Computers & Industrial Engineering 122 (2018) 283–291

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Preventive maintenance scheduling for serial multi-station manufacturing T

systems with interaction between station reliability and product quality
⁎
Xiaojun Zhou , Biao Lu
Department of Industrial Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

A R T I C LE I N FO A B S T R A C T

Keywords: Station reliability and product quality usually interact with each other stochastically in the multi-station man-
Multi-station manufacturing system ufacturing systems, which brings much trouble for the preventive maintenance (PM) scheduling of the system. In
Quality and reliability this paper, a station reliability evaluation method is developed firstly based on the introduced bidirectional
Preventive maintenance interaction mechanism between station reliability and product quality, in which the focus is on the derivation of
Opportunistic maintenance
the failure rate of the stations. Then, a dynamic opportunistic PM (OM) policy is presented for the series multi-
Cost saving
station systems with such quality integrated station reliability. The OM model is built up based on an extended
cost saving method, and the optimal PM scheme of the system is obtained by maximizing the short-term cost
savings which not only come from the stations conducting PM but also come from the downstream stations of
those PM stations. Finally, a numerical example is given to illustrate how this PM scheduling approach works,
and a numerical comparison on different OM policies is also given to show the effectiveness of the proposed OM
policy under such interaction between station reliability and product quality.

1. Introduction
Preventive maintenance (PM) is an important approach to ensure
the operation of the manufacturing systems. In reality, most manu-
facturing systems are composed of multiple stations whose operating
conditions interact with each other. This interaction makes the PM
scheduling of these systems become very complex. However, with the
increasing demand from industry, more and more research efforts are
being paid on the PM modeling for these multi-station manufacturing
systems with interactions between stations (Nicolai & Dekker, 2008).
Commonly, from the perspective of maintenance, there are three
types of interactions between stations: economic dependence, structural
dependence and stochastic dependence. Economic dependence usually
indicates that the preparatory or set-up work (i.e. the preparation cost
associated with erecting a scaffolding) can be shared when several
stations are preventively maintained simultaneously (Do Van, Barros,
Bérenguer, Bouvardd, & Brissaude, 2013; Zhou, Zhang, Lin, & Ma,
2013; Zhang & Zeng, 2015; Dao, Zuo, & Pandy, 2014). Structural de-
pendence often happens when the manufacturing system is in series or
series-parallel. In such an instance, the PM of a station may cause the
stop of the whole system or the series subsystem. Thus, grouping PM
activities for other system or subsystem stations can decrease the
number of system PM stops and then save the set-up cost for the system
(i.e. the downtime cost due to production loss) (Zhou, Xi, & Lee, 2009;
Wong, Chan, & Chung, 2014; Zhou, Lu & Xi, 2012; Xia, Xi, Zhou, & Lee,
2012). Stochastic dependence usually means that the deterioration of a
station may affect the deterioration of other stations in the system. That
implies under stochastic dependence, the shape of the lifetime dis-
tribution of a station depends not only on the deterioration state of this
station itself but also on the deterioration states of other stations in the
system (Sheu, Liu, Zhang, & Chien, 2008; Shi & Zeng, 2016; Khorshidi,
Gunawan, & Ibrahim, 2016; Zhang, Zhang, & Du, 2018).
Because stochastic dependence brings uncertainty on the dete-
rioration of system stations, it makes the PM modeling of the system
much more complex. For this sake, compared with economic depen-
dence and structural dependence, the efforts on the PM modeling for
the systems with stochastic dependence are relatively scarce. In the area
of stochastic dependence, Murthy and Nguyen (1985) firstly introduce
two types of failure interactions between different components in a
system. Type 1 is the failure of one component can lead to the failure of
the other component. For example, the fracture of the belt in a car may
lead to the breakage of the engine. Type 2 is the failure of one com-
ponent can affect the failure rate of the other component. For example,
in a k-out-of-n system with shared workload, the failure of one com-
ponent increases the workload of other components and subsequently
speeds up the deterioration of these components (Zhang et al., 2018).
Based on this definition on failure interaction, Golmakani and Moakedi
(2012) propose a PM model to find the optimal periodic inspection

⁎
Corresponding author.
E-mail address: zzhou745@sjtu.edu.cn (X. Zhou).

https://doi.org/10.1016/j.cie.2018.06.009
Received 29 September 2017; Received in revised form 2 March 2018; Accepted 6 June 2018
Available online 14 June 2018
0360-8352/ © 2018 Elsevier Ltd. All rights reserved.
X. Zhou, B. Lu Computers & Industrial Engineering 122 (2018) 283–291

Nomenclature c jp , c jcmaintenance cost for a PM and for a minimal repair for

station j respectively
PM preventive maintenance c jq cost for the appearance of an out of tolerance product
OM opportunistic PM Tij the ith PM interval for station j
PQD product quality deviation tij start time of the ith PM cycle for station j
QVs quality variables τij operation duration for station j since the (i − 1)th PM of
NVs noise variables this station
Xj (t ) deterioration of the QV related component in station j t1∗, t2∗, t3∗, ⋯, ts∗ happening moments of the upstream stations of sta-
Z= [Z1, Z 2, …, ZL]′ vector of noise variables tion j within the ith PM cycle of station j
E (Z) , Cov (Z) expected value and covariance matrix of Z t1, t2, ⋯, tn predetermined PM moments for station 1, 2, ⋯, n
Dj (t ) PQD from station j CSk cost savings for PM alternative k
Cj , Aj , Bj , Ψj coefficients for PQD function CSk Part1 cost savings for PM alternative k coming from the OM
E (Dj (t )) , Var (Dj (t )) expected value and variance of Dj (t ) stations
Qj (t ) product quality from station j CSk Part2 cost savings for PM alternative k coming from the down-
λj (t ) hard failure rate of station j stream stations of the OM stations
λij (τ ) hard failure rate of station j within PM cycle i SMCjl maintenance cost saving for station j
μj (t ) soft failure rate of station j SDCjl downtime cost saving for station j
ξj coefficient on the deterioration rate of the QVs of station j SQCjl quality cost saving for station j
without upstream quality impact PCjl penalty cost for the OM of station j
Ωj coefficient vector describing the effect of upstream quality TC total maintenance cost throughout the PM scheduling
on the QVs of station j horizon of the system
c pd , c ud downtime cost for a PM and for a minimal repair respec-
tively

interval for a two-component repairable system, Zhang, Wu, Li and Lee
(2015) study an Opportunistic PM (OM) policy for a multi-component
system, and Lai and Yan (2016) present a discrete replacement policy
for a two-unit system. These research results are good references for the
PM modeling of the multi-station systems. However, the focus of the
above efforts is on the components in a reparable system and these
components are usually physically connected or workload shared. This
makes it possible that the failure rate of a component is directly affected
by the deterioration of another component. Obviously, this circum-
stance may not be applicable for a series multi-station manufacturing
system.
In fact, the stochastic dependence between stations in a series multi-
station system is usually indirectly and it often happens through the
product. The deterioration of the upstream station can cause the de-
viation of the product quality from this station. This deviation will be
transferred to the downstream station along with the product and then
affects the deterioration process of that station. For example, the wear
of the drilling tool may lead to a smaller diameter of the hole. This
increases the workload of the downstream reaming station and subse-
quently speeds up the wearing of the reaming tool. From this aspect, the
product plays important roles in connecting the manufacturing stations,
and it is the medium which brings stochastic dependence between
stations. For this sake, it is necessary to consider the product quality
simultaneously when processing the PM scheduling for the multi-sta-
tion systems with such stochastic dependence.
In recent years, a lot of research results have been published on the
PM scheduling for manufacturing systems by integrating product
quality (Fakher, Nourelfath, & Gendreau, 2015; Rivera-Gomez, Gharbi,
& Kenné, 2013; Nafissa, Hayet, & Okba, 2013; Yin, Zhang, Zhu, Deng, &
He, 2015). However, most of the efforts just simply combine the
maintenance and the product quality in the cost function, and they do
not interpret how the deterioration of the station and the deviation of
the product quality interact with each other. In this research area, Chen
and Jin (2005) gives a quality-reliability chain model for multi-station
manufacturing systems. In this model, the failure rate of a station is
affected by the deterioration of the incoming product quality, and the
deterioration of the upstream station causes the deterioration of the
downstream product quality. Based on this interaction model, Sun and
Xi (2011) proposes an optimal PM policy for a degrading manufacturing
system, where the impact of the tool-degraded state on the product
quality is deeply discussed. Lu and Zhou (2016) develops a main-
tenance-quality jointed PM model for a single-unit deterioration
system, with considering how the quality related components in the
system impacts on the product quality. Chen and Jin (2005), Sun and Xi
(2011), Lu and Zhou (2016) are the few research results focusing on the
interaction between station reliability and product quality, and these
works greatly promote the development of the maintenance-quality
jointed modeling technique. However, the PM models proposed by Sun
and Xi (2011), Lu and Zhou (2016) are for the single-unit manu-
facturing systems, and they only consider the impact of the station
deterioration on the product quality. As proposed by Chen and Jin
(2005), for a multi-station manufacturing system, the upstream product
quality can also affect the deterioration of the downstream station. It
would be much more challenging to present an effective PM scheduling
approach for the multi-station systems with bidirectional interaction
between station reliability and product quality.
In this paper, a dynamic OM policy is proposed for the series multi-
station systems with such bidirectional interaction between station

Station 1 Station 2 Station 3 Station n

...... ......
QV1 , X 1 (t ) QV2 , X 2 (t ) QV3 , X 3 (t ) QVn , Xn (t)

PQD1 , D1 (t ) PQD2 , D2 (t ) PQD3 , D3 (t ) PQDn 1, Dn 1(t) PQDn , Dn (t )

Product Product Product Product Product

Q1 (t ) Q2 (t ) Q3 (t ) Qn 1(t) Qn (t )

Fig. 1. Interactions between station reliability and product quality in a serial n-station system.

284
X. Zhou, B. Lu
reliability and product quality. A station reliability evaluation method
is developed based on the introduced bidirectional interaction me-
chanism, and an extended cost saving method is proposed for the
modeling of the quality integrated OM policy. The rest of the paper is
organized as follows. Section 2 is the problem definition. Section 3
gives the station reliability model under the bidirectional interaction
between station reliability and product quality. The cost saving based
dynamic OM policy is proposed and modeled in Section 4. Finally, a
numerical example is given in Section 5 to illustrate the availability of
the proposed PM scheduling approach.
2. Problem definition
Consider a series n-station system with stable and fast production
rhythm as shown in Fig. 1. In this system, the deterioration of some of
the station components can cause the major Product Quality Deviations
(PQDs) which are denoted as D (t ) = [D1 (t ), D2 (t ), …, Dj (t ), …, Dn (t )]′. For
example, the wear of the drilling tool, which is a component of the
drilling station, may lead to a smaller diameter of the hole, which is a
major PQD. Here, Dj (t ) represents the PQD from station j. These com-
ponent deteriorations are defined as Quality Variables (QVs) which are
denoted as X(t ) = [X1 (t ), X2 (t ), …, Xj (t ), …, Xn (t )]′ where Xj (t ) represents
the deterioration of the component in station j. Meanwhile, the product
quality from station j, denoted as Qj (t ) , which can be derived from PQD,
can also affect the states of the downstream stations and cause addi-
tional deteriorations due to unexpected frictions, vibrations, etc. For
example, a smaller diameter of the hole from the upstream station may
increase the workload of the downstream reaming station and then
affect the wearing speed of the reaming tool.
Based on the interactions between products and stations, two kinds
of station failures need considering for the evaluation of the station
reliability. One is the hard failure which means the breakdown of sta-
tion components with an inevitable stop of the station. According to the
literatures, minimal repair is usually adopted to correct this kind of
failures, and it can restore the failed station to the operation status
without changing its failure rate (Barlow & Hunter, 1960). In this effort,
minimal repair is also considered when the hard failure occurs. All the
station components, including those QV related components (i.e. cut-
ting tools), may trigger the hard failures. The other is the soft failure
which means the out of tolerance of the product quality. Commonly,
when the soft failure happens, the station can keep operating since this
kind of failures has little damage to the station itself. Therefore,
minimal repair is not considered when the soft failure occurs. The
station will keep operating until the PM opportunity arises. However,
the cost for the unqualified product lost still need to be considered.
PM is performed whenever one of the system stations reaches its
predetermined PM interval which is obtained mainly on the status of
the hard failure rate λj (t ) and the soft failure rate μj (t ) . In this study,
PM includes the maintenance of the equipment, the replacement of the
cutting tools, the adjustment of the fixture locators, etc. Thus, the
system reliability and the product quality can be improved together
through PM. For a series multi-station system, the PM of a station
usually results in the stop of the whole system. At that moment, PM
opportunities arise for other system stations since combing PM activ-
ities can decrease the number of system PM stops and then save the
total set-up cost (i.e. downtime cost) (Zhou et al., 2009). Assume that
the PM for a whole station is imperfect, which means the PM can de-
crease the failure rate of the station but not all the way to zero, while
the PM for the QV related components is perfect since these compo-
nents (i.e. cutting tools) are usually all replaced or adjusted during the
PM.
Based on the above assumption and definition, the PM modeling for
the series multi-station system can be divided into two parts. First is the
station reliability evaluation with considering the bidirectional inter-
action between the QVs and the product quality. Thus, the station re-
liability, which is evaluated by λj (t ) and μj (t ) , can be determined.
285
Computers & Industrial Engineering 122 (2018) 283–291
Second is the modeling of the proposed OM policy, which includes the
PM modeling for the station level and for the system level. The station
level model is used to determine the original PM interval to perform PM
for each station, and the optimal system PM scheme can be further
obtained based on the system level model.
3. Station reliability evaluation under bidirectional interaction
3.1. Effect from QVs to product quality
For a series multi-station system, product quality is generally af-
fected by multiple process variables. The predefined QVs are the main
variables. Examples of QVs include cutting tools in machining pro-
cesses, and dies in stamping processes, etc. The description of QVs can
be determined through engineering analysis or design of experiments
(Jin & Shi, 2000). Another type of variables is the Noise Variables
(NVs). For example, the human factors, the environmental variations,
etc. The NVs usually vary randomly and are not controllable. However,
in a stable manufacturing environment, it is rational to assume that the
NV follows a stationary distribution which can be obtained on statistical
method. Thus, according to the response model given by Chen and Jin
(2005), the PQD from station j can be defined as
Dj (t ) = Cj + Aj X(t ) + Bj Z+ X′(t ) Ψj Z j = 1…n (1)
where Z= [Z1, Z 2, …, ZL]′ is the column vector of NVs (L is the number of
NVs), with mean E (Z) = 0 and covariance matrix Cov (Z) independent
of time. Assume Cov (Zi (t ), Zj (t )) = 0, i ≠ j , which means the NVs are
mutually independent. X (t ) = [X1(t ), X2 (t ), ⋯, Xn (t )]′ includes all the
QVs in the system. Cj is a baseline constant, and Aj and Bj are the row
vectors defining the linear effect of X(t ) and Z respectively. Ψj is the
matrix defining the effect of interaction between X(t ) and Z. The model
parameters, including Cj , Aj , Bj and Ψj , can be either fitted based on the
result from design of experiments (Jin & Shi, 2000) or obtained through
engineering analysis based on process and product design information
(Chen, Jin, & Shi, 2004).
Product quality is usually defined as the proximity of the target
requirement. Based on the above PQD function, the evaluation index for
the product quality can be defined as
Qj (t ) = E (Dj (t )−E (Dj |X (t ) = 0))2
= (E (Dj (t )−E (Dj |X (t ) = 0)))2 + Var (Dj (t )) (2)
Based on Eq. (1), it can be deduced that
⎧ E (Dj |X (t ) = 0) = Cj
⎪
E (Dj (t )) = Cj + Aj X(t )
⎨
⎪Var (Dj (t )) = Bj Cov (Z ) B′j + (X ′ (t )Ψj) Cov (Z )Ψ′j X (t )
⎩ (3)
Substitute Eq. (3) to Eq. (2), the index for the product quality can be
rewritten as
Qj (t ) = X ′ (t ) A′j Aj X (t ) + Bj Cov (Z ) B′j + (X ′ (t )Ψj) Cov (Z )Ψ′j X (t ) (4)
3.2. Effect from product quality to QVs
Equivalently, product quality also has similar impact on QVs.
Generally, the state of the QVs in station j, denoted as Xj (t ) , changes
over time due to their self-deterioration even without external impact.
Its changing trend is usually assumed to be linear in the literature (Al-
fawzan, & Rahim, 2001; Chen et al., 2004). With the deterioration of
the product quality from the upstream station, a more intensive dete-
rioration usually occurs on Xj (t ) (Chen & Jin, 2005). For example, the
worse surface roughness of the products from the upstream station
usually leads to the increasing wearing speed of the cutting tool in the
downstream station. In such an instance, Xj (t ) can be divided into two
parts. One is the self-deterioration part which is a linear function of
time, and the other is the external impact part which is a linear function
X. Zhou, B. Lu
of incoming product quality. Thus, the state of QVs in station j within
the ith PM cycle of station j (tij < t ⩽ t(i + 1) j ) is written as
⎧ Xj (tij ) + ξ j (t −tij ) + Ωj Q (t ) tij < t ⩽ t1∗
⎪ X (t ∗) + ξ j (t −t1∗) + Ωj Q (t ) t1∗ < t ⩽ t2∗
⎪ j 1
Xj (t ) = Xj (t2∗) + ξ j (t −t2∗) + Ωj Q (t ) t2∗ < t ⩽ t3∗
⎨
⎪ ⋮ ⋮
⎪ X (t ∗) + ξ j (t −ts∗) + Ωj Q (t ) ts∗ < t ⩽ t(i + 1) j
⎩ j s (5)
where Xj (tij ) is the initial deterioration state of the QVs in station j at
the beginning of the ith PM cycle of station j and usually Xj (tij ) = 0
because of the perfect PM on QVs, ξ j is a coefficient which reflects the
deterioration rate of the QVs without considering the quality impact
from upstream stations, Ωj is a coefficient vector to describe the effect
of the product quality on the QVs of station j,
Q (t ) = [Q1 (t ), Q2 (t ), ⋯, Qn (t )]′ denotes the quality output from every
system stations.
In Eq. (5), Xj (tij ) + ξ j (t −tij ) represents the deterioration resulting
from the QVs themselves, while Ωj Q (t ) represents the deterioration
resulting from the external impact from upstream stations. Similarly,
the model parameters ξ j and Ωj can also be fitted based on the result
from design of experiments. Furthermore, s is the number of PMs which
are performed on the upstream stations of station j within the ith PM
cycle of station j, and t1∗, t2∗, t3∗, ⋯, ts∗ represent the happening moments
of these PMs. Since each PM on an upstream station can restore the
deterioration state of the QVs in that station perfectly, the value of the
PQD from that station after each PM, which can be obtained on Eq. (1),
will decrease a certain degree. This indicates the quality indexes
Q1 (t ), Q2 (t ), ⋯, Q(j − 1) (t ) are piecewise functions within the ith PM cycle
of station j, which leads to the piecewise curve characteristic for Xj (t ) as
shown in equation (3). Fig. 2 gives the cure shape of Xj (t ) in such an
instance. In Fig. 2, station a, b, c, ⋯, y (a, b, c, ⋯, y ∈ (1, 2, ⋯, j−1))
represents the upstream stations of station j which are renumbered
according to their PM moments occurring within the current PM cycle
of station j. Whenever one of the upstream stations conducts PM, the
product quality from that station will be restored. That leads to a
slowdown of the deterioration rate of Xj (t ) .
3.3. Evaluation of station reliability
Conventionally, reliability can be described as a function of failure
t decision cycle of the system.
rate, and it satisfies the form of exp (− ∫0 f (t ) dt ) where f (t ) is the
failure rate function. As described in Section 2, two kinds of failure rate
need considering, which are the hard failure rate λj (t ) and the soft
failure rate μj (t ) . Because the trigger mechanism for the hard failure
and for the soft failure are different, it is assumed that λj (t ) and μj (t ) are
independent. In such an instance, according to the definition of the
bidirectional interaction in the above section, the impacts from the
product quality is mainly on the soft failure rate μj (t ) . Thus, the focus of
the modeling of station reliability is on the obtaining of μj (t ) under the
given bidirectional interaction.
In reality, the deviation threshold of the product quality is usually
predetermined in order to judge the qualified product. Assume [Ljl , Luj ] is
the tolerable value interval for the deterioration of product quality.
Then, μj (t ) can be described as
μj (t ) = 1−Pr(Ljl ⩽ Dj (t ) ⩽ Luj )
(6)
where the calculation of the possibility Pr(Ljl ⩽ Dj (t ) ⩽ Luj ) depends on
the distribution of Dj (t ) . If the noise variables satisfy Zj (t ) ∼ N (0, σz2j )
and Cov (Zi (t ), Zj (t )) = 0, i ≠ j , it can be deduced that Dj (t ) follows a
normal distribution. Thus, the value of μj (t ) can be determined based
on the probability density function of the normal distribution, and Eq.
(6) can be rewritten as
286
Computers & Industrial Engineering 122 (2018) 283–291
⎛ ⎛ Luj −E (Dj (t )) ⎞ l
⎛ Lj −E (Dj (t )) ⎞ ⎞
μj (t ) = 1−⎜Φ −Φ
⎜ Var (D (t )) ⎟ ⎜ Var (D (t )) ⎟ ⎟
j j (7)
⎝ ⎝ ⎠ ⎝ ⎠⎠
Because of the piecewise characteristic of Xj (t ) , μj (t ) is also a pie-
cewise function which divided by the PM moments of the upstream
stations of station j denoted as t1∗, t2∗, t3∗, ⋯, ts∗ .
4. PM scheduling for multi-station manufacturing systems
4.1. Dynamic OM policy
As defined in Section 2, PM is performed whenever one of the
system stations reaches its predetermined PM interval. At that moment,
PM opportunities arise for other system stations. Due to the dependence
between system stations, any PM operation on the upstream stations
will have an impact on the downstream stations, especially on the soft
failure rate μj (t ) as defined in Eq. (7). That brings uncertainty to the
future PM activities of the system and makes the PM scheduling more
complicate and changeable. In order to reflect this dynamic impact, a
rolling horizon method based OM policy is considered. Under this
policy, the PM scheduling horizon of the whole system is divided into a
lot of OM decision cycles, and every station only experience 1 PM ac-
tivity within a system OM decision cycle.
Fig. 3 gives the decision framework of the dynamic OM policy.
Fig. 3 shows that the decision process can be briefly divided into four
parts: parameter values assignment, original PM intervals determina-
tion for each station, rolling horizon based OM decision making, and
obtaining of the optimal PM scheme of the system throughout the
scheduling horizon. Especially, because of the imperfect effect of PM,
the original PM intervals of each station must be re-determined after
each OM is performed.
Fig. 4 further gives the brief description of a system OM decision
cycle. For a series manufacturing system composed of n stations,
whenever one of the stations stops to conduct PM, an OM decision cycle
starts. In Fig. 4, station l is the first station who reaches its pre-
determined PM interval in current OM decision cycle. At that moment,
other n-1 stations have the opportunities to conduct PMs simulta-
neously to save the setup-up cost. Whether the PM of a station can be
advanced and performed together with station l depends on how much
cost can be saved if the PM of that station is advanced. t1, t2, ⋯, tn re-
present the predetermined PM moments for station 1, 2, ⋯, n . It can be
seen that only 1 PM activity is considered for each station in one OM
Furthermore, when station l stops to perform PM, every other
system station can be opportunistically maintained. That implies there
are Cn0− 1 + Cn1− 1 + …+Cnn−−11 alternatives to conduct the OM. In order to
obtain the optimal alternative, the cost savings for every OM alter-
natives need to be determined. The larger of the cost savings for an
alternative, the more of the opportunity for this alternative to be se-
lected. Once the optimal OM alternative is determined, PM will be
performed based on this alternative. After that PM, the system will keep
operating until the next OM decision cycle starts. This procedure will
j (t )
Station Station Station . . . Station Station . . . .. .
a b c y j
PM PM PM PM
. . . .. . . . . .. .
Perfect
PM
tij t1* t2* t3* . . . .. . ts* t(i 1) j t
Fig. 2. Curve shape of Xj (t ) .
X. Zhou, B. Lu Computers & Industrial Engineering 122 (2018) 283–291

Start not stop until the predetermined PM scheduling horizon is reached.

Based on the proposed dynamic OM policy, the PM modeling of the
Assign values to parameters, determine the scheduling horizon T SH multi-station manufacturing systems is defined as follows.

Determine the original PM intervals for each station (1) PM modeling for station level. This PM model is established to
determine the original PM intervals for each station. These PM in-
tervals are used to judge if the system need to be stopped to conduct
The system starts to run, set the start time as t=0
PM.
(2) PM modeling for system level. This model is based on an extended
N cost saving method. Once one of the system station reaches its
Any station reaches its predetermined PM interval?
scheduled PM interval, the PM activities of other system stations
Y can be advanced and performed together with this station to save
A system OM decision cycle begins the total maintenance cost. The possible cost savings can be ob-
Obtain the optimal PM group and conduct PM on this group tained on this model and then the optimal PM scheme for the whole
system can be determined.
Re-determine the original PM intervals for each station
4.2. PM modeling for station level
The system keeps running
Station level PM model is used to determine the original PM inter-
N
t T SH ? vals for each station. Due to the dynamic nature of OM, a single PM
Y cycle cost-effective model is considered for the station level. It is shown
The optimal PM scheme of the system within the scheduling horizon as
t + Tij t + Tj
c pd + c jp + (c jc + c ud ) ∫t ij λj (t ) dt + +c jq ∫t ij μj (t ) dt
End cij =
ij ij

Tij (8)
Fig. 3. Decision framework of the dynamic OM policy.
where c pd
is the downtime cost for a PM, and c ud
is the downtime cost
for a minimal repair. It is assumed that the downtime cost is equal for
advance PM or not? PM
every station in the system since a stop of one station usually means the
t1 stop of the whole system for a series multi-station manufacturing
Station 1 system. c jp and c jc are the cost for a PM and for a minimal repair for
PM station j respectively, c jq is the cost for the appearance of an out of
t2
Station 2 tolerance product, Tij is the ith PM interval for station j, and tij is the
PM start time of the ith PM cycle for station j. Based on Eq. (8), the optimal
…

If advanced,
tl PM interval of station j within each PM cycle of this station can be
Station l how much cost determined dynamically by minimizing cij .
PM can be saved?
…

As defined in Section 2, the PM for the whole station is usually

tj imperfect and it cannot restore the station to be as good as new. To
Station j describe the imperfect effect of PM, an evolution rule for the failure rate
PM
…

of station j, denoted as λj (t ) , is adopted as defined by Zhou et al. (2009).

tn
Station n Based on this rule, the relationship between the failure rates before and
after the ith PM can be denoted as
Fig. 4. A system OM decision cycle under the OM policy.
λ (i + 1) j (τ ) = bij λij (τ + aij Tij ) τ ∈ (0, T(i + 1) j] (9)

Failure rate PM PM
where λij (τ ) represents the failure rate for station j within its PM cycle i,
and it is a segment of the failure rate λj (t ) . λij (τ ) can also be denoted as
j (t ), j (t ) λj (t ) for t ∈ (tij, tij + Tij]. 0 < aij < 1 and bij > 1 are the age reduction
factor and the hazard rate increase factor for station j within PM cycle i
the ith PM respectively, which can be deduced on the historical maintenance data
the (i-1)th cycle of of the station.
PM cycle of station j For the QV related components, it is assumed that the PM is perfect.
That implies the soft failure rate μj (t ) decreases to zero after each PM.
station j Based on the above consideration, Fig. 5 gives the evolution char-
j (t ) acteristic of λj (t ) and μj (t ) before and after the (i − 1)th PM of station j.
j (t )
imperfect 4.3. PM modeling for system level

j (t ) perfect
(t )
Commonly, OM means the PM moment of the station is ahead of the
original PM schedule obtained on the PM model for station level, and it
j

t results in cost savings. Different from the cost savings given by Zhou
et al. (2009), here the cost savings include two parts. One comes from
the stations themselves that conduct OM, denoted as CSk Part1. Ex-
t(i 1) j tij t(i 1) j amples of CSk Part1 include the cost saving for minimal repair because
of the advancing of PM which leads to a decrease of the number of
Fig. 5. Evolution characteristic of λj (t ) and μj (t ) . failures, and the cost saving for the stop of the system since some sta-
tions are preventively maintained together and subsequently the

287
X. Zhou, B. Lu Computers & Industrial Engineering 122 (2018) 283–291

Table 1
Initial values of parameters.
j (αj, βj ) [Llj , Luj ] (mm) ξ j (mm/day) Ωj c pd ($) c ud ($) c jp ($) c jc ($) c qj ($)

1 (3, 45) [−0.10, +0.22] 0.003 [0, 0, 0] 200 650 540 610 410
2 (2.5, 35) [−0.05, +0.15] 0.0023 [0.4, 0, 0] 200 650 320 830 430
3 (2.5, 25) [0, +0.23] 0.0026 [0, 0.4, 0] 200 650 330 920 450

Table 2
Original PM intervals for each station of system.
j (aij , bij ) T1j (days) T2j (days) T3j (days) T4j (days) T5j (days) T6j (days)

1 (0.09, 1.08) 17 17 17 17 17 16
2 (0.09, 1.07) 13 13 13 13 13 12
3 (0.06, 1.05) 11 11 11 11 10 10

Table 3
Intermediate results of cost savings (unit for cost: $).
k j l CSk CSk Part1 SDCjl SMCjl SQCjl PCjl CSk Part2 SQC2 SQC3

1 2 3 460.7 287.4 218.5 22.6 176.6 130.4 173.3 – 173.3

2 1 3 −19.4 −155.1 225.4 23.6 85.7 489.7 135.7 36.1 99.6
3 2, 1 3 375.1 132.3 – – – – 242.8 36.1 206.8
4 – 3 0 – – – – – – – –

Table 4 number of system PM stops can be decreased, etc. The other comes from
Optimal PM schedule for system. the downstream stations, denoted as CSk Part2 . The PM for a station can
j t
bring the state of the QVs of that station to be as good as new. Ac-
cording to the definition on the relationship between QVs and PQD
11 17 22 33 44 50 54 presented in Section 3, the restore of the QVs of that station also means
a slower deterioration speed for the QVs of the downstream stations.
1 – PM – OM – PM –
2 OM – OM OM OM – OM
This leads to a decreased gradient of the soft failure rate functions of the
3 PM – PM PM PM – PM downstream stations, which also means the cost saving for the im-
proved product quality. Thus, the total cost savings for OM alternative k
can be defined as
CSk = CSk Part1 + CSk Part 2 (10)

Assume there are m stations (including station l) in the OM alter-

native k, and station l reaches its predetermined PM interval first and at
that time t = tl . Then, CSk Part1 for OM alternative k can be defined as
m−1
CSk Part1 = ∑ (SMCjl + SDCjl + SQCjl−PCjl )
j=1 (11)

where SMCjl is the maintenance cost saving for station j, SDCjl is the
downtime cost saving for station j, SQCjl is the quality cost saving for
station j, and PCjl is the penalty cost for the OM of station j.
For the maintenance cost saving, it results from the reduction of the
Fig. 6. Total maintenance cost under different Ω . number of minimal repairs for station j. Suppose before this OM, station
j has experienced i-1 PM cycles. Then, the maintenance cost saving
within the ith PM cycle can be shown as
t +T t +τ
SMCjl = c jc ∫t ij ij λj (t ) dt − c jc ∫t ij ij λj (t ) dt
 ij
  originalcost  ij
  new cost
t +T
= c jc ∫t ij+ τ ij λj (t ) dt
ij ij (12)

where τij ∈ (0, Tij ) is the operation duration for station j at t = tl since
the (i − 1)th PM of this station.
For the downtime cost saving, it results from the reduction of the
number of PM stops and hard failure stops due to the OM of station j. It
can be defined as
tij + Tij
SDCjl = cud ∫t +τij ij
λj (t ) dt + cpd
(13)
Fig. 7. Total maintenance cost under different c pd .
For the quality cost saving, it results from the reduction of the

288
X. Zhou, B. Lu
number of soft failures, and it can be described as
tij + Tj
SQCjl = c jq ∫t +τ
ij ij
μj (t ) dt
For the penalty cost, the advancement of the PM action always
implies an insufficient utilization of the PM interval, resulting in an
increase of the number of PMs throughout the PM scheduling horizon
and subsequently an increase of the total PM cost of the station. Based
on this consideration, the penalty cost for the advancement of the PM
for station j is defined as
Tij
PCjl = cij (Tij−τij )
τij
Tij
where
τij
can be seen as an amplification coefficient of the average
penalty cost cij (Tij−τij ) . The definition form of cij (Tij−τij ) is given by Zhou
et al. (2009). However, in this study, this kind of definition always leads
to the circumstance that all the system stations are preventively
T
maintained together. Therefore, the coefficient ij is added to reduce the
τij
occurrence possibility of such circumstance.
As defined in Eq. (10), CSk Part2 comes from the downstream sta-
tions. The restore of the QVs of station j can slow the deterioration
speed of the QVs of the downstream stations. That also means the im-
provement of the product quality from the downstream stations. As-
sume the stations in the n-station system are numbered along with the
production process, and the first station occurred in OM alternative k,
except the station who reaches the predetermined PM interval first, is
station ω . Then, CSk Part2 for alternative k can be defined as
n
CSk Part 2 = ∑ SQCr
r=n−ω
where SQCr is the quality cost saving coming from station r which could
be any of the downstream stations of station ω , and it can be shown as
t +T t +T
SQCr = crq ∫t cr+ τ ir μrold (t ) dt −crq ∫t cr+ τ ir μrnew (t ) dt
cr ir cr ir
t +T
= crq t cr+ τ ir
∫ (μrold (t )−μrnew (t )) dt
cr ir
where tcr is the start time of the current PM cycle of station r, and
τcr ∈ (0, Tir ) is the operation duration for station r at time t = tl since the
last PM of this station. μrold (t ) is the old soft failure rate function for
station r if OM alternative k is not applied, μrnew (t ) is the new soft failure
rate function for station r if OM alternative k is applied without con-
sidering the PM for station r. The reason for not considering the PM for
station r is to evaluate the impact comes from the PMs of the upstream
stations of station r, although station r may be included in OM alter-
native k.
5. Numerical analysis
5.1. Original PM interval for each station
Consider a 3-station series manufacturing system which consists of
drilling, reaming and tapping procedures. With the deterioration of
drilling tools, the diameter of the hole gets smaller which increases the
difficulty of reaming and then intensifies the deterioration of reaming
tools. Equivalently, the deterioration of reaming tools has the same
impact on the deterioration of tapping tools. The process variables for
the three stations are described as X(t ) = [X1 (t ) X2 (t ) X3 (t )]′ where
the initial deterioration Xj (tij ) = 0, j = 1, 2, 3,
Z (t ) = [Z1 (t ) Z2 (t ) Z3 (t )]′ where Zj (t ) ∼ N (0, 0.001), j = 1, 2, 3.
Because the focus of this paper is on the OM policy, the value of the
parameters for the bidirectional interaction model is assigned sub-
jectively based on the research results given by Chen and Jin (2005), Lu
and Zhou (2016), and the reality of the 3-station system. Thus, the
product quality from each station is assumed as
Computers & Industrial Engineering 122 (2018) 283–291
D1 (t ) = 0.04 + [−1.974 0 0]X(t ) + [0.935 0 0] Z (t )
⎡−0.113 0 0⎤
(14)
+ X′(t ) ⎢ 0 0 0 ⎥ Z (t )
⎣ 0 0 0⎦ (18)
D2 (t ) = 0.07 + [−0.383 −1.836 0]X(t ) + [0.181 0.996 0] Z (t )
⎡−0.022 0 0⎤
+ X′1(t ) ⎢ 0 0.107 0 ⎥ Z (t )
⎣ 0 0 0⎦
(19)
D3 (t ) = 0.12 + [−0.085−0.415−1.931]X(t ) + [0.040.2220.918] Z (t )
(15) ⎡−0.005 0 0 ⎤
+ X′1(t ) ⎢ 0 0.024 0 ⎥ Z (t )
⎣ 0 0 0.183⎦
(20)
Because these stations are all machine tools, it is rational to assume
that the failure rate of the station follows a Weibull distribution which
αj
αj − 1
can be described as hj (t ) = ⎛ t ⎞ . The scheduling horizon of the
βj
⎝ βj ⎠
system is T SH = 60 days. The initial values of the other relevant para-
meters, which are also assigned according to the efforts by Zhou et al.
(2009), Chen and Jin (2005), Lu and Zhou (2016), and the reality of the
system, are listed in Table 1.
Within each PM cycle of a station, Eq. (8) is used to obtain the
original PM interval for the current PM cycle of that station. During this
term, the impact from the product quality of the upstream station is
ignored since this impact is dynamic throughout the scheduling hor-
izon. Once the station is preventively maintained, the current PM cycle
ends and then a new PM cycle starts. Because the PM is imperfect as
(16) defined in Eq. (9), the PM interval sequence for every station exhibits a
decrease trend. Table 2 gives the original PM intervals obtained on Eq.
(8) with the predetermined imperfect effect of PM denoted as (aij , bij ) .
5.2. Optimal PM schedule for system
(17) Based on the original PM intervals for each station, the optimal PM
schedule for the whole system can be obtained on the cost saving
scheme described in section 4. Assume all the system stations start their
operation from the beginning of their first PM cycle as shown in
Table 2.
Table 3 gives a segment of the intermediate results. In this case,
when t = 11, station 3 reaches the predetermined PM interval first.
Then, the whole system stops. There are four OM alternatives at that
moment. (1) station 2 with station 3, (2) station 1 with station 3, (3)
both station 1 and 2 with station 3, and (4) no stations with station 3.
To measure the benefit of each OM alternative, the cost savings are
calculated based on Eqs. (10)(17). Then, a PM decision is made based
on the value of CSk . In this case, as shown in Table 3, the cost savings
for alternative (1) is the largest one and it is 460.7. That implies a PM
for station 2 is also performed at t = 11.
It need to be explained that once station 2 is opportunistically
old
maintained, the real PM interval of this station changes from T12 = 13
new
to T12 = 11. Based on the definition of Eq. (9), the hard failure rate of
station 2 within its second PM cycle becomes λ22 (τ ) = bij λ12 (τ + aij ∗11) .
Thus, all the later PM intervals for station 2 will change, which is dif-
ferent from the intervals given in Table 2. Therefore, the second optimal
PM interval for station 2 should be re-optimized based on Eq. (8).
Besides, because of the PM of station 2 and station 3, the soft failure
and rates of these two stations are restored to be as good as new. Then, the
deterioration trends of these soft failure rates within the next system
OM decision cycle can be decided based on equation (5) and (7).
The above decision procedure will not stop until the predetermined
scheduling horizon (T SH = 60 ) is terminated. Table 4 gives the optimal
PM scheme throughout the predetermined scheduling horizon under
the above parameter values. In Table 4, PM means this station reaches
289
X. Zhou, B. Lu
the predetermined PM interval first, and OM means the PM of this
station is advanced and opportunistically performed together with the
station who reaches its PM interval first.
5.3. Comparison and discussion
Consider two other conventional policies to schedule the PM for the
3-station manufacturing system. (1) Never conduct OM. Each station of
the system is preventively maintained according to its predetermined
PM interval. (2) Always conduct OM. All the system stations are pre-
ventively maintained together whenever one of the system stations
reaches its predetermined PM interval. A comparison is implemented
between the proposed OM policy and the above two conventional OM
polies to show the effectiveness of the proposed OM policy. Because the
interactions between stations are the main factors considered in the
proposed OM policy, the focus of the comparison is on how the change
of the interactions impacts on different OM policies. According to the
proposed OM model, the correlation coefficient Ωj reflects the sto-
chastic dependence between stations, while the downtime cost c pd re-
flects the structural dependence between stations. In such an instance,
Ωj and c pd are chosen to do the parameter analysis for the comparison.
Let Ω varies from 0 to 2, and c pd varies from 50 to 600. Here Ω = x
means Ω1 = [0, 0, 0], Ω2 = [x , 0, 0], Ω3 = [0, x , 0]. Figs. 6 and 7 gives
the comparison results on the total maintenance cost throughout the
scheduling horizon between the proposed OM policy (called PC scheme
in the figures) and the other two conventional policies. The total
maintenance cost throughout the scheduling horizon (T SH = 60 ) can be
obtained based on the following equation which is written as
3 Transactions, 36, 827–839.
60 60
TC = ∑
j=1
( (c + c
c
j
ud )
∫0 λj (t ) dt + c jq ∫0
μj (t ) dt + Sj c jp + Mc pd
60 60
where (c jc
+ c ud )
∫ λj (t ) dt + ∫ μj (t ) dt +
0
c jq 0
represents the sum
of the total minimal repair cost, the total quality lost, and the total PM
cost throughout the horizon for station j, Sj is the number of PMs for
station j and M is the number of PM stops for the system throughout the
horizon.
The results show that under different Ω and c pd , the proposed OM
policy always suggests the optimal maintenance schedule with the
lowest cost. With the increase of Ω or c pd , the correlation between
stations becomes stronger. When Ω ⩾ 1.2 or c pd ⩾ 400 , the proposed
OM policy provides the same maintenance cost with “Always OM
policy”. That implies with the increase of the dependencies between
stations, whether it is stochastic or structural, the stations have more
opportunities to be preventively maintained together.
6. Conclusion
Station reliability and product quality usually interact with each
other in most multi-station manufacturing systems. This paper gives an
extended cost saving based dynamic OM policy for the series multi-
station manufacturing systems with considering the bidirectional in-
teraction between station reliability and product quality. The numerical
example shows that the proposed OM model can effectively reflect this
bidirectional interaction. It also shows that the total maintenance cost
under the proposed OM policy is always lower than the cost under the
other two conventional PM policies. With the increase of the value of Ω
or c pd , the system stations have more opportunities to be preventively
maintained together.
In reality, to apply the proposed dynamic OM policy to the PM
scheduling of the series multi-station manufacturing systems, the most
important preparation is the collection of some necessary data. First is
the manufacturing data including the quality state of the product before
and after it is processed by each station, and the deterioration states of
quality variables (QVs) and the values of noise variables (NVs) in the
Computers & Industrial Engineering 122 (2018) 283–291
station when the product is processed. Thus, the parameters in the bi-
directional interaction model between station reliability and product
quality can be estimated. Second is history failure, repair and main-
tenance information. The history failure records can be used to estimate
the parameters of the Weibull failure rate distribution, and the cost
parameter values can be obtained based on the history repair and
maintenance information.
However, with the constraint of the existing theory, some of the
definition is simplified in this study. For example, the effect from QVs to
station is just assumed to be linear based on the few existed efforts. In
the future work, more efforts could be paid on further modeling of the
effect from QVs to station. In addition, more attention also could be
paid on the PM modeling for other kind of systems with interaction
between station reliability and product quality. For example, series-
parallel multi-station systems.
Acknowledgement
This work is supported by the National Natural Science Foundation
of China (Nos. 51575356 and 71472125).
References
Al-fawzan, M. A., & Rahim, M. A. (2001). Optimal control of a deteriorating process with
a quadratic loss function. Quality and Reliability Engineering International, 17,
459–466.
Barlow, R., & Hunter, L. (1960). Optimum preventive maintenance policies. Operations
Research, 8, 90–100.
Chen, Y., Jin, J., & Shi, J. (2004). Integration of dimensional quality and locator relia-
bility in design and evaluation of multi-station body-in-white assembly processes. IIE
)
Chen, Y., & Jin, J. (2005). Quality-reliability chain modeling for system-reliability ana-
lysis of complex manufacturing processes. IEEE Transactions on Reliability, 54,
475–488.
(21) Dao, C., Zuo, M., & Pandey, M. (2014). Selective maintenance for multi-state ser-
ies–parallel systems under economic dependence. Reliability Engineering and System
Sj c jp Safety, 121, 240–249.
Do Van, P., Barros, A., Bérenguer, C., Bouvardd, K., & Brissaude, F. (2013). Dynamic
grouping maintenance with time limited opportunities. Reliability Engineering and
System Safety, 120, 51–59.
Fakher, H. B., Nourelfath, M., & Gendreau, M. (2015). Joint production-maintenance
planning in an imperfect system with quality degradation. In International con-
ference onindustrial engineering and systems management (IESM) (pp. 910–919).
IEEE.
Golmakani, H. R., & Moakedi, H. (2012). Periodic inspection optimization model for a
two-component repairable system with failure interaction. Computers and Industrial
Engineering, 63, 540–545.
Jin, J., & Shi, J. (2000). Diagnostic feature extraction from stamping tonnage signals
based on design of experiments. Journal of Manufacturing Science and Engineering, 122,
360–369.
Khorshidi, H. A., Gunawan, I., & Ibrahim, M. Y. (2016). A value-driven approach for
optimizing reliability-redundancy allocation problem in multi-state weighted k-out-
of-n system. Journal of Manufacturing Systems, 40, 54–62.
Lai, M., & Yan, H. (2016). Optimal number of minimal repairs with cumulative repair cost
limit for a two-unit system with failure rate interactions. International Journal of
Systems Science, 47, 466–473.
Lu, B., & Zhou, X. (2016). Joint modeling of preventive maintenance and quality im-
provement for deteriorating single-machine manufacturing systems. Computers and
Industrial Engineering, 91, 188–196.
Murthy, D. N. P., & Nguyen, D. G. (1985). Study of two-component system with failure
interaction. Naval Research Logistics Quarterly, 32, 239–247.
Nafissa, R., Hayet, M. L., & Okba, K. (2013). Desgin of multi-agent system for quality
control and machine maintenance. In International conference on control, decision
and information technologies (CoDIT) (pp. 794–797). IEEE.
Nicolai, R. P., & Dekker, R. (2008). Optimal maintenance of multi-component systems: A
review. Complex system maintenance handbook (pp. 263–286). London: Springer.
Rivera-Gomez, H., Gharbi, A., & Kenné, J. P. (2013). Joint control of production, over-
haul, and preventive maintenance for a production system subject to quality and
reliability deteriorations. The International Journal of Advanced Manufacturing
Technology, 69, 2111–2130.
Sheu, S., Liu, T., Zhang, Z., & Chien, Y. (2008). Extended optimal replacement policy for a
two-unit system with failure rate interaction and external shocks. IEEE Transactions
on Reliability, 39, 71–79.
Shi, H., & Zeng, J. (2016). Real-time prediction of remaining useful life and preventive
opportunistic maintenance strategy for multi-component systems considering sto-
chastic dependence. Computers and Industrial Engineering, 93, 192–204.
Sun, J., & Xi, L. (2011). Optimal maintenance policy for degrading manufacturing systems
using long-term average cost criterion. Production Planning and Control, 22, 124–136.
Wong, C., Chan, F., & Chung, S. (2014). Decision-making on multi-mould maintenance in
290
X. Zhou, B. Lu
production scheduling. International Journal of Production Research, 52, 5640–5655.
Xia, T., Xi, L., Zhou, X., & Lee, J. (2012). Dynamic maintenance decision-making for
series–parallel manufacturing system based on MAM–MTW methodology. European
Journal of Operational Research, 221, 231–240.
Yin, H., Zhang, G., Zhu, H., Deng, Y., & He, F. (2015). An integrated model of statistical
process control and maintenance based on the delayed monitoring. Reliability
Engineering and System Safety, 133, 323–333.
Zhang, T., Zhang, Y., & Du, X. (2018). Reliability analysis for k-out-of-n systems with
shared load and dependent components. Structural and Multidisciplinary Optimization.
http://dx.doi.org/10.1007/s00158-017-1893-z.
Zhang, X., & Zeng, J. (2015). Deterioration state space partitioning method for oppor-
tunistic maintenance modelling of identical multi-unit systems. International Journal
of Production Research, 53, 2100–2118.
291
Computers & Industrial Engineering 122 (2018) 283–291
Zhang, Z., Wu, S., Li, B., & Lee, S. (2015). (n, N) type maintenance policy for multi-
component systems with failure interactions. International Journal of Systems Science,
46, 1051–1064.
Zhou, X., Xi, L., & Lee, J. (2009). Opportunistic preventive maintenance scheduling for a
multi-unit series system based on dynamic programming. International Journal of
Production Economics, 118, 361–366.
Zhou, X., Lu, Z., & Xi, L. (2012). Preventive maintenance optimization for a multi-com-
ponent system under changing job shop schedule. Reliability Engineering and System
Safety, 101, 14–20.
Zhou, Y., Zhang, Z., Lin, T. R., & Ma, L. (2013). Maintenance optimisation of a multi-state
series–parallel system considering economic dependence and state-dependent in-
spection intervals. Reliability Engineering and System Safety, 111, 248–259.