EURO J Transp Logist (2019) 8:435–465

https://doi.org/10.1007/s13676-018-0117-z

RESEARCH PAPER

Risk-based optimal scheduling of maintenance activities

in a railway network

Alice Consilvio1 • Angela Di Febbraro1 • Rossella Meo1 •

Nicola Sacco1

Received: 15 July 2017 / Accepted: 30 January 2018 / Published online: 20 February 2018
 Springer-Verlag GmbH Germany, part of Springer Nature and EURO - The Association of European
Operational Research Societies 2018

Abstract In a railway system, maintenance activities need to be continuously

performed to ensure safety and continued rail operations. In this framework, while
on one hand unplanned corrective maintenance activities performed when a fault is
occurred are expensive and would cause low service quality, on the other hand
preventive maintenance that does not consider the actual asset condition is often
unnecessary and turns out to generate avoidable costs. To deal with this issue, in this
paper, a risk-based decision support system to schedule the predictive maintenance
activities is proposed. In such a framework, the interventions are planned by taking
into account the forecast degradation state of railway assets and performed when a
given threshold is reached, thus minimizing the probability of both sudden and
unnecessary operations. With the end of finding the optimal scheduling of predictive
maintenance, in this paper also the space-distributed aspect of railway infrastructure
is considered, defining the best path and the activities assignment for each main-
tenance team. The scheduling model is formulated as a Mixed Integer Linear
Programming (MILP) problem aiming at based on the risk minimization, according
to the ISO 55000 guidelines. A matheuristic solution approach is proposed and
applied to a real rail network. The relevant results show how the proposed
scheduling model can use the outputs of predictive tools and degradation models,
based on data from field, to mitigate the sudden failure risk by means of a cost-
effective maintenance plan at a network level.

Keywords Railway predictive maintenance  Railway assets degradation

modelling  Railway maintenance scheduling

& Alice Consilvio

[email protected]
1
Department of Mechanical, Energy, Management and Transportation Engineering, University
of Genoa, Via Montallegro 1, 16145 Genoa, Italy

123
436 A. Consilvio et al.

1 Introduction

Developing a resilient, consistent, cost-effective, high capacity and attractive rail

network is one of the main European target. To cope with this objective, an optimal
asset management and a suitable integration of rail operations and maintenance
actions are needed in order to reduce the infrastructure costs. In this framework, new
smart asset management strategies based on risk assessment and condition-based
approaches can improve rail reliability, availability, maintainability, safety, and in
particular reduce lifecycle costs. In this connection, the planning of suitable main-
tenance activities, in order to avoid faults and trains circulation disruptions as well
as unnecessary interventions, has become a challenging task that has to take into
account railways topological and operational constraints.
As a matter of fact, the railway infrastructure has the characteristic of being
scarcely redundant (with none or very few path alternatives), which implies that any
fault may result into dramatic drops of the system performances. Moreover, the
space-distributed aspect of railway infrastructure should be considered since such a
characteristic generates significant difficulties in the organization of maintenance
activities and in the management of the relevant supporting resources. Another
aspect that makes railway maintenance critical is the time constraint. In fact, the
available time for maintenance activities is strictly limited due to various factors
such as rail traffic, climate, fulfilment of fixed operation sequences, etc. Some of
these requirements result to be soft constraints, that is, whose violation can be
tolerated if no better choices exist, but other ones are hard constraints that can never
be violated.
Therefore, any railway asset needs very carefully planned maintenance activities,
aiming at guaranteeing its best performances at any time. To cope with this
problem, many maintenance approaches developed in the relevant literature have
shifted from corrective maintenance, performed only when a fault occurs, to
preventive maintenance strategies, which can be divided into:
• Planned maintenance, performed on a regular fixed time schedule. It may lead to
a significant reduction of the useful life of components, due to early replacement
and unnecessary maintenance activities;
• Condition-based maintenance, performed only when necessary, based on the
continuously monitored asset conditions. This approach allows a more efficient
utilization of the useful life of infrastructure components, but requires a regular,
sometimes very frequent, costly monitoring of the degradation state of the
railway assets;
• Predictive maintenance, performed only when necessary, based on model
estimations suitably integrated by few measures. In particular, the aim of
predictive railway maintenance is to minimize the probability of the occurrence
of the so-called mission-critical faults during train service, i.e., those failures
that prevent trains from circulating, or may lead to accidents, while keeping
maintenance costs as low as possible.

123
Risk-based optimal scheduling of maintenance activities… 437

The last approach, which defines the maintenance framework considered in this
work, guarantees low fault probabilities and the best cost reduction, because
maintenance is performed exactly when necessary and requires a limited number of
monitoring measures.
In this paper, a risk-based scheduling problem of railway maintenance activities
in a medium-size network is proposed with the aim of providing a decision support
system able to evaluate a schedule of the railway maintenance activities, which
guarantees to intervene before any asset reaches an intolerable fault risk due to a
highly degraded state.
The proposed approach is based on the risk minimization framework, according
to ISO 55000 guidelines (2014), and introduces in railway maintenance scheduling
the concept of risk, which implies that the activity priorities are based on criticality
indexes that take into account both the fault probability and the relevant losses at a
time. To this end, in this paper, the asset degradation is modelled as a stochastic
process, which provides the time-dependent probability that a fault occurs within a
certain interval, given that it is not occurred since the last maintenance activity. In
this framework, the time dependence of the considered stochastic variables is
necessary to represent the degradation process that introduces a time varying
memory in the fault probability.
Moreover, aiming at considering the significant space-distribution characteristic
of railways, the proposed approach determines the optimal path for maintenance
teams along the rail network and the relevant maintenance activities assignment,
optimizing the use of limited resources.
Summing up, the presented study shows how a scheduling model can use the
significant outputs of predictive degradation models with the twofold aim of
planning the risk-based maintenance at a network level and optimizing the resource
utilization.
The paper is organized as follows. In the next section, the existing literature on
maintenance scheduling models is analysed and compared to the proposed
approach. Then, the mathematical formulation of the considered approach is
described. Finally, the results of the model application to a real rail network together
with some discussions and indications about its future developments are provided.

2 Literature review

Preventive maintenance works are performed in order to reduce the probability of a

fault occurrence and to improve the overall reliability and availability of a system.
Poorly designed preventive maintenance schedules may incur high maintenance and
operation costs and loss of safety. Therefore, literature gives great attention to
railway maintenance scheduling problem. Authors agree that is essential to move
from a planned preventive maintenance to a predictive preventive maintenance
(Zoeteman and Esveld 2004). In rail world, automatic monitoring and diagnostic
systems, mounted on trains and along the lines, have become over years more and
more significant and sophisticated and new suitable models are necessary to use

123
438 A. Consilvio et al.

efficiently the new available data, for estimating the time when a fault is likely to
occur adapting maintenance interventions accordingly (Jimenez et al. 2010).
Therefore, new predictive maintenance scheduling models must be able to reduce
both the overall maintenance budget and the safety risk, evaluating the priority of
the maintenance tasks, taking into account not only assets degradation conditions
but also assets criticalities. A fault can be better tolerated by the system than
another, according to the related consequences.
In past literature, the most common optimization criteria for railway maintenance
are based on cost minimization, in order to reduce the overall maintenance budget.
Nowadays, the aim of predictive maintenance scheduling is to minimize time
duration and overall maintenance costs, planning the preventive maintenance
activities and assigning them to the different working teams, but also to maximize
the system reliability and availability, at the same time (Lidén 2015).
Moreover, as mentioned, railway maintenance must take into account the space-
distributed aspect of railway infrastructure. Railway assets are often not spatially
delimited to a point, and this implies difficulties in the organization of maintenance
activities and resources. The Travelling Repairman Problem (TRP) discussed in the
literature (Jothi and Raghavachari 2007) addresses repair scheduling of assets,
distributed in various locations, that have already failed, and the best route for the
repairman is sought to optimize the situation such that the minimum total asset
downtime is achieved.
In addition, the Vehicle Routing Problem (VRP), which aims to minimize the
total routing cost of K vehicles to n destinations, constrained by exactly one visit to
each destination, has been used in literature for maintenance operations. Mainte-
nance scheduling is modelled through VRP in situations in which multiple
maintainers need to repair broken assets in different positions, such as seasonal road
maintenance (Perrier et al. 2007).
On the other hand, the Travelling Maintainer Problem (TMP) aims to find the
most cost-effective routing for the maintenance operator to visit assets that have not
yet failed. Regarding this topic, Camci (2015) presents a model for the maintenance
scheduling of geographically distributed assets, incorporating the travel time
between assets in the scheduling problem, when one team is responsible for the
maintenance of multiple assets, taking into account also prognostic information.
As a general consideration coming from the relevant research, it is worth saying
that preventive maintenance problems in a large-scale railroad network involve
hundreds of activities and very complex relationships among them, which generates
a larger number of side constraints. In addition, such optimization problems have
been proven to be Non-deterministic Polynomial-time Hard (NP-Hard) problems,
that is, a class of problems to solve which, a polynomial-time algorithm has never
been found. Thus, they require a too long time to be solved and, for this reason,
many scheduling heuristic techniques have been used for managing preventive
maintenance activities (Soh et al. 2012), mainly based on:
• Genetic algorithms;
• Ontology-based modelling;
• Strategic gang scheduling;

123
Risk-based optimal scheduling of maintenance activities… 439

• Other specific heuristic approaches.

However, when the dimension of the problem admits it, a general purpose solver
of optimization problems can be sufficient. More in detail, Higgins (1998), Budai
et al. (2004), Peng et al. (2011), Borraz-Sanchez et al. (2011), Macedo et al. (2017),
develop mathematical models aiming at determining the assignment and the
schedules of maintenance teams to minimize the disruption of train operations and
the infrastructure possession time. Such models take into account budget
constraints, train schedules, working teams travel times, and various interrelations
among maintenance activities. The solution approaches consist of heuristic
approaches such as tabu search and simple greedy heuristic. Nevertheless, they
do not take into account data about the infrastructure conditions provided by
monitoring systems.
Some steps forward in this direction are done by Dell’Orco et al. (2008),
Umiliacchi et al. (2011), Jiménez-Redondo et al. (2014), Wen et al. (2016) and Su
et al. (2017). In fact, their models aim at minimizing the probability of a mission-
critical fault occurring during train service, keeping the railway assets at good safety
and comfort levels. They use data from special diagnostic trains and new inspection
technologies to evaluate rail track condition. Condition-based maintenance has been
usually applied to rail track (Quiroga et al. 2011) because its deterioration can be
directly measured and continuously perceived. Instead, deterioration could be
difficult to be measured for other rail assets, such as signalling equipment or
electrical equipment, since their deterioration does not progress continuously and
failures seem to occur suddenly. Recent advances in ICT, especially progress in data
analysis technologies, allow perceiving signs and causes of failures even in kind of
machinery that apparently breaks down suddenly, by continuously measuring
certain physical quantities (current, resistance, etc.) (Patra et al. 2010; Yokoyama
2015; Morant et al. 2016; Ghodrati et al. 2016).
Moreover, important steps from condition-based maintenance to predictive
maintenance have been done by Simson et al. (2000), Shingler et al. (2008), Burrow
et al. (2009), Zhang et al. (2013) and Letot et al. (2016) developing tools based on
predictive degradation models. In these approaches, asset condition data allow to
predict faults and to determine which maintenance work is required, for
guaranteeing safe train operation, and whether safety-related speed restrictions
are necessary.
These models enable long-term forecasts of asset condition and strategic
maintenance decisions, also considering the uncertainties of the deterioration
process (Baldi et al. 2016; Andrews et al. 2014), but they do not consider assets
criticalities and risk evaluation.
On the other hand, Carretero et al. (2003), Bharadwaj (2007), Tezuka et al.
(2015), Sun et al. (2017) apply reliability concepts to plan periodic maintenance,
taking into account the assets criticalities by means of a FMECA (Failure Mode
Effect and Criticality Analysis) procedure.
They develop risk-based methodologies to estimate the optimal time of
replacement or repair of a railway structure, but they do not take into account
predictive models of assets condition.

123
440 A. Consilvio et al.

Therefore, in literature many different approaches have been proposed to

schedule maintenance activities (Table 1).
Nevertheless, new steps forward are needed to adapt railway maintenance
schedules according to the inputs received from predictive models, especially for
technological assets.
Table 1 shows that the present model allows to take into account more aspects
than the previous ones, through the application of both risk-based and predictive
principles in the maintenance planning. More in detail, risk assessment and safety
are considered in the prioritization of maintenance activities, according to the
standard for risk-based asset management, the ISO 55000 guidelines, and the
prediction of the future asset condition is estimated through predictive models.
Moreover, the geographically distributed aspect of railway assets and the
maintenance teams availability are considered in order to optimize rail maintenance
performance and resources utilization.
In Consilvio et al. (2016) a toy case is described to validate a risk-based
scheduling model with stochastic deadlines for the planning of rail track
maintenance activities, by introducing a degradation curve for rail track. Instead,
in this paper, a risk-based scheduling model, characterized by deterministic
deadlines, is introduced to show the applicability of the approach to a real rail
network for the maintenance planning of technological assets.
This implies that a bigger instance of the problem is considered, taking into
account the time constraints related to real operations and the geographical
constraints of a real network.
In particular, the main advantages of the proposed approach, that will be
presented more in detail in the next sections, are:
• A reduction of the risk of critical failure;
• The reduction of corrective maintenance activities (needed only in case of
unpredicted failures);
• The elimination of unnecessary maintenance activities that would be instead
performed applying a preventive routine maintenance strategy;
• A more efficient use of the infrastructure possession time;
• An optimized management of maintenance teams.

Finally, it is worth mentioning that the proposed optimization model can be

applied to different kind of asset, by modifying the degradation curve that describes
the asset status behaviour.

3 The risk-based approach

In this section, the generic risk-based formulation of the optimal scheduling

problem is described. With the aim of clarity, the reader can refer to Table 2 for the
relevant notation.
To determine the risk-based model, the following assumptions are considered.

123
 Table 1 State of the art comparison
Considered Geographical Resource optimization Degradation Risk-based Interaction with
maintenance policy level model approach train circulation

Peng et al. (2011) Preventive/routine Rail network Paths of maintenance Time windows
teams
Macedo et al. (2017) Preventive/routine Rail line Maintenance teams and Train-free intervals
equipment
Umiliacchi et al. (2011) Condition-based Rail line Ontology based
Dell’Orco et al. (2008), Condition-based Rail line Neurofuzzy
Wen et al. (2016) Condition-based Rail line Linear degradation
Su et al. (2017) Condition-based Rail line Time slots
Sun et al. (2017) Risk-based Rail line Assignment and Weibull Time windows
routing of teams distribution
Carretero et al. (2003) Risk-based Rail line FMECA
Bharadwaj (2007) Risk-based Rail line Risk
assessment
Risk-based optimal scheduling of maintenance activities…

Tezuka et al. (2015) Risk-based Weibull Probability

of failure
Quiroga et al. (2011) Predictive Rail line Exponential
Zhang et al. (2013) Predictive Rail network Paths of maintenance Weibull
teams at network level
Camci (2015) Predictive Network Paths of maintenance Exponential Probability
teams at network level of failure
Letot et al. (2016) Predictive Rail line Wiener
Consilvio et al. (2015) Risk-based Rail line Assignment and Deterministic Train-free sub-
Predictive sequencing of activities soft and hard intervals
(on assets) to work risk
teams deadlines
441

123
 Table 1 continued
442

Considered Geographical Resource optimization Degradation Risk-based Interaction with

maintenance policy level model approach train circulation

123
Consilvio et al. (2016) Risk-based Rail network (toy case) Assignment and Stochastic soft Train-free sub-
Predictive sequencing and hard risk intervals
of activities (on assets) deadlines
to work teams and paths
Presented work Risk-based Medium-scale rail Assignment and Weibull Deterministic Train-free sub-
Predictive network sequencing soft and hard intervals
of activities (on assets) risk
to deadlines
work teams and paths
A. Consilvio et al.
Risk-based optimal scheduling of maintenance activities… 443

Table 2 Notation of the risk-based model

Sets and deterministic parameters
A Set of the considered assets and jAj the relevant cardinality
i ¼ 0; 1; . . .; jAj Index of assets
s Time at which the risk assessment is performed
D Generic time interval
Di Loss caused by the fault of the asset i
 i
tk k¼0;1;... Sequence of time instants at which the maintenance of the asset i are performed
t0i Time of the first installation of the asset i
Ri ðsÞ Failure risk at s of asset i
Ri Maximum tolerable risk of asset i
Stochastic variables
  
Fi s; h tki Cumulative distribution function (CDF) of the random fault process of the asset i, that
is, the probability that a fault occurs within the time s
 
h tki Vector of CDF parameters depending on previous maintenance activities performed in
tki
 i

ui s; D; tk Conditioned fault probability
Output parameters
 
dhi tki Hard deadline of the maintenance activity on asset i after the maintenance occurred in
tki
 i
dsi tk Soft deadline of the maintenance activity on asset i after the maintenance
occurred in tki
 
dgi tki Instant when it is necessary to start planning the maintenance of asset i, after the
maintenance occurred in tki . Such a parameter can be interpreted as the release time
of asset i in the scheduling problem

First, at any time s the risk Ri (s), related to a fault of the asset
 i occurring
 in the
interval ½s; s þ D given that it has not occurred in the interval tki ; s ; that is, since
the last maintenance activity performed in tki , is assumed to be
 
Ri ðsÞ ¼ ui s; D; tki  Di ð1Þ
 
being ui s; D; tki the probability that a fault will occur within s þ D, conditioned by
the fact that it has not occurred yet. Since this process has memory  and the
degradation state of an asset worsens with time, it is assumed that ui s; D; tki is a
non-decreasing function. With this hypothesis, and given the Cumulative Distri-
bution Function (CDF) F s; h tki of the stochastic variable representing the fault
 
occurrence time, the function ui s; D; tki can be written as the probability that the
fault occurs within a certain interval D starting in s, given that the fault has not
occurred in s, that is

123
444 A. Consilvio et al.

     
    i  F s þ D; h tki  F s; h tki
ui s; D; tki ¼ Pi t  s þ Djt  s; h tk ¼    : ð2Þ
1  F s; h tki

It is worth noting that thefault

 occurrence time results to be a stochastic process
since the CDF parameters h tki depend on time and, in particular, on the times at
which maintenance activities have been performed.
Note that the model in (2) does not introduce assumptions about the shape of the
fault occurrence time CDF, but only requires, in a very general form, that a time
dependence of the considered stochastic variables is able to represent the
degradation process.
Coming back to the model in (1), it is assumed that both the maximum tolerable
risk Ri for any asset i, and the consequent losses Di are externally defined and
known. With this assumption, it is easy to compute the maximum tolerable value of
the conditioned fault probability in (2), as ui ¼ Ri =Di . In addition, it is assumed
that, given s, it is possible to determine,
 analytically or numerically, the interval
duration D during which ui s; D; tki is below the value ui , that is
   
D tki ¼ u1
i ui ; tki  s; ð3Þ
 i
where the notation D tk points out that this interval depends on the last performed
maintenance activity. Nevertheless, aiming at keeping the notation simple, this
dependence is indicated only when necessary.
Focusing on technological assets, the fault occurrence time can be described by a
Weibull stochastic process:
0 !gðtki Þ 1
  i   i  t
F t; k tk ; g tk ¼ 1  exp@  i  A ð4Þ
k tk
   
with k tki and g tki being the scale and shape parameters determined by the
maintenance activity performed in tki :
This class of probability distributions is chosen since it best fits the considered
case study, although it is worth underlining that the proposed approach can be
applied using different kind of stochastic processes.
Moreover, it is well known that the Weibull distribution can be suitably used to
represent the time-dependent probability that a fault occurs within a certain interval,
given that it is not occurred since the last maintenance
  activity. In addition, the time
dependence of the parameters k tki and g tki allows to model the fact that the
maintenance activity performed in tki is not able to restore the initial state of an asset.
In other words, the decrease with the time sequence tki , k ¼ 0; 1; . . .; of the scale and
   
shape parameters k tki and g tki points out that it is impossible to restore the initial
  i    i 
conditions characterizing the new asset. In fact, k tki  k tk1 and g tki  g tk1
 
indicate that, given the same interval D, a fault in tki ; tki þ D is more probable than
i i

a fault in tk1 ; tk1 þD :

123
Risk-based optimal scheduling of maintenance activities… 445

 
Once the function ui s; Dh ; tki is determined and thanks to the assumption that it
is possible to compute the time s þ Dh at which it reaches the threshold ui , the so-
called hard deadline is determined as the instant
   
dhi tki ¼ s þ Dh tki ; ð5Þ

where in analogy with (3), the dependency of dhi on tki points out that they can vary
after each maintenance activity. Analogously, a soft deadline can be determined as
the instant
   
dsi tki ¼ s þ Ds tki ð6Þ
 
such that ui s; Ds ; tki ¼ as ui , being as 2 ð0; 1Þ a parameter indicating a tolerable
percentage of the fault probability ui or, equivalently, a tolerable percentage of the
fault risk. Finally, it is possible to define the required maintenance release time as
the instant
   
dgi tki ¼ s þ Dg tki ð7Þ
 
such that ui s; Dg ; tki ¼ ag ui , being ag 2 ð0; as Þ a parameter analogous to as . The
instant dg expresses the idea that it is possible to avoid considering, in the math-
ematical programming model, all the assets with release time greater than the
planning horizon, as their degradation state will not reach a non-negligible degraded
condition before the last maintenance activity of the considered assets has to be
concluded. Aiming at simplifying the notation, in the following, the explicit
dependence of the deadlines and of the release time on the time tki will be dropped
whenever not necessary.
Some examples of the soft and hard deadlines for a balise maintenance
(characterized by a Weibull fault probability) are reported in Fig. 1, where it is
possible to note that the slopes of the functions ui s; Ds ; tki increase with tki and,
consequently, the deadlines decrease.
It is worth mentioning that, considering the above described decrease of the
deadlines after the execution of the maintenance activities, the cost reduction can be
evaluated comparing the schedule defined by a planned routine maintenance and the
proposed approach. In this way, the number of unnecessary maintenance activities
that would be executed and the number of failure, that would not be prevented using
the cyclical approach, can be evaluated (Tezuka et al. 2015). In  particular,
  Fig.
 1
shows that considering the dark grey curve with parameters k t0i and g t0i and
 
performing the maintenance activity at s þ D t0i ¼ 500 days, the soft deadline and
the hard deadline are not exceeded. After the first maintenanceintervention,
   the
medium grey curve (characterized by the parameters k t1i  k t0i and
     
g t1i  g t0i ) shows that the execution of the intervention at s þ D t1i ¼
500 days implies that the soft deadline is exceeded while the hard deadline is not
reached. Finally, when the second maintenance
  intervention is executed, if a further
maintenance is performed at s þ D t2i ¼ 500, both the soft and the hard deadlines
are exceeded.

123
446 A. Consilvio et al.

       
Fig. 1 Comparison between the conditioned probabilities ui s; D t0i ; t0i ; ui s; D t1i ; t1i
  i i
and ui s; D t2 ; t2 evaluated in s ¼ tki þ 365 days, that is, a year after the maintenance

To cope with this problem, the inter-time between two consecutive maintenance
activities could be reduced. Nevertheless, fixed reductions may lead to an avoidable
reduction of the asset useful life and to an increase of maintenance costs.
Coming back to the proposed approach, being s  tki the instant at which the last
risk assessment has been performed, a generic asset i is in the following states:
• A good condition 8t 2 ðs; dgi Þ; that is, the asset is characterized by a negligible
degradation;
• A maintenance awaiting condition 8t 2 ðdgi ; dsi Þ; that is, the asset is charac-
terized by a small degradation, although its degradation process has already
begun;
• An acceptable degraded condition 8t 2 ðdsi ; dhi Þ; that is, the asset is charac-
terized by a non-negligible degradation, although it is still sufficiently far from a
high fault risk. In this state, the asset needs a maintenance activity;
• An unacceptably degraded condition 8t  dhi that is, the state of the asset is
characterized by a significant degradation that should never be reached.

In this way, the soft and hard deadlines are determined according to the asset
state, and in particular to the distance from the thresholds ui ; as ui ; and ag ui . This
means that an asset that is close to exceed the threshold will have a close deadline,
implying the possibility of considering different severity levels between assets in the
same state.
Once the condition of each asset is identified and the deadlines determined, it is
possible to evaluate the best schedule of the maintenance activities by means of the
Mixed Integer Linear Programming (MILP) problem described in Sect. 4 and only
considering the assets that have reached the maintenance awaiting condition.

123
Risk-based optimal scheduling of maintenance activities… 447

To conclude, it is worth remarking that the model, proposed in this section, can
be also applied considering different stochastic processes, provided that they are
able to capture the degradation dynamics of assets.
In this connection, an example of a different degradation process for rail
maintenance is proposed in Famurewa et al. (2016), where a model of the variance
of the longitudinal deformation of track is determined by means of measurements,
showing that it increases with time and cannot be restored to the initial value by
maintenance activities.

4 The maintenance scheduling problem

In this section, the MILP problem, to determine the optimal assignment of the
maintenance activities to the available maintenance teams and the relevant
schedule, is presented. In addition, the optimal path of each maintenance team
along the network is determined by taking into account the space-distributed aspect
of railway infrastructure and the necessary time to move between two given
different sites of the network. In doing so, the deadlines determined by means of the
model presented in Sect. 3 are considered as constraints or terms of the cost
function. As regards the factors that define the limited periods for maintenance, they
are taking into account by introducing train-free sub-intervals during which the train
circulation is forbidden and the maintenance activities can be performed.
In the next sections, the problem assumptions, formulation and solution approach
will be described. For what concerns the relevant notation, it is reported in Table 3
that integrates the definitions reported in Table 2.

4.1 Problem assumptions

In the following, without losing generality, it is assumed that the problem is stated
in a generic time instant together with the relevant risk modelling approach
described in Sect. 3.
Further assumptions considered in the MILP problem formulation are the
following:
• Only the assets that have hard deadlines in the interval ðs; s þ T Þ and release
time dgi  s þ T are considered, being T a suitably chosen horizon;
• The interval ðs; s þ T Þ is subdivided into non-continuous train-free sub-intervals
during which the train circulation is forbidden and maintenance activities can be
performed;
• All maintenance teams are available in s;
• The set-up time of each maintenance activity performed by each maintenance
team is sequence-dependent;
• The processing times of different maintenance teams on the same asset can be
different;
• Each maintenance team can perform only an activity at a time;
• Each asset is characterized by distinct soft and hard deadlines;

123
448 A. Consilvio et al.

Table 3 Notation of the scheduling model

Sets
T Set of maintenance teams, and jT j its cardinality
N Set of railway nodes
L Set of railway links
G ¼ fN ; Lg Graph representing the considered railway network
Htk ;T Set of the train-free sub-intervals of the interval ðtk ; tk þ TÞ
Indexes
i ¼ 0; . . .; jAj Indexes of assets
i=0 Dummy activity artificially introduced to correctly identify the first real one
j ¼ jAj þ 1 Dummy activity artificially introduced to correctly identify the last real one
D Matrix of the distances between the rail nodes
Matrices
A Adjacency matrix of the graph: element An,l is equal to 1 if node n 2 N and node l 2 N
are connected by link ðn; lÞ 2 L, and 0 otherwise
M Inclusion matrix: element Mi(n,l) is equal to 1 if asset i is located on link ðn; lÞ 2 L, and to
0 otherwise
D Matrix of the distances between the rail nodes
Constants
pi;g Processing time of maintenance activity on asset i by maintenance team g
T Maintenance planning horizon
‘h Length of the hth sub-interval in Htk ;T
Ih Initial time of the hth sub-interval in Htk ;T
si;j;m Set-up time for assigning activity on asset j after activity on asset i to maintenance team
þ
m. Such a set-up time can be interpreted as the sum si;j;m ¼ ci;j;m þ c i;m þ cj;m where
ci;j;m is the part of travel time along the rail network needed by team m to move from
the asset i to the asset j, c i;m is the removal time of the repair yard of the completed
maintenance activity on asset i by the team m, and cþ j;m is the preparation time of the
repair yard of the next scheduled maintenance activity on asset j by team m. Note that
the maintenance teams enter the railway network in correspondence of the access point
closest to the asset to be maintained and then they travel along the network to reach the
asset
xi Maintenance activity priority of the asset i
B Integer suitably chosen to approximate ?? in the constraints
aj Weight of the objective function jth terms, chosen according to planner intentions
Variables
ci;m Completion time of maintenance activity on the asset i performed by maintenance team
m
ti,m Starting time of maintenance activity on the asset i performed by maintenance team m
qi Tardiness of maintenance activity on asset i
xi;j;m Binary sequence variable equal to 1 if the activity on asset j is performed immediately
after the activity on asset i by maintenance team m, and 0 otherwise
yi;m Binary assignment variable equal to 1 if the activity on asset i is performed by
maintenance team m, and 0 otherwise
un;m Binary variable equal to 1 if node n 2 N is the origin of maintenance team m
vn;m Binary variable equal to 1 if node n 2 N is the destination of maintenance team m

123
Risk-based optimal scheduling of maintenance activities… 449

Table 3 continued

zðn;lÞ;m Binary variable equal to 1 if link ðn; lÞ 2 L belongs to the optimal path of maintenance
team m 2 T , and to 0 otherwise
wi;h Binary variable equal to 1 if the activity on asset i is performed in the time interval h, and
0 otherwise
Pm Total path among the sites of the assets travelled by the maintenance team m 2 T
DPm;g Difference between the paths travelled by the maintenance team m and g
DEm;g Difference between the number of jobs assigned to maintenance team m and g

• No pre-emption of maintenance activities is allowed;

• The maintenance activities can be tardy only with respect to the soft deadlines;
• The maintenance teams are available throughout all the considered horizon T;
• The maintenance activities on all the assets can be processed by any free
maintenance team.

4.2 Problem formulation

Given the above notation, the optimization problem can be formalized as

jT j X
jAj jAj
! jT j jT j X
jT j
X X X X  
mina1 xi ci;m þ x i qi þ a 2 Pm þ a3 DPm;g þ DEm;g
m¼1 i¼1 i¼1 m¼1 m¼1 g¼1;
g6¼m
ð8Þ
subject to:
jA
X jþ1
ci;m ¼ ti;m þ pi;m xi;j;m 8i 2 A; 8m 2 T ; ð9Þ
j¼1;j6¼i

 
ti;m  dgi tki 8i 2 A; 8m 2 T ; ð10Þ
 
ci;m  dhi tki 8i 2 A; 8m 2 T ; ð11Þ
  
qi ¼ max 0; ci;m  dsi tki 8i 2 A; 8m 2 T ; ð12Þ
 
tj;m  ci;m þ si;j;m  B 1  xi;j;m 8m 2 T ; 8j; i 2 A; ð13Þ
 
tj;m  s0;j;m  B 1  x0;j;m 8m 2 T ; 8j 2 A; ð14Þ
 
xi;j;m  min yi;m ; yj;m 8m 2 T ; 8i; j 2 A : i 6¼ j; ð15Þ

123
450 A. Consilvio et al.

jA
Xjþ1
x0;j;m  1 8m 2 T ; ð16Þ
j¼1

jT j X
X jAj
xi;j;m ¼ 1 8j 2 A; ð17Þ
m¼1 i¼0;i6¼j

jT j jA
X X jþ1
xi;j;m ¼ 1 8i 2 A; ð18Þ
m¼1 j¼1;j6¼i

jA
X jþ1 jAj
X
xj;e;m  xi;j;m ¼ 0 8j 2 A; 8m 2 T ; ð19Þ
e¼1;e6¼j i¼0;i6¼j

tj;m  Ih wj;h þ si;j;m xi;j;m 8i; j 2 A; 8h 2 Htk ;T ; 8m 2 T ; ð20Þ

 
ci;m  Ih þ ‘h þ B 1  wi;h 8i; 8h 2 Htk ;T ; 8m 2 T ; ð21Þ

jT j jH
X X tk ;T j

wi;h ¼ 1 8i 2 A; ð22Þ
m¼1 h¼1

jT j
X
yi;m ¼ 1 8i 2 A; ð23Þ
m¼1
 
DPm;g ¼ Pm  Pg  8g; m 2 T : g 6¼ m; ð24Þ
 
XjAj jAj
X 
 
DEm;g ¼ yj;m  yj;g  8g; m 2 T : g 6¼ m; ð25Þ
 j¼1 j¼1


jT j
X
ðzðn;lÞ;m þ zðl;nÞ;m Þ  1 8ðn; lÞ 2 L : An;l ¼ 1; ð26Þ
m¼1

jT j
X
ðzðn;lÞ;m þ zðl;nÞ;m Þ  jT j 8ðn; lÞ 2 L : An;l ¼ 1; ð27Þ
m¼1

zðn;lÞ;m þ zðl;nÞ;m  1 8ðn; lÞ 2 L : An;l ¼ 1; ð28Þ

X
Pm ¼ zðn;lÞ;m Dn;l 8m 2 T ; ð29Þ
ðn;lÞ2L

123
Risk-based optimal scheduling of maintenance activities… 451

jN j
X
un;m ¼ 1 8m 2 T ; ð30Þ
n¼1

jN j
X
vn;m ¼ 1 8m 2 T ; ð31Þ
n¼1

yi;m  zðn;lÞ;m  zðl;nÞ;m  0 8m 2 T ; 8i 2 A; 8ðn; lÞ 2 L : An;l ¼ 1; Mi;ðn;lÞ ¼ 1;

ð32Þ

jN j
X jN j
X
zðn;lÞ;m  zðl;nÞ;m ¼ un;m  vn;m 8n 2 N ; 8m 2 T ; ð33Þ
l¼1: l¼1:
anl ¼1 anl ¼1
un;m þ vn;m  1 8n 2 N ; 8m 2 T ; ð34Þ

ti;m ; ci;m ; qi 2 Rþ;0 8i 2 A; 8m 2 T ; ð35Þ

Pm 2 Rþ;0 ; DPm;g 2 R; DEm;g 2 Z 8g; m 2 T ; ð36Þ

xi;j;m 2 f0; 1g 8m 2 T ; 8j; i 2 A : i 6¼ j; ð37Þ

yi;m 2 f0; 1g 8m 2 T ; 8i 2 A; ð38Þ

wi;h 2 f0; 1g 8i 2 A; 8h 2 Htk ;T ; ð39Þ

zðn;lÞ;m 2 f0; 1g 8m 2 T ; 8ðl; nÞ 2 L; ð40Þ

un;m ; vn;m 2 f0; 1g 8m 2 T ; 8n 2 N ; ð41Þ

where
• The constraints in (9) define the completion times of the maintenance activities;
• The constraints in (10) define the release times of the maintenance activities;
• The constraints in (11) guarantee that the maintenance activity on each asset i is
completed before the relevant hard deadline;
• The constraints in (12) define the tardiness of the maintenance activities;
• The constraints in (13) define the precedence relation between the activities on
the assets i and j and the sequence-dependent set-up time;
• The constraints in (14) define the initial set-up times;
• The constraints in (15) impose that if the activity on asset j is performed soon
after the activity on asset i by maintenance team m both the asset i and the asset
j are assigned to the same maintenance team m;
• The constraints in (16) guarantee that at most one activity is scheduled as the
first work of each maintenance team;
• The constraints in (17) and (18) guarantee that every maintenance activity has
exactly one predecessor and one successor, respectively;

123
452 A. Consilvio et al.

• The constraints in (19) state that a predecessor/successor pair of activities has to

be assigned to the same maintenance team m;
• The constraints in (20) guarantee that if an activity is performed in the train-free
interval h, it starts after the beginning of such an interval;
• The constraints in (21) guarantee that all maintenance activities finish within the
train-free sub-interval;
• The constraints in (22) guarantee that each activity is only performed in one and
only one train-free sub- interval;
• The constraints in (23) guarantee that each activity is assigned to one and only
one maintenance team;
• The constraints in (24) define the path length differences between one
maintenance team and each other;
• The constraints in (25) define the differences in the number of activities assigned
to one maintenance team and the others;
• The constraints in (26) and (27) state that the link (n, l) has to be assigned at
least to one and at most to jT j maintenance teams. In any case, this constraint
allows reducing the number of teams on a link to 1 whenever the kind of
performed maintenance does not allow team’s overtaking, for instance when the
machinery occupies the rail;
• The constraints in (28) guarantee that each maintenance team can travel along
the link (n, l) in only one direction;
• The constraints in (29) define the length of the path for each maintenance team;
• The constraints in (30) and (31) guarantee that each maintenance team m has
only one origin and one destination, respectively;
• The constraints in (32) state that if asset i is located on link (n, l) and
maintenance team m crosses that link, i may be assigned to maintenance team
m or not;
• The constraints in (33) define all the possible paths for each maintenance team
m from its origin un,m to its destination vn,m. The peculiarity of such constraints
stands in the fact that the origin and the destination of all the maintenance teams
are optimally chosen by the model. In fact, these constraints depend on the
values assumed by the right-hand-side variables, whose values are constrained
by (34) ensuring that the same node n cannot be origin and destination of the
same maintenance team m;
• The constraints in (35–41) define the problem variables.

It is worth noting that the weights aj in the objective function are chosen by the
maintenance service provider, according to the strategic goals of its organization.
Anyway, this choice cannot affect the safety level of the system performance, since
the constraints in (11) guarantee that the hard deadlines are always fulfilled. If the
maintenance service provider gives a low value to the weight a1, making the soft
deadlines fulfilment negligible in the cost function, the safe condition is still
guaranteed. In other words, even if the maintenance can be performed very late with
respect to the soft deadlines, the assets still remain in an acceptable degraded
condition.

123
Risk-based optimal scheduling of maintenance activities… 453

Another aspect to be pointed out is that, even if the monetary cost of the
operations is not explicitly indicated in the objective function, it is implicitly
considered in the model, since planning the maintenance activities before the hard
deadline allows to avoid failure and costly service interruptions. Moreover, the cost
reduction is modelled in terms of reduction of the time needed to complete the
maintenance activities and optimization of resources utilization, through the
definition of the path of each maintenance team.

4.3 Problem solution

As mentioned, it is well known that the above optimization problem belongs to the
NP-hard class and, even for relatively small instances, requires a very long time to
be solved. Therefore, in this work, a matheuristic solution approach, based on the
general framework described in (Della Croce et al. 2014), is proposed to solve the
instances of the problem in (8–41). Although the details about the generic
matheuristic approach are beyond the scope of this paper, it is sufficient recalling
here that it is based on the idea of decomposing the original problem in smaller
problems, whose optimal (exact) solutions are used to refine heuristically the best
solution of the original problem. In doing so, matheuristic approaches combine the
strength of both approximated metaheuristic and exact methods. The customized
matheuristic algorithms, developed for solving the problem described in Sect. 4.2, is
reported in Figs. 2, 3, where it is possible to note that they are divided into several
main steps:

If Algorithm 1 provides a feasible solution, the scheduling plan improvement is

obtained by means of two further steps:

123
454 A. Consilvio et al.

As regards the specific optimization problems:

• In the first step of Algorithm 1, only the path travelled by the teams and the
balancing of the number of activities assigned to them are considered;
• In the second step of Algorithm 1 and in the first step of Algorithm 2 only the
completion times are optimized;
• Finally, the second step of Algorithm 2 considers the complete problem in
(8–41).

As concerns the comparison of the matheuristic performance with those provided

by a branch-and-bound approach implemented by the general purpose solver IBM-
Ilog Cplex, consider the outputs reported in Fig. 4 where the cost function values
for different instance sizes are reported in a logarithmic scale. Such results have
been obtained by implementing the matheuristic algorithm in Matlab and IBM-
Ilog Cplex on a 3.10 GHz PC with 16 GB RAM.
In addition, in such a figure, the computational times required by Cplex and by
the proposed algorithms to find the solutions or bounds are also indicated. Note that
the increasing values of the objective function are due to the increase of the problem
size and it does not indicate a worsening of performance. Indeed, the values of the
objective function are represented in order to show the gap from the lower bound.

Algorithm 1: Initial Solution Generation

Let be the maximum time for the solution of the task-team
Initialization assignment problem
Let be the maximum time for the activity sorting problem
Step 1: Task Assignment problem
1.a While ( or the optimum is found)
Set and minimize only with respect to the
constraints (23-34),(36),(38),(40-41)
End while
Let be the found solution, where is the vector
gathering the best task-team allocation variables and its
sub-vector associated with the team , is the vector
gathering the best link-team allocation variables
Let be the initial sequencing variable set
Step 2: Independent Sequences Assignment
Set
If the “TOP TO END” strategy is feasible, add the relevant
2.a
solution to go to Step 2.d
While ( or the optimum is found)
2.b Given , set and minimize (8) with respect to
the constraints (9-23), (35), (37-39)
End while
If a solution does not exist, go on Step 3.a, otherwise add
2.c
to and go to Step 2.d
2.d If , then is a feasible initial solution, otherwise
set and go to Step 2.a
End for
Step 3: Tabù strategy

3.a Define a new constraint making the solution

unfeasible and go to Step 1.a

Fig. 2 Initialization algorithm

123
Risk-based optimal scheduling of maintenance activities… 455

Algorithm 2: Solution refinement

Initialization Let be the maximum number of refinement iterations
Let be the maximum time for the whole solution refinement
Let be the maximum time for the matheuristic solution
refinement
While and
Step 1. For
Let the variables free to vary and fix all the
others. Set and minimize (8) with respect to
the constraints (9-23), (35), (37-39)
End for
Step 2. While
Randomly choose a subset
Fix the variables in
Minimize (8) with respect to , and
This problem is subject to and to the constraints (9-
41)
Let be the best found integer solution
If the solution is better than the solution

Set
End if
Set
End While
End While

Fig. 3 Matheuristic algorithm

Fig. 4 Cplex and matheuristic cost functions comparison. The labels represent the time required for
computing the values

As regards the results, Cplex is able to determine the optimal solution (defined as
‘Cplex optimum’ in the figure), in a short time, for instances that include up to 15
activities, while it is only able to compute feasible solutions (defined as ‘Cplex best
solutions’) for instances up to 25 activities. The reported cost function values are the
best found after 48 h together with the relevant lower bounds. Considering the gap,
it is clear that a good lower bound could not be found for big instances of the
considered problem. In addition, for instances with more than 25 activities, for
which even a feasible solution could not be found by Cplex, the reported values of

123
456 A. Consilvio et al.

the lower bound are the best obtained after 0.3 h, that is, the time needed by the
matheuristic to find its best solution.
For graphical reasons, in Fig. 4 only the instances until 40 activities are
considered even if the matheuristic is able to solve bigger size of the problem, as
shown in the following.
For what concerns the matheuristic optimization performance, the applications of
the proposed approach have shown, after a computational time of 1 h, an
improvement of about the 20% of the cost function with respect to the reference
initial solution provided by Algorithm 1. In Fig. 5, the comparison between the
initial solution and the final solution found by the matheuristic algorithm is
depicted. In looking at this result, it is worth keeping in mind that the values in the
graphs consider all the terms in the cost function defined in (8). A discussion about
the risk reduction is specifically reported in Sect. 6.

5 Model applicability and data requirements

In this section, the real world applicability of the mathematical programming

problem and of the relevant solution approach is briefly discussed, focusing in
particular on the input data requirements.
Then, with reference to the sets, the matrixes, and constants in Tables 2, 3, the
required data result to be:
• The failures occurred in the past;
• The maintenance activities executed in the past;
• The losses caused by the failures;
• The processing time of the maintenance tasks;
• The position of the asset;
• The speed of maintenance teams for moving along the network;
• The availability of maintenance teams;
• The available train-free sub-intervals for the maintenance interventions.

As regards the availability of this information, it is worth mentioning that:

• The failures occurred in the past are usually traced by the infrastructure manager
and the maintenance service providers. This means that the failure rates of the
assets are usually known;
• The executed maintenance activities are usually known as the infrastructure
manager and the maintenance service provider have to take trace of the
interventions performed by their teams;
• Each maintenance team has usually to communicate the starting and finishing
times of the intervention, making possible to know the duration of each
maintenance task and, for different kind of interventions, the relevant averages
considered in the following;
• The positions of the assets are known by means the railway line scheme;

123
Risk-based optimal scheduling of maintenance activities… 457

Fig. 5 Cost function improvement with respect to the initial solution considering 42 maintenance
activities (a) and 84 maintenance activities (b)

• The speed of the maintenance teams moving from an asset site to another
depends on the transportation mode they use to travel (working machines,
diagnostic trains, etc.);
• The number of available maintenance teams is a priori established by the
infrastructure manager and the maintenance service provider;
• The starting and final times of each train-free sub-interval are provided by the
traffic manager.

The losses caused by the failure represent the most difficult input parameter to be
defined. The infrastructure manager usually knows the criticality of the different
assets and the expected loss caused by their failures, although this information is

123
458 A. Consilvio et al.

sensitive and in general not shared. To cope with this problem, the maintenance
service provider can estimate these data from field by means of railway experts.
To conclude, it is worth noting that some of the above input data, such as the
processing time, are not deterministic. Therefore, work is in progress in order to
take into account these stochastic aspects in the considered model.

6 Application and results

In this section, the proposed approach is applied to two real world case studies and
the relevant results are discussed. To this aim, the North-Western Italian rail
network is considered. Its total length is 791 km and it can be represented by means
of a graph with 14 nodes and 21 links, as depicted in Fig. 6.
As regards the problem size in the two cases, the maintenance activities of 42 and
84 technological components, respectively, are considered to be scheduled. The
relevant risk modelling has been obtained by identifying suitable Weibull distri-
butions for all the asset kinds, whose parameters are reported in Table 4.
By applying the model described in Sect. 3, the soft and hard deadlines have been
calculated for each asset at time s ¼ t0i þ 365 days, that is 1 year after their
installation, taking into account a maximum tolerable value of risk of 0.510-4. Since
the values of the expected losses caused by the failures are sensitive data, only a
fictitious risk threshold is reported here.
As regards the parameters for the model in Sect. 4, it is assumed that four and six
maintenance teams mi are available in the first and second case study, respectively.
Moreover, train-free sub-intervals that start at each midnight and last 5 h are taking
into account. In addition, since a weekly schedule of maintenance teams activities is
required, the horizon T is set to 168 h so as to determine a short-term operational
planning. Finally, the travelling speed of the maintenance teams and the processing
time of the maintenance activities are given by the available data.
The optimal solution of the optimization problem is obtained via the matheuristic
approach above described by considering the following parameter values: tass ¼
Step1 Step2
0:05h; tord ¼ 0:03h; K ¼ 50; tmax ¼ 1h and tmax ¼ 0:2h: Note that K is the number
of iterations of the total cycle, while the number of iterations of the different steps
are into the order of hundreds. In any case, for all the instances the algorithms
stopped before reaching the threshold parameter values since no further improve-
ment was found.
The results depicted in Fig. 7a, b show the links assigned to maintenance teams
in the initial solution and final solution. In such figures, it is worth noting that some
teams cross the same link. In particular, in Fig. 7a team 2 and team 4 cross the same
links S. Giuseppe di Cairo-Savona and Arquata S. -Novi L., whereas in Fig. 7b:
• Team 1 and team 4 cross the links Turin–Novara, Novara–Milan, Milan–
Voghera;
• Team 1 and team 2 cross the link Turin–Asti;
• Team 2 and team 3 cross the link Alessandria–Novara;
• Team 2 and 4 cross the link S. Giuseppe di Cairo–Savona.

123
Risk-based optimal scheduling of maintenance activities… 459

Fig. 6 Railway network scheme

Table 4 Fault rate of railway technological assets: Weibull distributions after the first installation
   
Asset Number of considered Number of considered k t0i g t0i
assets case study 1 assets case study 2 [days]

Radio block centre 3 6 1690 3.2

Interlocking 3 6 1120 4.5
Balise 15 30 1378 5.6
Track circuits 10 20 1115 3.6
Axle bearings temperature 5 10 1251 3.8
detection systems
Switch and crossings (IT 6 12 1459 4.1
subsystem)

This is allowed by the constraints (27) where the parameter jT j is set to 2. In this
way, some assets of the same link could be assigned to up to 2 different maintenance
teams in order to satisfy the deadlines and reduce the total completion time. It is
interesting to note that the total path in the final solution is greater than the initial

123
460 A. Consilvio et al.

Fig. 7 Links assignment according to the initial (a) and final (b) solutions

solution one (with an increase in length of 16%) since, in order to execute the
maintenance activities in a shorter total time, the maintenance teams sometimes
need to cross the same links and travel for longer distances.
The results depicted in Fig. 8a, b show the activities assignment to maintenance
teams according to the initial and final solution. It is to be noted that in Fig. 8a and
in Fig. 8b:
• Each numbered rectangle represents the duration of a maintenance activity;
• A different colour is assigned to each team (grey, medium grey, light grey,
white);
• The black rectangles between each pair of numbered rectangles represent the
set-up times;
• The dark grey rectangles represent the time period in which maintenance is
forbidden due to the trains circulation.

The comparison between Fig. 8a and Fig. 8b shows a numerical reduction of the
total completion time of 20%. Graphically, such a result is represented by the fact
that in the final solution all the jobs are executed within the fifth train-free sub-
interval.
More in detail, in the final solution:
• Maintenance activities 12 and 42 are assigned to team 4 instead of team 1;
• Maintenance activity 9 is assigned to team 1 instead of team 4;
• Maintenance activity 15 is assigned to team 2 instead of team 3;
• Team 3 finishes its activities in the train-free interval 5, instead of interval 6 with
a correspondent completion time reduction of 19%;
• Team 1 finishes in the same time interval but early, with a completion time
reduction of 2%;
• Teams 2 and 4 finish in the same time interval with a not relevant increase of
their completion times (0.9%);

123
Risk-based optimal scheduling of maintenance activities… 461

Fig. 8 Maintenance activities assignment according to the initial (a) and final (b) solutions

Regarding the difference of maintenance teams workloads, it is possible to note

that:
• The maximum difference between the completion time of the maintenance
teams is the 25% in the initial solution and the 2% in the final solution;
• The sum of the differences of the number of maintenance activities assigned to
the maintenance teams and the differences of their path lengths, provided by the
PjT j PjT j  
term m¼1 g¼1; g6¼m DPm;g þ DEm;g ; shows a reduction of 80%.

123
462 A. Consilvio et al.

Focusing only on the risk reduction achievable with the proposed approach, the
following scenarios are considered:
• Scenario 1: all the maintenance activities are completed in correspondence of
the hard deadlines;
• Scenario 2: all the maintenance activities are completed in correspondence of
the soft deadlines;
• Scenario 3: the maintenance activities are executed according to the proposed
initial schedule depicted in Fig. 8a
• Scenario 4: the maintenance activities are executed according to the proposed
final schedule depicted in Fig. 8b

In the following discussion, the generated losses are neglected since they are
constant. Therefore, keeping in mind that the failure risk increases with time (see
Fig. 1):
• In Scenario 1 each asset has a risk proportional to the probability thresholds ui ;
• In Scenario 2 the risk for each asset would be proportional to as ui ;
• Finally, in Scenarios 3 and 4 the risk depends on the specific completion times in
the initial and best scheduling solution, respectively.

The risk in Scenario 4 turns out to be - 90% comparing to Scenario 1, - 80%

comparing to Scenario 2, and - 25% comparing to Scenario 3. Nevertheless, it is
worth underlining that the Scenario 1 and Scenario 2 schedules may be non-feasible
solutions.
The second case study consists of the scheduling of 84 maintenance activities
whose Weibull parameters are depicted in Table 4. In analysing this case study, for
the sake of space limitation, only the final schedule is reported and shown in Fig. 9.
Comparing to the previous case study, it is worth mentioning that the number of
maintenance teams should be increased from four to six in order to find feasible
solutions, whereas the algorithms parameters are the same. The results show an
average reduction of the completion time of the maintenance teams of around 11%
and a reduction of risk of 28% respect to the initial solution. In this case, a
comparison with the fictitious schedules of Scenario 1 and Scenario 2 is not
considered.
Summarizing, the presented case studies have shown that the proposed modelling
and optimization approaches:
• Improve the reliability and availability of railway transport in term of reduction
of failure risk and of the relevant cost;
• Improve the railway asset management efficiency through the optimization of
the use of the available time for maintenance with a equalization of the
maintenance teams completion times.
• Optimize the resources utilization avoiding unnecessary trips of the maintenance
teams, balancing the workloads, etc. In particular, the number of maintenance

123
Risk-based optimal scheduling of maintenance activities… 463

Fig. 9 Maintenance activities assignment and sequence for the instance 84 activities and 6 teams

activities assigned to each maintenance team is equalized and the difference of

their path lengths is reduced.

7 Conclusions and future work

This paper describes a risk-based optimal scheduling procedure for railway

predictive maintenance. To this aim, a general fault probability model is first
proposed. Then, after having formulated a MILP scheduling problem, a matheuristic
algorithm has been described. The proposed approach is proven to be a suitable tool
to manage the short-term planning of a set of maintenance activities, as well as to
optimize the resources assignment and utilization. In particular, the results of the
case study scenarios point out the capability of the proposed approach to provide
good results, finding a good admissible solution in a short time. Work is in progress
to generalize and suitably relax the Stochastic Process assumptions introducing
stochastic deadlines to the scheduling problem. Another aspect, which will be taken
into account in the future development of the presented work, is the stochastic
nature of some inputs of the problem, such as the processing time of the
maintenance activities, that are currently considered as deterministic parameters.

123
464 A. Consilvio et al.

References

Andrews J, Prescott D, De Rozières F (2014) A stochastic model for railway track asset management.
Reliab Eng Syst Safe 130:76–84
Baldi M, Heinicke F, Simroth A, Tadei R (2016) New heuristics for the stochastic tactical railway
maintenance problem. Omega 63:94–102
Bharadwaj U (2007) A risk based approach to maintenance optimisation of business critical railway
structures/equipment. Seminar on the application of new technologies to railways, 2007 Institution
of Engineering and Technology, p 53–60
Borraz-Śanchez C, Yang D, Klabjan D, (2011) Strategic Gang Scheduling for Railroad Maintenance.
Operations research and management science, railway application section
Budai G, Huisman D, Dekker R (2004) Scheduling preventive railway maintenance activities systems.
Proc IEEE Int Conf on Syst ManCyber 5:4171–4176
Burrow MPN, Naito S, Evdorides HT (2009) Network-level railway track maintenance management
model. Transp Res Rec 2117:66–76
Camci F (2015) Maintenance scheduling of geographically distributed assets with prognostics
information. Eur J Oper Res 245:506–516
Carretero J, Pérez JM, Garcı́a-Carballeira JM, Calderón A, Fernández J, Garcı́a J, Cardona L, Lozano A,
Cotaina N, Prete P (2003) Applying RCM in large scale systems: a case study with railway
networks. Reliab Eng Syst Safe 82(3):257–273
Consilvio A, Di Febbraro A, Sacco N (2015) A modular model to schedule predictive railway
maintenance operations. In: Proc. of 2015 International Conference on Models and Technologies for
Intelligent Transportation Systems (MT-ITS), p 426–433
Consilvio A, Di Febbraro A, Sacco N (2016) Stochastic scheduling approach for predictive risk-based
railway maintenance. In: 2016 IEEE International Conference on Intelligent Rail Transportation
ICIRT Birmingham UK, p 197–203
Della Croce F, Grosso F, Salassa F (2014) A matheuristic approach for the two-machine total completion
time flow shop problem. Ann Oper Res 213:67–78
Dell’Orco M, Ottomanelli M, Caggiani L, Sassanelli D (2008) New decision support system for
optimization of rail track maintenance planning based on adaptive neurofuzzy inference system.
Transp Res Rec 2043:49–54
Ghodrati B, Famurewa S, Hoseinie SH (2016) Railway switches and crossings reliability analysis. In:
Proceedings of IEEE International Conference on Industrial Engineering and Engineering
Management (IEEM), pp 1412–1416
Higgins A (1998) Scheduling of railway track maintenance activities and crews. J Oper Res Soc
49:1026–1033
ISO International Standardization Organisation (2014) ISO 55000:2014 Preview asset management—
overview, principles and terminology
Jimenez N, Barragan A, Cembrero P, Benitez FG, Caceres N, Schubert F, Simroth A, Santanera C (2010)
ACEM-RAIL W.P.1 Task1.1 D1.1 Report on the state of practice of railway infrastructure
maintenance. Seventh Framework Programme theme sst.2010.5.2.1
Jiménez-Redondo N, Escriba S, Benı́tez FG, Cores F and Cáceres N (2014) Towards automated and cost-
efficient track maintenance. Final developments of the ACEM-Rail project. Transport Research
Arena 2014
Jothi R, Raghavachari R (2007) Approximating the k-traveling repairman problem with repair times.
J Discret Algorithms 5(2):293–303
Juntti U, Nissen A, Kumar U (2016) Augmented utilization of possession time: analysis for track
geometry maintenance. Proc Inst Mech Eng 230(4):1118–1130
Letot C, Soleimanmeigouni I, Ahmadi A, Dehombreux P (2016) An adaptive opportunistic maintenance
model based on railway track condition prediction. IFAC-PapersOnLine 49–28:120–125
Lidén T (2015) Railway infrastructure maintenance—a survey of planning problems and conducted
research, 18th Euro Working Group on Transportation, EWGT 2015, 14–16 July 2015, Delft, The
Netherlands, Transportation Research Procedia Vol 10, p 574–583
Macedo R, Benmansour R, Artiba A, Mladenovic N, Urosevic D (2017) Scheduling preventive railway
maintenance activities with resource constraints. Electron Notes Discret Math 58:215–222
Morant A, Larsson-Kråik PO, Kumar U (2016) Data-driven model for maintenance decision support—a
case study of railway signalling systems. Proc Inst Mech Engineers 230:220–234

123
Risk-based optimal scheduling of maintenance activities… 465

Patra AP, Kumar U (2010) Availability analysis of railway track circuit. Proc Inst Mech Eng
224:169–177
Peng F, Kang S, Li X, Ouyang Y (2011) A heuristic approach to the railroad track maintenance
scheduling problem. Comput Aided Civ Infrastruct Eng 26(12):129–145
Perrier N, Langevin A, Campbell JF (2007) A survey of models and algorithms for winter road
maintenance Part III: vehicle routing and depot location for spreading. Comput Oper Res
34(1):211–257
Quiroga LM, Antoni M, Schnieder E (2011) Heuristic based track maintenance scheduling optimization.
In: WCRR 2011—9th World Congress of Railway and Research, Lille, France, May 2011
Shingler R, Fadin G, Umiliacchi P (2008) From RCM to predictive maintenance: The InteGRail
approach. In: 4th IET International Conference on Railway Condition Monitoring, p 1–5
Simson SA, Ferreira L, Murray MH (2000) Rail track maintenance planning: an assessment model.
Transport Res Rec 1713:29–35
Soh SS, Radzi NHM, Haron H (2012) Review on scheduling techniques of preventive maintenance
activities of railway. In: Proc. 4th International Conference on computational intelligence,
modelling and simulation (CIMSiM), p 310–315
Su Z, Jamshidi A, Núñez A, Baldi S, De Schutter B (2017) Multi-level condition-based maintenance
planning for railway infrastructures—a scenario-based chance-constrained approach. Transp Res
Part C 84:92–123
Sun J, Zhang R, Qin S (2017) Turnout maintenance scheduling problem considering reliabilities:
modeling and optimization. In: Proceedings of the 36th Chinese Control Conference July 26–28,
Dalian, China
Tezuka M, Munakata S, Sawada M, (2015) Maintenance schedule optimization based on failure
probability distribution. In: Proceedings of the 2015 International Conference on industrial
engineering and operations management Dubai, United Arab Emirates (UAE), March 3–5
Umiliacchi P, Lane D, Romano F (2011) Predictive maintenance of railway subsystems using an ontology
based modelling approach. In: 9th World Conference on Railway Research
Wen M, Li R, Salling KB (2016) Optimization of preventive condition-based tamping for railway tracks.
Eur J Oper Res 252:455–465
Yokoyama A (2015) Innovative changes for maintenance of railway by using ICT—to achieve Smart
maintenance. In: The Fourth International Conference on through-life engineering services, Procedia
CIRP, Vol. 38 pp. 24–29
Zhang T, Andrews J, Wang R (2013) Optimal scheduling of track maintenance on a railway network.
QualReliab Eng Int 29:285–297
Zoeteman A, Esveld C (2004) State of the art in railway maintenance management: planning systems and
their application in Europe. IEEE Int Conf Syst Man Cyber 5:4165–4170

123