Reliability Engineering and System Safety 142 (2015) 169–183

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

journal homepage: www.elsevier.com/locate/ress

Statistical modelling of railway track geometry degradation

using Hierarchical Bayesian models
A.R. Andrade a,n, P.F. Teixeira b
a
Institute of Railway Research, University of Huddersfield, Queensgate, Huddersfield, Yorkshire HD1 3DH, UK
b
CESUR, CEris, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, 1049-001 Lisboa, Portugal

art ic l e i nf o a b s t r a c t

Article history: Railway maintenance planners require a predictive model that can assess the railway track geometry
Received 4 June 2014 degradation. The present paper uses a Hierarchical Bayesian model as a tool to model the main two
Received in revised form quality indicators related to railway track geometry degradation: the standard deviation of longitudinal
2 May 2015
level defects and the standard deviation of horizontal alignment defects. Hierarchical Bayesian Models
Accepted 15 May 2015
(HBM) are flexible statistical models that allow specifying different spatially correlated components
Available online 30 May 2015
between consecutive track sections, namely for the deterioration rates and the initial qualities
Keywords: parameters. HBM are developed for both quality indicators, conducting an extensive comparison
Statistical model between candidate models and a sensitivity analysis on prior distributions. HBM is applied to provide
Railway track geometry
an overall assessment of the degradation of railway track geometry, for the main Portuguese railway line
Hierarchical Bayesian model
Lisbon–Oporto.
Railway infrastructure
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction This paper is structured in the following way: this first section
briefly introduces the motivation and idea behind it, Section 2
For railway Infrastructure Managers, predicting railway track provides some background on railway track geometry degrada-
geometry degradation is crucial to plan maintenance and renewal tion, focusing on current states of the Art and of the Practice, while
actions associated with it, and has become more and more identifying some current limitations in the statistical modelling of
relevant within their decision support systems. The use of more railway track geometry degradation. Section 3 introduces in a brief
complex predictive models that tackle important aspects of rail- way Hierarchical Bayesian Models (HBM). Then, Section 4 explores
way track geometry degradation, namely the spatial correlations the statistical modelling of railway track geometry degradation
between degradation models' parameters, may enhance decision- with HBM, focusing namely on the model assumptions, on the
making processes related to maintenance and renewal decisions, definition of prior distributions and the derivation of the joint
while preserving parsimonious in statistical modelling. posterior distribution, and the description and comparison of
Bayesian statistical models provide a flexible framework to different HBM specifications. Section 5 applies the HBM to a
combine prior information from past samples or from expert particular track segment of the main Portuguese line, with an
judgment with the new data. Moreover, they allow specifying extensive sensitivity analysis on the influence of prior distribu-
hierarchical probability structures that can capture uncertainty tions, and model comparison. Section 6 provides an overview of
associated with a model parameter. They also provide a learning the railway track geometry degradation of the main Portuguese
mechanism that can update information through time. Therefore, line using the best selected HBM model. Finally, Section 7
Bayesian statistical models seem a promising tool in transport discusses main conclusions and sketches directions for further
infrastructure management, namely in railway infrastructure. The research.
present paper intends to explore Hierarchical Bayesian Models
(HBM) as a predictive tool to assess maintenance needs given a
maintenance and renewal strategy.
2. Statistical modelling of railway track geometry degradation

2.1. State of the Art

n
Corresponding author. Institute of Railway Research, University of Huddersfield,
Queensgate, Huddersfield,Yorkshire, HD1 3DH, UK. Research on the topic of statistical modelling of railway track
E-mail address: [email protected] (A.R. Andrade). geometry has benefitted from a number of contributions in the

http://dx.doi.org/10.1016/j.ress.2015.05.009
0951-8320/& 2015 Elsevier Ltd. All rights reserved.
170 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

last three decades. First published works from the 80s focused ii) The rate of deterioration varies drastically from section to
on obtaining quantitative measures that described rail-vehicle section even for apparently identical sections carrying the
performance through two components: i) ride comfort and ii) same traffic;
safety (probability of derailment). Corbin and Fazio [1] called them iii) There is no proved effect on the quality and on the rate of
performance-based track-quality measures, and used simple linear deterioration by the type of traffic or track construction;
models to obtain response modes for a given vehicle to some track iv) The rate of deterioration appears to be a constant parameter
geometry irregularities, and thus computed envelopes of allowed for a section regardless of the quality achieved by the main-
profile (longitudinal level defect) deviations for different speeds tenance machine;
depending on the spatial frequency (cycles per metre) or on its v) Tamping machines improve the quality of a section of track to
inverse (i.e. the wavelength in metres). They used the power a more or less constant value.
spectral density of rail profile (longitudinal level defect) as a
statistical representation of railway track geometry, for ride
comfort and safety purposes. Therefore, note that the second conclusion above emphasises
Nearly at the same time, Hamid and Gross [2] also discussed the importance of modelling degradation at the track section level,
the need to objectively quantify track quality for maintenance whereas the fourth conclusion (together with the second conclu-
planning and ride comfort purposes, analysing rail track geometry sion) makes clear that the inherent uncertainty of the degradation
data collected over a period of one year on approximately 290 model parameters should be assessed with a learning mechanism
miles (467 km) of main-line, investigating statistical dependencies associated with it.
on track geometry defects (measured at 1-ft (0.3048-m) intervals Later on in the 90s, Iyengar and Jaiswal [6] proposed modelling
by the T-6 vehicle) and some indicators (e.g. mean, 99th percen- railway track irregularities as a non-Gaussian model and con-
tile, standard deviation and higher-order moments). Moreover, cluded that in terms of level crossing and peak statistics, the
they developed empirical degradation models through linear proposed non-Gaussian model was consistently better than the
autoregressive techniques that could describe the relationship Gaussian model in predicting the number of upward level cross-
between track quality indexes (defined as the standard deviation ings and peaks. Their non-Gaussian model to statistically repre-
of profile) and physical parameters. For instance, simple linear sent railway track irregularities used a finite series of uncorrelated
regressions were put forward to relate the track quality index with terms: the first term was a Gaussian process and the remaining
the root mean square of vertical acceleration. Nevertheless, as the terms were derived using a Gram–Schmidt procedure. Iyengar and
resulting empirical equations exhibited autoregressive terms, i.e. Jaiswal [7] eventually modelled track irregularities (both Absolute
they included previous values as explaining variables, these Vertical Profile (longitudinal level defects) and unevenness (stan-
expressions proved unreliable to predict track quality for the dard deviations)) as a stationary Gaussian random field so that the
medium- or long-term. Development of rail track degradation classical level-crossing and peak-statistics theory could be
models to predict future track quality indices for maintenance exploited to relate the sample deviation to the highest peak value
planning purposes had a major contribution with Bing and Gross in a simpler way than in their first non-Gaussian model. In a way,
[3], where they used a multiplicative form of model including as Iyengar and Jaiswal opted to simplify their initial approach so that
explaining variables: traffic information (equivalent train speed), they could make probabilistic approaches appealing to practical
track structure, maintenance (e.g. time since surfacing), or even engineers. Nevertheless, as it was discussed by Kumar and
ballast index in order to explain the rate of degradation at two Stathopoulos [8], unevenness data (standard deviations) indicated
consecutive time periods. Another important contribution was put a non-Gaussian character with skewness and kurtoses significantly
forward by Hamid and Yang [4], in which they followed an different from the typical values for Gaussian processes.
approach based on analytically describing typical variations of More recently, other approaches have also tried to capture the
track geometry, distinguishing random waviness, periodic beha- nonlinear characteristics of track quality deterioration [9–11],
viours at joints, and isolated variations, which occur occasionally though they typically would not consider all track geometry
but with regular patterns. In fact, they represented some track defects, and would model instead a quality index. An important
geometry defects as stationary random processes (modelled contribution was given in [12] for the case of a Spanish high speed
through the power spectral density) and others as periodic railway line, identifying the embankment height as a dependent
processes. variable on the density of maintenance works, though the study
Besides these first published works, there were previous investiga- analyses the past tamping actions and vertical accelerations rather
tions carried out in the 70s by the former Office for Research and than the track geometry records themselves. In that sense, that
Experiments (ORE) trying to understand the fundamentals of the work is analysing the ‘outputs’ of maintenance decisions, rather
deterioration mechanism and to control this phenomenon, which than their inputs (i.e. degradation) so that one may assess whether
examined data available from a number of administrations and or not that decision was a good one. Moreover, another relevant
showed that the factors governing the rate of deterioration were not issue in statistical modelling of railway track geometry is how to
obvious. They also showed that unknown factors in the track were the model the tamping/maintenance recuperation, i.e. the impact of
most critical in determining both the average quality and the rate of levelling and tamping operations in the railway track geometry
deterioration. Original tests on the rate of deterioration of track defects and their evolution. In fact, the specialized literature
geometry were carried out by ORE committee D 117. Although the usually does not cover this improvement in great detail and only
results were not very conclusive, track quality on relaying (i.e. the a few references have been found, such as [11] and [13]. A recent
initial quality) was identified as the most important factor. More work by Quiroga and Schnieder [14] has used accumulated
measurements on track geometry were recorded and other main tamping interventions as an explaining variable, showing higher
qualitative conclusions were drawn [5]: variances for higher number of accumulated tamping interven-
tions. Furthermore, a recent work by Vale and Lurdes [16] also
i) After the first initial settlement, both vertical quality and discussed a stochastic model for the geometrical railway track
alignment deteriorate linearly with tonnage (or time) between degradation process, focusing on the standard deviation of long-
maintenance operations; itudinal level defects and not on the standard deviation of
horizontal alignment. Other contributions focusing on different
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 171

prediction techniques used to assess future railway track geometry
IAL – Immediate Action Limit: refers to the value which, if
condition have been proposed, namely using artificial neural exceeded, requires imposing speed restrictions or immediate
networks [17], stochastic state space methods [18] or even Petri correction of track geometry;
net models [19]. A Bayesian approach is explored in [20], but this
IL – Intervention Limit: refers to the value which, if exceeded,
time following a nonparametric specification with a Dirichlet requires corrective maintenance before the immediate action
Process Mixture Model, focusing on the failure of different railway limit is reached;
components rather than specifically on railway track geometry
AL – Alert Limit: refers to the value which, if exceeded, requires
degradation. Finally, a very recent work by Gong et al. [21] has put that track geometry condition is analysed and considered in the
more focus on the deterioration of lateral alignment using vehicle regularly planned maintenance operations.
dynamic simulation and considering an elastic lateral model for
the railway track, while assessing the effect of different factors like
the vehicles, running speed, traffic mixes and different wheel/rail Although the IAL limits are considered normative, providing
contacts. the highest admissible limits to ensure safety and ride comfort;
the IL and the AL limits are purely indicative, reflecting common
practice among most European Infrastructure Managers. They are
2.2. State of the Practice even expressed as a range rather than a single value. In fact, the EN
13848-5 also directs each Infrastructure Manager to select their
According to a best practice guide for optimum track geometry own IL and AL limits according to their inspection and mainte-
durability [22], European Infrastructure Managers tend to trigger their nance systems, which in turn relate to different targets for safety,
preventive tamping actions based on a single indicator: the standard ride quality, lower life-cycle costs and track access availability.
deviation of longitudinal level defects. Nevertheless, the European
standard EN 13848-5 puts forward recommended Alert Limits 2.3. Current limitations in statistical modelling of railway track
for preventive maintenance actions based on two indicators: i) the geometry degradation
standard deviation of longitudinal level defects and also ii) the
standard deviation of horizontal alignment defects. Current statistical approaches tend to focus on track quality
In fact, recent research has discussed the use of these two indexes rather than on the standard deviations of longitudinal
indicators as predictors of other impacts associated with planning level defects (SDLL) and of horizontal alignment defects (SDHA).
maintenance of railway track geometry, namely the corrective/ These track quality indexes are sometimes dependent on the
unplanned maintenance needs and the delays imposed due to maximum permissible speed so they do not refer only to railway
temporary speed restrictions. Both works [23,24] have found that track physical degradation but also to its use. The statistical
not only the standard deviation of longitudinal level defects was a approaches that model the standard deviation of longitudinal
statistically significant predictor, but also the standard deviation of level defect (SDLL) do not consider the standard deviation of
horizontal alignment defects. Therefore, the present paper intends horizontal alignment defects (SDHA), arguing that current practice
to statistically model the evolution of these two quality indicators of maintenance decisions rely solely on the SDLL [22]. Never-
relative to railway track geometry using a HBM. theless, the SDHA indicator seems to play an important role as a
The degradation of railway track geometry is usually quantified by predictor of localised defects and corrective maintenance needs.
seven track geometry defects: the left and right longitudinal level Moreover, current statistical models have overlooked the spa-
defects (LLL and RLL), the left and right horizontal alignment defects tial correlations of the deterioration rates and the initial qualities
(LHA and RHA), the cant defects (C), the gauge deviations (GD) and the for consecutive track sections. In fact, this idea followed from an
track twist (T). These defects are measured through automated initial exploratory work previously conducted in [15], showing
measuring systems, typically integrated in inspection vehicles, and that spatial correlation between deterioration rates and initial
saved as signal data. Signal digital processing techniques are then used quality was statistically significant. In that sense, none of the
to align signals and derive indicators (for each type of defect) that can previous statistical models takes advantage of these spatial corre-
support maintenance and renewal decisions as part of planned lations between deterioration rates and the initial qualities for
maintenance or eventually as unplanned maintenance actions. Many consecutive track sections, in order to improve the current pre-
Infrastructure Managers tend to combine all these indicators into a dictive models. In the present paper, these spatial correlations are
track quality index, which is typically a function of the standard handled using HBM, so that the parameters for the deterioration
deviations of each defect and/or train permissible speed (as reported rates and the initial qualities (i.e. the slope β and the y-intercept α
in [25] or [26]). Nevertheless, the standard deviation for the short in simple linear regression models y ¼ α þ β:x þ ε) can be consid-
wavelength (3–25 m) of longitudinal level defects is still perceived as ered random quantities, and thus, Conditional Autoregressive
the crucial indicator for planned maintenance decisions for many (CAR) probability structures can be assigned to these parameters
European Infrastructure Managers. associated with consecutive track sections. Note that in classical
The use of the standard deviation of the short wavelength (3– statistical approaches, the spatial correlations would have to be
25 m) of longitudinal level defects (SDLL) as the crucial indicator tackled through CAR structures assigned to the random error ε, as
for planned maintenance on railway track geometry degradation the slope β and the y-intercept α are not random, and in fact they
may be attributed in our perspective to two reasons: i) due to the are assumed to be fixed but unknown, and estimated from
simplicity in the empirical expressions describing its evolution the data. And, therefore as a result, HBM is the mathematical
(which depends on the accumulated tonnage, usually in Million statistical method that let us model directly the spatial correlation
Gross Tons (MGT)), and ii) due to the fact that it correlates well between the deterioration rates and initial qualities and not on the
with the vertical force [13,27], which is a proxy or vertical random error.
acceleration felt by the passenger and thus, of ride quality. Our HBM approach models separately the two main indicators
Railway track geometry defects should be within certain limits (SDLL and the SDHA) and adds CAR probability structures to the
according to a given safety standard. The European Standard EN deterioration rates and the initial qualities parameters to handle the
13848-5 [28] provides limits for several indicators for each type of previously overlooked spatial correlations between consecutive tracks.
defect depending on the maximum permissible speed and for As it will be seen later on, this proved to provide a better fit to the data
three main levels: according to the Deviance Information Criterion (DIC).
172 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

3. A brief note on Hierarchical Bayesian Models (HBM) which contain v:

pðv j V\vÞ p pðvj parents½vÞ  ∏ pðwj parents½wÞ
This section briefly explores Hierarchical Bayesian Models w A children½v
(HBM) to predict the evolution of railway track geometry degrada-
Note that OpenBUGS, the most recent development of Win-
tion. Bayesian models are different than classical statistical models
BUGS, allows user-friendly inference for these HBM, especially if
in the fact that they assume parameters as random variables,
the adaptive rejection sampling algorithm from Gilks and Wild
whose uncertainty can be quantified by a prior distribution. This
[33] is needed.
prior distribution pðθÞ is then combined with the traditional
likelihood pðyj θÞ to obtain the posterior distribution of the
parameters of interest. The posterior distribution pðθ j yÞ of the 4. Modelling railway track geometry degradation with HBM
parameters θ given the observed data y can be computed accord-
ing to Bayes' rule as: Having briefly introduced the main concept of HMB in Sec-

     tion 3, let us explore the intricacies of the proposed modelling of

 p yj θ :pðθÞ ' ' '


p θj y ¼ R p yj θ :p θ dθ p p yj θ :p θ railway track geometry degradation using HBM. The associated
DAG for the more complicated models is provided at the end of
this section. We start by discussing the main assumptions on
The specification of the prior distribution constitutes a very statistical modelling of railway track geometry degradation that
important step in any Bayesian model, using for instance a non- our modelling approach will rely on, without forgetting the past
informative (or vague prior), or incorporating preceding known research findings previously discussed in Subsections 2.1 and 2.2.
information using old samples (hopefully under the same bound- The models developed in this section are specified in the same
ary conditions) or from expert judgment techniques. Further way for the main two quality indicators related to railway track
details on Bayesian statistics can be found in [29] and [30], or geometry degradation, i.e. for the standard deviation of long-
for a more practical approach [31]. However, in almost every case itudinal level defects (SDLL) and for the standard deviation of
in real applications, one finds that that the joint posterior horizontal alignment defects (SDHA). In the following, when we


distribution p θ j y has a reasonably high dimension, and integra- refer to the dependent variable as a quality indicator, the reader
tion through numerical methods must rely on Markov Chain should have in mind that the dependent variable can be either the
Monte Carlo (MCMC) methods, which are built in such a way that SDLL or the SDHA.
their stationary distribution is the desired posterior distribution.
HBM are a special case of Bayesian models. They benefit from a 4.1. Model assumptions
major property: they can be factorised through Directed Acyclic
Graphs (DAG) in a convenient way. This not only enables arbi-

Let us start by assuming that a quality indicator ysvkl at
trarily complex full probability models to be specified based on the inspection l for track section k from track segment v from area s
simple local components, but it also makes the identification of is normally distributed with mean msvkl and variance σ 2s , i.e.
full conditional distributions straightforward. Moreover, this hier- ysvkl  Nðmsvkl ; σ 2s Þ. Fig. 1 should support the reader to understand
archical construction is particularly useful, because once the full better meaning of indices s, v, and k for a typical double track line.
conditional distributions are identified/available, one can sequen- Then, some assumptions on its mean value result from a combina-
tially sample from them using the Gibbs sampler, and sample from tion of factors:
the posterior distribution of interest, or using any more general
sampling method (e.g. adaptive rejection sampling, Metropolis– 1) A constant linear evolution with accumulated tonnage (T svkl –
Hastings algorithm). accumulated tonnage since last tamping or renewal operation),
A DAG is a graph that is directed because each node is linked given by the deterioration rate – β svk (slope of a linear
through an arrow (i.e. with a defined direction) and it is acyclic regression), which assumes different values for each track
because no cycles are formed with the arrows (i.e. it is not possible section k in track segment v for area s.
to return to a given node once you left it). Parent nodes from a 2) An initial value for the quality indicator, given by the initial
given node (i.e. parents[v]) are all nodes that are connected quality – αsvk (y-intercept of a linear regression), which
through arrows from them to that node ‘v’; whereas children assumes different values for each track section k in track
nodes from a given node (i.e. children[v]) are all nodes that are segment v for area s.
connected through arrows from that node ‘v’ to them.
According to Lunn et al. [32] on the WinBUGS use of DAG
These first two assumptions presume a linear evolution of the
factorisation, let V denotes the set of all nodes ðvÞ in a DAG for a
quality indicator with accumulated tonnage since last mainte-
given HBM, it can be shown as:
nance or renewal action and an initial quality, allowing different


 values for each track section.
p θ j y p p θ; y ¼ pðV Þ ¼ ∏ pðvj parents½vÞ
v A V

3) A disturbance effect δsv of the quality indicator after each
Let V \ v denotes ‘all elements of V except v’. The full conditional tamping operation, i.e. the quality indicator does not recover to
pðv j V\vÞ is then proportional to the product of terms in pðV Þ its initial value αsvk , but it is affected by a rate 1 þ δsv , given by
δsv , which assumes different values for each track segment v for

Fig. 1. A typical double track line with indices from area s, segment v and track section k.
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 173

area s. Therefore, note that
at eachN new tamping cycle the of two additive components: an average value αsv and a spatially
initial quality would be αsvk 1 þ δsv svkl , in which Nsvkl is the correlated term εαsvk so that αsvk ¼ αsv þ εαsvk . For εα svk , we then ass-
number of tamping operations conducted since last renewal. ign a CAR probability structure such as εαsvk  εαsvð  kÞ ; σ 2αsv 
P
4) Distinction between renewed track sections ðRsvkl ¼ 1Þ and non- N εαsvk ; σ αsv =nsvk , in which εαsvk ¼
2
εαsvj =nsvk , N svk denotes
j A N svk
renewed track sections ðRsvkl ¼ 0Þ is assured through the the set of track sections which are considered neighbours to track
separation of the initial quality and the deterioration rates for section k (in segment v in area s), and nsvk is the number of
a non-renewed track section k from segment v for area
s – α 0
 svk neighbours of track section k (in segment v in area s), and finally
0
and βsvk respectively; whereas the disturbance effect δsv is εαsvð  kÞ is the vector with all components εαsvk from segment v in
considered the same for renewed and non-renewed track area s except the component related to track section k.
sections. Two well-known CAR structures were tested: the first-order
random walk ðRW ð1ÞÞ and the second-order random walk ðRW ð2ÞÞ
0
These next two assumptions provide an exponential effect of as hierarchical structures for parameters α; β; α0 ; β .
the number of tamping operations conducted since last renewal The first-order random walk ðRW ð1ÞÞ is defined by considering
and different parameters for renewed and non-renewed track as neighbour structure nsvk ¼ 1 for k ¼ 1 and k ¼ K s , and nsvk ¼ 2 for
sections. k ¼ 2; …; K s  1, and εαsvk ¼ εαsvðk þ 1Þ for k ¼ 1, εαsvk ¼ εαsvðk  1Þ þ εαsvðk þ
These four assumptions can be compiled in a single mathema- 1ÞÞ=2 for k ¼ 2; …; K s  1, and εαsvk ¼ εαsvðk  1Þ for k ¼ K s .
tical expression for the mean msvkl of the quality indicator ysvkl as: For the second-order random walk ðRW ð2ÞÞ these expressions
h
N i complicate a little more but they can be derived from the
msvkl ¼ αsvk 1 þ δsv svkl þ βsvk T svkl U Rsvkl following equivalence, where the symbol n denotes the different
h i 0

N 0 possible parameters α; β ; α0 ; β :
þ α0svk 1 þ δsv svkl þ β svk T svkl U ð1  Rsvkl Þ 
εnsvk  RWð2Þ 3 εnsvk εnsv  k ;
0 8  
In this expression, we should regard α; β; α0 ; β and δ as >
> N 2 εnsvðk þ 1Þ  εnsvðk þ 2Þ ; σ 2n ; k ¼ 1
>
>
parameters, to which there should be assigned a hierarchical >
> h i 
>
>
>
> N 2 εnsvðk  1Þ þ 4 εnsvðk þ 1Þ  εnsvðk þ 2Þ =5; σ 2n =5 ; k ¼ 2
probability structure, whereas N; T and R should be regarded as >
>
< h i 
known/explaining variables. Fig. 2 provides a graphical represen- σn  N 
2 εnsvðk  2Þ þ 4 εnsvðk  1Þ þ 4 εnsvðk þ 1Þ  εnsvðk þ 2Þ =6; σ 2n =6 ; k ¼ 3; …; K s  2
>
> h i 
>
>
tation of the intended behaviour of the mean of the quality >
> N  εnsvðk  2Þ þ 4 εnsvðk  1Þ þ εnsvðk þ 1Þ =5; σ 2n =5 ; k ¼ K s  1
>
>
>
>  
indicator expressed by the equation above (msvkl ). >
>
: N  εnsvðk  2Þ þ 2 εnsvðk  1Þ ; σ 2n ; k ¼ K s

4.2. Introduction of spatial correlations through CAR probability

structures 4.3. Definition of prior distributions

In this subsection, we discussed two CAR probability structures After assigning a CAR probability structure to tackle spatial
to handle the spatial correlation between the deterioration rates correlations in a parsimonious way, we define the prior distribu-
and the initial qualities of consecutive track sections. tions for all other parameters. Therefore, for the parameter δsv , i.e.
In order to be parsimonious in modelling, i.e. trying to keep the the disturbance effect of the initial quality after each tamping
number of parameters to a minimum, a good strategy is consider- operation, we define a typical probability structure expressing


ing (Gaussian) Conditional Autoregressive (CAR) terms. Besag [34] vague information on that parameter, i.e. δsv  N 0; σ 2δ . More-
showed in 1974 that conditional probability structures could deal over, for each variance component in each hierarchical structure,
with spatial interactions in the more complex structures. The CAR we finalise by assigning inverse gamma distributions to each
model is used to model the spatial dependencies in deterioration component, i.e. σ 2s  IGðc0 ; d0 Þ, σ 2δs  IGðc1 ; d1 Þ, σ 2αsv  IGðc2 ;
rates and initial qualities for consecutive track sections. For d2 Þ, σ 2β  IGðc3 ; d3 Þ, σ 2  IGðc4 ; d4 Þ and σ 2  IGðc5 ; d5 Þ,
sv α'sv β'sv
example, take the initial quality αsvk for a given track section k where IGðc; dÞ denotes an inverse gamma distribution with shape
in segment v in area s, we will consider that αsvk is a combination parameter c and scale parameter d, whose density is proportional

Fig. 2. Mean behaviour of the quality indicator for a given track section k in segment v in area s.
174 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183


to x  ðc þ 1Þ exp  dx , x 4 0. For all models (M1, M2, M3 and M4), are different for each area s. Of course, a more parsimonious
inverse gamma priors IGð0:5; 0:0005Þ were assigned for each modelling option could have been chosen, but then the simulation
variance parameter (σ 2s , σ 2αsv , σ 2β , σ 2 , σ 2 and σ 2δs ). Note that could not be conducted separately for each track segment s (which
sv α'sv β'sv
flat priors are improper distributions, i.e. do not integrate to one, favours for instance the possibility of applying parallel processing
but assume a constant value everywhere, attempting to describe techniques in the future).
vague or no prior information on that parameter.
As a side note, it is important to mention that the use of inverse 4.4. Joint posterior distribution
gamma prior distributions for the variance terms is not free of
criticism and has been discussed by Gelman [35], who recom- To derive the joint posterior distribution, prior independence is
mended starting with a non-informative uniform prior density, or assumed amongst the model parameters so that the joint posterior
when more prior information is desired, working within the half-t density is then proportional to:
family of prior distributions, which are more flexible and have (   !)
better behaviour near 0, compared to the inverse-gamma family, S Vs K s Ls 1 1 ysvkl  msvkl 2
∏ ∏ ∏ ∏ exp 
which are very sensitive in data sets where the variance terms s ¼ 1v ¼ 1k ¼ 1l ¼ 1 σ s 2 σs
may assume low values of σ 2 . However, the choice of assigning ( 
c þ 1  )
S 1 0
1
inverse gamma distributions ‘is an attempt at non-informativeness U ∏ exp  2 d0
within the conditional conjugate family’ [35], which mainly s¼1 σs2 σs
(  2 ! )
translates into full conditional posterior distributions for each S Vs 1 1 δsv
variance component within the same distributional family, i.e. U ∏ ∏ exp 
s ¼ 1v ¼ 1 σ δ s 2 σ δs
also inverse gamma distributions, as later seen in the Appendix for 8 !c 1 þ 1 !9
i), iii), v), vii), ix) and xi). This choice is not only attractive for S < 1 1 =
pedagogical purposes, but it is also a common choice in many U ∏ exp  2 d1
s ¼ 1: σ δ σδ ;
2
BUGS software applications. s s
(pffiffiffiffiffiffiffiffi   !)
Fig. 3 provides a DAG representing the proposed HBM, focusing S Vs Ks nsvk 1 εαsvk  εαsvk pffiffiffiffiffiffiffiffi 2
in 3 models explored later on: models M2, M3 and M4. Model M1 U ∏ ∏ ∏ exp  nsvk
s ¼ 1v ¼ 1k ¼ 1 σ αsv 2 σ αsv
o without the eight upper nodes (σ n and
2
has a simple n representation 8
' ! c2 þ 1 !9
εnsvk for n ¼ α; β; α ; β' ). Stochastic nodes are represented in S Vs < 1 1 =
ovals with solid line, whereas constants are in rectangular boxes. U ∏ ∏ exp  d
s ¼ 1v ¼ 1: σ αsv σ 2αsv 2
2 ;
Hyperparameters (c, d) of the variance components are not 8
pffiffiffiffiffiffiffiffi
0 ! 2 19
represented for clarity and ovals with dashed contours are S Vs Ks < n 1 ε  εβsvk pffiffiffiffiffiffiffiffi =
β
logically dependent on their parent nodes and are not stochastic. U ∏ ∏ ∏ svk
exp@  svk
nsvk A
s ¼ 1v ¼ 1k ¼ 1 : σ βsv 2 σ βsv ;
Note that the definition of the proposed HBM with completely
8 !c 3 þ 1 ! 9
independent DAG for each segment s facilitates simulation in a S Vs < =
1 1
separate manner for each segment s. This is guaranteed by U ∏ ∏ exp  d
s ¼ 1v ¼ 1: σ β σβ 3
2 2 ;
defining hyperparameters – σ 2s , σ 2δs , σ 2αsv , σ 2β , σ 2 and σ 2 that
sv α'sv β'sv sv sv

Fig. 3. Directed Acyclic Graph (DAG) for the proposed hierarchical Bayesian models M2, M3 and M4.
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 175

8 0 ! 19
S Vs Ks<pffiffiffiffiffiffiffiffi
nsvk 1 εα'svk  εα'svk pffiffiffiffiffiffiffiffi 2 = 5. HBM application to a particular area of the main Portuguese
U ∏ ∏ ∏ exp@  nsvk A line
s ¼ 1v ¼ 1k ¼ 1: σ α'sv 2 σ α'sv ;
80 1c 4 þ 1 0 19 This section will analyse some detailed results from the
S Vs< 1 1 =
U ∏ ∏ @ 2 A exp@  2 d4 A application of the proposed HBM on a particular area/sample from
s ¼ 1v ¼ 1: σ α' σ α' ; the historical data from the main Portuguese line (Lisbon–Oporto).
sv sv
8 0 !2 19
pffiffiffiffiffiffiffiffi
Ks < n ε  ε =
S Vs 1 β ' β ' pffiffiffiffiffiffiffiffi 5.1. A brief description of the sample
U ∏ ∏ ∏ svk
exp@  svk svk
nsvk A
s ¼ 1v ¼ 1k ¼ 1: σ β'sv 2 σ β'sv ;
80 This historical data mainly refers to: i) the inspection records
1c 5 þ 1 0 19
S Vs< 1 1 = from the EM 120 vehicle to get the standard deviation of long-
U ∏ ∏ @ 2 A exp@  2 d5 A itudinal level defects (SDLL) and the standard deviation of hor-
s ¼ 1v ¼ 1: σ σ β' ;
β'sv sv izontal alignment defects (SDHA) relative to 200-m long track
S Vs
0
sections, ii) the operation records to get the accumulated tonnage
U ∏ ∏ P ½αsv  U P β sv U P α0sv UP βsv ðT svkl Þ, and finally to iii) the maintenance records to get the past
s ¼ 1v ¼ 1
maintenance and renewal actions ðN svkl ; Rsvkl Þ. As the main con-
in which: tribution of this paper is the HBM itself rather than a comprehen-
h
N i sive analysis of the experience from the Lisbon–Oporto line
msvkl ¼ αsvk 1 þ δsv svkl þ βsvk T svkl URsvkl regarding inspection, operation and maintenance, we decided to
h
N i
0 focus on a particular area in this section, though in Section 6 an
þ α0svk 1 þ δsv svkl þ βsvk T svkl U ð1  Rsvkl Þ
overview of the all areas of the Lisbon–Oporto line is provided as
αsvk ¼ αsv þ εαsvk ; α0svk ¼ α0sv þ εα'svk ; βsvk an additional application of the present HBM. This first tested area
0 0 mainly involves a double-track ðV s ¼ 2Þ area of about 15 km
¼ βsv þ εβsvk ; and β sv ¼ β sv þ εβ'
svk ðK s ¼ 74Þ, from a total of 36 inspections ðLs ¼ 36Þ from February
2001 up to October 2009.
In order to ensure that the CAR model structures are identifi-
able, we follow the typical constraint suggested by Besag and
P 5.2. MCMC simulation details
Kooperberg [36] that is to impose that the εαsvk ¼ 0, and use a
flat prior for the constant αsv on the wholek real line. Note that
0 MCMC samples were of size 20,000, taking every tenth itera-
these CAR probability structures are also adopted for β, α0 and β
tion (thin ¼ 10) of the simulated sequence, after 10,000 iterations
parameters.
of burn-in period, for the four models. Initial values were set: for
As the joint posterior is rather complex, we present the full
the variance terms σ 2s , σ 2αsv , σ 2β , σ 2 , σ 2 and σ 2δs equal to 10
conditional posterior distribution in the Appendix, so that a Gibbs sv α'sv β'sv
(which is equivalent to precision (1=σ 2 ) equal to 0.1), for the
sampling strategy can iteratively draw for each parameter and use
spatially correlated terms εαsvk , εβsvk , εα' and εβ0svk equal to 0, and
them as current values for each conditional posterior distribution. svk 0
finally, for each parameter αsv , β sv , α0sv , βsv and δsv equal to 0. To
have a reasonable confidence on MCMC convergence, we ran an
4.5. Different model specifications and comparison alternative MCMC simulation for every model using different
initial values, and found very similar results for posterior esti-
For each quality indicator (SDLL and SDHA), mainly 4 models mates. We also used the diagnostic convergence tests from BOA
(M1, M2, M3 and M4) are explored whose differences are related program [37], which did not reveal any evidence against conver-
0
with hierarchical structures for α, β , α' and β : gence for all MCMC outputs.

 Model M1 assigns simple flat priors to each parameter 5.3. Discussion of model results
αsv ; βsv ; α0sv ; β0sv ;
 Model M2 assigns a flat prior to each parameter Table 1 provides a comparison of the four models based on the
αsv ; βsv ; α0sv ; β0sv and a normal prior with mean equal to zero DIC for each quality indicator (SDLL and SDHA). We can spot that
and variance σ 2αsv , σ 2β , σ 2 , σ 2 for each parameter model M3 with first-order random walk CAR structure has the
sv α'sv β'sv
εαsvk ; εβsvk ; εα0svk ; εβ0svk , respectively. lowest DIC value among the four models explored for the SDLL and
 Model M3 assigns a flat prior to each parameter for the SDHA. However, note that for each quality indicator (SDLL
αsv ; βsv ; α0sv ; β0sv and a first-order random walk for each para- and SDHA), the DIC for model M4, which has second-order random
meter εαsvk ; εβsvk ; εα0svk ; εβ0svk ; walk CAR structures, is not so distant from the DIC value for
 Model M4 assigns a flat prior to each parameter model M3.
αsv ; βsv ; α0sv ; β0sv and a second-order random walk to each Based on the MCMC simulations for each quality indicator and
parameter εαsvk ; εβsvk ; εα' ; εβ' . for each model, Table 2 provides estimates of the posterior
svk svk

parameters for the example track segment, for both dependent

In terms of model comparison, HBM tend to rely on a standard variables (SDLL and SDHA).
measure proposed by Spiegelhalter et al. [38] for comparing Regarding Table 2 for the dependent variable SDLL, note that
models of any degree of complexity. In fact, the Deviance Informa- both the parameters related to renewal track sections, i.e. initial
tion Criterion (DIC), whose assessment is straightforward using quality αsv and deterioration rate β sv , present very close values,
MCMC inference methods, balances two components: the especially for models M2, M3 and M4. This is also true for the
0
expected posterior deviance (i.e. the goodness of fit) and the parameters related to non-renewed track sections (α0sv and β sv ),
effective number of parameters

 (i.e.
 the complexity

 of the model). though for model M1 values differ very much particularly
for the
It is defined as: DIC ¼ 2D θ  D θ , where D θ is the posterior deterioration rate. Regarding the disturbance effect δsv , the
mean of deviance and θ is the posterior mean of the model standard deviations (s.d.) associated with its estimates are high
parameters. In practical terms, a lower DIC value means a better compared to its mean values for models M2, M3 and M4. More-
model in terms of relative comparison. over, note that all values for the disturbance effect seem to be very
176 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

close to zero. Besides, note that the deterioration rates for non- renewed track sections are (on average) at least 60% higher than
renewed track sections are (on average) at least four times higher the deterioration rates for renewed track sections for models M1,
than the deterioration rates for renewed track sections for models M2, M3 and M4; whereas the initial quality for non-renewed track
M2, M3 and M4, and only 40% higher for model M1; whereas the sections seems to be 3 times higher than for the renewed track
initial quality for non-renewed track sections seems to be at least sections (on average) for models M1, M2, M3 and M4.
four times higher than for the renewed track sections (on average) Moreover, comparing between SDLL and SDHA, one finds that,
for models M2, M3 and M4, and five times higher for model M1. on average, deterioration rates for renewed and non-renewed rail
Regarding Table 2 for the dependent variable SDHA, note that track sections are lower for the horizontal alignment than for the
both the parameters related to renewed track sections, i.e. initial longitudinal level defects, whereas the initial qualities are higher
quality αsv and deterioration rate β sv , also present very close for renewed rail track sections and lower for non-renewed rail
values, especially for models M2, M3 and M4. Again, this is also track sections. Regarding the disturbance effect ðδsv Þ, it assumes
true for the parameters related to non-renewed track sections higher values for the model relative to the standard deviation of
(α0sv and β'sv ), though for model M1 values differ a little bit more, horizontal alignment defects (SDHA) than for the model relative to
particularly for the initial quality. Regarding the disturbance effect the standard deviation of longitudinal level defects (SDLL).


δsv , the standard deviations (s.d.) associated with its estimates
are relatively high compared to its mean values for models M2, M3
and M4. Moreover, all values for the disturbance effect seem to be
close to zero. Besides, note that the deterioration rates for non- 5.4. A sensitivity analysis on the influence of prior distributions

To analyse the influence of priors' specifications, a sensitivity

Table 1
Comparison of the DIC for different models M1–M4 for both quality indicators: the analysis is conducted in this subsection, assuming different inverse
standard deviation of longitudinal level defects (SDLL) and the standard deviation of gamma priors IGðc; dÞ. For this purpose, we followed the same
horizontal alignment defects (SDHA). experimental design from a similar investigation contained in Silva
et al. [39] on Bayesian Hierarchical models to analyse revascular-
Models defined fromnsvk ¼ αsvk ; βsvk ; α0svk ; β0svk DIC
ization odds, using only inverse gamma distributions but different
SDLL SDHA combinations: ðc; dÞ ¼ ð0:5; 0:0005Þ, ð0:001; 0:001Þ, ð0:01; 0:01Þ,
ð0:1; 0:1Þ, ð2; 0:001Þ, ð0:2; 0:0004Þ and ð10; 0:25Þ, which are
M1: nsvk ¼ nsv 7192 8324 denoted by A, B, C, D, E, F and G, respectively. As it is noted in
M2: nsvk ¼ nsv þ εnsvk ; εnsvk  Nð0; σ 2n Þ  2794 3857
[39], priors C and D (with c ¼ d) are variants of prior B with a
M3: nsvk ¼ nsv þ εnsvk ; εnsvk  RWð1Þ  3274 3406
M4: nsvk ¼ nsv þ εnsvk ; εnsvk  RWð2Þ  3234 3481
larger associated dispersion (compared to B) in increasing order,
whereas
prior E and F correspond to distributions with the same
Note: The symbol n intends to represent the parameters α; β; α0 ; β0 . mode d=ðc þ 1Þ as prior A, but with lower (E) and larger

Table 2
Estimates of the posterior parameters for an example track segment (s ¼1, v¼1) for both quality indicators: SDLL and SDHA.

Quality indicator Models αsv (mm) βsv (mm/100MGT) α0sv (mm) β0sv (mm/100MGT) δsv

Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d.

SDLL M1 0.3407 0.020 1.203 0.143 1.787 0.038 1.680 0.329  0.1016 0.019
M2 0.3065 0.024 1.469 0.133 1.379 0.067 6.406 0.488  0.0035 0.008
M3 0.3102 0.012 1.460 0.081 1.381 0.015 6.247 0.179  0.0015 0.009
M4 0.3034 0.013 1.499 0.090 1.375 0.016 6.169 0.170 0.0055 0.009
SDHA M1 0.4353 0.029 0.542 0.191 1.524 0.041 1.123 0.375 0.0172 0.024
M2 0.4243 0.020 0.602 0.133 1.345 0.074 1.032 0.443 0.0262 0.016
M3 0.4263 0.019 0.606 0.132 1.342 0.255 1.069 0.273 0.0279 0.015
M4 0.4291 0.019 0.595 0.128 1.343 0.025 0.959 0.272 0.0304 0.016

Table 3
Estimates of the spatial variance components based on model M3 with different inverse gamma prior for both quality indicators (SDLL and SDHA) for the particular test
segment (s ¼1, v¼1).

QI Prior σ 2s (mm2) σ 2αsv (mm2) σ 2α0sv (mm2) σ 2βsv ((mm/100 MGT)2) σ 2β0 ((mm/100 MGT)2) σ 2δs
sv

Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d.

SDLL A 0.0323 0.00141 0.0531 0.00816 0.3478 0.04848 1.4520 0.2379 21.950 3.609 0.0304 0.105
B 0.0318 0.00069 0.0522 0.00818 0.3679 0.04834 1.6220 0.2489 23.100 3.758 0.1732 0.787
C 0.0327 0.00199 0.0454 0.00729 0.3551 0.04917 3.3140 0.4299 24.610 4.025 0.2596 4.513
D 0.0312 0.00067 0.0429 0.00723 0.3618 0.04591 16.380 1.9770 42.070 5.544 0.7139 6.716
E 0.0326 0.00241 0.0506 0.00833 0.3522 0.04694 1.5470 0.2432 21.610 3.565 0.0121 0.013
F 0.0318 0.00073 0.0535 0.00856 0.3535 0.04529 1.4480 0.2358 22.380 3.592 0.0947 0.725
G 0.0316 0.00067 0.0360 0.00569 0.3048 0.03832 32.700 3.6110 57.510 6.801 0.0274 0.010
SDHA A 0.1114 0.00238 0.0052 0.00259 0.3140 0.03985 0.4275 0.0923 7.245 1.757 0.0023 0.005
B 0.1119 0.00230 0.0040 0.00209 0.2884 0.04409 0.6264 0.1122 8.026 1.842 0.1276 1.406
C 0.1108 0.00227 0.0038 0.00133 0.3180 0.04218 2.3540 0.3045 11.50 2.069 0.1262 1.236
D 0.1104 0.00227 0.0097 0.00232 0.3155 0.04316 15.220 1.8170 31.50 4.277 0.5429 3.340
E 0.1117 0.00233 0.0028 0.00156 0.3037 0.04087 0.6045 0.1056 6.914 1.609 0.0010 0.001
F 0.1115 0.00238 0.0063 0.00274 0.3065 0.04239 0.3762 0.0845 7.481 1.762 0.0108 0.057
G 0.1104 0.00236 0.0114 0.00206 0.2650 0.04256 31.790 3.5250 48.940 5.856 0.0269 0.009
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 177

(F) dispersion. Prior G is a less vague prior from this experimental the experimental set used, and model M3 has a lowest DIC for less
design set. vague (more precise or more informative) priors.
Table 3 provides for model M3 and for both quality indicators
SDLL and SDHA some estimates of the variances σ 2s , σ 2αsv , σ 2β , σ 2 ,
sv α'sv
σ 2β' and σ 2δs for different parameters and the sensitivity of these 6. Applying HBM: an overview of the Portuguese main line
sv
estimates using different priors. Note that for both the SDLL and degradation
the SDHA cases, the mean values for σ 2s , σ 2α and σ 2 are not very
α'
sensitive to different inverse gamma priors; whereas σ 2β and σ 2 This section explores the application of the HBM model to all
β'
seem to be very sensitive for priors D and G and σ δ seems to vary
2
the track segments of the Lisbon–Oporto line. In Section 5, the
very much for all priors with associated standard deviation (s.d.) HBM were validated for a particular area with a double-track line,
higher than its mean values, which indicates that priors are and a detailed sensitivity analysis was conducted. In the current
conveying a lot of information to the variance of the disturbance section, the HBM for the SDLL and for the SDHA will be explored for
effect – σ 2δ . all the track segments without the detailed sensitivity analysis
We also conducted a sensitivity analysis on model selection, i.e. previously conducted. Model M3, with a first-order random walk
we analysed if different priors would affect the selection of model CAR structure for the deterioration rates and the initial quality
3 as the best model according to the Deviance Information parameters, was the selected model explored in the following
Criterion (DIC) of both quality indicators. Table 4 explores the analysis, as it exhibited the lowest values in the Deviance Informa-
sensitivity of DIC for different models using different priors. For tion Criterion (DIC) for that particular example.
this analysis, we decided to leave out model M1 because its DIC For the application of the HBM model to all track segments,
value for Prior A was high compared to the others. The last two some adaptations were conducted: i) in non-renewed track seg-
columns of the table present the selected model according to the ments the parameters αsv and βsv , associated with renewed track
lowest DIC for each prior. segments were excluded from the equation to express the mean
Regarding Table 4 on the SDLL, note that model M3 seems to be value (msvkl ) in the HBM, as well as the corresponding CAR
the best model for priors A, B, D, F and G; whereas model M4 structures and ii) in renewed rail track segments (i.e. renewed
0
seems to present a lowest DIC value for priors C and E. Therefore, before 2001), the parameters α0sv and βsv were excluded.
the selection between models M3 and M4 is sensitive to the prior The Lisbon–Oporto railway line has a total length of 336.2 km
used for the variance components for the quality indicator SDLL. and links the most populated cities in Portugal by using passenger
Apart from prior C, for which DIC values are almost the same, it trains running at a maximum speed of 220 km/h and freight trains
seems that model M4 has a lowest DIC for more vague priors in running at 80 km/h. The maximum permissible load of trains is
the experimental set used, and model M3 has a lowest DIC for less equal to 22.5 t and the renewal works have been conducted since
vague (more precise or more informative) priors. Regarding 1996. Approximately 2/3 of the line was renewed, whereas the
Table 4 on the SDHA, model M3 seems to be the best model for remaining 1/3 will be renewed in the next years. These renewal
priors A, B, E, F and G; whereas model M4 seems to present a operations included a thorough improvement of the track-bed
lowest DIC value for priors C and D. Similar to what happen for the bearing capacity and a complete renewal of the track super-
standard deviation of the longitudinal level defects, the selection structure incorporating monoblock concrete sleepers (with a
between models M3 and M4 is sensitive to the prior used for the length of 2.60 m and a seating surface of 3125 cm2 per rail seat)
variance components for the horizontal alignment case. Again, it spaced at 600 mm, rail UIC 60 and Vossloh fastening system with
seems that model M4 has a lowest DIC for more vague priors in plastic railpads ZW 687 (vertical stiffness 450 kN/mm). The
Lisbon–Oporto line is a double-track line in its majority, and
between Braço de Prata (4.0 km) and Alverca (21.8 km), it is a
multiple track segment, more specifically a four-track segment:
Table 4
two tracks (Rapid and Slow Lines – RL and SL) in the Oporto
Sensitivity analysis of the model selection based on DIC for both quality indicators
SDLL and SDHA and for different inverse gamma priors. direction (Northwards) and two tracks (RL and SL) in the Lisbon
direction (Southwards). This multiple track configuration is parti-
Prior Models DIC Selected model cularly useful as it allows faster trains to overtake slower trains.
For clarity, only the two Rapid tracks (RL) of the multiple track
SDLL SDHA SDLL SDHA
segments between Braço de Prata (4.0 km) and Alverca (21.8 km)
A M1 7192 8324 M3 M3 are represented in the following figures.
M2  2794 3857 In terms of maintenance and renewal actions performed in the
M3  3274 3406 analysed period, i.e. between 2001 and 2009, Fig. 4 provides an
M4  3234 3481
overview of the annual number of track sections which benefited
B M2  2796 3857 M3 M3
M3  3337 3382 from planned maintenance/tamping actions (N) and the ratio of
M4  3230 3520 renewed tracks sections (R) from 2001 to 2009. This ratio
C M2  2790 3851 M4 M4 quantifies the number of renewed track sections over the total
M3  3106 3429 number of track sections; or in other terms, the number of
M4  3107 3398
D M2  2775 3856 M3 M4
renewed kilometres over the line total length. Note that the ratio
M3  3275 3449 of renewed track sections has increased from 0.33 in 2001 (i.e.
M4  3154 3431 approximately 1/3 of total length) to 0.64 in 2009 (i.e. approxi-
E M2  2794 3859 M4 M3 mately 2/3 of total length) that has been renewed since 1996.
M3  3083 3467
In terms of MCMC simulation details, the initial values, burn-in
M4  3217 3552
F M2  2795 3865 M3 M3 period, and number of samples were set equal to the MCMC
M3  3232 3400 simulation runs for the exemplifying area in Section 5, i.e. MCMC
M4  3189 3556 samples were of size 20,000, taking every tenth iteration
G M2  2779 3853 M3 M3 (thin ¼10) of the simulated sequence, after 10,000 iterations of
M3  3288 3209
M4  3214 3266
burn-in period, with the initial values set for the variance terms
σ 2s , σ 2αsv , σ 2α' , σ 2βsv , σ 2β' and σ 2δs equal to 10 (which is equivalent to
sv sv
178 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

Fig. 4. Information on the number of track sections tamped (N) and on the ratio of renewed track sections (R) for the Lisbon–Oporto line between 2001 and 2009.

Fig. 5. Initial qualities of the standard deviation of longitudinal level defects for renewed (α) and non-renewed (α0 ) track segments in Lisbon–Oporto line. (For interpretation
of the references to color in this figure, the reader is referred to the web version of this article)

precision (1=σ 2 ) equal to 0.1), for the spatially correlated terms Both directions are presented in the following figures: the
εαsvk , εβsvk , εα0svk and εβ0svk equal to 0, and finally, for each parameter Oporto direction (towards Oporto) is presented in solid lines for
αsv , βsv , α0sv , β0sv and δsv equal to 0. renewed (in blue) and non-renewed track sections (in red) and the
Figs. 5, 6 and 7 provide the posterior mean for all the Lisbon direction (towards Lisbon) is presented in dashed lines for
degradation parameters for the HBM for the SDLL: for the initial renewed (in blue) and non-renewed track sections (in red).
standard deviations for renewed track sections α and for non- Moreover, some track segments do not exhibit values for the
renewed track sections α0 for both directions (in Fig. 5), for the renewed or non-renewed track segments. This happens because
deterioration rates for renewed track sections β and for non- some track segments, for the analysed period between 2001 and
0
renewed track sections β for both directions (in Fig. 6), and for the 2009, were already renewed or still need renewal. Track sections
disturbance effect due to tamping δ for both directions (in Fig. 7). which present both parameters for renewed and non-renewed
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 179

Fig. 6. Deterioration rates of the standard deviation of longitudinal level defects for renewed (β) and non-renewed (β0 ) track segments in Lisbon–Oporto line. (For
interpretation of the references to color in this figure, the reader is referred to the web version of this article)

Fig. 7. Disturbance effect (δ) of the initial quality after each tamping operation for the standard deviation of longitudinal level defects. (For interpretation of the references to
color in this figure, the reader is referred to the web version of this article)

track sections are the segment that benefitted from a renewal renewed track segments also exhibit the higher values compared to
action in that analysed period. the other track segments. This might indicate that the data associated
Fig. 5 provides the posterior mean for the initial quality with this particular track segment is not reliable. Fig. 6 exhibits the
0
parameters of the Hierarchical Bayesian model for the standard posterior means for the degradation parameters βsv and β sv , respec-
deviation of longitudinal level defects, i.e. for the renewed track tively the deterioration rates for renewed and non-renewed track
segments (αsv ) and for the non-renewed track segments (α0sv ). segments. Again the two directions of the Lisbon–Oporto line are
Non-renewed track segments represented in red tend to exhibit incorporated in the same figure: the Oporto direction in solid lines,
larger values than renewed track segments represented in blue. For and the Lisbon direction in dashed lines. Non-renewed track segments
the first track segment between Lisbon – Sta. Apolónia (0.0 km) and (marked in red) exhibit higher deterioration rates than renewed track
Bifurcação Xabregas (1.6 km), non-renewed track sections exhibit segments (marked in blue) for both directions. Note that some
extremely high values compared to the other track segments, whereas segments do not exhibit values for the renewed or non-renewed track
180 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

segments. For instance, the track segments comprehended between Similar to previous Figs. 5–7, following Figs. 8–10 provide the
locations Pombal (169.6 km) and Pampilhosa (231.2 km), only the posterior mean for all the degradation parameters, but for the
0
parameter βsv for non-renewed track sections is represented for the HBM for the SDHA: for the initial standard deviations for renewed
Oporto and the Lisbon directions, because no renewed track sections track sections α and for non-renewed track sections α0 for both
are within those locations for the analysed time period and thus, no directions (in Fig. 8), for the deterioration rates for renewed track
value for β sv is represented. sections β and for non-renewed track sections β ' for both direc-
Fig. 7 provides values for the disturbance effect (δ) of the initial tions (in Fig. 9), and for the disturbance effect due to tamping δ for
quality after each tamping operation for the SDLL. It exhibits a both directions (in Fig. 10).
maximum value of 0.021 for the track segment between locations Similar to Fig. 5, Fig. 8 exhibits the posterior means of the initial
287.4 km and 290.2 km for the Lisbon direction, and a minimum qualities, but for the SDHA, i.e. for the renewed track segments
value of 0.003 for the track segment between locations 106.4 km (αsv ) and for the non-renewed track segments (αsv' ). Again, the
and 114.4 km. parameter referring to non-renewed track segments exhibits a
larger value than the same parameter referring to non-renewed

Fig. 8. Initial qualities of the standard deviation of horizontal alignment defects for renewed (α) and non-renewed (α0 ) track segments in Lisbon–Oporto line.

Fig. 9. Deterioration rates of the standard deviation of horizontal alignment defects for renewed (β) and non-renewed (β0 ) track segments in Lisbon–Oporto line.
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 181

Fig. 10. Disturbance effect (δ) of the initial quality after each tamping operation for the standard deviation of horizontal alignment defects.

track segments. Again the track segment near Lisbon – Sta. (336.2 km). The standard deviations associated with the posterior
Apolónia (0.0 km) exhibits a considerable high value for αsv' in estimates for the degradation parameters are much higher when
the Lisbon direction, which might indicate some reliability pro- compared with other areas.
blems in the model for that particular segment.
Fig. 9 exhibits the posterior means for the deterioration rates
0
for renewed (β sv ) and non-renewed (βsv ) track segments in 7. Conclusions and further research
Lisbon–Oporto line for the SDHA. Deterioration rates for non-
renewed track segments tend to exhibit higher posterior means This paper has discussed statistical modelling of railway track
than for renewed track segments, though this contrast is not as geometry degradation using Hierarchical Bayesian models, in
visible as in Fig. 6 for the longitudinal level defects case. Never- order to predict the evolution of the main quality indicators for
theless, the posterior means for the deterioration rates tend to planned maintenance, namely the standard deviation of long-
exhibit higher variability among track segments for the horizontal itudinal level defects (SDLL) and the standard deviation of hor-
alignment indicator (Fig. 9) than for the longitudinal level case izontal alignment defects (SDHA). Particular attention was given to
(Fig. 6), specially for the parameters associated with non-renewed the need to insert spatial statistical dependencies between model
track segments. parameters for consecutive track sections, namely make them
Finally, Fig. 10 exhibits the posterior means for the disturbance follow random walk priors, capturing the spatial correlation
effect (δ) of the initial quality after each tamping operation, but in between deterioration rates and initial qualities, which proved to
this case for the SDHA. For the horizontal alignment case, in be a better model based on the Deviance Information Criterion
contrast with the longitudinal level case, it exhibits only positive (DIC) comparison. Moreover, we have also provided an inference
values with a considerable larger range. It exhibits a maximum method through the derivation of the full conditional posterior
value of 0.178 for the track segment between Gaia (332.4 km) and distribution and the associated Markov Chain Monte Carlo
Oporto-Campanhã (336.2 km) for the Lisbon direction, and a (MCMC) or Gibbs sampling procedure is provided in the Appendix.
minimum value of 0.001 for the track segment between locations In general terms, the application of the model to a sample of
9.6 km and 13.8 km. railway inspection, operation and maintenance data, showed that
Note that in order to preserve some clarity in the exposition of the HBM exhibit a worse fit of the quality indicator SDHA compared
the previous Figs. 5–10, the corresponding standard deviations for to the quality indicator SDLL, suggesting that the horizontal
0
each degradation parameter (α, α0 , β , β and δ) were excluded from alignment defects seems to be less predictable.
these figures, neither a credible interval was included for that For further research, the present HBM can be extended by
purpose. It is also important to mention that similar to what assigning a typical transportation demand model to the future
happened in the exemplifying area explored in Section 4.3, the tonnage usage T, relaxing the assumption that T is known quantity.
standard deviations for these estimates are relatively low com- Additionally, the HBM can serve as a simulation predictive tool and
pared to the posterior mean of those parameters, indicating the to compare different maintenance and renewal strategies, and
statistically significance of those degradation parameters. Never- ideally find a strategy that minimises life-cycle costs and safety
theless, for the disturbance effect δ, the associated standard impacts, while improving ride comfort and track access availabil-
deviations exhibit for some segments high values compared to ity. We have submitted a paper on that topic elsewhere [40].
the posterior mean for that parameter, especially for the HBM for Moreover, an important unexplored branch regards the use
the SDLL. Another important issue regarding the exploration of the of more complicated correlation structures than the first- and
degradation parameters is related with the reliability of the track second-order random walks. In fact, other spatial formulations,
geometry indicators for the initial segments, i.e. the stations areas using for example power exponential functions to model the
near Lisbon – Sta. Apolónia (0.0 km) and near Oporto-Campanhã decline of correlation depending on the distance between track
182 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183

sections may bring more inside in the spatial correlation between  nsvk  2 1 X
2
εα0svk θ  ε0svk p exp  εα0svk  εα0svk  2 ysvkl  msvkl Þ;
degradation models. These extensions should also be compre- 2σ α 0
2 2σ s l
hended within a multivariate Hierarchical Bayesian model to
predict SDLL and SDHA in a joint model instead of the separate s ¼ 1; …; S; v ¼ 1; …; V s ; k ¼ 1; …; K s ;
HBM for each indicator. Finally, another future direction worth
pursuing would be trying to let the model comparison be  1 1X  2
σ 2α0 θ  σ 2  IG c2 þ K s ; d2 þ n ε  εα'svk Þ;
conducted for each area s, and then select as an overall predictive α' 2 2 k svk α'svk
model the combination of all models selected at a particular s, s ¼ 1; …; S; v ¼ 1; …; V s ;
rather than running model selection for all the areas. In that sense,
we would be letting the overall prediction model select which  nsvk  2 1 X
2
model (i.e. M3 or M4 for example), it would use for each area s. εβ0svk θ  εβ0 p exp  εβ0svk  εβ0svk  2 ysvkl  msvkl Þ ;
svk 2σ β0
2 2σ s l

s ¼ 1; …; S; v ¼ 1; …; V s ; k ¼ 1; …; K s ;
Acknowledgements
 1X  2
σ 2β' θ  σ 2  IG c2 þ K s ; d2 þ
1
nsvk εβ'  εβ' Þ;
We would like to thank the support and collaboration of the β' 2 2 k svk svk

Portuguese Railway Infrastructure Manager, REFER, E.P.E., as well s ¼ 1; …; S; v ¼ 1; …; V s ;

as the support of the Portuguese Foundation for Science and
!
Technology, FCT (Project reference PTDC/SEN-TRA/112975/2009) X
2
and MIT Portugal (FCT PhD Grant SFRH/BD/33785/2009). We are αsv jθ  αsv p exp ysvkl  msvkl U P ½αsv ;
k;l
also grateful to Professor Carlos Paulino for letting the first author
attend his classes on Bayesian Statistics and to Professor Giovani s ¼ 1; …; S; v ¼ 1; …; V s ;
Silva for his suggestions and comments on a previous version of !
 X
2
the manuscript. Needless to say, that any errors or lapses are βsv θ  βsv p exp ysvkl  msvkl U P βsv ;
totally our responsibility. k;l
s ¼ 1; …; S; v ¼ 1; …; V s ;
!
Appendix  X
2
α0sv θ  α0sv p exp ysvkl  msvkl U P α0sv ;
k;l
Let θ be the vector of the model parameters, with elements σ 2s ,
s ¼ 1; …; S; v ¼ 1; …; V s ;
δsv , σ 2δ , εαsvk , σ 2α , εβsvk , σ 2β , εα'svk , σ 2α' , εβ'svk , σ 2β' , αsv , βsv , α0sv and β0sv ,
with s, v and k varying for: s ¼ 1; …; S; v ¼ 1; …; V s ; k ¼ 1; …; K s . !
0 
X
2 0
From the joint posterior, one can derive the full conditional β θ  β0sv p exp
sv ysvkl  msvkl U P βsv ;
posterior distributions (denoted below by jθ  j ), which are given k;l
by: s ¼ 1; …; S; v ¼ 1; …; V s ;
!

2 1 1X
2
σ s θ  σ2s  IG c0 þ V s K s Ls ; d0 þ y  msvkl ; s ¼ 1; …; S;
2 2 v;k;l svkl References

! [1] Corbin JC, Fazio AE. Performance-based track-quality measures and their

 1 2 1 X
2
δsv θ  δsv p exp  2 δsv  2 y msvkl ; application to maintenance-of-way planning. Transp Res Rec: J Transp Res
2σ δ 2σ s k;l svkl Board 1981;802:19–26.
[2] Hamid A, Gross A. Track-quality indices and track-degradation models for
s ¼ 1; …; S; v ¼ 1; …; V s ; maintenance-of-way planning. Transp Res Rec: J Transp Res Board
1981;802:2–8.
[3] Bing AJ, Gross A. Development of railroad track degradation models. Transp
 1 1X 2
σ 2δ θ  σ2δ  IG c1 þ V s ; d1 þ δ Þ; s ¼ 1; …; S; Res Rec: J Transp Res Board 1983;939:27–31.
2 2 v sv [4] Hamid A, Yang T-L. Analytical descriptions of track-geometry variations.
Transp Res Rec: J Transp Res Board 1981;838:19–26.
! [5] Esveld C. Modern Railway Track. 2nd ed., 2001. isbn:90-8004-324-3-3.
 nsvk
2 1 X
2 [6] Iyengar RN, Jaiswal OR. A new model for non-Gaussian random excitations.
εαsvk θ  εαsvk p exp  εα  εα  y m ; Probab Eng Mech 1993;8:281–7.
2σ 2α svk svk
2σ 2s l svkl svkl
[7] Iyengar RN, Jaiswal OR. Random field modelling of railway track irregularities.
J Transp Eng 1995;4:303–8.
s ¼ 1; …; S; v ¼ 1; …; V s ; k ¼ 1; …; K s ; [8] Kumar KS, Stathopoulos T. Discussion of random field modelling of railway
track irregularities 1997; 123: 245.
! [9] Riessberger K. Extending maintenance cycles and ballast life. In: Proceedings
 1 1X
2 of the 7th international heavy haul conference, Brisbane, Australia; 2001.
σ 2α θ  σ2α  IG c2 þ K s ; d2 þ nsvk εαsvk  εαsvk ; [10] Ubalde L, López-Pita A, Teixeira P, Bachiller A, Gallego I. Track deterioration in
2 2 k
high-speed railways: influence of stochastic parameters. In: Proceedings of the
s ¼ 1; …; S; v ¼ 1; …; V s ; 8th international conference railway engineering, London, UK; 2005, p. 54–62.
0-947644-56-3.
! [11] Veit P. Track quality – luxury or necessity? Maintenance & renewal J Railw
 nsvk  2 1 X
2
εβsvk θ  εβsvk p exp 
Tech Rev 2007:8–12.
εβsvk  ε βsvk  y  m ;
2σ 2β 2σ 2s l svkl svkl [12] López-Pita A, Teixeira PF, Casas C, Ubalde L, Robusté F. Evolution of track
geometric quality in high-speed lines: ten years experience of the Madrid-
Seville line. Proc Inst Mech Eng F: J Rail Rapid Transit 2007;221:147–55.
s ¼ 1; …; S; v ¼ 1; …; V s ; k ¼ 1; …; K s ; [13] Lichtberger B. Track compendium: formation, permanent way, maintenance,
economics. 1st ed., 2005. 3-7771-0320-9.
 1X  2 [14] Quiroga LM, Schnieder E, Monte Carlo simulation of railway track geometry
σ 2β θ  σ 2β  IG c2 þ K s ; d2 þ
1
nsvk εβsvk  εβsvk Þ; deterioration and restoration, In: Proceedings of the Institution of Mechanical
2 2 k Engineers, Part O: Journal of Risk and Reliability 2012;226(3):274-282.
[15] Andrade AR, Teixeira PF. Uncertainty in rail track geometry degradation:
s ¼ 1; …; S; v ¼ 1; …; V s Lisbon–Oporto line case study. J Transp Eng 2011;137:193–200.
 A.R. Andrade, P.F. Teixeira / Reliability Engineering and System Safety 142 (2015) 169–183 183

[16] Vale C, Lurdes SM. Stochastic model for the geometrical rail track degradation [28] EN 13848-5:2008. Railway applications – Track – Track geometry quality –
process in the Portuguese railway Northern Line. Reliab Eng Syst Saf Part 5: Geometric quality levels.
2013;116:91–8. [29] Bernardo JM. Bayesian statistics. Updated and abridged version of the Chapter
[17] Guler H. Prediction of railway track geometry deterioration using artificial ‘Bayesian Statistics’ published in probability and statistics of the Encyclopedia
neural networks: a case study for Turkish state railways. Struct Infrastruct Eng of Life Support Systems (EOLSS) (R Viertl, editor), Oxford, UK: UNESCO; 2003.
2014;10(5):614–26. [30] Paulino CD, Turkman MAA, Murteira B. Estatística Bayesiana. Fundação
[18] Prescott D, Andrews J. Stochastic state space methods for railway network Calouste Gulbenkian, 2003. 972-31-1043-1.
asset management modelling. Adv Risk Reliab Technol Symp 2013:421–33 [31] Ntzoufras I. Bayesian modelling using WinBUGS. Wiley series in computa-
ISBN 978-1-907382 61 1. tional statistics, 2009. 978-0-470-14114-4.
[19] Andrews J, Prescott D, Rozières F. A stochastic model for railway track asset [32] Lunn DJ, Thomas A, Best N, Spiegelhalter D. WinBUGS – A Bayesian modelling
management. Reliab Eng Syst Saf 2014;130:76–84. framework: concepts, structure, and extensibility. Stat Comput
[20] Mokhtarian P, Namzi-Rad MR, Ho TK, Suesse T. Bayesian nonparametric
2000;10:325–37.
reliability analysis for a railway system at a component level. In: Proceedings
[33] Gilks WR, Wild P. Adaptive rejection sampling for Gibbs sampling. J R Stat Soc,
of the IEEE International Conference on Intelligent Rail Transportation (ICIRT)
Ser C (Appl Stat) 1992;41:337–48.
2013. p. 197–202.
[34] Besag J. Spatial interaction and the statistical analysis of lattice systems. J R
[21] Gong C, Iwnicki S, Bezin Y. The effect of railway vehicle dynamics on the lateral
Stat Soc, Ser B (Methodological) 1974;36:192–236.
alignment of track. Proc Inst Mech Eng F: J Rail Rapid Transit. http://dx.doi.
[35] Gelman A. Prior distributions for variance parameters in hierarchical models
org/10.1177/0954409714536548.
(Comment on Article by Browne and Draper). Bayesian Anal 2006;3:515–34.
[22] UIC. Best practice guide for optimum track geometry durability, 2008.
[36] Besag J, Kooperger CL. On conditional and intrinsic autoregressions. Biome-
isbn:2-7461-1456-9.
[23] Andrade AR, Teixeira PF. Unplanned maintenance needs related to rail track trika 1995;82:733–46.
geometry. ICE Transp - Proc Inst Civil Eng – Transp 2014;167:400–10. [37] Smith B. Bayesian output analysis program (BOA) for MCMC. R package
[24] Andrade AR, Teixeira PF. A negative binomial regression model to predict version 1.1, 2005.
temporary speed restrictions in railway infrastructure. J Risk Reliab, under [38] Spiegelhalter DJ, Best NG, Carlin BP, Van Der Linde A. Bayesian measures of
review. model complexity and fit. J R Stat Soc, Ser B (Methodological) 2002;64(4):583-
[25] Zhao J, Chan AHC, Roberts C, Stirling AB. Optimizing policies of railway ballast 639.
tamping and renewal. Transp Res Rec: J Transp Res Board 2006;1943:50–6. [39] Silva GL, Dean CB, Niyonsenga T, Vanasse A. Hierarchical Bayesian spatiotem-
[26] El-Sibaie M, Zhang Y. Objective track quality indices. Transp Res Rec: J Transp poral analysis of revascularization odds using smoothing splines. Stat Med
Res Board 2004;1863:81–7. 2008;27:2381–401.
[27] Esveld C. Vehicle reactions versus track geometry. railroad conference, ASME/ [40] Andrade AR, Teixeira PF. Exploring different Alert Limit strategies in the
IEEE Joint. 1990. p. 171–174. maintenance of railway track geometry. J Transp Eng, in review.