
Corporation.

CONCLUSIONS

The siting of nuclear waste facilities has been very

difficult in all countries. It is a well known fact that
radioactivity in general has associated with it a highly
significant dread factor. This may be an older and
more basic reaction to dangers which cannot be seen
or heard or smelled. This fear of radiation in general
is carried over (although in a less extreme manner) to
radioactive wastes.
Bearing in mind that risk in radioactive waste issues
is perceived to exceed substantially acceptable levels,
it is necessary to proceed so as to reduce its perception.
It could be useful to take some actions in order to
gain public consensus on dangerous activities and on
radioactive waste repository in particular.

Some activities are suggested below:

to use a two-way communication trying to actively
involve lay people and stakeholders already in site
selection (this was not done in Scanzano);
before a referendum on a site, selected by scientists, takes place, population should be correctly
informed about risks and benefits coming from the
repository;
to bet on voluntary approval of the pre-selected
sites by population, offering noteworthy benefits
as compensative measures. They should be offered
both to municipalities and to population: i.e. a discount on electricity price, tax exemption such as ICI
(a tax on property) and tax on urban solid waste.
All these benefits should not be extortionate, as in
Gyeongju case, so as not to arouse suspicion about
the existence of a high risk;
independence of the institution or company responsible for site construction and for site management
and control.
It is not possible to ignore the need to have public
consent to operate hazardous processes or to simply
dismiss public views as misguided, irrational and illinformed. Furthermore it is obvious by now that the
problem of radioactive waste repository could not be
solved by experts behind closed doors or by means of
an authoritarian decree, as in Scanzanos case.

REFERENCES
Bier, V.M. 2001. On the state of the art: risk communication
to the public. Reliability Engineering and System Safety,
71: 139150.
Casal, J., Montiel H., Planas-Cuchi E., Vilchez J.A.,
Guamis J., & Sans J. 1997. Information on the risks of
chemical accidents to the civil population. The experience of Baix Llobregat. Journal of Loss Prevention in the
Process Industries, Vol. 10 (3): 169178.
Covello, V.T., Sandman, P.M. & Slovic P. 1988. Risk Communication, Risk Statistics, and Risk comparison: a manual
for plant managers. Chemical Manufacturers Association.
Gough, J. & Hooper G. 2003. Communicating about risk
issues.
Http://www.europe.canterbury.ac.nz/conferences/tech
2004/tpp/Gough%20and%20Hooper_paper.pdf.
Heath, R.L., 1995. Corporate environmental risk communication: cases and practices along the Texas gulf coast.
Communication Yearbook, 18: 255277.
Kasperson, R.E., Golding D. & Tuler S., 1992. Social
distrust as a factor in siting hazardous facilities and
communicating risks. Journal of Social Issues, 48 (4):
161178.
National Research Council. 1989. Improving risk communication. National Academy Press.
Nuclear Decommissioning Authority. 2007. Managing radioactive waste safely: literature review of international
experiences of community partnership.

1348

OECD. 2003. Public information, consultation and involvement in radioactive waste management. An International
Overview of Approaches and Experiences.
Rowan, K.E., 1991. Goals, obstacles, and strategies in risk
communication: a problem-solving approach to improving communication about risks. Journal of Applied
Communication Research, 19: 300329.
Slovic, P., 1993. Perceived risk, trust, and democracy. Risk
Analysis, 13: 675682.

Slovic, P., 1996. Perception of risk from radiation. Radiation

Protection Dosimetry, 68 (3/4): 165180.
U.S. Nuclear Regulatory Commission. 2004. Effective Risk
Communication.
Walker, G., Simmons, P., Irwin, A. & Wynne B., 1999. Risk
communication, public participation and the Seveso II
directive. Journal of Hazardous Materials, 65: 179190.

1349

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk management measurement methodology: Practical procedures

and approaches for risk assessment and prediction
R.B. Duffey
Atomic Energy of Canada Limited, Chalk River, Ontario, Canada

J.W. Saull
International Federation of Airworthiness, East Grinstead, UK

ABSTRACT: We have discovered a factor that precisely represents the information needed in decision-making,
skill acquisition and increasing order. The information entropy given by the H-factor is a function of the
learning rate or depth of experience, and increases so does the value of the H-factor declines, demonstrating
increased order. The relative value of the information entropy H-factor at any experience depth is a direct
measure of organizational learning and the effectiveness of safety management systems. The definition of
the Information Entropy, H, is in terms of probabilities based on the frequency distribution of outcomes at
any given depth of experience. We suggest here a simple step-by-step Methodology to applying the H-Factor
to managing risk. To distinguish it from other standards and existing risk management techniques (e.g. PSA,
FMEA, bow-tie etc., etc.), we call this safety management and risk prediction measurement process the Risk
Management Measurement Methodology, or the RM 3 for short.

INTRODUCTION

We want to reduce our risk, so we need to apply a

method. We have explained the observed outcomes
(which appear as errors, accidents and events) as the
average behavior of a complex human-technological
system (HTS), whose instantaneous microscopic or
detailed behavior cannot be described (Duffey and
Saull, in prep.). We utilized the well-known concepts and methods of statistical physics to explain
the stochastic occurrence of failures in technological systems, coupled with the systematic key role
of human error and learning through experience.
The method also potentially quantifies the impact of
safety management schemes, and explains why the
impact of learning changes the predictions we want
to make.
In statistical physics, the very impossibility of predicting the instantaneous behavior of complex physical
systems lead to Boltzmanns successful invention of
deriving the average observed behavior from the distribution of unobserved statistically fluctuations. The
method is available in standard physics textbooks,
Sommerfeld (1956) and Sherwin (1961), and is the
foundation of thermodynamics and information theory. Using this same idea translated to experience
space, we derive a physically based theory based on
the distribution of observed outcomes (error states) for

a system, where the error states define one distribution

(or microstate) out of many possible ones. Using this
physical model, we can therefore reconcile the apparently stochastic occurrence of accidents and errors
with the observed systematic trend from having a
learning environment. It is further shown that learning rates depend on experience, and that a formal
experience shaping factor emerges naturally during the total experience. We address the key question
of whether and how they are indeed preventable using
safety management systems (SMS).
We reconcile the apparent randomly observed outcomes with the systematic effect of learning. We
describe the inherent unpredictability (and inevitability) of accidents and the large contribution due to
unforeseen human errors. We adopt and adapt the principles and methods of statistical physics to the observation and prediction of risk, accidents and the
contribution of human error. The idea here is to find
the simplest physical theory and systematically obtain
the most useful results, validated against data.
The theory provides a consistent interpretation
between learning models and the statistical behavior
of error occurrence, the impact of context and performance shaping on learning. Thus, we observe order
emerging from chaos in HTS, precisely as postulated
by Prigogine for physical and chemical systems. Based
on physics and physical reasoning, we obtain a general

1351

accident theory that is able to provide real predictions

and be validated by and explain actual risk data. It is
also testable against actual event data.
Firstly, we invoke the Learning Hypothesis (Duffey
and Saull (2002)). We have already shown that generally humans learn from their mistakes, and as a
result the error rates in a learning environment follow an exponential curve that is decreasing towards
a minimum with increasing accumulated experience.
The Learning Hypothesis is also entirely consistent
with Ohlssons psychological error correction theory developed to describe individual human learning
(Ohlsson (1996)). Thus, we invoke the same model
and interpretation as applicable all the way from the
micro-individual learning process and behaviors to
the observed macro-system outcomes. In this sense,
homo-technological systems behave just like a collective of Homo sapiens when learning and making
errors. The model hypothesis is also an emergent one
that does not attempt to describe the inner unobserved
myriad details behind human decision-making and
actions. The concept is entirely consistent with the
ideas of complexity as the the amount of information
that is discarded during the process that brings a physical object into being (Tor Norretranders (1998)).
2

INFORMATION ENTROPY

We now want to know how probable things like outcomes are, and how the arrangements that we might
have and the number of outcomes vary with experience. For each ith number of microstates, ni , representing the location of a possible outcome, the probability,
pi , of an outcome at any experience depth, i , for any
total number of, Nj , outcomes is given by the Laplace
(1814) relation:
pi (ni /Nj )

(1)

Now all microstates are equally likely: any one can

occur. There are many, many combinations or distribution of the total number of outcomes, Nj , that are
possible amongst the many different microstates, ni .
The number of possible combinations, Wj , is given by
the standard relation:
Wj Nj !/ni !

(2)

where conventionally the product ni ! = n1 !n2 ! n3 !

. . .ni !. The measure of the overall State of Order
(or Disorder), H, in the system is described by the
usual definition, being the probability of the, ni ,
distributions (Greiner et al (1997)):
Hj = Nj !(pi /ni !)

(3)

Following Jaynes (2003) and Pierce (1980), we have

distinguished the information entropy by using a
different symbol, H, from the usual one used in thermodynamic use, S. As defined here, H, has nothing
to do with thermodynamics: we have simply used the
same model assumptions to derive the distributions
and the probabilities. In Jaynes (2003) (p.353) terminology, H, is here the most likely measure of
the uniformity of a probability distribution, or the
amount of uncertainty. It is a direct measure of the
unobserved complexity of the observed system outcomes. In Information Theory the common usage of
the entropy, H, is as a measure of the amount of choice
in the rate of the number of possible binary combinations or ways that any message, signal or symbol can
be encoded, encrypted, transmitted, or presented to a
receiver/observer (see Pierce (1980) p.78105). Conversely, it is a measure of the uncertainty as to what
will actually be received or, as we shall see later, how
it may be actually perceived. In our interpretation, this
uncertainty is a measure of the probability of error,
recalling that this is the converse of the probability
of success, which we can now call certainty. With
our intent and interest in managing risk and safety,
the entropy, H, then simply reflects how we observe,
interpret or perceive the distribution of HTS outcomes
occurring in our experience interval.
Interestingly, the entropy, H, has also been proposed
as a natural measure of knowledge, or a scoring rule
or calibration in the elicitation of expert judgment
(Woo book, Chapter 6, p. 161 et seq.). Differences
between experts are adjusted by weighting the value
of H assigned to the each experts judgment of the
probability of assumed or presumed outcomes.
We are adapting this method of reasoning to human
behavior and risk. Now only specific distributions
of the microstates can satisfy the physical and hence
mathematical restraints. The conditions we seek for
the risk case are here given in mathematical form:
1. finding a maximum or a minimum number of
microstates, as a condition for learning, (dn/d =
0), so that the observed distribution is the most
likely;
2. conserving the total experience (ni i = constant,
j ), since we have a closed system and an arbitrary
but fixed amount of knowledge and learning; and
3. conserving the number of outcomes (ni =
constant, Nj ) as we only have a fixed number of
actual outcomes, and hence being invariant for
small changes.
Note that the outcomes themselves may be either
distinguishable or indistinguishable, as for quantum
statistics, in the sense that the outcomes may be unique
or may be a repeat event. They are independent. This
distinction has the effect of changing the number of
allowed/observed microstates (or distributions) as the

1352

population of outcomes as a function of experience,

ni (i ), changes.
Now, as before the number of possible arrangements are,
Wj = Nj !/ni !

(4)

where, from Stirlings formula for large numbers,

Nj ! (Nj /e)Nj

and

ni ! (ni /e)ni

(5)

Using the exact same derivation as used in statistical

mechanics by Sommerfeld (1956) and Sherwin (1961),
for example, substituting the above estimate leads to
the number of possible combinations,
ln Wj Nj ln Nj ni ln ni

Provided we can now find the probability distribution of the outcomes, we can measure the entropy. In
the special case of the outcomes being described by a
continuous random variable, as for a learning curve,
we can replace the summation over discrete intervals
with an integral function:

Hj = pi ln pi dp = p2i (1/4 1/2 ln pi )
(9)

(6)

where the summation with depth is given conventionally by:

The information entropy is then given by this

H-factor. As the learning rate or depth of experience
increases so does the value of the H-factor decline,
demonstrating increased order. The relative value of
the information entropy, H-factor, at any experience
depth is a direct measure of organizational learning
and the effectiveness of safety management systems.
Hence, we can quantify and manage risk, without
needing shelves full of purely paper management
edicts, policies, procedures and regulations.

ni ln ni = n1 ln n1 + n2 ln n2 + + ni ln ni (7)
3
This simple looking expression for the possible
combinations leads directly to the measure of disorder,
the information entropy, Hj , per observation interval.
In terms of probabilities based on the frequency
of microstate occupation, ni = pi Nj and using Stirlings approximation we find the classic result for the
information entropy:
Hj = pi ln pi

(8)

and the maximum value occurs for a uniform distribution of outcomes. Interestingly, this is of course also the
Laplace-Bayes result, when p(P) 1/N for a uniform
risk.
These two classic results for, Wj and Hj , are literally
full of potential information, and are also in terms of
parameters we may actually know, or at least be able
to estimate.
The above result for the Information Entropy, H,
thus corresponds to the famous thermodynamic formula on Boltzmanns grave, S = k ln W, where, k,
corresponded to the physical proportionality between
energy and temperature in atomic and molecular systems. But there the similarity but not the analogy ends:
the formulae both look the same, but refer in our case
to probability distributions in HTS, not in thermodynamic systems. In our case, for the first time, we
are predicting the distribution of risk, errors, accidents, events, and decisions and how they vary with
depth of experience and learning, in a completely
new application. In fact, since, H, is a measure of
Uncertainty, we can now define the Certainty, C, as
now given by, say, C = 1 H, which is parallel to the
definition of Exergy as the converse of Entropy.

THE EXPONENTIAL DISTRIBUTION

OF OUTCOMES, RISK AND ERROR STATES

Now we are observing an accumulated experience

interval with a fixed and finite number of outcomes, in
a macrostate arrangement that is one of many possible
ones. We hypothesize that the most likely distribution
is the one we happen to observe! This is equivalent
to the maximization or minimization of (information)
entropy principle, which is discussed at length by
Jaynes (2003). From the separate conditions of constancy of experience and outcomes, for any small
change in the distribution, total experience and number
of outcomes are insensitive or invariant so:
d(ln Wj ) = 0, for no change in the probability of a
given microstate distribution.
d( ni i ) = i i dni = 0, for no change in total
experience with invariant ith increments.
d ni = 0, for no change in the total number of
outcomes, Nj , and hence in the total microstate
distribution.
The distribution of microstates with experience is
then given by the formula:
d(ln W) = d ni ln(ni + i ) = 0

(10)

or solving,
ni = n0 exp( i )

(11)

where, and , are constants which are unknown at

this stage, and, n0 , the number of microstates or distributions at the initial or zeroth level, (i = 0), of

1353

experience, 0 . The microstate occupancy number, ni ,

is now defined as a function of the depth of experience, and is an exponential. This exponential form is
exactly what we observe, for example with data for
auto deaths and ship sinkings: it is the distribution
of risk for a given experience interval. As measured
by the occupancy of states, risk decreases as we gain
depth of experience.
This form of exponential occupancy distribution
relation is called the Gibbs function or the Boltzmann
distribution in the scientific fields of thermodynamics, statistical physics and quantum mechanics. All we
have done is extend the use and application of statistical models from the microscopic world of energy
flow to the treatment of macroscopic outcomes, as
Jaynes (2003) had suggested it might. Any microstate
distribution is governed by this expression for the
exponential distribution: it is itself universal, although
the chosen parameters may vary.
As we saw, the equivalent probability of this or
any ith microstate is just the same as Laplaces definition, being the ratio of the microstate distribution
occupancy to the total outcomes:
pi = ni



ni = ni /Nj

(12)

We summarize the main results as follows. The

instantaneous failure rate, , which gives the universal learning curve (ULC) is given from the DSM
expression that fits the world data (Duffey and Saull,
in prep.):
() = m + (0 m )ek

where k 3, 0 = n/0 , and m 5.10 per unit

accumulated experience, and initial experience has
been taken as small. The probability, p, for the bathtub
curve, by the double exponential:
p() = 1 exp {( m )/k (0 )}

pi = p0 exp( i )

(13)

The probability of risk also decreases with increasing depth of experience. The form of the distribution
is similar to that adopted for the fractional population
change during species extinction (Gott (1993)), where
in that case the parameter, , is a constant extinction
rate rather than a learning constant.

THE FUNDAMENTAL BASIS FOR RISK

PREDICTION

We have discovered the H-factor that precisely represents the information needed in decision-making,
skill acquisition and increasing order. It is called the
Information Entropy. The properties of the H-factor
are as a fundamental measure of the predictability
of a random event, (that) enables intercomparisons
between different kinds of events. Hence, we need
in any HTS to quantitatively assess a Safety Management Systems effectiveness in reducing outcomes,
and managing risk. This is the prudent and necessary
measure that management needs to create order from
disorder.

(15)

The definition of the information entropy, H, is in

terms of probabilities based on the frequency distribution of outcomes at any given depth of experience,
Hj = pi ln pi

(16)

Given the distribution of outcomes with depth of

experience is simply an exponential, a reasonable fit
to the available outcome data is given by (Duffey and
Saull, in prep.):
pi = p0 exp aN ,

The expression for the probability of any ith

microstate that follows is then, with the usually
adopted exponential model for a given total, Nj ,

(14)

(17)

where the value of the exponent, a, illustrates, measures and determines how fast we are achieving order
and our safety goals. Here, N , the non-dimensional
measure of the depth of experience, is just a convenient measure for the experience interval divided by
the total observed, /M , in the experience interval.
The information entropy is then given by the nondimensional form of the H-factor:

Hj = 1/2{p0 eaN }2 {aN + 1/2}

(18)

A typical best value for the parameter a 3.5, as

given by the aircraft near-miss and auto death distributions with experience (i.e. the number of flights, and
the years of age above 18 of the driver, respectively).
The relative value of the information entropy,
H-factor, at any experience depth is a direct measure
of organizational learning and the effectiveness of
safety management systems. Hence, we can quantify and manage risk, without needing shelves full of
purely paper management edicts, policies, procedures
and regulations.
5

RM3: PROCEDURE FOR EVALUATING

THE H-FACTOR

According to the science, mathematics, data, and

trends from modern systems, we just need to measure

1354

practices, trends and values obtained in other systems.

The final step is to make a prediction based on all
that is known, and to amend, revise and change the
management measures if needed.
The obvious objection to this approach is why
not just fit some arbitrary curves and make some
projections? The answer is that the form of the relations used, and the constants derived, are soundly
based on observation, data and learning. To demonstrate the application of the method, the direct use
of this approach to a practical example is given in a
companion paper by Duffey and Skjerve (2008).

Data Collection ( past and present )

Step 1 :Tabulate Event Data +
Choose Experience Measure
and Observation Interval

DSM 1:Calculate rate,IR

Plot chart
(Plot Failure rate on ULC )
Calculate k-value and compare to k=3

H1 :Plot data
Fit exponent "a" value in Equation 2
(plot distribution on Depth graph)
Compare to a = 3.5

DSM 2: Calculate Probability, p

Equation 3

H2: Information Entropy

Calculate H using Equation 3

DSM 3:Bathtub Curve

Choose non-dimensional experience
Plot data on bathtub plot and compare

H 4:SMS Measure
Plot H versus non-dimensional experience,N*

DSM4: Check learning rate

Compare to Goals
Refine goals and measures

H5: Check risk managment

Compare to Goals
Refine goals and measures

Make Risk Prediction

Measure again

Make Risk Prediction

Change systems

REFERENCES

Figure 1. The Risk Management Measurement Methodology (RM3) step-by-step procedure using a double or
side-by-side flow.

and manage the right things. We can now show how to

do this by providing an example of a working procedure. We suggest here a simple step-by-step Methodology to applying the H-Factor to managing risk. To
distinguish it from others standard and existing risk
management techniques (e.g. PSA, FMEA, bow-tie,
etc.,) we call this Safety Management and Risk Prediction Measurement Process the Risk Management
Measurement Methodology, or the RM 3 for short.
We suggest the following procedure or outline, as
shown in the diagram (Fig. 1) to make it easier to
follow.
The two process flow paths descend as boxes of the,
DSM and H, sequences on each side of the diagram, in
complementary and fairly obvious steps in both intent
and scope. The LHS of the diagram, or DSM, follows
the use of event and outcome rates to determine the
learning rate and probability, and hence to determine
position on the ULC (DSM path). The RHS of the diagram, or H, follows the use of the numbers of events
or outcomes to determine the distributions with experience, in order to establish the information entropy
(the H path).
Each step is a progression from the previous step,
utilizing the outcome data to fit the constants in
the general theory, and intercompare to the best

Duffey, R.B. & Saull, J.W, 2002. Know the Risk, First Edition,
Boston, USA, Butterworth and Heinemann.
Duffey, R.B. & Saull, J.W. Manuscript in preparation to be
published. Managing Risk: The Human Element, West
Sussex, UK, John Wiley & Sons Limited.
Duffey, R.B. & Skjerve, A.B. 2008. Risk Trends, Indicators
and Learning Rates: A New Case Study of North Sea Oil
and Gas, Proc. ESREL 2008 and 17th SRA Europe Annual
Conference, 2225 September, Valencia, Spain.
Goth III. & J. Richard, 1993. Implications of the Copernican principle for our future prospects, Nature, Vol. 363,
pp. 315319.
Greiner, W., Neise, L. & Stocker, H., 1997. Thermodynamics and Statistical Mechanics, Springer, New York, pp.
150151.
Jaynes, E.T. 2003. Probability Theory: The Logic of Science, First Edition, Edited by G.L. Bretthorst, Cambridge
University Press, Cambridge, UK.
Laplace, P.S., 1814. Essai philosophique sur les probabilities,
extracted as Concerning Probability in J.R. Newmans
The World of Mathematics, Vol. 2, p. 1327, 1956.
Pierce, J.R. 1980. An Introduction to Information Theory,
Dover, New York.
Sherwin, C.W. 1961. Basic Concepts of Physics, Holt,
Rinehart and Winston, New York, USA, 1st Edition,
pp. 308338.
Sommerfeld, A, 1956. Thermodynamics and Statistical
Mechanics: Lectures on Theoretical Physics, Academic
Press, New York, NY.
Tor Norretranders, 1991. The User Illusion: Cutting consciousness Down to Size, Penguin Books, London, UK,
p. 290.
Woo, G., 1999, The Mathematics of Natural Catastrophes,
Imperial College Press, London, UK, Chapter 6, p.161
et seq.

1355

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Risk perception and cultural theory: Criticism and methodological

orientation
C. Kermisch
Universit Libre de Bruxelles, Brussels, Belgium
Fonds de la Recherche Scientifique

P.-E. Labeau
Universit Libre de Bruxelles, Brussels, Belgium

ABSTRACT: This paper presents briefly cultural theory applied to risk perception and examines how it has
been empirically tested. It shows that the method used so far for the tests is flawed and thus does not allow to
(in)validate cultural theory. Therefore, we suggest a research direction that would enable an effective examination
of the validity of cultural theory for risk perception.
1

INTRODUCTION

The development of risky technologies such as nuclear

power plants has led to numerous conflicts between
technical risk estimates and lay perceptions of risk.
Henceforth, the justification of these technologies
requires both aspects to be taken into account. It is
thus fundamental for the manager, who is eager to
optimize his/her work, to secure a good understanding
of the risk perception process.
To this end, the psychometric paradigm occupies,
since the end of the seventies, a dominant position
(Slovic, 1992, McDaniel, 1997). This theory presents
two specificities.
On the one hand, it focuses on risk characteristics, which are likely to influence its perception. It
is, for example, the case of dread or involuntary
risks, newness, immediacy of the effects or threats
to future generations. These characteristics allow for
instance to explain why, in the United States in the
eighties, nuclear power was perceived as one of the
most threatening technologies, although its scientific
risk assessment was quite low.
On the other hand, the psychometric paradigm proposes a universal conception of risk perception. Hence,
it does not permit to measure its variations at the
individual and collective levels.
This supposed universality is the most criticized
aspect of the psychometric paradigm and can be overcome by adopting cultural theory (Douglas, 1982,
Thompson, 1982, Rayner, 1992). The latter supplies
an alternative interpretation of risk perception and
addresses the issue of differences of individual or
collective perception in a more relevant way. However,
its empirical testing does not seem to be satisfactory

and needs improvement. The purpose of this paper is

to provide one direction of research allowing such an
improvement.
The present communication is composed of three
distinct parts. The first part presents briefly cultural theory; the second part exposes the attempts
of empirical confirmation this theory has instigated
and demonstrates their methodological flaws; finally,
the last part outlines an approach which could be
implemented for this theory to be tested in a relevant
manner.

2
2.1

CULTURAL THEORY OF RISK:

PRESENTATION
Cultural biases and social organizations

The origin of this theory is to be found in the cultural analysis initiated by the British anthropologist
Mary Douglas. Indeed, in Purity and danger, the
examination of pollution and taboo leads Douglas to
observe that social organizations select beliefs related
to dangers in order to secure their own stabilization
(Douglas, 1966).
The generalization of this finding constitutes the
basic hypothesis of cultural theory as it will be developed later: cultural biases and social organizations
sustain each other mutually without any bond of causal
nature. In other words, a type of social organization
is maintained by a set of cognitive and axiological
contents, and conversely (Fig. 1).
While, in Douglas cultural theory, the notion of
social organization deals only with the interpersonal
relationship patterns, the concept of cultureor

1357

Figure 1. Interaction between cultural biases and social

organizations.

Figure 2. Grid and group dimensions; four types of social

organizations (Douglas, 1979, p. 7).

cultural biasdesignates the beliefs, values, and

ways to perceive the world and to react to it. More precisely, Mary Douglas conceives culture as a framework
that individuals develop in the course of the interactions they are engaged into. She considers culture
neither as a closed system nor as a simple external
passiveinfluence of the social environment on individuals, and even less so as a set of mental productions
such as values, perceptions, beliefs, norms or rationalizations. Such a conception of culture rehabilitates the
individual in as much as he constructs the sense of
his actions in conjunction with the requirements of the
social context within which he operates.
In Natural symbols (Douglas, 1970), Mary Douglas defines four types of social organizations based
on two independent dimensions, grid and group
(Fig. 2).
Group corresponds to the experience of a bounded
social unit. Grid refers to rules which relate one person to others on an ego-centred basis (Douglas, 1970,
p. viii). Group represents thus the level of social
incorporation of the individual in a social unit; it

describes the range of the interactionswithin the

social structureas absorption of individual life. Grid
takes into account the nature of these interactions; it
reliesin the anthropological traditionon an egocentred foundation, that is on age, sex or heredity of the
concerned individual. This dimension integrates thus
the institutionalized classifications, which regulate the
interactions of the members of any social structure.
Together, grid and group enable Douglas to highlight,
in Cultural bias, four types of social organizations: the
hierarchical, the insulated, the individualistic, and the
egalitarian forms (Douglas, 1979).
Strong grid and group dimensions define the hierarchical form, where the roles are clearly specified
and the related distribution of resources is often
inequitable.
The insulated form corresponds to a weak group
dimension and a strong grid dimension. Individuals do
not have the possibility to make personal transactions,
their autonomy is minimal and their social roles are
completely determined.
Weak grid and group dimensions define the
individualistic form. In this case, the conditions are
strongly competitive, individual autonomy is important and its members exercise control over individuals
from other social forms.
Finally, the egalitarian form corresponds to a strong
group dimension and a weak grid dimension. The roles
there are not defined and the fundamental social constraints are related to the protection, by the group, of
the outside border against foreigners and to the control
exercised in the name of the group over the behaviour
of individuals.
In Douglas interactive system, individual choices
are not determined, this enabling the individual to
move from one social organization to the other.
A dynamics exists thus between the individual and his
environment. As mentioned before, to each of these
social forms corresponds a cultural bias covering the
whole cognitive and axiological contentsincluding
risk perceptions.
The next section is more specifically devoted to the
study of the link between risks perceptions and social
organizations.
2.2 Risk perceptions and social organizations
From the cultural hypothesis, which suggests that
any kind of social organization produces its own
conception of nature and its inherent dangers in order
to stabilize itself, Douglas and Wildavsky point out
in Risk and culture that the perception of dangers
is related to the social form in which they manifest
themselves: each cultural bias has a propensity to minimize some dangers and to maximize others, it operates
therefore a selection among the multiple surrounding
dangers (Douglas & Wildavsky, 1982). Risk is

1358

constructed according to a social process, in which

individuals are implicated as active subjects, excluding
thereby the possibility of access to a raw risk.
The originality of Risk and culture lies in the
determination of the types of risks that emerge as stabilizing principles in each type of social organization
Douglas and Wildavsky are only interested in three of
the social forms presented in Cultural bias: the hierarchical, the individualistic and the egalitarian forms.
The hierarchical form judges the rationality of its
members in terms of submission to the rules enacted by
the system. The collapse of the system and the risks,
which threaten it, mainly riots and wars, constitute
the prevailing risks. In this context the risk to take
or to avoid is selected by a process rather than by an
individual. The hierarchical form believes that it provides the best protection to the future generations by
maintaining as such the system and its related benefits.
The individualistic form develops a type of rationality in compliance with the principles of utility theory.
The main risk is the collapse of the market as defined
in economic terms. However, individualists are not
scared by uncertainties, which are perceived as opportunities within a logic of intentional risk taking as a
function of expected profits. They tolerate change and
are willing to assume responsibilities as long as they
are the beneficiaries in case of profit.
A third type of rationality appears in the context
of voluntary adhesion, which is specific to egalitarian forms. It can be understood through the strategies
required for their survival founded on equality, purity,
the exclusive nature of the group and the importance
of not being excluded from it. The risks highlighted
in the egalitarian form are those that are conflicting
with its demand for purity. Examples are pollution or
technological risks, which are tied to the relations with
the outside world.
Let us now examine how cultural theory has been
empirically tested.
3

EMPIRICAL TESTING OF CULTURAL

THEORY

A good number of psychologists and sociologists have

attempted in the nineties to devise an empirical validation of this theory. Karl Dake and Aaron Wildavsky
were the first authors to publish such a study
(Wildavsky & Dake, 1990, Dake, 1991, Dake,
1992), but they were soon followed by a lot of other
researchers (Marris et al., 1998, Brenot et al., 1998,
Sjberg, 1998).
In order to establish the influence of social cultural factors on risk perception, the basic methodology
underlying most of these studies can be summarized in
three steps: the determination of risk perceptions, of
cultural biases, and the intercorrelation of the results.

3.1 Determination of risk perceptions

First, a questionnaire is proposed for the purpose of
determining risk perceptions. It is divided in two parts.
On the one hand, it is dedicated to the study of risk
perceptions by using variables chosen from a list of
36 concerns that people have about society today
(Wildavsky & Dake, 1990, p. 46). The interviewee
has to give his/her opinion about problems concerning technology, environment, war, economy and social
deviance. For instance, he/she has to evaluate the
importance of risks such as National Debt, dangers
associated with nuclear power, political corruption,
federal overregulation, decline in moral values or
restriction of civil liberties, etc.
On the other hand, Dake and Wildavsky also study
the level of approval or rejection of societal risks in
order to establish if technological risk-taking and its
management are seen as an opportunity for advancement or as a source of catastrophe. Interviewees have
to evaluate propositions such as American society is
becoming overly sensitive to risk; we now expect to be
sheltered from almost all dangers, or every citizen
should become more informed about the great risks
we are taking with our technological developments
(Dake, 1991a, p. 68).
It has to be mentioned that these results are more
about social, political and economical preferences
than about differences concerning risk perceptions
strictly speaking, which are, in principle, the focus
of their works.
Other researchers such as Marris et al., or Sjberg
will take this criticism into account and readjust the
focus in order to evaluate perceptions of technologies
such as genetic engineering, car driving, terrorism,
nuclear power, AIDS, food additives, etc. (Marris
et al., 1998, p. 647).
3.2 Determination of the cultural biases
A second questionnaire is submitted with the objective to establish the cultural bias of each interviewee.
It includes three indexes, dedicated respectively to
the hierarchical, the individualistic and the egalitarian
bias. In another paper, Dake also takes into account
the insulated bias (Dake, 1992, p. 31).
The questionnaire measuring the egalitarian bias
refers to equality of conditions, to equity, as well as
to the perception of unfairness associated with other
cultural biases.
The criteria allowing to establish the belonging to
the hierarchical bias are founded on the importance
given to patriotism, centralization, order, law and
ethical norms.
The index dedicated to the evaluation of the individualistic bias refers to the support of economic growth
and private profit.

1359

Finally, the questionnaire concerning the insulated

bias is founded on the determination of a lack of trust
towards other people and on the feeling of being treated
unfairly.
Table 1 shows representative items from the British
Edition of the Cultural Biases Questionnaires, which
also takes into account the insulated bias (Dake, 1992,
p. 31).
Besides the fact that the samples used by Dake and
Wildavsky or by Dake are too small to be representative, some other criticisms may be formulated.
Indeed, these questionnaires do not allow to
take into account the social cultural polyvalence of
interviewees, who could belong to different cultural
biases depending on specific domains of their life,
such as family life, professional life, etc. (note that

Table 1. Items selected from the British Edition of the

Cultural Biases Questionnaire (Dake, 1992, p. 31).
Hierarchy
We should have stronger armed forces than we do now.
I think there should be more discipline in the youth of today.
I would support the introduction of compulsory National
Service.
People should be rewarded according to their position in
society.
I am more strict than most people about what is right and
wrong.
Individualism
If a person has the get-up-and-go to acquire wealth, that
person should have the right to enjoy it.
In a fair system people with more ability should earn more.
A free society can only exist by giving companies the
opportunity to prosper.
In this country, the brightest should make it to the top.
People who are willing to work hard should be allowed to
get on.
Egalitarianism
I support government efforts to get rid of poverty.
If people in this country were treated more equally we
would have fewer problems.
The government should make sure everyone has a good
standard of living.
Those who get ahead should be taxed more to support the
less fortunate.
I would support a tax change that made people with large
incomes pay more.
Fatalism
A person is better off if he or she doesnt trust anyone.
There is no use in doing things for peopleyou only get it
in the neck in the long run.
Cooperating with others rarely works.
The future is too uncertain for a person to make serious
plans.
I have often been treated unfairly.

all the authors do not agree on the fact that people

may belong to several biases at the same time).
If individuals do indeed belong to several biases, we
can ask ourselves to which of these biases the individual refers to when he/she answers the questionnaire.
In this respect, Oltedal suggests that the interviewee
should specify his/her bias within the survey (Oltedal
et al., 2004, p. 27).
Moreover, a more general criticism concerns the
fact that answers can be influenced by the content and
the admitted objective of the questionnaire, as well as
by the behaviour of the examiner.
3.3 Interpretation of the results
In a third step, the answers of the questionnaires
are intercorrelated in order to identify relationships
between on the one hand, the perception of risks, and,
on the other hand, the cultural biases.
To put it briefly, Dake and Wildavsky establish that
the correlation coefficient r, proves to be positive when
linking:
the hierarchical bias and the perception of risks from
social deviance;
the individualistic bias and the perception of economic risks;
the egalitarian bias and the perception of technological and environmental risks.
These results lead them to conclude that strong links
exist between cultural biases and risk perceptions as
suggested by cultural theory. They conclude thus that
cultural theory is a good predictor for risk perception.
However, the sole use of the correlation coefficient
in order to express the link between two variables may
be questioned. Indeed, r corresponds to the strength
of the relationship existing between two variables, but
does not give information about the relationship in
itself. Presenting only r values does not seem very
rigorous: a superficial reading of these results leads to
interpret the relationship as if it was linear by default,
which is not necessarily the case.
Moreover, Dake and Wildavsky state that the correlations between cultural biases and risk perceptions
are strong. However, the highest r value in all these
early studies never exceeds 0,46which is at best
an average correlation according to Cohen (Cohen,
1988).
Finally, it has to be mentioned that a few years
later, researchers such as Sjberg or Marris et al.,
do not reach the same conclusion: according to them,
cultural biases do not influence risk perceptions significantly. These differences can be explained by the
use of different questionnaires and/or indicators.
The variations in the interpretation of the results do
not allow the formulation of any conclusive validation.
But what seems the most important is that the study of

1360

Figure 3. Dakes and Wildavskys results express the link

between risk perceptions and cultural biases.

Figure 4. Testing interactions between cultural biases and

social organizations.

this experimental research leads to the conclusion that

it does not verify the validity of cultural theory. Let us
examine why.

4.1 Determination of cultural biases

3.4

Omission of social aspects

The fundamental hypothesis of cultural theory lies

in the existence of a strong link between social
organizations and cultural biases. However, Dakes
and Wildavskys methodology completely neglects the
social aspects. Indeed, by studying the link between,
on the one hand, risk perceptionswhich are part
of cultural biasesand, on the other hand, cultural
biasesthrough the study of general values, they
test the internal cohesion of cultural biases rather than
the links between social organizations and cultural
biases (Fig. 3). For instance, they verify if individuals
belonging to the hierarchical group do focus on social
disorder, or if egalitarians focus on environmental and
technological risk.
Douglas main hypothesis being that each specific
cultural bias is linked to its social organization model,
they should have tested instead the link between cultural biaseseventually including risk perceptions
and the type of social organization individuals are
belonging to in order to really (in)validate cultural
theory.
That is why this work does not allow either to validate or to invalidate cultural theory. A reorientation of
the research becomes therefore necessary.

A NEW METHOD FOR EMPIRICAL TESTING

With this prospect, we propose a methodology, which

evaluates on the one hand the cultural bias of the
individual (including risk perceptions), based on psychometric questionnaires, and on the other hand the
social organization he/she is belonging to, according
to the model developed by Jonathan Gross and Steve
Rayner (Gross & Rayner, 1985) (Fig. 4).

The first step should be based on questionnaires such

as the ones that have been used by Dake and Wildavsky aiming at determining the cultural bias to
which individuals belong. These questionnaires would
thus include questions about general values, as well as
more precise questions concerning risk.
This method does not allow to resolve the problems
mentioned above. However at this point, it appears to
be the only way to identify to which cultural bias an
individual belongs.
4.2 Determination of social organizations
The second step should consist in determining the type
of social structure in which an individual is evolving.
With this prospect, the method developed by Gross and
Rayner in Measuring culture would be helpful. Indeed,
it seems to allow to quantify grid and group dimensions
and thus to identify the type of social organization of
an individual.
In their model, both grid and group dimensions
are defined through several predicates. We have to
insist on the fact that these predicates are not empirically observed properties but formal characteristics
belonging to the model, and thus, determined by the
analyst.
Their model, called EXACT, integrates five
elements:
X: the whole set of members belonging to the social
organization;
A: the whole set of activities during which the
members interact;
C: the whole set of rolespublicly recognized
assumed by the members during their activities;
T: the representative duration of a set of activities;
E: the whole set of non-members who could be
admitted as members during T.
Gross and Rayner propose to define group thanks to
five predicates and grid thanks to four predicatesbut

1361

it has to be kept in mind that this choice of predicates

is not the only possible choice.
The predicates leading to the estimation of the
group dimension are:
proximity, measuring the intimacy level between
members of a social organization and being a
function of the number of their interactions;
transitivity of the social interactions between
members;
frequency, as the proportion of time during which
a member interacts with other members, relative to
the total time T;
range of interactions, measuring the commitment of
the individual in his social organization, compared
to his commitments outside of it;
impermeability, evaluating the probability that a
non-member joins the social organization.
The predicates allowing the estimation of the grid
dimension are:
specialization, as the proportion of possible roles a
member assumes during T;
asymmetry, corresponding to the impossibility to
exchange roles between members;
enabling of a member, corresponding to the ratio
between the number of roles to which the individual
can access and the total number of roles;
liability, corresponding to the ratio between the
number of interactions in which a member is
dominant or subordinate and the total number of
interactions.
In order to obtain the grid or group value, we need
to average the values of the former predicates.
Even if this second step integrates the social aspect,
which was omitted in Dakes and Wildavskys studies,
it does not seem entirely satisfactory. Indeed, it seems
difficult to assign a number to some of the predicates
mentioned above. For instance, how should we evaluate the probability that a non-member joins the social
structure (impermeability)?
Further research seems to be needed in order to
improve the model developed by Gross and Rayner
and make it operational.
4.3

Testing cultural theory

Once a proper model aiming at determining the

belonging to a social organization will be developed,
the final step would consist in examining the interrelations between cultural biases and social organizations, in order to verify if some relationships can be
established, which would (in)validate cultural theory.
From a theoretical point of view, this method seems
much closer to the basic principles of cultural theory,
which intends to study the link between cultural biases
and a social organization.

CONCLUSION

After a brief presentation of cultural theory applied

to risk perception and its empirical testing, we have
shown why the test methodology used insofar is flawed
from a theoretical point of view.
To overcome this problem, we suggest a procedure in three steps. First, psychometric questionnaires
should be submitted aiming at establishing the cultural bias of an individual. Secondly, a method similar
to the one developed by Gross and Rayner should
be implemented. However, this method needs some
improvement and further research should be conducted in order to develop a more precise and rigorous
way of measuring the belonging to a social organization. Finally, the results should be analyzed and
correlated through statistical techniques.
This approach would enable an effective examination of the validity of the cultural paradigm.
However, it has to be noted that further research is
needed in order to provide an empirical basis to this
method.

REFERENCES
Brenot J., Bonnefous S. & Marris C., Testing the cultural
theory of risk, Risk Analysis 18(6), 1998, p. 729739.
Cohen J., Statistical power analysis for the behavioral
sciences, Hillsdale, Erlbaum, 1988.
Dake K., Myths of nature: culture and the social construction of risk, Journal of social issues 48(4), 1992,
p. 2137.
Dake K., Orienting dispositions in the perception of risk:
an analysis of contemporary worldviews and cultural
biases, Journal of cross-cultural psychology 22(1), 1991,
p. 6182.
Douglas M. & Wildavsky A., Risk and culture, Berkeley,
University of California Press, 1982.
Douglas M., Cultural bias, London, Royal Anthropological
Institute (Occasional paper No. 35), 1979.
Douglas M., Natural symbols: explorations in cosmology,
London, Barrie & Rockliff, 1970.
Douglas M., Purity and danger: an analysis of concepts of
pollution and taboo, London, Routledge & Kegan Paul,
1966.
Gross J. & Rayner S., Measuring culture, New York,
Columbia University Press, 1985.
Marris C., Langford I. & ORiordan T., A quantitative test
of the cultural theory of risk perceptions: comparison with
the psychometric paradigm, Risk analysis 15(5), 1998,
p. 635647.
McDaniels T. et al., Perception of ecological risk to water
environments, Risk analysis 17(3), 1997, p. 341352.
Oltedal S. et al., Explaining risk perception. An evaluation of
cultural theory, Trondheim, Rotunde, 2004, available on
http://www.svt.ntnu.no/psy/Torbjorn.Rundmo/ Cultural_
theory.pdf.

1362

Rayner S., Cultural theory and Risk Analysis, in Krimsky

S. & Golding D. (eds), Social Theories of Risk, Westport,
Praeger, 1992, p. 83115.
Sjberg L., World views, political attitudes and risk perception, Risk: health, safety & environment 137, 1998,
p. 137152.
Slovic P., Perception of risk: reflections on the psychometric paradigm, in Krimsky S. & Golding D.
(eds.), Social theories of risk, Westport, Praeger, 1992,
p. 117152.

Thompson M. & Wildavsky A., A proposal to create a cultural theory of risk, in Kuntreuther H. & Ley E. (eds.),
The risk analysis controversy: an institutional perspective,
Berlin, Springer-Verlag, 1982, p. 145161.
Wildavsky A. & Dake K., Theories of risk perception: who
fears what and why?, Daedalus 119(4), 1990, p. 4160.

1363

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Standing in the shoes of hazard managers: An experiment

on avalanche risk perception
C.M. Rheinberger
WSL Swiss Federal Institute for Snow and Avalanche Research, Davos, Switzerland
Institute for Environmental Decisions, ETH Zurich, Switzerland

ABSTRACT: This paper deals with the question of how societal impacts of fatal accidents can be integrated
into the management of natural hazards. In Switzerland, the well-known -model, which weights the number of
expected fatalities by the function N (with > 1), is used across governmental agencies and across hazards to
give additional weight to the number of potential fatalities in risk estimations. Although the use of this function
has been promoted in several federal publications, it is not evident whether laypeople want public managers to
overweight less frequent accidents of larger consequences against more frequent ones of smaller consequences.
Here we report on a choice experimental task that required respondents to decide on three risky situations. In
each situation they were asked, which of two roads they would protect from avalanche risk, given that they
were the responsible hazard manager and had only resources to protect one of the two roads. The results of
this experiment show that laypeople behave on average risk seeking ( < 1) rather than risk averse ( > 1),
when they make decisions on avalanche risks involving fatalities. These results support earlier research, finding
experts to behave risk prone when deciding on multi-fatality risks (N > 100). In conclusion, people tend to give
more weight to the probability of fatalities than to the number of expected fatalities.

INTRODUCTION

In a frequently cited article published in Management

Science, Paul Slovic and colleagues (1984: 464) asked
the question: How should a single accident that takes
N lives be weighted relative to N accidents, each of
which takes a single life? More than 30 years later,
the question of how such societal impacts of fatal accidents can be integrated into hazard management is still
looming. In Switzerland, the well-known -model,
which weights the number of expected fatalities by
the function N (with > 1), is used across governmental agencies and across hazards to give additional
weight to the number of potential fatalities in risk estimations. The use of this function has been promoted
in several federal publications including the recent
Strategy for the Protection against Natural Hazards
(PLANAT 2005).
But do laypeople want public managers to overweight less frequent accidents of larger consequences
against more frequent ones of smaller consequences?
Within the framework of expected utility theory
(Von Neumann & Morgenstern 1947), overweighting of consequences implicates risk averse behavior
of the decision maker. This risk averse behavior
requires the societal impact of a hazard to be modeled

adequately as a function of the number of fatalities N ,

implying a weighting of N so that a single large accident becomes more seriousand thus its prevention
more worthythan many small accidents causing the
same total number of fatalities (Slovic et al. 1984).
However, risk averse behavior in managing fatal accidents may not benefit the public. As Keeney (1980a, b)
pointed out: (i) a certain loss of N fatalities is generally perceived worse than a 5050 chance of either
0 or 2 N fatalities; (ii) above a certain threshold
of N , the marginal impact on society declines with
each additional victim (people often state that a catastrophe is catastrophe, no matter if it costs 2,000 or
2,100 lives); (iii) risk equity as prescribed by individual safety goals (see e.g. Pate-Cornell 2002) leads to
risk seeking decision behavior (Keeney 1980b).
In this paper, we report on a recent experiment
from Switzerland that asked people about their perception of avalanche risk to roads, and in particular on
their weighting of accident frequencies vs. the associated number of fatalities. We therefore conducted
a choice experimenta method, which has recently
gathered the attention of researchers in the field of
mortality risk (Alberini et al. 2007; Itaoka et al. 2006;
Tsuge et al. 2005)putting laypeople into the shoes
of natural hazard managers.

1365

2
2.1

METHOD

9 blocks of participants provided answers on 27

different situations.

Design of the experimental task

Choice experiments are based on random utility

theory, which describes discrete choices in a utility
maximizing framework (Louviere et al. 2000). Basically, random utility theory assumes that the utility
U of avoiding a specific risk can be split into a valuebased cognitive part expressed by the value function V
and a random component that is determined by factors subsumed under the notion of the affect heuristic
(Finucane et al. 2000; Slovic et al. 2004). Therefore,
individual i will prefer to protect the road A over the
road B, if it holds that:
Pr(A|i, C) = Pr{Vi (A) + i (A) > Vi (B) + i (B)} (1)
for A and B being alternatives of the choice set C.
Equation 1 was used to elicit preferences for the
prevention of multi-fatality accidents caused by snow
avalanches. To this end, we presented each participant
with four choice situations. These situations differed
with regard to the number of expected accidents over
the next 20 years and the associated number of fatalities per accident. Figure 1 illustrates one example
situation. The frequency and the associated fatalities
per accident were changed iteratively among alternative roads. In each situation respondents were asked,
which road they would protect from avalanche hazards,
given that they were the responsible hazard manager
and had only resources to protect one of two roads (i.e.
C = {A, B}).
Using a fractional design, we collected choices on
annual mortality risks ranging from 0.2 to 0.9 expected
fatalities per year (i.e. 4 to 18 expected fatalities over
the 20-year period). Each block of participants had to
decide on two choice situations, in which the expected
fatalities were equal for both alternatives; one situation, in which the expected fatalities were more on the
more frequently affected road; and one situation, in
which the expected fatalities were more on the road
with the larger consequences per accident. In total,
Which of these two roads would you protect from avalanches ?
Road A

Road B

Expected accidents
During 20 years

Expected number of
fatalities per accident

Road A

Road B

I would protect:

Figure 1.

Example of one choice situation.

2.2 Participants
Based on a random sample from the telephone book,
we recruited respondents for a survey from two areas
of Switzerland. About half of the respondents were
selected from the mountainous region around Davos;
the other half of the respondents lived in the city of
Zurich. These samples were chosen to represent highand low-exposed people.
To collect the data, we conducted a questionnaire
survey by mail. The questionnaire with cover letter was
sent to those individuals, which agreed in the phone
recruitment to participate in this study. The experimental task, which we present here, was completed
and returned by 471 persons. The response rate for the
mountaineers was 50% (n = 227), and for the urbanities 54% (n = 244). It is crucial that participants
responses can be linked to their experiences with natural hazards. Therefore, we asked participants whether
they had ever been affected by any kind of natural
hazard. A description of the demographic characteristics of the survey sample is given in Table 1. The
samples were well balanced for gender and for origin.
Furthermore, there were expected differences between
mountain and urban areas concerning the experience
with natural hazards.
The questionnaire comprised 16 pages and contained two different choice tasks (here we only report
on one of them). Besides, we asked several questions on the perceived risk from natural hazards and
from ordinary road accidents in mountainous regions
of Switzerland. Participants provided answers on their
familiarity with the life in the mountains, past experiences with natural hazards, the severity of these
threats, and some attitudinal questions related to
behavior under risk. In addition, respondents answered
how sure they were in their decisions. Data on
socio-demographics were also collected.

Table 1.

Descriptive statistics for the sample areas.

Setting

Urban area

Mountain area

N
Age, M (SD)
Gender
Males
Females
Experience
Yes
No

227
49.0 (16.0)

244
48.4 (16.5)

52%
48%

50%
50%

29%
71%

46%
54%

M = Mean; SD = Standard Deviation.

1366

2.3

Data analysis

Having collected many choices on these risky situations, one can calculate for each option the probability
to be selected by an individual i, given the observed
attributes of the option under choice (namely the frequency of accidents and the number of fatalities) and
the alternative option. The analysis of these choices
calls for a flexible choice model such as the conditional
logit model:
exp Vi (A)
Pr(A|i) =
exp Vi (A) + exp Vi (B)

(2)

where the value function V may take on any functional form (Greene 2003). This flexibility allows for
testing different value functions and is therefore an
appropriate model for our purpose.
3
3.1

RESULTS
Qualitative results

Participants who provided qualitative answers on their

choice making could be divided into three distinct
groups: (i) about half of the respondents pointed out
that they had emphasized the probability of accidents;
(ii) a quarter of respondents stated that they would
particularly avert accidents causing multiple fatalities;
and (iii) another quarter of respondents indicated that
they had made their decisions based on the expected
number of fatalities over the whole 20-year period.
When asked how certain they were in their decisions on a four-staged category scale (uncertain, rather
uncertain, rather certain, or certain), respondents differed depending on their individual characteristics. We
found that men were more certain in deciding than
women, younger respondents were more certain than
older ones, and particularly older women were highly
uncertain. A one-way ANOVA confirmed these significant differences in decision certainty for gender
(F = 6.73, p < 0.01), age (F = 6.78, p < 0.01), and
the interaction between gender and age (F = 9.93,
p < 0.01). Other characteristics had no significant
impact on respondents certainty.

prospect theory (Tversky & Kahneman 1992), nor

preference reversals due to changes in the reference
state (Tversky & Kahneman 1991) seem to affect the
choice situations we presented in this experiment.
Therefore, we used a EUT-compatible value function for the model estimation:
Vi = R + Xi

Quantitative results

Vi = R + Xi

i I

(4)

where the product of the accident probability and the

weighted number of fatalities is referred to as the
perceived risk vector R (Bohnenblust & Slovic 1998).
Table 2 presents the estimation of the value function specified in Equation 4 for the best-fit model.
The -value was selected based on the Akaike Information Criterion (AIC), which is computed as the log
likelihood minus the number of parameters to be estimated. According to Agresti (2002), minimizing the
AIC yields the best fit for generalized linear models.
Figure 2 shows the AIC-values of the estimated models depending on the underlying -values ( [0, 2]).
The best-fit model results from = 0.72.
The results reported in Table 2 show that only the
perceived risk has significant impact on the appraisal
of the choice situations, i.e. we can drop the subscript for the individual i and write a generalized value
Estimates for the best-fit -model.

Variable

The use of the conditional logit model (Eq. 2) requires

selecting a value function V , which reflects utility theory. Though expected utility theory (EUT) has been
shown to be problematic with regard to a number of
aspects (Kahneman & Tversky 2000), these drawbacks might be neglected here for good reasons. First,
the experimental setting deals solely with decisions
under risk in the domain of losses. Neither misperceptions of probabilities as proposed by cumulative

(3)

where and are coefficient vectors. R is a vector of risk, which is defined as the product of the
accident frequency over 20 years and a weighting function W (N ) of the number of fatalities per accident.
Consequently, this product is equal to the disutility
caused by the expected fatalities, which in the case
of W (N ) = N becomes the statistical mortality risk
per year. Xi is a vector of personal characteristics
including age, gender, education, experience with and
exposure to natural hazards for each individual i out
of the response population I .
In order to test the validity of the -model, we modified the basic model of Equation 3, using the weighting
function W (N ) = N :

Table 2.

3.2

i I

Coefficient t stat.

Best-fit specification, AIC = 1.1397

Perceived risk ( = 0.72)
0.4993
Gender = male
0.0442
Age
0.0019
0.0970
Education = High School

Exposure = weekly
0.1102
0.1771
Experience = yes

1367

Dummy (1 if true, 0 if false).

17.977
0.404
1.059
0.866
1.051
1.640

Figure 2. AIC-values of the estimated models depending

on the underlying -values.

Figure 3. Probability to protect road A (N accidents, each

of which takes a single life) over protecting road B (a single
accident that takes N lives). Curve 1 represents a choice situation of equal expected fatalities, i.e. the number of accidents
in A equals the number of fatalities in B; Curves 2 and 3 represent choice situations, in which more fatalities are expected
in B than in A, i.e. the number of accidents in A equals the
number of fatalities in B + 2 (curve 2), and +5 (curve 3)
respectively.

function as V 0.5(p N 0.72 ). Inserted into Equation 2, this function predicts the probability that people
choose to protect road A (with N accidents, each of
which takes a single life) over road B (with a single
accident that takes N lives).
Already at balancing ten accidents with one fatality each (on road A) vs. one single accident with ten
fatalities (on road B), people prefer to protect road A by
90% (curve 1 in Fig. 3). Even in situations, in which the
number of expected fatalities to be avoided on road B
is larger than on Athe risk neutral agent would thus
prefer to protect Bour decision model predicts that
A is protected until the number of expected fatalities
falls below a certain threshold N (curves 2 and 3
in Fig. 3).

CONCLUSIONS

In this study, we present first results of a recent choice

experimental task, which aims to improve the understanding of decision-making on the mitigation of snow
avalanches in the public space. The experiment gathered more than 1,400 decisions from laypeople on
the hypothetical protection of avalanches endangering
roads in the Swiss Alps.
The results of this study indicate that people perceive avalanche risk neither as the mere number of
expected fatalities (see Fig. 3), nor do they overweight the number of fatalities as suggested e.g. by
the Swiss strategy for the protection against natural
hazards. In fact, the majority of the respondents put
more emphasis on the probability of accidents.
This finding is in line with the psychophysical
principle of decreasing sensitivity to fatalities. This
principle, also referred to as psychophysical numbing (Fetherstonhaugh et al. 1997), implies that the
subjective value of saving a specified number of lives
is greater for a smaller tragedy than for a larger one and
may have large impact on the choices that respondents
made in our experiment.
Comments by respondents suggest that the preference for avoiding frequent accidents is fostered by
different perceptions of the risk on road A and B.
Particularly, several respondents stated that they feel,
it would be possible to temporarily close roads on
which accidents are only expected in times of increased
avalanche danger. Thus, they opted for protecting
the road with multiple accidents, as these are more
difficult to predict and to avoid by road closures.
The non-significance of the socio-economic variables in the -model estimation (Table 2) indicates that
this finding does not depend on specific characteristics
of the respondents, such as their own exposure to or the
experience with avalanches, but can be generalized.
Coming back to our initial question whether laypeople want public managers to overweight less frequent
accidents of larger consequences against more frequent ones of smaller consequences: The results of
this study show that laypeople would not do such overweighting when put into the shoes of hazard managers.
Furthermore, results of recent research from Sweden
(Abrahamsson & Johansson 2006) suggest that risk
experts also weight the probability of fatal accidents
more than the number of potential fatalities in their
decisions about risky situations. Though this might be
partly explained by their responsibility toward society
in any case of fatal accident, even experts do not seem
to accept the paradigm of risk aversion in live saving
interventions.
In conclusion, policy makers in the field of natural hazards are well advised to rethink the paradigm
of risk aversion. The imprudent use of the -model
does neither lead to decisions upon risk reductions,

1368

which are optimal from a societal perspective, nor does

it take into account the different perceptions of risky
situations.
REFERENCES
Abrahamsson, M. & Johansson, H. 2006. Risk Preferences
Regarding Multiple Fatalities and Some Implications
for Societal Risk Decision Making-An Empirical Study.
Journal of Risk Research 9(7): 703715.
Agresti, A. 2002. Categorical data analysis. Wiley series in
probability and statistics. New York: Wiley.
Alberini, A., Tonin, S., Turvani, M. & Chiabai, A. 2007.
Paying for permanence: Public preferences for contaminated site cleanup. Journal of Risk and Uncertainty 34(2):
155178.
Bohnenblust, H. & Slovic, P. 1998. Integrating technical
analysis and public values in risk-based decision making.
Reliability Engineering & System Safety 59(1): 151159.
Fetherstonhaugh, D., Slovic, P., Johnson, S.M. & Friedrich, J.
1997. Insensitivity to the value of human life: A study of
psychophysical numbing. Journal of Risk and Uncertainty
14(3): 283300.
Finucane, M.L., Alhakami, A., Slovic, P. & Johnson, S.M.
2000. The affect heuristic in judgments of risks and
benefits. Journal of Behavioral Decision Making 13(1):
117.
Greene, W.H. 2003. Econometric Analysis. Upper Saddle
River, NJ: Prentice Hall.
Itaoka, K., Saito, A., Krupnick, A., Adamowicz, W. &
Taniguchi, T. 2006. The effect of risk characteristics on
the willingness to pay for mortality risk reductions from
electric power generation. Environmental & Resource
Economics 33(3): 371398.

Kahneman, D. & Tversky, A. (eds), 2000. Choices, values

and frames. New York: Cambridge University Press.
Keeney, R.L. 1980a. Equity and Public Risk. Operations
Research 28(3): 527534.
Keeney, R.L. 1980b. Utility-Functions for Equity and Public
Risk. Management Science 26(4): 345353.
Louviere, J.J., Hensher, D.A. & Swait, J.D. 2000. Stated
choice methods: analysis and applications. Cambridge:
Cambridge University Press.
Pate-Cornell, M.E. 2002. Risk and uncertainty analysis
in government safety decisions. Risk Analysis 22(3):
633646.
PLANAT 2005. Strategie Naturgefahren Schweiz. Biel:
Swiss National Platform for Natural Hazards.
Slovic, P., Finucane, M.L., Peters, E. & MacGregor, D.G.
2004. Risk as analysis and risk as feelings: Some thoughts
about affect, reason, risk, and rationality. Risk Analysis
24(2): 311322.
Slovic, P., Lichtenstein, S. & Fischhoff, B. 1984. Modeling the Societal Impact of Fatal Accidents. Management
Science 30(4): 464474.
Tsuge, T., Kishimoto, A. & Takeuchi, K. 2005. A
choice experiment approach to the valuation of mortality.
Journal of Risk and Uncertainty 31(1): 7395.
Tversky, A. & Kahneman, D. 1991. Loss Aversion in Riskless Choicea Reference-Dependent Model. Quarterly
Journal of Economics 106(4): 10391061.
Tversky, A. & Kahneman, D. 1992. Advances in
Prospect-TheoryCumulative Representation of Uncertainty. Journal of Risk and Uncertainty 5(4): 297323.
Von Neumann, J. & Morgenstern, O. 1947. Theory of
Games and Economic Behavior. Princeton, N.J.: Princeton University Press.

1369

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

The social perception of nuclear fusion: Investigating lay understanding

and reasoning about the technology
A. Prades, C. Oltra & J. Navajas
CIEMAT, Barcelona, Spain

T. Horlick-Jones
Cardiff School of Social Sciences, Cardiff University, UK

J. Espluga
Universitat Autnoma de Barcelona, Spain

ABSTRACT: This paper reports on the development of a novel methodology to investigate lay perceptions of
nuclear fusion as an energy source. Our approach has concentrated on generating a learning and deliberation
process that allows groups of lay participants to engage with information about fusion in a quasi-naturalistic
fashion. In this way, we have attempted to avoid the terms of this deliberation being pre-framed according to
the technical logics of nuclear engineering and risk management. We provide some early observations on the
implementation of this methodology.

INTRODUCTION

Previous research into the social perception of Large

R&D Programmes has clearly shown that there is no
such thing as a uniform and unchanging public opinion towards nuclear fusion. In common with other
emergent technologies, fusion is not well understood
among the lay public and, consequently, historical
associations with the fission programme have a dominant role in shaping views towards fusion, where they
do exist (Prades et al, 2007; 2008).
In order to enhance the understanding of lay publics,
they need to engage with information about the specific technology. To provide such information in a
suitably balanced way, and to promote engagement
with it, is a non-trivial problem, and indeed a research
problem in itself. There is a danger here of creating
perceptions by providing information in such a way
that it drives the recipients to obvious and perhaps
ill-considered conclusions. Rather, research participants need to be taken through a learning process,
which allows sometimes quite complex information
to be assimilated, and examined in a considered way,
avoiding the imposition of a correct technical framing of the facts. In this way it will be possible to
examine the changing modes of reasoning adopted by
lay people as they gain a better understanding of fusion
technology and related issues.

In this paper, we report on the development of

a hybrid methodology, designed to accomplish such
a process of engagement, and to gain insights into
changes in lay reasoning as people proceed through a
learning and discussion process. We also present some
very early observations on the implementation of this
methodology.
2

THE METHOD

From a methodological perspective, we recognised

there was a need to adopt an approach that allowed
groups of lay citizens to discuss and deliberate upon
the issues surrounding fusion in quasi-naturalistic
ways. Research participants needed to be able to gain
sufficient knowledge in order to engage with the complexities of these issues, but without the terms of this
deliberation being framed according to the technical
logics of nuclear engineering and risk management.
We decided to use hybrid discussion groups, that
each met twice, with an interval between meetings that
allowed participants to reflect upon matters raised at
the first meeting. Here, our objectives were to allow the
assimilation of the detail of the issues, and to enable
group participants to explore these issues carefully,
and in a considered and well-informed way. Similar
reconvened groups were used in implementing part of

1371

the UKs recent public debate on genetically modified

(GM) crops, where they were effective in promoting
processes of learning and deliberation (Horlick-Jones
et al, 2007a; 2007b). They have been also used
successfully in a recent pilot engagement exercise carried out for the UK railway industry (Horlick-Jones,
2008).
The reconvened discussion groups were to be similar to focus groups in nature; however their design
differed from classical focus groups in that they
would allow the presentation of new material to participants, and the discussion of this material in a structured
way (cf. Bloor et al, 2001). In this way, we proposed
to generate a sensitisation, learning and deliberative
process, rather than simply seeking to tap pre-existing
understandings (or lack of understanding). The groups
were to have a hybrid nature, in the sense that they
would combine elements of research and engagement.
Importantly, this methodology concentrates on the use
of ordinary language, and the elicitation of patterns of
demotic practical reasoning, rather than pre-framing
the issues in terms of technical categories (HorlickJones, 2005; 2008; Myers, 2007; Timotjevic and
Barnett, 2006).
Rather than receiving formal tuition on the technicalities of the issues to be discussed, the participants
would be encouraged to reflect upon the nature of
fusion-related issues by taking part in exercises that
promote awareness of the character of fusion, seen
in terms of everyday knowledge and reasoning. We
planned to accomplish these objectives by the use
of structured facilitation, stimulus materials (including a vignette scenario: Hughes, 1998), and component parts of some problem structuring methods
(Rosenhead and Mingers, 2001). Group participants
would also be invited to consider information materials
we would provide during the interval between group
meetings, and to keep a diary of their impressions.
These information materials would provide a number
of perspectives on fusion, drawn from industry, media
and environmental group sources.

DATA GATHERING

A pilot study, comprising a single pair of groups, was

carried out at the end of September 2007. The main
finding was that the group-based design worked well,
and showed its potential to generate quasi-naturalistic
collective discourses, and to promote a rich discussion
and learning process among our research participants.
Subsequently, in October 2007, we ran six discussions groups (each meeting on two occasions),
composed of lay citizens. An agency specialising in
providing such services for businesses and research
organisations was employed to recruit our participants.
As our focus is on lay citizens, those likely to have specialist knowledge or strong views about fusion were
excluded. In practice the recruiters excluded members of environmental groups or people who (or whose
immediate family) work in energy-related employment. The recruitment company was asked to segment
our groups along age (18 to 25; 26 to 40; over 40) and
socio-economic lines (AB: High & Medium High status; CD: Low & Medium Low status). All groups were
of mixed gender.

Age

1825

2640

40+

Socioeconomic
status

Group 1
Group 3

Group 2
Group 5

Group 4
Group 6

AB
CD

When recruited, participants were told that they were

being invited to take part in discussions about energy
issues, and that the work was being sponsored by the
European Union. Participants were paid an honorarium for taking part in the groups and participating in
the interval exercise.

THE DATA ANALYSIS

Our analysis of data generated by the groups is still

taking place, as we prepare this conference paper.
Therefore, by necessity, here we can only discuss our
approach to data analysis, and to present some highly
preliminary observations on the implementation of the
methodology.
In order to analyse the data, an approach based
essentially upon the technique of analytic induction
has been adopted (Bloor, 1978; Silverman, 1993). We
have also drawn on some elements of grounded theory,
in the sense that our analysis has been open to emergent findings (Dey, 1993; Seale, 1999). The process
also utilises a broad approach to discourse analysis,
which is sensitive to both the detail of conversational
interaction, and the resources deployed in such talk

1372

(Antaki, 1994; Horlick-Jones et al, 2007a; Myers,

2007). Additionally, as the group process has generated not only qualitative but also some quantitative
data, the research team may have an important opportunity for comparisons and data triangulation (Fielding
and Fielding, 1986).
Finally, the group process is also being evaluated
as an engagement exercise. An evaluation questionnaire was administered to participants at the end of
the second meetings of each group. This questionnaire
uses the Rowe-Frewer (2000) framework for evaluating the effectiveness of participative and deliberative
processes (see also detailed discussion in HorlickJones et al, 2007b). The process evaluation will
draw on an analysis of the questionnaire data together
with observational notes on the conduct of the group
meetings.
5

PRELIMINARY INSIGHTS

This section sets out some very preliminary impressions on the discussion and learning processes that we
staged, based on observational notes. It is important
to bear in mind the status of this material. These are
initial sketches that need to be checked, confirmed (or
possibly rejected) in the next steps of the data analysis
process, as described in the previous sections.
As might be expected, and in line with empirical
evidence on the lay perception of nuclear fusion, there
was a general and widespread lack of awareness and
knowledge about the technology among our research
participants.
In an effort to encourage a preliminary discussion
about fusion and with the aim to explore lay perceptions in the light of generic and contextual information,
a very short news article was given to participants.
Even after reading this short news article, most participants felt that they were still not able to discuss fusion
(How am I going to talk about it if I still do not know
what it is?). It could be said that fusion was perceived
as very obscure, not only in spatial or temporal terms,
but also from a conceptual perspective, as it appeared
to be far from my understanding and difficult to
be dealt with.
According to the first impression of the observers,
further reactions to this short news article included
a combination of surprise and perplexity. Participants
were surprised to find out that such an important scientific project was going on and they knew nothing about
it. Some participants experienced this surprise as a
positive issue (scientific progress, hope, etc); while
others were sceptical (is this real?) or incredulous.
As far as the learning process is concerned, the
first point to be made is thatin some groups
participants seemed to have considerable difficulties in coping with the technical knowledge. At first

sight, only some groups seemed to be able to deal

with the technical details. Nevertheless, it seems
thatafter the discussion and the reading of the informative materialssome changes took place in our
participants modes of reasoning: from a simple categorisation (what is the nature of this thingthat
we are discussing) to one based upon attempt to
infer aspects of fusion from contextual information
(structural-calculative reasoning: see Horlick-Jones
et al, 2007a).
A detailed analysis of the main corpus of the data
(group discussions and diaries) will be necessary to
properly address the effectiveness of the informative
materials that were provided to participants at the end
of the first session. At the current stage of familiarization with the data, one could say that the perceived
discrepancies between the informative materials produced by the fusion community and by the environmental groups generated perplexity, scepticism and
confusion among our research participants.
We are interested in investigating what we have
termed elsewhere nuclear branding (Prades et al,
2008): a labelling of fusion technology with negative
associations, connected with the shortcomings of the
fission programme. We expected to find a high prominence of such framings and associations, especially
during the first sessions of the groups, when levels
of awareness and understanding about fusion were
expected to be low. According to the observers notes,
it seems that some groups were more sensitive than
others to the nuclear branding effect. Perhaps surprisingly, whilst in some of the groups the nuclear branding
effect was not as prominent as expected during the first
sessions, in others it became even more pronounced
during the groups second sessions of discussion.
Finally, with regards to the possible relationship
between increasing knowledge about fusion and its
acceptability, at this stage of the data analysis process (familiarization with the data), it seems that the
arguments underlying the acceptability (or otherwise)
of fusion seem to change through the learning and
discussion process.
6

EVALUATION OF THE PROCESS

As mentioned earlier, at the end of the second meetings of each group, a questionnaire was administered
to participants. The aim of the questionnaire was
to evaluate the effectiveness of the group process
as an exercise in citizen engagement. The questions
dealt with issues such as perceived representativeness,
impartiality, freedom to express personal opinions,
understanding of the process and level of interest of
the group discussions. All the evaluation items were
rated very highly by participants. They perceived that
the process was unbiased, interesting, that it allowed

1373

expressing their opinions, and that their role in the

process was clear. The item rated highest was the perceived freedom to express personal opinions. So, the
level of participants satisfaction with the engagement
process was very high.
7

DISCUSSION

Our preliminary assessment of the group-based investigation is that the methodology has worked well;
in that it is showing its capacity to generate quasinaturalistic collective discourses, and to promote a rich
discussion and learning process among our research
participants. It is also worth mentioning that the
method has demonstrated its capacity to generate a
significant amount of rich qualitative and quantitative
data.
ACKNOWLEDGMENTS
The work reported in this paper was supported by the
European Fusion Development Agreement (EFDA)
(contract TW6-TRE-FESS-B1) and UKs Economic &
Social Research Council (award RES-000-22-2664).
Parts of this work, supported by the European Communities under the contract of Association between
EURATOM/CIEMAT, were carried out within the
framework of the European Fusion Development
Agreement. The views and opinions expressed herein
do not necessarily reflect those of the European
Commission.
REFERENCES
Antaki, C. (1994) Explaining and Arguing: the Social
Organization of Accounts. London: Sage.
Bloor, M. (1978) On the analysis of observational data: a
discussion of the worth and uses of inductive techniques
and respondent validation, Sociology, 12(3) pp.545557.
Bloor, M., Frankland, J., Thomas, M. and Robson, K. (2002)
Focus Groups in Social Research. London: Sage.

Dey, I. (1993) Qualitative Data Analysis: a User-Friendly

Guide for Social Scientists. London: Routledge.
Fielding, N. and Fielding, J. (1986) Linking Data, Qualitative
Research Methods Series No. 4. Newbury Park Cal.: Sage.
Horlick-Jones, T. (2005) Informal logics of risk: contingency and modes of practical reasoning, Journal of Risk
Research, 8(3) pp.253272.
Horlick-Jones, T. (2008) Reasoning about safety management policy in everyday terms: a pilot study in citizen
engagement for the UK railway industry, Journal of Risk
Research, in press.
Horlick-Jones, T., Walls, J. and Kitzinger, J. (2007a) Bricolage in action: learning about, making sense of, and
discussing issues about genetically modified crops and
food, Health, Risk & Society, 9(1) pp.83103.
Horlick-Jones, T., Walls, J., Rowe, G., Pidgeon, N.,
Poortinga, W., Murdock, G. and ORiordan, T. (2007b)
The GM Debate: Risk, Politics and Public Engagement.
London: Routledge.
Hughes, R. (1998) Considering the vignette technique and
its application to a study of drug injecting and HIV Risk,
and Safer behaviour Sociology of Health and Illness, 20
(3) pp.381400.
Myers, G. (2007) Commonplaces in risk talk: face threats
and forms of interaction, Journal of Risk Research, 10(3)
pp.285305.
Prades Lpez, A., Horlick-Jones, T. Oltra, C. and Sol, R
(2007) Social Perception of Large Research & Development Programs: A review of Social Science Research.
Coleccin Documentos Ciemat. Madrid: Ciemat.
Prades Lpez, A., Horlick-Jones, T. Oltra, C. and Sol, R.
(2008) Lay perceptions of nuclear fusion: Multiple
modes of understanding. Science and Public Policy in
press.
Rosenhead, J. and Mingers, J. (eds.) (2001) Rational Analysis
for a Problematic World Revisited: Problem Structuring Methods for Complexity, Uncertainty and Conflict.
Chichester: Wiley.
Rowe, G. and Frewer, L. (2000) Public participation methods: a framework for evaluation, Science, Technology &
Human Values, 25(1) pp.329.
Seale, C. (1999) The Quality of Qualitative Research.
London: Sage.
Silverman, D. (1993) Interpreting Qualitative Data: Methods
for Analysing Talk, Text and Interaction. London: Sage.

1374

Safety culture

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Us and Them: The impact of group identity on safety

critical behaviour
R.J. Bye, S. Antonsen & K.M. Vikland
Norwegian University of Science and Technology, Studio Apertura, Trondheim, Norway

ABSTRACT: This paper discusses the significance of identity on safety critical behavior. Investigations of
accidents and near misses find that the majority of incidents may be seen as a result of human failure. These
wrongdoings are often described as quiet aberration or short cuts, which can be said to be standardized or
institutionalized actions. In This paper the concepts of identity and identification is used in an attempt to explain
how standardized quiet aberration or short cuts is established and maintained in an organization. Identities
are seen as products and producers of relations and belonging conventions. We propose that identity is an
important aspect of safety because it conveys models of self and others as well as models of how work should
be performed.
1

INTRODUCTION

Safety management has traditionally been preoccupied with the formal aspects of organizations. In the
last decades, however, the search for ways to improve
safety has expanded to include also the informal
aspects of organization. The concept of safety culture
has been subject to much research and most recently,
the role of interpersonal trust has been thoroughly
investigated in a special issue of Risk Analysis1 . In
this paper we are going to direct attention to a topic
which has been largely neglected in both these strands
of research, but which is nevertheless crucial in understanding the way safety-critical work is performed.
This is the issue of group (or occupational) identity.
The goal of this paper is to explore different ways identity may influence a groups work practice, and thereby
also influence safety. We argue that, in terms of safety
critical work practice, identity offers explanatory force
because, firstly, identities may be seen as constituting models of selves, secondly by involving models
of others, i.e. constituting borders between us and
them, and thirdly, by providing models of work, i.e.
conventions for work performance.

WHAT IS IDENTITY?

Identity might be defined as a persons self-conception

(Seymour-Smith 1986). Identity has been a central

1 Risk

Analysis vol. 26 (5) pp. 10971407.

concept in sociology, social psychology, social anthropology, and more recently in organizational studies
(Hatch & Schultz 2004). In social science, the term
has been applied and developed through studies of
human interaction by Cooley (1902), Mead (1934) and
Goffman (1959).
Group identity or collective identity is usually seen
as distinct from individual identity [9]. Brewer &
Gardner (1996) claim that personal identities differentiate the self from the others. Collective identity,
on the other hand, reflects assimilation to a certain
social group. Both personal and collective identities
are dependent on interaction with other people, by
emphasizing similarities and dissimilarities with other
individuals. Seen as a product of interaction, the term
identity has been criticized because it indicates a fixed
stable state. Due to this some writers prefers to use the
term identification rather than identity to underline the
transitory characteristic of the phenomena. According
to Weick, (1995) changes in interactional patterns or
changes between interactional relations implies a shift
in definitions of selves. An individual may as a member of an organization hold a portfolio of overlapping
identities that are mobilized in different interactional
situations.
In a working situation, as in any other social situation, humans construct identities. This construction
process involves a notion of we-ness (Hatch &
Schultz 2004), which defines a certain community
of people. The constitution of identities may partly
be seen as a result of an orientation towards shared
problems (Lysgaard 1961). In an organization these
identity defined communities may coincide with formal work groups in the organizational design, such as

1377

the we-ness of a department, a certain task force

etc, but identity may also constitute communities and
belonging activities that run across the formal design
of an organization, e.g. the workers or the craftsmen. Group identity is thus a central element of the
informal organization.
In addition to defining the boundaries of a community, the identities of a working community or
occupational community also imply a set of characteristics of the group members, and these characteristics
often constitute elaborate ideologies which itemize the qualities, interests, and capabilities of those
identified (Becker 1971:179).
Moreover, identities influence how people make
sense of information about the world. In the words
of Weick (1995:20): Depending on who I am, my
definition of what is out there will also change. This
implies that the concept of identity not only deal with
the self-concept, the we-ness of the community, but
also include a more extended system of beliefs, i.e.
what we may call worldviews. The self-image of
a community reflects and reproduces the definitions
of what is out there. For instance, organizational
studies of informal groups have shown how shop-floor
workers construct holistic images of the company that
may contradict with managers view (Lysgaard 1961,
Crozier 1964).
Identity also influences how people act. By taking on a particular identity, people tend to behave in
a manner that is proper according to the notion of
we-ness, i.e. as a proper worker you perform your
tasks in a specific manner. For instance, a certain attitude, way of thinking or work practice is legitimized
due to this sense of we-ness. Identities represent
rules for how to interpret and act in specific situations. Due to this identities contribute to a behavioral
alignment; the development of a set of shared conventions between people who share a notion of common
identity. Furthermore, this behavioral alignment also
contributes to a potential sanctioning of behavior that
is not in accordance with the notion of proper conduct.
By deviating from behavioral norms, one is running
the risk to exclude oneself from the community.
According to Wenger (1998), identities serve as
trajectories for learning and change. This means that
identities limit and direct the activities of a certain
community. Attempts of non-members of the identity constituted community to influence the behavior
of its members are likely to fail. Especially, introductions of demands that are incoherent with the collective
self image of a group, will probably have a minimal
success rate. Introduction of demands from the others
may actually strengthen both the notion of we-ness
and the behavioral alignment that are considered as
threatened by the others.
By constructing a notion of a collective, an
us, one also creates distinctions between us

and them. The border marking aspects of identity

have by some authors been addressed as ethnicity
(Tambs-Lyche 1976). We use the term ethnicity to
denote the process of border marking between
groups of people. This is somewhat different to the
traditional meaning of ethnicity as referring to groups
of different national or tribal origin.
Identity and ethnicity is expressed by a certain
repertory of signs, such as for instance a specific type
of behavior, language, dress code, usually named as
identity marks. By sharing certain identity marks one
express the we-ness and ones membership in a certain community. The identity marks can also express
the uniqueness of the community. The uniqueness of
identity marks, can in relation to other groups, enable
a bilateral construction of the typical other, or what
we with reference to Tambs-Lyche (1976) may call the
construction of stereotypes. A stereotype is socially
constructed and is composed by certain signs and
symbols that are interpreted as ethnic, and given
a certain meaning. Such stereotypes appear in pairs.
A community A has its stereotype of a group B and
group B has its stereotype of group A. A certain community, for instance a community of blue collar workers will have their typified images of other groups,
e.g. white collar workers, and the other way around.
The exclusive attributes and signs of the others are
interpreted as expressions of the stereotypical images
of the others, i.e. the others are attributed qualities that are symbolized by certain common attributes
among the others. For instance, a specific dress
code shared among a group of others become a
symbol and an expression of the typical others.
The process of border marking may not necessary
imply a hierarchy of rank between the groups, but hierarchies are often present in most social situations, and
have an impact on the identification process. Formal
hierarchies based on responsibility and authority in a
company is often accompanied by informal constituted
hierarchies between formal or informal communities.
The informal hierarchies are likely to be based on additional criteria, such as e.g. a notion of importance,
cleverness trustworthiness etc. When communities see themselves in competition with each other, the
experience of rank and power does not necessarily correspond with the expectations. Such mismatches can
generate very different actions among the group members (see e.g.) that not necessary are in accordance with
the official activities and goals of the organization.

CASE DESCRIPTIONS

This paper is based on two case studies in the

Norwegian petroleum industry: 1) offshore oil and gas
installations and 2) offshore service vessels. The work

1378

groups we are focusing on in these two settings are

involved in very different work processes.
1) Oil and gas installations in the North Sea differ
in terms of physical construction, size and age. The
larger installations, like the ones studied here, usually have 200 to 350 workers onboard. The majority of
those working on a platform have a two weeks shift
with twelve hour watches, followed by four weeks
off. The production of oil and gas involves several
occupations and work groups. In this paper, large production installations are the context with a focus on one
occupation, process operators, as an identity defined
community. The kinds of installation we here are using
as the empirical example have approximately 26 process operators onboard the platform, divided between
day- and nightshifts. The process operators core tasks
are to observe and control the production of oil and
gas on the installation. The systems they are working
with are closed, which means that the process operators do not directly see the potential dangers related to
their work. They detect dangers or things that need to
be fixed by alarms and computers (Nvestad 2008).
In the central control room, 23 process operators
supervise every system at the platform. Other process
operators assist the control room by making observations of their designated area out on the installation.
Process operators out in the plant communicate with
the central control room by radio.
2) The offshore service vessels provide the oil and
gas installations with services related to supply, anchor
handling and emergency preparedness. The ships are
usually manned by 1015 seamen, working six hour
watches for 28 days at a time. The supply vessels call
on port around three times a week, while the standby
vessels sail for the full 28 days. The anchor handling
vessels have a more varying sailing schedule, but it
is not uncommon for these vessels to sail for several
weeks without coming to port.

on three installations in the North Sea. During this

period one researcher had two weeks shifts followed by
four weeks off. The shifts were divided between three
installations, where the researcher observed various
work operations and took part in simple work tasks. In
addition, informal interviews were conducted, covering themes from work processes, safety to the life as
a platform worker. Being part of the social environment where the actual work was performed yielded an
understanding of both explicit and tacit dimensions of
work and community.
2) A quantitative survey on offshore service vessels
(N = 493 response rate 56, 7). The survey data was
analyzed through analysis of variance and frequency
distributions.
3) Interviews with employees on offshore installations and service vessels. Platform workers were interviewed as part of 1) a causal analysis after a serious gas
blowout. 152 people from different parts of the organization, both onshore and offshore, were interviewed
individually and 2) as part of an evaluation of an organizational changing process with an alternation from
occupational departments to integrated teams on three
installations. 726 persons were interviewed, covering
all the shifts on these three platforms.
4) Accomplishment of search conferences and
workshops aiming at safety improvement on installations and service vessels.
All these sources of data have been analyzed by
looking for 1) the way the workers perceive their own
workwhat constitutes a good seaman or what
does it mean to work on a platform, 2) Identity marks
that express belonging and 3) linguistic categories that
define us from others. The description of identity presented here constitutes patterns of concurrent
descriptions of these three analytical dimensions.

5
4

RESEARCH METHODS

Collective identity is hardly a concept that can be

easily operationalized into clear-cut analytical dimensions. Since a groups identities may consist of tacit
dimensions, to be able to describe this identities
requires an in-depth qualitative approach. At the same
time, the desire to analyze larger communities requires
a more wide-angled approach. The study of collective
identity, therefore, requires the use of multiple methods, both quantitative and qualitative. It also requires
the researcher to take the time to get to know the field
of study.
Our analysis relies on four different sources of data:
1) Ethnographic field work/participant observation
was carried out on offshore installations. The fieldwork was carried out over a period of seven months

FINDINGS

5.1 Identity on an oil and gas platform

occupational community
Our empirical findings from the installations show that
the identity as an offshore worker is of current interest in contexts outside the installation, and it is very
often people that do not work offshore themselves that
use the label offshore worker. The workers, on the
other hand, are more likely to define who they are
related to the actual work they are doing, in terms of an
occupation or a working community: I am a mechanic
working on a platform; I am a driller working offshore etc. When interacting with people working on
other installations, the object of identification can be
either the department, installation or the company.
On larger oil and gas installations the platform organization was traditionally divided into occupational

1379

departments (mechanical department, logistic department etc). In recent years there have been organizational changes which have altered the organizational
structure from departments to integrated teams, with
a mix of occupations in each team. At least this is the
case for installations in our empirical material. Even
so, our data show that the primary identity defined
communities are connected to the occupational departments, and not to the teams. For the process operators
it is still the production department with the rest of
the process operators, which continues to be the community they identify with. They have a strong notion
of we-ness and as an occupational group, the process operators generally think that they are doing the
most important work on the installation: Without us
the platform will stop, we own the equipment, are
statements from process operators that underline this.
These expressions are strengthened further by the fact
that process operators are assisting the central control
room, which is defined to be the heart on the platform both when it comes to normal operations and
emergency situations. And to underline the importance of the control room and the role of the control
room operators, the control room operators have to
have thorough knowledge of every system and area on
the platform, due to their work in watching over and
controlling the whole oil and gas production on the
installation.
The radio that process operators carry with them
to communicate and report directly with the control
room, when doing their observing and controlling out
in the plant, is an identity mark that underlines this
responsibility and identification with the community
together with other process operators. In addition to
the flagmen assisting the crane operators on deck, the
process operators are the only occupational group that
is carrying a radio out in the plant.
Process operators look upon themselves as being
strong personalities with an individualistic attitude.
These characteristics are to a great extent attributed to
the responsibility the process operators have in watching over and controlling the oil and gas production
process.
The expressions of we-ness among process operators can also be looked upon as a set of shared
conventions that represent attitudes and definitions on
how to behave in a working situation. Due to their role
in controlling and watching over the production, there
is a prevailing understanding among process operators
that they are observing a working process, i.e. when
mechanics are changing a packing on the well head.
An important aspect in this observing and watching
over the production is that the process operators are
handling gas under high pressure in their work. Most
of this work is done by checking out what alarms and
computers have been detecting. But during the fieldwork there were many stories told by process operators

about colleagues that had this excellent nose which

was able to smell the gas out in the production plant,
before it was detected by alarms, computers and gas
detectors. These intuitive qualities are highly respected
among the occupational community of process operators, and are regarded as a combination of expert
knowledge and personal skills. This can be seen as
tacit knowledge due to years of practical experience
being a process operator on a platform. It is knowledge and judgments that represent a contrast to safety
management systems, which are seen as attempts to
reduce the leverage for necessary improvisation and
professional judgement. The process operators view
themselves as the experts of the installations. Maybe
this is why process operators express a more critical stance towards their safety procedures than other
occupational groups offshore (Nvestad 2008:521).
A possible consequence is that introduction of new
procedures from the management will not be accepted
and followed up: As long as the community of process
operators regards these new procedures as unsuitable
according to their notions of how the work should be
done and a notion of their own superiority regarding
knowledge, there will probably be an acceptance in
the community to disregard them, i.e. an acceptance
of quiet aberration.
5.2 When are the occupational community
made explicit?
All work that is to be executed on the process plant has
to have a work order and a work permission signed
by the process operator responsible for the specific
area, before it can be started. Every work order that
is completed must also be signed by the responsible
process operator for the specific area to be closed.
Therefore, one might say that the community of process operators is made explicit in the interaction with
everyone else doing maintenance work on the process
plant. Everyone need the approval from process operators to do their work, and this can sometimes involve
waiting if there are no process operators available to
sign the document.
The process operators role in signing, observing
and controlling attribute hierarchy of rank. From the
process operators point of view they are the most
important occupational group on the installation due
to the content of their work. Observing and signing
work orders and permissions are symbols of power.
This self concept may not only generate a hostility
towards new procedures, but also limit the interaction with other work groups which the community of
process operators rank as inferior compared to them
selves. This may have crucial impact on the information flow from other work groups, which in certain
situation may be crucial in their work of running the
production process.

1380

5.3

Perspectives on hazards and safety

Offshore oil and gas production involves great potential for hazardous situations, both in terms of occupational accidents, major accidents and environmental
damage. On an oil and gas installation there is gas
under high pressure, which represent a potential for
leakage and explosions and must therefore be handled with care and knowledge. Process operators to
a great extent have the responsibility to monitor and
control these hazards. Our findings indicate that the
process operators take this responsibility seriously,
and they are fully aware of the potential crisis if
hydrocarbons go astray. Due to this, the process
operators emphasize the importance of knowing their
designated area on the installation. The process operators have an awareness of their responsibility for
the potential crisis related to their work. They are
handling hydrocarbons under high pressure which,
if something goes wrong, have the potential resulting in gas leakages, fires, explosionshazardous
situations with consequences for all the 200 to 350
workers onboard the platform, in addition to environmental damages. The awareness and responsibility
for these potential big crises are to a great extent
representing the process operators interpretation of
hazards.
Their focus on hazardous and their responsibility
are used in the interaction with other work groups
on the platform to legitimate their informal rank as
superior to the others. Any how, this view and attitude towards the other limits the information flow
between the work groups, and may lead to bounded
information regarding e.g. the state of the production
systems.
5.4

Identity on offshore supply

vesselsBeing a good seaman

The survey was intended as a broad monitoring of the

organizational aspects of safety on the service vessels,
but included one item that is highly relevant to the
purpose of this paper. The respondents were asked to
state five words that they felt described the term good
seamanship. 402 persons (81,5 per cent of the sample) answered this question. The notion of seamanship
has previously been a described as the very definition of the properties a good seaman should possess
(Knudsen 2005). The words used to specify this notion
can therefore be taken to express core characteristics
of the seamens identity.
There is, of course, great variety in the words
402 seamen use to describe the notion of seamanship. However, a rough categorization of the description reveals a highly concurrent pattern. The most
frequently mentioned aspects of good seamanship
related to the maintenance of the social community

on the vessels. Words like community, trust, support,

cooperation and well-being were mentioned by 281
respondents. The second most frequent category pertains to the ability to work safely. This must be seen in
relation to the third largest category of descriptions,
which relates to the ability to make wise and cautious decisions. The ability to make sound judgments
of work related risks falls within this category. Other
characteristics that stand out as central parts of the
notion of seamanship are being independent, responsible, getting the job done, being able to take care
of novices, and in particular: having sufficient practical work experience. The ideals conveyed through the
notion of seamanship is thus of an autonomous man
(the ideal is predominately masculine) who is able to
make safe decisions, based on his own professional
skill and judgment. Furthermore, the ability to act in
such a way is not something which is learned through
formal education alone, it is based on having practical
experience. Interestingly, there are few differences in
the words used to describe the notion of seamanship
by the different groups on board. The characteristics
of a good seaman thus seem to be the same, irrespective of whether the seaman is a captain, an engineer or
a sailor.
In addition to analyzing the words the seamen
use to describe their ideals of work, it is also worth
considering the words that are not used, and which
consequently fall outside the notion of seamanship. 44
of the 402 respondents mentioned working according
to safety procedures as a characteristic of seamanship,
and only three persons saw the reporting of incidents
as a part of being a good seaman.
This is also reflected in the results on other items in
the survey. For instance, 45 percent agreed that it was
sometimes necessary to violate procedures in order to
get the job done. Also, nearly 30 percent of the respondents stated that near misses and minor accidents were
rarely or just sometimes reported.
The interviews confirm and elaborate this pattern.
Especially, the seamen defined the doing of paper
work to lie outside their conception of a seamans
duties: You know, us seamen, were not so keen on
filling out a whole lot of forms. Thats not where
our job lies. Our job is to sail a boat, were not
office workers whose job is to fill out a bunch of
forms.
This citation, and our data material is full of similar
statements, illustrates two things: First, it shows that
the seamens model of work is centered on practical
work. Relating to regimes of written documentation,
work permissions and other formalized work requirements are not part of what the seamen perceive as
their core tasks. Second, the citation illustrates the way
the seamens identity is expressed in contrast to their
stereotyped image of other groups, here the office
workers.

1381

5.5

Seamen talking about the others

The use of the collective categories of seamen and

good seamanship becomes a topic in relation to
interaction and discussions regarding: 1) newcomers
on board, and 2) actors outside the ship.
The collective categories are made relevant when
newcomers enter the vessels and are supposed to be
integrated in the daily interaction. The newcomers
are evaluated by the existing crew members based on
the notion of good seamanship. The newcomers are
expected to prove themselves and earn the reputation
of possessing good seamanship.
When the crew members interact with actors outside the ship (e.g. personnel in the oil company or the
shipping company) they often refer to themselves as
seamen. This collective and seemingly egalitarian category seems to be used in some situations instead of
the more restricted and hierarchal ranked categories
such as sailor, mate, captain etc. The most general
aspect of being a seaman is to have first-hand experience of working at sea. The most general function of
the word is thus to separate between those who know
what it means to work on a ship, and those who do
not. The counterpart to being a seaman seems to be
an office worker, who has no knowledge about life
at sea. This label is used to denote personnel at both
the oil company and the shipping company, who do
office work, i.e. contracting, development of procedures, HSE management, cost surveillance etc. The
office workers that are not necessarily involved in
the seamens daily activities, but which nevertheless
influence the premises of their work. As we shall discuss further in section 6.2, the fact that the office
workers do not possess seamanship serves to undermine their legitimacy and authority in their attempts to
govern and regulate the way work is to be performed
on the vessels.
To describe people, external to the ship, that are
involved in the daily operative work activities, the
following categories are used: 1) supply base workers, 2) traffic control workers and 3) platform/rig
workers. Members of category 1 and 3 are not necessarily restricted to employees of the operating oil
company. Among members of category 3, the seamen also address more specific groups, such as crane
operators and drilling managers.
Among the different categories of others, the platform/rig workers are especially attributed negative
qualities. This is partly expressed in a rich portfolio
of narratives that are frequently told to newcomers
(and researchers) to explain the vessels status in
the logistics chain. Central themes in these stories
are a) disrespects towards the vessel crew b) focus
on their own interests and c) neglect of the vessels.
One often told story among the crew members on
standby vessels is about a Christmas Eve. The vessel

was supposed to receive mail, presents and food via

the platform. According to the story, they did not
receive any thing this specific Christmas eve because
the platform workers where occupied with their own
grandiose Christmas celebration on the platform.
Several stories are about platform workers that
deliberately throw objects (trash) down on the boats.
These stories are often told when talking about
episodes where rigs and platforms soiled the vessels
with spill water, drilling mud and chemicals, which
happen rather frequently. One of these stories ends
with a description of a platform/rig workers standing many feet above, urinating down on the vessel.
Another group of narratives is about fat crane
operators which do not respond when the vessels
arrive at the platform because they are, according to
the stories, preoccupied with frequent coffee brakes,
meals or naps, There are also several tales about
drilling managers calling the vessels and jaw at the
captain that he should hurry up and do the job according to his preferences. Some of these tales are about
captains that have overruled the drilling managers
demands, due to legitimate reasons (e.g. weather conditions, operational hazards), and that this decision
had resulted in various sanctions from the oil company later on (i.e. reprimands, bad reputation, scold,
loss of contract etc.).
Of course, such narratives may be exaggerated and
changed as they are told. Irrespective of their objective truth, however, the symbolic value for the ones
telling the story remains the same. The organizational
folklore conveys a form of informal knowledge about
the ideals of work and the way the group relate to
other groups. As such, this folklore expresses the
explicit and tacit contents of a groups identity. The
socialization of novices will be done partly through
introducing the newcomer to stories conveying the
basic assumptions of the organization.
5.6 Seamen talking about hazards and causes
for accident and near misses
In the interviews with the vessel crews there were
some major themes that concerned the the crew members. These main themes may be conceptualized as
1) high physical and psychological pressure related to
the work, 2) heavy work load, 3) lack of coordination
between different organizational units of the operating
oil company, and 4) noise and lack of rest due to poor
facilities on board. These conditions are attributed to
The oil company in general: The oil companys
demands to efficiency and sailing time are tremendously high. We are under constant pressure due to
the work speed. We feel that it is a mortal sin to spend
time at the pier. We never know when the telephone is
ringing.

1382

The seamens work situation is commonly compared to the platform workers situation, in terms of
shifts, salary and work load: The motivation decrease
when you look at the big differences between our shifts,
and what they have on the platforms. We have one
month on and one month off. They have two weeks
on and one month off. I feel like a second-rate human
compared to them.

6
6.1

DISCUSSION
Models of selves and models of work

To be a good seaman means to be an independent,

responsible man, capable of making wise and cautious decisions, and taking pride in the ability to get
the work done. At the same time it is expected that
he should be supportive and cooperative with his fellow seamen. Much of these characteristics can also
be said to be fitting descriptions of the process operators on the platforms: They see themselves as the
autonomous experts controlling and watching over the
production processes on the plant. Both the seamen
and the process operators have a strong sense of ownership towards their respective work contents, and work
context (i.e. the plant and the vessel). Although there
are massive differences regarding the types of work
performed by the two groups, there seem to be highly
similar patterns in their models of selves, i.e. the way
they see themselves.
At this point in the paper, it is relevant to ask the
following question: how does all this relate to safety?
The link between identity and safety is to be found
in the fact that the models of selves to a great extent
also represent models of work. Identification processes
serve to align the work practice among the members of
the identity constituted communities by providing ideals for work performance, i.e. being a good seaman
or a good process operator. These ideals form the
basis of social conventions for behavior and communication. In the communities studied here, there seem to
be considerable friction between such informal models
of work, and the prescribed procedures. This friction may, in turn, generate the conventional deviations
often termed by the others as wrong attitude,
quiet aberration or short cuts. This image of an
independent, wise and experienced worker seems to
be in contradiction with literary knowledge, paperwork and written procedures, which especially seem
to contradict with the ability to get the work done.
Why does this happen? We propose that the selfconceptions of being autonomous experts, which
characterize the two communities studied here, are
inconsistent with the attempts of safety management
systems to regulate, control and standardize work
processes in the two communities. As described in

section 5, both the seamen and the process operators

placed much emphasis on having practical work experience as a basis for being a capable worker. This means
that any statements regarding the nature of their work
should be grounded in practical experience, in order
to be taken seriously by the community of workers.
Importantly, the procedures and requirements inherent in a safety management system are often perceived
as not being based on such practical experience: Consequently, demands from safety management systems
do not always correspond with, and may even challenge, the established models of work. The focus on
paper work, incident reporting and work procedures
thus seems to be incommensurable with the established
models of selves and models of work.
6.2 Models of others and worldviews
In addition to constituting conventions for behavior
and communication within the group, a groups identities also directs and limits interaction and information
flow between different identity constituted communities, i.e. between us and them. The seamens
interpretation of being treated as inferior by the oil
company is one example. This interpretation of
the oil company is regarded as undeserved seen
in relation to the seamens models of themselves as
autonomous experts. This comprehension seems to be
used as a premise in the interpretation of the daily
interaction in the operative work activities. By only
receiving orders, getting asked to speed up the work,
getting spill water or mud in their heads, work a month
(not 14 days as them) and getting less paid than them,
becomes symbolic evidence that their comprehension
of being treated as underdogs is correct. These daily
experiences support the mythological stories the crew
members tell each other, which in turn contribute
to strengthen these views. Together this compose a
world view of a system which is predominantly hostile
against the seamen, regarding them as inferior, where
the only concern is that the seaman do their work as
efficiently as possible. This world view of a hostile oil
company seems to strengthen the feeling of we-ness
among the seamen. The adaptation to their definition
of what is out there seems to be to become more
seamen. Based on the dominant world view one may
not expect gratitude from the others. The identity
boundary constitutes a ghetto of world views which
maintains and strengthens the established model of a
good seamanship.
The situation for the process operators are in many
ways opposite to the seamens: in most situations
the process operators are those giving orders to other
groups and they look upon themselves as superior to
other working communities on the platform due to the
long tradition in watching over, observing, controlling
and signing work orders and permissions. This has

1383

also constructed their view of the other occupational

communities on the platform; to put it in a direct manner, the others are doing the dirty work. This
world view became very obvious in the changing process were integrated teams replaced the former strong
production department, and where the process operators were the occupational group that protested the
most against this change. They felt that this was a
treat to their profession and that this was part of a
plan from the management to break up and destroy the
strong occupational community of the process operators. This changing process strengthened the feeling
of weness among the process operators. And in the
same way as the seamen became more seamen, the
process operators became even more process operator
as a respond and adaption to what is out there.
The conception of identity as world-views also
implies another and more abstract relationship
between identity and safety. When identity influences
on ones conception of what is out there, this also
pertains to which risks that can be acknowledged by the
members of an identity constituted community. Identity may therefore, in the words of Turner and Pidgeon
(1997) involve both a way of seeing and a way of
not seeing.

CONCLUSION

The impact of identification processes on an organizational safety level may be considered on, at least,
three levels.
First, group identity serves to align the work practice among the members of the community. Models of
selves are accompanied by a set of conventions about
how to act in a specific situation. This means that models of selves also include models of work, i.e. models of
how work should ideally be performed. These conventions may be more or less congruent with the formal
structures of safety management. Identity thus influences safety by influencing behaviour, e.g. the level
of compliance with procedures. Second, identity not
only involves a model of oneself, but also models of
others. This means that identity also influences the
way different groups perceive others. These relations
may direct and limit interaction and information flow
between different identities constituted communities,
i.e. between us and them. This aspect of identity may influence on safety by either facilitating or
serving as barriers for communication and information flow regarding risks. Third, identities are crucial

ingredients in sensemaking processes. This means that

identities influences on which risks that are recognized
as such by the members of a community. Which risks
are seen or recognized will of course determine
which risks one seeks to prevent.
These three aspects of identity are important pieces
of the puzzle to understand the mechanisms that influence work practice. It should be stressed that identities
may be deeply rooted in the culture of a community. As such, it is not something which can be easily
changed. Nevertheless, group identity is something
which should be taken into account in order to determine which intervention efforts are likely to resonate
with the community involved.
REFERENCES
Becker, H.S. (1971): Sociological work: method and substance, London.
Brewer, M.B. & Gardner, W. (1996): Who Is This We?:
Levels of Collective Identity and Self-representation.
Journal of Personality and Social Psychology, vol. 71,
pp. 837, 913.
Cooley, C.H. (1902): Human nature and the social order,
New York, Charles Scribners sons.
Crozier, M. (1964): The bureaucratic phenomenon, Chicago,
University of Chicago Press.
Goffman, E. (1959): The presentation of self in everyday life,
New York, Doubleday.
Goffman, E. (1961): Asylums: essays on the social situation of mental patients and other inmates, New York,
Doubleday.
Hatch, M.J. & Schultz, M. (2004): Organizational Identity,
New York, Oxford University Press.
Knudsen, F. (2005): Smandskab. Arbejdsidentitet og sikkerhedsregler hos danske smnd. Tidsskriftet Antropologi.
Lysgaard, S. (1961): Arbeiderkollektivet. Oslo, Universitetsforlaget.
Mead, G.H. (1934): Mind, Self and Society, Chicago,
University of Chicago Press.
Naevestad, T.-O. (2008): Safety understandings among crane
operators and process operators on a Norwegian offshore
platform. Safety Science, 46, 520534.
Seymour-Smith, C. (1986): Macmillan Dictionary of Anthropology, London, Macmillan Press.
Tambs-Lyche, H. (1976): On the Nature of Etnich Stereotypes, Bergen, Institutt fors sosialantropologi.
Turner, B.A. & Pidgeon, N.F. (1997): Man-made Disasters.
2nd edition, Oxford, Butterworth Heinemann.
Weick, K.E. (1995): Sensemaking in organizations, Thousand Oaks, SAGE Publications.
Wenger, E. (1998): Communities of Practice. Learning,
Meaning, and Identity, New York, Cambridge University
Press.

1384

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Does change challenge safety? Complexity in the civil aviation

transport system
S. Hyland & K. Aase
University of Stavanger, Stavanger, Norway

ABSTRACT: The paper describes our attempt at mapping the complexity of the civil aviation system, in terms
of how changes affect safety and interactions between actors within the system. Three cases were selected to
represent three distinct levels of the transport system: the civil aviation authority case, the air traffic control/airport
operation case and the maintenance case. Through the complexity perspective, we identified several positive
system characteristics or mechanisms that contributed to maintain and improve safety during change processes,
including a strong safety consciousness, collective coordination among individuals, and safety practices based
on flexibility and knowledge sharing. These system characteristics were strengthened by changes with a positive
perceived influence on safety, such as new technology improving flight safety. However, changes involving
efficiency, merging and relocation were perceived to influence safety negatively by creating conflicts of priorities
and reducing safety margins. The mixed effects of changes represented a common challenge across the three
case organizations.

INTRODUCTION

To understand change it is necessary to understand the

particular characteristics of the system to be studied.
One trait of the civil aviation system is how the practices of various actors in the system, from operators to
regulatory and political bodies, depend on current and
emerging influences from national and international
legislation, regulations and technology (in terms of
methods, procedures and standards), and current and
emerging market conditions (such as deregulation and
changes in competition). Another trait of the civil aviation system is how the core tasktransportation of
people and goods from one destination to another
involves a high degree of cooperation and interaction
between actors at every level, from the task of the aviation technician to ensure aircraft and flight reliability
and the role of the air traffic controller to monitor and
guide air traffic, to the responsibility of the regulator
to ensure that the practices of the technician and controller are in accordance with current safety procedures
and regulations.
In sum, the above traits of the civil aviation system
illustrate a complex system comprised of a number
of human, technological, organisational, and societal
interactions, interfaces and dependencies. To understand the safety implications of changes occurring
within such a complex system, the following two main
objectives have guided our efforts:

1. To explore whether changes in different parts of the

Norwegian aviation system affect the safety work
within each organisation (internally)
2. To explore whether changes in the aviation system affect the relations between organisations that
contributes to aviation safety (crosswise)
According to the objectives, we attempt to map
the simultaneousness of changes and their possible
implications for safety in the Norwegian civil aviation sector, ranging from a legislative level to ground
level operators.

2
2.1

THEORIZING ON COMPLEXITY
What is complexity?

Complexity can be seen as a systemic property, permeating every level and actor of for example the civil
aviation system, from the particular work tasks of the
aviation technician to the changes implemented at a
national and international political level. The following four categories of complexity reflect this view
(Pettersen et al. 2008):
Technological complexity arises from increased specialization, increased interactions, and reduced
transparency that may contribute to hide critical
system connections and dependencies, making the

1385

system more vulnerable (Wolf 2001, Perrow 1984).

Applied at the aviation system, advanced computer
systems may complicate observation and correction
of error signals for the aviation technician.
Work situation complexity is a product of cognitive
aspects of the individual and the particular work
situation or process (Rasmussen 1982, Goodstein
et al. 1988, Woods 1988, Amalberti 1996, Hollnagel
1996). Transferred to the aviation system, complexity arises from how the work situation of
the aviation technician is shaped by interactions
between the particular process (maintenance of
flight engine), user interfaces (visualization capability of tools), the individual itself (experience,
competency, process knowledge), and the context of the individual (organizational structure,
colleagues).
Organisational complexity can be defined in a number of ways. One view is that organizations are
open systems, where individuals, technology, and
structures interact within and outside the organization, seeking ways of adapting to the environment
(Weick 1977, Morgan 1986, Sandelands & Drazin
1989, Drazin & Sandelands 1992). Another view is
that organizational complexity arises from attempts
at achieving higher reliability between technology
and individuals through redundancy of individuals
and systems, and operational/collective sensitivity,
thus increasing not only resilience to errors and
accidents, but also the overall organizational complexity. Within the aviation system this can be seen
through aircraft designs utilizing several computer
systems that support the same functions (technical redundancy), and through pilots and technicians
conferring with one another and thus maintaining
the reliability of these systems (human redundancy
and collective sensitivity).
Political/economical complexity occurs when
impulses and decisions at a macro level influence
the premises and practices at meso (organization)
and micro (operational) levels. Within the civil aviation system this can be observed through changes
in national and international regulations and market conditions (deregulation), facilitated by national
authorities and key international actors (like the
European Aviation Safety Agency, EASA). Combined, the varying origin and changing nature of
decisions and impulses challenge the ability of
the particular organization to adapt, interpret and
adjust, thus introducing political complexity. In
addition, deregulation introduces competition and
new actors, placing increased pressure on profit
margins and operations, giving rise to economical
complexity.
In addition, several case specific theories provide insights into particular complexity aspects that

can improve our understanding of the aviation

system:
a. Snook (2000) sees the accidental shoot down
of two Black Hawk helicopters over Northern
Iraq in 1994 (case) as a combination of interactions between individual and organizational factors. Specifically, due to perceptions of global
procedures as static and impractical local adaptations of rules and procedures had occurred
both at an individual level (pilots of the Black
Hawk helicopters and F-15s) and at an organizational level (practices within the control central).
According to Snook (2000), over time this created a situation where local practices gradually
detached itself from written procedures. Finally,
the complex interactions between locally adapted
practices of several actors both at an individual and organizational level contributed to the
accident.
b. Similar to Snook (2000), Vaughan (1996; 2005)
sees the Challenger accident as a culmination
of individual, organizational, political and economical factors that over time had shaped the
practices of the NASA organization in a particular way. Specifically, NASA had developed a
tendency to normalize technical irregularities and
deviations (individual and organizational aspects).
The organization was also affected by an institutional culture (structural secrecy) and a production oriented culture that prioritized continued
launching above thorough risk analyses (individual and organizational aspects). In addition, the
organization struggled for continued support and
resources from Congress, and a need for publicity (political and economical aspects). According
to Vaughan (1996; 2005), the accident was a
result of complex interactions between all these
aspects.
c. Comparable to the complexity picture provided by
Snook (2000) and Vaughan (1996; 2005), Woods
(2005) identifies five patterns within NASA that
are characteristic of the complexity involved in
the Colombia accident: (1) a focus on production pressure, eroding defences and creating a
drift towards failure, (2) a focus on previous success rather than failure anticipation, (3) a loss
of the big picture due to fragmented and uncertain basis for decision making, (4) a disregard
for new evidence in current risk assessments, and
(5) a focus on the chain-of-command rather than
coordination and overlap between organizational
units. According to Woods (2005), the complex
interactions between these organizational aspects
affected individual decisions and actions within
NASA in an adverse way, ultimately leading to the
accident.

1386

2.2

Constructing a descriptive framework

The complexity discussion above suggests that both

processes specific to the individual organization itself
(e.g. work practices, safety principles) and factors
outside the control of the organization (e.g. national
and international regulations), affect and determine
the impacts of changes on safety and the interactions
within and between actors. Since we are interested in
understanding the complexity of the system as a whole
this implies studying micro, meso and macro levels
(Rasmussen 1997). Hence, we have chosen to operationalise socio-technical complexity using categories
that reflect both individual and systemic properties:
context, history, changes, change perception, safety
perception, and safety practice. Context includes internal aspects (e.g. work environment, technical tools)
and external aspects (e.g. competition, influence from
authorities, international trends), while history covers
habits and patterns of the particular organization and
the system as a whole. Changes can be of both internal
(e.g. priorities of material and personnel) and external
(e.g. national and international influences) character,
while safety perception represents the expressed view
on safety and safety practice and observable safety
activities as they are practiced. Finally, change perception covers the expressed view on changes and change
aspects.

METHODOLOGY

The paper is based on empirical data gathered in three

case studies within Norwegian civil aviation:
The civil aviation authority case consists of 26 interviews with inspectors, advisors and managers. The
object of this study was to describe perceptions of
safety, safety practices and changes.
The air traffic control/airport operation case contains a study of five airports with 126 informants
(interviews), aimed at diagnosing the safety culture
as a means for improvement. The case also includes
qualitative free text data concerning changes and
safety aspects from a questionnaire survey, with
231 respondents (managers, planners, engineers, air
traffic controllers).
The maintenance case was carried out as an
exploratory study of a line maintenance department, with participant observation, 15 interviews
and a number of informal discussions. The goal
was to gain insight into how safety is created and
maintained through work practices at an individual/group level. The case also includes free text
data from the described questionnaire survey, with
283 respondents within maintenance (managers,
planners, engineers, aviation technicians).

The descriptive framework for analyzing empirical

data was achieved by systematic reviewing, structuring
(through categorization), and describing the informants and respondents perceptions of change and
safety aspects in each of the three cases, utilizing quotations and descriptions as close to the data material
as possible. In all cases, research quality was ensured
using trustworthiness criteria (Guba & Lincoln, 1981).

RESULTS

After empirical categorization according to our

descriptive framework, the categories safety perception, safety practice and change perception are used
both within and between the three cases. Below we
present results according to these categories, within
each case first (research objective 1) followed by a
cross case comparison (research objective 2).
4.1 Results within cases
4.1.1 The Civil Aviation Authority (CAA) case
Safety perception: Informants demonstrate a strong
awareness of how formal control mechanisms can
support safety, emphasizing the role of market control (with operators), use of regulations to create
predictability and the desire to standardize the relationship between operators and the authority. However,
many informants describe safety as a general concept,
in terms of being beneficial to society, achieving a
107 risk level etc. A lack of precision in the safety
definitions was supported by informants impressions
that CAA lacked a clearly articulated and communicated safety philosophy, and a collective coordination,
discussion and communication of regulatory practice.
Safety practice: The results suggest that a strong
individual safety consciousness exists within the CAA,
evident from statements like we all have an individual safety policy, we handle safety in a responsible
way and you are never fully qualified when it
comes to safety. On the positive side, these quotations are examples of professional pride, responsibility
and willingness to learn and evolve (professionalism
traits). However, an approach to safety based on the
values and norms of the individual also runs the risk
of facilitating variations in safety practice, as these
quotations demonstrate: the inspection practice is
fragmented, there exists a high individual consciousness, but no collective coordination and the
organization lacks an overall inspection practice that
is standardized rather than dependent of individuals.
Change perception: Informants typically describe
the recent focus on risk based inspections (focusing on identifying and addressing system weaknesses
and risks), opposed to the previous practice of

1387

detailed inspections. The transition to such inspection

methodology has been problematic due to variations in concept definitions and interpretations, a
lack of responsibility, and the creation of a riskbased approach in an organization based on individual
practice and tacit knowledge. The change from a
single-provider (monopolist) market to deregulation,
resulting in increased competition and focus on the
individual airline, marks another important change for
the informants. This change has demanded increased
vigilance and inspection openness/focus, and it has
also created a certain lack of clarity between the
desire of political authorities to lower the requirements for operators within the market and the focus
of the CAA on thorough evaluation of existing and
new operators. International rules and regulations are
commonly discussed by many informants. The transition to EASA (European Aviation Safety Agency)
is described as particularly problematic, since the
existing national legislation has to coexist with the
new EASA regulation for several years. This is seen
as time consuming, further complicated by descriptions of how political authorities restrict the resources
for updates on changes and translation of changes
from English to Norwegian. According to one informant, the problem of being updated is symptomatic
of the many and simultaneous international changes
in general. A fourth important change is the relocation of CAA from Oslo (south of Norway) to Bod
(north of Norway). During the relocation process,
loss of competency and knowledge transfer (tacit
knowledge) occurred when much of the existing (experienced) personnel were unwilling to move to Bod
and had to be replaced by new personnel. Other
informants describe the effects of this situation as
lessened by the professionalism of the individuals
in Oslo, and their determination to transfer existing
competency.
4.1.2 The air traffic control/airport
operation case
In this case, we discuss safety perception related to
the questionnaire survey, and safety practice related
to the study of five airports. In addition, a selection
of change categories identified from the questionnaire
survey is discussed.
Safety perception: To achieve safety within aviation, many respondents focus on all measures and
efforts aimed at reducing or avoiding incidents and
accidents. Other respondents define safety through
the creation of barriers against adverse events, through
reduced risk of incidents and accidents or through personnel, routines and systems that can reduce risk or
affect the safety level. The focus on achieving aviation
safety is also reflected through the view that all individuals, equipment, procedures, rules, etc involved in

the transportation of passengers from point A to B, as

a whole, contribute to aviation safety. The following
quotations illustrate such views: to achieve acceptable flight safety all parts of the organizations involved
must have safety as a first priority and [aviation
safety involves] physical and human aspects that are
relevant to air traffic, and all factors that directly or
indirectly affect task performance.
Safety practice: Several results indicate variations
in the practice of safety. Informants in three of five airports describe procedures as characterized by having
high respect, status and normative value, as illustrated
by one informant: when the work entails routines we
do not know, we consult the procedure. In contrast,
the two remaining airports emphasize everyday work
routines as less ruled by procedures, and where procedures are considered more of a guidance for instance
to drive ways, speed limits and security/access. However, among all airports deviation from procedures
occur due to security interfering with the performance
of everyday work tasks, or due to the need of getting the job done, or due to a lack of training and/or
procedural knowledge. The procedural deviations to
a lesser degree occur among the former three airports. Variations in safety practice are also evident in
how airports handle conflicting goals. Within three
of the airports, conflicts between safety and other
concerns are handled positively through (a) managements way of prioritizing safety, (b) prioritizing safe
job performance above efficiency, and (c) the utilization of necessary resources to support safe practices.
Informants at the other two airports emphasize that
there exists conflicts between safety and the need of
getting the job done, and that they are not being properly resolved. Contributing factors include the need
to resolve traffic load in tight areas to ensure production and turn around time, resulting in procedural
deviations.
Change perception: 8 change categories are identified in the questionnaire survey. (1) Restructuring, (2)
development in reporting systems, (3) development in
technology, systems and equipment, (4) development
in competency, (5) development in quality and management systems, (6) development in international
regulation, (7) development in security, and (8) economy and efficiency. In this paper, category 1, 4, 6 and
8 are discussed to illustrate main results.
Among descriptions of restructuring, the privatization of the air traffic control/airport operation
organization and a recent efficiency program are seen
as negative change influences, in combination with
a negative view on restructuring processes in general
among respondents, in terms of insecurity, unrest, malcontent, stress etc. Negative associations also relate
to the relocation of the CAA to Bod, the merging
of airlines and the reduction of air traffic control stations. A positive restructuring process emphasized by

1388

many respondents is the creation of the CAA as an

independent organization. Other positive associations
to changes include competence development, such
as introducing requirements to practical and theoretical competency updates and testing through courses,
and implementing crew resource management (CRM)
programs. Positive descriptions also include development in international regulations in the direction of
harmonization of rules and procedures and more
uniform use of procedures and systems. According to one respondent, the focus on standardization
has resulted in improved predictability of air traffic flow. Of negative change influences the focus on
economy and efficiency are particular strong among
respondents, including views that a recent efficiency
program (downsizing, cost-cutting) have shifted the
priorities in direction of economy, creating a conflict
between economy and safety in general and specific
issues like lack of education and a double moral (speak
safety but act opposite).
4.1.3 The maintenance case
In the maintenance case, we discuss safety perception
related to the questionnaire survey, and safety practice related to the line maintenance study. In addition,
a selection of change categories identified from both
studies is discussed.
Safety perception: Many respondents see aviation
safety as a product of the entire system, meaning
everyone involved in the process of achieving a safe
flight from A to B. This attitude is supported by
descriptions such as quality in all parts of the process, same attitude in all areas and focus on
quality/safety in all parts, for employees, customers
and authorities. Respondents, and in particular aviation engineers, also define aviation safety as performing work tasks in accordance with regulations,
procedures, documents and routines from authorities,
airlines and aviation fabricants. This focus on technical diligence is supported by respondents valuing
task qualification, training and motivation. Overall,
the value of a safe flight and the protection of human
lives and materials involved are strongly expressed by
all respondents.
Safety practice: Within line maintenance, informants describe three distinct mechanisms that support
the safety practices in the organization. One is the
utilization of practical and experience-based informal knowledge, where technicians through trial and
error seek the right person with required knowledge
to address grey zones in formal documentation, for
example in terms of resolving what a components
normal situation is. In other words, informal competency and practical skills compensate for grey zones in
formal documents, procedures and thus a lack in the
basis for decision-making. Another safety mechanism

is the use of slack resources, where experienced

conflicts between work pressure and safety concerns are resolved by the technicians by valuing and
employing a slower pace to address safety issues
and increase reliability, particular in critical situations, even though this may lead to flight delays.
The third mechanism, flexibility, is a product of
the two former, in that the focus on slack resources
and knowledge seeking creates a work organization
with flexibility and ability to judge errors and solve
problems.
Change perception: The following 7 change categories are described by informants and respondents:
(1) Restructuring, (2) development in reporting systems, (3) development in technology, systems and
equipment, (4) development in competency, (5) development in international regulation, (6) development in
security, and (7) economy and efficiency. In this paper,
categories 1, 2 and 3 are discussed to illustrate main
results.
Within the line maintenance organization, several restructuring processes seemed to affect safety
practices. The organization went from being a technical division integrated in an airline to sale and
merger with a competing airline as a result of falling
income and increased competition. As a result, the
line maintenance organization became part of a larger
independent technical unit separate from the airline.
A consequence of the separation from the airline was
reduced communication between the line organization
and the maintenance control center of the airline, a
coordinating unit responsible for overview and specialized information flow. Reduced communication in
turn affected the competency base of the line organization, reducing access to experience, knowledge and
resources that were crucial elements in technicians
safety practices. According to several informants,
fewer refreshing courses also followed the restructuring processes, reducing technicians knowledge on
both technological developments and on new failure
types and solutions. Finally, restructuring of the organization also meant that the line base was relocated
from the airport itself (3 minute drive), forging a physical detachment between line technicians and operative
personnel and thereby reducing the previous ability to
exchange information between technicians and pilots.
As a contrast to the negative changes within the line
maintenance organization, respondents from the questionnaire survey describe improvements in routines for
reporting and reporting in general and improvements
in technical (computer-based) solutions for reporting.
These changes seem to have affected reporting and
follow-up positively. Technical developments within
other areas, like new equipment in airplanes and
helicopters and new airplanes, navigation and radar
systems, are also perceived as safety enhancing by
respondents.

1389

4.2

Results across cases

4.2.1 Safety perception

In all three cases procedures, rules and regulations are
seen as beneficial to safety, although the understanding differs between informants and respondents in each
case. For example, within the civil aviation authority
case and the air traffic control/airport operation case
rules and regulations are seen as a market stabilizing mechanism and an important factor in achieving
a safer air traffic respectively, while within maintenance the work on safety is concerned primarily with
how regulations, procedures and documents regulate
job descriptions and practices. Despite variations, all
informants and respondents share common descriptions and definitions of safety as ensured through (a)
the absence of incidents or accidents and (b) the common focus on achieving a performance or safety level
of flight traffic that do not pose a threat to individuals and materials. The main difference between cases
are evident from the more generally oriented understanding of safety within the civil aviation authority
case (beneficial to society, 107 risk level) and the
more specific understanding of safety in the air traffic control/airport operation and maintenance cases
(what safety should include and how to achieve it).
These variances in safety understanding can naturally
be attributed to the supervisory tasks of the former and
more specific tasks of the latter cases.
4.2.2 Safety practice
The enactment and practice of safety in the civil aviation system as a whole have characteristics that are
common for the three cases. However, the characteristics also involve aspects that are unique to the
particular case. Among the shared characteristics, the
strong sense of responsibility, professionalism and
safety awareness are common occurring traits across
the cases, contributing to a higher understanding of
safety challenges as well as aiding the integration of
safety in actual work tasks. However, the civil aviation authority case differs from the two other cases,
in that the safety responsibility has a tendency to be
dependent on the individual rather than the collective. Another shared characteristic is the high respect,
status and normative value of procedures, rules and
regulations, making these aspects an important part of
the informants and respondents safety perceptions
and practices in all three cases. However, all cases
demonstrate that the transition from the formal system (rules, regulations etc.) to practice is problematic,
caused by for example problems of enactment (procedural deviations), sometimes insufficiently defined
procedures (grey zones) and conflicts between safety
and other priorities (conflicting goals). To manage the
transition from theory to practice, all case organizations demonstrate informal mechanisms that aid the

transition process. For example, in line maintenance

practical experience transfer, easy access to individuals and competencies and learning across organizational units have facilitated improvisation, adaptation
and flexibility as means of handling increased time
pressure and conflicting priorities.
4.2.3 Change perception
The perception of changes both varies and co-varies
across the three cases, where for example changes
related to technology are given solely positive associations while changes related to restructuring almost
exclusively are viewed negatively. In the middle of
these extremes, the change categories of competency and international developments receive mixed
descriptions by respondents. Specifically, negative
perceptions of restructuring changes are attributed to
relocation of the civil aviation authority and to processes of efficiency and airline merging, while positive
perceptions of changes to technology are attributed
to improvements in reporting systems, equipments
and tools. Furthermore, in all three cases downsizing,
cost-cuts and increased demands on profit represent
negative associations to economy and efficiency. Concerning mixed change impressions, these relate to for
example how competency in air traffic control/airport
operation is seen both as beneficial to safety (via
updates, courses, CRM) and as a threat to safety
(lack of education). Similarly, international developments of rules and regulations are described both as
a strength to safety in the air traffic control/airport
operation and maintenance cases, improving standardization and harmonization, and as a negative influence
on practices in the civil aviation authority case, creating conflicts between existing and new regulations in
terms of translation delays and resources required to
update legislation.
4.3 Lessons learned
From the results within and across the three cases,
we identify four lessons of importance to the civil
aviation system:

The management of change and safety is achieved

through a combination of both formal rules and
procedures and informal mechanisms.
An individual approach to safety is important to
handle specialized roles and functions, while a
collective approach improves change coping, coordination and goal awareness among individuals
and organizations.
The focus on professionalism and professional
pride helps individuals and organizations to adapt
to changes and stabilize consequences of changes
(e.g. the safety mechanisms within the aviation
maintenance case).

1390

 Processes of restructuring, such as relocations,

mergers and downsizing, represents a challenge to
safety in all three cases, and requires higher awareness both by the individuals and organizations
initializing them and by those affected.
5

ANALYSIS

Based on the results, we discuss the different complexity categories within our civil aviation transport
case. In addition, we point to aspects our study shares
in common with previous case studies within aviation
(Snook 2000, Vaughan 1996; 2005, Woods 2005).
Technological complexity and work situation complexity are best illustrated through the aviation maintenance case, where technicians interact with specialized
(aircraft) computer systems and equipment. To handle
this complexity, the case organization utilizes practical
and experience-based knowledge and resource slack.
This flexibility means that when the organization faces
a new situation where systems or components act in
unforeseen ways, potentially hiding critical system
functions and dependencies, the right person for the
particular task is searched for throughout the organization and across organizational boundaries. In other
words, the informal coping mechanisms of seeking
individuals with specific knowledge and/or of reducing the pace of a particular operation help improve
the transparency of the interface between individuals
and technology. This reduces work situation complexity and thereby also prevents possible unintended or
unwanted side effects of the particular operation
that might adversely affect safety.
All cases show organizational complexity elements
that fall within the definition of organizations as open
systems, where individuals, technology and structures
interact within and outside the organization, seeking
ways of adapting to the environment. For example, in
all cases individuals and organizations seek adaptation
to changes in national and international regulations
and legislations, by integrating new safety protocols
in current operations (the air traffic control/airport
operation and line maintenance cases) or by implementing additional rules and regulations for inspection
(the civil aviation authority case). Similarly, new market conditions, such as increased competition and
demands for profit, has lead to the development of new
inspection methodologies (the civil aviation authority
case) and a focus on efficiency and cost-cutting (the air
traffic control/airport operation and line maintenance
cases). Overall, both examples illustrate how the particular organization interacts with local, national and
global aspects of its environment, and seeks way of
countering the effects of changes in the environment.
In all three cases, political/economical complexity plays a major role. One example is how changes

in international rules and regulations require the civil

aviation authority to handle two parallel and coexisting legislations (national and international). Another
example is how deregulation at a national level has
created new market conditions, with restructuring
demands (downsizing, cost-cutting etc.) occurring in
both the air traffic control/airport operation and line
maintenance cases. A more specific political complexity aspect can be seen in how an increase in the
number of actors competing within the civil aviation
market has lead to the development of a risk based
inspection methodology, aimed at capturing systemic
safety concerns. Results of the civil aviation authority case suggest that the development of the new
methodology is challenged by changes at a national
(deregulation) and international (new/modified rules
and legislations) level, thus adding further layers to the
picture of political complexity.
Overall, both our results and analysis support a
methodological approach to civil aviation, based on
identifying and understanding unique traits of the individual parts of the system (the particular case) as
well as traits common to two or more parts (across
cases). Specifically, by comparing the unique individual, organizational and economical/political aspects
of the particular case, we gain better understanding of
how unique and common elements interact, creating a
complex system that affects change perceptions, safety
perceptions and practices and change implementation.
In sum, the use of complexity to understand change
and the effects of change on safety is similar to and
support the complexity perspectives applied by Snook
(2000), Vaughan (1996, 2005) and Woods (2005) in
understanding the factors involved in several specific
accident scenarios.

CONCLUSION

Next, we address the research objectives of the paper,

followed by our future concerns for safety in the civil
aviation transport system of Norway.
6.1 The research objectives
Objective 1: To explore whether changes in different parts of the Norwegian aviation system affect the
safety work within each organisation (internally)
Within each case, the results above suggest that
safety work within all three organisations has been
influenced by one or several changes, in both positive and negative direction. Negative tendencies are
first and foremost illustrated by views on relocation
processes and how they affect competence transfer
and development of risk-based supervision (regulator), and by views on restructuring processes and how

1391

they increase the focus on costs and downsizing, and

how they change flexibility and safety practice (airport
operation/air traffic control, maintenance). However,
data also suggests that negative tendencies are counteracted by positive change initiatives, where learning
abilities, reporting frequency, competence level, and
technical standard are strengthened by changes related
to reporting routines/systems, equipment, technical
facilities, technology, competence, and harmonisation
and standardisation.
Objective 2: To explore whether changes in the
aviation system affect the relations between organisations that contributes to aviation safety (crosswise)
Across the studied cases, we find that a number of
qualities contribute to unite and strengthen the safety
work within the civil aviation system. Characteristic
features are employees professionalism, safety consciousness, and responsibility related to occupational
practice. In addition, there is a strong emphasis on
regulation, procedures, and new technology combined
with a focus on learning and improving existing competence. At the same time, and as a contrast to these
features, the empirical material indicate that relations
across organisations are influenced by new organisational and physical interfaces (relocation of regulatory
authority, separation of maintenance from airline company, relocation of line maintenance base away from
airport) that may reduce the conditions for how safety
work is practiced in the aviation system. More specific,
this concerns the premises for how each organisations
and the aviation system as a whole are able to attend
the collective approaches and mechanisms shown to
be essential for safety.
6.2 Future concerns
In our opinion, new organisational and physical interfaces that affect the relations between the different
actors of the Norwegian aviation system represent the
main challenge regarding implications of changes for
aviation safety in the future. In addition, the results
clearly document how societal characteristics in forms
of developments within market and technology, safety
regulation, and internationalisation have both direct
and indirect consequences for the different actors
safety work and in turn for the overall aviation safety.
The results also demonstrate that a combination of
internal and cross-wise studies of actors and levels
provide insight into system traits as well as individual traits, with implications for the system as a whole.
An example is how our results indicate that informal
mechanisms for coping with change (system trait),
like flexibility, slack, utilization of knowledge (line
maintenance case) and transfer of tacit knowledge
(the aviation authority case), can become threatened
by larger structural changes, such as relocation or

efficiency programs. This supports a system approach to understanding changes and implications to
safety that we believe should be continued in future
research endeavours focusing on understanding and
managing changes.

ACKNOWLEDGEMENTS
We wish to thank the Norwegian Research Council
for financing this work as part of the RISIT (Risk and
Safety in the Transportation Sector) program. We also
wish to thank our network within Norwegian civil aviation that contributed with enthusiasm and knowledge
in planning and accomplishing the project. Special
thanks go to Avinor and the Sola Conference for their
financial contribution to the project. Access to a questionnaire survey within the Norwegian civil aviation
transport system, was made available to us by the
Accident Investigation Board Norway and Institute of
Transport Economics. Finally, we wish to thank our
research colleagues at DNV Norway and the International Research Institute of Stavanger (IRIS) for inputs
and discussions, and Preben Linde who stepped in
as a project manager in the difficult initial research
phase.

REFERENCES
Amalberti, R. 1996. La conduite de systmes risques [The
control of systems at risk]. Paris: Presses Universitaires
de France.
Drazin, R. & Sandelands, L. 1992. Autogenesis: A Perspective on the Process of Organizing. Organization Science
3(2): 230249.
Goodstein, L.P. et al. 1988. Task, erors and mental models.
London, Taylor & Francis.
Guba, E.G. & Lincoln, Y.S. 1981. Effective evaluation:
Improving the usefulness of evaluation results through
responsive and naturalistic approaches. Josey-Bass, San
Francisco, USA.
Morgan, G. 1986. Images of Organization. Newbury Park,
CA, Sage.
Hollnagel, E. 1996. Reliability Analysis and Operator
Modelling. Reliability Engineering and System Safety 52:
327337.
Perrow C. 1984. Normal Accidents: Living with High-Risk
Technologies. Basic Books, NY.
Pettersen, K.A. et al. 2008. Managing the social-complexity
of Socio-technical systems (paper draft).
Rasmussen, J. 1982. Human Factors in High Risk Technology. In Green. E.A (ed.), High Risk Safety Technology.
London, John Wiley and Sons.
Rasmussen, J. 1997. Risk Management in a dynamic
society: A modelling problem. Safety Science 27(2/3):
183213.
Sandelands, L. & Drazin, R. 1989. On the Language of Organization Theory. Organization Studies 10(4): 457477.

1392

Snook, S.A. 2000. Friendly fire: The Accidental Shootdown of U.S. Black Hawks over Northen Iraq. Princeton,
Princeton University Press.
Vaughan, D. 1996. The Challenger launch decision: Risky
technology, culture, and deviance at NASA. Chicago:
University of Chicago Press.
Vaughan, D. 2005. System Effects: On Slippery Slopes,
Repeating Negative Patterns, and Learning from Mistake?
In Starbuck, W.H. & Farjoun, M. (eds), Organization at
the Limit: Lessons from the Columbia Disaster. Malden,
MA, Blackwell Publishing.
Weick, K. 1977. Organization design: organizations as
self-designing systems. Organizational Dynamics 6(2):
3046.

Wolf, F.G. 2001. Operationalizing and testing normal accident theory in petrochemical plants and refinerie. Production and Operations Management 10(3): 292305.
Woods, D.D. 1988. Coping with Complexity: The Psychology of Human Behaviour in Complex Systems. In
Goodstein, L.P., et al. (ed.), Task, erors and mental
models. London, Taylor & Francis.
Woods, D.D. 2005. Creating Foresight: Lessons for Enhancing resilience from Columbia. In Starbuck, W.H. &
Farjoun, M. (eds), Organization at the Limit: Lessons
from the Columbia Disaster. Malden, MA, Blackwell
Publishing.

1393

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Electromagnetic fields in the industrial environment

J. Fernndez, A. Quijano, M.L. Soriano & V. Fuster
Institute of Electric Technology, Paterna, Spain

ABSTRACT: The potential consequences of the human exposure to electromagnetic fields have generated an
increasing interest for both, public and relevant authorities, in terms of health and safety at work. It is worth
highlighting the adoption of the Directive 2004/40/CE of the European Parliament and the Council of 29 April
2004, which is being transposed into national legislation.
The directive establishes the minimal health and safety requirements for the protection of workers from the
health and safety risks derived, or that could be derived, from the exposure to electromagnetic fields (0 Hz to
300 GHz) during their work. Therefore, in fulfilment of the directive, the employers will have to evaluate and,
if necessary, measure and/or calculate the electromagnetic field levels that workers are exposed to: on the one
hand, by the measure of the electromagnetic fields at the place of work, and, on the other hand, faced with the
appearance of high electromagnetic field levels, to carry out the appropriate actions by means of electromagnetic
shieldings or other measures to reduce the electromagnetic field level. The efficiency of the aforementioned
measures is realized by means of electromagnetic fields simulations by finite elements.

INTRODUCTION

Electronic technology and systems have contributed to

significant improvements in quality of life and communications in society, and have given rise to great
advances in standards of living, as well as providing
the means to perform certain highly specialized medical treatments. Nevertheless, these systems generate
electromagnetic (EM) fields which, at excessive levels
of exposure, can be dangerous to humans.
The possible physiological effects of electromagnetic radiation came to be widely studied following
various studies made in the 1970s. People became
increasingly concerned as some studies derived links
between these fields and illnesses such as cancer,
depression etc.
According to the special Eurobarometer poll on
electromagnetic fields carried out in 2006, only 14% of
Europeans declared themselves Not at all concerned
by the potential health risks of electromagnetic fields,
as opposed to 35% Not very concerned, a further
35% Fairly concerned and 13% Very concerned.
Furthermore, the Eurobarometer also indicates
that concern about adverse effects of electromagnetic
fields has been on the increase in recent years.
Changes in work practices have been a crucial
factor in the increased exposure of persons to electromagnetic fields. The authorities have taken this into
account, and this has resulted in the creation of a European Directive (2004/40/EC) to guarantee the safety of
workers with respect to electromagnetic fields.

Figure 1. European citizens concern about potential health

risks due to electromagnetic fields.

DEVELOPMENT OF DIRECTIVE
2004/40/EC

In order to guarantee a high level of protection of

public health, the Commission monitors advances in
scientific research as well as national and international
standards regarding electromagnetic fields. At the EU
level, the Council Recommendation of the 12th of July

1395

1999 on limiting exposure of the population to electromagnetic fields (0 Hz300 GHz) establishes a system
of minimum restrictions and reference levels.
The Recommendation is based on the directives of
the International Commission on Non Ionizing Radiation Protection (ICNIRP), approved in 1998 by the
Scientific Director Committee advising the European
Commission, and, in 2001, by the Scientific Committee on Toxicity, Ecotoxicity and the Environment
(CSTEE). It is confirmed and recorded in the Scientific Committee on Emerging and Newly Identified
Health Risks (SCENIHR) decree of the 29th of March
2007, and is a European reference standard in worker
safety (2004/40/EC).
In 1989 the European Parliament and the Council adopted the Councils Directive 89/391/CEE, dated
the 12th of June 1989, concerning the application
of measures to promote improved worker health and
safety. In September 1990 the European Parliament
adopted a resolution on its action plan for the application of the EU Charter of workers fundamental
social rights; in this framework the Commission was
solicited to develop a specific Directive pertaining
to risks related to persons physically present in the
workplace.
First were developed the directives pertaining
to vibrations and noise, and in 2004 Directive
2004/40/EC, concerning the minimum safety standards concerning risks due to electromagnetic fields
(eighteenth specific Directive with order in section 1
of article 16 of Directive 89/361/CEE). The present
Directive establishes minimum requirements pertaining to EM fields (0 Hz to 300 GHz), to be fulfilled
by companies, both in terms of limit exposure values and values giving rise to a required action. The
present Directive applies only to short-term, but not
long-term, known negative effects on workers health
and safety.

If charge is represented as q, and the Electric Field

Strength as E, the following force is exerted on the
charge:

F = q E

Electric field strength (E): vector magnitude corresponding to the force exerted on a charged particle,
independent of its movement in space [V/m].
Magnetic field: region of space in which a point
electric charge with velocity v experiences a
force exerted perpendicular and proportional to the
velocity vector and to a field parameter called
magnetic induction, according to the following
expression:

F = q (v B)

FUNDAMENTAL THEORY
OF ELECTROMAGNETIC FIELDS

 =H

B

In order to correctly identify different sources

of EM fields in the industrial or work environment, and perform measurements to be compared
with legal standards currently in effect, a minimum understanding of EM field theory is necessary. Fuller knowledge of this area may be
obtained through consulting the available bibliography (Cardama et al. 1989), (Stutzman et al.
1998).
Electric field: region of space surrounding a
charged space and forming a vector space such
that any point of the region may be characterized by a vector called Electric Field Strength.

(2)

Magnetic field strength (H): vector quantity which,

along with magnetic induction, determines the
magnetic field at any point in space [A/m].
Magnetic flux density or magnetic induction (B):
Vector function of the magnetic field, defined by
the following expression:

(1)

(3)

where is the magnetic permeability of the propagation medium, expressed in [Weber/A m]. Magnetic induction is measured in the Tesla [T], which
corresponds to 1 [Weber/m2 ].
Electromagnetic field: Field associated with EM
radiation. described by two vectors, one electrical
and the other magnetic, which advance, mutually
perpendicular, in the direction of propagation.
Specific Absorption Ratio (SAR): Power absorbed
per unit of mass of body tissue, whose average value
is computed for the whole body or parts thereof in
[W/kg].
Contact current (IC ): current between a person and
an object [A]. A conducting object in an electric
field can be charged by the field.
Current density (J): current flowing through a
unit section perpendicular to the current direction, within a three-dimensional volume such as the
human body or a part thereof [A].
Power density (S): Power per unit area normal to
the direction of propagation [W/m2 ]. Appropriate
magnitude for use at very high frequencies, and
whose penetration of the human body is low.

Electromagnetic phenomena may be described

through Maxwells equations, which include Faradays
equations (Equation 5) and those of Ampre-Maxwell

1396

(Equation 4) and Gauss (Equations 6 and 7).


 = J + D
H
t

 = B
E
t

D =

(4)
(5)
(6)

 =0
B

(7)

where D is displacement electric field [Coulombs/m2 ],

J is the current density [A/m2 ] and is the charge
density [Coulombs/m3 ].
Ampre-Maxwells and Faradays equations demonstrate the interdependence of time-varying electric and
magnetic fields; in such cases the two fields are interrelated. When the fields do not vary with time (static
fields), the electric and magnetic fields are mutually
independent.
The relationship between electrical and magnetic
fields also depends on the distance of the point of
measurement from the source of EM radiation. This
distance (d) is expressed in terms of number of
wavelengths (). Wavelength is defined as follows
(Equation 8):
=

c
f

Figure 2.

LEGAL SPECIFICATIONS OF EXISTING

LIMIT EXPOSURE VALUES AND VALUES
GIVING RISE TO A REQUIRED ACTION

On the 29th of April 2004, directive 2004/40/EC of

the European Parliament and Council delineated the
minimum health and safety requirements concerning
exposure of workers to risks from physical agents
(EM fields). This Directive is pending transposition
to Spanish law. Table 1 shows Limit Exposure Values
and Table 2 Values Giving Rise to a Required Action.
In both tables, f is the frequency of the electric and
magnetic fields given in the Frequency Range column.
It is important to note that the magnitudes appearing
in Table 2 can be measured directly; those in Table 1
cannot, as their values are obtained indirectly through
calculations in which the human body is approximated
by means of simple geometries.

(8)
5
8

where c is the speed of light, whose value is 310 m/s,

and f the frequency of the emitted radiation.
Established norms define a practical approximation
criterion, in order to establish the limits of the near
field and the far field, for frequencies below 1 GHz:
the near field is where d < 3, and the far field is
where d > 3.
The above gives rise to the following conclusions:
At low frequencies, measurements of both the electrical and magnetic fields must be performed, as
these two are independent.
At high frequencies, the two fields are interdependent. Established standards indicate that if the
measurement point is in the near field, the intensities of the electric field E[V/m] and magnetic
field H[A/m] must be recorded, in order to compare these with the established values in each case.
If the measurement point is in the far field, it is
only necessary to measure one of the parameters, as
the other may then be deduced using Equation 9.
|H | =

Representation of EM waves.

|E|
0

IDENTIFICATION AND CLASSIFICATION

OF THE MAIN ENERGY
INFRASTRUCTURES PRESENT
IN THE LABOUR MARKET

In the following table (Table 3) is displayed the emissions spectrum regulated by Directive 2004/40/EC
according to frequency.
In this last table, N is the parameter determining
the width of each band between 0.1 10N 3 10N Hz
(International Telecommunications UnionITU).
In the following section, specific types of equipment in work environments generating EM radiation
are listed for each frequency range (Vargas et al.):
Sources generating fields at frequencies below
3 kHz (0 Hz f < 3 kHz)

(9)

where 0 = 120  is the impedance of empty space.

1397

Static fields (0 Hz) (0 Hz)




Magnetic levitation trains

Magnetic resonance systems for medical diagnosis
Electrolytic systems applied in experimental
industrial settings

FEB (30 Hz f < 300 Hz)



Equipment for generation, transmission or use

of electrical energy at 50 Hz

Table 1.

Limit exposure values.

Current density
in head and
trunk, J

Average SAR
for entire body

Local SAR
(head and trunk)

Local SAR
(extremities)

Power density,
S

Range of
frequencies

mA/m2 (rms)

W/Kg

W/Kg

W/Kg

W/m2

Up to 1 Hz
14 Hz
41000 Hz
1000 Hz100 KHz
100 KHz10 MHz
10 MHz10 GHz
10300 GHz

40
40/f
10
f/100
f/100

0.4
0.4

10
10

20
20

50

Table 2.

Values giving rise to a required action.

Electric field
strength, E

Magnetic field
intensity, H

Magnetic
induction, B

Plane wave
equivalent
power density,
Seq

Range of
frequencies

V/m

A/m

W/m2

mA

mA

01 Hz
18 Hz
825 Hz
0.0250.82 kHz
0.822.5 kHz
2.565 kHz
65100 kHz
0.11 MHz

20000
20000
500/f
610
610
610
610

1.63 105
1.63 105 /f 2
2 104 /f
20/f
24.4
24.4
1600/f
1.6/f

2 105
2 105 /f 2
2.5 104 /f
25/f
30.7
30.7
2000/f
2/f

1.0
1.0
1.0
1.0
1.0
0.4f
0.4f
40




High- and medium-voltage lines

Household appliances (fridges, hair-dryers
etc.)




Induction stoves
Modulated radio-wave transmission antennas
Arc-welding equipment




VLF (3 kHz f < 30 kHz)








Nautical radio-telephones
AM radio broadcasting
Heat-sealing machines

Antennas for radio broadcast and navigation

Computer monitors
Anti-theft systems

Amateur radio operators

Heat-sealing machines
Surgical diathermy machines
Anti-theft systems

VHF (30 MHz f < 300 MHz)






Radio broadcast antennas

Frequency modulation
TV station antennas
Anti-theft systems

UHF (300 MHz f < 3 GHz)

LF (30 kHz f < 300 kHz)



Sources of radio-frequency fields (3 kHz f <

300 GHz)

Current induced
in extremities, IL

3 MHz f < 30 MHz

300 Hz f < 3 kHz



Contact current,
IC

Screens and monitors

Radio broadcast antennas
Nautical and aeronautical telecommunications
Radio tracking devices






Mobile telephones
Mobile telephone base stations
Microwave ovens
Surgical diathermy machines
Anti-theft systems

SHF (3 GHz f < 30 GHz)

HF (300 kHz f < 3 MHz)

1398

Table 3. Nomenclature of frequencies regulated by

Directive 2004/40/EC.
N

Banda

11

EHF
Extremely high
frequencies
SHF
Super high
frequencies
UHF
Ultra high
frequencies
VHF
Very high
frequencies
HF
High
frequencies
MF
Medium
frequencies
LF
Low
frequencies
ELF
Extremely low

300 GHz
30 GHz

1 mm
10 mm

30 GHz
3 GHz

10 mm
100 mm

3 GHz
300 MHz

100 mm
1m

300 MHz
30 MHz

1m
10 m

30 MHz
3 MHz

10 m
100 m

3 MHz
300 kHz

100 m
1 km

300 kHz
30 kHz

1 km
10 km

30 kHz
0 Hz

10 km

10
9
8
7
6
5








Field produced by 3 current lines.

Figure 4.

Field produced by 3 current lines with screen.

Figure 5.

Relative positions of current lines and screen.

Satellite telecommunications antennas

Radars
Microwave transmission links

EHF (30 GHz f < 300 GHz)



Figure 3.

Radio navigation equipment

Radars
Radio broadcast antennas

MEASURES TO BE ADOPTED BY
COMPANIES TO COMPLY WITH
DIRECTIVE 2004/40/EC

In the first instance, companies should perform studies

to identify sources of EM fields on their premises, and
determine how their workers safety can be guaranteed
given these measurements.
These measurements in the EM fields present constitute a fundamental step in the verification of compliance with Directive 2004/40/EC in various work
environments.
One useful tool for such identification and diagnosis is EM field simulation software. These programs
employ finite element resolution of partial derivative
equations representing the magnetic fields to correctly
model sources of EM radiation and their surroundings.
They enable the simulation of those magnitudes which
may be measured directly, as well as the derivation of
indirect measurements (such as current density, for

example), to be compared with those indicated in the

Directive.
These software programs enable the implementation of studies of the best configuration and location

1399

of different equipment deployed in industrial environments, in order to minimize their combined effect on
the workers. All of these simulations are always tested
through comparisons with actual medical data in the
workplace. As an example, consider the magnetic field
shown in Figure 3, which can be reduced through
screening structure as shown in Figure 4. Figure 5
shows the physical location of the current lines creating the magnetic field with respect to the screen.
7

PRESENT AND FUTURE OF WORK

SAFETY WITH RESPECT TO EM FIELDS

The deadline for transposition of Directive 2004/40/EC

into the legal systems of the Member States was 30
April 2008.
In 2006, the medical community informed the
Commission of its concerns regarding the implementation of this Directive, claiming that the exposure limit
values laid down therein would limit to a disproportionate extent the use and development of magnetic
resonance imaging (MRI). Moreover, the International
Commission for Non-ionising Radiation Protection
(ICNIRP) is currently reviewing the guidelines on
static magnetic fields and loss-frequency time-varying
fields on which the Directive is based.
In order to allow a full analysis of the studies,
including that launched by the Commission, regarding the potential negative impact of the Directive
on the medical use of MRI, and pending the results
of the review of the ICNIRP recommendations, the
postponement of the deadline for the transposition of
Directive 2004/40/CE was recently officially approved
by European Parliament.
The European Directive 2004/40/EC was transposed into Italian law by decree on the 19th of November 2007. This decree will be enforced beginning the
30th of April 2008. Italy is the first major country
that has decided to implement the requirements of
Directive 2004/40/EC (in light of the postponement
of the deadline for the transposition of the Directive,
recently approved, the Italian decree will soon be modified to introduce some relaxation for MRI systems
and similar devices); however there are a number of
smaller countries that have also adopted the directive:
Austria, Slovakia, Czech Republic, Slovenia, Estonia,
Latvia and Lithuania.
It is plain to see that a number of European countries have already taken the decision to transpose the
EU Directive into their own legal systems. The new
deadline for transposition throughout the EU is 2012;
therefore companies throughout the continent must
progressively adapt in order to be in compliance when
the legislation comes into definitive effect.

CONCLUSIONS. FUTURE OF WORK

SAFETY WITH RESPECT TO EM FIELDS

One most salient point in the history of work safety

with regard to EM fields is the fact that they are normally imperceptible to humans; this has meant that
through the years their possible effects on workers have
been minimized by businesses and industries.
Nonetheless, the transposition of the European
Directive concerning the limits of exposure of workers
to EM radiation into Spanish law is, if not imminent,
due to take place within a short timeframe; for this reason, Spanish businesses and industries should begin to
take the measures necessary for compliance.
One means of effectively reducing exposure to
EM fields is through the application of preventive
measures in work station design, in the selection
of equipment and work methods and procedures. In
order to give impetus to established industries adopting
the new legislation and effectively guide new businesses in this regard, studies of the sources of EM
radiation in industrial environments will have to be
performed; these studies should carry out thorough
measurements of field strengths and enable the identification of points of risk. Where necessary, existing
systems should be reconfigured to include screening
systems, in order to protect workers from the risks
associated with exposure to EM fields.
REFERENCES
Cardama, A. et al. 1998. Antenas. Barcelona: Edicions UPC.
Stutzman, L. et al. 1998. Antenna theory and design. New
York: John Wiley & Sons.
European Commission 2007. Special EUROBAROMETER
272a. Electromagnetic fields.
Directive 2004/40/EC of the European Parliament and of
the Council of 29 April 2004 on the minimum health and
safety requirements regarding the exposure of workers to
the risks arising from physical agents (electromagnetic
fields) (18th individual Directive within the meaning of
Article 16(1) of Directive 89/391/EEC).
2007/0230/COD. Proposal for a DIRECTIVE OF THE
EUROPEAN PARLIAMENT AND OF THE COUNCIL
amending Directive 2004/40/EC on minimum health and
safety requirements regarding the exposure of workers
to the risks arising from physical agents (electromagnetic fields) (eighteenth individual Directive within the
meaning of Article 16(1) of Directive 89/391/EEC).
Vargas, F. et al. Campos electromagnticos y salud pblica.
Informe tcnico elaborado por el comit de expertos.
Subdireccin General de Sanidad Ambiental y Salud Laboral. Direccin General de Salud Pblica y Consumo.
Ministerio de Sanidad y Consumo.

1400

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Electrostatic charges in industrial environments

P. LLovera, A. Quijano, A. Soria & V. Fuster
Institute of Electric Technology, Paterna, Spain

ABSTRACT: Electrostatics discharges are a dangerous ignition source in industrial environment. Many
regulations and standards help industry to design equipment and installations to avoid the presence of electrostatic charge. However, very often the presence of static charge in equipments or processes is hard to detect
and, even more difficult, to eliminate. Several strategies must be followed to avoid hazards and damages in
industrial processes or products: good equipment design, deep comprehension of materials properties, staff
training, strict maintenance and very careful modification of any process or equipment. In this paper, we will
focus on materials characterization and on-site diagnostic.

generation is unavoidable, it is necessary to apply some

elimination techniques.

INTRODUCTION

Electrostatics discharges are a dangerous ignition

source in industrial environment. Many regulations
and standards help industry to design equipment and
installations to avoid the presence of electrostatic
charge. However, very often the presence of static
charge in equipments or processes is hard to detect
and even more difficult to eliminate. Several strategies must be followed to avoid hazards andr damages
in industrial processes or products: good equipment
design, deep comprehension of materials properties,
staff training, strict maintenance and very careful modification of processes and equipments. In this paper,
we will focus on materials characterization and onsite diagnostic. The former can be performed in an
Industrial Electrostatics Laboratory and can be applied
to several materials (polymers, textiles, composite
materials), charging processes (triboelectric charging,
corona charging or contact charging) and different
kind of samples; from small materials to elements
such as chairs or even floors. An analysis of the
measurement techniques, the magnitudes (resistance,
charge dissipation constant, etc.) and charge generation mechanisms is necessary in order to understand
the electrostatic processes. Closely related to laboratory tests, on-site diagnostics are the best tool for
electrostatics charges detection. This can be applied to
industrial environment. In general, on-site measurements are required for new installations where electrostatic charge presence was not taken into account,
or to check the validity of the designed measures. In
case of existing installations, after process modifications, in situ measurements are recommended to detect
the presence of new unexpected electrostatic charge.
In addition, for those cases where electrostatic charge

2
2.1

BASIC ELECTROSTATIC PROCESSES

Charge generation

Static electricity was first noticed by the effects of

frictional charging. The properties of amber where
already known by Thales of Miletus (600 BC), but they
were only linked to general effects of electricity in the
course of 18th and 19th centuries. From an industrial
point of view, static electricity became a real concern
due to the extensive use of plastic materials during
the 20th century. There are different processes leading
to the accumulation of static charge on the surface of
materials: frictional charging, contact charging, induction charging, chemical charging and high voltage
charging. Additionally, piezoelectric or photoelectric
charging can also occur. However frictional and contact charging can be considered the major causes in
industrial environments.
Contact charging is the most basic process. It happens between almost any different surfaces. It is due to
a redistribution of charge in the surface layer to reach
potential energy equilibrium. Charge carriers in the
highest energy levels of one surface are transferred to
lowest energy in the opposite surface, thus leading to
a net charge in surfaces if they are separated. Contact charge is limited, since it depends on the available
energy levels on the surface (Blythe 1997). Transferred charge increases with the number of contacts
until it reaches a maximum value.
Frictional charging or tribocharging is due to the
sliding, rolling, impact, vibration or deformation of

1401

The main dissipation cause is conduction. Surface

or volume conduction should be considered as two different properties of the materials. Surface conduction
depends on the state of the surface (dust, humidity,
roughness, etc.) and the path to ground length. Surface conductivity shows big variations and it cannot
be considered an intrinsic property of the material.
Each material should be tested with the working operation surface conditions to determine surface resistivity.
Usually, insulating materials increase their surface
conduction with ageing. On the opposite, volume
conduction is closely related to material properties
and will depend on the composition (polymer, additives, etc.) and the humidity content of the sample. Of
course, both conductivities are temperature dependent.
Insulating materials exhibit very often an internal
polarization leading to a partial screening of surface
(or volume) charges. Polarization induces a potential
decay (in absolute value) and, in general, it is difficult
to distinguish between volume conduction or dielectric
polarization.

Surface
conduction

Volume
conduction

Figure 1.

Surface and volume (or bulk) conduction.

3
Figure 2.

Volume polarization.

surfaces. Charge generation depends on relative velocity and pressure. Charging mechanism is not well
understood, but it is commonly accepted that the physical process is the same as contact charging. Pressure
and velocity should modify the total amount of contact
points and thus the transferred charge. Tribocharging
is present in many industrial processes: extrusion lines,
moulding, packaging, unrolling, conveying, etc.
Induction charging is created by high voltage conductors. The electric field can generate static charges
in floating (non-grounded) conductive elements or a
permanent polarization on insulating materials.
Chemical or electrochemical charge generation
occurs during electrolytic charge separation or corrosion processes. However, charge is usually dissipated
by conductive paths. Piezoelectric or photoelectric
charging are particular phenomena which can induce
some electrostatic charge. They are used in some specific applications (sensors, printers) but they should
not represent a hazard.

2.2

Charge dissipation

Electrostatic charge is accumulated on insulating

materials (or conductive materials isolated from
ground by an insulating material). The stability of the
charge depends on dissipation processes.

MATERIALS CHARACTERIZATION

The main intrinsic property used to characterize antistatic materials is surface or volume resistivity. Commonly, a material is considered dissipative if surface
resistivity is in the range 106 to 1012 . But conductivity in itself is not enough to guarantee dissipation
of electrostatic charges, as long as a conductive path
to ground is not provided.
There are many standards describing the test procedures for different applications. The situation is,
in fact, quite puzzling and sometimes it is difficult
to choose the appropriate system. One of the most
general standard is the IEC 60093 which describes
a measurement configuration for resistivity measurements based on plane parallel electrodes, with a guard
ring to distinguish between surface or volume conduction (Fig. 3). The resistance is the ratio between
the applied voltage and the measured current. Experimental setup allows to measure surface or volume
current. Besides, due to slow polarization, the current decreases continuously during measurement so
resistance increases with time. It is a compromise to
establish the resistance as the value at a maximum time
or the value which remains almost constant within the
precision of the measurement. In fact, a slow polarization current can be decreasing for very long time
periods. This is also the reason why samples are short
circuited before measurements for a long period.
However, for high resistances or inhomogeneous
materials (like textiles with conducting fibers) resistivity measurements are not sufficient to ensure a
dissipative behaviour. Charge dissipation is also quantified by means of a potential decay measurement or

1402

Surface potential mapping

Volume resistivity
V

IS

IV

Corona charging system

Surface potential
measurement

IS + IV

Sample

IS
V

Figure 3.

Figure 5. Charge dissipation setup with a surface potential

probe and surface mapping.

Surface Voltage [V]

IV
IS + IV

Grounding plane

Surface resistivity

2000

insulating material

1500
1000
500

Time [s]

Volume and surface resistivity measurement.

0
0

10

20

30

40

50

60

-500

Sensing aperture

antistatic material
-1000

Corona charging

-1500
-2000

Figure 6. Potential decay curves of a dissipative material for

packaging and an insulating material (the polarization decay
can be observed at the beginning of the curve).

Sample
Grounded enclosure
Enclosure for
surface dissipation
samples

Surface
Potential (V)

Figure 4. Charge dissipation setup according to IEC

61340-5-1.

charge dissipation constant. Static charge is created

by ions deposition from a DC corona discharge. This
method generates a static charge on the surface of the
material which is very similar to tribocharging charge.
The dynamics of charge dissipation can be followed
with a surface potential probe or an electric field meter.
The standard IEC 61340-5-1 suggests the use of an
electric field meter which gives an average surface
potential of the charged area and the average dynamics
of dissipation. A corona charging system with needles
charges the sample at a determined voltage (Fig. 4).
Then, it is quickly removed from the measuring system allowing an electric field meter to measure the
electric field through a sensing aperture. Instead of
using a field meter, an electrostatic surface potential
probe follows the dynamics of dissipation (see Fig. 6)
with a great spatial resolution (less than 1 mm). It also
allows the scanning of the surface for a potential mapping, in some cases showing irregular areas or defects
(Fig. 7).
In an Industrial Electrostatics Laboratory testing
can be performed on objects like tools, furniture, floor
samples, shoes, clothes, etc. Testing on objects instead
of materials is necessary to ensure the final product

Y (cm)

X (cm)

Figure 7. Surface potential scanning of an insulator surface

after corona charging. The figure shows some anomalous
potential decay.

is antistatic, since the production process can modify

materials properties. The object design can also stop
charge dissipation due to coatings or grounding path
interruption.

INDUSTRIAL TESTING

There are four main steps to achieve a good control

of static electricity in industry: design, verification,
maintenance and training.

1403

The presence of electrostatic charges should be considered at the earlier stage of industry installations
design. According to the processes and associated
risks (personal safety, product or process damages)
some initial measures should be taken. For example:
identification of ignition risks and possible charging
situations (rolling, sliding, impact, powder conveying, etc.) or antistatic measures (dissipating materials,
grounding points, neutralizers, etc,).
Besides, in-situ verification should be performed to
check possible static charge sources. Some important
points for in.situ measurements are:
floor resistance: it is measured with special electrodes in two ways, point to point (see fig. 8) or
point to building electrical ground.
electrostatic potential of people: the setup of the
standard EN 1815 provides a simple setup to measure electrostatic potential of people walking on
resilient floors. Figure 9 shows an example of
measurement on an insulating floor. Potentials can
easily reach thousands of volts. It is considered that

Resistance meter

25 kg

25 kg
10 cm

Floor resistance measurement.

Human Voltage [V]

Figure 8.

Electrodes

3000 V is the minimum electrostatic potential to feel

electrostatic discharge effects.
high local electrostatic fields: detection by means
of a hand held field meter or similar. This should be
done carefully under explosive atmospheres using
appropriate equipment. For example, standard EN
1718 shows a measurement setup for electrostatic
field generated by a running light conveyor which
could be applied also in a diagnostic. Equipment
manufacturers offer a wide range of measurement
systems.

CHARGE ELIMINATION TECHNIQUES

There are several techniques to eliminate electrostatic

charges in processes and installations. Some of them
are oriented to material properties and others to the
processes.
Regarding material properties, choosing conductive materials or replacing the existing by conductive
ones is a well know elimination technique. Some external antistatic additives can also be used to increase
material conductivity, although they must be applied
regularly in order to keep conductivity. In some cases,
additives are inappropriate for certain processes or
products. Other solutions are based on the modification of the process. For instance, a reduction on the
speed of movements (belts, powders flow, etc.) can
reduce static charge under a critical level (which is,
of course, depends on the application). In other cases,
like offices or plastic storage, increasing air humidity could sometimes be enough to avoid electrostatic
charges.
However, in many situations, those previous techniques are not enough, and it is necessary to use
an external ion source to compensate electrostatic
charge. This is why ionizers are employed to discharge
surfaces.
Bipolar ionizers generate a flow of different polarity ions by corona discharge (AC or square voltage).

Electrostatic field meter

Belt sample under test

Time [s]

Figure 9. Human potential evolution walking on a resistive

floor. Each peak corresponds to a step.

Grounding plate

Figure 10. Example of experimental setup to measure static

fields in a conveyor (EN 1718).

1404

Finally, but of paramount importance, staff training should be seriously considered as the major tool
allowing to avoid problems with electrostatic charge.
A good understanding of electrostatics and elimination
techniques is the best way to succeed in electrostatic
elimination.

Bipolar corona ion generation

Negative ion flow

Positive ion flow

Figure 11.
ionizer.

Charge neutralization by an air-blown bipolar

Passive electrode

Figure 12.

Passive ionizer principle.

The train of positive and negative ions cancels the surface charge according to the polarity of the charges on
the material (Fig. 11). When surface is neutralized a
balanced flow of charge reaches the surface. The efficiency of ionizers decreases with the distance to the
object.
Sometimes a passive ionizer is enough to reduce
the charge down to an acceptable level. In that case,
the electric field of the static charge itself is used to
induce a corona discharge of the opposite sign in a
needle electrode. Generated ions partially discharge
the surface. This technique cannot completely eliminate the surface charge. Some commercial systems are
able to combine a surface potential measurement with
a DC driven corona discharge to ensure electrostatic
discharge. There are also some ionizers equipped with
a gas flow (nozzles and guns) to increase the efficiency.

CONCLUSIONS

Very often, electrostatic charges are not detected until

an accident or problem happens. Furthermore, sometimes it is not easy to understand the related charging
mechanisms or to identify the source. Basic material
properties involved in industrial electrostatics are surface and volume conductivity and polarisation. These
properties can be tested in an Industrial Electrostatics Laboratory. Testing is also possible with objects
instead of materials. But it must be kept in mind
that conductive materials are not always an efficient
solution to avoid electrostatic charges. Some simple
elimination techniques (process velocity, air humidity) can be applied, although sometimes ionizers are
required. In-situ measurements and maintenance are
very important to keep a process clean of static
charge. Nevertheless, all these solutions must be completed by staff training, as the major warrant for an
optimal security level in relation to static charge.
REFERENCES
J.S. Chang, A.J. Kelly & J.M. Crowley. 1995. Handbook of
electrostatic processes. Marcel Dekker Inc. New York.
Blythe, A.R. 1977. Electrical properties of polymers. Cambridge University Press. Cambridge.
IEC. 1980. Methods of test for volume resistivity and surface
resistivity of solid electrical insulating materials.
IEC. 61340-5-1. 2007. ElectrostaticsPart 5-1: Protection
of electronic devices from electrostatic phenomenaGeneral requirements.
EN 1815: 1998. Resilient and textile floor coverings. Assessment of static electrical propensity.
EN 1718: 1999. Light conveyor belts. Test method for the
measurement of electrostatic field generated by a running
light conveyor belt.

1405

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Empowering operations and maintenance: Safe operations with the

one directed team organizational model at the Kristin asset
P. Nsje & K. Skarholt
SINTEF Technology and Society, Trondheim, Norway

V. Heps & A.S. Bye

StatoilHydro, Trondheim, Norway

ABSTRACT: How operations and maintenance is organized and carried out is a central factor in the secure and
reliable operation of an offshore installation. In the Kristin asset on the Norwegian continental shelf, operations
and maintenance is conducted with a lean, highly skilled organization. To date, operations are characterized
by a safe and effective execution, and high regularity. The paper explores one central aspect of the safe and
effective execution observed on board; namely how the organizational model used in the asset influences safe
and effective day-to-day operations. The asset uses one directed team as a central concept in the operating model,
where the operations crew is empowered and synchronizes tasks between functions. The paper discusses how
this leads to shared situational awareness and self-synchronization as critical concepts and issues for safe and
effective operations.

INTRODUCTION

Kristin is a new installation on the shelf, set in production in 2005. It is a semi-submersible, producing from
the Kristin reservoir over four templates. The reservoir is marked by high temperatures and high pressure
(HT/HP). Its production capacity is 125,000 barrels
of condensate and just over 18 million cubic meters
of rich gas per day. The platform is also designed for
later tie-ins, such as Tyrihans (under development).
Now, how operations and maintenance is organized
is a central factor in the secure and reliable operation of an offshore installation. In the last decade,
a set of interesting developments on the Norwegian
continental shelf has occurred. Most large installations had their peak production in the mid- to late
nineties. New installations have traditionally been
smaller, yet located on technically more challenging
fields. Also, on the organizational side, developments
have been made. First, by using multi-skilled and
multi-functional teams as a central way of organizing work on the platform, exemplified in the sgard
field development. These basic insights are based on
organizing work to enhance flexibility and innovation,
seen in lean manufacturing (Womack 1990), and selfdirected work teams (Trist and Bamforth 1951; Fisher
and Kimball 2000). Second, a set of technological
door-openers are emerging; improved communication
and the possibility of online support between onshore

and offshore functions (i.e. Integrated Operations),

and knowledge systems, such as SAP, being easier to
use and integrate in work flows.
Capitalizing on these developments, several important developments were made in terms of the organization of work processes on the Kristin asset.
To promote better day-to-day planning and on-thefly problem-solving operator and technician workstations are set-up in an open office layout, close to
central control room, management, and workshops.
This is to promote Self-synchronization.
To increase involvement, operators and technicians
are given the responsibility to plan, execute and
report work orders on day-to-day basis. This is to
promote Empowerment.
To increase problem-solving capacity, operators
and technicians are to maintain close relationships
with their respective system/discipline specialist
in technical support. (Less different from other
installations,).
To promote effective decision-making, the offshore
and onshore managements are using an on-line
collaboration-room as their permanent office.
These developments have aimed to increase safe
operations uptime and are core elements in the
one directed team operation model used on Kristin.
As indicated, the key concepts in the model

1407

are self-synchronization and empowerment. Unlike

hierarchical synchronization, self-synchronization is
a mode of operation where functions share understanding and awareness in situations, and where resources
are coordinated between functions according to situations as they arise, based on this shared awareness.
Details of this are discussed later on. First, the
asset uses empowerment to enhance shared awareness by delegating responsibility for day-to-day and
weekly operations to the function itself. This means
responsibility for planning and rescheduling. Second,
technicians and operators are responsible for fullloop work processes, retrieving tasks in SAP, ordering
parts, executing tasks, and for reporting them when
work is done. This reduces hand-offs in the crew and
between shifts to a minimum. In the following key
elements of the model are discussed, especially the
concepts of empowerment and self-synchronization in
the operation and maintenance crew.
2

METHODS

The basis for the analysis of the operating model has

been:
Identifying work practices (informal practices as
well as collaboration forms)
Analyzing communication arenas (use of collaboration rooms, other forms of information) and communication content (problem solving, planning,
prioritizing etc.)
The scope of the study is the entire asset. The asset
is operated by a crew of 31 persons on board offshore
and by the onshore operations management, onshore
technical support, and production support, (equaling
a total of 18 man-years). Operations are organized
as one team, sharing collaboration facilities. The
Discipline function is organized as a matrix organization, where the technical lead discipline engineers
are part of a pool in Technical Engineering Support
(DVM), and are assigned to the different assets.
To cover the research questions, a qualitative
research design was used, including field observations, informal talk, and semi-structured interviews.
All functions in the asset were interviewed, both
onshore and offshore. All three shifts offshore were
covered. Functions were covered until saturation
(Strauss and Corbin 1990). The empirical material
consists of 69 interviews.
The researchers involved in the project maintained a
close relationship with management and stakeholders
during the project period. Management and stakeholders were involved in the analysis and discussion
of the importance of issues. All findings, proposed
goals, and suggested actions were discussed informally and formally with these actors, thus mutually

shaping and co-constructing the concepts used (e.g.

one directed team). This gives a high degree of reliability in the analysis, heightens the relevance of the
study, and strengthens ownership of the analysis itself
and suggested actions.
3

ONE DIRECTED TEAM: KEY FEATURES

AND CAPABILITIES

The organizational model for Kristin covers three

specific functions; management, operations, and technical support, and it covers how work practices/flows
are shaped. In the collection of data we were interested in both the internal practices of the function
and the relations between functions. The focus in this
paper is, however, the workings of the operation and
maintenance crew.
On the installation, the workspaces have been
designed to promote interactions and visibility. There
are three central work arenas relevant for the operation
(cf. Figure 2):
The collaboration rooms onshore/offshore
The office landscape area, including small conference rooms and two quiet rooms offshore
The central control room (CCR) offshore
The collaboration room and the office landscape
area lend themselves to increased interaction both
between and within functions. The physical distance
between the CCR, the landscape and the collaboration
room is less than 5 meters. As important activities,
such as planning, are to take place in the office landscape and the crew is physically close, face-to-face
communication is promoted.
The character and type of interaction in the operational model varies between functions. It is a goal
that these relations are played out in a dynamic way,
where know-how is used and decisions are made across
functional borders. Problems are to be solved by combining experience and analyses from operations and
technical support disciplines. Also, functional roles
are to be crossed during inspections, reviews, or just
straight-forward problem-solving. Through this combination, know-how on particular problems or issues is
enhanced and transaction costs are reduced. Because
of this way of operating and collaborating, it takes
less time for a new operator or discipline specialist
to understand the specifics of a valve, electric motor,
or a system trip. A goal for the operational model is
that the crossing of functional borders and relational
dynamics is combined into a higher degree of shared
situational awareness at the asset.
Given the organizational set-up described above,
the quality of the relations is critical for competent execution. In the material, observations of how interaction
takes place are an issue of concern for the individual,

1408

Figure 1.

Three dimensions of management and decision-making in the One Directed Team operation model.

and an area of organizational focus. We see that when

relations are discussed, dynamic aspects are central.
Close and familiar relations are positive. Co-solving
of problems and mutual use of different know-how are
central elements. And lastly, having clear roles and
responsibilities is crucial.
The organizational model aims for One Directed
Team. This model uses empowerment of separate functions and work arenas (such as the landscape, the
collaboration rooms) for problem-solving, and tries
to develop shared situational awareness as a means
to achieving the four mentioned operational goals of
HSE, uptime, cost-control, and knowing the technical
condition.
Management and decision-making in the operation
model moves along three dimensions as indicated in
Figure 1.
First, in the O&M-crew decisions are made by
the function when permitted (cf. empowerment). For
problem-solving and coordination, the larger O&Mcrew is used as an extended knowledge base, and as
a pool for extra hands when necessary. For example, many maintenance tasks need an electrician to be
completed, to cut off power supply before the

dismantling of equipment. When work is done

effectively, operators and electricians schedule and
prioritize their tasks according to the needs of the
different functions in the crew. At the base of this,
the operators/functions need to have a certain level of
self-synchronization.
Second, technical lead-discipline engineers in
DVM supports technical problem-solving, acts as a
knowledge base between shifts and between installations, and decides on and prioritizes tasks within the
specialty. Examples of the latter are decisions concerning modifications or replacements, or setting new
limits in control systems. Decisions are made by the
specialist together with management, but decisions
are informed by the crew. When it comes to arising
situations it is an advantage to have close and handson relations between the specialist and the crew. In
practice, the specialist employs a type of situational
leadership in such situations.
The third dimension is between the O&M-crew
and management (LED). The O&M supervisor is the
first line leader of the crew, but management is in
effect done by the whole management team. Management is not and cannot be involved in solving

1409

Figure 2. Layout of work-planning areas (D&V), CCR and Collaboration room (LED) offshore with link to onshore
functions indicated. Distance between D&V, CCR and LED is less than 5 meters.

(all) technical questions, but is involved in giving

priority to which tasks to be solved. Also, management coordinates between functions when necessary,
and between technical support, contractors and the
O&M-crew when necessary. The emerging management style in the asset is one of developing capabilities,
empowering operators on the one hand, and supporting
self-synchronization on the other.
The key elements in existing work practices are
described in the following

3.1

Work practices in the Operations

and Maintenance (O&M) Crew

The landscape is designed with workstations combined into a number of islands with three stations
together down the middle of the space, and a continuous set of workstations along the walls (c.f. Figure 2).
Process operators together with the SAS-technician
and the Security/Maritime Coordinator are located at
stations close to the collaboration room and CCR,
while the other functions are placed further along in

the landscape. Contract workers on shorter assignments (such as rust-cleaning and repainting) have work
stations in the far end of the office landscape. Guest
stations can be found at this end as well.
In the landscape all work preparations, planning
and reporting takes place, for all disciplines, including F&A and contractors. Thus, it is an integrated work
area for all for the planning, coordination and reporting phases of operational tasks. Work preparation and
planning is done early in the day, before moving out
into the process facilities, where most of the work-day
is spent.
In the landscape there is discussion and on-going
talk on work orders, tasks, and assignments for the
different disciplines. Operators can and will hear what
the other functions are planning or discussing. As
such the landscape is a source of informal information.
The operators see the landscape as a collective asset.
One operator says: It is easy get hold of persons,
it easy to discuss and coordinate tasks, and we feel
that we are connected . . . that we get information,
decisions, first hand and directly . . . No physical barriers in the landscape makes it possible to bring in other

1410

operators when a discussion takes place. Also, operators will join ongoing discussions if they hear that there
are issues or elements where they have relevant information. Most planning activities take place right after
the 7 a.m. morning meeting in the shared coffee bar,
where all the personnel on board are informed about
HSE, production status, and upcoming operational
issues.
The work process at Kristin is thus as follows:
After a planning/preparation session in the landscape,
personnel move out to workshops, stores, and the
process plant. Work orders are retrieved from SAP,
and for each role or discipline there is a set of programmed/scheduled activities and a set of corrective
activities. These are all part of the planned activities
for the week. There is one planning meeting for the
O&M-crew, held every Saturday, with a plan responsible person within each discipline. Now, giving leeway
to other tasks, the operators choose which work orders
to complete/start.
There are many examples of how this promotes
higher self-synchronization. When discussing the
days work with one technician, he described how his
planned taskthe scheduled refurbishment of a large
valvewas moved to after lunch. This was agreed at
the planning/preparation session in the morning. Then
he went back to SAP to find something less complex to fill his day until lunch. He pointed out that
he always had a set of work orders at hand, in order to
be doing something useful precisely for such situations. Accordingly, this work practice makes the
organization more robust in order to withstand changes
and arising situations.
The operators havegiven the premises of what
goes on in the larger O&M-crew and what is prioritized from managementa certain amount of sway
over which work orders to pick up. These decisions
are made on the lowest level possible. Personnel have
full responsibility for the task, including its planning,
execution, and reporting. This means that material
must be found or ordered before execution and work
is reported in one integrated loop. Correspondingly,
SAP-proficiency in the asset is very high. Says one
informant: Those who want a list of tasks you have
to do today put in their hand, will not apply for work
on Kristin.
4

DISCUSSION

The one directed team operation model promotes a

high degree of ownership of tasks, and a high degree
of transparency around work. This is critical for both
the dynamic between functions (between operations,
management, and technical support) and for the selfsynchronization experienced in the crew. Due to the
visibility of the different functions in the O&M-crew,

it is easy to quickly grasp the status and assignment

of the function. Also, transparency makes the connection between tasks and functions visible; meaning
that if a task is not completed it is easy to see who
is responsible. Personnel lend each other a helping
hand and/or find suitable tasks if planned activities
must be rescheduled. Of course, scheduled activities or
work orders in SAP, etc., are at hand for the functions.
As mentioned in the introduction, scheduled activities are completed in an orderly and effective way.
But it is important to note that in the asset and in the
O&M-crew, work processes lend themselves to quick
rescheduling of tasks when other tasks are more important. These are decisions made on the fly between
functions in the crew. Therefore, slack is minimized
and operators have a short turn-around time, and are
quick to handle arising situations. What stands out,
then, is that the work processes in the asset enables
self-synchronization in the function and between the
functions in the O&M crew.
The combined effect of the shared work areas and
full-loop work processes in the operations and maintenance crew is central in the operation model. The
combination enables:
collective knowledge base (such as help with problem solving)
self-synchronization (such as rescheduling of tasks,
handling of arising situations)
shared situational awareness (resulting in lower
transaction cost, enabling effective decisionmaking)
These observations resemble how teams and technical work is organized in lean enterprise models,
exemplified by Toyota (cf. Womack 1990). In lean
manufacturing, teams are multi-disciplinary and organized around processes and products, decisions are to
be made on lowest possible level, and hand-offs are to
be minimized. Empowerment is a central aspect of the
lean enterprise.
Furthermore, developments such as Intelligent
Manufacturing Systems (IMS) and Holonic Manufacturing systems (HMS) push issues such as agility
and adaptation as critical organizational goals for
team organization. In the center of HMS is that
of a holon, an autonomous, co-operative building
block of a manufacturing system (van Leeuwen and
Norrie 1997). This resembles what otherwise could
be called a team, but adds information and technologies as integral to the team. Thus information,
sensors etc. is an important organizational parameter.
In the Engineering phase of the Kristin development
information availability, online instrumentation of
process equipment was central. In the design of operation model, this has been a enabler for the One Directed
Team.

1411

Self-directed teams are further down on the continuum from low to high empowerment than are examples
from lean manufacturing (Fisher and Kimball 2000;
Lee and Koh 2001). As mentioned in the introduction,
the self-directed team model used in the sgard license
is one example on the NCS (Trist and Bamforth 1951;
Emery and Thorsrud 1976; Qvale and Karlsen 1997).
The one directed team model, however, does not use
such elements as explicit group meetings and consensus decision-making found in self-directed teams.
In the Kristin case, the generic insights on operations have been moved one step further, trying to
improve the understanding of the workings of these
organizational principles. Following Barley (1996),
Barley and Kunda (2001), and Orlikowski and Barley
(2001), a more grounded research design increases
the understanding of work processes and knowledge
flows (Orlikowski 2002). Also, the generic organizational capabilities suggested by Teece, Pisano et al.
(1997) can be given concrete substance in these settings. This is then what Teece, Pisano et al. (1997)
would call dynamic capabilities, specific to the organization. Henderson has tried to develop some examples
relevant dynamic capabilities (Baird, Henderson et al.
1997; Balasubramanian, Nochur et al. 1999). Selfsynchronization through shared situational awareness
and empowerment are key concepts in the dynamic
capabilities he prescribes.
Other contributions have developed new insights
into the type of knowledge used in technical work
processes, especially (Ryle 1949; Orlikowski 2002),
trying to include the type and nature of transactions
between actors as a part of effective problem-solving.
This fits nicely into the notion of shared situational
awareness mentioned above.
In the one directed team operational model shared
situational awareness and empowerment is used to promote self-synchronization. This is discussed in the
following.
First, how is shared situational awareness (SA) in
teams promoted and managed in dynamic environments? Rousseau et al. (2004:1415), Artman (2000),
and Patrick and James (2004) argue that there is an
increasing interest in studying team cognition, based
on the fact that teamwork, or working towards a shared
goal requires information sharing and coordination.
Depending on the tasks at hand, awareness is distributed in the team to one member or shared between
team members. Artman (2000: 113) defines TSA as;
Two or more agents active construction of a situation
model, which is partly shared and partly distributed
and, from which they can anticipate important future
states in the near future. Patrick and James (2005: 68)
support this perspective of TSA, because such a perspective . . . not only emphasizes the interpretative
and distributed nature of the types of knowledge that
constitute the awareness of the team but also hints

at the importance of communication and other team

processes that underpin the development of that awareness. These are particularly important as the nature
of the required awareness, both shared and individual, will fluctuate with the nature of the operations
undertaken by the control room team together with
the changing state of the plant. Individual SA will rely
on a number of cognitive processes as perception and
projection but also new activities like coordination and
communication.
Let us be more specific on how team SA is understood. Team SA can be defined as the degree to which
every team member possesses the SA required for his
or her responsibilities, or the SA developed in the
whole team to achieve its mission. Shared SA is the
overlap between team members, a subset of team SA
(French et al. 2004: 268) or the degree to which team
members possess the same SA or SA requirements.
In dynamic systems the state of the system changes
both autonomously as well as a consequence of
the decision makers actions. Several interdependent
decisions in real time are required if the system
is to be controlled (Brehmer 1992, in Garbis and
Artman 2004: 275). Events happen, are complex in
nature, contain several components, need analysis and
decision-making under pressure, and involve teamwork with considerable amounts of communication.
They argue that in order to understand how SA
develops, we must focus on three issues. First, one
must go beyond the individual operators and include
whole teams, or the concerted efforts of several team
members and their actions. Second, one must consider that cooperative work is based on the existence
of mutually shared and interdependent goals between
team members, and, finally, team situation awareness
relies on the use of various artifacts and presenting information to the team. Various team members
can have different information resources that must
be combined and coordinated to develop a shared
understanding. An important point here is that the
meaning of information embodied in these artifacts
is not always clear and must be interpreted and
negotiated between team members. Therefore team
situation awareness calls for negotiation and communication between team members with and through
these artifacts. They focus on the role of artifacts
and the interaction and negotiating practices as part
of artifacts as an important part of SA. It is a distributed cognitive process for two reasons, because
it is distributed between the operators of the team
and distributed in the sense as something in-between
the operators and the informational artifacts that
are used.
Correspondingly, in the Kristin case, we have seen
that the elements supporting SA are the features intrinsic to the one directed team. First, SA is supported with
the use of shared work areas discussed in the previous

1412

chapter. Second, establishing shared goals has been

and is an ongoing concern in the operation. In the
early phases of operation, a flying start was achieved
by having technicians and operators involved in the
engineering, procurement, and construction phases of
the installation. Later on, this was promoted by establishing a close collaboration between the technicians
and the onshore technical support. Through this the
crew establishes a higher degree of knowledge about
the artifacts, particularities, and history of the installation, and a higher degree of knowledge of the priorities
of the installation. It promotes know-whatknowing
other operators and technicians in the crew and the
installation, c.f. Orlikowski (2002). Third, the combination of existing, singular pieces of information into
larger maps is happening on the fly in the work area,
by the mechanisms indicated above. In Kristin, this
type of combination is a result of current work practices, where information sharing between functions
and in the crew is an ongoing feature of the process.
However, the flip side of the coin here is that some decisions are made with the persons available regardless of
who should be available. For example, the involvement
of technical support can suffering when or if system
specialists are unavailable, and there are examples of
disciplines where the technical support-role has developed to only rubber stamp decisions made on lower
levels. Another issue is to log decision made informally or in the collaboration room. As on put it; you
need to be there to know what is happening. There
are examples of decisions being made several times on
the same subject. On the whole, however, these negative aspects mentioned are minor issues in the larger
workings of the organization, not undermining critical
decision-making.
The work processes in the one directed team are
full-loop. A central feature of the operations at Kristin
is that operators and technicians are given the responsibility for planning, executing and reporting all tasks
within the team. The combination of work-form
and shared team awareness promotes empowerment.
Empowerment has been defined as a function of
information, authority, resources, and accountability (Fisher and Kimball 2000). The SA discussed
above makes it easier to share and get information. The full-loop work process promotes the three
other elementsauthority, resources, and accountability. These elements resemble the definitions of
self-directed teams, for example by the Association
for Quality and Participation. However, in this case,
there are important differences between SDT and
ODT. The team at Kristin is not fully cross-functional
in the sense that discipline-responsibilities will be
crossed; e.g. Pressure Safety Valves (PSV) are the
sole responsibility of the PSV-mechanic, a responsibility that will not be delegated. Also, the given tasks
and their prioritization are set by management and in

long-term plans (i.e. scheduled maintenance or the

steering of wells). There are no forums for discussing
strategies, and no consensus decision-making.
What we see, thus, in the Kristin case is that
shared situational awareness and empowerment promotes self-synchronization as an alternative to traditional coordination. The traditional coordination form
in a hierarchical organization is a task given to operational leaders (e.g. foremen), managing resources,
and personnel. It is a precise coordination form, but it
does little to promote knowledge sharing and motivation. On the other hand, self-directed teams promote
high levels of motivation and knowledge sharing, but
is dependent on discipline among all actors to keep
transaction costs down, if not it will make decisionmaking cumbersome and actions difficult to agree on.
Kristin Operations tries to combine the best of these
two worlds. It uses SA and empowerment to enhance
knowledge sharing and promote motivation. But it also
increases responsibility, in comparison to self-directed
teams, by keeping responsibility within the discipline.
The status of tasks for each discipline/function is
highly transparent. This increases the drive in the team
to coordinate material and personnel by itself. This
mode of coordination is self-synchronization.
5

CONCLUSION

In the operations on Kristin there have been no serious personal injuries (red or yellow incidents) to date.
Absences due to illness were as low as 2.7 percent in
2006. The asset has had higher regularity (up-time)
than other assets and better than planed. In the first
year of operation, non-planned shut-downs were 1.27
percent of the total time available. There have been
no gas leakages (on level 1 to 4) to date. Corrective
maintenance: Outstanding work orders of high and
medium level priority is about 500 hours, well below
the average of 1800 hours for installations on the NCS.
Scheduled maintenance is more difficult to assess in
this early phase of operations, but at the end of 2006
there was about 400 hours outstanding.
The central operational feature promoting this operational excellence is the one directed team operational
model. As argued, the combination of empowerment and shared situational awareness enables the
operations and maintenance crews to be proactive in
problem-solving. Second, this keeps transaction cost
in functions, in the team, and between teams and external functions low, securing effective problem-solving.
Third, with full-loop work processes and a high level
of transparency around who is responsible for tasks,
the number of hand-offs are reduced and motivation
strengthened. By this self-synchronization is achieved,
reducing the need for coordination from management
and promoting safe and reliable operation.

1413

REFERENCES
Artmann, H. (2000). Team Situation Assessment and Information Distribution, Ergonomics, vol.43, pp.11111128.
Baird, L., J.C. Henderson, et al. (1997). Learning
from action: An analysis of the Center for Army
Lessons Learned (CALL). Human Resource Management 36(4): 385.
Balasubramanian, P., K. Nochur, et al. (1999). Managing process knowledge for decision support. Decision
Support Systems 27(12): 145162.
Barley, S.R. (1996). Technicians in the workplace: Ethnographic evidence for bringing work into organization
studies. Administrative Science Quarterly 41(3): 404.
Barley, S.R. and G. Kunda (2001). Bringing work back in.
Organization Science 12(1): 76.
Emery, F.E. and E. Thorsrud (1976). Democracy at work: The
report of the Norwegian Industrial Democracy Program.
Leiden, Netherlands, Martinius Nijhoff.
Fisher, K. and A. Kimball (2000). Leading self-directed work
teams. New York, McGraw-Hill.
French, H.T., Matthew, M.D. and Redden E.S. Infantry Situation awareness in: Banbury, S and Tremblay (ed) (2004),
A Cognitive Approach to Situation Awareness: theory and
application, Ashgate Publishing Company, Burlington
VT: USA.
Garbis, C. and Artmann, H. Team Situation Awareness as Communicative Practices in: Banbury, S and
Tremblay (ed) (2004), A Cognitive Approach to Situation
Awareness: theory and application, Ashgate Publishing
Company, Burlington VT: USA.
Lee, M. and J. Koh (2001). Is empowerment really a new
concept? The International Journal of Human Resource
Management 12(4): 684695.

Orlikowski, W.J. (2002). Knowing in practice: Enacting a collective capability in distributed organizing.
Organization Science 13(3): 249.
Orlikowski, W.J. and S.R. Barley (2001). Technology and
institutions: What can research on information technology
and research on organizations learn from each other?
MIS Quarterly 25(2): 145.
Patrick, J. and James, N. (2004). A Task-Oriented Perspective
of Situation Awareness in Banbury, S and Tremblay (ed)
(2004), A Cognitive Approach to Situation Awareness:
theory and application, Ashgate Publishing Company,
Burlington VT: USA.
Qvale, T. and B. Karlsen (1997). Praktiske forsk skal gi
metoder for bedriftsutvikling. Bedre Bedrift 4.
Rousseau, R., Tremblay, S. and Breton, R. Defining and
Modeling Situation Awareness: A Critical Review in:
Banbury, S and Tremblay (ed) (2004), A Cognitive
Approach to Situation Awareness: theory and application,
Ashgate Publishing Company, Burlington VT: USA.
Ryle, G. (1949). The concept of mind. London, Hutchinsons.
Strauss, A. and J. Corbin (1990). Basics of Qualitative
Research: Grounded Theory Procedures and Techniques.
Newbury Park, SAGE Publications.
Teece, D.J., G. Pisano, et al. (1997). Dynamic Capabilities and Strategic Management. Strategic Management
Journal 18(7): 509533.
Trist, E.L. and K. Bamforth (1951). Some social and
psychological consequences of the Longwall method of
coal-getting. Human Relations 4: 338.
van Leeuwen, E.H. and D. Norrie (1997). Holons and
holarchies [intelligent manufacturing systems]. Manufacturing Engineer 76(2): 8688.
Womack, J.P., D.T. Jones, et al. (1990). The machine that
changed the world. New York, Macmillian.

1414

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Leadership and safety climate in the construction industry

J.L. Meli & M. Becerril
School of Psychology, University of Valencia, Valencia, Spain

S.A. Silva
Instituto Superior de Cincias do Trabalho e da Empresa, Lisboa, Portugal

K. Mearns
School of Psychology, University of Aberdeen, Aberdeen, UK

ABSTRACT: The psychosocial model of work-related accidents adopts an agents safety response approach
to safety climate and shows the chain of safety influences from the Organizational Safety Response (OSR)
to accident indicators, such as micro-accidents (MA), through the Supervisors Safety Response (SSR), the
Co-Workers Safety Response (CSR), the Worker Safety Response (WSR) and the perceived risk (PR). This
paper analyzes the influence of leadership, role clarity, and monotonous and repetitive work (MRW) on the
psychosocial model variables in a Spanish construction sample (N = 789). Leadership and role clarity appear as
significant antecedents of the OSR, SSR and CSR, while MRW acts as an antecedent of PR and MA. Leadership
also shows direct effects on MRW and PR. Results suggest that leadership not only affects the chain of safety
responses that influence the safety outcomes, but it also directly affects the main safety properties of the work,
the levels of risk and micro-accidents.

ROLE THEORY PERSPECTIVE,

LEADERSHIP AND SAFETY
RESPONSES

In spite of the many studies showing the heavy economic, social and human costs of lack of safety, (e.g.,
HSE, 2005), year after year figures show high rates
of work-related accidents, even in the so-called established market economies that have enough economic
and legal resources to cope with this phenomenon
(Hmlinen et al. 2006). There are many sources of
explanations for this rather surprising fact. Many companies are unaware of the costs of accidents, and many
others, especially medium and small sized businesses,
may have limited resources for safety, including a lack
of sufficient available knowledge and material opportunities. However, the generalized lack of success on
this obvious economic matter suggests that the main
processes of organizational management should be
reviewed from a safety point of view.
Organizational behaviour, regardless of the hierarchical level of the performer, can be described as a
resultant vector compound of three main vectors: productivity, quality and safety. Although the profitable
results of the company simultaneously depend on the
three vectors, traditionally many organizations emphasize productivity or productivity and quality leading

to competition between the three vectors rather than

collaboration. For each employee at his or her workplace, the direction and strength of each of these three
vectors can be seen as the result of a complex set of
socialforcesprovidedbytheorganization, themanagers,
supervisors, co-workers and the employee him/ herself.
Itisassumedthatleadershiparrangesandgivesmeaning
to all of these forces that contribute to the definition
of the employees organizational behaviour.
Merton (1957) depicted this complete set of social
actors that can influence the organizational behaviour
of a focal role as the role-set, and Kahn et al. (1964)
analyzed role ambiguity, role stress and other role-set
dysfunctions as obstacles to the positive development
of the valued organizational behaviour. Role-set influences provide a theoretical basis for the description and
understanding of the contributions of leadership and
role claritythe opposite of role ambiguityto the
formation of safety behaviour. Role theory, through
its developments related to role stress and leadership influence, can help to understand why some role
expectations oriented toward productivity, quality or
safety can have greater probabilities of becoming organizational safety behaviour (Meli, 2004a). From a
role theory point of view, safety climate can be reinterpreted as the set of safety perceptions encompassing
the safety demands coming from each organizational

1415

actor inside the organization that arranges and gives

meaning to the safety situation. Following role theory,
safety climate can be understood as a main affective
and cognitive antecedent of safety behaviour and a
result of leadership and a clear role definition.
Different approaches to the conceptualization and
measurement of leadership in different organizational
settings have suggested that positive leadership tends
to improve subordinates organizational behaviour.
Where safety is valued, positive leadership will also
improve this facet of organizational behaviour known
as the safety response, thus affecting risk and accident
indicators. A limited number of studies have presented
some results consistent with this line of reasoning. For
example, high positive leadership conceptualized as
a leader-member exchange appeared to be associated
with subordinate safety commitment, better safety
communication and fewer accidents in 49 supervisorgroup leader dyads working in manufacturing settings
characterized by high-risk environments and high
safety concern (Hofmann & Morgeson, 1999). The
relationship between the leader-member exchange and
safety behaviour, particularly safety citizenship, when
evaluated in 127 transportation team members in the
U.S. Army, suggests that not only do leaders need to
boost effective working relationships with their subordinates, but they also must build a climate within the
team that emphasizes safety concerns. Positive leadership influences safety citizenship behaviour, but safety
climate moderates this relationship through role definition (Hofmann et al. 2003). These studies apply a
global definition of safety climate and focus on a particular definition of leadership and a facet of safety
performance, showing consistent but complex relationships among these variables. There is also some
evidence, from a study with 534 metal industry production workers, that positive leadership contributes
to a positive safety climate when safety climate is analyzed at the supervisor and group level (Zohar, 2000).
In 411 production workers, also in a metal processing
plant, transactional and transformational leadership
influences safety behaviour, but the effectiveness of
supervision safety depends on the priorities assigned
by upper management as a part of the organizational
safety response (Zohar, 2002). Also in a services sector
context, specifically in 174 restaurant workers, safety
specific transformational leadership predicted workrelated accidents through the effects of safety climate,
the workers safety response represented by safety
consciousness, safety related events and minor near
misses and micro-accidents (Barling et al. 2002). The
use of measures of micro-accidents as a more stable
safety outcome has been defended as a way of minimizing distributional problems associated with the use
of accident records (Zohar, 2000, Chmiel, 2005).
Safety climate is recognized as a multidimensional construct that should be relevant in the safety

diagnosis and intervention actions (Clarke, 2006,

DeJoy et al. 2004, Flin et al. 2000). Meli et al.
(2006) identified four main agents that perform or are
responsible for each safety climate question. Meli
& Becerril (2006) reviewed the safety climate literature from the agent point of view and identified
five main levels: organization-management, supervisors, co-workers, worker and perceived risk. Hsu
et al. (2007) divide organizational factors in safety into
four categories: organizational level factors, safety
management factors, work group factors, and individual level factors. Taking these agents into account,
safety climate can be split up into five main safety
climate variables: Organizational Safety Response
(OSR), Supervisors Safety Response (SSR), Coworkers Safety Response (CSR), Worker Safety
Response (WSR), and Perceived Risk of Accidents
(PR). The psychosocial model of work-related accidents describes and tests the role process, circumscribed to the safety responses vector, of management,
supervisors, co-workers and the worker him/herself as
focal role (Meli, 1998). This model involves an analytic view of safety climate that keeps the measure of
the safety response of each main social agent separate. The separate identification of the safety response
of each safety actor makes it possible to show the chain
of social influences from the OSR to the safety results,
such as accidents or micro accidents (MA), through
the SSR, the CSR and the WSR. The chain of safety
responses affects the PR, considered as a subjective
perception of the probability of accidents, and, finally,
the safety results (Meli, 2006).
Organizations present a huge variability in context
and structure, risk exposure and safety concern, and
these factors can affect the function of leadership and
role definition in safety. The construction industry
is, without doubt, one of the sectors with the highest accident rates and the most serious accidents in
terms of the injuries produced (e.g., HSC/HSE, 2003).
Although great progress has been made as a result
of improving preventive measures, in the EU-15 construction is the sector that shows the highest levels
of occupational injuries with more than three days of
absence, and one of the sectors with the highest rate
of fatal accidents at work (Lundholm, 2004). Previous results on the relationship between safety climate
and leadership have been obtained in formalized and
relatively stable organizational contexts, such as the
manufacturing industry or the army, where there is
a clear chain of formal power and well-defined risk
settings. The construction sector is characterized by
loose organizational chains of influence across many
subcontracting levels working in continually changing work settings where low routinization and low
behaviour control suggest the presence of many situational contingencies not easily covered by specific
safety instructions or procedures. In this situation,

1416

positive leadership and good role definition seems

particularly important, in order to achieve a positive
safety response at each level of the organization (Meli
et al. 2007).
In the EU-27, 32.7% of construction workers
reported doing short repetitive tasks of less than
one-minute, 49.1% do repetitive tasks of less than
10 minutes and 49.1% perform monotonous tasks
(EUROFOUND, 2007). In Spain, 64.6% of construction workers reported repetitive hand and arm
movements (MTAS, 2007). In many blue-collar construction jobs, a continually changing work setting
and low routinization and control are compatible
with frequent episodes of monotonous and repetitive
work (Sobeih et al. 2006; Spielhotz et al. 2006).
Monotonous and repetitive work can be understood
as a work property affected by the decisions about
technology and the organization of work adopted by
managers, and, thus, affected by OSR and positive
leadership.
Taking into account the psychosocial chain of safety
influences, positive leadership can be expected to contribute to a clear definition of roles. In this paper, both,
positive leadership and role clarity are conceived as
functional foundation climate properties, so we can
expect both to have a positive effect on the safety
responses. Additionally, positive leadership and the
OSR can be expected to have a direct negative effect on
monotonous and repetitive work, which in turn affects
PR and MA.
The purpose of this study is to investigate the influence of leadership and role clarity on the psychosocial
chain of safety influences defined by the psychosocial
model of work-related accidents in the construction
sector.

2
2.1

METHOD

2.2

Measures

All variables in this study were measured using the

HERC Battery (Meli, 2006), a set of questionnaires
designed for the assessment of the psychosocial and
behavioural risks in the construction sector. All items
were answered on an 11-point Likert-type scale, where
0 expresses the minimum level and 10 the maximum.
Table 1 presents the mean, standard deviation and
alpha coefficient of the measures considered in this
paper. All reliabilities, ranging from 0.72 to 0.94, can
be considered satisfactory.
The OSR was measured using an 8-item questionnaire that referred to tangible safety actions by
the company. These safety actions included: 1) The
presence of safety structures, 2) Organizational fulfilment of safety norms, 3) Safety inspections, 4)
Safety training and information, 5) Safety meetings,
6) Promotional campaigns and 7) Safety incentives
and sanctions.
The SSR scale consists of 5 items measuring
selected supervisors safety actions that can affect
workers safety behaviour: 1) Providing models to the
worker through the supervisors own safe or unsafe
behaviour, 2) Reactions, reinforcements or punishments for the workers safe or unsafe worker behaviour,
and 3) Active encouragement of worker safety
behaviour through instructions and promotion of
safety issues.
The CSR was considered by means of a 5-item
scale. The co-workers safety response embraces three
aspects: 1) Supplying models of safe or unsafe behaviour through the co-workers own safe or unsafe
behaviour, 2) Reactions to the safe or unsafe behaviour
of the peer, and 3) Active support of safety.
The 5-item WSR scale measures the safety response
of the worker. This response covers three aspects: 1)
The workers safe or unsafe behaviour, 2) The workers
reactions to their own safe or unsafe behaviour, and 3)
The workers encouragement of safety.

Sample

The sample is composed of 789 Spanish construction workers who voluntarily agreed to participate in the study. Data was obtained during
professional construction training sessions performed
by a Spanish occupational foundation for the construction sector in 10 Spanish provinces. 90.6% of
the respondents were male, and 9.4% were female.
Average age of respondents was between 30 and 39
years. 43.1% fell into this category. 33.5% were
between 18 and 30 years of age, 18.2% between
40 and 49 years of age, and 5.2% were over 50
years of age. The sample performed seven construction jobs: 21.0% bricklayers, 11.8% mates, 11.1%
moulding carpenters, 6.4% installers, 11% crane
operators, 14.2% foremen, and 24.5% construction
technicians.

Table 1. Descriptive statistics (means and standard deviations) and reliabilities (coefficient alpha) for each scale.
Scale

Mean SD

Organizational Safety Response

(OSR)
Supervisors Safety Response (SSR)
Co-workers Safety Response (CSR)
Workers Safety Response (WSR)
Monotonous and Repetitive Work
(MRW)
Perceived Risk of accidents (PR)
Positive Leadership (LEA)
Role Clarity (CLR)
Micro-Accidents (MA)

6.06

2.43 0.90

6.59
5.59
7.69
4.31

2.64
2.53
1.89
2.43

0.94
0.93
0.84
0.72

3.97
6.52
3.19
3.57

1.96
1.89
2.01
2.20

0.89
0.73
0.77
0.83

1417

Alpha

The MRW scale has 3 items directly related to the

frequency of monotonous and repetitive work (e.g.,
In my job I have to do repetitive and monotonous
tasks over and over again).
The PR was measured using an 8-item scale that
refers to the workers perception about his or her
own probability of suffering a work-related accident
or illness, including occupational illness. The scale
explores this through several degrees of injury severity
and through several classes of injuries and illness.
The LEA scale measures positive leadership by
means of a 5-item Likert scale. The items measured
the following positive leadership properties: 1) professionalism and knowledge, 2) facing up to and coping
with problems, 3) correct treatment of subordinates,
4) supporting subordinates, and 5) motivating their
participation.
The CLR scale measures role claritythe opposite
of role ambiguity- by means of 5 items. A high score
on this variable means that the employee knows clearly
the objectives, methods and consequences of his/her
performance.
The MA scale contains 2 items that measure the frequency of micro-accidents, such as small cuts, scrapes
or abrasions. Micro-accidents have been considered a
favourable final safety indicator, compared with accidents or other classical indicators, due to its greater
sensitivity and lesser skewness (Chmiel, 2005).
The effects of LEA and CLR and the other relationships involved in the psychosocial model of workrelated accidents was tested by means of an equations
structure model using EQS software.

.48**

LEA

.40**.23**
.32**

OSR

.33**

RESULTS

Figure 1 shows the results of the structural equations

model. The model shows an acceptable global fit
(NFI = .99; NNFI = .97; CFI = .99, AASR =
.0161). Figures on the unidirectional arrows are standardized path coefficients.
The Psychosocial Model. Results confirm the
psychosocial chain of effects described from the organizational safety response to the indicator of microaccidents. The OSR has a main set of positive and
significant effects on the SSR, the CSR and the WSR,
and it also has a negative effect on the PR and the
MA. The SSR positively affects the CSR and also
the WSR. The CSR positively affects the WSR. The
PR negatively affects the psychosocial chain of safety
influence through the WSR and the OSR; however,
the effect of the WSR does not raise the level of significance. Finally, MA is affected significantly and
negatively by the organizational safety response and
positively by the PR.
The role of LEA, CLR and MRW. LEA plays a major
role, having a direct and positive effect on the CLR,

.14**

-.21**
SSR

.33**
.07**

.07*

.25**

CSR

.30**
.16**

-.21**
MRW

.11**

.22**

WSR

-.16**
.40**

-.05

PR

-.09*

.32**
.44**

MA

CLR

-.09**

Figure 1. Psychosocial Model of Leadership and Safety Climate. Figures on the unidirectional rows are standardized
path coefficients. = p < 0.05; = p < 0.01.

the OSR and the SSR, and to a minor degree the CSR.
LEA also affects the MRW and the PR directly and
negatively. CLR positively affects the OSR, the SSR,
the CSR and the WSR. Finally, MRW has a direct and
positive effect on PR and MA.

CONCLUSIONS AND DISCUSSION

These results confirm the main flow of safety influences between the organization and the individual
behaviour, and the importance of leadership, role
clarity and monotonous and repetitive work in the
assessment of the psychosocial set of variables that
make up the safety climate.
The path analysis coefficients corroborate the main
social process of influence between the organization

1418

and the workers safety response, through the supervisors safety response and the group influences represented by the co-workers safety response. This result
confirms previous research testing the psychosocial
model in general samples (Meli, 1998; 2004b) and in
construction samples (Meli et al. 2007). In this construction sample, the paths linking the various safety
responses are strong, as is the step between the perceived risk and the micro-accidents. However, the
paths linking the safety responses to the perceived
risk and the micro-accidents are weak, especially the
path between the workers safety response and the
perceived risk.
Positive leadership emerges as a powerful exogenous variable directly affecting role clarity, the chain of
safety responses and monotonous and repetitive work.
Positive leadership contributed to explaining a clear
definition of role, and both positive leadership and role
clarity contributed to a positive safety response by the
organization, the supervisors and, to a lesser degree,
the co-workers. Role clarity also directly affects the
workers safety response.
The role of perceived risk is clearly affected by
positive leadership, with a negative sign, and by
monotonous and repetitive work, with a positive sign.
This chain of social influences gives structure to the
different safety responses contained under a holistic safety climate perspective, and it also confirms
the expected relationship between perceived risk and
micro-accidents.
In this research, a realistic view of perceived
risk is assumed; that is, the risk that is perceived
reflects the degree of safety of the work setting
given the behaviour of the social agents involved.
From this point of view, risk represents an estimation of the perceived probability of undesired
events, such as accidents or micro-accidents, and is
a result of the safety behaviours of all the social
agents. However, perceived risk can also be understood and analyzed from a subjective point of view,
as a result of the perceived level of undesired events,
such as accidents and micro-accidents, and as an
antecedent of the social chain of safety responses.
Both views can be accepted as partially true, and
risk can be understood both as a perceptive estimation of the probability of accidents and other
undesired events given the safety conditions and the
safety responses, and as a cognitive antecedent of
these safety responses influenced by the perception
of accidents and other undesired events. However,
it is hard to assume a subjective view of perceived
risk when we focus on the relationships with positive leadership and monotonous and repetitive work.
The results provided by the tested model in this construction sample suggest that the perceived risk is
strongly affected by the presence of monotonous and
repetitive worka characteristic of the tasks and the

organization of workand also by poor leadership.

Monotonous and repetitive work also directly affects
micro-accidents.
The social distance between the worker and the
company that characterizes the construction sector
should explain lower bonds between the organizational
safety response and the worker safety response. However, results show that in this sample the organizational
safety response and the supervisors safety response
are relevant predictors of the workers safety response
together with the co-workers safety response. All the
sets of social responses are consistently and significantly related, and there is little evidence of the effect
of social distance on the workers safety response.
An unexpected gap occurs between the workers
safety response and the perceived risk. This path works
in the expected direction, but it is both very low and
non-significant. This gap cannot be explained by the
social distance characteristics of the construction sector. In fact, the only safety response that shows a
significant negative effect on the perceived risk is
the organizational safety response. Perhaps a defensive attribution bias prevents the expected relationship
between the workers safety response and the perceived risk. The means and standard deviations of
the variables clearly reveal a positive self-report bias
that is probably also a perception and attribution bias.
While the mean OSR is 6.06, the mean SSR is 6.59,
and the CSR is 5.59, all three with a SD of around 2.5,
the WSR is 7.69 with a noticeable shrinkage in the
SD to 1.89. That is, employees attribute to themselves
a safety response 1.05 points better than the safety
responses of their supervisors (t = 12.37, p < 0.001),
1.62 points better than the safety responses of their
company (t = 20.95, p < 0.001) and 2.07 points better than their co-workers (t = 26.46, p < 0.001). This
response bias makes it difficult to find a relationship
between WSR and PR. Our data do not allow us to
decide whether we should interpret this as a defensive
report or as a perceptive or attribution bias. In any case,
from a preventive point of view, this bias has negative
consequences, precluding a realistic involvement in
safety.
There are two main sources of explanations for
the perceived risk. One is a task characteristic,
monotonous and repetitive work, and the other is an
organizational characteristic, poor leadership. Results
suggest that perceived risk for construction employees is mainly a matter of the characteristics of the
task and the social properties of the organization, such
as leadership and organizational safety response. The
role of monotonous and repetitive work as a main predictor of the perceived risk and micro-accidents can
be understood in two ways. First, monotonous and
repetitive work is a source of risk by itself, facilitating some kind of accidentse.g., muscle skeletal
disorders (Sobeih et al. 2006; Spielholz, et al. 2006).

1419

Second, the presence of monotonous and repetitive

work acts as a marker, helping to identify kinds of
work characterized by salient risks. Thus, in the construction sector where there is more monotonous and
repetitive work, there are also other associated types of
risk. Moreover, micro-accidents are mainly predicted
by perceived risk and monotonous and repetitive work.
Our results can be considered coherent with the
conclusions of Hofmann and Morgeson (1999), Zohar
(2000, 2002), Barling et al. (2002) and Hoffmann et al.
(2003). According to these authors, positive leadership
influences safety climate and safety climate moderates
the effects of leadership directly and through role definition. Moreover, positive leadership plays a direct
role in the perceived risk, and it influences work characteristics, such as monotonous and repetitive work,
which has direct effects on safety indicators, such as
micro-accidents.
The results confirm the psychosocial chain of influence among the safety responses of the different
organizational agents, and they also emphasize a main
contribution of positive leadership, not only as an
antecedent of the safety responses, but also as a direct
predictor of perceived risk and monotonous and repetitive work. These results suggest that a systematic
evaluation of leadership, role clarity and monotonous
and repetitive work should be incorporated into the
risk assessment procedures, not only due to their
importance in the psychosocial health and stress facets
of work, but also because of their direct and main
contributions to the safety perspective.
ACKNOWLEDGEMENTS
This paper was developed within the CONSTOOLKIT Project, [BIA2007-61680]. Financial support was
provided by the Ministerio de Educacin y Ciencia
(Spain) and The European Regional Development
Fund (FEDER). [Website: www.uv.es/seguridadlaboral].
REFERENCES
Barling, J., Loughlin, C. & Kelloway, E.K. 2002. Development and test of a model linking safety-specific transformational leadership and occupational safety. Journal of
Applied Psychology, 87(3), 488496.
Chmiel, N. 2005. Measuring the invisible: Self-report minor
injuries, and their relationship to safety behaviour and
safety climate in the workplace. Paper presented at the
XII European Congress of Work and Organizational
Psychology. Istanbul.
Clarke, S. 2006. The relationship between safety climate and
safety performance: a meta-analytic review. Journal of
Occupational Health Psychology, 11(4), 315327.
Dejoy, D.M., Schaffer, B.S., Wilson, M.G., Vandenberg,
R.J. & Butts, M.M. 2004. Creating safer workplaces:

assessing the determinants and role of safety climate.

Journal of Safety Research, 35(1), 8190.
EUROFOUND. 2007. Fourth European Working Conditions
Survey. Dublin: European Foundation for the Improvement of Living and Working Conditions.
Flin, R., Mearns, K., OConnor, P. & Bryden, R. 2000. Measuring safety climate: identifying the common features.
Safety Science, 34, 177192.
Hmlinen, P., Takala, J. & Saarela, K.L. 2006. Global
estimates of occupational accidents. Safety Science, 44,
137156.
Hofmann, D.A. & Morgeson, F.P. 1999. Safety-related
behavior as a social exchange: The role of perceived organizational support and leader-member exchange. Journal
of Applied Psychology, 84(2), 286296.
Hofmann, D.A. Morgeson, F.P., & Gerras, S.J. 2003. Climate
as a moderator of the relationship between leader-member
exchange and content specific citizenship: Safety climate
as an exemplar. Journal of Applied Psychology, 88(1),
170178.
HSE 2004. Improving health and safety in the construction
sector. London: Health and Safety Executive. http://www.
nao.org.uk/publications/nao_reports/03-04/0304531.pdf
HSE 2005. Accident costs. http://www.hse.gov.uk/costs/
Kahn, R.L., Wolfe, D.M., Quinn, R.P., Snoek, J.D. &
Rosenthal, R.A. 1964. Organizational Stress: Studies in
role conflict and ambiguity. New York: Wiley and Sons.
Hsu, S.H., Lee, C., Wu, M. & Takano, K. 2007. A crosscultural study of organizational factors on safety: Japanese
vs. taiwanese oil refinery plants. Accident Analysis &
Prevention, 40(1), 2434.
Lundholm, L. 2004. European statistics on accidents at work;
The number of work accidents in the member stateshas
it decreased in the last 10 years? with particular focus
on the construction area. Swedisk Work Environment
Authority.
Meli, J.L. 1998. A psychosocial causal model of
work related accidents. Anuario de Psicologia, 29(3),
2543.
Meli, J.L. 2004a. A Methodology for the Measurement of
Complex Power Interactions and Social Positions in Organizations. In Miguel Pina e Cunha and Lourdes Munduate
(Eds). Poceedings of the Fourth Symposium on Power
Dynamics and Organizational Change. ISPA Edioes.
Lisboa. 147204.
Meli, J.L. 2004b. The Psychosocial Causal Model of workrelated accidents of the University of Valencia. Perspectives and new developments. In R. Mondelo, P. Mattila,
M.; Karwowski, W., & Hale, A. (eds.), Proceedings of
the Third International Conference on Occupational Risk
Prevention. Santiago de Compostela.
Meli, J.L. 2006. The psychosocial model of work-related
accidents: A test in the construction industry. In Scott
McIntyre & Jonathan Houdmont: Occupational Health
Psychology: Key papers of the European Academy of
Occupational Health Psychology. 7th Conference of the
European Academy of Occupational Health Psychology.
Dublin: Edioes ISMAI. 171.
Meli, J.L. & Becerril, M. 2006. Safety climate dimensions from the agent point of view. In Mondelo, P;
Mattila, M., Karwowski, W., Hale, A. (Eds.) Proceedings
of the Fourth International Conference on Occupational
Risk Prevention.

1420

Meli, J.L., Mearns, K., Silva, S. & Lima, L. 2007. Safety climate and the perceived risk of accidents in the construction
industry. Safety Science, doi: 10.1016/j.ssci.2007.11.004.
Meli, J.L., Silva, S., Mearns, K. & Lima, L. 2006. Exploring
the dimensionality of safety climate in the construction
industry. In Mondelo, P., Mattila, M., Karwoski, W. &
Hale, A. Proceedings of the Four International Conference on Occupational Risk Prevention.
Merton, R.K. 1957. Social theory and social structure (2nd
Edn.). Glencoe, IL: Free Press.
MTAS. 2007. VI Encuesta Nacional de Condiciones de
Trabajo.
Sobeih, T.M., Salem, O., Daraiseh, N., Genaidy, A.
Shell, R. 2006. Psychosocial factors and musculoskeletal

disorders in the construction industry: a systematic review.

Theoretical Issues in Ergonomics Science, 7(3), 329344.
Spielholz, P., Davis, G. & Griffith, J. 2006. Physical Risk Factors and Controls for Musculoskeletal Disorders in Construction Trades. Journal of Construction Engineering
and Management, 132, 10591068.
Zohar, D. 2000. A group-level model of safety climate:
Testing the effect of group climate on microaccidents
in manufacturing jobs. Journal of Applied Psychology,
85(4), 587596.
Zohar, D. 2002. The effects of leadership dimensions, safety
climate, and assigned priorities on minor injuries in
work groups. Journal of Organizational Behavior, 23(1),
7592.

1421

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Local management and its impact on safety culture and safety

within Norwegian shipping
H.A. Oltedal
University College Stord/Haugesund, Norway

O.A. Engen
University of Stavanger, Norway

ABSTRACT: This paper addresses safety culture on tankers and bulk carriers and which factors affect the
safety culture onboard vessels. The empirical setting for the study is the Norwegian shipping industry. Safety
management is a challenging issue within shipping for several reasons. First of all, life and work onboard a vessel
is a 24 hour activity and the crew has few possibilities of interacting with the surrounding society. Secondly
the geographical distance between the on-shore organization and the vessel may affect both the quality of those
systems and plans developed on shore and their implementation on the vessels. The ship management is thus
identified as a key factor to a sound safety culture along with the on shore crewing strategy.

INTRODUCTION

In this paper we will discuss the safety culture within

the Norwegian shipping industry with tankers and bulk
carriers, and identify which organizational factors may
affect this particular safety culture.
In Norway, shipping has for several centuries
been the principal trade, and Norway as a maritime
nation has roots way back in the Viking age. Today
Norway is one of the five largest shipping nations in
the world, after Greece, Japan, Germany and China. In
the third quarter of 2007 the Norwegian foreign-going
fleet comprised 1,795 ships, the highest number ever
in Norwegian history, of which about 49 percent are
flying under the Norwegian flag (Nrings- og handelsdepartementet 2007). The remaining 51 percent may
register in any of the worlds more than 150 flag states.
Norwegian shipping companies employ some 57,000
seamen from more than 60 different nationalities and
of which about 30 percent are Norwegian Nationals
(Norwegian Shipowners Association). The crew may
be recruited and managed by the shipping company
itself, or by one of the worlds many professional crew
hiring companies. Within the Norwegian fleet, most
sailors are contract-employees working on different
vessels during each enrolment, which results in continually shifting working groups. The situation today
is a result of a structural change dating back to the
60s and 70s when technical development allowed for
bigger vessels with more automation and monitoring,

along with the need for reorganization to improve efficiency. This resulted in a cut in the crewing level. Later
in the 80s a global recession caused further structural
changes; flagging-out, use of external crewing agencies and signing on crew from developing countries
and lower wages (Bakka, Sjfartsdirektoratet 2004).
However, the shipping industry is today facing new
manning related challenges as there is a global shortage of manpower, this is due to three main challenges:
First, it is less attractive nowadays to work in the shipping industry. Second, the recruitment for ship crews
has been slow. This has resulted in the third situation
where the liquefied natural gas (LNG) shipping sector is drawing crew from the tanker industry, and the
tanker industry in turn is drawing people from the dry
bulk sector.
In 1894 the British Board of trade carried out a
study which showed that seafaring was one of the
worlds most dangerous occupations, and it still is
(Li, Shiping 2002). Regulations in order to reduce
the risk at sea were introduced about 150 years ago.
These regulations initially encompassed measures to
rescue shipwrecked sailors, and further requirements
for life-saving equipment, seaworthiness and human
working conditions. Traditionally the safety work has
focused on technical regulations and solutions even
though experience and accident statistics indicate that
most of the accidents at sea somehow were related
to human performance (Bakka, Sjfartsdirektoratet
2004). However, a few very serious accidents at seathat

1423

occurred in the late 80s resulted in a change towards

how safety was organised, and more focus was given
to the human barriers and how the seafarers working conditions were affected by organisational and
managerial factorsboth on shore and at sea. Along
with this the term safety culture started to gain a
foothold also within shipping. The idea of safety culture within shipping was officially introduced on the
4th November 1993 by the adoption of a new resolution, the present SOLAS Convention 1974 Chapter
IX, entitled Management for the Safe Operation of
Ships and for Pollution Prevention, also known as the
International Safety Management Code (ISM Code)
(Le Meur 2003).
Hence, the main purpose of this paper is to elaborate
the following questions:

Person
Safety
Climate

CONTEXT
External Observable Factors

Situation
Organisational
l Factors

APPROACH TOWARDS SAFETY CULTURE

There seems to be no clear consensus concerning

the ontological, epistemological, and methodological
questions related to the topic of safety culture. The
main differences seem to be
1. Definition of the scope of safety culture and the
relationship between culture and climate.
2. Which methods are regarded as most suitable for
measurement.
3. The relationship to other organisational (safetyrelated) aspects (Cooper 2000, Guldenmund 2000,
Neal, Griffin & Hart 2000, Peterson, Ashkanasy &
Wilderom 2000, Sorensen 2002, Yule 2003).
However, it is not the scope of this paper to problematise the concept of safety culture. As a point
of departure we will apply Scheins definition of
organisational culture:
A pattern of shared basic assumptions that the
group learned as it solved its problems of external

Behavior
Safety
Behavior

Figure 1. Reciprocal safety culture model (adopted from

Cooper, 2000).

What characterises safety culture on tankers and

bulk carriers?
Which factors affect the safety culture on board
vessels?
With reference to shipping, this article will more
concretely analyse crewing strategies such as outsourcing of crewing management and the extended use
of contract employment instead of permanent employment. Our hypothesis is that these conditions may contribute to an unfavourable and error-inducing working
environment, i.e. poor communication between shore
management and the ship management and the remaining crew, unworkable procedures, lack of loyalty to
the organisation, dysfunctional interaction, fear of
reprisals, which again counteract the development of
a safety culture.

Internal
Psychological
Factors

adoption and internal integration, that has worked

well enough to be considered valid and, therefore, to be taught to new members as the correct
way to perceive, think and feel in relation to those
problems(Schein 2004).
Further we have decided to use a methodological framework presented by Cooper (2000), and the
application of this framework will be discussed below.
Cooper (2000) introduces a reciprocal model of safety
culture that allows the multi-faceted and holistic nature
of the concept to be fully examined by using a triangular methodology approach, depicted in Figure 1.
Coopers (2000) model contains three elements:
1. The subjective internal psychological factors i.e.
attitude and perceptions.
2. Observable on-going safety related behaviour.
3. The organisational situational features.
According to Cooper (2000) these elements reflect
those accident causation relationships found by a
number of researchers, such as John Adams, Herbert
William Heinrich and James Reason. The investigation of several serious shipping accidents such
as Herold Free Enterprize (Department of Transport
1987), Exxon Valdes (National Transportation Safety
Board 1990) and Scandinavian Star (Justis- og politidepartementet 1991) is also congruent with their
findings. The Herold Free Enterprize accident was
partly caused by members of the crew not following
best practice, but was also due to managerial pressure from the organizations upper level to sail as early
as possible, along with other mistakes made by the
on-shore management. Two years later when the US
tanker Exxon Valdes grounded, the accident investigation determined several probable causes linked to
human errors induced by managerial faults in the upper
levels of the organisation. At the vessel, the third mate
failed to properly manoeuvre the vessel, possibly due
to fatigue and excessive workload. The master failed

1424

to provide a proper navigation watch, possibly due to

impairment from alcohol. At the onshore part of the
organisation, the shipping company fails to supervise
the master and provide a rested and sufficient crew
for the Exxon Valdez. In addition to this effective
pilot and escort services were lacking. The following year, in 1990, there was a fire on the passenger
liner Scandinavian Star. In the aftermath of this
accident the investigation brought into focus organisational and managerial faults with regard to a lack
of competence and training, but also weaknesses in
the wider social-technical system. These weaknesses
consisted of ownership separated from management,
unsatisfactory control routines by the flag state and,
in general, an overall maritime system with a lack of
transparency. Further, Coopers (2000) three elements
will be outlined more in detail, starting with safety
related behaviour.
2.1

The importance of safety related behaviour

Herbert William Heinrich work (published in 1931

Industrial Accident Prevention) is the basis for the theory of Behaviour-Based Safety (BBS), which holds
that as many as 8090 percent of all workplace accidents are caused by human error and unsafe acts
(Tinmannsvik, 2008). Scheins (2004) definition of
culture does not clearly address observable behaviour
patterns, but behaviour is regarded to be partly
determined by a persons perceptions, feelings and
thoughts. However, Schein (2004) regards behavioural
patterns as a manifestation of a culture existing at a
higher level in the organisation, and not as culture
itself. When it comes to BBS, the current theories
posit that safety culture, and a reduction of accidents
may be achieved through continuous attention to three
domains:
1. Environmental factors such as equipment, tools,
physical layout procedures, standards, and temperature.
2. Person factors such as peoples attitudes, beliefs,
and personalities.
3. Behaviour factors, such as safe and at-risk work
practices, referred to as the Safety Triad (Geller
2001).
When adopting this approach humans are seen
as a cause of accidents, whereupon interventions
to enhance safety are aimed at changing attitude
or behavior (i.e. poster campaigns, training, procedures and so on, or changing the technology they
operate). This orientation towards risk and safety management has traditionally been and still is adopted from
the majority of the shipping companies. The BBSapproach has been criticised for placing too much
responsibility on the people operating the systems,
assuming that they are responsible for the outcome

of their actions (Dekker & Dekker 2006). An alternative view is to recognise human error not as a cause
of accidents, but as a consequence or symptom of
organisational trouble deeper within the organisation,
arising from strategic or other top level decisions.
This includes resource allocation, crewing strategy
and contracting (Dekker, Dekker 2006, Reason 2001,
Reason, Hobbs 2003). An organisation is a complex
system balancing different, and often also conflicting,
goals towards safety and production in an aggressive
and competitive environment (Rasmussen 1997), a
situation that to a large extent is current within shipping. The BBS approach towards safety often implies
that more automation and tighter procedures should be
added in order to control the human actions. However,
the result may be that more complexity is added to the
system. This in combination with the organisations
struggle to survive in the competitive environment,
leads to the system becoming even more prone to accidents (Perrow 1999) (Dekker & Dekker 2006, Reason
2001, Reason & Hobbs 2003). However, the concept
of focusing on the human side of safety is not wrong.
After all, the technology and production systems are
operated, maintained and managed by humans, and
as the final barrier towards accidents and incidents
they are most of the time directly involved. The proponents of the BBS approach argue that behaviour
control and modification may bring a shift in an organisations safety culture, also at the upper level, but this
is most likely if the focus is not exclusively addressing
observed deficiencies at the organisations lower levels (DeJoy 2005). DeJoy (2005) calls attention to three
apparent weaknesses related to the BBS approach:
1. By focusing on human error it can lead to victimblaming.
2. It minimises the effect of the organisational
environment in which a person acts.
3. Focusing on immediate causes hinders unveiling the basic causes, which often reside in the
organisational environment.
Due to this, we will also include the organisational
environment in the safety culture concept, as proposed
by Cooper (2000).
2.2

The relation to organisational factors

When human error is not seen only as a cause of accidents, but as a symptom and consequence of problems
deeper inside the organisation, or what Reason (2001,
2003) refers to as latent organisational factors, emphasis is placed on weaknesses in strategic decisions made
at the top level in the organisation. These strategic
decisions may reflect an underlying assumption about
the best way to adapt to external factors and to achieve
internal integration, and if they are common for most
shipping companies an organisational culture may also

1425

be revealed (Schein 2004). Schein (2004) also stresses

the importance of leadership. The top management
influences the culture as only they have the possibility of creating groups and organisations through their
strategic decisions. And when a group is formed, they
set the criteria for leadership and how the organisation
will support and follow up their leaders. The leaders, at
all levels on shore and at the vessel, are also key figures
in the development of a safety culture. It is their task
to detect the functional and dysfunctional elements of
the existing culture, and to channel this to the upper
levels of the organisation. In return, the upper levels of
the organisation should give their leaders the support
necessary in order to develop the desired culture.
3

METHODOLOGICAL IMPLICATIONS

Development of questionnaire items

The survey instrument was developed by Studio

Apertura in collaboration with DNV and SINTEF. The
development was based on an evaluation of seven
already existing questionnaires in comparison with
various theoretical views of the safety culture concept (Studio Apertura 2004). Minor adjustments were
made after a pilot for use within the tanker and bulk
carrier sector. This resulted in a questionnaire with
constructs and accompanying number of items as presented in table 1. All items were measured on a 5 point
likert scale ranging from strongly disagree to strongly
agree, or very seldom/never to very often/always.
3.2

Questionnaire constructs and number of items.

Construct

Number of items

Top managements safety priorities

Local management
Procedures & guidelines
Interaction
Work situation
Competence
Responsibility & sanctions
Working environment
Reporting practices

3
7
7
18
8
5
7
9
10

response rate of 80% and a vessel response rate of

91.5%. The survey was carried out in 2006.

Coopers (2000) framework put forward the importance of methodological triangulation in order to grasp
all facets of the cultural concept. The internal psychological factors are most often assessed via safety
climate questionnaires. Our approach is to start with
such a survey in order to gain insight into the seafarers
perceptions and attitudes related to safety, along with
self-reported work behaviour related to risk taking,
rule violation and accident reporting. The survey also
includes questions related to crewing strategy, which
opens up the possibility of assessing the relationship between the organisational situation and actual
behaviour. The survey results are used to determine
which organisational factors are most likely affect the
safety culture, and to define research areas for a further
qualitative study.
3.1

Table 1.

Questionnaire sample

A total of 1574 questionnaires were distributed to 83

randomly selected Norwegian controlled tankers and
bulk carriers. All vessels were flying a flag on the Paris
MOU white or grey list. 76 vessels returned a total
of 1262 completed forms, which gives an individual

3.3 Statistical analysis

The Statistical Package for the Social Sciences (SPSS)
v.15.0 was used to perform all of the analysis, which
included descriptive statistics, exploratory factor analysis (EFA), confirmatory factor analysis (CFA) and
bivariate correlation analysis.
With regard to the EFA, the principal component
analysis with Varimax rotation was carried out. The
factors were extracted based on the three following
analytical criteria: (1) Pairwise deletion, (2) Eigen
value more than 1, and (3) factor loading more than
0.50. Of the extracted factors, all factors with 2 or
fewer items were removed, based on the notion that a
factor should be comprised of at least three items to be
robust (Pett, Lackey & Sullivan 2003). A confirmatory
factor analysis (CFA), using a one-factor solution for
each construct, has also been performed. The advantage of the CFA is that it allows for more precision in
evaluating the measurement model (Hinkin 1995), and
the results were compared with the EFA for providing
validity evidence based on the hypothesis that a valid
instrument should produce similar results.
Each factor was then evaluated using the KaiserMeyerto-Olkin (KMO) parameter, and only factors
with KMO value at 0.60 or above were included in
the further analysis (Hair 1998).
This was followed by a scale-reliability test. For that
purpose, the Cronbachs Alpha coefficient of internal
consistency was calculated, and evaluated along with
inter-item statistics. Cronbachs Alpha is a measure
of scale reliability concerned with the proportion of a
scales total variance that is attributable to a common
source, presumed to be the true score of the latent
construct being measured. In our case that will be
the safety culture. Usually a value above 0.7 is considered acceptable, although some advocate an alpha
level of 0.8 or better (Netemeyer, Sharma & Bearden
2003). As the alpha value is a function of, inter
alia, the average inter-item correlation; the inter-item

1426

correlation and item-total statistics have also been

evaluated. Rules of thumb suggest that the item-total
correlation should exceed .50, and the inter-item correlation should exceed .30 (Hair 1998), but it should
not exceed .80. An inter-item correlation exceeding
.80 suggests that items are duplicates of one another
(Pett, Lackey & Sullivan 2003). Then the remaining
items went through a last CFA, a five-factor solution,
in order to provide each factors explained variance.
Finally, correlation analysis has been carried out
in order to evaluate the construct validity, which is
viewed as the extent to which an operational measure
truly reflects the underlying safety culture concept,
and if they operate in a consistent manner.
Based on this analytical process, five factors (1)
interaction, (2) reporting practices, (3) competence,
(4) local management, and (5) work situation were
found to be reliable and valid. The aforementioned
factors are presented further in detail in the next
section.
4
4.1

RESULTS
Results from descriptive analysis

Regarding demographics, 21 different nationalities are

represented. The Filipino contingent forms the largest
group constituting 63% of the sample, followed by
the Norwegian group with almost 11%, and the Polish
which represents 9%. The last major group was the
Russians with 6%. The other remaining 17 nationalities were represented in a range from 3% to 1%.
There is also great variation with regard to employment conditions. All in all, 12% of the sample
consists of permanent employees, of whom 80% are
Norwegian and 16% from the European Union. 91% of
the Norwegians are permanent employee. The remaining 9% are apprentices, substitutes or newly employed
on probation. Only 3% of the non Norwegian sailors
are permanent employees. With regard to the Filipino
seafarers, the largest nationality, 99.6% are contract
employees, most on 9 month contracts (62%), followed by 6 month contracts (27%). The extended use
of contract employment is reflected in their experience.
All in all, 85% had three years or more experience
within shipping in general. However, 69% of the sample had worked on the current vessel for only 1 year
or less.
The employment terms were in general different for
the captains. The captains normally do not have sailing
periods that exceed 6 months. The most typical sailing
period for the captains is 3 months or less.
4.2

Results from factor analyses

From the 9 theoretical safety culture constructs, a

five factor solution was derived, (1) interaction, (2)

reporting practices, (3) competence, (4) local management, and (5) work situation. With regard to the local
management, competence and work situation
factor both EFA and CFA result in final solutions consisting of the same items, but with minor differences
in factor loading. The CFA included three more items
in the interaction factor than the EFA, and the final
factor, reporting practices resulted from only the
CFA.
Four of the constructs did not pass the reliability
tests. The first, top managements safety priorities,
was excluded due to low representative reliability
across subpopulations. This construct also consisted of
too few items. The remaining three constructs, procedures and guidelines, responsibility and sanctions
and working environment were excluded due to
low validity, mostly resulting from poor theoretical
relationship within the items of each construct.
For the further analysis the results from the CFA are
used. The 5 factors in question are presented in Table 2
along with number of items and explained variance.
Each factors Cronbachs alpha value and inter item
statistics is presented in table 3.
The alpha values range from .808 to .878, and the
internal item statistics are all within the recommended
levels. The five factors are therefore considered to be
a reliable and valid reflection of the underlying safety
culture concept.
Further, table 4 presents the correlation coefficients
between the factors, or safety culture dimensions. All
correlations are significant at the 0.01 level (2-tailed)

Table 2. Final factors, number of items and explained

variance.
Factor

Number
of items

Explained
variance

Interaction
Reporting practices
Competence
Local management
Work situation

8
5
4
3
3

35.63%
9.77%
7.12%
5.96%
5.08%

Table 3. Final factors, Cronbachs alpha and inter-item

statistics.
Factor

Alpha

Inter-item
range

Item-total
range

Interaction
Reporting practices
Competence
Local Management
Work situation

.878
.808
.839
.866
.817

.360.606
.335.761
.497682
.692.716
.512.749

.520.724
.491.668
.628.712
.724774
.554.739

1427

Table 4.

Factor correlation matrix. Pearsons r.

F1

F1: Interaction
F2: Reporting
practices
F3: Competence
F4: Local
management
F5: Work
situation

F2

F3

F4

.352
.639

1
.323

.474

.362

.367

.494

.322

.441

.444

F5

The five safety culture dimensions correlate in

a positive direction, which is consistent with the
theoretical concept, and they are therefore considered
to be a valid reflection of the underlying safety culture
construct.

DISCUSSION

All three constructs have a good alpha level, and as

the alpha levels are concerned with the variance that
is common among the items, these constructs also
reflect the areas where it is possible to speak about
safety culture. With reference to Coopers framework
towards safety culture, we will further discuss how
the organisations factors such as crewing strategy,
witch includes employment terms, rotations system
and policy towards the on board shipping management,
may affect the on board safety culture and climate
represented by the identified dimensions. The organisations structural factors are all to be found within
Coopers element of situation, while the identified
safety culture dimensions are to be found within the
elements of person and behaviour.
Interaction is the dimension accounting for the
largest proportion of the total explained variance, with
35.63%, meaning that with regard to safety culture
most of the variance in the original data is explained
by this dimension. When taking into account how
distinctive a ship is as a work place, this is no surprise. A ship may be characterised as a total institution
since both work and leisure time happen at the same
place and with few possibilities to interact with the
surrounding world (Goffman 1968). In such a setting the crew members are socialised into a common
culture and rules of interaction. Schein (2004) refers
to this as internal integration. The interaction climate
is characterised by lack of stability within the crew
due to different terms of employment. First of all,
permanent employment seems to be reserved for the
Norwegian sailors. Sailors of other nationalities are
almost all contract employees. In addition, the length
of contract varies and all crew members have different

dates for signing on and signing off. Schein (2004)

points out that lack of stability may be a threat to the
possibility of developing a culture: ( . . . ) there must
be a history of shared experience, which in turn implies
some stability of membership in the group. Even if
the crew as a group is in constant change, they all have
common history as seafarers. So even if lack of stability within the group indicates that a common culture
should not develop on the ship, a common culture of
how to act and interact may have developed amongst
the seafarers, and when a new crewmember is signed
on a new vessel, he knows what is expected from him.
However, the question is if such a culture is a safe
culture? Reason (2001, 2003) emphasize that to reach
a safe culture, the organisation should strive for an
informed culture where those who manage and operate
the system, both on board and on shore, have current
knowledge about the factors that determine the safety
of the system as a whole, which again depends on that
the crew on board are prepared to report their errors
and near misses, and the reporting practice is one of
the dimensions deriving from the analyses, explaining
9.77% of the variance. This dimension also includes
feedback on reported events. In order to attain good
reporting practices, the organisation should strive to
create an atmosphere of openness, trust and loyalty.
Integrating into the group is also a survival mechanism, and every crewmember will most likely make an
effort to integrate. If not, he would most likely have a
hard time during his contract period with no possibility
to leave the vessel and the other crewmembers. However, to compromise oneself and be open about ones
own mistakes is not always an easy task, especially not
in an unknown working environment. Something that
may reinforce the crewmembers fear of reporting their
own mistakes is the ongoing practice that each crew
member is evaluated by their senior officer/captain,
and based on this report get recommended or not
recommended for re-hire. Interviews have revealed
that this evaluation practice differs. Some practise
an open evaluation where all parts are involved, with
focus on how to improve the evaluated crews shortcomings, and where the shore organisation seeks to
ensure that the evaluation is conducted in as objective a way as possible. At other vessels, the evaluation
is closed for insight by the evaluated and may also
be highly subjective. Some of the respondents have
expressed that by reporting, their next contract may be
at stake, or they may meet with other negative consequences. So, lack of stability and constantly changing
working groups may sacrifice a trusting and open environment, and thus also the sailors commitment to
safety.
A crew committed to safety is essential, but not
enough. Lack of competence may cause a situation
where the crew do not identify potential dangerous situations, or create them. Competence, which accounts

1428

for 7.12% of the total variance, is in this setting comprised of activities performed on board the vessel, and
is all under the control of the captain, training, drills
and familiarisation when signing on. Also, the competence dimension does correlate strongly with the
interaction dimension with a correlation coefficient
at .639. This indicates that a situation when the sailors
are feeling confident with the nature of their task also
results in a better interaction climate where conflicts
are more likely to be absent. As with the interaction
climate, competence will also be affected by the crew
stability. A crew member that is constantly signing on
new vessels and that has to interact with new crew
members and leaders, uses more effort adapting to
the new situation, working out how things are done
at that specific vessel, the informal structure onboard
and so on. When more stability is provided, more effort
may be placed on upgrading their competence, and the
competence will be kept within the vessel and organisation. Both the training activities and crewing strategy
may be controlled by the ship management, and thus
these safety culture dimensions are also, to a certain
degree, controllable.
The dimension of work situation consists of proactive activities as Safe Job Analysis (SJA), safety
evaluations and the possibility they have to prioritize
safety in their daily work. So how may the organisation affect this? For one, they may supply sufficient
crew. Today many vessels are sailing with a smaller
crew at the same time as new standards and resolutions
like the ISM-code increase the amount of paperwork
to be done. Both own observations and interviews
reveal that inter alia check lists and SJA are done in a
mechanical manner. This may originate from various
reasons such as an overload of work, no understanding
of the importance of those activities, lack of feedback
or improper planning by the local management.
The local management dimension, accounts for
5.96% of the explained variance, and the direct effect
of local management is relatively small. However,
local management is considered to have an indirect
effect on the safety climate through the managers,
or senior officers, affect on the interaction climate,
competence and training activities, reporting practices
and the work situation. Again we wish to focus on
the importance of stability within the work group.
Most captains have a sailing period of 3 months or
less, while most of the non Norwegian ratings have
a sailing period of 9 months. Most senior officers
also have a shorter sailing period then an ordinary
rating. Then a rating possibly has to deal with several different leaders during his stay. And each captain
and department managers leadership style may vary,
and are sometimes even destructive, as shown by following comment from a Pilipino engineer. The only
problem on board is the treatment of senior officers to
the lowest rank. (. . . ) There are some senior officers

who are always very insulting towards jr. officers and

rating. Schein (2004) regards the leader as an important key figure in the cultural development. At sea
the captain holds a key role. The captain is the one
in command at every vessel, and according to Schein
(2004) the captains orientation will affect the working climate, which precedes the existence of a culture.
So, in a situation where lack of crew stability impedes
the development of a safety culture, the role of the
captain is even more vital. Also, it is important to
take into account that the leadership style that is practised on board not only affect the sphere of work, but
also time off. However the Captains themselves may
not be aware of their own importance, or how they
affect safety. Most Captains, or other department leaders for that matter, do not have managerial training
or education. When adopting a cultural view towards
safety, as in this research, in as opposed to a behaviour
based view, more emphasis is placed on organisational
factors. Decisions regarding crewing strategy, employment terms and managerial development programmes
are all strategic decisions made on shore. With reference to Scheins culture definition, we will argue
that the safety culture originates within the organisation on shore. Based on Scheins definitions of
culture there ought to exist a pattern of shared basic
assumptions that may solve the problems the shipping
industry is facing. Our case however has revealed an
offshore practise characterised by extended use of contract employment, lack of stable working conditions
on board the vessels, and little or no use of managerial
training and development. This practice does not promote a good safety culture and is considered to has a
negative effect on the overall safety level.
6

CONCLUSIONS

The aim of this paper was to analyse the characteristics of the safety culture on Norwegian tankers and
bulk carriers, and identify what organisational factors
may affect the safety culture on board vessels. Statistical analysis identified five safety related dimensions
on board the vessels: interaction climate, reporting
practices, competence, local management and work
situation. Within shipping the interaction climate is
characterised by unstable working conditions. Under
such conditions it is difficult to achieve and maintain
a stable crew, and proper management becomes even
more important. Also the Captain has a vital role, as
he has the possibility to directly affect all the other
safety related aspects through his own leadership style.
The Captains, officers and ratings normally have different employment terms and shift terms. This may
jeopardise the development of a sound safety culture
as the crew has a poor possibility of developing common behaviour practices and a mutual understanding

1429

of how to do things right. As neither the Captains

nor the officers normally have any managerial training, their leadership styles often affect the safety in
a negative direction. The on board situation is to a
large extend considered to be created by the on-shore
crewing strategy and management policy.
In order to develop a sound safety culture on-board,
the shipping companies should go in new directions
and pursue a crewing strategy which offers more
favourable employment terms and fixed shifts for all
nationalities, and strive for a more stable workforce.
Another measure would be to accept the Captains
and department managers roles as leaders, and offer
managerial development. A final measure will be
to develop a policy and system that ensure proper
onboard management.
REFERENCES
Bakka, D. & Sjfartsdirektoratet 2004, Hundre r for sikkerhet til sjs: Sjfartsdirektoratet 19032003, Direktoratet,
Oslo.
Cooper, M.D. 2000, Towards a model of safety culture,
Safety Science, vol. 36, no. 2, pp. 111136.
DeJoy, D.M. 2005, Behavior change versus culture change:
Divergent approaches to managing workplace safety,
Safety Science, vol. 43, no. 2, pp. 105129.
Dekker, S. & Dekker, S. 2006, The Field guide to understanding human error, Ashgate, Aldershot.
Department of Transport 1987, mv Herald of Free Enterprise.
Geller, E.S. 2001, The psychology of safety handbook, Lewis
Publishers, Boca Raton, FL.
Goffman, E. 1968, Asylums: essays on the social situation of mental patients and other inmates, Penguin,
Harmondsworth.
Guldenmund, F.W. 2000, The nature of safety culture: a
review of theory and research, Safety Science, vol. 34,
no. 13, pp. 215257.
Hair, J.F. 1998, Multivariate data analysis, 5th edn, Prentice
Hall, Upper Saddle River, N.J.
Hinkin, T.R. 1995, A review of scale development practices
in the study of organizations, Journal of Management,
vol. 21, no. 5, pp. 967988.
Justis- og politidepartementet 1991, Scandinavian Star
ulykken, 7 april 1990, Justis- og politidepartementet,
Oslo.

Le Meur, C. 2003, Maritime Safety Culture.

Li, K.X. & Shiping, Z. 2002, Maritime professional safety:
prevention and legislation on personal injuries on board
ships.
Nrings- og handelsdepartementet 2007, St kurs: regjeringens strategi for miljvennlig vekst i de maritime
nringer, Nrings- og handelsdepartementet, Oslo.
National Transportation Safety Board 1990, Grounding of
U.S. Tankship EXXON VALDEZ on Bligh Reef, Prince
William Sound Near Valdez, AK March 24, 1989.
Neal, A., Griffin, M.A. & Hart, P.M. 2000, The impact
of organizational climate on safety climate and individual
behavior, Safety Science, vol. 34, no. 13, pp. 99109.
Netemeyer, R.G., Sharma, S. & Bearden, W.O. 2003, Scaling
procedures: issues and applications, Sage Publications,
Thousand Oaks, Calif.
Norwegian Shipowners Association, www.rederi.no.
Perrow, C. 1999, Normal accidents: living with high-risk
technologies, Princeton University Press, Princeton, N.J.
Peterson, M.F., Ashkanasy, N.M. & Wilderom, C.P.M. 2000,
Handbook of organizational culture & climate in Sage,
Thousand Oaks, Calif., pp. 193.
Pett, M.A., Lackey, N.R. & Sullivan, J.J. 2003, Making sense
of factor analysis: the use of factor analysis for instrument
development in health care research, Sage, Thousand
Oaks, Calif.
Rasmussen, J. 1997, Risk management in a dynamic society: a modelling problem, Safety Science, vol. 27, no.
23, pp. 183213.
Reason, J. 2001, Managing the risks of organizational
accidents, Ashgate, Aldershot.
Reason, J. & Hobbs, A. 2003, Managing maintenance error:
a practical guide, Ashgate, Aldershot.
Schein, E.H. 2004, Organizational culture and leadership,
3rd edn, Jossey-Bass, San Francisco.
Sorensen, J.N. 2002, Safety culture: a survey of the stateof-the-art, Reliability Engineering & System Safety,
vol. 76, no. 2, pp. 189204.
Studio Apertura 2004, Sikkerhetskulturer i transportsektoren. Metoder for kartlegging av sikkerhetskultur:
Evaluering av noen eksisterende verkty.
Tinmannsvik, R.K. 2008 Robust arbeidspraksis: hvorfor
skjer det ikke flere ulykker p sokkelen? Tapir akademisk
forl., Trondheim.
Yule, S.J. 2003, Senior managers transformational leadership behaviours for safety. Australian Journal of
Psychology, vol. 55, no. Supplement.

1430

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Quantitative analysis of the anatomy and effectiveness of occupational

safety culture
P. Trucco, M. De Ambroggi & O. Grande
Department of Management, Economics and Industrial Engineering, Politecnico di Milano, Italy

ABSTRACT: Despite increasing technological advancement and the strict regulation on occupational safety,
the number of accidents and injuries at work is still high. According to many studies, the best way to reduce these
occurrences is to promote a better safety culture. To this end many models have been proposed to identify the
complex nature and structure of safety culture: some of them take into consideration only professional aspects,
others take into consideration socio-cultural aspects. The paper proposes a new model, based on Bayesian
Belief Networks, for representing, analysing and improving safety culture. Thanks to the direct involvement
of employees, the model is able to provide a quantitative representation of the real anatomy of the companys
safety culture and its impact on the expected probability of safe behaviours performed by workers. To test and
prove the validity of the proposed approach a case study in a leading agricultural tractor manufacturer has been
carried out.
1

INTRODUCTION

Despite increasing technological advancement and the

strict regulation on occupational safety, the number of
accidents and injuries at work is still high. In order to
the overall reduction of workplace accidents, workplace safety has been studied from different points
of view (Silva et al. 2004), but only in the last
decade the promotion of a positive safety culture has
been largely regarded as the most important prerequisite for continuous improvement. To this end many
models have been proposed which try to identify
the complex nature and structure of safety culture.
Some of them asserts that accidents stem almost
entirely from the system (Deming 1986) and model
only professional aspects, others suggest that nearly
every accident may be traced to an employees unsafe
act (Burk et al. 1990) and take into consideration
socio-cultural aspects. Nevertheless, most industrial
accidents are caused by an interacting system of social
and technical forces (Brown et al. 2000). Therefore,
a companys safety culture is based on both technical
and social, systemic and personal aspects; hence, some
socio-technical models have been developed. Most of
these models theorize the idea that the characteristics of a single worker directly affect safety, while the
characteristics of the system affect safety indirectly,
through the behaviour of workers. The pre-eminent
socio-technical model was developed by Brown et al.
(2000): the study aimed only at analysing the relevant
factors of safety culture according to a pre-defined

model, verifying empirically the correlations among

these factors.
However, an effective exploitation of the beneficial effects of safety culture on safety improvement
requires a detailed representation of the actual anatomy
of safety culture within a company, and eventually
compare it against a more efficient paradigm able to
positively affect the safe behaviour of workers.
The paper proposes a new approach, based on
Bayesian Belief Networks (BBNs), for representing,
analyzing and improving the actual anatomy of companys safety culture and its impact on the expected
probability of safe behaviours performed by workers.
The validity of BBNs in supporting the modeling of
safety culture and the evaluation of potential strategies
for safety improvement has also been demonsttrate by
Zhou et al. (2007).
By means of extensive interviews of managers and
workers, based on properly designed questionnaires
and elicitation processes, a large and representative
set of data on socio-technical factors has been gathered. Then, by means of a specific algorithm, K2
(Cooper et al. 1992), the Bayesian structure underlying
the relationships among socio-technical factors was
returned. The resulting BBN is a model of the companys safety culture that can be exploited to identify
and analyse the effectiveness of different organizational and behaviour-based measures for improving
occupational safety. An application of the proposed
approach is presented referring to a large agricultural
tractor manufacturing company.

1431

SAFETY CULTURE AND SAFETY CLIMATE

Notwithstanding the key role played by safety culture and safety climate in preventing or reducing the
incident rate of an organization, there is no apparent consensus on how to define the terms safety
culture and safety climate and how to describe
them. Guldenmund (2000) and Choudhry et al. (2007)
can be considered as reference for a the state-of-theart of definitions, theories, researches and models
related to safety culture. Guldenmund (2000) reviewed
the literature on safety culture and safety climate
in terms of definitions, sources, causal models and
goals from Zohar (1980) to Cabrera et al. (1997)
and Williamson et al. (1997). Choudhry et al. (2007)
focused their literature review on safety culture definitions and researches undertaken from 1998 onwards.
Furthermore, Guldenmund (2000) gives an overview
of the amount of questions, surveyed population and
dimensions of safety culture and climate researches.
Even though many models take into account only
technical aspects or only socio-cultural aspects, we
think that the socio-technical models are more suitable for description of the culture of real organizations.
Indeed, the paper describes a model for assessing the
anatomy and the effectiveness of safety culture considering both social and technical factors. The next
section briefly introduces the socio-technical model
developed by Brown et al. (2000) that has been used as
theoretical reference for our analyses on safety culture
mapping.

2.1

Cavalier attitudes toward safety riskthe extent to

which an employee feels that he or she can ignore
safety procedures without incurring the risk of an
accident or injury.
Safety-efficacyemployees confidence that he or
she has the skill to work safely in the context of a
specific environment.
Safe work behavioursthey choose this as the
criterion variable for a predictive model.
Based on literature and interviews in manufacturing
plants Brown et al. (2000) initially hypothesized the
relationships among the selected constructs (Figure 1).
Using data from a survey on 551 employees working in
the steel industry, they then examined and tested the
structural model using covariance structure analysis
procedures in Bentlers EQS program (1995). Figure 2
shows the resulted indirect effects model based on their
hypothesized system-person sequence.
Results provided evidence that a chain reaction
of technical and social constructs operates through
employees to influence safe behaviours. Therefore, the
model underlines the fact that the characteristics of the
system and the human factors are both affecting safety,
but system effects show themselves indirectly through
people.

Browns model

The purpose of the authors was to define more clearly

the antecedents to safe and unsafe work behaviours in
manufacturing setting (Brown et al. 2000). According to the evidence coming from personal intervews
with plant managers and safety professionals in 10
manufacturing companies across several industries,
they choose to investigate six variables. They defined
these variables as follows.

Figure 1. Preliminary socio-technical model predicting safe

work behaviour.
Source: Brown et al. 2000.

Safety hazardstangible factors in the work environment (e.g. heavy items that must be carried,
repetitive motion, moving objects) that may pose
risks for possible injuries or ailments.
Safety climatethe more appropriate description of
the construct that captures employees perceptions
of the role of safety within an organization.
Pressure for expediency over safetyan employees
perception that an organization encourages him or
her to work around safety procedures in order to
meet production quotas, keep up with the flow
of incoming work, meet important deadlines, or
continue getting paychecks.

Figure 2. Resulted indirect effects model. All structural

parameters estimates were significant (p < 0.05) and in predicted direction.
Source: Brown et al. 2000.

1432

Some shortcomings of Browns approach can be

pointed out. The analysis aims only at empirically verifying postulated relevant factors of safety culture, and
the strength of the links among the factors is based
only on correlations. The recognized validity of the
socio-technical model aims further than Browns analysis, trying to find a quantitative method to be used
as an instrument for safety culture management and
for safety actions planning. The authors of the present
paper are convinced that research on safety culture
cannot be based only on postulated paradigms and patterns of factors, but should be based also on methods
for the observation of the factors within companies
in order to find their actual safety culture, expressed
in terms of mutual influences among relevant shaping
factors.
The paper aims at using Bayesian Belief Network to
model and analyse the anatomy of the real safety culture of an organization and to provide a methodology
to identify potential strategies for safety improvements. While Zhou et al. (2007) has already proved
the possibility of using BBN to describe and analyse
a postulated model of safety culture, any research has
been carried out on the analysis of the safety culture
anatomy and effectiveness starting from real data.

METHODOLOGY

The paper proposes a three-steps methodology.

Step 1. Data gathering to select and describe the
variables playing a role in the safety culture under
study. The design, validation and administration of
an ad hoc questionnaire is proposed to this end.
Step 2. Network inference. Starting from gathered
data, by means of Bayesian software (e.g. Bayesware
Discover, Genie, Hugin) it is possible to infer the
network of influences among selected factors and
compute all related conditional probabilities.
Step 3. Reasoning process. With a Bayesian Network, four types of reasoning can be performed
(Zhou et al. 2007): diagnostic, predictive, inter-causal,
and combined reasoning. This BN distinctive feature
allows the use of the resulted network describing the
organization safety culture as a real managing tool.
The next section presents the application of this
methodology to a case study carried out in a leading
agricultural tractor manufacturing company.

CASE STUDY

To test and prove the functionality of the developed approach a case study in SAME DEUTZ-FAHR
Italia has been carried out. The group is one of the

worlds main manufacturers of tractors, combines,

farm machinery and engines. SAME DEUTZ-FAHRs
manufacturing is distributed over three plant in Europe
(Italy, Germany and Croatia), and one in India.
Another factory is currently being constructed in
China. The case study was focused on the Italian plant.
Indeed, an in-depth investigation of the industrial accidents occurred from 1999 to 2004 has revealed a lack
of safe work behaviours.

4.1

Variables

We chose eleven variables to analyse the companys

safety culture: the six variables of the Browns
model and five variables already studied by different authors and came out as important from accident
investigation and preliminary interviews with workers
and managers. The six variables of the Brown model
have been already presented in the section 2.1. The
remaining five variables with their definitions are:
Communicationeffective communication between managers and the workforce regarding health
and safety issues (Clarke 2006).
Management commitmentmanagement promotion of a positive safety culture (e.g. participation
in risk assessments and consultative committee
meetings, allocation of resources).
Priority of safety over other issuesperception of
the employees of the safety versus demands to reach
production goals (Rundmo et al. 1998).
Safety efficiencyinvolvement in safety-related
decisions perceived by employees.
Employee fidelitythis factor should influence the
employee tendency to notify minor incidents and /or
injuries.
According to the socio-technical model, the
selected variables can be grouped into three macrocategories as shown in Table 1.

4.2

Safety culture questionnaire

To gather the data needed to investigate the selected

variables a questionnaire composed of three different
sections was developed. The first section investigated personal characteristics of interviewees. The
second part included 31 questions related to 10 social
and technical factors (see Table 2). It should be pointed
out that each factor is related to several questions. The
interviewees were asked to respond to the questions
using a three-point Likert scale.
The third part included 4 questions related to safe
work behaviours. In this section, the interviewees were
asked to respond questions estimating the percentage
of time spent in safe related behaviours.

1433

Table 1.

Table 3. Internal consistency of the questionnaire

(Cronbachs alpha).

Variables of the proposed model.

Macro category

Note

Variable

Safe work
behaviours

The criterion
variable for a
predictive model
Employee
perception of the
technical system
factors

Safe work
behaviours

System-related
factors

Human Factors

Employee
attitudes and
behaviours

Safety hazards
Safety climate
Pressure
Management
commitment
Priority of safety
over other issues
Safety efficacy
Cavalier attitude
Communication
Employee
involvement
Employee fidelity

Variables

Alfa-Cronbach

Safety hazard
Safety climate
Pressure
Management commitment
Priority of safety over other
issues
Safety efficacy
Cavalier attitude
Communication
Employee involvement
Employee fidelity
Safe Behaviour

0.75
0.77
0.72
0.78
0.83
0.81
0.88
0.86
0.75
0.71
0.83

Table 2. Examples of questions (extract from the proposed

questionnaire).
Personal data
Age:
Years of employment at SDFG:
Department/Tech. Unit:
...

...
...
...

SocioTechnical factors
Hazard
Do you lift too heavy loads?
Do you consider dangerous the
transit of trucks/trolleys?
Are the tools/machine tools you
currently use too dangerous?
Is the work environment safe?
...
Cavalier attitude
Do you think safety rules
are useful?
Would your job be safe also
without procedures?
...

(1 = Never, 2 =
Sometime, 3 = At all)
2

Figure 3.

Bayesian network of company safety culture.

1
2
2
(1 = Never, 2 =
Sometime, 3 = At all)
2
1

The questionnaire was designed to ensure that all

participants understood the questions and was distributed to 130 employees randomly extracted from
the target population of 815 employees. Referring
to the sampling theory explained by Tryfos (1996)

the selected sample can be considered statistical significant. The response rate was about 95% (123
responses).
Since the factors are relevant to several questions in
the questionnaire, values are calculated as the average
scores of all relevant questions with the same weight.
For example the value of the variable Hazard, derived
from the responses in Table 2, has been calculated as
the average scores of the 4 questions, that is Hazard =
(2 + 1 + 2 + 2)/4 = 1.75. Due to this, the value of
the 10 social and technical factors lies in the interval
between 1 to 3, whereas the value representing Safe
Work Behaviour lies in the interval between 0 to 100
(percentage of working time).
The internal consistency of the questionnaire pertinent to the eleven factors is measured by Cronbachs
alpha (Table 3), related to scale reliability that was
defined as the extent to which a measure produces similar results over different occasions of data collection
(McMillan et al. 1989).

1434

Figure 4a. The anatomy of safety culture in the manufacturing area.

to set the states of the variables starting from the results

of the questionnaire. Parents variables were defined
referring to the percentage of compliance, e.g. for the
Safe Behaviour variable the intervals [70.090.0;
90.095.0; 95.0100.0] were defined.
The best Bayesian network, emerged from the analysis of the data related to the whole company, is shown
in Figure 3. The results shown that not all the 11 factors play a role in the company safety culture, e.g. the
variable Employee involvement was not linked to any
other factors. Safe work behaviour is directly affected
only by the cavalier attitude. The lack of directly connected positive factors (communication and employee
fidelity) was initially supposed due to the width and
heterogeneity of the field of study. Therefore, we
decided to analyse the safety culture of more homogeneous areas, both referring to the nature of the tasks
performed and to the type and number of accidents.
This was quite a radical change of Browns view,
as the different safety cultures of the company have
been elaborated empirically, instead of imposing an
established unique model.
Two more homogeneous areas were separately considered: manufacturing unit and assembling unit.
Analysing the data from the manufacturing area (31%

Figure 4b. Value of critical shaping factors of safety culture

in manufacturing area.

4.3

Network development

The collected data were processed through the software Bayesware Discoverer, that implements the
algorithm K2 (Cooper et al. 1992), to find out the
Bayesian structure underlying the relationships among
socio-technical factors and also to estimate conditional probabilities. The arcs in the BBN were directed
considering the experts opinion in cause-effect relationships (it is apparent, for example, that the age
of the worker affects the safety climate and not the
reverse) and the Browns model (Brown et al. 2000).
The data pre-processing (using algorithms as K2,
NPC, PC) allowed the reduction of CPTs complexity:
during the pre-processing it is convenient to have the
discretisation of continuous intervals aiming at reducing the number of possible states. In the applied case
the learning algorithm and data processing were used

Figure 5a.
area.

The anatomy of safety culture in the assembling

Figure 5b. Value of critical shaping factors of safety culture

in assembling area.

1435

of the samples), the resulting anatomy of the safety

culture is described by the Bayesian network reported
in Figure 4a. It is apparent that, in accordance with the
socio-technical model developed by Brown (2000), on
top of safety culture there are only systemic factors,
specifically safety hazards, safety climate and management commitment. Personal factorsi.e. safety
efficacy, employee fidelity and cavalier attitudeplay
the role of mediators between systemic factors and
workers safe behaviours.
The probability distribution of each root node
(safety hazard, safety climate and management commitment) and of the safe work behaviour is shown in
Figure 4b.
Considering the current setting of systemic factors
(e.g. 27.2% probability of having poor safety climate,
30.8%, for good safety climate and 42% for optimal
safety climate), the rate of safe work behaviours was
estimated about 93.6%. Even though this value may
seem high (6 unsafe behaviours out of 100), the high
value of gravity index of incidents occurred at workers
operating in manufacturing area suggests the need for
an improvement of compliance with safe behaviours.
The same analysis was then performed in the assembly area (69% of the samples); Figures 5a and 5b show
the estimated anatomy of its safety culture. As already
shown for the manufacturing area, the systemic factors are largely present on top of the network. The
main characteristic of the safety culture in this area
is the presence of three personal factorscommunication, employee fidelity and cavalier attitudein an

intermediate position, between systemic factors and

workers safe behaviours. In this area the rate of safe
behaviours was estimated about 73.8%, lower than in
the manufacturing area, and this is mainly due to the
different structure of the two safety cultures.
4.4

Discussion of the results

Based on Bayesian theory (Korb et al. 2004) belief

on both relationships and events can be updated when
new evidences or cases are available. Theoretically,
the more evidence a Bayesian network has, the more
accurate the beliefs achieved for the network. When
the states of one or a few nodes in the Bayesian network are given, the beliefs pertinent to all other nodes
are then assessed through the probabilistic relationship
of the networkthe so-called reasoning. In other
words, the effectiveness of different safety management strategies for controlling the shaping factors of
companys safety culture can be evaluated in terms of
outcomes at the safe behaviour node. Specifically,
a variation in the probabilities of state of the safe
behaviour variable, due to changes in the states of
the root nodes, can be estimated.
Tables 4 and 5 show the effectiveness of different
strategies for improving safe work behaviours related
to the safety culture in the manufacturing and assembly areas respectively. To better understand how the
values are calculated we explain in detail the analysis of the scenario in which we decided to perform
strategies to limit safety hazards in the manufacturing

Table 4. Summary of the effectiveness of strategies to improve safe work behaviours in manufacturing area. All values
are percentage. Negative values means decreases.

Joint strategy

Simple strategy Simple strategy

(10% factor (best scenario)
improvement)

Strategy

Safety
hazards
(%)

Safety
climate
(%)

Management
commitment
(%)

Safe work
behaviour
(%)

Unsafe work
behaviour
(%)

0,8

11, 9

0,0

0, 2

0,2

3, 1

0,3

4, 4

0,0

0, 1

0,2

2, 3

1,5

21, 3

13,7

1,0

14, 7

16,3

13,7

0,3

3, 9

16,3

13,7

1,8

25, 9

27, 2
16,3
13,7
10
10
10
27, 2

16,3

27, 2

27,2

1436

Table 5. Summary of the effectiveness of strategies to improve safe work behaviours in assembling area. All values are
percentage. Negative values means decreases.

Strategy
Simple
strategy (best scenario)
Simple
strategy (10% factor
improvement)
Joint strategy

Priority
of safety
(%)

Safe
workplace
behaviour
(%)

Unsafe
workplace
behaviour
(%)

38,9

1,5
6
0,6

4,2
16,8
1,9

10,0
38,9

1,5
7,3

4,3
20,6

Pressure
(%)

22,6
10
27,2

Figure 6. Effectiveness of strategies to limit hazard at

minimum level in the manufacturing area.

area at the minimum level (Fig. 6). In this scenario the

average value of safety hazards changed from 2.06 to
1.5. The safe work behaviours increased to 94.36%,
meaning that the employees have unsafe behaviours
for 5.64 hours over 100 working hours, corresponding
to a decrease of 11.9% of unsafe work time.
Considering either simple and joint strategies
a comparison of the increased percentage of safe
behaviour would not be enough to make an evaluation
of the best strategy to implement. Indeed, the economic
effort required by any action or improvement (Level
of Effort) must also be considered. A data comparison made by using a theoretic curve of risk reduction
may be useful to discover the best option. We assumed
an exponential relation between the level of effort and
the corresponding improvement of a specific factor in
safety culture (Cagno et al. 2000).
Considering as target in manufacturing area the
improvement of safe work behaviours related to a strategy that aims at improving management commitment,
the required decrease of safety hazards to gain the
same improvement has been calculated. Then, the two
correspondent levels of effort has been compared to

Figure 7. Comparison of levels of effort on different

factors able to decrease unsafe work behaviours of 3.1% in
the manufacturing area.

suggest the best option (Fig. 7). To achieve an improvement of 3.1% on the rate of safe behaviours it is more
efficient to mitigate safety hazards (8.3% of reduction) than promote management commitment (13.7%
of improvement). Therefore, the analysis suggests to
focus on the elimination or minimisation of hazards.
The same approach has been applied in the assembling area where a reduction of 4.2% of unsafe
behaviours is explored thanks to a higher priority of
safety (Fig. 8). In this case the lowest level of effort
required to achieve the same target, i.e. the most efficient action, is guaranteed by reducing the pressure for
compliance with production scheduling over safety.
The proposed analysis to improve safe behaviours
is regarded as suitable to suggest how and where focus
actions for improvement only on the short-medium
term. Indeed, the effectiveness of these strategies on
middle-long term is not demonstrated. We think that
in the middle-long term it is more effective to concentrate on improving the anatomy of safety culture, by
removing incoherent relationships and promoting the
inclusion of positive shaping factors.

1437

Figure 8. Comparison between levels of effort related to a

decrease of 4.2% of unsafe work behaviours in the assembling
area.

CONCLUSIONS

The paper proposes a method based on Bayesian Belief

Networks (BBNs) able to assess the anatomy of the real
safety culture of an organisation, eventually to be compared against theoretical models. Furthermore, the
proposed approach allows to identify alternative strategies to promote safe work behaviours and to evaluate
their relative effectiveness.
The case study carried out in SAME DEUTZFAHR Italia has also shown the possibility of a
practical application of socio-technical models for
supporting the safety improvement process in large
manufacturing companies. The case study also confirmed that personal factors play a critical role in
mediating the relationships between systemic factors
(technology and organisation related) and safe work
behaviours.
REFERENCES
Bentler, P.M. 1995 EQS Structural Equation Program Manual. Encina, CA: Multivariate Software.
Brown, K.A., Willis, A.G. & Prussia G.E. 2000. Predicting
safe employee behaviour in the steel industry: Development and test of sociotechnical model. Journal of
Operations Management 18(4): 445465.
Burk, A.F. & Smith, W.L. 1990. Process safety management
within Du Pont. Plant/Operations Progress 9(4).

Cabrera, D.D., Isla, R. & Vilela, L.D. 1997. An evaluation

of safety climate in ground handling activities. Aviation
safety, Proceedings of the IASC-97 International Aviation
Safety Conference. Netherlands: Soekkha.
Cagno, E., Di Giulio, A. & Trucco, P. 2000. Economic
Evaluation Model of Safety Management. Ergonomics
and Safety for Global Business Quality and Productivity, Koradecka, D., Karwowski, W., Das, B. (eds), CIOP.
Varsavia, Polonia.
Choudhry, R.M., Fang, D. & Mohamed, S. 2007. The nature
of the safety culture: A survey of the state-of-the-art.
Safety Science 45: 9931012.
Clarke, S. 2006. Safety climate in an automobile manufacturing plant: The effects of work environment, job
communication and safety attitudes on accidents and
unsafe behaviour. Personnel review 35(4).
Cooper, G.F. & Herskovitz, E.A. 1992. Bayesian method
for the introduction of probabilistic networks from data.
Machine learning 9: 309347.
Deming, W.E. 2000. Out of the Crisis. MIT Center for
Advanced Engineering Study. Cambridge, MA.
Guldenmund, F.W. 2000. The nature of the safety culture:
a review of theory and research. Safety Science 34(8):
215257.
Korb, K.B. & Nicholson, A.E. 2004. Bayesian Artificial
Intelligence. FL, USA: Chapman & Hall/CRC Press.
McMillan, J.H. & Schumacher, S. 1989. Research in Education: A Conceptual Introduction, second ed. New York:
HarperCollins.
Rundmo, T., Hestad, H. & Ulleberg, P. 1998. Organisational
factors, safety attitudes and workload among offshore oil
personnel. Safety Science 29: 7587.
Silva, S., Lima, M.L. & Baptista, C. 2004. OSCI: an organizational and safety climate inventory. Safey Science 42:
205220.
Tryfos, P. 1996. Sampling for Applied research: Text and
Cases. John Wiley & Sons.
Williamson, A.M., Feyer, A.M., Cairns, D. & Biancotti, D.
1997. The development of a measure of safety climate:
the role of safety perceptions and attitudes. Safety Science
25: 1527.
Zhou, Q., Fang, D. & Wang, X. 2007. A method to identify
strategies for the improvement of human safety behaviour
by considering safety climate and personal experience.
Safety Science. doi:10.1016/j.ssci.2007.10.005.
Zohar, D. 1980. Safety climate in industrial organizations:
theoretical and applied implications. Journal of Applied
Psycology 65(1): 96102.

1438

Safety, Reliability and Risk Analysis: Theory, Methods and Applications Martorell et al. (eds)
2009 Taylor & Francis Group, London, ISBN 978-0-415-48513-5

Safety management and safety culture assessment in Germany

H.P. Berg
Bundesamt fr Strahlenschutz (Bf S), Salzgitter, Germany

ABSTRACT: Experience with accidents in different branches of industry, also in railways companies and the
nuclear industry, has shown the importance of safe operation management. Recent events in German nuclear
power plants such as a transformer fire underlined the necessity of an appropriate and regularly reviewed safety
management system. Such a system has two major aims, firstly to improve the safety performance through
planning, control and supervision of safety related activities through all operational phases and secondly to
foster and support a strong safety culture through the development and reinforcement of good safety attitudes
and behaviour in individuals and teams. Methods of evaluating safety culture can be divided into two main
groups: the approach mainly based on interviews with plant management and staff to get direct information
about the plant safety culture and an approach that uses information derived from plant operation, incidents and
accidents. The current status of the still ongoing discussion in Germany is presented.

INTRODUCTION

The beginning of the safety culture period of accident

investigation and analysis can be traced back to the
nuclear accident at Chernobyl in 1986 in which a poor
safety culture was identified as a factor contributing to the accident by both the International Atomic
Energy Agency and the OECD Nuclear Agency. Since
then, safety culture has been discussed in other major
accident inquiries and analyses of system failures,
such as the Kings Cross Underground fire in London
and the Piper Alpha oil platform explosion in the North
Sea [1], as well as the crash of Continental Express
Flight 2574 [2], the Columbia Space Shuttle accident
[3], and the explosion at the British Petroleum refinery
in Texas City [4].
In general, safety culture has been found to be
important across a wide variety of organizations and
industries. For example, safety culture has been
associated with employees safety-related behavior in
industries such as manufacturing [5], shipping [6],
chemical processing [7], and building maintenance
[8]. Safety culture also appears to predict on-the
job injury and accident rates in manufacturing firms
[910], offshore oil and gas companies [11], and
also in broad cross-organizational studies of workers
in general [12]. While initial studies of safety culture took place in jobs that have traditionally been
considered high-risk, organizations in other areas are
increasingly exploring how safety culture is expressed
in their fields. Overwhelmingly, the evidence suggests that while safety culture may not be the only
determinant of safety in organizations [13], it plays

a substantial role in encouraging people to behave

safely. Accordingly, safety professionals working in
traffic safety have also begun contemplating the role
that safety culture might play in mitigating risk within
their settings [14].
In the past, the German nuclear regulations did
not contain any explicit requirements for comprehensive safety management nor for structuring the utilitys
organisation under the aspect of nuclear safety that
would allow the comprehensive planning and review
of a safety management system. However, a number
of sometimes rather general individual requirements
to be fulfi