Decision Support Systems 55 (2013) 126–134

Contents lists available at SciVerse ScienceDirect

Decision Support Systems

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

ELECTRE III to dynamically support the decision maker about the periodic
replacements configurations for a multi-component system
Antonella Certa ⁎, Mario Enea, Toni Lupo
Dipartimento di Ingegneria Chimica, Gestionale, Informatica, Meccanica, Università degli Studi di Palermo, 90128 Palermo, Italy

a r t i c l e i n f o a b s t r a c t

Article history: The problem tackled by the present paper concerns the selection of the elements of a repairable and stochastically
Received 7 May 2012 deteriorating multi-component system to replace (replacements configuration) during each scheduled and
Received in revised form 23 October 2012 periodical system stop within a finite optimization cycle, by ensuring the simultaneous minimization of both the
Accepted 30 December 2012
expected total maintenance cost and the system unavailability. To solve the considered problem, a combined
Available online 17 January 2013
approach between multi-objective optimization problem (MOOP) and multi-criteria decision making (MCDM)
Keywords:
resolution techniques is proposed. In particular, the ε constraint method is used to single out the optimal Pareto
Multi-objective optimization frontier whereas the ELECTRE III multi-criteria decision support method is proposed to support the selection
ELECTRE III, periodic maintenance policy of the replacements configuration that represents the best compromise among the considered objectives. The
Multi-component system proposed approach is sequentially applied at each scheduled system stop by allowing the dynamic updating of
Non-homogeneous Poisson process the information about the decisional context in which the decision maker has to operate. To illustrate the
whole procedure a numerical case study is solved for different hypothesized scenarios related to the importance
attributed by the decision maker to the system unavailability and the maintenance cost objectives.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction interventions for each element. Caldeira, Duarte et al. [4] propose a
mathematical programming model for a series system, with elements
A particular kind of preventive maintenance policy for multi- characterized by a Weibull hazard function, to calculate the optimum
component systems is the periodic one. In such a maintenance policy, frequency to perform preventive maintenance actions for each element.
each system element can be replaced during scheduled and periodic The aim is the minimization of the maintenance cost so that the total
stops of the system disregarding of its failure history and repaired downtime, in a certain period of time, does not exceed a fixed value.
when the failure occurs. The periodic maintenance policy represents Certa et al. [10] formulate a constrained mathematical model to solve
for its easy implementation, especially in large and complex multi- the problem of determining both the optimal number of periodic
component systems, the most used maintenance policy [23]. For exam- inspections within a finite time frame and the system elements to
ple, systems such as power plants, gas turbine plants, and petroleum replace during each scheduled inspection. Furthermore, the mainte-
refineries are commonly subjected to scheduled and periodic stops dur- nance policy has to ensure a very high system reliability level with the
ing which preventive maintenance can be eventually performed and, in minimum total maintenance cost. Taghipour et al. [20] develop a
addition, the potential and hidden failures are detected and repaired. model able to find out the optimal periodic inspection intervals over a
In the literature, in order to determine the parameters characteriz- finite time horizon considering a system composed of elements
ing a periodic maintenance policy, different models have been devel- subjected to soft and hard independent failures. Hard failures are
oped considering one or more objectives to reach. The most treated instantaneously detected and fixed while soft failures can be detected
objective in the periodic maintenance policy models is the minimiza- and fixed only at the scheduled inspections. The model takes into
tion of the total maintenance cost. For example Bris et al. [3] develop account as objective to optimize the expected total cost arising from
an optimization model, considering the system elements periodically soft and hard failures incurred over the optimization time horizon and
inspected and maintained, aiming to find out the optimal maintenance a recursive procedure is adopted to solve it. More recently, Taghipour
policy for each element by minimizing the cost function and respecting et al. [19] develop a model aiming to determine the optimal periodicity
an availability constraint. In order to solve the problem, the Authors of the inspection intervals of a reparable system subjected to hidden
propose a genetic algorithm, whose structure includes the inspection failures over finite and infinite optimization time. The proposed objec-
time and the time length between two consecutive maintenance tive function is the minimization of the expected total cost. A failed
system element may be replaced or minimally repaired with a probability
⁎ Corresponding author.
dependent on the elements age. Recursive procedures are developed to
E-mail addresses: [email protected] (A. Certa), [email protected] (M. Enea), calculate the probabilities of failures, the expected number of minimal
[email protected] (T. Lupo). repairs and the expected downtime of the system.

0167-9236/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.dss.2012.12.044
 A. Certa et al. / Decision Support Systems 55 (2013) 126–134 127

Considering production systems, a suitable parameter for evaluat- The most widely used approaches to describe the Pareto frontier are
ing the maintenance policy effectiveness is its stationary availability, the weighted sum method and the ε constraint method [12]. The
also defined as the expected time percentage in which the system is greatest difficulty in applying the first one consists of setting suitable
functioning. Therefore, for this kind of systems, also the stationary weight vectors to obtain optimal non-dominated solutions and more-
availability maximization represents another objective to be reached. over it is able to describe only the convex region of the frontier. To the
Vaurio [22] studies the average unavailability of standby units under contrary, the second approach overcomes the limitations of the previ-
several sets of assumptions concerning the renewal efficiency of tests ous one and it allows the complete description of the Pareto frontier.
and repairs. Cassady et al. [7] tackle the problem to single out the set In any case, the determination of just a single point of the frontier
of elements on which to operate at periodic scheduled breaks, in using these methods requires the resolution of a mathematical pro-
order to maximize the system availability and to respect a mainte- gramming problem. For the previous reasons the ε constraint method
nance cost constraint. The maintenance activities must be completed is herein chosen to describe the optimal Pareto frontier at each system
within an allotted time. The problem is formulated by a mathematical stops. Also Certa et al. [9] propose the use the ε constraint method to
model and solved for two simple systems. Also Tsai et al. [21] develop find out the optimal solutions in a multi-objective context related to a
an algorithm to find out an optimal maintenance policy for a medium-term scheduling of periodic maintenance actions.
multi-component system. In particular, the Authors consider a pre- In order to select the best maintenance strategy the Analytic Hierar-
ventive maintenance policy that takes simultaneously into account chy Process (AHP) multi-criteria method [18] is the decision support
three actions: the mechanical service, aiming at alleviate the strength method more considered in the literature. Bevilacqua and Braglia [2]
degradation; the repair activity, addressed to partially restore a de- propose the use of the AHP for selecting the best maintenance strategy
graded component and the replacement activity, settled to recover a among preventive, predictive, condition-based, corrective and opportu-
component to its original condition. For each element the preventive nistic maintenance alternatives for an important Italian oil refinery.
maintenance intervals are investigated on the base of the system Carnero [6] propose a model that carries out the decision regarding
availability maximization and the minimum interval value is chosen the selection of the diagnostic techniques and instrumentations in the
for programming the periodic maintenance policy. predictive maintenance programs. The proposed resolution approach
In the light of the previous considerations, the expected total main- consists of a combination of tools belonging to the MCDM techniques,
tenance cost and the system unavailability are herein considered as just like the AHP and the factor analysis (FA). Pariazar et al. [16] suggest
objective functions to be optimized in order to develop a periodic main- the use of the AHP to select the maintenance strategy by considering
tenance policy for a multi-component system over a finite optimization cost, safety, execution capability as criteria of evaluation. Given the
cycle. Moreover, since the cost and the system unavailability objectives high number of considered criteria, they propose the AHP to reduce
could take a different relative importance depending on the operative the inconsistency of the judgments of comparison that affects this
context in which the decision maker has to operate, the problem is for- methodology. Recently, Arunraj and Maiti [1] propose an approach
mulated and faced by means of tools belonging to the multi-objective based on the AHP and goal programming for the maintenance policy se-
optimization problem (MOOP) and the multi-criteria decision making lection and they show a case study related to a benzene extraction unit
(MCDM) techniques. In particular, being the two objective functions of a chemical plant. Papakostas et al. [15] propose a methodology to
contrasting one each other, it is not possible to find out a single optimal support the decision maker about maintenance actions to carry out on
solution corresponding to the best result for both the considered objec- an aircraft, aiming at high fleet operability and low maintenance cost.
tives but an optimal Pareto frontier, that is a set of non-dominated In particular, a multi-criteria mechanism evaluates a set of maintenance
trade-off solutions, can be described. Each optimal solution represents plan alternatives.
an optimal replacements configuration for the related objectives The AHP is a compensative method, that is, with relation to a solu-
trade-off level. Clearly, the optimization procedure should generate a tion, a very bad score value in an objective can be compensated by
number of non-dominated solutions among which the decision maker high score values in the other objectives. This compensation feature
should select the most attractive one. Therefore, as stressed by could not be approved by the decision maker in the case in which
Campanella [5], it is important to support the decision maker in the the scores difference overcomes a value of threshold fixed by she/
choice of the solution that represents for she/he the best compromise he. On the contrary, the ELECTRE III [17] is a non-compensative
among the considered objectives. For such motivation an efficient com- MCDM method and thus it is herein proposed as decision support
bined procedure based on integration between the MOOP and the method to select the solution that represents the best compromise
MCDM techniques is herein proposed. among the considered objectives.
In the literature, in order to describe the Pareto frontier different The present paper is organized as follows: in the next section the
approaches, with relation to maintenance problems, have been pro- problem description is given; in Section 3 the mathematical formula-
posed. For example heuristic approaches like evolutionary algorithms tion and the solution approach of the multi-objective problem is
have been developed. Those approaches often imply high computa- shown; in Section 4 the ELECTRE III method for the selection of the
tional time and they do not ensure the determination of the overall best maintenance strategy is presented; in Section 5 a numerical anal-
Pareto frontier. Hilber et al. [13] propose a multi-objective model to ysis related to a case study is presented and the related results are
maximize customer satisfaction and minimize maintenance costs for commented; finally, the conclusions close the work. Furthermore,
a power delivery system. The aim is the determination of the optimal all the steps necessary to execute the proposed combined procedure
trade-off between preventive and corrective maintenance. Chang et related to the case study are reported in the Appendix A.
al. [11] develop a multi-objective optimization model in order to
carry out a proper preventive maintenance scheduling for a power 2. Problem description
supply substation that is able to provide an effective trade-off be-
tween reliability and maintenance cost. In particular, in order to As mentioned before, it is herein considered a reparable and sto-
solve the considered problem, they propose a multi-objective genetic chastically deteriorating multi-component system consisting of series
algorithm and the minimum cut set method. Certa et al. [8] propose components which can be composed by a single element or by series
an exact algorithm addressed to the determination of the set of elements arranged in branches. In the first case, a failure (hard fail-
elements on which maintenance actions must be executed so as the ure) leads to the system downtime, whereas a failure of an element
system operates with a required reliability until the next scheduled (soft failure) of a component composed by series elements arranged
inspection, minimizing both the maintenance total cost and the in branches does not imply the system failure. It is considered that
total maintenance time. during the system operative time both the hard and the soft failures
128 A. Certa et al. / Decision Support Systems 55 (2013) 126–134

are immediately detected and minimally repaired, during the system computational time that rapidly grows up at increasing of the number
downtime for the hard type or while the system is working for the of considered stops.
soft type. Instead, the replacement of system elements can be Successively, among the Pareto frontier solutions previously found
performed only during the scheduled and periodic system stops, out, the choice of the replacements configuration for the stop j=1,
whose duration has not been greater than the allotted time Ta. In that represents the best compromise between the cost and unavailability
fact, as in many real systems, such as chemical processing facilities, objectives, is made by means of the ELECTRE III method.
and power plants, the replacement of system elements can be The optimization procedure will be again executed at t = tp for the
performed only during the scheduled system stops and within the al- stop j = 2 by taking into account the replacements activities performed
lotted time Ta. In fact, the presence of some operative and/or some on the system during the previous one j = 1 and considering one more
technical constraints can make economic disadvantage or even im- time the optimization time frame T. In such a way the optimal replace-
possible to perform the maintenance actions at any time [7]. For ex- ments configurations set for the scheduled stop j = 2 is singled out and,
ample in many industrial systems, in order to assure the respect of one more time again, among which the selection of the replacements
the production scheduling related to a given time interval, the dura- configuration for the stop j = 2 is made by means of the ELECTRE III.
tion of the maintenance activities, at each scheduled system stop, Such optimization procedure will be sequentially repeated until the
has not overcome the allotted time. Also the value of the interval tp, last stop j = (Q-1) will be reached.
that is the time between two consecutive stops, arises from aspects Finally, at the end of the optimization cycle j = Q, it may be
linked to functional features of the system that, considering large assumed that a major system overhaul will be performed and a new
and complex multi-component systems, is generally fixed for the cycle will start again, then a new optimization procedure should be
whole lifetime of the system. applied.
Considering a finite optimization cycle of length T*, that could be
equal, for example, to the period covered by a contract of mainte- 3. Mathematical problem formulation
nance services for the providing of services for T*/tp number of sched-
uled system stops, the aim of the proposed approach is to single out The following basic nomenclature is hereafter used.
the replacements configuration, that is the selected system elements i = 1,2,…,N index representing a generic element;
to replace during each scheduled stop. b = 1, 2, …Bi index representing a generic branch of the component i;
The proposed optimization methodology consists of splitting the so- k = 0,1,…,(P-1) index representing the last stop in which the
lution procedure into two phases: the first step consists of obtaining the replacement of the generic element has been performed. k takes
set of non-dominated trade-off solutions (Pareto optimal solutions) value equal to zero if the generic element has never been replaced
within the whole space of the feasible ones. In the second step, the so- until the current stop;
lutions belonging to the Pareto frontier are evaluated and compared xi,j binary variable taking 1 if the element i is replaced at the stop j.
in order to select the best one. The two-step approach is sequentially re- In order to obtain the optimal replacements configurations set,
peated at each system stop, since particular aspects as further informa- that is the Pareto frontier solutions for each system stop, the two
tion about the decisional context, or particular needs, or knowledge of following objective functions have to be optimized:
practicable solutions, could be provided to the decision maker only in
a second phase. So, it will be easier to select the solution representing min C ð1Þ
the best compromise for all the considered objectives by choosing it
from a restricted set of the non-dominated ones at the moment in min U ð2Þ
which the decision has to be made. In such way, the proposed optimiza-
where C and U are the functions of the expected values of the total
tion approach is configured as a dynamic optimization method in the
maintenance cost and the system unavailability respectively over
time domain.
the time frame value T that covers (P-1) system stops.
In order to explain how the developed approach works, Fig. 1 shows
The cost function C can be expressed as:
the sequential optimization approach. In particular, j = 1 is the current
instant in which the optimization procedure is started. The optimal re- X
P −1  
placements configurations set, that is the Pareto frontier solutions, for C¼ C j þ C m;j ð3Þ
the first scheduled stop j =1 is singled out in order to minimize both j¼1
the expected total maintenance cost and the system unavailability in
an optimization time frame T that includes a certain number of next sys- where Cm,j is the cost of the minimal repair activities performed in the
tem scheduled stops. The choice of the number of the scheduled system operative time between the stop j and the next one j + 1 and Cj is the
stops is made in order to reach a good trade-off between the goal of cost arising from the replacement of the selected elements performed
extending the optimization temporal horizon to obtain solutions closer during the stop j. The latter is given by the following relationship:
to the long-term optimal ones and the goal to assure acceptable
N 
X 
Cj ¼ cri þ C cp ⋅t ri ⋅xi;j ð4Þ
i¼1

T
t=(Q-1).tp with, cri, Ccp and tri the spare parts cost, the unit time cost of the main-
tenance crew for replacement activities and the replacement time of
the element i respectively.
In order to evaluate the duration of each system stop, it is herein
T assumed that the maintenance crews can operate in parallel during
t=tp
the replacement activities. That is, the maintenance crews can oper-
t=0 T
ate simultaneously each one on a different element belonging to the
elements set to be replaced. In [14] an example of such strategy,
j=1 j=2 j=P-1 j=P j=Q-1 j=Q that allows the reduction of the system stops duration, is described
T* with relation to the maintenance global service contract between a
maintenance services provider and the public company of the waste
Fig. 1. Sequential optimization approach. management of the City of Palermo (Italy).
 A. Certa et al. / Decision Support Systems 55 (2013) 126–134 129

In such condition, considering the stop j, the longest replacement mi(t), considering the time period [0,t], is given by the following
time among those of the elements replaced determines the duration equation:
Tj of the considered system stop. Such condition is ensured by the
t t
following constraint: f i ðt Þ
mi ðt Þ ¼ ∫ λi ðt Þ⋅dt ¼ ∫ dt ¼ − ln Ri ðt Þ ð11Þ
n o 0
R
0 i
ðt Þ⋅
T j ¼ max t ri ⋅xi;j ∀j; ∀i ð5Þ
being λi(t), fi(t) and Ri(t) the failure rate, the probability density
Moreover, in the considered optimization cycle of length T* it is function and the reliability of the considered element respectively.
supposed that the duration of each scheduled system stop has to be The reliability of the element i, Ri,j(t) calculated at the generic
less or equal to the allotted time Ta. Therefore the following constraint scheduled stop j until the next one j + 1, can be expressed as:
has to be satisfied: j−1   
X
Ri;j ¼ r i;j;j ⋅xi;j þ r i;j;k ⋅zi;j;k ⋅ 1−xi;j ∀k=0≤kbj ð12Þ
Tj ≤ Ta ∀j; ∀i ð6Þ k¼0

The constraint (6) implies that, in the considered optimization Whenever xi,j is equal to 1, i.e. the element i is replaced during the
cycle, the system elements characterized by a replacement time stop j, its reliability value is given by ri,j,j(t). On the contrary, if xi,j is
greater than the allotted time Ta cannot be replaced. equal to zero, therefore the replacement of the element i has been
In the Eq. (3), Cm,j is the cost of the minimal repair activities performed during the stop k, with k b j, assuming k = 0 if the element i
performed in the operative time between the stop j and the next has not been replaced until the stop j. In such condition the reliability
one j + 1. It includes the costs of consumed materials, the allocated value of the element i is given by ri,j,k(t).
maintenance crews and, only in the case of hard failures, also those According to the theory of the probability, the reliability value ri,j,j(t)
arising from the system downtime. in Eq. (12) is expressed by:
Instead, the cost of the minimal repair activities performed on the    
system components composed of series elements arranged in branches r i;j;j ¼ Ri;j t jþ1 −t j ¼ Ri;j t p ∀j ð13Þ
does not include the system downtime cost, since the failure on an own
element (soft failure) does not imply the system failure. whereas the reliability value ri,j,k(t), is given by the following
In particular, considering a generic component iz, belonging to the equations:
set Z of the system components composed of series elements ar-    
ranged in branches constituted of Bi branches (whose generic branch Ri;j t jþ1 −t k Ri;j ðj þ 1−kÞ⋅t p
b is composed of eb series elements), the related cost for the minimal r i;j;k ¼   ¼   ∀j; ∀k j k≠0 ð14Þ
Ri;j t j −t k Ri;j ðj−kÞ⋅t p
repairs Cm,iz,j is given by the following equation:

Bi X
X eb   and
C m;iz ;j ¼ mk;j ⋅ cm;k þ t m;k ⋅C cm ∀k∈iz ð7Þ
   
b¼1 k¼1
Ri;j t jþ1 Ri;j j⋅t p þ t start;i
r i;j;k ¼   ¼   ∀j and with k ¼ 0 ð15Þ
with, mk,j, cm,k and tm,k the expected number of failures, the cost of Ri;j t j Ri;j ðj−1Þ⋅t p þ t start;i
the consumed materials related to the minimal repair activities and
the minimal repair time of the element k respectively and Ccm, as be-
where tstart,i is the age of the element i at the beginning of the optimiza-
fore said, the unit time cost of the maintenance crew for minimal
tion procedure. The previous relationships do not take into consideration
repair.
the system downtime caused by hard failure since the latter is negligible
To the contrary, if the component i belongs to the set S of the sys-
with respect to the interval tp.
tem components composed by a single element, in addition the cost
By considering the Weibull probability distribution, with the shape
arising from the system downtime, that is CDT, has to be considered
parameter βi >1 and the scale parameter ηi >0, the relations (13), (14)
in the expression of its minimal repairs cost:
and (15) become:
C m;is ;j ¼ tcm;i s þ ðC DTm þ C cm Þ⋅t m;is b⋅mis ;j ð8Þ   β 
tp i
r i;j;j ¼ exp − ∀j ð16Þ
ηi
Therefore, the cost of the minimal repair activities performed in
the operative time between the stop j and the next one j + 1 Cm,j is h   i
ðjþ1−kÞ⋅t p βi
exp − ηi
given by:
r i;j;k ¼ h   i ∀j; ∀k j k≠0 ð17Þ
ðj−kÞ⋅t p βi
exp − η
C m;j ¼ ∑ C m;iz ;j þ∑ C m;is ;j ð9Þ i

∀iz ∀is
h  β i
j⋅t þt i
exp − p η start;i
Finally, the function of the expected system unavailability U over r i;j;k ¼ h 
i
β i ∀j and with k ¼ 0 ð18Þ
ðj−1Þ⋅t p þt start;i i
the time frame T can be expressed as: exp − η i

X
P −1
U¼ ∑ t m;is ⋅mis ;j ð10Þ Finally, the term zijk in Eq. (12) is a binary dummy variable that makes
j¼1 ∀is mutually exclusive the two conditions previously described, i.e. the two
terms of the expression 12 related to the instant in which the element
3.1. Expected number of failures and reliability calculation has been replaced, that is during the current stop j or during a previous
stop. The previous requirements are fulfilled by the following condition:
The expected number of failures of the system elements, under
the hypothesis of minimal repair of the hard and the soft failures,
j  
follows a non-homogeneous Poisson process (NHPP) with a power zi;j;k ¼ xi;k ⋅ ∏ 1−xi;n ∀k=0 ≤ k b j ð19Þ
law intensity function. In such a case, its value for the element i, n¼kþ1
130 A. Certa et al. / Decision Support Systems 55 (2013) 126–134

In such way, the following conditions are satisfied: relation is very important: a solution outranks another solution if
sufficient reasons exist to assert that the first is as good as the second
• When xij takes value equal to 1, zijk is equal to zero for any k; and good reasons to reject such assertions do not exist. The outranking
• When xij is equal to zero, zijk is equal to 1 only for the stop k in relation is based upon a concordance/discordance principle. This princi-
which the last replacement of the element i has been performed. ple consists of the verification of the existence of a concordance of
criteria in favor of the assertion that one solution is as good as another
3.2. Description of the Pareto frontier and that a verifiably strong discordance among the score values that
may reject the previous assertion does not exist.
In order to find out the extreme Pareto optimal solutions, the Lexico- For each considered criterion, in our case cost and system un-
graphic Goal Programming (LGP) method is initially used (Deb, 2001). availability, the following thresholds are introduced:
This method separately considers the two objective functions, thereby
reducing the multi-objective problem into a mono-objective one. • q indifference threshold;
For example, these sequential steps determine the extreme solution of • p preference threshold;
minimum cost: • v veto threshold.
• Minimizing C as a single objective problem obtaining C min;
where q b p b v.
• Minimizing U, by imposing the value of C not greater than C min,
obtaining U max.
Those threshold values can be expressed in term of percentage of
The procedure is analogously applied changing the objectives hierar- the score differences assumed by the solutions, with respect to the
chy to find the other two bounds Cmax and Umin of the extreme solution worst one under the considered criteria.
of minimum system unavailability. Once the extreme points of the Pareto Therefore, the decision maker is involved to supply his/her prefer-
frontier are determined, the ε constraint method is used to describe the ences by choosing the indifference, preference and veto thresholds for
whole Pareto optimal frontier. By using this method the multi-objectives each objective function. Furthermore, the method allows to assign a rel-
optimization model described in Section 2.1 is modified in a single objec- ative weight to the considered criteria and this fact permits to represent
tive optimization model in which one of the two objectives has to be different scenarios related to the importance attributed by the decision
minimizing and restricting the remaining one within user specified maker depending on the operative context in which she/he operates.
values ε. In particular, to describe the Pareto frontier a multi-step
optimization procedure has to be applied: in the first step, minimizing 5. Case study
C and imposing that the U function has to take a value less than Umax.
In this way an optimal solution belonging to the Pareto frontier can be In order to show how the combined approach works a numerical
found. In the next step the procedure is repeated minimizing C and case study is solved. In particular, it is considered a multi-component
imposing that the U function has to take a value less than the one system that is periodically stopped to perform inspection and replace-
corresponding to the solution previously found. The procedure is ment activities for which the stop interval tp is equal to 800 time
repeated until the other Pareto frontier extreme solution is obtained. In units (T.U.) and the allotted time Ta is equal to 20 T.U. The system is
such way, the ε constraint method ensures the determination of the composed, from a reliability point of view, of seven components
whole Pareto frontier also in the presence of non-convex regions. arranged in series (see Fig. 2). The first component is constituted by
two branches each one composed of two elements; the second compo-
4. Selection of the Pareto solution by means of ELECTRE III nent is a simple parallel component with three branches, while the
other components are constituted by a single element. Totally, the
In the present section the fundamental features that characterize system is composed of 12 elements.
ELECTRE III are given. As said before ELECTRE III (Roy, 1990) [17] is As said before, the reliability behavior of the elements is modeled by
the multi-criteria method proposed for choosing the solutions, which Weibull distributions whose parameters η and β are reported in Table 1.
are the replacements configurations to perform during each system Table 1 also shows the cost data, expressed in monetary units (M.U.),
stop. ELECTRE III is a multi-criteria decision-making method that the time data expressed in time units (T.U.) and the age of the system
reflects the decision maker's preferences and it can be applied when a elements tstart,i at the beginning of the optimization cycle of length
set of alternatives must be ranked according to a set of criteria equal to T*= 2400 T.U. All the data have been randomly generated
conflicting each other or when just the preferred has to be selected. within opportune ranges. For each step, the model is solved by the com-
The method is based upon pseudo-criteria. Using proper thresholds, a mercial software Lingo.
pseudo-criterion takes into account the uncertainty and ambiguity The chosen optimization time frame value T is supposedly equal to
that can affect the evaluation of the objectives, so that, if the difference the optimization cycle of length T* and therefore, the number of
in the performance of two solutions is minimal, according to a certain the scheduled system stops in which to find out the replacements
criterion, such solutions can be considered indifferent with respect to configurations and the number of the stops covered by the time frame
that criterion. Another peculiarity that differentiates ELECTRE III from T is equal to 3 (Fig. 3).
other multi-criteria methodologies as the AHP, is that, as previously
said in the introduction, it is not compensative, which means that a
very bad score in one objective function is not compensated by good 5
scores in the other one. In other words, the decision maker will not 1 3
choose a solution if it is very bad compared to another one, even on …
6 8 12
a single criterion. It is understandable as in the treated case that the
decision maker can put on this kind of behavior since she/he could 2 4
not be unwilling to choose a solution with respect to another one if it 7
presents an objective function value (cost or unavailability) higher of
a certain threshold with respect to the considered objective function
value of the second solution, independently from the objective function
value assumed by the other objective function. This threshold is called
veto threshold. Therefore, in ELECTRE III the concept of outranking Fig. 2. Multi-component system reliability block diagram.
 A. Certa et al. / Decision Support Systems 55 (2013) 126–134 131

Table 1 Table 2
Case study data. Replacements configurations.

Element tstart,i tri tmi Cri Cmi ηi βi j=1 j=2 j=3

1 100.00 9.07 1.57 7.33 1.46 987.96 1.92 Scenario 1: (8,10,12) Scenario 1: (8,10,11,12) Scenario 1: (8,10,11,12)
2 300.00 11.19 1.98 9.22 1.86 2772.96 1.38 Scenario 2: (8,9,10,11,12)
3 300.00 10.82 2.13 7.50 1.71 2591.76 1.23 Scenario 2: (8,10,11,12) Scenario 1: (8,10,11,12)
4 100.00 8.65 3.63 10.20 4.24 883.06 1.95 Scenario 2: (8,9,10,11,12)
5 200.00 17.48 3.35 10.52 2.69 1983.67 1.38 Scenario 2: (8,10,11,12) Scenario 1: (8,10,11,12) Scenario 1: (8,10,11,12)
6 300.00 22.18 3.89 11.82 3.51 1448.87 1.41 Scenario 2: (8,9,10,11,12)
7 100.00 12.17 2.06 7.79 2.01 813.92 1.33 Scenario 2: (8,9,10,11,12) Scenario 1: (8,10,11,12)
8 300.00 18.75 3.63 10.96 3.06 1218.47 1.81 Scenario 2: (8,9,10,11,12)
9 100.00 15.91 2.97 9.05 2.50 3226.68 1.33
10 100.00 11.58 2.21 10.66 1.83 1350.61 1.68
11 200.00 17.63 3.26 11.48 2.71 1955.64 1.20
12 300.00 9.83 1.72 9.35 1.52 1531.56 1.67
CDTm 50.00 cycle, are determined. Table 2 shows all the replacements configurations
Ccm 2.00 obtained for each system stop by reporting the elements to be replaced.
Ccp 1.00 The total cost and the system unavailability values over the opti-
T 2400
mization cycle are shown in the Fig. 4 for all the decisions that can
be made during the considered optimization cycle. In particular,
each point is characterized by a triple of numbers in which the first,
Two scenarios have been hypothesized and each one is linked to the second and the third indicate the scenario hypothesized at the
the relative importance that the decision maker assigns to the objec- stop j = 1, j = 2 and j = 3 respectively.
tives. The relative importance, translated in weights, is chosen equal In Appendix A the sequential procedure applied to solve the case
to Wc = 0.9 and Wu = 0.1 (scenario 1) and Wc = 0.1 and Wu = 0.9 study is reported with more details.
(scenario 2) for the cost C and the system unavailability U respective-
ly, since those values represent extreme decisional contexts, that is
extreme operative needs, in which the decision maker could make 6. Conclusions
the decisions. The other values that the decision maker has to choose
are those of the thresholds. In this case it is assumed that, for each The present paper aims to offer a useful methodological approach
objective, the threshold values are chosen with relation to the range to dynamically support the decision maker in the selection of the
of the objective values of the extreme Pareto solutions analyzed for elements of a multi-component system to replace during each scheduled
the particular stop and scenario. However, the choice of the threshold and periodic stop.
values is explained in more detail in Appendix A. Two objectives are considered to be reached, both to minimize
Referring to Fig. 3, among the Pareto solutions found out for j = 1, over a finite optimization cycle, and in particular they are the
the choice of the replacements configuration is made by means of the expected total maintenance cost and the system unavailability. The
ELECTRE III method. In such way, the two replacements configura- developed approach combines the MOOP and the MCDM techniques.
tions will be chosen with relation to the two possible hypothesized The optimal replacements configurations for each system stop, i.e. the
scenarios. optimal Pareto frontier, are determined considering a finite optimiza-
The optimization procedure will be again executed for the next tion time frame that goes over the considered system stop, so to
stop j = 2. The Pareto solutions are singled out taking into account obtain the optimal solutions closer to the long-term optimal ones.
the two replacements configurations that can be performed on the Among these, the decision maker will choose the one that represents
system at j = 1, considering one more time the optimization time the best compromise between the considered objectives by means of
frame T. The replacements configurations for the stop j = 2 will the ELECTRE III method. The procedure is repeated for each scheduled
be chosen considering one more time again the two possible hypoth- system stop within the considered optimization cycle so as the infor-
esized scenarios. mation about the decisional context can be dynamically updated and
By iterating such procedure until the system stop j=3 is reached, all therefore allowing, at each system stop, the selection of the replace-
the decisions, that can be made during the considered optimization ments configuration most suitable to the decisional context.
The short computational time required by the combined approach
makes the proposed procedure suitable to solve maintenance problems
in real operative contexts in which the analysis focuses on large multi-
component system.

t=1600 T=2400

T=2400
t=800

t=0 T=2400

j=1 j=2 j=3 j=4

T*

Fig. 3. Case study sequential optimization approach. Fig. 4. Total cost and system unavailability values over the optimization cycle.
132 A. Certa et al. / Decision Support Systems 55 (2013) 126–134

Appendix A By applying the ELECTRE III for each scenario the chosen solutions
are those reported in Table A6.
A.1. System stop j = 1

The Pareto solutions are reported in Table A1. Table A6

Selected solutions for j = 2 — scenario 1 at j = 1.

Scenario Chosen solution

Table A1
1 2
Pareto solutions related to j = 1.
2 2
Pareto solution U C Replacements configuration

1 28.63 1857.53 8,12

2 27.66 1909.93 8,10,12
Scenario 2 at the system stop j = 1
3 26.92 2017.70 8,10,11,12
4 26.64 2241.59 8,9,10,11,12 The Pareto solutions are reported in Table A7.

Table A7
The objective values related to the extreme solutions, that is solution Pareto solutions related to j = 2 — scenario 2 at j = 1.
1 and solution 4 in Table A1, differ in term of cost value of 17% with
Pareto solution U C Replacements configuration
respect to worst value and of 7% in term of unavailability value. Therefore
the following thresholds, reported in Table A2, are assumed. 1 29.98 2144.08 8,10,12
2 28.10 2192.75 8,10,11,12
3 26.64 2355.16 8,9,10,11,12

Table A2
Thresholds for ELECTRE III related to j = 1.
The objective values related to the extreme solutions, that is solu-
Cost thresholds (%) Unavailability thresholds (%) tion 1 and solution 3 in Table A7, differ in term of cost value of 9%
Indifference Preference Veto Indifference Preference Veto with respect to worst value and of 11% in term of unavailability
3 10 15 1 4 6 value. Therefore the following thresholds, reported in Table A8, are
assumed.

By applying the ELECTRE III for each scenario the chosen solutions
are those reported in Table A3. Table A8
Thresholds for ELECTRE related to j = 2 — scenario 2 at j = 1.

Cost thresholds (%) Unavailability thresholds (%)

Table A3 Indifference Preference Veto Indifference Preference Veto
Selected solutions for j = 1.
3 6 8 3 6 8
Scenario Chosen solution

1 2
2 3 By applying the ELECTRE III for each scenario the chosen solutions
are those reported in Table A9.

A.2. System stop j = 2

Table A9
Selected solutions for j = 2 — scenario 2 at j = 1.
Scenario 1 at the system stop j = 1
Scenario Chosen solution
The Pareto solutions are reported in Table A4.
1 2
2 3
Table A4
Pareto solutions related to j = 2 — scenario 1 at j = 1.

Pareto solution U C Replacements configuration

A.3. System stop j = 3
1 30.24 2157.82 8,10,12
2 28.10 2194.53 8,10,11,12
Scenarios 1 and 2 at j = 2 — scenario 1 at j = 1
3 26.64 2355.16 8,9,10,11,12
The Pareto solutions are reported in Table A10.

The objective values related to the extreme solutions, that is solution

Table A10
1 and solution 3 in Table A4, differ in term of cost value of 9% Pareto solutions related to j = 3 — for both the scenarios at j = 2 — scenario 1 at j = 1.
with respect to worst value and of 12% in term of unavailability value.
Therefore the following thresholds, reported in Table A5, are assumed. Pareto solution U C Replacements configuration

1 30.69 2282.58 8-10-12

2 28.82 2331.25 8-10-11-12
Table A5 3 28.52 2407.74 8-9-10-12
Thresholds for ELECTRE III related to j = 2 — scenario 1 at j = 1. 4 26.64 2456.40 8-9-10-11-12

Cost thresholds (%) Unavailability thresholds (%)

Indifference Preference Veto Indifference Preference Veto

The objective values related to the extreme solutions, that is solution
1 and solution 4 in Table A10, differ in term of cost value of 7% with
2 4 6 3 6 9
respect to worst value and of 13% in term of unavailability value.
 A. Certa et al. / Decision Support Systems 55 (2013) 126–134 133

Therefore the following thresholds are assumed. Since, as it is possible Scenario 2 at j = 2 — scenario 2 at j = 1
to note from the Table A6, the chosen solutions at j = 2 by hypothesizing The Pareto solutions are reported in the Table A16.
both the scenarios are coincident, the only following objective thresh-
olds set has to be assumed (Table A11).
Table A16
Pareto solutions related to j = 3 — scenario 2 at j = 2 — scenario 2 at j = 1.

Pareto solution U C Replacements configuration

Table A11
Thresholds for ELECTRE III related to j = 3 — for both the scenarios at j = 2 — scenario 1 1 29.87 2239.62 8-10-12
at j = 1. 2 27.99 2305.55 8-10-11-12
3 26.64 2456.40 8-9-10-11-12
Cost thresholds (%) Unavailability thresholds (%)

Indifference Preference Veto Indifference Preference Veto

2 4 6 4 8 12 The objective values related to the extreme solutions, that is

solution 1 and solution 3 in Table A16, differ in term of cost value of
9% with respect to worst value and of 11% in term of unavailability
By applying ELECTRE III for each scenario the chosen solutions are value. Therefore the following thresholds, as shown in the Table
those reported in Table A12. A17, are assumed.

Table A17
Thresholds for ELECTRE III related to j = 3 — scenario 2 at j = 2 — scenario 2 at j = 1.
Table A12
Selected solutions for j=3 — for both the scenarios at j=2 — Cost thresholds (%) Unavailability thresholds (%)
scenario 1 at j=1.
Indifference Preference Veto Indifference Preference Veto
Scenario Chosen solution 3 6 8 3 6 9
1 2
2 4
By applying the ELECTRE III for each scenario the chosen solutions
are those reported in Table A18.

Scenario 1 at j = 2 — scenario 2 at j = 1 Table A18

The Pareto solutions are reported in Table A13. Selected solutions for j = 3 — scenario 2 at j = 2 — scenario
2 at j = 1.

Scenario Chosen solution

Table A13
Pareto solutions related to j = 3 — scenario 1 at j = 2 — scenario 2 at j = 1. 1 2
2 3
Pareto solution U C Replacements configuration

1 30.69 2282.58 8-10-12

2 28.82 2332.00 8-10-11-12
3 28.52 2410.00 8-9-10-12
References
4 26.64 2456.40 8-9-10-11-12
[1] N.S. Arunraj, J. Maiti, Risk-based maintenance policy selection using AHP and goal
programming, Safety Science 48 (2010) 238–247.
[2] M. Bevilacqua, M. Braglia, The analytic hierarchy process applied to maintenance
The objective values related to the extreme solutions, that is solution strategy selection, Reliability Engineering and System Safety 70 (2000) 71–83.
[3] R. Bris, E. Châtelet, F. Yalaoui, New method to minimize the preventive maintenance
1 and solution 4 in Table A13, differ in term of cost value of 7% with
cost of series–parallel systems, Reliability Engineering and System Safety 82 (2003)
respect to worst value and of 13% in term of unavailability value. 247–255.
Therefore the following thresholds, reported in Table A14, are assumed. [4] J.A. Caldeira Duarte, J.C. Taborda, J.T. Craveiro, T.P. Trigo, Optimization of the
preventive maintenance plan of a series components system, International Journal
of Pressure Vessels and Piping 83 (2006) 244–248.
[5] G. Campanella, A framework for dynamic multiple-criteria decision making, Decision
Table A14
Support Systems 52 (1) (2011) 52–60.
Thresholds for ELECTRE III related to j = 3 — for scenario 1 at j = 2 — scenario 2 at j = 1. [6] M.C. Carnero, Selection of diagnostic techniques and instrumentation in a predictive
maintenance program, a case study, Decision Support Systems 38 (4) (2005)
Cost thresholds (%) Unavailability thresholds (%)
539–555.
Indifference Preference Veto Indifference Preference Veto [7] C.R. Cassady, E.A. Pohl, W.P. Murdock Jr., Selective maintenance modelling for indus-
trial systems, Journal of Quality in Maintenance Engineering 7 (2) (2001) 104–117.
2 4 6 4 8 12 [8] A. Certa, G. Galante, T. Lupo, G. Passannanti, Determination of Pareto frontier in
multi-objective maintenance optimization, Reliability Engineering and System
Safety 96 (2011) 861–867.
By applying ELECTRE III for each scenario the chosen solutions are [9] A. Certa, M. Enea, G. Giacomo, T. Lupo, A multi-objective approach to optimize a pe-
riodic maintenance policy, in: H. Pham (Ed.), Proceedings of the 18th ISSAT Interna-
those reported in Table A15. tional Conference on Reliability and Quality in Design, July 26–28, 2012, (Boston) -
ISBN: 978-0-9763486-8-9.
[10] A. Certa, G. Galante, T. Lupo, A. Passannanti, G. Passannanti, A model for a periodic
preventive maintenance policy for a series parallel system, Proceedings of the
Table A15 19th FAIM International Conference on Flexible Automation and Intelligent
Selected solutions for j=3 — for scenario 1 at j=2 — scenario Manufacturing, July 6–8, 2009, pp. 1026–1033, (Middlesbrought, UK), ISBN
2 at j=1. 978-0-9562303-3-1.
[11] C.S. Chang, F. Yang, Inspection frequencies for substation condition-based main-
Scenario Chosen solution tenance, NPTS 2007: Proceedings of the Naval Platform Technology Seminar,
1 2 Singapore May 16–17.
[12] K. Deb, Multi-objective optimization using evolutionary algorithms, John Wiley & Sons,
2 4
New York, 2001.
134 A. Certa et al. / Decision Support Systems 55 (2013) 126–134

[13] O. Hilber, V. Miranda, M. Matos, L. Bertling, Multiobjective optimization applied Antonella Certa received the Master's degree in Management Engineering in 2004 and
to maintenance policy for electrical networks, IEEE Transactions on Power System the Ph.D. degree in the Economic Analysis, Technology Innovation and Management of
22 (4) (2007). the Development Policy Studies Program in 2008 from the University of Palermo. She is
[14] T. Lupo, An effective opportunistic maintenance policy for a global service, currently a researcher and she is a professor of Industrial Plants at the Department of
International Journal of Service Science 3 (2010) 179–193. Production Engineering of the University of Palermo (divison of Agrigento). Dr. Certa's
[15] N. Papakostas, P. Papachatzakis, V. Xanthakis, D. Mourtzis, G. Chryssolouris, research interests are mainly focused on project management, multi-criteria analysis
An approach to operational aircraft maintenance planning, Decision Support and maintenance optimization models.
Systems 48 (4) (2010) 604–612.
[16] M. Pariazar, J. Shahrabi, M.S. Zaeri, Sh. Parhizi, A combined approach for mainte- Lupo Toni received the Master's degree in mechanical engineering from the University
nance strategy selection, Journal of Applied Sciences 8 (23) (2008) 4321–4329. of Palermo, Italy in 1996, and the Ph.D. degree in Industrial Engineering from the same
[17] B. Roy, The outranking approach and the foundations of ELECTRE methods, in: university, in 2003. He is currently a researcher at the Department of Production Engineer-
C.A. Bana e Costa (Ed.), Reading in Multiple Criteria Decision Aid, Springer Verlag, ing of the University of Palermo and he is a professor of Quality Industrial Management.
Berlin, 1990, pp. 155–183. His research interests are in reliability optimization, and statistical method of reliability
[18] T.L. Saaty, Fundamentals of decision making and priority theory with the analytic modeling. Dr. Lupo is a member of the Italian Association of Mechanical Technology.
hierarchy process, RWS Publication, Pittsburg (PA), 2000.
[19] S. Taghipour, D. Banjevic, Periodic inspection optimization models for a reparable Mario Enea is a full professor at the University of Palermo in Project Management,
system subjected to hidden failure, IEEE Transactions on Reliability 60 (1) (2011) Industrial Plant Management and Applied Economics for Engineering. His research fields
275–285. include production management, scheduling problems related to project management,
[20] S. Taghipour, D. Banjevic, A.K.S. Jardine, Periodic inspection optimization model optimal layout problems, environmental impact evaluation and waste management.
for a complex repairable system, Reliability Engineering and System Safety 95
(2010) 944–952.
[21] Y.T. Tsai, K.S. Wang, L.C. Tsai, A study of availability-centered preventive maintenance
for multi-component systems, Reliability Engineering and System Safety 84 (2004)
261–270.
[22] J.K. Vaurio, Availability and cost functions for periodically inspected preventively
maintained units, Reliability Engineering and System Safety 63 (1999) 133–140.
[23] H. Wang, A survey of maintenance policies of deteriorating systems, European
Journal of Operational Research 139 (2002) 469–489.