Reliability Engineering and System Safety 216 (2021) 107940

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

A mathematical programming model to select maintenance strategies in

railway networks
Claudia Fecarotti a ,∗, John Andrews b , Raffaele Pesenti c
a Department of Industrial Engineering & Innovation Science, Eindhoven University of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands
b
Department of Mechanical Engineering, The University of Nottingham, University Park, NG72DR, Nottingham, UK
c
Department of Management, University Ca’Foscari Venezia, Cannaregio 873, 30121 Venezia, Italy

ARTICLE INFO ABSTRACT

Keywords: This paper presents a nonlinear integer programming model to support the selection of maintenance strategies
Maintenance optimisation to implement on different segments of a railway network. Strategies are selected which collectively minimise
Railway networks the impact of sections’ conditions on service, given network availability and budget constraints. Different
Mathematical programming
metrics related to the network topology, sections’ availability, service frequency, performance requirements
Availability
and maintenance costs, are combined into a quantitative approach with a holistic view. The main contribution
is to provide a simple yet effective modelling approach and solution method which are suitable for large
networks and make use of standard solvers. Both an ad hoc heuristic solution and relaxation methods are
developed, the latter enabling the quality of the heuristic solution to be estimated. The availability of railway
lines is computed by exploiting the analogy with series–parallel networks. By varying the model parameters,
a scenario analysis is performed to give insight into the influence of the system parameters on the selection
of strategies, thus enabling more informed decisions. For its simple structure, the model is versatile to address
similar problems arising in the maintenance of other types of networks, such as road and bridges networks,
when deciding on the strategic allocation of maintenance efforts.

1. Introduction This paper introduces an optimisation model based on mathematical

programming to support the selection of the maintenance strategies
Infrastructure asset maintenance for transportation networks is a to implement on different regions of a railway network. The resulting
complex decision making process involving the allocation of limited combination of strategies enables to obtain the best value from the
maintenance resources to achieve a trade-off between costs and ser- assets from a network perspective. Its intended use is subsequent to
vice performance. Railway networks are an example of large scale
the development of the actual strategies, which are here given as
transportation networks consisting of a variety of heterogeneous assets
input to the model. The model is performance-oriented, it assumes that
spatially distributed and interdependent. The network topology cou-
pled with the operational use of the infrastructure are such that a failure decisions are aimed at minimising the expected number of trains per
of an individual section might affect the availability of an entire line. time unit affected by a speed restriction, while availability targets for
Maintenance planning is addressed at strategic, tactical and oper- different railway lines have to be met and budget constraints have to be
ational level. Strategic maintenance planning involves first, at asset respected. The main contribution of this paper is to present a simple yet
level (e.g. track, signalling and telecoms, bridges, earthworks, electrical effective modelling approach and solution method which are suitable
power), the development of maintenance strategies which define fixed for large networks and make use of standard solvers, thus making
rules to maintain the assets. These rules include type of inspection its use in practice more attractive. From a modelling perspective, the
and its frequency, the waiting time to maintenance, the thresholds on value of the proposed model is to provide a quantitative approach
time, age, usage and/or conditions which trigger maintenance actions which combines different metrics related to the network topology,
with different levels of urgency. Then, at network level, strategies must sections availability, service frequency, performance requirements and
be chosen for different network segments. Maintenance strategies as
maintenance costs, in a holistic view with a clear and simple structure
defined above, form the basis for the planning of maintenance activities
rather than a black box. As the approach is quantitative in nature, a
and intervention programs defined at tactical and operational planning
levels, which govern the maintenance operations. wide study of the sensitivity of the solution with respect to the model

∗ Corresponding author.
E-mail address: c.fecarotti@tue.nl (C. Fecarotti).

https://doi.org/10.1016/j.ress.2021.107940
Received 2 September 2020; Received in revised form 19 July 2021; Accepted 20 July 2021
Available online 24 July 2021
0951-8320/© 2021 Published by Elsevier Ltd.
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

parameters is performed. The value of this study is to give insight on the the budget allocation and maintenance planning problem for multi-
relative and combined contribution of different factors on the selection asset systems is addressed again as a finite-horizon Markov decision
of strategies. These factors include track redundancy, centrality of sec- process for each individual asset. Then the system level optimisation
tions, service frequency, budget and lines availability targets. The final is formulated as an integer programming problem which minimise the
contribution of this paper is to have developed both a heuristic solution total expected maintenance cost of all assets under a budget constraint
approach and an ad hoc relaxation method, thus enabling the quality formulated for each decision period. The shared budget is the only
of the heuristic solution to be estimated. The model addresses a real interaction considered across the facilities. The problem is nonlinear,
problem in practice faced by infrastructure owners/managers related to and a standard linearisation method based on auxiliary binary variables
the strategic maintenance planning of their networked infrastructures. is adopted. Through Lagrangian relaxation of the budget constraint,
Although the model is developed and demonstrated here for a railway the system-level optimisation can be decomposed into multiple Markov
network, its simple structure makes it versatile and suitable to address decision problems, one for each asset, and a lower bound to the primal
similar problems arising in the maintenance of other types of networks problem is obtained. This decomposition approach is only possible
when deciding on the strategic allocation of maintenance efforts. Ex- because the shared budget is the only dependency among facilities. A
amples are road and bridges networks, where long-term maintenance linear optimisation model is presented in [11] to determine optimal
strategies must be chosen for individual or groups of road segments or intervention programs consisting of the interventions to be performed
bridges respectively. each year over a finite time horizon for a system of assets. The above
mentioned contributions all focus on sequential decision making and
1.1. Literature review formulate the problem as a Markov decision process, mostly solved via
dynamic programming. The sequence of interventions to be executed
for each asset per time interval (e.g. each year) during a finite time
Literature proposes different optimisation models to support main-
horizon is obtained. All models assume knowledge of the state of the
tenance planning for different infrastructures and planning levels. A
asset following intervention.
recent overview of the literature on asset management and mainte-
In [12] and [13] the authors address a similar problem as the above
nance of multi-unit systems is given in [1,2], and previously in [3,4].
cited contributions but with a different modelling approach. They aim
The survey in [1] distinguishes between contributions dealing with
at determining optimal intervention programs for railway networks
multi-component and multi-asset systems, focusing more specifically
with multiple assets, where interventions are selected among a given
on the latter. While authors usually refer to multi-component systems
set under budget and structural constraints. They adopt a network flow
to describe a single asset (e.g. a machine or a bridge) and its con-
model approach resulting in a mixed integer linear program which is
stituent components, multi-asset systems are systems of multiple assets
solved via the simplex and branch and bound methods. Interventions
otherwise connected. The present paper focuses on this last category.
are selected based on their cost, duration and reduction of failure
In the infrastructure sector, maintenance planning of systems of
risk. The approach is extended in [14] which focuses particularly on
assets mainly focuses on the optimisation of sequential decisions. Typi-
the estimation of intervention costs and the effects of intervention
cally these decisions concern the selection of one or more interventions
programs on service.
out of a given set based on the current state of the assets. Example
Other maintenance planning problems for multi-asset systems ap-
of interventions are rehabilitation, renovation, routine maintenance, pearing in the literature deal with the prioritisation of assets inter-
or do nothing. Decisions are often constrained by limited budgets and ventions, the grouping of works and the scheduling of maintenance
capacity targets, while the objective is to optimise either maintenance activities. In [15] a linear integer programming model is proposed
costs and/or some measure of the system performance. The authors for the scheduling of preventive small routine maintenance activi-
in [5] address the problem of planning maintenance for a system of ties and major jobs on a single railway link. The dynamic grouping
heterogeneous assets undergoing stochastic deterioration over a finite and scheduling of maintenance activities for multi-component sys-
time horizon. They develop a two-stage bottom-up approach. First, the tems is addressed in [16,17]. Jobs are grouped together to minimise
optimal maintenance activities are determined for each asset among a possession costs based on given maintenance frequency and which
set of feasible interventions. Then, a system-level optimisation selects routine jobs can be combined. The optimal grouping of components
the best combination of interventions given budget constraints. All is also addressed in [18] in the context of preventive maintenance of
assets in the system are independent except for the shared budget. A multi-component systems via a multi-level preventive decision-making
similar approach has been used in [6] with application to a hypothetical model. A maintenance grouping approach for a series system with
railway system. A number of assets are associated to each link in availability constraints is proposed in [19], where the maintenance
the network, each with a set of available maintenance activities. For planning is updated in a dynamic context. A multi-objective optimisa-
each asset the best activity is obtained by solving a Markov deci- tion model is presented in [20] for scheduling renewal works of ballast,
sion problem via dynamic programming. At system-level the budget rails and sleepers while seeking a trade-off between life-cycle costs and
constraint includes the cost reduction achieved through opportunistic the track availability. The problem is solved via a genetic algorithm.
maintenance of adjacent assets. A threshold on the minimum capacity The optimisation of maintenance schedule for railway power supply
to be guaranteed between an origin and a destination node is also system is addressed in [21,22] where the authors develop a biobjective
imposed. In [7], the authors develop a model with the objective of optimisation model which considers the trade-off between reliability
determining the optimal set of maintenance interventions for a system and maintenance costs. A biobjective optimisation model is also pre-
of bridge decks. Again, they suggest a two-step approach where they sented in [23] to optimise planned maintenance and renewal activities
use the optimal cost of maintenance and replacement for each bridge for track geometry. The model determines whether a track section is
resulting from a facility-level optimisation problem previously defined tamped or renewed in a given trimester or not, and what level of
in [8]. Then, at system level, they select the optimal combination of speed restriction is imposed. The contributions reviewed above address
maintenance and replacement activities which minimise the probability maintenance planning problems at a tactical and/or operational level.
of system failure subject to budget constraints. A two-stage bottom- At a strategic level instead, a typical problem is to determine long-
up optimisation approach is proposed in [9] for timely maintenance term maintenance policies defining the fixed rules for inspection and
planning in heterogeneous systems. The specific maintenance action maintenance such as inspection frequency and condition (or alter-
and time of execution are optimised per component based on a sys- natively age or usage) thresholds triggering interventions. In [24] a
tem perspective. The approach is demonstrated with application to a non-homogeneous Markov model solved via a numerical procedure is
simple imaginary railway network. In a more recent contribution [10] used to determine the probability of rail cracks, and it is combined

2
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

with a Genetic Algorithm to find rail inspection intervals and waiting contribution of individual track sections to the availability of railway
time for maintenance which minimise both costs and the probability lines along which service run. The model requires the knowledge of the
of rail cracks. In [25] the authors consider a series–parallel systems of effects of each strategy on the long-term behaviour of the assets. Input
components of different types. They use a Markov model to describe to the model can be given by state-based models as suggested in [31],
the degradation and on-condition maintenance processes, and recour which combine degradation and maintenance processes to predict the
to Monte Carlo simulation to obtain the average system availability long-term assets response to maintenance strategies. Petri nets [32] can
and the probability of being under maintenance at any given time t. be used to this aim, as they can assess the probabilities of different
A genetic algorithm is then used to find the maintenance thresholds states of interest, including section closures and speed restrictions, as
per component type, which simultaneously optimise profit from system well as the average work volumes (see e.g. [33–36] for railway tracks
operation and system availability. In [26] a life-cycle cost model is pro- and bridges). Alternative methods which model the assets as multi-
posed to investigate tamping and track renewal strategies to minimise state systems (e.g. [37,38,24]) can serve the same purpose, as well as
ballast life-cycle costs. In particular, an iterative search algorithm is historical data.
used to determine the best strategy among a set of alternatives includ- The rest of the paper is organised as follows. After providing the
ing fixed intervals or fixed condition’s thresholds triggering a tamping problem description and problem statement in Section 2, the model
intervention. A similar decision problem is addressed in [27] for multi- is presented in Section 3, where the approach to model the unavail-
component systems, where the optimal frequency of scheduled visits ability of railway lines based on the theory of reliability networks is
and the preventive maintenance threshold for a component within a described. Then Section 4 provides the solution approach, including
series system are sought. Here renewal theory is used to evaluate the both the method developed to find an approximate solution and a lower
long-term average maintenance costs, and an iterative procedure is bound to estimate the quality of the approximate solution. Finally the
used to find the near optimal values of visits’ frequency and mainte- proposed modelling and solution methods are tested through a wide
nance thresholds that minimise the long-term average cost. The above scenario analysis on an illustrative example considering a real portion
contributions look at the development of maintenance strategies by of the UK railway network.
optimising maintenance parameters such as inspection intervals and
maintenance thresholds simultaneously for multiple components within
2. Problem description
a system. However they assume that the “structure” of a maintenance
strategy is the same for all component types, and the only difference is
The problem addressed in this paper is to optimise the selection of
in the values taken by the maintenance parameters.
long-term maintenance strategies to implement on different segments
This paper can be framed among those works aiming at the selec-
of a given network. The aim is to obtain the best value from the assets
tion of long-term maintenance strategies for multi-asset systems with
from a network perspective based on a trade-off between maintenance
a networked structure. However, unlike the above contributions, it
costs, network availability requirements and delivered level of ser-
presents an approach that allows for fundamentally different mainte-
vice. Indeed, the infrastructure owner wants to maintain its portfolio
nance strategies to be compared based on their costs and resulting
of assets to ultimately deliver the required network availability and
availability and level of performance of the assets. Indeed in reality,
service levels (e.g. permissible speed) [39], and he is bounded to do
the assets comprising an infrastructure system are quite diverse and
complex, and this reflects into equally diverse and complex mainte- so with a limited budget. In railway practice, long-term maintenance
nance strategies. From here the choice to separate the problem of strategies are normally developed per asset type (e.g. track, signalling
developing individual strategies, from the one of selecting the best and telecoms, bridges, earthworks, electrical power) [39–41]. Different
strategies to be used in combination for a network of assets and allocate strategies result in different long-term costs and conditions profiles, and
the maintenance budget accordingly. This paper focuses on the latter lifecycle costs analysis tools are used to determine such profiles per
problem, thus contributing to the wider literature of portfolio decision strategy and per asset. The assets’ conditions in turn have an impact on
analysis which aims at supporting decisions for the selection of projects the proportion of time the section they insist on is closed or subject to a
and the allocation of resources [28]. Recent contributions to portfolio speed restriction. Here it is assumed that a set of maintenance strategies
decision analysis in infrastructure maintenance are [29,30]. In [29], a is given, and that the effect of each strategy on the unavailability and
sequential portfolio selection approach on a multi-period horizon is pre- speed reduction of a section is known. These values can be obtained
sented to identify the optimal risk-based maintenance portfolios for gas from multiple models available in the literature among which are [33–
networks. The maintenance portfolio consists of a set of pipe segments 38,24], as well as from historical data. Railway services run along
which will undergo complete replacement. The problem of allocating railway lines for which a minimum threshold of availability must be
prognostic and health management capabilities in power transmission achieved to ensure the use of the infrastructure as agreed with the
networks to maximise the network global reliability efficiency under train operating companies (TOCs). When a speed restriction is imposed
budget constraints is addressed in [30]. on a track section, then trains are delayed. Strategic planning looks at
In the present paper, given a portfolio of potential maintenance medium/long-term planning horizons for which no details are available
strategies and a network of assets, one wants to select the strategies that on trains timetable and maintenance execution schedule. It is therefore
collectively enable a trade-off between maintenance costs, system avail- not possible at this level of planning to obtain more precise measures of
ability and service impact. This corresponds to allocate maintenance train delays. For this reason, here it is assumed that the proportion of
efforts among different sections of the network. The problem is for- time that a section is subject to a speed restriction combined with the
mulated here for a railway network and is modelled via mathematical train frequency, is representative of the impact of a chosen strategy on
programming. A nonlinear combinatorial optimisation problem is pre- service delay per section. The superposition of the effects for all sections
sented, where the nonlinearities are due to the availability constraints is then considered at network level. The logic behind the assignment
imposed on railway lines which account for the network topology. It of strategies to segments of the network is based on the network
further develops the work in [31] by refining an ad hoc linearisation segmentation approach used in practice for strategic planning purposes.
approach to approximate the optimal solution from above, and by pre- The UK railway network is split into 19 Strategic Routes, which in
senting a relaxation method to quantify the quality of the solution and turn are partitioned into Strategic Route Sections (SRSs) [42], each
therefore the performance of the proposed heuristic. The model takes including multiple tracks and stations. A SRS is a section of the network
into account economic, functional and operational dependencies be- characterised by broadly homogeneous traffic and infrastructure type,
tween different sections of the network by considering a shared limited thus the same maintenance strategy applies to an entire SRS. The lines
budget, and by making use of reliability network theory to model the along which railway services are provided, consist of one or more SRSs.

3
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Fig. 1. A generic series–parallel system.

2.2. Problem statement

We represent a Strategic Route as a set of SRSs, 𝑅 ∶= {𝑟 ∶ 𝑟 =

1, 2, … , 𝑛𝑅 }. We define 𝐿 ∶= {𝑙 ∶ 𝑙 = 1, 2, … , 𝑛𝐿 } as the set of the
𝑛𝐿 served railways lines, where 𝑙 ⊆ 𝑅, for all 𝑙 ∈ 𝐿. We also define
𝑆 ∶= {𝑠 ∶ 𝑠 = 1, 2, … , 𝑛𝑆 } as the set of the 𝑛𝑆 (maintenance) strategies
available for each SRS. The same strategy 𝑠 applies to an entire SRS 𝑟.
Each element of 𝑆 corresponds to a different combination of individual
asset type sub-strategies. In addition, we use the following notation:

• 𝑥 is a Boolean vector of decision variables whose components 𝑥𝑠𝑟

indicate whether strategy 𝑠 is implemented on SRS 𝑟, namely
{
1 if strategy 𝑠 is selected for SRS 𝑟
𝑥𝑠𝑟 = (3)
0 otherwise;

• 𝛿𝑙𝑟 is a Boolean parameter describing whether railway line 𝑙 ∈ 𝐿

includes SRS 𝑟 ∈ 𝑅, namely
{
1 if 𝑟 ∈ 𝑙
𝛿𝑙𝑟 = (4)
Fig. 2. Schematic representation of a railway line as a series of SRSs. 0 otherwise;

• 𝑑𝑠𝑟 is the proportion of time that SRS 𝑟 ∈ 𝑅 is subject to a speed

restriction following implementation of strategy 𝑠 ∈ 𝑆;
2.1. Modelling the unavailability of railway lines
• 𝑐𝑠𝑟 is the cost per unit of time of strategy 𝑠 ∈ 𝑆 implemented on
SRS 𝑟 ∈ 𝑅;
In this paper railway lines are assimilated to series–parallel systems
• 𝑓𝑟 is the nominal frequency of trains travelling on a SRS 𝑟 ∈ 𝑅;
of the type shown in Fig. 1. It consists of a set of subsystems 𝐼 ∶= {𝑖 ∣
• 𝑄𝑙 (𝑥) is a function 2𝑅×𝑆 → [0, 1] which maps the strategies
𝑖 = 1, 2, … , 𝑛𝐼 } connected in series. Each subsystem consists of a set
implemented on the SRSs of a line 𝑙 ∈ 𝐿 into the unavailability of
of components 𝑃𝑖 ∶= {𝑗 ∣ 𝑗 = 1, 2, … , 𝑚𝑖 } connected in parallel. The
line 𝑙. Given that the same maintenance strategy is implemented
unavailability of a series–parallel system 𝑄𝑆𝑒𝑟𝑖𝑒𝑠−𝑃 𝑎𝑟𝑎𝑙𝑙𝑒𝑙 is
throughout a SRS, Eq. (2) can be rewritten as
∏ ∏
𝑄𝑆𝑒𝑟𝑖𝑒𝑠−𝑃 𝑎𝑟𝑎𝑙𝑙𝑒𝑙 = 1 − (1 − 𝑞𝑗 ), (1) ∏ ∑ ∏ ∏
𝑄𝑙 (𝑥) = 1 − {1 − 𝛿𝑙𝑟 𝑥𝑠𝑟 [1 − (1 − 𝑝𝑗𝑠 )]}, (5)
𝑖∈𝐼 𝑗∈𝑃𝑖
𝑟∈𝑅 𝑠∈𝑆 𝑖∈𝐼𝑟 𝑗∈𝑃𝑖
where 𝑞𝑗 is the unavailability of the generic component 𝑗. A railway where 𝑝𝑗𝑠 is the unavailability of track section 𝑗 if strategy 𝑠 is
line is modelled as a sequence of portions of SRSs connected in series. implemented;
Each portion of SRS included in a line is in turn composed by a number • 𝑄∗𝑙 is the maximum tolerable unavailability of line 𝑙 ∈ 𝐿;
of track sections connected as a series–parallel system, where each
• 𝑏 is the available budget per unit of time.
component is a continuous section of track between two consecutive
stations/junctions. An example of such a representation is given in We are now ready to formally state the problem of interest.
Fig. 2, which depicts a railway line including two SRSs, 𝑥 and 𝑦. This
representation reflects the topology of a railway line as a sequence Problem 1 (RMSP — Railway Maintenance Strategies Problem). Consider
of single, double (or more) track sections connecting stations and a Strategic Route consisting of a set 𝑅 of Strategic Route Sections
junctions. (SRSs), a set 𝑆 of strategies available for each SRS and a budget
By exploiting the analogy with series–parallel systems, the unavail- 𝑏. Choose a strategy 𝑠 ∈ 𝑆 for each SRS 𝑟 ∈ 𝑅 so that the total
ability of a railway line, 𝑄𝑙 , can be written as expected number of delayed trains per unit of time is minimised, the
∏ ∏ ∏ total cost of the chosen strategies does not exceed the budget 𝑏 and
𝑄𝑙 = 1 − [1 − 𝛿𝑙𝑟 (1 − 𝑝𝑗 )] (2) their implementation keeps the unavailability of each line 𝑙 below the
𝑟∈𝑅 𝑖∈𝐼𝑟 𝑗∈𝑃𝑖
acceptable threshold 𝑄∗𝑙 .
where 𝛿𝑙𝑟 is a Boolean parameter which takes value 1 if line 𝑙 includes
SRS 𝑟, and 0 otherwise. Index 𝑖 ∈ 𝐼𝑟 indicates the generic section 3. Model formulation
between two consecutive stations/junctions within SRS 𝑟. Section 𝑖
may be single track, or include more than one parallel track. It can The problem is formulated as follows:
be therefore assimilated to a parallel system with components 𝑗 ∈ 𝑃𝑖 , ∑∑
min 𝑧(𝑥) = 𝑑𝑠𝑟 𝑓𝑟 𝑥𝑠𝑟 𝑠.𝑡. (6)
where 𝑗 is the generic track with unavailability 𝑝𝑗 .
𝑟∈𝑅 𝑠∈𝑆

4
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

∑
𝑥𝑠𝑟 = 1 ∀ 𝑟 ∈ 𝑅, (7) 𝑧∗𝐿𝑀 ∶= min 𝑧𝐿𝑀 (𝑥).
𝑥
𝑠∈𝑆
∑∑ ̃𝑙 (𝑥) is an upper bound to the non
𝑐𝑠𝑟 𝑥𝑠𝑟 ≤ 𝑏, (8) Since the linear approximation 𝑄
𝑟∈𝑅 𝑠∈𝑆 ̃𝑙 (𝑥), any solution to the linearised
linear 𝑄𝑙 (𝑥), namely 𝑄𝑙 (𝑥) ≤ 𝑄
∏ ∑ ∏ ∏ problem 𝐿𝑀 will also be feasible for the original nonlinear problem.
𝑄𝑙 (𝑥) = 1 − {1 − 𝛿𝑙𝑟 𝑥𝑠𝑟 [1 − (1 − 𝑝𝑗𝑠 )]} ≤ 𝑄∗𝑙 ∀𝑙 ∈ 𝐿, (9)
𝑟∈𝑅 𝑠∈𝑆 𝑖∈𝐼𝑟 𝑗∈𝑃𝑖 Solving 𝐿𝑀 therefore provides a suboptimal solution for the original
nonlinear problem.
𝑥𝑠𝑟 ∈ {0, 1} ∀ 𝑟 ∈ 𝑅, 𝑠 ∈ 𝑆. (10)

The objective function 𝑧(𝑥) represents the long-run expected number 4.2. Associated relaxed problem: lower bound solution
of trains affected by a speed restriction per unit time. Each term
𝑑𝑠𝑟 𝑓𝑟 𝑥𝑠𝑟 is a measure of the contribution of each SRS to the overall Standard relaxation methods include relaxation by elimination of a
service disruption. The train frequency is used to weight each SRS subset of constraints, Lagrangian and continuous relaxation [44]. How-
proportionally to the normalised number of trains travelling on the SRS. ever, given some of the properties of the linearised problem 𝐴𝐿, it is
The frequency also implicitly weights each SRS based on its centrality, possible to build a relaxed problem with a much simpler approach than
namely its role in serving more than one line. This is because sections the standard relaxation methods mentioned above. This is explained in
serving multiple lines usually carry higher train frequencies. The set of the following. The error induced by replacing 𝑄𝑙 (𝑥) with its linear form
constraints (7) indicates that only one strategy can be selected for each ̃𝑙 (𝑥) is 𝑄
𝑄 ̃𝑙 (𝑥) − 𝑄𝑙 (𝑥) = 𝐸𝑄 (𝑥). The error 𝐸𝑄 (𝑥) becomes smaller, and
𝑙 𝑙
SRS. Constraint (8) adds a bound on the overall costs according to the therefore the approximation is more accurate, as the terms 𝑝𝑗𝑠 decrease.
available budget. The set of constraints (9), one for each line 𝑙, puts Given the set of potential strategies 𝑆, it is therefore possible to identify
a threshold on the unavailability of each line. Here, Eq. (5) is used to a lower bound 𝐸𝑄𝑙 ∣𝑥𝑚𝑖𝑛 and an upper bound 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 to the error
express the unavailability of each line 𝑙. Constraints (9) are strongly incurred in the approximation of the constraint for each railway line.
nonlinear, which makes the problem a nonlinear integer optimisation 𝐸𝑄𝑙 ∣𝑥𝑚𝑖𝑛 is obtained when the strategy providing the lowest value of
problem. section unavailability 𝑝𝑗𝑠 is implemented to all the SRSs. It is calculated
from Eq. (5) by using the lowest value of 𝑝𝑗𝑠 for all sections 𝑗. 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥
4. Solution approach is obtained if the strategy with the highest value of 𝑝𝑗𝑠 ) is implemented
to all SRSs. Hence the error depends on the current value of the decision
To solve the nonlinear problem we propose an ad hoc linearisation variables and is bounded as follows
procedure to find an approximate solution. We then formulate a relaxed 𝐸𝑄𝑙 ∣𝑥𝑚𝑖𝑛 < 𝐸𝑄𝑙 (𝑥) < 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 , (17)
counterpart to the nonlinear problem to determine a lower bound to the
optimal solution. The upper and lower bounds thus obtained are used where 𝑥𝑚𝑖𝑛 and 𝑥𝑚𝑎𝑥 are the values of the decision variable vector if the
to calculate an upper bound to the percentage error, 𝜀+ %, which is used best and worst strategies are selected respectively for all SRSs.
to measure the level of suboptimality of the approximate solution. If Constraints (9) are relaxed by adding 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 to the right-hand
side, the solution to the relaxed problem thus obtained is a lower bound
4.1. Linearisation: upper bound solution not only to the linearised problem but also to the original non-linear
problem. Let us name such relaxed problem 𝑅𝑀, formally defined as
For the problem at hand the unavailability of a railway line as ∑∑
(𝑅𝑀) min 𝑧𝑅𝐴𝐿 (𝑥) ∶= 𝑑𝑠𝑟 𝑓𝑟 𝑥𝑠𝑟 𝑠.𝑡. (18)
given in Eq. (5) is strongly non-linear. However, two assumptions are 𝑥
𝑟∈𝑅 𝑠∈𝑆
made: (i) the occurrences of sections’ closure are random variables ∑
𝑥𝑠𝑟 = 1 ∀ 𝑟 ∈ 𝑅, (19)
independently distributed, and (ii) the corresponding probability is 𝑠∈𝑆
small. In such circumstances it is acceptable to approximate the un- ∑∑
𝑐𝑠𝑟 𝑥𝑠𝑟 ≤ 𝑏, (20)
availability of a series system with an upper bound given by its Rare 𝑟∈𝑅 𝑠∈𝑆
Event Approximation [43]. This is a first order approximation according
̃𝑙 (𝑥) ≤ 𝑄∗ + 𝐸𝑄 ∣𝑥
𝑄 ∀𝑙 ∈ 𝐿, (21)
to which the unavailability of a series system can be approximated with 𝑙 𝑙 𝑚𝑎𝑥

the sum of the unavailability of the individual components. It follows 𝑥𝑠𝑟 ∈ {0, 1} ∀ 𝑟 ∈ 𝑅, 𝑠 ∈ 𝑆. (22)
that the unavailability of a railway line can be simplified as:
∑ ∑ ∏ ∏ The optimal solution 𝑥∗𝑅𝑀
and optimal value of the objective function
̃𝑙 (𝑥) =
𝑄𝑙 (𝑥) ≤ 𝑄 𝛿𝑙𝑟 { 𝑥𝑠𝑟 [1 − (1 − 𝑝𝑗𝑠 )]}. (11) 𝑧∗𝑅𝑀 for problem 𝑅𝑀 are:
𝑟∈𝑅 𝑠∈𝑆 𝑖∈𝐼𝑟 𝑗∈𝑃𝑖
𝑥∗𝑅𝑀 ∶= arg min 𝑧𝑅𝑀 (𝑥)
If the left-hand side of Constraint (9) is replaced with its Rare Event 𝑥
Approximation (11), the original nonlinear problem (6) is transformed 𝑧∗𝑅𝑀 ∶= min 𝑧𝑅𝑀 (𝑥).
𝑥
into a linear integer programming model. This linearised model will be
Once a solution to the linearised problem 𝐿𝑀 and the relaxed
called from now on 𝐿𝑀 and its formal definition is
∑∑ problem 𝑅𝑀 are obtained, the upper bound to the percentage error,
(𝐿𝑀) min 𝑧𝐿𝑀 (𝑥) ∶= 𝑑𝑠𝑟 𝑓𝑟 𝑥𝑠𝑟 𝑠.𝑡. (12) 𝜀+ %, can be calculated as
𝑥
𝑟∈𝑅 𝑠∈𝑆
∑ 𝑧∗𝐿𝑀 − 𝑧∗𝑅𝑀
𝑥𝑠𝑟 = 1 ∀ 𝑟 ∈ 𝑅, (13) 𝜀+ % = %. (23)
𝑧∗𝑅𝑀
𝑠∈𝑆
∑∑
𝑐𝑠𝑟 𝑥𝑠𝑟 ≤ 𝑏, (14) 𝜀+ % is the maximum error one can commit wrt the optimal solu-
𝑟∈𝑅 𝑠∈𝑆 tion, by taking the approximate solution resulting from the linearised
̃𝑙 (𝑥) ≤ 𝑄∗
𝑄 ∀𝑙 ∈ 𝐿, (15) problem.
𝑙

𝑥𝑠𝑟 ∈ {0, 1} ∀ 𝑟 ∈ 𝑅; 𝑠 ∈ 𝑆. (16) 5. Numerical study: application to the East Midlands Strategic
The optimal solution 𝑥∗𝐿𝑀
and optimal value of the objective function Route
𝑧∗𝐿𝑀 for problem 𝐿𝑀 are:
The optimisation approach is applied here to select the best com-
𝑥∗𝐿𝑀 ∶= arg min 𝑧𝐿𝑀 (𝑥) bination of maintenance strategies for a set of SRSs comprising one of
𝑥

5
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Table 3
Model parameters.
SRS 𝑐1𝑟 𝑐2𝑟 𝑐3𝑟 𝑑1𝑟 𝑑2𝑟 𝑑3𝑟
01 70 80 95 0.05 0.005 0.0005
02 70 80 95 0.05 0.005 0.0005
03 50 70 85 0.05 0.005 0.0005
04 45 65 80 0.05 0.005 0.0005
05 40 60 70 0.05 0.005 0.0005
06 40 60 70 0.05 0.005 0.0005
07 40 60 70 0.05 0.005 0.0005

Table 4
Maximum error 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 for each line.
Line 𝐸𝑄𝑙 ∣𝑋𝑚𝑎𝑥
𝑙1 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 = 1.2610 × 10−10
1
𝑙2 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 = 7.0854 × 10−8
2
𝑙3 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 = 2.3153 × 10−7
3
𝑙4 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 = 2.0201 × 10−8
4
𝑙5 𝐸𝑄𝑙 ∣𝑥𝑚𝑎𝑥 = 9.4174 × 10−5
5

Table 5
Unavailability thresholds for each line.
Set 𝑄∗1 𝑄∗2 𝑄∗3 𝑄∗4 𝑄∗5
𝑆𝐴11 0.001 0.008 0.01 0.05 0.01
𝑆𝐴12 0.0001 0.0008 0.001 0.005 0.008
𝑆𝐴13 0.00001 0.00008 0.0001 0.0005 0.0008

Fig. 3. Simplified map of part of the East Midlands Route including seven of its eleven
SRSs (11.01 to 11.07) based on [42].
each with 9 sections in series. Three potential maintenance strategies
Table 1 have been considered, 𝑆 = {𝑠1, 𝑠2, 𝑠3}. Table 3 provides the cost 𝑐𝑠𝑟 and
SRSs and trains frequency. the probability of a speed restriction 𝑑𝑠𝑟 when strategy 𝑠 is implemented
SRS Trains per hour to SRS 𝑟. The values of unavailability 𝑝𝑗𝑠 corresponding to each strategy
01 London St. Pancras–Bedford 20 are defined per track section 𝑗 (as explained in Section 2.1). For
02 Bedford–Nottingham 8 this numerical analysis it is assumed that for all track sections 𝑗, the
03 Wichnor Jn/Long Eaton–Chesterfield 8 unavailability 𝑝𝑗𝑠 takes values 0.01, 0.001 and 0.0001 for strategy 𝑠1,
04 Chesterfield–Nottingham 4
𝑠2 and 𝑠3 respectively.
05 Nottingham–Newark Castle 1
06 Matlock–Ambergate 1 The maximum error 𝐸𝑄𝑙 ∣𝑋𝑚𝑎𝑥 in the evaluation of the unavailability
07 Netherfield–Grantham 2 of each railway line induced by linearisation is given in Table 4 for each
railway line.
Table 2 The error is very small, with the highest value achieved for line 𝑙5 .
Service types and SRSs included within each railway line. This is due to the fact that in line 𝑙5 , each section connected in series
Service name Service type SRSs consists of a single or a double track (see Fig. 4), while the other lines
London St. Pancras to Nottingham LDHS {01, 02} all have two, three and four parallel tracks, thus making each term
London St. Pancras to Sheffield (via Derby) LDHS {01, 02, 03} in the series smaller. For each railway service, different availability
Norwich to Liverpool Interurban {02, 04, 07} requirements can be considered depending on the performance targets
Nottingham to Leeds Interurban {02, 04}
set by the infrastructure manager. Similarly, different levels of budget
Newark Castle–Nottingham–Derby–Matlock Local {02, 03, 05, 06}
can be investigated. A scenario analysis study has been conducted by
varying the value of the available budget 𝑏 and the threshold values on
the unavailability of railway lines 𝑄∗𝑙 . For each scenario, problems 𝐿𝑀
the UK Strategic Routes, the East Midlands (EM) Route. Details of the and 𝑅𝑀 are solved to provide an upper and a lower bound to the global
EM route and its SRSs can be found in [42]. A simplified graphical optimum of the original nonlinear problem respectively. The estimation
illustration of part of the EM route showing seven SRSs is given in of the optimality gap 𝜀+ % is also calculated by applying Eq. (23). The
Fig. 3. optimisation models have been implemented in Matlab R2018b and a
The set of SRSs considered in this example are listed in Table 1 along standard solver based on the Branch and Bound algorithm has been
with the train frequency. used. The computational time to solve 90 instances (both 𝐿𝑀 and 𝑅𝑀
Railway services running along the EM Route which have been for 45 scenarios) was 2.768 s, run on a Intel Core i5-8350U processor,
considered here are listed in Table 2 along with the service type (Long CPU 1.70 GHz 1.90 GHz, 8 GB RAM, 64 bit Operating System.
distance high speed — LDHS, interurban and local) and lists the SRSs
included within each service.
The series–parallel representation of each railway line is given in 5.1. Scenario analysis 𝑆𝐴1
Fig. 4. The circles represent sections of track between two consecutive
stations. For example, line 1 includes SRS 01 and SRS 02. SRS 01 For this first scenario analysis 𝑆𝐴1, it has been assumed that dif-
consists of four parallel tracks, each divided into 14 sections connected ferent lines have different thresholds on their unavailability, thus re-
in series. Connection between the parallel lines is at the two ends. SRS flecting different criticality values. Three sets of scenarios have been
02 comprises: a portion of four parallel tracks, each with 9 sections in analysed, for each set the thresholds on the availability of the railway
series, a portion of three parallel tracks and a portion of double track, lines are given in Table 5.

6
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Fig. 4. Railway lines as series parallel systems.

For each set, the following values of the available budget 𝑏 have for sets of scenarios 𝑆𝐴11 , 𝑆𝐴12 and 𝑆𝐴13 respectively. The tables
been tested: detail the strategy selected for each SRS and the corresponding value
𝑏 = {350, 375, 400, 425, 450, 475, 500, 525, 550, 575, 600, 625, 650, 675, 700}. of the objective function 𝑧∗𝐿𝑀 . The optimal value 𝑧∗𝑅𝑀 obtained by
The number of scenarios analysed in 𝑆𝐴11 is therefore 3 × 15 = 45. solving the corresponding relaxed problem 𝑅𝑀 and the upper bound
Tables 6, 7, 8 show the results obtained by solving problem 𝐿𝑀 to the percentage error, 𝜖+ % calculated using Eq. (23) are also given.

7
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Table 6
Results for set 𝑆𝐴11 .
Set 𝑆𝐴11 : 𝑄∗1 = 0.001, 𝑄∗2 = 0.008, 𝑄∗3 = 0.01, 𝑄∗4 = 0.05, 𝑄∗5 = 0.01
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – – s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – – s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
03 – – s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
04 – – s1 s1 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – – s1 s1 s1 s1 s1 s2 s2 s3 s3 s3 s3 s3 s3
06 – – s2 s2 s2 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3
07 – – s1 s1 s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
𝑧∗𝐿𝑀 – – 1.255 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – – 1.255 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝜖+ % – 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Table 7
Results for set 𝑆𝐴12 .
Set 𝑆𝐴12 : 𝑄∗1 = 0.0001, 𝑄∗2 = 0.0008, 𝑄∗3 = 0.001, 𝑄∗4 = 0.005, 𝑄∗5 = 0.001
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – – s1 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – – s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
03 – – s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
04 – – s1 s2 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – – s1 s1 s1 s1 s1 s2 s2 s3 s3 s3 s3 s3 s3
06 – – s2 s2 s2 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3
07 – – s1 s1 s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3

𝑧∗𝐿𝑀 – – 1.795 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – – 1.795 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
+
𝜖 % – – 0 0 0 0 0 0 0 0 0 0 0 0 0

Table 8
Results for set 𝑆𝐴13 .
Set 𝑆𝐴13 : 𝑄∗1 = 0.00001, 𝑄∗2 = 0.00008, 𝑄∗3 = 0.0001, 𝑄∗4 = 0.0005, 𝑄∗5 = 0.0001
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – – – – – s1 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – – – – – s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
03 – – – – – s2 s2 s3 s3 s3 s3 s3 s3 s3 s3
04 – – – – – s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – – – – – s2 s2 s2 s2 s3 s3 s3 s3 s3 s3
06 – – – – – s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
07 – – – – – s2 s2 s3 s2 s3 s3 s3 s3 s3 s3
𝑧∗𝐿𝑀 – – – – – 1.1155 1.1225 0.0805 0.0355 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – – – – – 1.1155 1.1225 0.0805 0.0355 0.022 0.022 0.022 0.022 0.022 0.022
𝜖+ % – – – – – 0 0 0 0 0 0 0 0 0 0

A comparison between the expected number of delayed trains for all that enables to achieve the required levels of availability for all lines.
scenarios are also compared in Fig. 5. For any further increment of the available budget, better strategies are
No feasible solution to problem 𝐿𝑀 can be found for budgets 𝑏 = selected first for the SRSs with higher train frequency. These are in fact
350 and 𝑏 = 375 as the strategies that are affordable do not ensure the SRSs with a higher impact on the objective function.
the required level of availability for each railway line. The increased For set of scenarios 𝑆𝐴12 , the availability thresholds are decreased
budget 𝑏 = 400 is enough to find a feasible solution. Such solution for all lines. Results are given in Table 7. As before, no feasible solution
consists of strategy 𝑠1 selected for all SRSs except SRS 06 and 01 for exist until budget is increased to 𝑏 = 400, when the optimal solution
which a better strategy 𝑠2 is selected. SRS 01 belongs to the line with consists of the cheapest strategy (𝑠1) for all SRSs except SRSs 02 and 06.
the most restrictive availability requirement, which justify the selection While SRS 01 only belongs to line 𝑙1 , SRS 02 belongs to all lines which
of a better strategy. Even though SRS 06 belongs to the line with the now have to meet more restrictive availability targets with respect to
less restrictive availability requirement, it is given priority over other scenarios 𝑆𝐴12 . For further increases of the budget, better strategies
sections. The explanation is that, since SRS 06 is single track, a less are selected first for the SRSs with higher train frequency.
performing strategy than 𝑠2 would not enable to meet the availability Finally, results for scenarios 𝑆𝐴13 are given in Table 8. None of the
threshold. This solution is also the cheapest combination of strategy strategies affordable with a budget up to 𝑏 = 450 constitute a feasible

8
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Fig. 5. Expected number of delayed trains due to speed restrictions per unit time for Fig. 6. Expected number of delayed trains due to speed restrictions per unit time for
scenario analysis 𝑆𝐴1. scenario analysis 𝑆𝐴2.

solutions. Strategy (𝑠3) is needed for single tracked SRS 06, and at least corresponding value of the objective function 𝑧∗𝐿𝑀 . The optimal value
strategy 𝑠2 for SRSs 02, 03 and 05 to meet the unavailability threshold 𝑧∗𝑅𝑀 and the upper bound to percentage error, 𝜖+ % are also given.
on line 𝑙5 . The approximate solutions obtained by solving problem 𝐿𝑀 For budget value 𝑏 = 350 no feasible solution to problem 𝐿𝑀 exists
and the lower bound given by relaxed problem 𝑅𝑀 coincide for all given the available strategies, regardless of threshold 𝑄∗ . As it can
scenarios. This means that the linearised problem 𝐿𝑀 yields the global be observed from Table 9, when the budget is increased to 𝑏 = 375
optimum for the considered scenarios. This can be explained by the an optimal solution is found for 𝑄∗ = 0.08 only, and it consists of
fact that the unavailability of all lines is of a bigger order than the strategy 𝑠1 for all SRSs; 𝑄∗ = 0.08 is permissive enough that the
corresponding error 𝐸𝑄𝑙 |𝑚𝑎𝑥 (see Table 4). cheapest strategy is sufficient for all lines to meet the availability target.
𝑖
The analysis of the three sets of scenarios has helped identify the For any further increment of the budget, the optimal solutions are
factors affecting the solution, and their importance. Such factors are: selected based on their impact on the objective function only and better
(i) the level of redundancy of the SRSs, (ii) the centrality of the SRSs, strategies are selected first for the SRSs with higher train frequency.
namely if a SRSs belongs to one or more lines, (iii) the required level of When the threshold 𝑄∗ is decreased to 0.008, strategy 𝑠1 selected
availability, and (iv) the train frequency. The level of redundancy and for all SRSs no longer provides a feasible solution. Budget 𝑏 = 375
the centrality of the SRSs are indicative of the influence of the network is not enough to select a better strategy, hence no feasible solution
topology. The first three factors are the features to look at first as they exists that meets both budget and line availability constraints. If the
determine whether the affordable solutions are actually feasible for the budget is increased to 400, strategy 𝑠2 can be selected for 𝑆𝑅𝑆06,
availability targets. The more restrictive these availability targets are, which combined with strategy 𝑠1 for all other SRSs provides a feasible
the more important are the level of redundancy and centrality of each solution. However, the budget is enough to select 𝑠2 also for 𝑆𝑅𝑆01,
SRS. The fourth factor affecting the solution is the train frequency. thus producing a solution which is not only feasible but also optimal.
As long as both budget and availability constraints are satisfied, the For further increases of the budget, better strategies are selected first
selection of the strategies is only based on the effect on the objective for SRSs with higher train frequencies. It is worth noticing that for
function. Better strategies are applied first to SRSs with higher train 𝑆𝑅𝑆06 strategy 𝑠1 is never selected. Furthermore, if solutions obtained
frequency. If two or more SRSs have the same train frequency, then for 𝑏 = 475, 𝑄∗ = 0.08 in Table 9 and 𝑏 = 475, 𝑄∗ = 0.008 in Table 10
the order in which these SRSs are attributed the best strategy does not are compared, one can see that the only difference is that when the
matter as they would result in equivalent optimal solutions. threshold is made more restrictive, strategy 𝑠2 is selected for 𝑆𝑅𝑆06
rather than 𝑆𝑅𝑆05. One could deduct that line 𝑙5 can only reach the
5.2. Scenario analysis 𝑆𝐴2 availability target if at least strategy 𝑠2 is selected for 𝑆𝑅𝑆06 as the
latter is single track.
In scenario analysis 𝑆𝐴1 different lines had different availability Table 11 shows that no feasible solutions exist for budget 𝑏 = 350
thresholds. It is therefore difficult to evaluate the order in which the and 𝑏 = 450 when the threshold is 𝑄∗ = 0.0008 as at least strategy 𝑠3
first three important factors (redundancy, centrality and availability for the single tracked SRS 06, and 𝑠2 for most of the others SRSs are
targets) influence the solution, and the impact of the features indicative needed to meet the availability targets. A first feasible solution is found
of the network topology regardless of different requirements set for for 𝑏 = 475, which enables strategy 𝑠3 to be selected for SRS06 and
each line. A second scenario analysis has been carried out here, 𝑆𝐴2, double tracked SRS 05, and strategy 𝑠2 for SRSs 02 and 03 so that line
where all lines have the same availability target, so that no line has 5 achieves the required availability target. Strategy 𝑠2 is also selected
priority over the others. Values of lines unavailability considered are for the other SRSs, except 07, to satisfy the availability requirements.
𝑄∗ = {0.08, 0.008, 0.0008}, while 15 different budget values are consid- Higher budgets lead to better strategies to be chosen for higher fre-
ered as in scenario analysis 𝑆𝐴1. Three sets of scenarios result from the quency sections first so to yield better values of the objective function.
combinations of the above values, one set for each value of threshold The approximate solutions obtained by solving problem 𝐿𝑀 and the
𝑄∗ . Each set contains 15 scenarios, one for each budget value from lower bound given by relaxed problem 𝑅𝑀 coincide for all scenarios.
𝑏 = 350 to 𝑏 = 700. Tables 9, 10, 11 show the results obtained by This means that the linearised problem 𝐿𝑀 yields the global optimum
solving problem 𝐿𝑀 for values of 𝑄∗ equal to 0.08, 0.008 and 0.008 for the considered scenarios. This can be explained by the fact that the
respectively. The tables detail the strategy selected for each SRS, the unavailability of all lines is of a bigger order than the corresponding

9
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Table 9
Results for set 𝑆𝐴21 .
Set 𝑆𝐴21 : 𝑄∗1 = 𝑄∗2 = 𝑄∗3 = 𝑄∗4 = 𝑄∗5 = 0.08
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – s1 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – s1 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
03 – s1 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
04 – s1 s1 s2 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – s1 s1 s1 s1 s2 s1 s2 s2 s3 s3 s3 s3 s3 s3
06 – s1 s1 s1 s1 s1 s2 s2 s2 s3 s3 s3 s3 s3 s3
07 – s1 s1 s1 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
𝑧∗𝐿𝑀 – 2.2 0.58 0.4 0.22 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – 2.2 0.58 0.4 0.22 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝜖+ % – 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Table 10
Results for set 𝑆𝐴22 .
Set 𝑆𝐴22 : 𝑄∗1 = 𝑄∗2 = 𝑄∗3 = 𝑄∗4 = 𝑄∗5 = 0.008
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – – s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – – s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
03 – – s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
04 – – s1 s1 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – – s1 s1 s1 s1 s1 s2 s2 s3 s3 s3 s3 s3 s3
06 – – s2 s2 s2 s2 s2 s2 s2 s3 s3 s3 s3 s3 s3
07 – – s1 s1 s1 s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
𝑧∗𝐿𝑀 – – 1.255 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – – 1.255 0.535 0.265 0.175 0.103 0.058 0.031 0.022 0.022 0.022 0.022 0.022 0.022
𝜖+ % – – 0 0 0 0 0 0 0 0 0 0 0 0 0

Table 11
Results for set 𝑆𝐴23 .
Set 𝑆𝐴23 : 𝑄∗1 = 𝑄∗2 = 𝑄∗3 = 𝑄∗4 = 𝑄∗5 = 0.0008
Budget 𝑏
SRS 350 375 400 425 450 475 500 525 550 575 600 625 650 675 700
01 – – – – – s2 s3 s3 s3 s3 s3 s3 s3 s3 s3
02 – – – – – s2 s2 s3 s3 s3 s3 s3 s3 s3 s3
03 – – – – – s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
04 – – – – – s2 s2 s2 s3 s3 s3 s3 s3 s3 s3
05 – – – – – s3 s2 s2 s2 s3 s3 s3 s3 s3 s3
06 – – – – – s3 s3 s3 s3 s3 s3 s3 s3 s3 s3
07 – – – – – s1 s2 s3 s2 s3 s3 s3 s3 s3 s3
𝑧∗𝐿𝑀 – – – – – 0.301 0.1255 0.0805 0.0355 0.022 0.022 0.022 0.022 0.022 0.022
𝑧∗𝑅𝑀 – – – – – 0.301 0.1255 0.0805 0.0355 0.022 0.022 0.022 0.022 0.022 0.022
𝜖+ % – – 0 0 0 0 0 0 0 0 0 0 0 0 0

error 𝐸𝑄𝑙 |𝑚𝑎𝑥 (see Table 4). In Fig. 6 the optimal values of the objective with no redundancy and higher centrality so to satisfy the availability
𝑖
function obtained in scenario analysis 𝑆𝐴2 are compared. constraints. When a feasible solution is still to be found, the impact
From Fig. 6 it is possible to deduce the following: on the objective function which depends on the train frequency, is of
secondary importance. Once a feasible solution satisfying both budget
• for fixed 𝑄∗ the optimal number of trains affected by a speed and availability constraints is found, then the remaining budget is used
restriction decreases for increasing values of the available budget; to improve the strategy selected for SRSs with higher train frequency.
• for fixed budget, more restrictive availability thresholds lead to
higher numbers of trains affected by a speed restriction. 5.3. Sensitivity analysis

This last result might seem contradictory but it is in fact not. An In this section the sensitivity of the optimal solution with respect to
explanation to this behaviour can be found by looking at the network the unavailability threshold 𝑄∗𝑙 for each line is analysed. Parameter 𝑄∗𝑙
topology. The SRSs with higher train frequency are also the ones with is decreased from 0.08 to 0.0008 one line at a time, while the threshold
higher levels of redundancy as they all contain two, three and four for the remaining lines is kept at 0.08. Results are shown in Fig. 7.
parallel tracks. When the availability target is made more restrictive The purpose of this analysis is to identify the most stringent con-
(lower values of 𝑄∗ ), better strategies are needed first for those SRSs straint(s), namely the lines which are most sensitive to an increase

10
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

Fig. 7. Optimal solutions 𝑧∗𝐿𝑀 when unavailability threshold is varied one line at a time, and for different budget values.

in the availability requirement. These will be the lines for which constraint on the availability of line 𝑙5 is always stringent. This is
an increase in the availability target will determine a change in the because SRS 06 is single track, thus constituting a bottleneck for the
optimal solution. Figs. 7(a), 7(b), 7(c), 7(d) and 7(e) show the optimal network. Indeed, if any section of track within SRS 06 fails, than the
solution 𝑧∗𝐿𝑀 plotted against the 15 values of budget from 𝑏 = 350 entire line will be unavailable. The maximum level of availability that
to 𝑏 = 700, and obtained by varying the unavailability threshold of can be ensured for each line is of the same order as the minimum
line 𝑙1 , 𝑙2 , 𝑙3 , 𝑙4 and 𝑙5 respectively. From Figs. 7(a), 7(b) and 7(d) it availability of the SRS with lowest redundancy. The insight obtained
can be seen that no change in solution is registered when decreasing through this sensitivity analysis can be useful to the infrastructure
the unavailability threshold for lines 𝑙1 , 𝑙2 and 𝑙4 . Fig. 7(c) shows manager when decisions on a redistribution of maintenance resources
how the optimal solution changes only when the threshold on line across different SRSs has to be made.
𝑙3 is decreased from 0.008 to 0.0008 and for budget lower than 𝑏 = The scenario and sensitivity analysis have shown how the level of
425. Finally, Fig. 7(e) shows a change in the optimal solution when redundancy and the centrality of the sections determine whether the
decreasing the unavailability threshold for line 𝑙5 from 0.08 to 0.008, affordable solutions are actually feasible for the availability targets. The
and again from 0.008 to 0.0008. Results of this sensitivity analysis more restrictive these availability targets are, the bigger the influence
show that when varying the budget and the threshold values 𝑄∗𝑙 the of redundancy and centrality of each section. Then, as long as both

11
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

budget and availability constraints are satisfied, better strategies are Acknowledgements
applied first to sections with higher train frequency. The maximum
level of availability that can be ensured in each line is of the same order John Andrews is the Network Rail Professor of Infrastructure As-
of the maximum availability of the section with lower redundancy. set Management and Director of Lloyd’s Register Foundation (Lloyd’s
Single track sections constitute a bottleneck for the network because Register Foundation supports the advancement of engineering-related
any failure here would result in the entire line shutting down. education, and funds research and development that enhances safety
of life at sea, on land and in the air) Resilience Engineering Re-
search Group at the University of Nottingham. The authors gratefully
6. Conclusions acknowledge the support of these organisations.

This paper presents a nonlinear integer-programming model and References

a tailored solution approach to support the selection of maintenance
strategies for a railway network. The logic behind the assignment of [1] Petchrompo S, Parlikad A. A review of asset management literature on multi-asset
systems. Reliab Eng Syst Saf 2019;181:181–201.
strategies to track sections makes use of the network segmentation and
[2] de Jonge B, Scarf P. A review on maintenance optimisation. European J Oper
sections aggregation approach used in practice for strategic planning Res 2020;3(285):805–24.
purposes. Strategies are selected which collectively minimise the im- [3] Keizer MO, Flapper S, Teunter R. Condition-based maintenance policies for
pact of section conditions on service, given network availability and systems with multiple dependent components: A review. European J Oper Res
2017;261:405–20.
budget constraints. Different metrics related to the network topology, [4] Nicolai R, Dekker R. Optimal maintenance of multi-component systems: A
sections’ availability, service frequency, performance requirements and review. In: Complex system maintenance handbook. Springer series in reliability
maintenance costs, are combined into a quantitative approach with a engineering, London: Springer; 2008.
holistic view and a clear and simple structure rather than a black box. [5] Yeo H, Yoon Y, Madanat S. Algorithms for bottom up maintenance optimization
for heterogeneous infrastructure systems. Struct Infrastruct Eng 2012;1:1–12.
The approach is simple yet effective to model large networks and makes [6] Furuya A, Madanat S. Accounting for network for effects on railway asset
use of standard solvers, thus facilitating its implementation in practice. management. J Transp Eng 2013;139:92–100.
Based on the Rare Event Approximation principle, both an ad hoc heuris- [7] Robelin CA, Madanat SM. Reliability-based system-level optimization of bridge
maintenance and replacement decisions. Transp Sci 2008;42(4):1–6.
tic solution and relaxation methods are developed. The latter enables
[8] Robelin C, Madanat S. History-dependent bridge deck maintenance and re-
the quality of the heuristic solution to be estimated by calculation of placement optimization with markov decision process. J Infrastruct Syst
the upper bound to the percentage error. With this information the 2007;3(13):195–201.
decision maker can decide whether the current approximate solution is [9] Verbert K, Schutter BD, Babuška R. Timely condition-based maintenance planning
for multi-component systems. Reliab Eng Syst Saf 2017;159:310–21.
acceptable or a better solution needs to be searched for. Furthermore it
[10] Shi Y, Xiang Y, Xiao H, Xing L. Joint optimization of budget allocation and
enables to quantify the performance of the proposed heuristic. To show maintenance planning of multi-facility transportation infrastructure systems.
the potential for implementation to real world scenarios, the model European J Oper Res 2021;288:382–93.
and its relaxed counterpart have been implemented on an illustrative [11] Lethanh N, Adey B, Fernando D. Optimal intervention strategies for multiple
objects affetcted by manifest and latent deterioration. Struct Infrastruct Eng
example considering a portion of the UK railway network, and solved
2015;11(3):389–401.
for a number of scenarios where the available budget and the thresholds [12] Burkhalter M, Martani C, Adey B. Determination of risk-reducing intervention
on lines availability are varied. This scenario analysis gives insight on programs for railwy lines and the significance of simplification. J Infrastruct
the robustness of the solution with respect to the model parameters, as Syst 2018;24:1:04017038.
[13] Burkhalter M, Adey B. A network flow model approach to determining opti-
well as on the relative and combined contribution of different factors mal intervention programs for railway infrastructure networks. Infratsructures
on the selection of strategies, thus enabling more informed decisions. 2018;3(3):31.
These factors include track redundancy, centrality of sections, service [14] Burkhalter M, Adey B. Modelling the complex relationship between interven-
frequency, budget and lines availability targets. tions, intervention costs and the service provided when evaluating intervention
programs on railway infrastructure networks. Infratsructures 2020;5(12):113.
The infrastructure manager can leverage on this model to explore [15] Budai G, Huisman D, Dekker R. Scheduling preventive railway maintenance
how a portfolio of strategies can affect lines availability and service, activities. J Oper Res Soc 2006;57(9):1035–44.
and how performance requirements can best be achieved by an op- [16] Horenbeek AV, Pintelon L. A dynamic predictive maintenance policy for
multi-component systems. Reliab Eng Syst Saf 2013;120:39–50.
timal distribution of the maintenance budget across different railway
[17] Van PD, Barros A, Bérenguer C, Bouvard K, Brissaoud F. Dynamic grouping
sections. Due to its simple structure, the model can be easily adapted maintenance with time limited opportunities. Reliab Eng Syst Saf 2013;120:51–9.
to address similar problems arising in the maintenance of other types [18] Nguyen K, Van PD, Grall A. Multi-level predictive maintenance of multi-
of networks, such as road and bridges networks, where long-term component systems. Reliab Eng Syst Saf 2015;144:83–94.
[19] Van PD, Vu HC, Barros A, Bérenguer C. Maintenance grouping for multi-
maintenance strategies must be chosen for individual or groups of component systems with availability constraints and limited maintenance teams.
road segments or bridges respectively, when deciding on the strategic Reliab Eng Syst Saf 2015;142:56–67.
allocation of maintenance efforts. [20] Caetano L, Teixeira P. Availability approach to optimizing railway track renewal
operations. J Transp Eng 2013;139:941–8.
[21] Chen S, Ho T, Mao B. Maintenance schedule optimisation for a railway power
CRediT authorship contribution statement supply system. Int J Prod Res 2013;51:4896–910.
[22] Chen S, Ho T, Mao B, Bai Y. A bi-objective maintenance scheduling for power
feeding substations in electrified railways. Transp Res C 2014;44:350–62.
Claudia Fecarotti: Conceptualization, Methodology, Software, Val- [23] Andrade AR, Teixeira PF. Biobjective optimisation model for maintenance and
idation, Writing – original draft, Writing – review & editing, Visu- renewal decisions related to rail track geometry. Transp Res Rec J Transp Res
Board 2011;2261:163–70.
alization. John Andrews: Writing – review & editing, Visualization.
[24] Podofillini L, Zio E, Vatn J. Risk-informed optimisation of railway tracks
Raffaele Pesenti: Writing – review & editing, Visualization. inspection and maintenance procedures.. Reliab Eng Syst Saf 2006;91:20–35.
[25] Marseguerra M, Zio E, Podofillini L. Condition-based maintenance optimization
by means of genetic algorithms and Monte Carlo simulation. Reliab Eng Syst Saf
Declaration of competing interest 2002;77:151–66.
[26] Zhao J, Chao AHC, Stirling A, Madelin K. Optimising policies of railway ballast
tamping and renewal. Transp Res Rec 2006;1943:50–6.
The authors declare that they have no known competing finan-
[27] Zhu Q, Peng H, Timmermans B, van Houtum G. A condition-based maintenance
cial interests or personal relationships that could have appeared to model for a single component in a system with scheduled and unscheduled
influence the work reported in this paper. downs. Int J Prod Econ 2017;193:365–80.

12
C. Fecarotti et al. Reliability Engineering and System Safety 216 (2021) 107940

[28] Liesiöa J, Salo A, Keisler JM, Morton A. Portfolio decision analysis: Recent [36] Fecarotti C, Andrews J. A petri net approach to assess the effects of railway
developments and future prospects. European J Oper Res 2021;293:811–25. maintenance on track availability. Infrastruct Asset Manag 2020;3(7):201–20.
[29] Sacco T, Compare M, Zio E, Sansavini G. Portfolio decision analysis for risk-based [37] Bai L, Liu R, Sun Q, Wang F, Xu P. Markov-based model for the prediction of
maintenance of gas networks. J Loss Prev Process Ind 2019;60:269–81. railway track irregularities. Proc Inst Mech Eng F 2013;229(2):150–9.
[30] Compare M, Bellani L, Zio E. Optimal allocation of prognostics and health [38] Prescott D, Andrews J. Investigating railway track asset management using a
management capabilities to improve the reliability of a power transmission Markov analysis. Proc Inst Mech Eng F 2015;229(4):402–16.
network. Reliab Eng Syst Saf 2019;184:164–80. [39] Skinner AKM, Williams J. Challenges of developing whole life cycle cost models
[31] Fecarotti C, Andrews J. A mathematical programming approach to railway for network rail top 30 assets. In: IET and IAM asset management conference.
network asset management. In: Haugen, et al., editor. Proceedings of ESREL London, UK: 2011.
conference: Safety and reliability - Safe societies in a changing world. Trondheim, [40] Track Asset Policy, Tech. rep.. Network Rail; 2012.
Norway; 2018. [41] Electrical power asset policy. Internal network rail documentation. Tech. rep,
[32] David R, Alla H. Discrete, continuous and hybrid petri nets. Springer Science & Network Rail; 2012.
Business Media; 2010. [42] Route specifications: London North Eastern and East Midlands, Tech. rep..
[33] Andrews J, Prescott D, De Rozieres F. A stochastic model for railway track asset Network Rail; 2018.
management. Reliab Eng Syst Saf 2014;130:76–84. [43] Andrews JD, Moss TR. Reliability and risk assessment. Professional Engineering
[34] Le B, Andrews J, Fecarotti C. A petri net model for railway bridge maintenance. Publishing; 2002.
Proc Inst Mech Eng Part O 2017;3(231):306–23. [44] Hillier FS, Lieberman GJ. Introduction to operations research. 9th ed..
[35] Fecarotti C, Andrews J. Producing an effective maintenance strategy to control McGrow-Hill Higher Education; 2009.
railway risk. In: Proceedings of probabilistic safety assessment and management.
UCLA Los Angeles; 2018, p. 16–21.

13