Computers & Industrial Engineering 109 (2017) 335–346

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Approximate evaluation of average downtime under an integrated

approach of opportunistic maintenance for multi-component systems
Hao Peng a,⇑, Qiushi Zhu b
a
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China
b
Vanderlande Industries, Veghel, The Netherlands

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

Article history: Since the quality of after-sales services is becoming more and more important for users of capital goods,
Received 22 November 2016 the average downtime per year becomes one of the key performance indicators of the systems. For multi-
Received in revised form 20 March 2017 component systems, the average downtime per year is not only determined by the maintenance policies
Accepted 26 April 2017
of individual components, but also can be saved by the joint maintenance decisions. Considering this
Available online 19 May 2017
dependency of system downtime, we proposed a unified opportunistic maintenance policy for systems
under a mix of different individual maintenance policies (i.e., age-based, condition-based, and failure-
Keywords:
based policies). Components can employ different maintenance policies. We developed approximation
Multi-component systems
Opportunistic maintenance
methods to evaluate the average downtime per year and a heuristic solution procedure to determine
Condition-based maintenance the interval length for the scheduled downs, the control limits or the age limits of components. The
Average downtime numerical study shows the accuracy of the approximation methods. Heuristic solutions are also obtained
to demonstrate the use of our model.
Ó 2017 Published by Elsevier Ltd.

1. Introduction with a focus on the performances or services of the machines, in

High-tech capital goods (e.g., oil-gas refineries, lithography
machines, and baggage handling systems) nowadays become more
and more important for production and services. Users of such
advanced engineering systems usually require the systems to be
available 24 h a day, 7 days a week. For instance, the lithography
machines in chip factories are often the bottlenecks of the produc-
tion lines, and run 24 h every day to create more chips. The down-
times of the lithography machines are very costly, i.e., millions of
euros of reduced production output. The baggage handling system
in a busy hub airport also works almost continuously to serve the
flights. The out-of-service hours of the baggage handling system
can result in delays of the flights, which is unacceptable for passen-
gers. Hence, the average downtime per year becomes the key per-
formance indicator of these advanced engineering systems
(Timmermans, 2012; Trivedi, Kim, & Ghosh, 2013). While buying
these advanced engineering systems, the tolerance level of the
downtime per year will be specified by the users in the service con-
tract to guarantee that the systems can run properly most of the
time. The original equipment manufacturers (OEM) thus have to
shift from a pure product-oriented strategy to a business strategy
⇑ Corresponding author at: Zhongguancun east 55, 100190 Beijing, China.
E-mail address: penghao@amss.ac.cn (H. Peng).
order to satisfy the customers’ demand and increase the market
share (Kim, Cohen, & Netessine, 2007). Due to this trend towards
the so-called ”performance-based logistics”, the OEMs are moti-
vated to evaluate the key performance indicators (e.g., average
downtime per year) for the service schedules (e.g., maintenance
schedule) and optimize the service schedules accordingly.
In this research for multi-component systems, we focus on the
type of dependencies that the total downtime can be saved by
jointly maintaining several components prone to failures. For
example, if one failed component is maintained together with
some other component prone to failure, the maintenance actions
will be performed in a parallel way so that the total downtime is
smaller than conducting the two maintenance actions sequen-
tially; moreover, the setup time of production/services or the wait-
ing time of maintenance resources can be reduced if several
components can be repaired/replaced simultaneously, since only
one setup or one delivery is needed. Thus, how to evaluate the
key performance indicator (i.e., the average downtime per year)
considering these dependencies is our research question for
multi-component systems under different maintenance policies.
We aim to build up a probability model to characterize the coordi-
nation of the different maintenance policies of multiple compo-
nents so that the average downtime per year can be evaluated
and minimized by optimizing the joint maintenance decisions.

http://dx.doi.org/10.1016/j.cie.2017.04.043
0360-8352/Ó 2017 Published by Elsevier Ltd.
336 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346
For single-component systems, the key performance indicators
related to the expected downtime per year have been studied by
many researchers (e.g., the average downtime per unit time,
instantaneous availability, limit availability, or interval availabil-
ity); see (Ebeling, 2009) for a brief introduction. For example,
Christer and Lee (2000) revised the delay time model to evaluate
the expected downtime per unit time, by including the downtimes
in the calculation of the expected number of failures over an
inspection period. Jiang, Kim, and Makis (2013) investigated the
structure of the optimal control policy that maximizes the long-
run expected availability under partial observations. van
Dijkhuizen and van der Heijden (1999) maximized the interval
availability, instead of the usual limiting availability, to determine
the optimal preventive maintenance policy. For order-driven man-
ufacturing systems, the interval availability is often seen as a more
appropriate performance measure.
For multi-component systems, Monte Carlo simulation is one
approach to evaluate system downtimes or availabilities, e.g.,
Marquez, Heguedas, and Iung (2005) provided a case study of
cogeneration plants. Faulin, Juan, Serrat, and Bargueno (2008) pro-
posed the use of discrete-event simulation as an efficient method-
ology to obtain estimates of availability functions in time-
dependent real systems. The systems can present multiple states,
dependencies among failure/repair times or non-perfect mainte-
nance policies. Naseri, Baraldi, Compare, and Zio (2016) modeled
the time-dependent effects of environmental conditions on the
system availability of oil and gas processing facilities. Due to the
complexity of the problem, direct Monte Carlo simulation was
used. It is worth noting that a calendar-based preventive mainte-
nance policy was considered, which is similar to our case.
Borgonovo, Marseguerra, and Zio (2000) presented a Monte Carlo
approach for the evaluation of plant maintenance strategies and
operating procedures under economic constraints. A model of
obsolescence is introduced to evaluate the convenience of substi-
tuting a failed component with a new, improved one.
Compared with Monte Carlo simulation, analytical methods can
save the computation times significantly. Trivedi et al. (2013)
reviewed analytic modeling techniques such as non-state-space
models, state-space models, and hierarchical models to evaluate
average system downtime or system availability, with case studies
applying these methods from Motorola, Cisco and Sun Microsys-
tems. Average downtime (in minutes per year) has been used as
a performance indicator for the Cisco router case and Sun
Microsystems case. Cochran, Murugan, and Krishnamurthy
(2001) presented a generic Markov model to reduce the computa-
tional effort of availability evaluation. The method was used and
compared with a Petri-net simulation in a case study of a reactor
regenerator system in a Fluid Catalytic Cracking Unit in a petro-
leum refinery. Raje, Olaniya, Wakhare, and Deshpande (2000)
assessed the availability of a critical pumping system of the Crude
Distillation Unit of a refinery using a three-state Markovian model.
If the maintained system evolves according to a semi-Markov pro-
cess, Bloch-Mercier (2000) identified the optimal preventive main-
tenance policy to maximize the stationary availability. Csenki
(1994) derived the joint availability of a system modelled by a
semi-Markov process, which is defined as the probability of the
system being functional in both t and t þ x. Wang and Pham
(2006) investigated the availability of the series system whose
components are subject to imperfect repair, correlated times to
failure and repair. The shut-off rules are optimized. de Smidt-
Destombes, van der Heijden, and van Harten (2004) analyzed the
availability of a k-out-of-N system with identical components con-
sidering the spare part stock level, the maintenance policy and the
repair capacity. Maintenance is initiated when the number of failed
components exceeds some critical level. Kreimer and Mehrez
(1998) computed the availability of a real-time queueing system
when both service and maintenance times are exponentially dis-
tributed. Du, Cui, and Lin (2016) obtained the closed-form expres-
sions of four availability indexes using the technique of aggregated
stochastic process. In most of the above works, the type of down-
time dependencies among components has rarely been considered,
which refers to the savings on the total downtime by joint mainte-
nance decisions; see (Do, Vu, Barros, & Berenguer, 2015 and Tsai,
Wang, & Tsai, 2004).
If we do not consider downtime as a key performance indicator,
many works have been done recently to propose maintenance poli-
cies for degrading systems. Alaswad and Xiang (2017) reviewed
the latest works of condition-based maintenance. Ahmadi and
Fouladirad (2017) maximized the expected gain of a deteriorating
production system under random inspections. Verbert, De
Schutter, and Babuska (2017) proposed a new strategy for timely
maintenance planning in multi-component systems. Lee and Cha
(2016) considered periodic preventive maintenance policies and
assumed that the failure process between two PMs follows a new
counting process which is a generalized version of the nonhomoge-
neous Poisson process. Zhou, Lin, Sun, and Ma (2016) optimized a
maintenance policy of a parallel-series system considering both
stochastic and economic dependence among components as well
as limited maintenance capacity. El Hajj, Castanier, Schoefs, and
Yeung (2016) constructed a degradation model for maintenance
of reinforced concrete structure subjected to cracking. Liu, Xie,
and Kuo (2016) analyzed the system reliability subject to shared
constant or cumulative loads.
Our research differs from the existing works by considering a
mixture of different maintenance policies (i.e., age-based policy,
condition-based policy and failure-based policy) for multi-
component systems. The coordination of maintenance actions
under the different policies has rarely been discussed in the liter-
ature. However, for a complex engineering system in practice, dif-
ferent maintenance policies are employed for different
components due to the diverse characteristics of components.
For example, some electronic parts (e.g., circuit board, current
adapter) are under the failure-based maintenance policy (FBM),
since their failure times follow exponential distributions. For parts
that have increasing failure rates, the age-based maintenance pol-
icy (ABM) is often employed. If the conditions of parts can be mea-
sured easily, the condition-based maintenance policy (CBM) is
often used. We proposed a unified opportunistic maintenance
framework for such kind of multi-component systems under
mixed maintenance policies. In this framework, when the degra-
dation level of a CBM component exceeds its control limit, we will
take the first appeared (scheduled/unscheduled) opportunity from
other components and jointly maintain this CBM component with
other components. Similarly, when the age of an ABM component
exceeds its age limit, we will take the appeared opportunity to
conduct joint maintenance on this ABM component. The average
long-run downtimes per year for components and system are
approximately evaluated and minimized by optimizing the control
limits for CBM components and the age limits for ABM
components.
The outline of this paper is as follows. The problem description
and formulation are given in Section 2. In Section 3, approximation
methods are proposed to evaluate the expected downtime of a
multi-component system under a mix of different maintenance
policies. Based on the approximation models, in Section 4, a heuris-
tic solution procedure is given to minimize the expected down-
time. In Section 5 numerical study is performed to check the
accuracy of the approximation models. The heuristic solutions of
the optimization model are also presented in Section 5, as well
as a sensitivity analysis.
 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346
2. Model description
2.1. Notation
s: interval of scheduled downs (yrs).
Ai : age limit on the age of component i 2 IABM (decision
variable).
C i : control limit on the degradation level of component i 2 ICBM
(decision variable).
Hi : CM threshold on the degradation of component i 2 ICBM .
li : expected life time of component i 2 IFBM (yrs).
Z i : average downtime per year of component i 2 I (hrs/yr).
Z syst : average downtime per year of the system (hrs/yr).
PMUSD
di : PM time of component i 2 fICBM [ IABM g at unscheduled
system downs (hrs).
di
PMSD
: PM time of component i 2 fICBM [ IABM g at scheduled sys-
tem downs (hrs).
Rep
di : corrective replacement time of component i 2 I (hrs).
SUSD
: setup time at unscheduled downs (hrs).
SSD : setup time at scheduled downs (hrs).
CM Rep
di : CM time (SUSD þ di ) of component i 2 I (hrs).
2.2. System under a mixture of maintenance policies
Consider a series system consisting of multiple components
that are subject to failures. The set I denotes the set of all compo-
nents, and the components are numbered as f1; . . . ; jIjg. Within the
system, components follow different maintenance policies. We
denote components that are under condition-based maintenance
policy by subset ICBM . Components that are under age-based main-
tenance policy are denoted by IABM . Components that are under
failure-based maintenance policy are denoted by IFBM . It holds that
ICBM \ IABM \ IFBM ¼ £ and ICBM [ IABM [ IFBM ¼ I. We are in particular
interested in the cases that have many components in a system
(i.e., a high jIj), for which we have large-scale optimization prob-
lems at the system level.
For such multi-component systems, we often see two types of
system downs, especially in the high-tech industry.
 Scheduled downs: For many systems, periodic maintenance
actions for components/systems (e.g., inspection, cleaning, and
lubrication) are executed at predetermined time points with
fixed time intervals s (yrs) in between. It facilitates the planning
and coordination of maintenance resources, including service
engineers, maintenance tools, or spare parts. For example, the
production lines of food processing industry often schedule pre-
ventive maintenance activities periodically (van Dorst, 2014).
 Unscheduled downs: Beside the scheduled downs, we will also
have unexpected system downs when failures of components
occur. The unscheduled downs (USDs) may come from the three
different types of components. For example, the electronics of a
system are often under failure-based maintenance policies.
When a short-circuit failure occurs, the system goes down
due to power loss. The loading structures, such as trusses, are
often under age-based policies. When a failure due to material
fatigue occurs, a truss fails before its preventive maintenance
action. The system then goes down due to a structure deforma-
tion. For CBM components, if their conditions pass the failure
thresholds, the engineers will shut down the system immedi-
ately, which constitutes a proportion of the unscheduled downs
(Timmermans, 2012). We approximate the arrivals of unsched-
uled downs by a Poisson process, which is reasonable when the
number of components in the system is large.
337
Every scheduled/unscheduled down will cause a period of
downtime (hrs) and a setup time (hrs) for new productions. At
the system level, in order to reduce the total system downtime
and the setup time, a smart clustering of maintenance actions is
required. In our opportunistic maintenance policy, the scheduled
or unscheduled downs become the opportunities of preventively
maintaining the components that have not failed yet in the system.
To better illustrate the opportunistic maintenance policy, the
maintenance cycles of an age-based component and a condition-
based component are demonstrated in Fig. 1. The maintenance
cycles are the intervals between two consecutive maintenance
actions for every component.
For component i 2 IFBM , the lifetime of component i is a random
variable with probability density function f i ðxÞ and expected value
li . A corrective maintenance action will be conducted after the
component fails. This leads to a setup time SUSD for the unsched-
Rep
uled down and a time di for the corrective replacement of com-
ponent i.
For component i 2 IABM , the lifetime of component i is a random
variable T i with probability density function f i ðxÞ. We introduce an
age limit Ai and execute a preventive maintenance action if the life-
time is larger or equal to Ai when the opportunities of joint main-
tenance arrive. Since these preventive maintenance actions are
performed simultaneously with the scheduled downs or unsched-
uled downs from other components, the downtime can be partially
saved for component i 2 IABM . The fixed setup time for component i
is also saved by this joint maintenance, due to the fact that we only
need one setup for this joint maintenance activity. Thus, the corre-
PMSD
sponding preventive maintenance time is equal to di or
PMUSD
di ,
for scheduled downs or unscheduled downs respectively.
If component i fails, a corrective maintenance action is taken on
Rep
component i, which leads to a corrective replacement time di
and a fixed setup time SUSD for this unscheduled down (namely,
CM Rep
the corrective maintenance time di ¼ SUSD þ di ). Notice that
CM
di is larger than
PMSD
di and
PMUSD
di .
The age limit Ai ¼ 1 implies
that we will never have a preventive maintenance action, which
makes the ABM policy degenerate to a FBM policy.
For component i 2 ICBM , the degradation level is measured con-
tinuously and the degradation level at time t is denoted by
X i ðtÞ; t P 0. We assume that a higher level of X i ðtÞ corresponds to
a higher level of degradation. A failure threshold Hi is given by
the experts in the design and engineering department. If the degra-
dation level X i ðtÞ reaches level Hi , one is not allowed to continue
the operation and has to do an immediate repair. Hence, when
X i ðtÞ reaches level Hi , we see this as a failure and the repair is seen
as a corrective maintenance action. This corrective maintenance
CM
action of component i leads to a time di , which is composed of
Rep
a corrective replacement time and a fixed setup time SUSD for
di
this unscheduled down. In order to avoid these relatively long
down times, we introduce a control limit C i (6 Hi ). We replace
component i preventively if X i ðtÞ P C i when an opportunity of
scheduled/unscheduled downs appears at time t. The time of a pre-
ventive replacement at a scheduled or unscheduled down is equal
PMSD PMUSD CM
to di or di respectively. Both times are smaller than di .
Notice that C i ¼ Hi implies that we will never have a preventive
replacement, which makes the CBM policy degenerate to a FBM
policy.
We assume the degradation processes of CBM components and
the lifetimes of FBM and ABM components are independent of each
other. The components are as good as new after repair or replace-
ment. The service life of the system is assumed to be much longer
than the lifetimes of the components, and the time of maintenance
338 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346
Fig. 1. The maintenance cycles for ABM and CBM components with SDs planned at fixed time points.
is not at the same magnitude as the lifetimes of the components
(yrs). Let C denote the vector of the control limits fC 1 ; . . . ; C jICBM j g
for CBM components and A denote the vector of the age limits
fA1 ; . . . ; AjIABM j g for ABM components. Under these assumptions,
the long-run average downtime (hrs) per year of the system
Z syst ð:Þ is thus given as
SSD X Fig. 1, the unexpected failure of the ABM component at the end
Z syst ðs; C; AÞ ¼ þ Z i ðs; C; AÞ;
s i2I
where Z i ð:Þ is the downtime (hrs) per year of maintenance incurred
by component i. The downtime (hrs) per year of component i; Z i ð:Þ,
includes the opportunistic preventive maintenance time and cor-
rective maintenance time of component i.
The evaluation of Z i ð:Þ has two difficulties. The first difficulty is
that renewal theory cannot be directly applied to evaluate the
average downtime per year Z i ð:Þ because when component i is pre-
ventively or correctively maintained at the end of a maintenance
cycle, the scheduled downs of the next maintenance cycle will not
be rescheduled to make the first SD appears s (yrs) away from
the beginning of the nextmaintenance cycle, which violates the
renewal property of a cycle. We develop approximation methods
for ABM and CBM components to improve the downtime per year
evaluations, considering the fact that the scheduled downs are at
fixed time points, as introduced in Section 3. The second difficulty
of evaluating Z i ð:Þ comes from the arrival process of USDs, which is
a superposition of the failure processes of all the components. The
superposition process of jIj non-renewal processes is not, in gen-
eral, a renewal counting process because the interoccurrence times
are not independent and certainly not identically distributed. How-
ever, scholars have established the result that under certain condi-
tions, the superposition of indefinitely many uniformly sparse
renewal processes tends to a Poisson process (Ross, 1996). We thus
also use Poisson process to approximate the arrival process of
USDs.
Furthermore, the failure processes of ABM and CBM compo-
nents are dependent through the USDs generated by all the com-
ponents in the system, although the degradation processes of
CBM components and the lifetimes of FBM and ABM components
are assumed to be independent of each other. For instance, in
ð1Þ
of the third maintenance cycle (see Fig. 1) becomes an opportu-
nity of preventive maintenance for the still-working CBM compo-
nents; the CBM component does not take this opportunity since
the degradation level is below the control limit C. At the end of
the first maintenance cycle of the CBM component, the degrada-
tion level reaches the failure threshold H and a corrective main-
tenance action has to be conducted; this thus becomes an
opportunity of preventive maintenance for the ABM component;
since the age of the ABM component at this very moment is lar-
ger than the age limit A, the opportunity of preventive mainte-
nance is taken by the ABM component. Hence, considering the
dependencies of the failure processes of ABM and CBM compo-
nents, we need to have an iterative procedure to evaluate the
arrival rates of USDs generated by all the components, as well
as Z i ð:Þ.
The optimal solution ðs ; C ; A Þ can be obtained by minimizing
Eq. (1). The difficulties of solving this optimization problem arise
from the evaluation of Z i ð:Þ and the optimization procedure cou-
pled with it, which will be introduced in Section 3.
2.3. Assumptions
(1) The life time of each ABM component is independent of
scheduled and unscheduled downs caused by other components
in the system.
 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346
(2) The degradation of each CBM component is independent of
scheduled and unscheduled downs caused by other components in
the system.
(3) The time horizon is infinite.
(4) The downtimes are not at the same magnitude as the life-
times of components. (hrs vs. years)
(5) Maintenance actions restore components as new.
(6) The system is composed of a large number of components.
3. Approximate evaluation
Due to the difficulties of evaluating the system average down-
time per year Z syst ðs; C; AÞ in Eq. (1), we provide the approximation
methods in this Section. For FBM component i (i.e., i 2 IFBM ), the
average downtime per year Z i ð:Þ does not depend on the decision
variables ðs; C; AÞ and can be evaluated by Z i ¼ di =li according
CM
to renewal theory (Ross, 1996). For a CBM or ABM component i
(i.e., i 2 ICBM [ IABM ), assuming that the arrival rate of USDs ki is
known, the average downtime per year Z i ð:Þ can be approximated
by the methods proposed in Section 3.1. The unknown arrival rates
of USDs for all the CBM and ABM components, k ¼ fki ; 8i 2
ICBM [ IABM g, are dependent on the decision variables ðs; C; AÞ and
have to be evaluated through an iterative procedure proposed in
Section 3.2 due to the dependencies of the failure processes of
CBM and ABM components.
3.1. Approximate evaluation of a CBM (or ABM) component
As we mentioned before, the renewal property of the mainte-
nance cycles for component i is violated due to the SDs at fixed
time points. A maintenance cycle will start at n time units (yrs)
away from the previous SD (0 6 n < s), as shown in Fig. 2. This
deviation of a renewal point from SDs, n, is changing from mainte-
nance cycle to maintenance cycle. But given this deviation n of a
certain maintenance cycle, we can derive the cycle downtime
(hrs) and cycle length (yrs) by analyzing the random events.
Fig. 2. The deviation of the renewal point f.
339
The degradation process can be described by many different
kinds of stochastic processes, e.g., random coefficient model,
Gamma process, Brownian Motion or Markov Chain. If the degra-
dation process is monotonic, the probability that the degradation
at time t exceeds a threshold v is equal to the probability that
the passage time T v (yrs) of the threshold v is less than time t, i.e.,
PrfT v 6 tg ¼ PrfX i ðtÞ P vg; ð2Þ
which is the cumulative density function (c.d.f.) of the passage time
T v (yrs), F T v ðtÞ. Hence, the c.d.f. and p.d.f. (probability density func-
tion) of the passage time T Ci and T Hi (yrs) can be derived based on
the degradation process X i ðtÞ, given C i and Hi respectively. Since
we assume the degradation X i ðtÞ is monotonic, X i ðtÞ will first cross
the control limit C i before reaching Hi (i.e., T C i < T Hi ). The CBM com-
ponent is eligible for preventive maintenance, only if C i < X i ðtÞ < Hi .
In other words, if there are opportunities between T Ci and T Hi for the
CBM component to do preventive maintenance together with the
SDs or USDs, we will take the first opportunity to maintain the
CBM component preventively. If no opportunity appears between
T C i and T Hi , we have to maintain the CBM component correctively,
once X i ðtÞ crosses the failure threshold Hi (i.e., at the time point T Hi ).
Without loss of generality, we consider T C i occurs in an interval
between the ðn  1Þth SD and the nth SD (yrs), i.e., ðn  1Þs  n 6
T C i < ns  n; n 2 N. Namely, X i ðtÞ reaches C at a time point
u 2 ½ðn  1Þs  n; ns  nÞ. The probabilities of the different mainte-
nance actions at the end of the maintenance cycle are derived
under the following two scenarios.
Scenario 1: ðn  1Þs  n 6 T C i < ns  n and T Hi < ns  n
Given ðn  1Þs  n 6 T C i < ns  n, if X i ðtÞ passes Hi at a time
point v before the next SD ns  n, i.e., T Hi ¼ v and v 2 ½u; ns  nÞ,
there will be no opportunity of preventive maintenance from
SDs. Hence, it is possible either to take the first opportunity of pre-
ventive maintenance from USDs or to conduct corrective mainte-
nance at the end of the maintenance cycle. The preventive
maintenance action with an USD will take place if the first USD
after T C i occurs before the degradation X i ðtÞ reaches the failure
340 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346

Z sn Z sn
threshold Hi . Since the arrival process of USDs from other compo-
P 2;i ðC i jnÞ ¼ ð1  eki ðv uÞ Þf T H jT C ðv juÞ dv f T C ðuÞ du
nents I n i is assumed to be a Poisson process (with arrival rate ki ), 0 u i i i

the probability that the preventive maintenance action with an Z sn Z 1 

USD takes place at the end of the maintenance cycle is þ ð1  eki ðsnuÞ Þ f T H jT C ðv juÞ dv f T C ðuÞ du
0 sn i i i

Z nsn Z nsn Z nsn Z nsn

X
1
ð1  eki ðv uÞ Þf T H jT C ðv juÞ dv f T C ðuÞ du; ð3Þ þ ð1  eki ðv uÞ Þf T H jT C ðv juÞ dv f T C ðuÞ du
ðn1Þsn u i i i
n¼2 ðn1Þsn u i i i

Z nsn Z !
where f T C ðuÞ is the p.d.f. of the passage time T C i and f T H jT C ðv juÞ is
1
i i i þ ð1  eki ðnsnuÞ Þ f T H jT C ðv juÞ dv f T C ðuÞ du :
the conditional p.d.f. of the passage time T Hi given that T C i ¼ u. In ðn1Þsn nsn i i i

Eq. (3), ð1  eki ðv uÞ Þ is the probability that the first USD after T C i ð8Þ
arrives before T Hi . On the other hand, if no opportunity of preven-
The conditional probability of a corrective maintenance action
tive maintenance is available before T Hi , a corrective maintenance
P3;i ðC i jnÞ is given as
action will take place with a probability
Z sn Z sn
Z nsn Z nsn P3;i ðC i jnÞ ¼ eki ðv uÞ f T H jT C ðv juÞ dv f T C ðuÞ du
eki ðv uÞ f T H jT C ðv juÞ dv f T C ðuÞ du: ð4Þ 0 u i i i

ðn1Þsn u i i i 1 Z nsn
X
þ
Scenario 2: ðn  1Þs  n 6 T C i < ns  n and T Hi P ns  n n¼2 ðn1Þsn

Given ðn  1Þs  n 6 T C i < ns  n, if X i ðtÞ passes Hi at a time Z nsn

point v after the next SD ns  n, i.e., T H ¼ v and v 2 ½ns  n; 1Þ,  eki ðv uÞ f T H jT C ðv juÞ dv f T C ðuÞ du: ð9Þ
i i i
u
it is not possible to have a corrective maintenance action since
there’s a SD for component i to be preventively maintained before Notice that the sum of P1;i ðC i jnÞ; P2;i ðC i jnÞ and P3;i ðC i jnÞ is equal to
failure. Our opportunistic maintenance policy will take the first one, since
opportunity (from either SDs or USDs) appears after T C i to per- Z sn 1 Z nsn
X
form preventive maintenance action. Hence, a preventive mainte- f T C ðuÞ du þ f T C ðuÞ du ¼ 1: ð10Þ
0 i
n¼2 ðn1Þsn i
nance action with a SD takes place if there’s no USD appeared
before the SD, ns  n. The probability that the preventive mainte- Based on the above analysis of random events at the end of the
nance action with a SD happens at the end of the maintenance maintenance cycle, the expected cycle length of component i condi-
cycle is tioned on n can be obtained by
Z nsn Z Z sn Z sn  
1
1 
eki ðnsnuÞ f T H jT C ðv juÞ dv f T C ðuÞ du; ð5Þ Li ðC i jnÞ ¼ u þ 1  eki ðv uÞ f T H jT C ðv juÞdv
ðn1Þsn nsn i i i 0 u ki i i
Z 1  
1 
where eki ðnsnuÞ is the probability that no USD after T C i appears þ uþ 1e ki ðsnuÞ
f T H jT C ðv juÞdv f T C ðuÞdu
sn ki i i i

before the SD, ns  n. On the other hand, if the first USD after T C i X1 Z nsn Z nsn  
1 
arrives before the SD ns  n, we will take the opportunity to con- þ u þ 1  eki ðv uÞ f T H jT C ðv juÞdv
n¼2 ðn1Þ s n u k i i i
duct preventive maintenance. In this case, the probability that the Z 1  
1 
preventive maintenance action with an USD takes place at the þ u þ 1  eki ðnsnuÞ f T H jT C ðv juÞdv gf T C ðuÞdu:
end of the maintenance cycle is nsn ki i i i

Z nsn Z 1
ð11Þ
ð1  e ki ðnsnuÞ
Þ f T H jT C ðv juÞ dv f T C ðuÞ du: ð6Þ Notice that since the maintenance cycle is at the magnitude of year
ðn1Þsn nsn i i i

whereas the downtime is at the magnitude of hours, the downtime

To summarize, under our opportunistic maintenance policy for
component i, the occurrence of the three possible maintenance
actions at the end of a maintenance cycle depends on which
event happens first, the failure of component i (i.e., the degrada-
tion level X i ðtÞ reaches Hi ), the first SD after T C i , or the first USD
after T C i . Considering the fact that the degradation level X i ðtÞ may
cross the control limit C i in any interval of SDs, the conditional
probability of a preventive maintenance action with a SD,
P1;i ðC i jnÞ (given that the deviation of a renewal point is n), is thus
given as
Z sn Z 1
P1;i ðC i jnÞ ¼ eki ðsnuÞ f T H jT C ðv juÞ dv f T C ðuÞ du
0 sn
1 Z nsn
X USD is taken at the end of the previous maintenance cycle. More-
þ eki ðnsnuÞ
n¼2 ðn1Þsn
Z 1
 f T H jT C ðv juÞ dv f T C ðuÞ du:
nsn i i
The conditional probability of a preventive maintenance action with
an USD P2;i ðC i jnÞ is
at the end of each maintenance cycle is ignored for the evaluation of
cycle lengths. As mentioned previously, the deviation of n is chang-
ing from maintenance cycle to maintenance cycle. We use a random
variable D to describe the distribution of the deviations among all
the maintenance cycles. The distribution of D depends on the ran-
dom events at the end of every maintenance cycle. If a preventive
maintenance action with a SD is taken at the end of a maintenance
cycle, the n for the next maintenance cycle will be equal to 0. If a
preventive maintenance action with an USD or a corrective mainte-
nance action is taken, the n for the next maintenance cycle can be
any possible value in ð0; sÞ. Since the arrivals of USDs are assumed
to follow a Poisson process, the n can take any value in ð0; sÞ with
i i i
equal chances when a preventive maintenance action with an
over, when a corrective maintenance action is taken, the n will be
approximately evenly-distributed over ð0; sÞ if the intervals of SDs
ð7Þ are relatively small compared with the average value of the lifetime
i
T Hi . Hence, the following probability density function can be used to
approximately describe the random variable D,
 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346 341

8
>
< q; ifn ¼ 0; maintenance action with an USD conditioned on the deviation
f D ðnÞ ¼ ð1qÞ
s ; if0 < n < s; ð12Þ n; P2;i ðAi jnÞ, is given as
>
: Z
0; otherwise: ni sn
P2;i ðAi jnÞ ¼ ð1  eki ðuAi Þ Þ f i ðuÞ du
We can see that D is uniformly-distributed on ð0; sÞ except that Ai
Z 1
there is a positive probability mass q at n ¼ 0. After specifying the þ ð1  eki ðni snAi Þ Þ f i ðuÞ du: ð18Þ
probability density function f D ðnÞ (or equivalently the parameter ni sn
q), the unconditional probability of a preventive maintenance action
The probability of a corrective maintenance action conditioned on
with a SD, P 1 , can be obtained by the law of total probability, (sim-
n; P3;i ðAi jnÞ, is given as
ilarly, the unconditional probabilities of a preventive maintenance
Z Ai Z ni sn
action with an USD, P2 , and a corrective maintenance action, P3 ,
can be obtained.)
P3;i ðAi jnÞ ¼ f i ðuÞ du þ eki ðuAi Þ f i ðuÞ du: ð19Þ
0 Ai
Z s
ð1  qÞ Notice that the sum of P 1;i ðAi jnÞ; P 2;i ðAi jnÞ and P 3;i ðAi jnÞ is equal to
P1 ¼ P1;i ðC i j0Þq þ P1;i ðC i jnÞ dn;
0 s one. The expected cycle length of component i conditioned on n
Z s
ð1  qÞ can be obtained by
P2 ¼ P2;i ðC i j0Þq þ P2;i ðC i jnÞ dn; ð13Þ
0 s Z Ai
Z s
ð1  qÞ Li ðAi jnÞ ¼ ufi ðuÞdu
P3 ¼ P3;i ðC i j0Þq þ P3;i ðC i jnÞ dn:
s
0
0 Z ni sn Z uAi 
þ ðAi þ sÞki eki s ds þ ueki ðuAi Þ f i ðuÞdu
Obviously, the accuracy of the approximation method is dependent Ai 0
on the selection of the probability density function f D ðnÞ or the Z 1 Z ni snAi 
parameter q. Notice that the parameter q represents the probability þ ðAi þ sÞki eki s ds þ ðni s  nÞeki ðni snAi Þ f i ðuÞdu:
ni sn 0
that a preventive maintenance action with a SD is taken at the end
ð20Þ
of the previous maintenance cycle, which should be equal to the
unconditional probability P 1 . Hence, the parameter q can be deter- The rest of the analysis to obtain Z i ðAi Þ is the same as the analysis
mined by solving the first equation in Eq. (13), which is given as for the CBM component.
Rs
P1;i ðC i jnÞ dn
q¼ 0
Rs : ð14Þ 3.2. Approximate evaluation of the arrival processes of USDs
s  sP1;i ðC i j0Þ þ P ðC i jnÞ dn
0 1;i
In the approximation methods of evaluation for either individ-
The expected cycle length of component i without the condition can
ual CBM components or ABM components (see Section 3.1), the
be derived by
arrival process of USDs for component i is assumed to be a Poisson
Z s
ð1  qÞ process with rate ki ; 8i 2 ICBM [ IABM . As we discussed in Section 2,
Li ðC i Þ ¼ Li ðC i j0Þq þ Li ðC i jnÞ dn: ð15Þ
0 s this is an approximation since the superposition of the failure pro-
cesses of all the other components I n i is not a Poisson process or a
The expected cycle downtime of component i can be obtained by
renewal process. Also, we have mentioned that under our oppor-
K i ðC i Þ ¼ P 1 di
PMSD
þ P2 di
PMUSD CM
þ P 3 di : ð16Þ tunistic maintenance policy, the failure processes of ABM and
CBM components are dependent through the USDs generated by
Using the concept of renewal theory (Ross, 1996), the expected all the components in the system. The unscheduled downs from
downtime per year of component i; Z i ðC i Þ, is approximately equal failed components are the opportunities of preventive mainte-
to K i ðC i Þ=Li ðC i Þ. Therefore, given the interval of scheduled downs nance for the still-working components. Due to these dependen-
s and the arrival rate of unscheduled downs ki , we can approxi- cies, it is not possible to evaluate the arrival rates of USDs for all
mately evaluate the expected downtime per year of component the CBM and ABM components (fki ji 2 ICBM [ IABM g) separately.
i; Z i ðC i Þ, by Eqs. (15) and (16). Therefore, we develop an iterative procedure to evaluate the arri-
Similar to the approximate evaluation of a CBM component, the val rates of USDs for all the CBM and ABM components, given that
average downtime per year of component i 2 IABM ; Z i ð:Þ, can be the decision variables are specified as ðs; C; AÞ.
approximately evaluated; see (Zhu, 2015) for the details. We pro-
vide the model in (Zhu, 2015) here for the convenience of the audi- Algorithm.
ence. Given the deviation n of a certain maintenance cycle, we can
derive the cycle downtime and cycle length by analyzing the ran- Step 1 Initiation (k ¼ 1):
dom events at the ends of maintenance cycles: a preventive main- 8i 2 IABM , set Ai ¼ 1, and calculate the failure rate of com-
tenance action with a SD, a preventive maintenance action with an ponent i at the first iteration, hi;1 ¼ 1=Li;1 , where Li;1 equals
USD and a corrective maintenance action. Under our opportunistic the expected lifetime of component i. (hi;1 refers to the
maintenance policy for component i, the occurrence of the three USDs generated by component i, whereas ki refers to the
possible maintenance actions at the end of a maintenance cycle USDs generated by components I n i). 8i 2 ICBM , set
depends on which event happens first, the failure of component C i ¼ Hi , and calculate the failure rate of component i at
i, the first SD after the age limit Ai , or the first USD after Ai . The the first iteration, hi;1 ¼ 1=Li;1 , where Li;1 equals the
probability of a preventive maintenance action with a SD condi- expected failure time of component i, i.e., Li;1 ¼ E½T Hi .
tioned on n; P 1;i ðAi jnÞ, is equal to 8i 2 IFBM , calculate the failure rate of component
Z 1 i; hi ¼ 1=li , where li is the expected value of the lifetime
P1;i ðAi jnÞ ¼ eki ðni snAi Þ f i ðuÞ du; ð17Þ of component i. (since the failure rate of a FBM component
ni sn
i does not depend on the failure rates of other components,
where ni ¼ dAisþne is the index of the first SD after Ai counted from the hi will not change for different iterations).
beginning of each maintenance cycle, and f i ð:Þ is the p.d.f. of the life- Step 2 Repeat the following iteration with the updated hi;k until
time distribution for component i. The probability of a preventive jhi;k  hi;k1 j < e for all the CBM and ABM components:
342 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346
 For all i 2 IABM : By using the rate of USDs
P
ki ¼ j2fInfigg hj;k , we can evaluate the probability of a
corrective maintenance action P 3i;k ðs; Ai Þ and the
expected cycle length Li;k ðs; Ai Þ. Then we update
hi;kþ1 ¼ P3i;k ðs; Ai Þ=Li;k ðs; Ai Þ;
(see Eqs. (19) and (20) in Section 3.1). For all i 2 ICBM : By
P i; hi ¼ 1=li , where li is the expected value of the lifetime
using the rate of USDs ki ¼ j2fInfigg hj;k , we can evaluate the
probability of a corrective maintenance action P 3i;k ðs; C i Þ and
the expected cycle length Li;k ðs; C i Þ. Then we update
hi;kþ1 ¼ P3i;k ðs; C i Þ=Li;k ðs; C i Þ;
(see Eqs. (13) and (15) in Section 3.1).
 Let k :¼ k þ 1.
P
Step 3 Obtain the ~k ¼ fki ; i 2 IABM [ ICBM jki ¼ j2fInfigg hj;k g.
In step 2 of the above algorithm, the iteration calculates the val-
ues hkþ1 given the values of the previous iteration hk , which can be
seen as a map L. Let D be the domain of L; it is a closed, bounded
and convex subset ½0; UB1   ½0; UB2   . . . ½0; UBjIABM jþjICBM j  of
RjIABM jþjICBM j , where UBi (i 2 f1; 2; . . . ; jIABM j þ jICBM jg) is a sufficiently
large real number to cover the range of L. The range of L is bounded
in RjIABM jþjICBM j since in step 2 the probability of a corrective mainte-
nance action is bounded by ½0; 1 and the expected cycle length is
RA
lower-bounded by EðT C i Þ for CBM components and by 0 i ufi ðuÞdu
for ABM components. Furthermore, the map L is continuous on
D, which can be easily verified based on Eq. (9), (11), (19) and
(20). Thus according to Brouwer’s fixed point theorem (Carl &
Heikkil, 2011), the continuous map L : D ! D has a fixed point at
which h0 ¼ Lðh0 Þ. The above algorithm provides one method to
approach to a fixed point and terminates when hkþ1 is close enough
to hk .
After obtaining the approximated arrival rates of USDs for all
the CBM and ABM components, ~ k, the approximated downtime
per year for all the CBM and ABM components,
fZ i ðs; C; AÞ; i 2 IABM [ ICBM g, can thus be determined through the
approximation methods proposed in Section 3.1, given that the
decision variables are specified as ðs; C; AÞ.
4. Solution approach
The optimal solution ðs ; C ; A Þ for a system consisting of mul-
tiple components under difference maintenance policies
fICBM [ IABM [ IFBM g can be obtained by minimizing Eq. (1). As we
discussed before, the optimization problem is not decomposable
P
with respect to i2I Z i ðs; C; AÞ, due to the dependency of the failure
processes of the ABM and CBM components. Hence, we develop a
heuristic approach based on our approximation method. In this
heuristic approach, we choose the values of ðs; C; AÞ in a nested
e and age limits
way. For a given s, we search for the control limits C
e P
A that minimize i2I Z i ðs; C; AÞ by an iterative procedure. Then we
specify the value of s to minimize Z syst ðs; C; e by enumeration.
e AÞ
Algorithm.
Step 1 For every s 2 ðsLB ; sUB Þ, evaluate and minimize the down-
time per year of Z i ðs; C; e (ðsLB ; sUB Þ is the range of search
e AÞ.
for s)
Step 1.1 Initiation (k ¼ 1):
8i 2 IABM , set Ai ¼ 1, and calculate the failure rate of
component i at the first iteration, hi;1 ¼ 1=Li;1 , where Li;1
equals the expected lifetime of component i. (hi;1 refers
to the USDs generated by component i, whereas ki refers
to the USDs generated by components I n i).
8i 2 ICBM , set C i ¼ Hi , and calculate the failure rate of com-
ponent i at the first iteration, hi;1 ¼ 1=Li;1 , where Li;1 equals
the expected failure time of component i, i.e., Li;1 ¼ E½T Hi .
8i 2 IFBM , calculate the failure rate of component
of component i. (since the failure rate of a FBM compo-
nent i does not depend on the failure rates of other com-
ponents, hi will not change for different iterations).
Step 1.2 Repeat the following iteration with the updated hi;k until
jhi;k  hi;k1 j < e for all the CBM and ABM components:
 For all i 2 IABM : By using the rate of USDs
P
ki ¼ j2fInfigg hj;k , we can find the optimal age limit
Ai;k to minimize Z i ð:Þ in a certain range
ðAi LB ; AUB
i Þ(by enumeration), and update
hi;kþ1 ¼ P3i;k ðs; Ai;k Þ=Li;k ðs; Ai;k Þ;
where P3i;k ðs; Ai;k Þ is the probability of a corrective mainte-
nance action and Li;k ðs; Ai;k Þ is the expected cycle length (see
Eqs. (19) and (20) in Section 3.1).
P
For all i 2 ICBM : By using the rate of USDs ki ¼ j2fInfigg hj;k , we
can find the optimal control limit C i;k to minimize Z i ð:Þ in a
certain range ðC i LB ; C UB
i Þ(by enumeration), and update
hi;kþ1 ¼ P3i;k ðs; C i;k Þ=Li;k ðs; C i;k Þ;
where P3i;k ðs; C i;k Þ is the probability of a corrective mainte-
nance action and Li;k ðs; C i;k Þ is the expected cycle length (see
Eqs. (13) and (15) in Section 3.1).
 Let k :¼ k þ 1.
Step 1.3 Obtain the A e ¼ fA ; i 2 IABM g and C
e ¼ fC  ; i 2 ICBM g for
i;k i;k
e is minimized for each s.
e AÞ
each s ; Z i ðs; C;
Step 2 Optimize s with respect to Z syst ðs; C; AÞ. Given the optimal
e and C
e can be obtained from the results of
interval s ; A
Step 1.
Other genetic algorithms can be good for problems with no spe-
cial structure. For this particular problem, the dependency of indi-
vidual components is through the arrival of unscheduled downs
and scheduled downs. This unique feature can help us partially
decompose the overall optimization problem into sub-
optimization problems for individual components, which can
increase the efficiency of searching. Note that the upper bound
and the lower bound of s can be determined by the industry regu-
lations and workforce management.
5. Numerical Study
To demonstrate the usage of the model, we start with a simple
example: a system consisting of 5 condition-based components
(ICBM ¼ f1; 2; 3; 4; 5g) and 5 age-based components (IABM ¼ f6; 7;
8; 9; 10g) and 40 failure-based components (IFBM ¼ f11; 12; . . .
49; 50g). To show our model is able to deal with systems consisting
of both identical and non-identical components, we consider the
case where (i) component i 2 IABM and i 2 ICBM are non-identical.
(ii) components i 2 IFBM consisting of 4 identical subgroups
IFBM1 ¼ f11; . . . ; 20g; IFBM2 ¼ f21; . . . ; 30g; IFBM3 ¼ f31; . . . ; 40g and
IFBM4 ¼ f41; . . . ; 50g. Each of these identical subgroups consists of
10 non-identical components with their expected life time (yrs)
fl11 ; . . . ; l20 g = fl21 ; . . . ; l30 g = fl31 ; . . . ; l40 g = fl41 ; . . . ; l50 g =
 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346 343

Table 1
The parameter setting.

Parameters Explanation Value for i ¼ f1; . . . ; 5g IABM Value for i ¼ f6; . . . ; 10g ICBM
PMUSD
di Preventive maintenance time with an USD [hour] 2:00; 2:10; 2:20; 2:30; 2:40 2:50; 2:60; 2:70; 2:80; 2:90
PMSD
di Preventive maintenance time with a SD [hour] 1:00; 1:05; 1:10; 1:15; 1:20 1:25; 1:30; 1:35; 1:40; 1:45
Rep Corrective replacement time [hour] 5:00; 5:50; 6:00; 6:50; 7:00 7:50; 8:00; 8:50; 9:00; 9:50
di
USD Unscheduled setup time [hour] 5 5
S
SSD Scheduled setup time [hour] 3 3
aABM
i
Scale parameter of Weibull distribution for component i’s lifetime (i 2 IABM ) 1:13; 1:24; 1:35; 1:46; 1:58 –
bABM
i
Shape parameter of Weibull distribution for component i’s lifetime (i 2 IABM ) 2:10; 2:31; 2:52; 2:73; 2:94 –
aCBM
i
parameter of Gamma process for component i’s degradation process (i 2 ICBM ) – 2:00; 2:00; 1:35; 1:46; 1:58
bCBM
i
parameter of Gamma process for component i’s degradation process (i 2 ICBM ) – 1:85; 1:25; 2:21; 1:81; 2:41
Hi Failure threshold (for condition-based components) – 3:00; 3:50; 3:50; 4:00; 2:41

Table 2
The base values of the four factors in the test bed

Factor Parameters in a factor Base values

D fC i ; i 2 ICBM g [ fAi ; i 2 IABM g
s fsg[yrs]
l fli ; i 2 IFBM g[yrs]
r frðT i Þ; i 2 IABM g [ frðT Hi Þ; i 2 ICBM g
{2, 2.33, 2.33, 2.67, 1.61} [{1.13, 1.24, 1.35, 1.46, 1.58}
{0.5}
{1.33,1.47,1.60,1.73,1.87,2,2.13,2.27,2.4,2.53;
1.33,1.47,1.60,1.73,1.87,2,2.13,2.27,2.4,2.53;
1.33,1.47,1.60,1.73,1.87,2,2.13,2.27,2.4,2.53;
1.33,1.47,1.60,1.73,1.87,2,2.13,2.27,2.4,2.53;}
{0.50, 0.50, 0.51, 0.51, 0.52} [{1.17, 1.04, 2.05, 1.83, 1.51}

Table 3
The average gaps between the simulation results and approximation results in the test bed

Factor Three levels The average gaps of fDZ i ; i 2 IABM [ ICBM g The average percentage of
overestimated components
D 75% f0:0045; 0:0050; 0:0060; 0:0061; 0:0071g [ f0:0037; 0:0038; 0:0050; 0:0039; 0:0045g 54.07%
100% f0:0037; 0:0039; 0:0034; 0:0042; 0:0043g [ f0:0041; 0:0037; 0:0052; 0:0059; 0:0041g 54.81%
125% f0:0031; 0:0032; 0:0040; 0:0040; 0:0045g [ f0:0080; 0:0067; 0:0060; 0:0062; 0:0062g 57.04%
s 75% f0:0038; 0:0046; 0:0050; 0:0044; 0:0045g [ f0:0049; 0:0049; 0:0052; 0:0050; 0:0043g 57.04%
100% f0:0036; 0:0036; 0:0046; 0:0046; 0:0052g [ f0:0048; 0:0039; 0:0053; 0:0056; 0:0057g 52.22%
125% f0:0038; 0:0039; 0:0039; 0:0053; 0:0062g [ f0:0061; 0:0054; 0:0058; 0:0054; 0:0048g 56.67%
l 75% f0:0044; 0:0049; 0:0045; 0:0042; 0:0053g [ f0:0049; 0:0046; 0:0062; 0:0047; 0:0048g 54.81%
100% f0:0030; 0:0035; 0:0045; 0:0048; 0:0059g [ f0:0049; 0:0045; 0:0053; 0:0039; 0:0052g 53.70%
125% f0:0038; 0:0036; 0:0045; 0:0053; 0:0047g [ f0:0059; 0:0053; 0:0047; 0:0074; 0:0048g 57.41%
r 75% f0:0037; 0:0037; 0:0041; 0:0038; 0:0046g [ f0:0045; 0:0040; 0:0055; 0:0042; 0:0048g 50.00%
100% f0:0034; 0:0040; 0:0048; 0:0050; 0:0060g [ f0:0049; 0:0048; 0:0057; 0:0049; 0:0045g 57.41%
125% f0:0041; 0:0044; 0:0046; 0:0056; 0:0052g [ f0:0062; 0:0055; 0:0050; 0:0069; 0:0055g 58.52%

f1.33, 1.47, 1.60, 1.73, 1.87, 2.00, 2.13, 2.27, 2.40, 2.53 g (exponen- 5.1. Accuracy of the approximate evaluation
tially distributed) and their corrective maintenance time (hrs)
CM CM CM CM CM CM CM CM
fd11 ; . . . ; d20 g =fd21 ; . . . ; d30 g = fd31 ; . . . ; d40 g = fd41 ; . . . ; d50 g = f The accuracy of our approximate evaluation is assessed by the
6.00, 6.10, 6.20, 6.30, 6.40, 6.50, 6.60, 6.70, 6.80, 6.90g. Then the relative gap between the simulation result fZ ^ i ðs; C; AÞ; i 2 ICBM
average downtime per year of these 40 FBM components is [IABM g and the approximate result fZ i ðs; C; AÞ; i 2 ICBM [ IABM g, i.e.,
P
i2IFBM Z i ¼ 137:82 hours per year.
^ i ðs; C; AÞ  Z i ðs; C; AÞj=Z
fDZ i ¼ jZ ^ i ðs; C; AÞ; i 2 ICBM [ IABM g. We vary
For component i 2 ICBM , we use a Gamma process (van four factors in our test bed: the decision variables fC; Ag; the inter-
Noortwijk, 2009) with parameter aCBM i and bCBM
i to model its degra- val of scheduled downs s; the expected lifetime of FBM components
dation path (Gamma process is a continuous time and continuous fli ; i 2 IFBM g; the standard deviations of the lifetime distributions
state Markov process with independent and stationary incre- for ABM components and the standard deviations of the passage
ments). The initial degradation value is assumed to be zero time distributions over the failure thresholds for CBM components
(X i ð0Þ ¼ 0) and the failure threshold is Hi . For component i 2 IABM , frðT i Þ; i 2 IABM g [ frðT Hi Þ; i 2 ICBM g 1. For each of the four factors,
we assume the lifetime distribution is a Weibull distribution with three different levels are obtained by multiplying base values with
its scale and shape parameter aABM
i and bABM
i . The input parameters a set of coefficients, f75%; 100%; 125%g. The base values of the four
are given in Table 1. As mentioned previously in Section 2,the fixed factors are given in Table 2. The rest of the parameters are set at
the fixed values given in Table 1 and Section 5.
setup time of scheduled down SSD is smaller than the fixed setup
time of unscheduled down SUSD ; corrective maintenance time
CM Rep
di ¼ di þ SUSD must be higher than preventive maintenance time 1
The standard deviations are obtained by setting aABM
i ; bABM
i ; aCBM
i ; bCBM
i ; Hi at the
PMUSD PMSD
(di or di ). values given in Table 1.
344 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346

Table 4
The maximum gaps between the simulation results and approximation results in the test bed

Factor Three levels The maximum gaps of fDZ i ; i 2 IABM [ ICBM g

D 75% {0.0116, 0.0134, 0.0120, 0.0137, 0.0175} [{0.0089, 0.0108, 0.0131, 0.0159, 0.0098}
100% {0.0132, 0.0142, 0.0087, 0.0086, 0.0136} [{0.0112, 0.0126, 0.0122, 0.0229, 0.0126}
125% {0.0088, 0.0084, 0.0108, 0.0105, 0.0120} [{0.0165, 0.0184, 0.0187, 0.0170, 0.0147}
s 75% {0.0116, 0.0138, 0.0120, 0.0116, 0.0105} [{0.0145 ,0.0183, 0.0145, 0.0229, 0.0147}
100% {0.0084, 0.0142 ,0.0116, 0.0137, 0.0145} [{0.0158, 0.0154 ,0.0135, 0.0147, 0.0130}
125% {0.0132, 0.0134, 0.0113, 0.0113, 0.0175} [{0.0165, 0.0184, 0.0187, 0.0162, 0.0126}
l 75% {0.0116, 0.0138, 0.0120, 0.0137, 0.0145} [{0.0165, 0.0184, 0.0187, 0.0147, 0.0130}
100% {0.0085, 0.0103, 0.0116, 0.0116, 0.0147} [{0.0129, 0.0154, 0.0104, 0.0146, 0.0126}
125% {0.0132, 0.0142, 0.0099, 0.0113, 0.0175} [{0.0158, 0.0183, 0.0135, 0.0229, 0.0147}
r 75% {0.0075, 0.0073, 0.0120, 0.0137, 0.0145} [{0.0145, 0.0089, 0.0135, 0.0147, 0.0126}
100% {0.0132, 0.0122, 0.0110, 0.0101, 0.0175} [{0.0128, 0.0183, 0.0131, 0.0146, 0.0147}
125% {0.0116, 0.0142, 0.0113, 0.0116, 0.0147} [{0.0165, 0.0184, 0.0187, 0.0229, 0.0130}

Table 5
The heuristic solution of one instance.

The age limits for IABM The control limits for ICBM s [yr] Z syst [hrs per yr]

f0:4181; 0:4588; 0:43875; 0:4088; 0:4424g f2:01; 2:1875; 3:0555; 3:024; 1:72315g 0.46 27.46707026
P1 f0:34500:37600:40680:86720:8895g f0:54590:54550:55440:55480:5341g
P2 f0:28330:32780:39580:08280:0791g f0:39400:39300:39920:39980:3890g
Li f0:66680:69480:66980:47190:4714g f1:81541:28723:20142:13472:9566g

Table 6
The heuristic solutions under different parameter settings

Factor Three levels The control limits and age limits fAi ; i 2 IABM g [ fC i ; i 2 ICBM g s Z syst

r 75% f0:42375; 0:4092; 0:38475; 0:4161; 0:8058g [ f1:875; 2:345; 3:465; 3:492; 1:6147g 0.430769231 21.60865739
100% f0:4181; 0:4588; 0:43875; 0:4088; 0:4424g [ f2:01; 2:1875; 3:0555; 3:024; 1:72315g 0.46 27.46707026
125% f0:4633; 0:4526; 0:432; 0:4672; 0:4345g [ f2:01; 2:1875; 2:695; 3:08; 1:72315g 0.469230769 34.17628123
SSD 1 f0:41810:45880:438750:40880:4424g [ f2:01; 2:1875; 3:0555; 3:024; 1:72315g 0.46 27.46707026
3 f0:4746; 0:4712; 0:459; 0:4964; 0:474g [ f1:95; 2:135; 2:625; 3; 1:6629g 0.507692308 31.88488385
5 f0:4746; 0:5208; 0:50625; 0:4818; 0:5214g [ f1:95; 2:135; 2:485; 3; 1:6629g 0.525 35.52731326
PMSD 100% f0:4181; 0:4588; 0:43875; 0:4088; 0:4424g [ f2:01; 2:1875; 3:0555; 3:024; 1:72315g 0.46 27.46707026
di
150% f0:4972; 0:496; 0:54; 0:5256; 0:5056g [ f2:01; 2:1875; 2:7825; 3:024; 1:72315g 0.54 32.61658097
200% f0:4068; 0:4464; 0:54; 0:6424; 0:632g [ f2:01; 2:345; 2:765; 3; 1:72315g 0.68 36.35415815
di
CM 50% f0:678; 0:62; 0:675; 0:657; 0:632g [ f2:34; 2:73; 3:465; 3:648; 1:9762g 0.34 21.30054877
100% f0:4181; 0:4588; 0:43875; 0:4088; 0:4424g [ f2:01; 2:1875; 3:0555; 3:024; 1:72315g 0.46 27.46707026
200% f0:30962; 0:33976; 0:3105; 0:3358; 0:29388g [ f1:71; 1:8375; 3:465; 3:648; 1:48215g 0.34 35.15965602

The results of the test bed are summarized in Tables 3 and 4. On
average, the relative difference DZ i (8i 2 IABM [ ICBM ) between the
simulation results and approximation results for all the instances
in the test bed is 0.48%, which is small. This shows that the approx-
imation approach proposed by us is accurate. The maximum rela-
tive difference among all the components and all the instances in
the test bed is 2.29% (see Table 4), which is also small.
5.2. Heuristic policies based on the approximate evaluation
The heuristic solutions can be obtained through the solution
approach proposed in Section 4.
To give a numerical example of the heuristic policy, we set the
parameters at the values given in Table 7. The heuristic solution is
provided in Table 5.2 By using this heuristic policy, the probabilities
of preventive maintenance with a SD for each component are given
in the row beginning with P1 ; the probabilities of preventive main-
tenance with an USD are given in the row beginning with P2 . The
probabilities of preventive maintenance with a SD are higher than
the probabilities of preventive maintenance with an USD, since in
2 ing the usage time of a system. Furthermore, when the standard
In this table, the average downtime per year of FBM components are not included
in Z syst since they are not influenced by the decision variables.
this numerical example PM with an USD are more time consuming
than PM with a SD, i.e., di
PMUSD
> di
PMSD
.
We investigate the changes of the heuristic solutions under dif-
ferent parameter settings, i.e., the standard deviations of the life-
time distribution for ABM components and the standard
deviations of the passage time distributions for CBM components
frðT i Þ; i 2 IABM g [ frðT Hi Þ; i 2 ICBM g, the fixed setup time of sched-
uled down SSD , the preventive maintenance time for scheduled
PMSD
downs fdi ; i 2 IABM [ ICBM g, and the corrective maintenance
CM
time fdi ;
i 2 IABM [ ICBM g. The heuristic solutions for each factor
under three levels are presented in Table 6. The value of one factor
is changed for each instance, while the rest of the parameters are
set at the base values as in Table 7.
From Table 6, we can observe that the changing patterns of the
heuristic solutions are sophisticated due to the fact that there are
multiple decision variables jointly influencing the average down-
time per year of a system. When the standard deviations of the life-
times for ABM and CBM components r increase, the average
downtime per year increases, since the failure processes become
more difficult to control and more unexpected failures appear dur-
deviations r increase, the frequency of scheduled downs decreases
 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346 345

Table 7
The setting of base values in SubSection 5.2.

Parameters Explanation Value for i ¼ f1; . . . ; 5g Value for i ¼ f6; . . . ; 10g

IABM ICBM

di
PMUSD Preventive maintenance time with an USD [hour] 2:00; 2:10; 2:20; 2:30; 2:40 2:50; 2:60; 2:70; 2:80; 2:90
di
PMSD Preventive maintenance time with a SD [hour] 1:00; 1:05; 1:10; 1:15; 1:20 1:25; 1:30; 1:35; 1:40; 1:45
Rep Corrective replacement time [hour] 5:00; 5:50; 6:00; 6:50; 7:00 7:50; 8:00; 8:50; 9:00; 9:50
di
SUSD Unscheduled setup time [hour] 5 5
SSD Scheduled setup time [hour] 1 1
aABM
i
Scale parameter of Weibull distribution for component i’s lifetime (i 2 IABM ) 1:13; 1:24; 1:35; 1:46; 1:58 –
bABM
i
Shape parameter of Weibull distribution for component i’s lifetime (i 2 IABM ) 2:10; 2:31; 2:52; 2:73; 2:94 –
aCBM
i
parameter of Gamma process for component i’s degradation process (i 2 ICBM ) – 2:00; 2:00; 1:35; 1:46; 1:58
bCBM
i
parameter of Gamma process for component i’s degradation process (i 2 ICBM ) – 1:85; 1:25; 2:21; 1:81; 2:41
Hi Failure threshold (for condition-based components) – 3:00; 3:50; 3:50; 4:00; 2:41
Parameters Explanation Value for i ¼ f11; 12 . . . 50g; IFBM
li The expected lifetime for component i [year] {26.6, 29.4, 32, 34.6, 37.4, 40, 42.6, 45.4, 48, 50.6;
26.6, 29.4, 32, 34.6, 37.4, 40, 42.6, 45.4, 48, 50.6;
26.6, 29.4, 32, 34.6, 37.4, 40, 42.6, 45.4, 48, 50.6;
26.6, 29.4, 32, 34.6, 37.4, 40, 42.6, 45.4, 48, 50.6;}

and the probability of preventive maintenance with USDs
increases, which may due to the increased randomness of failure
processes. When the fixed setup time of scheduled downs SSD
increases, the interval of scheduled downs increases and the prob-
ability of corrective maintenance increases, which leads to the
increase of the average downtime per year. For CBM components,
when SSD increases, the control limits are nonincreasing in order
to compensate for the reduced opportunities from scheduled
downs. When the preventive maintenance time for scheduled
PMSD
downs di increases, the frequency of scheduled downs
decreases and the average downtime per year increases, since
the benefit of using the opportunities from scheduled downs
CM
becomes less. When the corrective maintenance cost di increases,
the average downtime per year increases and the age limits of ABM
components decreases to avoid the higher corrective maintenance
time.
structure of the optimal contract can be obtained. Also, other types
of dependencies among components can be investigated. For
example, components can be integrated into modules so that the
process of diagnosis or replacing the failed module can be much
faster, which will further reduce the downtime. A drawback of
modular design is that the cost of holding spare parts rises because
modules are in general subject to higher failure rates than their
constituent elements. Thus, the optimal modular design should
be investigated.
Acknowledgement
This research was partially funded by a grant from the Chinese
Ministry of Science and Technology (Grant No. 2016YFF0203801),
and was supported by the President Fund of the Academy of Math-
ematics and Systems Science and the one-hundred plan (class C) of
the Chinese Academy of Sciences (No. Y629141CC1-4199900).

6. Conclusions Appendix A. Derivation of the expected cycle length

For a multi-component system under a mix of condition-based, L_ i ðC i jnÞ can be obtained by.
age-based, and failure based policies, we have proposed a class of
opportunistic maintenance policies using scheduled and unsched- Z sn Z sn Z v u Z 
1
uled downs. In the framework of this opportunistic maintenance ðu þ sÞki eki s ds þ v ki eki s ds f T H jT C ðv juÞdv
policy, we developed approximate methods to evaluate the 0 u 0 v u i i

Z 1 Z snu Z 1  
expected downtime per year. The shifts of the renewal points have þ ðu þ sÞki e ki s
ds þ ðs  nÞ ki e ki s
ds f T H jT C ðv juÞdv f T C ðuÞdu
sn 0 snu i i i
been considered in the approximate models, which are caused by Z nsn Z v u 
1 Z nsn
X Z 1
the periodic schedules of preventive maintenance. Based on the þ ðu þ sÞki e ki s
ds þ v ki e ki s
ds f T H jT C ðv juÞdv
approximation models, we proposed a heuristic procedure to n¼2 ðn1Þsn u 0 v u i i

Z 1 Z nsnu Z 1  
determine the interval length for the scheduled downs, the control þ ðu þ sÞki eki s ds þ ðns  nÞ ki eki s ds f T H jT C ðv juÞdv f T C ðuÞdu:
limits of the condition-based components and the age limits of nsn 0 nsnu i i i

age-based components, in order to minimize the expected down- ð21Þ

time per year. The proposed approaches will help the decision Using integration by part, we have
makers approximately evaluate the expected downtime per year Z v u Z 1
1 
of a service plan, when the joint maintenance decisions are speci- ðu þ sÞki eki s ds þ v ki eki s ds ¼ u þ 1  eki ðv uÞ ;
fied. Moreover, the heuristic optimization procedure will help the 0 v u ki
decision makers find out a service plan that minimizes the and
expected downtime per year of a system. Z nsnu Z 1
In the current trend of performance-based contracts, future
ðu þ sÞki eki s ds þ ðns  nÞ ki eki s ds
research can be conducted on the contracting issues between the 0 nsnu
OEMs and the customers. Due to the significant uncertainties in 1 
repair processes, it is very hard to guarantee a certain service level ¼uþ 1  eki ðnsnuÞ :
ki
or quote a price for providing it. Cost-sharing arrangements are
often employed for the maintenance support. Insights into the Hence, L_ i ðC i jnÞ can be rewritten as
346 H. Peng, Q. Zhu / Computers & Industrial Engineering 109 (2017) 335–346

Z sn Z sn  
1  Christer, A. H., & Lee, C. (2000). Refining the delay-time-based pm inspection model
uþ 1  eki ðv uÞ f T H jT C ðv juÞ dv with non-negligible system downtime estimates of the expected number of
0 u ki i i
Z 1   failures. International Journal of Production Economics, 67(1), 77–85.
1  Cochran, J. K., Murugan, A., & Krishnamurthy, V. (2001). Generic markov models for
þ uþ 1e ki ðsnuÞ
f T H jT C ðv juÞ dv f T C ðuÞ du availability estimation and failure characterization in petroleum refineries.
sn ki i i i
Computers & Operations Research, 28(1), 1–12.
X1 Z nsn Z nsn  
1  Csenki, A. (1994). Joint availability of systems modeled by finite semi-markov
þ uþ 1  eki ðv uÞ f T H jT C ðv juÞ dv processes. Applied Stochastic Models and Data Analysis, 10(4), 279–293.
n¼2 ðn1Þsn u ki i i
de Smidt-Destombes, K. S., van der Heijden, M. C., & van Harten, A. (2004). On the
Z 1   
1  availability of a k-out-of-n system given limited spares and repair capacity
þ uþ 1  eki ðnsnuÞ f T H jT C ðv juÞ dv f T C ðuÞ du: ð22Þ under a condition based maintenance strategy. Reliability Engineering & System
nsn ki i i i Safety, 83(3), 287–300.
Do, P., Vu, H. C., Barros, A., & Berenguer, C. (2015). Maintenance grouping for multi-
component systems with availability constraints and limited maintenance
teams. Reliability Engineering & System Safety, 142, 56–67.
Appendix B. The base values of the numerical experiment Du, S. J., Cui, L. R., & Lin, C. (2016). Some reliability indexes and sojourn time
distributions for a repairable degradation model. Proceedings of the Institution of
The base values of the numerical experiment in SubSection 5.2 Mechanical Engineers Part O-Journal of Risk and Reliability, 230(3), 334–349.
Ebeling, C. E. (2009). An introduction to reliability and maintainability engineering.
are given in Table 7. Waveland Pr Inc.
El Hajj, B., Castanier, B., Schoefs, F., & Yeung, T. (2016). A risk-oriented degradation
model for maintenance of reinforced concrete structure subjected to cracking.
Appendix C. Simulation procedures Proceedings of the Institution of Mechanical Engineers Part O-Journal of Risk and
Reliability, 230(5), 521–530.
To evaluate the accuracy of the approximation, we run a simu- Faulin, J., Juan, A. A., Serrat, C., & Bargueno, V. (2008). Predicting availability
functions in time-dependent complex systems with saedes simulation
lation to compare with the approximate evaluation results in Sec- algorithms. Reliability Engineering & System Safety, 93(11), 1761–1771.
tion 5. There are m runs in the simulation (e.g., m ¼ 100). For each Jiang, R., Kim, M. J., & Makis, V. (2013). Availability maximization under partial
run i 2 f1; 2; . . . ; mg, we generate (1) a Poisson process with a rate k observations. OR Spectrum, 35(3), 691–710.
Kim, S. H., Cohen, M. A., & Netessine, S. (2007). Performance contracting in after-
and random arrival time points D ¼ fd1 ; d2 ; . . .g; (2) a set of random sales service supply chains. Management Science, 53(12), 1843–1858.
failure times T ¼ ft 1 ; t 2 ; . . .g; and (3) a constant set B ¼ fs; 2s; . . .g, Kreimer, J., & Mehrez, A. (1998). Computation of availability of a real-time system
using queueing theory methodology. Journal of the Operational Research Society,
on a time horizon ð0; T max Þ that is sufficiently large(e.g., 106 times
49(10), 1095–1100.
larger than the mean value of the lifetime). Then by using the Lee, H., & Cha, J. H. (2016). New stochastic models for preventive maintenance and
method of discrete event simulation, we record the number of pre- maintenance optimization. European Journal of Operational Research, 255(1),
80–90.
ventive maintenance actions with USDs, the number of preventive
Liu, B., Xie, M., & Kuo, W. (2016). Reliability modeling and preventive maintenance
maintenance actions with SDs and the number of corrective main- of load-sharing systems with degrading components. IIE Transactions, 48(8),
tenance actions separately for each run, as well as the average 699–709.
cycle length. The probabilities of preventive maintenance actions Marquez, A. C., Heguedas, A. S., & Iung, B. (2005). Monte carlo-based assessment of
system availability. a case study for cogeneration plants. Reliability Engineering
with USDs, preventive maintenance actions with SDs and correc- & System Safety, 88(3), 273–289.
tive maintenance actions can be calculated based on these results Naseri, M., Baraldi, P., Compare, M., & Zio, E. (2016). Availability assessment of oil
for each run. We also can compute the expected downtime per and gas processing plants operating under dynamic arctic weather conditions.
Reliability Engineering & System Safety, 152, 66–82.
year for each run Z^i after obtaining the probabilities of the three Raje, D. V., Olaniya, R. S., Wakhare, P. D., & Deshpande, A. W. (2000). Availability
types of actions and the average cycle length. The final result of assessment of a two-unit stand-by pumping system. Reliability Engineering &
System Safety, 68(3), 269–274.
the simulation is the mean value of the m runs, with a
Ross, S. M. (1996). Stochastic processes (Second ed.). John Wiley & Sons.
100ð1  aÞ% confidence interval Timmermans, B. (2012). Development and application of decision model for
sffiffiffiffiffi synchronizing condition-based maintenance. Master of science thesis,
Eindhoven University of Technology.
S2
Z^  tð1  a=2; m  1Þ Trivedi, K. S., Kim, D. S., & Ghosh, R. (2013). System availability assessment using
m stochastic models. Applied Stochastic Models in Business and Industry, 29(2),
94–109.
P ðZ^i ZÞ
^ 2 Tsai, Y. T., Wang, K. S., & Tsai, L. C. (2004). A study of availability-centered
where S ¼ m i¼1 m1 and tð1  a=2; m  1Þ is the upper 1  a=2 crit- preventive maintenance for multi-component systems. Reliability Engineering &
ical point for the t-distribution with ðm  1Þ degrees of freedom (in System Safety, 84(3), 261–270.
our case, m ¼ 100 and a ¼ 5%). van Dijkhuizen, G., & van der Heijden, M. (1999). Preventive maintenance and the
interval availability distribution of an unreliable production system. Reliability
Engineering & System Safety, 66(1), 13–27.
References van Dorst, N. (2014). Component condition monitoring and remaining useful life
prediction. Master of science thesis, Eindhoven University of Technology.
van Noortwijk, J. M. (2009). A survey of the application of gamma processes in
Ahmadi, R., & Fouladirad, M. (2017). Maintenance planning for a deteriorating
maintenance. Reliability Engineering & System Safety, 94(1), 2–21.
production process. Reliability Engineering & System Safety, 159, 108–118.
Verbert, K., De Schutter, B., & Babuska, R. (2017). Timely condition-based
Alaswad, S., & Xiang, Y. S. (2017). A review on condition-based maintenance
maintenance planning for multi-component systems. Reliability Engineering &
optimization models for stochastically deteriorating system. Reliability
System Safety, 159, 310–321.
Engineering & System Safety, 157, 54–63.
Wang, H., & Pham, H. (2006). Availability and maintenance of series systems subject
Bloch-Mercier, S. (2000). Stationary availability of a semi-markov system with
to imperfect repair and correlated failure and repair. European Journal of
random maintenance. Applied Stochastic Models in Business and Industry, 16(3),
Operational Research, 174(3), 706–1722.
219–234.
Zhou, Y. F., Lin, T. R., Sun, Y., & Ma, L. (2016). Maintenance optimisation of a parallel-
Borgonovo, E., Marseguerra, M., & Zio, E. (2000). A monte carlo methodological
series system with stochastic and economic dependence under limited
approach to plant availability modeling with maintenance, aging and
maintenance capacity. Reliability Engineering & System Safety, 155, 137–146.
obsolescence. Reliability Engineering & System Safety, 67(1), 61–73.
Zhu, Q. (2015). Maintenance optimization for multi-component systems under
Carl, S., & Heikkil, S. (2011). Fixed point theory in ordered sets and applications. New
condition monitoring. PhD thesis, Eindhoven University of Technology.
York: Springer.