Progress in Nuclear Energy 52 (2010) 710e714

Contents lists available at ScienceDirect

Progress in Nuclear Energy

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

A Particle Swarm Optimization (PSO) approach for non-periodic preventive

maintenance scheduling programming
Cludio M.N.A. Pereira a, b, *, Celso M.F. Lapa a, Antnio C.A. Mol a, b, Andr F. da Luz a
a
b

Comisso Nacional de Energia Nuclear e IEN/CNEN R. Hlio de Almeida, 75, 21941-972, P.O. Box 68550, Ilha do Fundo, Rio de Janeiro, Brazil
Universidade Gama Filho, Departamento de Cincia da Computao, Rua Manoel Vitorino 553, Rio de Janeiro, Brazil

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 9 December 2009
Received in revised form
5 April 2010
Accepted 29 April 2010

This work presents a Particle Swarm Optimization (PSO) approach for non-periodic preventive maintenance scheduling optimization. The probabilistic model, which is focused on reliability and cost
evaluation, is developed in such a way that exible intervals between maintenance interventions are
allowed. Due to the fact that PSO is typically skilled for real-coded continuous spaces, with xed
dimension (number of search parameters), a non-straightforward codication for solution candidates has
been developed in order to allow PSO to deal with variable number of maintenance interventions. To
evaluate the proposed methodology, the High Pressure Injection System (HPIS) of a typical 4-loop
Pressurized Water Reactor (PWR) has been considered. The optimization problem consists in maximizing
the systems average availability for a given period of time, considering realistic features such as: i) the
probability of needing a repair (corrective maintenance), ii) the cost of such repair, iii) typical outage
times, iv) preventive maintenance costs, v) the impact of the maintenance in the systems reliability as
a whole, vi) probability of imperfect maintenance, etc. Obtained results demonstrated good capability of
proposed PSO approach for automatic expert knowledge acquisition (without any a priori information),
which allowed it to nd optimal solutions.
2010 Elsevier Ltd. All rights reserved.

Keywords:
Particle Swarm Optimization
Preventive maintenance
Nuclear power plant

1. Introduction
Preventive maintenance (PM) scheduling planning is an important issue in nuclear power plants (NPP) operation. As PM directly
affects operations reliability, availability and cost, PM optimization
has been the focus of many investigations by nuclear engineers and
researchers in the last 15 years.
According to Duthie et al. (1998), since the beginning of the last
decade, researchers have been publishing papers addressing
preventive maintenance optimization of nuclear power plant
systems. This may be classied in three main groups. The rst one has
the focus on systems reliability (Hilsmeier et al., 1995; Vaurio, 1997).
The second one focuses on probabilistic models and perform tests
among some standard policies (Martorell et al., 1996; Duthie et al.,
1998; Martorell et al., 2004). Finally, we can mention those, which
apply expert knowledge to determine good maintenance policies
(Van Noortwijk et al., 1992). In order to avoid the optimization

* Corresponding author. Comisso Nacional de Energia Nuclear e IEN/CNEN

R. Hlio de Almeida, 75, 21941-972, P.O. Box 68550, Ilha do Fundo, Rio de Janeiro,
Brazil. Tel.: 55 21 2173 3897.
E-mail address: cmnap@ien.gov.br (C.M.N.A. Pereira).
0149-1970/$ e see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.pnucene.2010.04.009

difculties inherent to huge search spaces, many applications have

considered systems with few components (Harunuzzaman and
Aldemir, 1996; Billinton and Pan, 1998).
From the probabilistic point of the view, Park et al. (2000)
contributed to the solution of the class of problem under discussion by including components with very small degradation degrees,
Chiang and Yang (2001) have proposed approaches for maintenance optimization problems aiming at obtaining systems availabilities by Markovian methods and Dijkhuizen and Heijden (1999)
have optimized the distribution of availability intervals instead of
optimizing the preventive maintenance policy.
In the practice, establishing a good maintenance policy is often
a hard task, due to the large number of components and operational
constraints found in industrial systems. Aiming to better deal with
such complexity, Munz and Martorell (1997) proposed the use of
genetic algorithm as the optimization tool.
Lapa et al. (2000) proposed an innovative approach, considering
the possibility of non-periodic PM scheduling. The objective was to
maximize systems average availability during a given operation
period. To accomplish that, a PSA model has been developed. As
a consequence of considering non-periodic PM interventions,
problems complexity increased very much. In order to deal with
such complexity, a robust optimization tool was required and

C.M.N.A. Pereira et al. / Progress in Nuclear Energy 52 (2010) 710e714

genetic algorithms (GA) has been used. In that work, optimized nonperiodic scheduling demonstrated to t better to optimizations
objective and constraints than conventional periodic PM.
Lapas approach for non-periodic interventions has also been
successfully applied to surveillance tests (Lapa et al., 2002, 2003;
Pereira and Lapa, 2003; Sacco et al., 2008).
In further work (Lapa et al., 2006), the model has been extended,
considering maintenance costs. Those works considered realistic
features such as: i) the probability of needing a repair (corrective
maintenance), ii) the cost of such repair, iii) typical outage times, iv)
preventive maintenance costs, v) the impact of the maintenance in
the systems reliability as a whole, vi) probability of imperfect maintenance, etc. The genetic modeling used in such works, however,
presented a limitation due to the proposed solution candidate
(genotype) encoding, which considered a 15-days time step, allowing
only dates which are multiples of 15, and, consequently, restricting
the search space.
In this work, a Particle Swarm Optimization (PSO) approach for
non-periodic preventive maintenance scheduling optimization is
proposed. Such proposed model, eliminate the above mentioned
time-step limitation, allowing interventions to be scheduled with
one-day interval (if required). Such approach propitiates a better
tting of the schedules to components characteristics and maintenance policy objectives, as could be observed in preliminary results
(Pereira et al., 2009).
The proposed PSO is intended to search for the optimum maintenance policy considering several relevant features such as: i) the
probability of needing a repair (corrective maintenance), ii) the cost
of such repair, iii) typical outage times, iv) preventive maintenance
costs, v) the impact of the maintenance in the systems reliability as
a whole and vi) probability of imperfect maintenance.
PSO (Kennedy and Eberhart, 1995) is a population-based metaheuristic (PBM), in which solution candidates are enhanced through
the simulation of a simplied social adaptation model. Several
successful applications of PSO to nuclear problems reported in
literature (Siqueira et al., 2005; Domingos et al., 2006; Pereira et al.,
2007; Waintraub et al., 2009), in which PSO demonstrated advantages over other well-established PBM, motivated this work.
Considering that PSO works in continuous space, with xed
length real-coded vectors (to encode solution candidates) and the
proposed problem is discrete and may allow variable number of
maintenance interventions for each system component, a non-trivial
encoding of solution candidates has been developed in this work.
In order to evaluate the proposed methodology, the High Pressure Injection System (HPIS) of a typical 4-loop PWR has been
considered. Obtained results demonstrated good capability of
proposed PSO approach for automatic expert knowledge acquisition
(whithout any a priori information), which allowed it to nd optimal
solutions.

711

Fig. 1. Standard PSO pseudo code.

respectively, the position and the velocity of particle i in time t, in

an
n-dimensional
search
space.
Considering
that



!
pBesti t fpBesti;1 t; .; pBesti;n tg is the best position already
found
by
particle
i
until
time
t
and



!
gBesti t fgBesti;1 t; .; gBesti;n tg is the best position already
found by a neighbor until t, the PSO updating rules for velocity and
position are given by (Kennedy and Eberhart, 1995):



vi;n t 1 w:vi;n t c1 :r1 : pBesti;n t  xi;n t


c2 :r2 : gBesti;n t  xi;n t

(1)

xi;n t 1 xi;n t vi;n t 1

(2)

where r1 and r2 are random numbers between 0 and 1. Coefcients

!



!
c1 and c2 are given acceleration constants towards pBest and gBest
respectively and w is the inertia weight.
The inertia weight, w, is the responsible for the scope of the
exploration of the search space. High values of w promote global
exploration and exploitation, while low values, lead to local search.
A common approach to provide balance between global and local
search is to linearly decrease w during the search process.
The swarm is randomly initialized. Then, while stopping criterion is not reached, particles move according velocity and positions
equations (Eqs. (1) and (2)). The PSO algorithm pseudo code can be
seen in Fig. 1.
2.2. Optimization problem modeling
In this work a typical PWR High Pressure Injection System
(HPIS) has been considered. The HPIS can be represented by seven

V1

2. Proposed methodology
2.1. Particle Swarm Optimization
PSO (Kennedy and Eberhart, 1995) is Population Based Metaheuristic (PBM) inspired by the behavior of biological swarms and
social adaptation. In PSO, a swarm of structures encoding solution
candidates (particles) y in the n-dimensional search space of
the optimization problem looking for optima or near-optima
regions. The position of a particle represents a solution candidate
itself, while the velocity attribute, provides information about
direction and changing rate. Particles are guided by two components: i) cognitive information based on particles own experience
and ii) social information based on observation of neighbors. Let
!
!
X i t fxi;1 t; .; xi;n tg and V i t fvi;1 t; .; vi;n tg be,

Line 1
B1

V3

B2

V4

Line 2
B3

V2

Fig. 2. High pressure injection system.

712

C.M.N.A. Pereira et al. / Progress in Nuclear Energy 52 (2010) 710e714

V1
X0

...

X22

V2
X23

...

...

X45

!
i) X i t elements are real numbers ranging from 540 to 540;
ii) each element is rounded to the closes integer;
iii ) values from 1 to 540 represent valid days for maintenance
interventions;
iv) non-positive values represent no intervention;
v) duplicate values are removed;

B3

...

X138

...

X160

Vector X i (t )
!
Fig. 3. Vector X i t.

Fig. 4 exemplies such decoding procedure.

Eq. (3) shows the objective function, which should be
minimized.

main components: three pumps and four valves as shown in Fig. 2.

In normal operation its function is to complete the inventory of the
primary loop through the reactor coolant system, as well as to
guarantee the seal of the pumps of this system. Under accident
situations, in which the steam generators are unavailable or there is
a rupture in the primary system, the HPIS is used for removing the
decay heat. Considering that the reactor in operating with power
above 60% and at least 2 of the 3 pumps must be available during
the mission time, the top event is the unavailability to supply the
inventory by both feeders.
In this optimization problem, optimum maintenance scheduling
for components B1, B2, B3, V1, V2, V3 must be found. Therefore,
solution candidates must encode all possible scheduling combinations for all components.
Considering that PSO works with real-coded xed length
vectors, its application to a discrete problem, in which solutions
may have variable length (components variable number of maintenance interventions) is not straightforward. Firstly, a maximum
of 23 interventions for each component has been established. So,
a solution candidate may comprise at most 161 (23 maintenance  7 components). Hence, it should be used a vector of 161
elements.
Then,
PSO
position
vector
is
given
by
!
!
X i t fxi;1 t; .; xi;161 tg. Fig. 3 illustrate X i t which present
23 elements for each component.
Note that, to allow variable scheduling length (components may
!
have less than 23 interventions), decoding X i t is not
straightforward.
Considering a total operation period of 540 days and a time step
!
of 1 day, the following steps are required to decode X i t into valid
scheduling.

F Wd :U Wc :C

(3)

where U is the average systems unavailability, given by:

1
T

Utdt

(4)

where T is the total mission time, U(t) is the systems unavailability

at time t, calculated according to Lapa et al. (2006).
C is the total cost for a given scheduling, given by:
X
X

Q 1

0/T
CTQ

(5)

where Q is the component index and j is the maintenance intervention index. C0/T
TQ is the total cost for Q during interval [0,T], also
calculated according to Lapa et al. (2006).
Wd and Wc are weights to be applied to emphasize the importance of each sub-objective (U or C).
3. Computational experiments and results
In the present investigation, 3 case studies have been carried
out. The rst one is a hypothetical situation used as benchmark, in
which global optimum is well known. The second and third ones
are situations closer to a real problem.

Vector X:
V1
x0

x1

X2

x3

x4

x5

x6

x7

x8

x9

x10 x11 x12 X13 x14 x15 x16 x17 x18 x19 x20 x21 x22

32.1 -18 95.7 200 -30 -9.1 -123 301 260 -11 -521 123 498 -23

-91 -510 15.9 350 200 -10

-89 380

10

...

B3

...

...

...

...

...

B3

...

...

...

B3

...

...

Valid days for maintenance

V1
32

96

200

301 260

123 498

16

350 200.

380

10

Maintenance scheduling without repetitions (sorted)

V1
10

16

32

96

123 200 260 301 350 380 498

Decoding X i (t ) into a valid scheduling.

!
Fig. 4. Decoding X i t into a valid scheduling.

C.M.N.A. Pereira et al. / Progress in Nuclear Energy 52 (2010) 710e714

713

0.025

Table 1
Results for the second case study.
Exp.

Particles

Seed

Cost

Unavailability

Fitness

1
2
3
4
5
6
7
8
9
10

20
20
20
20
20
50
50
50
50
50

123,456,789
7622
777
13
13,987
123,456,789
7622
777
13
13,987

0.0056313
0.00576458
0.00571318
0.00570311
0.00583616
0.00576754
0.00570256
0.00584039
0.00570006
0.00562846

0.00448557
0.00423274
0.00429454
0.00437837
0.00408374
0.00423455
0.00430842
0.00400797
0.00432445
0.00455739

0.0101169
0.00999732
0.0100077
0.0100815
0.0099199
0.0100021
0.010011
0.00984836
0.0100245
0.0101859

Fitness (F)
Unavailability (U)

0.02

Cost (C)

0.015
0.01
0.005
0
0

500

1000

1500

2000

Iterations
3.1. First case study

Fig. 5. PSO evolution in second case study.

In order to test efciency and consistency of proposed methodology, an investigation considering the following characteristics
have been done:
i) component is not under aging and failure rates are constant,
so, maintenance does not improve reliability;
ii) all maintenance is perfect.
Under such conditions maintenance does not provide any gain.
It only increases cost. Therefore the expected optimum scheduling
is no maintenance interventions for all components.
Ten experiments have been made with different random seeds
and typical values for C1 and C2 (both set to 2.0). Inertia weight, w,
decreased from 0.8 to 0.2 in 2000 generations.
In all cases no maintenance have been proposed, demonstrating the efciency and consistency of the proposed approach.
3.2. Second case study
In this scenario, more realistic values for failure rates, costs for
maintenance and repair, etc., have been used according to
Harunuzzaman and Aldemir (1996).
Again, ten experiments have been made with different random
seeds and typical values for C1 and C2 (both set to 2.0). Inertia
weight, w, decreased from 0.8 to 0.2 in 1000 iterations. The stopping criterion was the number of iterations equal to 2000. Tables 1
and 2 show obtained results.
In Table 1 it can be observed the consistency of the method in
nding solutions very close to each other (a near-optimum region).
Note that Valve 1 and Pump 1 which are inline to each other stop
at coincident time (t 228). The PSO was able to discover itself that
such fact improves availability. The same occurs with Valve 2 and
Pump 3 (in italic bold). Also, pumps undergo fewer interventions
due to the higher cost and outage time for maintenance.
Fig. 5 shows the PSO evolution during 2000 generations.
In Fig. 5 it can be observed a stagnation (or very slight minimization) of the unavailability (U). This stagnation occurs due to the

relatively high value of Wc (when compared to Wd), which

emphasizes the minimization of the cost term (C). Such relatively
high value of Wc also made maintenance not so interesting (high
cost/availability relation), leading to a relatively low number of
maintenance interventions.
3.3. Third case study
Aiming to decrease the relatively higher inuence of cost on the
overall tness value and, therefore, allowing unavailability to be
minimized, Wc has been decreased 25 times. The optimization
parameters were the same used in the second case study.
Fig. 6 shows the PSO evolution (for the best run) during 2000
iterations. In this case, the unavailability (U) inuence over the
tness is more accentuated, and, although the compromise
between U and C imposes some oscillation in evolution, U could be
decreased along the time.
Results for 5 experiments are shown in Table 3 while best
solution found by the PSO can be seen in Table 4.
As expected, the lower weight (Wc) applied to the cost (C) made
maintenance more interesting and led to a higher number of
(maintenance) interventions. Although such fact increases problems
complexity, the proposed PSO could deal with that, nding good
schedules. Again, in this more complex situation, PSO ratied its
great capacity of knowledge discovery: as occurred in the second
case study, all interventions in Pump 1 occur together with (at same
time) Valve 1 (see underlined bold values in Table 4) and most
interventions on Pump 3 are coincident with Valve 2 (see italic bold
values in Table 4).

0,0035

Fitness (F)
Unavailability (U)
Cost (C)

0,003
0,0025
0,002

Table 2
Best solution found for second case study.

0,0015

Fitness 0.00984836
Component

Number of interventions

Days of the interventions

Valve 1
Valve 2
Valve 3
Valve 4
Pump 1
Pump 2
Pump 3

4
4
3
3
1
2
1

72-146-228-326
73-148-226-331
108-218-343
107-216-342
228
132-281
226

0,001
0,0005
0
0

500

1000

1500

Iterations
Fig. 6. PSO evolution in third case study.

2000

714

C.M.N.A. Pereira et al. / Progress in Nuclear Energy 52 (2010) 710e714

Table 3
Results for the third case study.
Exp.

Particles

Seed

Cost

Unavailability

Fitness

1
2
3
4
5

50
50
50
50
50

123,456,789
7622
777
13
13,987

0.000422884
0.000498632
0.000497344
0.000459859
0.000450358

0.00206422
0.00168114
0.00157986
0.00175532
0.00171411

0.0024871
0.00217977
0.0020772
0.00221518
0.00216447

Table 4
Best solution found for third case study.
Fitness 0.0020772
Component

Number of
interventions

Days of the interventions

Valve 1
Valve 2
Valve 3

6
9
12

Valve 4
Pump 1
Pump 2
Pump 3

8
4
5
6

48-92-145-202-287-368
47-91-136-177-208-240-273-310-360
46-90-138-178-212-250-288-319-347385-426-463
57-112-174-223-277-329-382-447
92-202-287-368
55-111-171-239-329
60-115-177-240-273-360

4. Conclusions
This work demonstrates the feasibility of using PSO for preventive maintenance optimization. The non-conventional approach to
encode solution candidates into xed length real-coded vector
demonstrated to be efcient to deal with a 1-day step scheduling
optimization with variable number of interventions, allowing PSO to
better t schedules to optimization objectives and, therefore, to nd
good solutions.
The efciency and consistency of the method has been observed,
not only in terms of the achieved tness values, but principally in the
knowledge acquired in different situations to which it has been
submitted. Such fact demonstrated a great ability of the proposed
PSO in dealing with complex, deceptive and multimodal search
spaces.
As a future work, such approach could be extended to a more
useful but complex approach, involving Pareto based multi-objective optimization in order to provide a more effective and practical
tool for multi-criteria optimization problems.
References
Billinton, R., Pan, J., 1998. Optimal maintenance scheduling in a two identical
component parallel redundant system. Reliability Engineering and System
Safety 59, 309e316.
Chiang, J.H., Yang, J., 2001. Optimal maintenance policy for a Markovian system under
periodic inspection. Reliability Engineering and System Safety 71, 165e172.

Dijkhuizen, G., Heijden, M., 1999. Preventive maintenance and the interval availability distribution of an unreliable production system. Reliability Engineering
and System Safety 66, 13e27.
Domingos, R.P., Schirru, R., Pereira, C.M.N.A., 2006. Particle Swarm Optimization in
reactor core designs. Nuclear Science and Engineering 152, 1e7.
Duthie, J.C., Robertson, M.I., Clayton, A.M., Lidbury, D.P.G., 1998. Risk-based
approaches to ageing and maintenance management. Nuclear Engineering and
Design 184, 27e38.
Harunuzzaman, M., Aldemir, T., 1996. Optimization of standby safety system
maintenance scheduling in nuclear power plants. Nuclear Technology 113,
354e367.
Hilsmeier, T.A., Aldemir, T., Vesely, W.E., 1995. Time-dependent unavailability of
aging standby components based on nuclear plant data. Reliability Engineering
and System Safety 47, 199e205.
Kennedy, J., Eberhart, R.C., 1995. Particle Swarm Optimization. In: Proceedings of
IEEE International Conference on Neural Networks, vol. 4, pp. 1942e1948.
Lapa, C.M.F., Pereira, C.M.N.A., de Barros, M.P., 2006. A model for preventive
maintenance planning by genetic algorithms based in cost and reliability.
Reliability Engineering and System Safety 91 (2), 233e240.
Lapa, C.M.F., Pereira, C.M.N.A., Frutuoso e Melo, P.F., 2003. Surveillance test policy
optimization through genetic algorithms using non-periodic intervention
frequencies and considering seasonal constraints. Reliability Engineering and
System Safety 81 (1), 103e109.
Lapa, C.M.F., Pereira, C.M.N.A., Frutuoso e Melo, P.F., 2002. An application of genetic
algorithms to surveillance tests optimization of a PWR auxiliary feed-water
system. International Journal of Intelligent Systems 17, 813.
Lapa, C.M.F., Pereira, C.M.N.A., Mol, A.C.A., 2000. Maximization of a nuclear system
availability through maintenance scheduling optimization using genetic algorithm. Nuclear Engineering and Design 196, 95e107.
Martorell, S., Munz, A., Serradell, V., 1996. Age-dependent models for evaluating
risks & cost of surveillance & maintenance of components. IEEE Transactions on
Reliability 45 (3), 433e441.
Martorell, S., Sanchez, A., Carlos, S., Serradell, V., 2004. Alternatives and challenges
in optimizing industrial safety using genetic algorithms. Reliability Engineering
and System Safety 86 (1), 25e38.
Munz, A., Martorell, S., Serradell, V., 1997. Genetic algorithms in optimizing
surveillance and maintenance of components. Reliability Engineering & System
Safety 57 (2), 107e120.
Park, D.H., Jung, G.M., Yum, J.K., 2000. Cost minimization for periodic maintenance
policy of a system subject to slow degradation. Reliability Engineering and
System Safety 68, 105e112.
Pereira, C.M.N.A., Schirru, R., Lapa, C.M.F., Canedo, J.A.C., Waintraub, M.,
Meneses, A.A.M., Baptista, R.P., Siqueira, N.N., 2007. Particle Swarm Optimization applied to nuclear engineering problems. International Journal of Nuclear
Knowledge Management 2 (3), 313e332.
Pereira, C.M.N.A., Lapa, C.M.F., Mol, A.C.A., Luz, A.F., 2009. A PSO approach for
preventive maintenance scheduling optimization. In: International Nuclear
Atlantic Conference.
Pereira, C.M.N.A., Lapa, C.M.F., 2003. Parallel island genetic algorithm applied to
a nuclear power plant auxiliary feedwater system surveillance tests policy
optimization. Annals of Nuclear Energy 30, 1665e1675.
Sacco, W.F., Lapa, C.M.F., Pereira, C.M.N.A., Filho, H.A., 2008. A metropolis algorithm
applied to a nuclear power plant auxiliary feedwater system surveillance tests
policy optimization. Progress in Nuclear Energy 50, 15e21.
Siqueira, N.N., Pereira, C.M.N.A., Lapa, C.M.F., 2005. The Particle Swarm Optimization algorithm applied to nuclear systems surveillance test planning. In:
Proceedings of the International Nuclear Atlantic Conference (INAC).
Van Noortwijk, J., Dekker, R., Mazzuchi, T., 1992. Expert judgment in maintenance
optimization. IEEE Transactions on Reliability 41, 427e432.
Vaurio, J.K., 1997. On time-dependent availability and maintenance optimization of
standby units under various maintenance policies. Reliability Engineering and
System Safety 56, 79e89.
Waintraub, M., Schirru, R., Pereira, C.M.N.A., 2009. Multiprocessor modeling of
parallel particle swarm optimization applied to nuclear engineering problems.
Progress in Nuclear Energy 51, 680e688.