Computers & Industrial Engineering 98 (2016) 102–112

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Optimal maintenance policy for a system with preventive repair

and two types of failures
Zhibin Zheng, Wenhui Zhou, Yanfang Zheng, Yongzhong Wu ⇑
School of Business Administration, South China University of Technology, Guangzhou 510640, China

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

Article history: Maintenance policy is one of the most critical issues in operations management. In reality, the proper
Received 30 July 2015 functioning of productive systems is affected by so many complicated factors that even preventive repair
Received in revised form 31 March 2016 cannot eliminate the possibility of system failures of different types. A cost-effective maintenance strat-
Accepted 2 May 2016
egy is usually desired. In this paper, a geometric process maintenance model with preventive repair and
Available online 4 May 2016
two types of failures (repairable failure and unrepairable failure) is studied. A maintenance policy ðT; NÞ is
proposed, where preventive repair will be conducted when the successive operating time reaches T, and
Keywords:
the system will be replaced by a new one when an unrepairable failure or the Nth repairable failure
Maintenance policy
Geometric process
occurs. The optimal policy ðT  ; N Þ is obtained such that the average cost rate (i.e., the long-run average
Preventive repair cost per unit time) is minimized. The model is generalized to reflect three different types of maintenance
Failure repair systems. An algorithm is proposed to obtain the optimal policy ðT  ; N  Þ. Numerical experiments are con-
Average cost rate ducted to examine the impacts of system parameters on the optimal maintenance policy.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction
The proper maintenance of productive systems is vital to ensure
the normal operation of productive systems in various industries,
including manufacturing, healthcare and power industries. It
extends equipment life, improves equipment availability, retains
equipment in proper condition and thus contributes to the overall
performance of the company (Swanson, 2001; Waeyenbergh &
Pintelon, 2002). The growing importance of maintenance manage-
ment has generated increasing interests in academic research. In
the past decade, maintenance and replacement problems have
been studied extensively in the literature.
In the earliest studies of the maintenance problem, repair-
replacement models were commonly considered as perfect repair
models, which assumed that a failed system would be ‘‘as good
as new” after repair. In practice, most systems cannot be consid-
ered ‘‘as good as new” after repair. Barlow and Hunter (1960) intro-
duced a minimal repair model, in which minimal repair does not
alter the aging of the system. Brown and Proschan (1983) consid-
ered the imperfect repair model, which involves two types of
repair, i.e., perfect repair with probability p, and minimal repair
with probability ð1  pÞ. Following the introduction of this model,
relevant studies have been conducted by Park (1979), Phelps
⇑ Corresponding author.
E-mail address: okwyz@hotmail.com (Y. Wu).
(1981), Block, Borges, and Savits (1985), and Kijima (1989). In real-
ity, many systems are considered to be degenerative because of
aging effects and accumulated wearing. This means that the suc-
cessive operating time between failures monotonously decrease
with the times of repair. Given this fact, Yeh (1988) and Lin
(1988) presented the geometric process (GP) model to describe a
degenerative system. This model proposed a new replacement pol-
icy N  , in which the system would be replaced when failure occurs
for N  times. Thereafter, the GP model has been extensively applied
to the maintenance problem for its capability in describing real
maintenance data sets. Yeh (1992) and Yeh and Chan (1998)
applied a GP model to fit three real data sets by using nonparamet-
ric and parametric methods. Lam, Zhu, Chan, and Liu (2004) used
the GP model to analyze more real data sets. Through the analysis
of the data of aircraft, computer, car and so on, they found that on
average the GP model was the best model among four homoge-
neous and nonhomogeneous Poisson models for fitting these real
data from a series of events. Subsequently, Tang and Lam (2006),
Zhang and Wang (2007), Wang and Zhang (2009), and
Finkelstein (2010) conducted further studies to extend the geomet-
ric process model.
All aforementioned models assume that a system has unique
modes of failure. However, a system can have two or more failure
modes in many circumstances. Some system failures can be classi-
fied based on causes. For instance, an electronic system may fail
because of a short circuit or an open circuit. In a manual control

http://dx.doi.org/10.1016/j.cie.2016.05.007
0360-8352/Ó 2016 Elsevier Ltd. All rights reserved.
 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112
system, a failure may be caused by a manmade mistake or machine
trouble. Some system failures can be classified by severity. Lam,
Zhang, and Zheng (2002) adopted a monotone process model to
study a one-component degenerative system with k þ 1 states (k
failure states and one working state). Their research showed that
the studied system was equivalent to a geometric process repair
model with a unique failure mode based on replacement policy
in that both systems have the same long-run average cost per unit
time and the same optimal policy. Furthermore, Zhang and Wang
(2010) examined the same problem under optimal policy T. In
some cases, slight failures can be easily fixed by a repairman. In
other cases, serious failures are impossible to repair, or repairing
them is not valid because of the extremely high cost. Recently,
Wang and Zhang (2013) investigated the optimal replacement pol-
icy for a system with two types of failures, i.e., repairable and unre-
pairable failures, where unrepairable failure leads to an immediate
replacement of the system.
In many industries, the preventive repair or preventive main-
tenance (a scheduled and regular machine maintenance practice)
is commonly adopted. By detecting the potential problem of the
system, preventive repair could extend system lifetime, availabil-
ity and reliability. Morimura (1969) studied various types of pre-
ventive repair policies. Zhang (2002) investigated the optimal
replacement policy N  for a geometric process repair model with
‘‘as good as new” preventive repair. The author determined the
replacement policy N  while assuming that the cycle time of pre-
ventive repair T was fixed. Furthermore, Lam (2007) examined a
combination policy ðT; NÞ for a GP maintenance model with pre-
ventive repair. In this paper, the proposed model also determines
the optimal cycle for preventive repair instead of replacement
policy N. Recently, Wang and Zhang (2014) examined an optimal
bivariate policy ðT; NÞ for a GP maintenance model with inspec-
tions and preventive repair in which failures are detected only
by periodic inspections.
Lam (2007) proposed a practical combination policy ðT  ; N  Þ for
the GP maintenance model with preventive repair. However, this
model does not consider the case of unrepairable failure. In reality,
both repairable and unrepairable failures may occur in most sys-
tems. Hence, considering both repairable and unrepairable failure
types in the GP model system is reasonable and appropriate. A
repair-replacement policy that considers both failure types should
be investigated.
In this paper, the optimal repair-replacement policy for a degen-
erative system with preventive repair and both repairable and
unrepairable failures is studied. The GP model is used to describe
the degenerative characteristics of the system. In the proposed
maintenance policy, the system will be repaired when a repairable
failure occurs and be replaced with a new system when an unre-
pairable failure or the Nth repairable failure occurs. In addition, pre-
ventive repair is conducted when the successive operating time
reaches T. The optimal maintenance policy ðT  ; N  Þ is determined
to minimize the long-run average cost per unit time.
The remaining of the paper is organized as follows. Section 2
describes the problem and the proposed GP maintenance model
in details, followed by the analysis of the model in different cases
in Section 3. In Section 4, an algorithm for achieving the optimal
maintenance policy is proposed, with which some meaningful the-
orems are obtained and discussed. Numerical examples are con-
ducted and discussed in Section 5. Finally, the paper concludes in
Section 6.
2. The GP maintenance model
For the ease of exposition, we first summarize all the notations
and symbols adopted in Table 1.
103
Table 1
Notations and symbols.
X in Potential successive operating time of the system after ði  1Þth
preventive repair in the nth period
Y in The ith preventive repair time in the nth period
Zn Failure repair time in the nth period
Xs Total operating time in the sth replacement cycle
Ys Total preventive repair time in the sth replacement cycle
Zs Total failure repair time in the sth replacement cycle
Ws Replacement time in the sth replacement cycle
a The ratio parameter of geometric process (GP) {X 1n ; n ¼ 1; 2; . . .}
(a P 1)
b1 The ratio parameter of geometric process (GP) {Y 1n ; n ¼ 1; 2; . . .}
(0 < b1  1)
b2 The ratio parameter of geometric process (GP) {Z n ; n ¼ 1; 2; . . .}
(0 < b2  1)
k Expectation of X 11
b Expectation of Y 11
g Expectation of Z 1
x Expectation of W s
T Cycle time of preventive repair
N Number of repairable failures
p Probability of repairable failure
q Probability of unrepairable failure
pf Probability of the successive operating time shorter than T
qf Probability of the successive operating time longer than T
r System working reward per unit time
c0 The basic replacement cost of the system
c1 Cost rate (cost per unit time) of preventive repair
c2 Cost rate of failure repair
c3 Variable replacement cost per unit time
Mn Total number of preventive repairs in the nth period
Fs Total number of failures (including repairable and unrepairable)
until an unrepairable failure occurs in the sth replacement cycle
CðT; NÞ Long-run average cost per unit time given policy ðT; NÞ
2.1. Problem description
This paper considers a new productive system, which yields
working reward per unit time during the course of operation.
Due to wear and tear, environment influence and other factors,
system failures may occur, some of which are repairable while
the others are unrepairable. The probability of the repairable fail-
ure is p and the probability of the unrepairable failure is q ¼ 1  p.
To deal with these two types of failures and improve the relia-
bility of the system, three different types of maintenance are
adopted.
 Preventive repair. Preventive repair is a routine practice and
may involve inspections, measurements, adjustments, and
minor parts replacement in order to postpone failure from
occurring. The cost rate(cost per unit time) of preventive repair
is c1 .
 Failure repair. Failure repair is conducted when a repairable
failure occurs. Failure repair often involves changing of certain
damaged parts to restore functionality of the system. The cost
rate of failure repair is c2 .
 System replacement. When a failure occurs, the system may
also be replaced by a new and identical one. The replacement
cost consists of two parts, i.e., the fixed cost c0 to purchase a
new system and the variable cost which is proportional to the
replacement time at rate c3 .
A maintenance policy ðT; NÞ is proposed and studied, in which a
preventive repair is performed when the successive operating time
reaches T, and the system is replaced by a new one when an unre-
pairable failure or the Nth repairable failure occurs. It is noted that
if a failure occurs before a scheduled preventive repair, the preven-
tive repair will be postponed to the time when another cycle time T
104 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112

passes after the completion of the failure repair or the replacement  Let F s (s ¼ 1; 2; . . .) be the total number of failures (including
of the system. repairable and unrepairable failures) until an unrepairable fail-
For convenience of discussion, the following definitions are ure occurs in the sth replacement cycle.
given.
With the above definitions, the possible courses of the system
 The sth replacement cycle ðs ¼ 1; 2; 3 . . .Þ denotes the time are illustrated in Figs. 1 and 2.
interval between the completion of the ðs  1Þth replacement
and sth replacement.
2.2. Geometric process formulation
 Within a particular replacement cycle, there are different peri-
ods. The nth period of failure repair cycle ðn ¼ 1;2; 3. .. ;N  1Þ
In this paper, the geometric process formulation of a degenera-
denotes the time interval between the completion of the
tive system with preventive repair and two types failures has been
ðn  1Þth failure repair and the nth failure repair. In particular,
established. First, the definitions of a geometric process are given
if the first failure is an unrepairable failure, then the first period
below.
is the time interval between system installation and the com-
pletion of the first replacement. The Nth period is the time
Definition 1. Given two random variables X and Y; X is said to be
interval between the completion of the ðN  1Þth failure repair
stochastically greater than Y, or Y is stochastically less than X, if
and the following replacement in the replacement cycle.
 Let X in (n ¼ 1; 2; . . . ; i ¼ 1; 2; . . .) be the potential successive PðX > tÞ P PðY > tÞ for all real t:
operating time of the system after ði  1Þth preventive repair This is denoted by X P Y or YPst X (see, e.g., Ross, 1996). Further-
in the nth period, or equivalently, the time interval between more, a stochastic process fX n ; n ¼ 1; 2; . . .g is stochastically
the ði  1Þth preventive repair and the next failure. It should decreasing (increasing) if X n Pst ð6st ÞX nþ1 for all n ¼ 1; 2; . . .
be noted that if X in is larger than the cycle time of preventive
repair, T, the preventive repair takes place before any
Definition 2. A stochastic process fX n ; n ¼ 1; 2; . . .g is a geometric
potential system failure and the actual successive operating
time equals T. process if there exists a real a > 0 such that fan1 X n ; n ¼ 1; 2; . . .g
forms a renewal process. The real a is called the ratio of the geo-
 Similarly, let Y in ðn ¼ 1; 2; . . . ; i ¼ 1; 2; . . .Þ be the ith preventive
metric process (see, e.g., Yeh, 1988 & Lin, 1988 for more details).
repair time(time spent for the nth preventive repair of the sys-
Obviously, if a > 1, then fX n ; n ¼ 1; 2; . . .g is stochastically
tem) in the nth period. Furthermore, let Z n (n = 1, 2, . . .) be the
decreasing, i.e.
failure repair time (time spent for the failure repair) of the
system in the nth period. X n Pst X nþ1 ; n ¼ 1; 2; . . . :

Fig. 1. A possible diagram of a replacement cycle of the system when F s > N.

Fig. 2. An example of a replacement cycle of the system when F s 6 N.

 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112
If 0 < a < 1, then fX n ; n ¼ 1; 2; . . .g is stochastically increasing, i.e.
X n 6st X nþ1 ; n ¼ 1; 2; . . . :
If a = 1, then the geometric process becomes a renewal process.
With the definitions of geometric process, the following
assumptions can be made for the degenerative system.
Assumption 1. The successive operating times {X 1n ; n ¼ 1; 2; . . .},
the preventive repair times {Y 1n ; n ¼ 1; 2; . . .} and the failure repair
times {Z n ; n ¼ 1; 2; . . .} form geometric processes with ratio a; b1 ,
and b2 and expectation of EðX 11 Þ ¼ k; EðY 11 Þ ¼ b, and EðZ 1 Þ ¼ g,
respectively, whereby a P 1, 0 < b1 &b2  1. These variables
have cumulative distribution function F n ðxÞ ¼ Fðan1 xÞ,
n1 n1
Gn ðyÞ ¼ Gðb1 yÞ, and Hn ðzÞ ¼ Hðb2 zÞ and probability density
n1 n1
function f n ðxÞ ¼ an1 f ðan1 xÞ, g n ðyÞ ¼ b1 gðb1 yÞ, and
n1 n1
hn ðzÞ ¼ b2 hðb2 zÞ, respectively.
Assumption 2. X in (i ¼ 1; 2; . . .) and Y in ði ¼ 1; 2; . . .Þ are independent
and identically distributed (i.i.d.) random variables in the nth per-
iod of a replacement cycle.
Assumption 3. Replacement time W s in the sth replacement cycle
is a random variable with EðW s Þ ¼ x.
Assumption 4. {X in ; i ¼ 1; 2; . . . ; n ¼ 1; 2; . . .}, {Y in ; i ¼ 1; 2; . . . ;
n ¼ 1; 2; . . .} and {Z n ; n ¼ 1; 2; . . .} are all independent of W s .
Assumption 4 is realistic because every replacement cycle
forms a renewal process. The replacement time of the system
depends on external factors like the worker’s proficiency rather
than the state of the system (Tang & Lam, 2006).
Fig. 3 is an example showing how the characteristics of the
degenerative system can be described by the geometric process
formulation.
 At the beginning, a new system starts operation and the 1st per-
iod begins. The potential successive operating time, X 11 , follows
the distribution ðEðX 11 Þ ¼ k; VarðX 11 Þ ¼ r2 Þ. This means that the
next potential failure is expected to happen after X 11 , if no pre-
ventive repair is taken in between and interrupts it. Here, sup-
pose that the value of X 11 is larger than the cycle time of
preventive repair T. Then, a preventive repair is taken after time
T before any failure actually happens. (For convenience, the
repair times are not shown in the figure.) After the preventive
repair, the potential successive operating time, X 21 , also follows
the distribution ðEðX 21 Þ ¼ k; VarðX 21 Þ ¼ r2 Þ, and X 11 and X 21 are i.i.
d. In a sense, the preventive repair restores the system back
to the state at the beginning of the current period, i.e., State 1,
and the next failure is expected to be postponed. After that,
the 2nd preventive repair takes place as X 21 also turns out to
be larger than T, and restores the system back to State 1 again.
Then, suppose that X 31 is smaller than the preventive repair
cycle T, and a repairable failure occurs, which incurs the 1st fail-
ure repair, after which the 2nd period begins.
 After the completion of the 1st failure repair, the potential suc-
cessive operating time X 12 follows the distribution


EðX 12 Þ ¼ ak ; VarðX 12 Þ ¼ ra2 ða P 1Þ. As a P 1, X 12 stochastically
2
decreases comparing to those in the 1st period. As X 12 turns
out to be longer than T, the 1st preventive repair in the 2nd per-
iod takes place and restores the system to the system state at
the beginning of the 2nd period, i.e., State2. Following that,
105
the potential successive operating time X 22 follows the distribu-


tion EðX 22 Þ ¼ ak ; VarðX 22 Þ ¼ ra2 . Subsequently, the 2nd repairable
2
failure occurs as X 22 turns out to be smaller than T, which incurs
the 2nd failure repair and brings the system to the 2nd period.
 Likewise, the system degenerates as the number of failure
repairs increases. After ðn  1Þth failure repair, the potential
successive operating times in the nth period follow the distribu-


tion EðX in Þ ¼ an1
k r2
; VarðX in Þ ¼ a2ðn1Þ ða P 1Þ and X in ði ¼ 1; 2; 3 . . .Þ
are i.i.d.
 When the Nth repairable failure occurs, the system replace by a
new one.
In the geometric process formulation, the failure repair is not
perfect: it can fix the failure but not avoid the deterioration of
the system, in that the potential successive operating times are
stochastically decreasing with the times of failure repairs. In com-
parison, the preventive repair restores the system to the ‘‘better
than old” state (the system state when last failure repair com-
pletes) and the next failure expects to be postponed.
The maintenance policy ðT; NÞ is established to minimize the
average cost rate, i.e., the long-run average cost per unit time
CðT; NÞ. With the renewal theorem (see Ross, 1996), the long-run
average cost per unit time CðT; NÞ is given by the following:
The expected cost of a replacement cycle
CðT; NÞ ¼ : ð1Þ
The expected length of a replacement cycle
Let X s , Y s , and Z s be the total operating time, the total preven-
tive repair time and the total failure repair time in the sth replace-
ment cycle, respectively. The above objective function can be
expressed as follows:
c1 EY s þ c2 EZ s  rEX s þ c0 þ c3 EW s
CðT; NÞ ¼ : ð2Þ
EX s þ EY s þ EZ s þ EW s
Eq. (2) contains a negative term ‘‘rEX s ”, in which rEX s is
the expected revenue of the system when it operates normally
in the sth replacement cycle. This is because the maintenance
not only needs maintenance cost but also cause lost of
revenue.
3. Model analysis
With high modularization and integration, some systems only
encounter unrepairable failures (i.e., p ¼ 0), and the system never
encounters repairable failures or failure repairs. In such a case,
the policy ðT; NÞ is not applicable because the replacement policy
N of these systems is not well-defined. Furthermore, since both
the operating time and preventive repair time form renewal pro-
cess in different replacement cycles, the geometric process is no
longer suitable for these systems. As a result, the analysis of the
case of p ¼ 0 should be different from that of the other case of
p > 0. In the following, we present the analysis of the model under
two cases: the case with two failure types (p > 0) and the case with
only unrepairable failures (p ¼ 0).
3.1. Case I: The system with two failure types
We first discuss the case with two failure types, i.e., p > 0. Here,
we denote X n (n ¼ 1; 2; . . .) and Y n (n ¼ 1; 2; . . .) as the total operat-
ing time and the total preventive repair time of the system in the
nth period, respectively. For the two possible cases, the total
operating time of the sth replacement cycle is expressed by:

X1 þ X2 þ    þ X Fs ; F s 6 N
Xs ¼ ; ð3Þ
X1 þ X2 þ    þ XN ; Fs > N
106 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112

and the total preventive repair time in the sth replacement cycle is For convenience of analysis, we define the following new
expressed by: function:

Y 1 þ Y 2 þ    þ Y Fs ; F s 6 N Z
Ys ¼ : ð4Þ 1 T
Y 1 þ Y 2 þ    þ Y N; Fs > N CðT; nÞ ¼ EðX n Þ ¼ f1  Fðan1 tÞgdt: ð8Þ
Fðan1 TÞ 0
Similarly, the total failure repair time of the sth replacement
cycle is expressed by: The expected total operating time of a replacement cycle can
 then be calculated as follows:
Z 1 þ Z 2 þ    þ Z F s 1 ; F s 6 N
Zs ¼ : ð5Þ
Z 1 þ Z 2 þ    þ Z N1 ; Fs > N
X
N X
k X
N

We note that the F s forms a geometric distribution. We then have EðX s Þ ¼ PðF n ¼ kÞ EðX n Þ þ PðF n > NÞ EðX n Þ
k¼1 n¼1 n¼1
the following two probabilities:
( X
N X
k X
N
PðF s ¼ kÞ ¼ pk1 q ¼ ðpk1 qÞ CðT; nÞ þ pN CðT; nÞ
:
PðF s > NÞ ¼ pN k¼1 n¼1 n¼1

X
N X
k Z
Given that M n (n ¼ 1; 2; . . .) is the total number of the preventive 1 T
¼ ðpk1 qÞ f n1 f1  Fðan1 tÞgdtg
repairs in the nth period, and forms a geometric distribution, we Fða TÞ 0
k¼1 n¼1
have the following:
X
N Z T
PðMn ¼ iÞ ¼ qif pf ; 1
þ pN f f1  Fðan1 tÞgdtg: ð9Þ
n¼1
Fðan1 TÞ 0
where pf is the probability of the successive operating time shorter
than T and given by: Similarly, the expectation of the total preventive repair time in
pf ¼ PðX 1n 6 TÞ ¼ Fðan1 TÞ; the nth period is:
!
and qf is the probability of the successive operating time longer X
Mn
b b qf
EðY n Þ ¼ E Y in ¼ n1
EðM n Þ ¼ n1
: ð10Þ
than T and given by: i¼1 b1 b1 pf
qf ¼ PðX 1n > TÞ ¼ 1  pf :
Hence, the expected total preventive repair time of a replacement
Thus, cycle is:

qf ð1  Fðan1 TÞÞ X
N X
k X
N
EðM n Þ ¼ ¼ : ð6Þ EðY s Þ ¼ PðF n ¼ kÞ EðY n Þ þ PðF n > NÞ EðY n Þ
pf Fðan1 TÞ
k¼1 n¼1 n¼1
( ) ( )
The total operating time in the nth period can be expressed by XN Xk
b qf XN
b qf
the following: ¼ ðp qÞ
k1
n1 p
þp N
n1 p
: ð11Þ
k¼1 n¼1 b1 f n¼1 b1 f
X
Mn
Xn ¼ T þ X nMn þ1 ¼ M n T þ X nMn þ1 ; Similarly, the expected total failure repair time of a replacement
i¼1
cycle is:
According to conditional expectation properties, the expected
X
N X
k1 X
N1
total operating time of the system in the nth period can be com- EðZ s Þ ¼ PðF n ¼ kÞ EðZ n Þ þ PðF n > NÞ EðZ n Þ
puted as follows: k¼1 n¼1 n¼1

n þ1
! !
EðX n Þ ¼ EfEðX n jMn Þg ¼ EfEðMn T þ X M
n jM n Þg X
N X
k1
g X
N1
g
¼ ðpk1
qÞ þp N
¼ EfEðM n TjM n Þ þ EðX nMn þ1 jMn Þg k¼1 n¼1 b2
n1
n¼1
n1
b2
Z T
1 X
¼ EfM n T þ tdFðan1 tÞg N
gqpk1 ðbk1  1Þ gpN ðbN1  1Þ
Fðan1 TÞ 0 ¼ k2
2
þ N2
2
: ð12Þ
Z T k¼1 b2 ðb2  1Þ b2 ðb2  1Þ
1  Fðan1 TÞ 1
¼ Tþ tdFðan1 tÞ
n1
Fða TÞ n1
Fða TÞ 0 With substitution of Eqs. (9), (11), and (12) into Eq. (2), the
Z T long-run average cost per unit time of the system is given as fol-
1
¼ f1  Fðan1 tÞgdt: ð7Þ lows:
Fðan1 TÞ 0
in which

hP P
 PN
i hP P

N b qf b qf g
c1 k¼1 ðp
k1
qÞ kn¼1 bn1 pf
þ pN n¼1 bn1 pf þ c2 N
k¼1 ðp
k1
qÞ k1
n¼1 bn1
P
g i hP i
1 1 2
P P
Eðc1 Y s þ c2 Z s  rX s þ c0 þ c3 W s Þ þpN N1n¼1 bn1 r N
k¼1 ðp
k1
qÞ kn¼1 CðT; nÞ þ pN Nn¼1 CðT; nÞ þ c0 þ c3 x UðT; NÞ
CðT; NÞ ¼ ¼ hP 2
Pk PN i hP Pk
b qf  ¼ WðT; NÞ þ c1 ;
EðX s þ Y s þ Z s þ W s Þ N
ðp k1
qÞ C ðT; nÞ þ pN
CðT; nÞ þ N
ðpk1
qÞ
k¼1 n¼1 n¼1 k¼1 n¼1 bn1 pf
PN
b qf i hPN Pk1
g  PN1
g i
1

þp N
n¼1 bn1 pf
þ k¼1 ðpk1
qÞ n¼1 bn1 þ p N
n¼1 bn1 þ x
1 2 2

ð13Þ
 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112 107

XN X
k1
g Xk bqf
UðT; NÞ ¼ ðpk1 qÞ½ðc2  c1 Þ ð n1 Þ  ðr þ c1 Þ CðT; nÞ EðY s Þ ¼ EðY 1 Þ ¼ ; ð18Þ
pf
k¼1 n¼1 b2 n¼1

X
N1
g XN
EðZ s Þ ¼ 0: ð19Þ
þ pN ½ðc2  c1 Þ ð n1 Þ  ðr þ c1 Þ CðT; nÞ þ c0
n¼1 b2 n¼1 With substitution of Eqs. (17)–(19) into Eqs. (2), the long-run
þ ðc3  c1 Þx; ð14Þ average cost per unit time of the system is given by the following:
bqf RT
and c1 pf
 r FðTÞ
1
0
f1  FðtÞgdtg þ c0 þ c3 x UðTÞ
" # CðT; NÞ ¼ CðTÞ ¼ RT ¼ þ c1 ;
XN X k Xk
b qf X
k1
g 1 bq
f1  FðtÞgdtg þ p f þ x WðTÞ
WðT; NÞ ¼ ðpk1 qÞ CðT; nÞ þ ð n1 Þ þ ð n1 Þ FðTÞ 0 f

k¼1 n¼1 n¼1 b1

pf n¼1 b2 ð20Þ
!
XN X b qf
N X g
N1
þ p ½ CðT; nÞ þ
N
ð n1 Þ þ  þ x: ð15Þ in which
n1
n¼1 n¼1 b1
pf n¼1 b2 Z T
1
UðTÞ ¼ ðr þ c1 Þ f1  FðtÞgdt þ R þ ðc3  c1 Þx; ð21Þ
FðTÞ 0

and
Remark 1. According to the equations above, if the probability of
Z
the repairable failure p ¼ 1, then our model will be simplified into 1 T
bqf
Eq. (16). The GP model with preventive repair and only repairable
WðTÞ ¼ f1  FðtÞgdtg þ þ x: ð22Þ
FðTÞ 0 pf
failures is found to be the same as that in Eq. (4.1) of Lam (2007).
PN n o P PN1
g
 For both Case I and II, the optimal replacement policies can be
 r Nn¼1 CðT; nÞ þ c0 þ c3 x
b qf
c1 n¼1 bn1 pf
þ c2 n¼1 bn1
determined analytically and numerically, as discussed in the
CðT; NÞ ¼ PN n b qf o PN1
g 
1 2
PN following sections.
n¼1 CðT; nÞ þ n¼1 bn1 p þ n¼1 bn1 þ x
1 f 2

ð16Þ 4. Optimal maintenance policy

In this section, an algorithm is proposed to obtain optimal

3.2. Case II: The system with only unrepairable failures
maintenance policy for the GP maintenance model. In practice,
an interval time T min should be given for the repairman to prepare
In this case, the maintenance policy (T; N) is not applicable
for the next repair and recovery period for the repair tools between
because the replacement cycle N of these systems is not well-
the two preventive repairs. We assume that the preventive repair
defined. It should be noted that the case with only unrepairable
cycle has a lower bound T min . If the preventive repair cycle
failures is different from the standard age replacement models in
T ¼ 1, then our model becomes a GP model without preventive
the literature. Under standard age replacement model, the system
repair. Therefore, we also assume an upper bound T max for the
is always replaced at its age t or failure, whichever occurs first,
preventive repair cycle. The lower bound T min and upper bound
where t is a constant. In our model, we consider a preventive repair
T max can be chosen such that
policy T, where preventive repair may postpone the unrepairable
failure. PðT min 6 X 11 6 T max Þ P h ð23Þ
Fig. 4 shows a possible diagram of this special case.
with a large enough h ¼ 0:95 as an example.
In this case, the system will be replaced by a new and identical
Obviously, the problem of the optimal maintenance policy is
one whenever a failure occurs. Hence, no failure repair time occurs,
transformed into the following equation:
and every replacement cycle consists of only one period. We then
have the following: CðT  ; N Þ ¼ minCðT; NÞ
T;N
Z T
1
EðX s Þ ¼ EðX 1 Þ ¼ f1  FðtÞgdtg; ð17Þ s:t: T min 6 T 6 T max :
FðTÞ 0

Fig. 3. A possible process with the sequence of events on the timeline.

108 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112

Solving Case I is considered to be difficult. Therefore, the following repairable failure in our problem. It is reasonable that the possible
algorithm is proposed to obtain the optimal maintenance policy for occurrence of unrepairable failure will even reduce the optimal
Case I, preventive repair cycle because of the relatively high cost of
system replacement.
Algorithm 1.
Theorem 2. If the operating time of a new system has an exponential
Step1: Segmenting the interval (T min ; T max ) into k intervals distribution, i.e., X i1  ExpðkÞ, (i ¼ 1; 2; . . .), then the optimal mainte-
with a small enough equal step length h by nance policy ðT  ; N  Þ for case I is determined by the following:
ti ¼ T min þ hiði ¼ 0; 1; . . . ; kÞ.  
Step2: For each i ¼ 1; 2; . . . ; k, calculate the average cost CðT  ; N Þ ¼ min minCðT min ; NÞ; minCð1; NÞ :
N N
Cðt i ; nÞ and find a ni such that
Similarly, for case II,
Cðti ; ni Þ ¼ minCðti ; nÞ:
n CðT  Þ ¼ minfCðT min Þ; Cð1Þg:

Step3: For each t i , compare the value of Cðti ; ni Þ and determine

the following: Proof. For Case I, if the operating time of a new system has an
exponential distribution ExpðkÞ, then, Eqs. (14) and (15) recast as
Cðt ; n Þ ¼ min Cðti ; ni Þ: follows
" ! #
16i6k

X
N X
k1
g Xk
k
Then, the policy (t ; n ) is an approximate optimal UðT; NÞ ¼ ðpk1 qÞ ðc2  c1 Þ  ðr þ c1 Þ
n1 an1
maintenance policy for the system. k¼1 n¼1 b n¼1
" ! 2 #
X
N1
g XN
k
þ p ðc2  c1 Þ
N
n1
 ðr þ c 1 Þ
Furthermore, it has been found that the two theorems obtained n¼1 b2 n¼1
an1
by Lam (2007) apply to our model as well with some subsidiary þ c0 þ ðc3  c1 Þx;
conditions. These theorems can help obtain an optimal mainte-
nance policy ðT  ; N  Þ conveniently. The two theorems are given in and
the following. " ! !#
XN X k
k Xk
b qf X
k1
g
WðT; NÞ ¼ ðp qÞ
k1
þ þ
k¼1 n¼1
an1 n¼1 bn1 1
pf n¼1 b2
n1

Theorem 1. If the lifetime distribution F 2 ERBLE, for case I, " ! !#

UðT; NÞ P 0, then for all T P T min and N P 1, we have: XN
k XN
b qf X
N1
g
þ pN n1
þ n1 p
þ n1
þ x:
a n¼1 b1 n¼1 b2
CðT  ; N  Þ ¼ minCðT min ; NÞ: n¼1 f
N
It is noted that UðT; NÞ does not depend on T, then
Similarly, for case II, UðTÞ P 0, then for all T P T min and N P 1, we
have the following: UðT; NÞ ¼ UðT min ; NÞ ¼ UðT max ; NÞ:
 qf
CðT Þ ¼ CðT min Þ: Furthermore, from Lemma 1 of Lam (2007), pf
is nonincreasing
in T. Therefore, WðT; NÞ is nonincreasing in T.
Proof. In our model, we have the same assumption as Lam (2007) Obviously, WðT; NÞ > 0. Then, when UðT; NÞ < 0, we have
q
on the preventive repair. It results in that pf (see Eq. (6)) and CðT; nÞ UðT; NÞ UðT; NÞ UðT max ; NÞ
f
CðT; NÞ ¼ þ c1 P þ c1 ¼ þ c1
(see Eq. (8)) are the same as Eqs. (4.3) and (4.4) in Lam (2007)’s WðT; NÞ WðT max ; NÞ WðT max ; NÞ
work, respectively. Thus, it follows from Lemma 1 and 2 of Lam ¼ CðT max ; NÞ P minCðT max ; NÞ:
q N
(2007) that both pf and CðT; nÞ are nonincreasing in T.
f
When UðT; NÞ P 0, we have
For case I, recall Eqs. (14) and (15), it follows that UðT; NÞ is
nondecreasing in T, while WðT; NÞ is nonincreasing in T. Then, for
UðT; NÞ UðT; NÞ UðT min ; NÞ
CðT; NÞ ¼ þ c1 P þ c1 ¼ þ c1
UðT; NÞ P 0, we have WðT; NÞ WðT min ; NÞ WðT min ; NÞ
¼ CðT min ; NÞ P minCðT min ; NÞ:
UðT;NÞ UðT min ;NÞ UðT min ;NÞ N
CðT;NÞ ¼ þ c1 P þ c1 P þ c1 ¼ CðT min ;NÞ
WðT;NÞ WðT;NÞ WðT min ;NÞ For Case II, the proof is similar to above. Thus, this completes
ð24Þ the proof of Theorem 2. h
for any T P T min . Theorem 2 means that if the operating time of a new system has
RT an exponential distribution, then either using the most frequent
For Case II, EðX s Þ ¼ FðTÞ
1
0 f1  FðtÞgdtg is nonincreasing in T.
preventive repair or not using the preventive repair at all is most
Thus, UðTÞ is nondecreasing in T, while WðTÞ is nonincreasing in T. economical.
Then, the rest of the proof is similar to Eq. (24). This completes the Theorems 1 and 2 are the same as those in Lam (2007) although
proof of Theorem 1. h the latter does not consider repairable failure. In spite of the simi-
According to Lam (2007), the ERBLE lifetime distribution is larity of the theorems, different properties from those in Lam
denoted by the expected total operating time (2007) have been found through numerical analysis, which will
Rt
lðtÞ ¼ FðtÞ
1
0
f1  FðxÞgdx, which does not increase in t. be discussed in Section 5.
Theorem 1 shows that if the lifetime distribution F 2 ERBLE and
the long-run average cost per unit time CðT; NÞ is nonnegative, then 5. Numerical example
the minimum preventive repair cycle for the system is cost-
optimal. The difference between our problem and that of Lam Three numerical examples are conducted to illustrate the opti-
(2007) is the consideration both unrepairable failure and mal maintenance policies obtained in the previous section.
 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112
Example 1. Bus transit is one of the essential and convenient ways
for people to travel in a modern city. Hundreds or thousands of
buses serve passengers in the city every day. Whenever a bus
breaks down on the road, the bus company needs to deploy
another spare bus which leads to additional cost to the company
and a possible traffic congestion. Most breakdowns are due to bus
engine failure. Therefore, an appropriate engine maintenance
policy is vital for a bus company to improve reliability and
availability.
In this example, the established model is employed to analyze
the maintenance policy of bus engines. The parameter values in
our example have been selected so that they resemble those in real
life. The readers may refer to Leung and Lee (1998), Leung and Fong
(2000) and Lam et al. (2004) in which the methods for estimating
parameters of GP model are described in details.
The parameters of the system are given as follows: (1) the oper-
ating time of a new system forms Weibull distribution Wðj; cÞ
with j ¼ 2, c ¼ 1, (2) the GP ratio of the successive operating time,
preventive repair time and failure repair time are a ¼ 1:01,
b1 ¼ 0:98 and b2 ¼ 0:96 respectively, (3) the expectation value of
the operating time, preventive repair time and failure repair time
in the first period are b ¼ 0:05 Month, g ¼ 0:5 Month and
x ¼ 0:9 Month, respectively, (4) the purchasing price of a new
bus engine is c0 = 18,500 USD, the expectation of preventive repair
cost, failure repair cost, replacement cost and working reward are
c1 ¼ 3300 USD/Month, c2 ¼ 7231 USD/Month, c3 ¼ 4569 USD/
Month, and r ¼ 9179 USD/Month, respectively.
Function (23) is used to determine a lower bound T min and an
upper bound T max of preventive cycle T as follows:
PðT min 6 X 11 6 T max Þ P 0:95:
As the operating time of a new system forms a Weibull distribution
Wð2; 1Þ, we have the following:
2
PðX 11 6 TÞ ¼ 1  eT :
Reasonably setting T min ¼ 0:1 Month and T max ¼ 2 Months, we have
the following:
Pð0:1 6 X 11 6 2Þ ¼ 0:9717 P 0:95:
For Case I, let the probability of the repairable failure p ¼ 0:98 and
the probability of unrepairable failure q ¼ 0:02. Then, the method
established in Section 4 is adopted to determine the optimal main-
tenance policy as follows:
Step1: Segmenting the interval (T min ; T max ) = (0.1, 2) into 20
intervals with equal length ti ¼ 0:1i (i ¼ 1; 2; . . . ; 20).
Step2: For each i ¼ 1; 2; . . . ; 20, calculate the average cost Cðti ; nÞ
and choose a ni as follows:
Cðt i ; ni Þ ¼ minCðt i ; nÞ:
n N  and CðT  ; N  Þ are non-increasing in p. When the probability of
Fig. 4. An example of a replacement cycle of a replacement cycle of the system when p ¼ 0.
109
The numerical results of this step are shown in Table 1(a). The col-
umns of ni and Cðt i ; ni Þ are the values satisfying Eq. (25), so that
Cðt i ; ni Þ are the minimum long-run average costs per unit time cor-
responding to the fixed t i and varying ni .
Step3: Compare the values of Cðt i ; ni Þ to find an approximate
optimal policy Cðt  ; n Þ, so that:
Cðt  ; n Þ ¼ min Cðt i ; nÞ:
16i6k
Table 2(a) shows that the minimum value of Cðt i ; ni Þ is
5008:92 USD/Month. It should be noted that the negative value
of the cost means that there is profit gained by running of the sys-
tem. An approximate optimal maintenance policy is
ðt  ; n Þ ¼ ð0:2; 8Þ, i.e., the optimal preventive repair cycle is
T ¼ 0:2 Month = 6 Days and the system should be replaced when
an unrepairable failure or the 8th repairable failure occur.
It is also noted that an approximate optimal maintenance policy
lies in the interval ð0:1; 0:3Þ. Hence, we can segment the interval
ð0:1; 0:3Þ with an equal length of 0.01 and use steps 2 to 3 again
to obtain an accurate approximate optimal maintenance policy.
The numerical results are given in the Table 2(b). The minimum
value of Cðti ; ni Þ is 5019.00 USD/Month. Hence, an approximate
optimal maintenance policy is ðt  ; n Þ ¼ ð0:22; 9Þ, i.e., the optimal
preventive repair cycle is T ¼ 0:22 Month 7 Days and the system
should be replaced when the 9th repairable failure occurs.
The 3D graph of the Cðti ; ni Þ is shown in Fig. 5. We can also find
that the Cðt i ; ni Þ is jointly convex in ti and ni . This property guaran-
tees the existence of optimal maintenance policy.
For Case II with only unrepairable failures, only the cycle of the
preventive repair needs to be determined. The same parameters
above are used. It shows that the optimal policy exists with
ti ¼ 0:1 (as shown in Table 3) and the minimum average cost
Cðt  Þ ¼ 3329:62 USD/Month. Therefore, the optimal preventive
repair cycle is T ¼ 0:1 Month = 3 days.
Example 2. In this example, the same maintenance problem for
bus engine is used to examine the influence of parameters a, b1 , b2 ,
p, and c0 on the optimal maintenance policy ðT  ; N  Þ. Different sets
of values are assigned to the five parameters and the approximate
optimal maintenance policies are obtained accordingly. The com-
putational results are presented in Tables 4 and 5.
Sensitivity analysis shows the following: (1) T  and N  are non-
increasing in a, while CðT  ; N  Þ is non-decreasing in a. N  is non-
decreasing in b1 and b2 , while CðT  ; N  Þ is non-increasing in b1
and b2 . The above result is straightforward because by definition,
parameter a; b1 and b2 determines the degenerative degree of the
system and the system degenerates faster after a failure with
greater a or smaller b1 and b2 . (2) T  is non-decreasing in p, while
ð25Þ
110 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112

Table 2 Table 4
The values of Cðti ; ni Þ of case I. Sensitivity of optimal policy to a; b1 and b2 .

ti ni Cðt i ; ni Þ ti ni Cðti ; ni Þ Parameter T N CðT  ; N  Þ

(a) The values of Cðti ; ni Þ in interval ð0:1; 2:0Þ a = 1.01 0.22 9 5019.00
0.1 6 4238.97 1.1 13 2165.76 a = 1.06 0.18 6 4637.66
0.2 8 5008.92 1.2 13 1993.85 a = 1.11 0.17 4 4396.19
0.3 9 4870.54 1.3 13 1854.34 a = 1.16 0.16 4 4227.62
0.4 10 4497.51 1.4 14 1744.39 a = 1.21 0.15 3 4095.18
0.5 11 4067.16 1.5 14 1659.66 b1 = 0.96 0.22 9 5019.00
0.6 11 3649.12 1.6 14 1594.86 b1 = 0.91 0.20 6 4870.65
0.7 11 3263.40 1.7 14 1546.34 b1 = 0.86 0.19 5 4765.89
0.8 12 2922.29 1.8 14 1510.79 b1 = 0.81 0.19 5 4677.65
0.9 12 2627.08 1.9 14 1485.31 b1 = 0.76 0.18 4 4618.00
1.0 12 2374.77 2.0 15 1467.48 b2 = 0.98 0.22 9 5019.00
b2 = 0.93 0.22 6 4772.20
(b) The values of Cðti ; ni Þ in interval ð0:1; 0:3Þ b2 = 0.88 0.22 5 4566.87
0.11 7 4408.85 0.21 9 5017.19 b2 = 0.83 0.22 4 4386.82
0.12 7 4549.81 0.22 9 5019.00 b2 = 0.78 0.23 4 4215.02
0.13 7 4664.14 0.23 9 5014.77
0.14 7 4756.42 0.24 9 5005.47
0.15 8 4830.88 0.25 9 4991.59
0.16 8 4890.30 0.26 9 4973.59 Table 5
0.17 8 4936.17 0.27 9 4952.16 Sensitivity of optimal policy to p and c0 .
0.18 8 4970.03 0.28 9 4927.50
0.19 8 4993.83 0.29 9 4900.28 Parameter T N CðT  ; N  Þ Parameter T N CðT  ; N  Þ
0.20 8 5008.92 0.30 9 4870.54
p = 0.98 0.22 9 5019.00 c0 = 18,500 0.22 9 5019.00
p = 0.95 0.22 10 5029.99 c0 = 15,500 0.22 8 5083.96
p = 0.92 0.22 11 5039.54 c0 = 12,500 0.23 8 5152.51
p = 0.89 0.22 14 5047.09 c0 = 9500 0.23 7 5231.76
p = 0.86 0.23 20 5051.65 c0 = 6500 0.24 6 5320.04
Table 3
The values of Cðti ; ni Þ of case II.

ti Cðti Þ ti Cðt i Þ ti Cðt i Þ ti Cðti Þ Example 3. In this example, we consider the Magnetic Resonance
0.1 3329.62 0.6 2421.38 1.1 6139.28 1.6 7610.54 Imaging (MRI), which is a most important testing tool in hospitals.
0.2 2730.55 0.7 3415.18 1.2 6568.98 1.7 7747.65 The MRI machine is one of the most expensive medical equipment
0.3 1390.64 0.8 4271.55 1.3 6921.07 1.8 7851.37 and may cost about USD 1 million to USD 5 million. The radio
0.4 1.60 0.9 5002.59 1.4 7205.78 1.9 7928.33
frequency unit (RFU) plays significant role in MRI machine, and it is
0.5 1281.54 1.0 5621.04 1.5 7432.65 2.0 7984.28
also one of the parts which cause equipment failure in most cases.
Thus, an appropriate maintenance policy for RFU could improve
the reliability and availability of MRI machine.
the repairable failure is high, the preventive repair cycle should be
increased to reduce the cost of preventive repair. In addition, the The established model is used to obtain the optimal mainte-
replacement cycle should be increased to balance the increasing nance policy for RFU. The parameters of the system are as follows:
cost of failure. (3) Similarly, T  is non-decreasing in c0 , but N  (1) the operating time of the RFU forms exponential distribution
and CðT  ; N  Þ are non-increasing in c0 . As the purchase price of Exp(k) with parameter k ¼ 0:01, (2) the GP ratio of the successive
the system is relatively high, it makes more sense to increase the operating time, preventive repair time and failure repair time are
frequency of preventive repair and reduce the frequency of system a ¼ 1:02, b1 ¼ 0:97 and b2 ¼ 0:95, respectively, (3) the expectation
replacement. of the operating time, preventive repair time and failure repair

8000

6000

4000

2000
C (t , n i )
i

−2000

2.0
−4000
1.5
1.0
−6000 0.5 ti
0 5 10 15 20 0
25 30 35 40 45 50
ni

Fig. 5. The plot of Cðti ; ni Þ.

 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112 111

Table 6
The values of CðT min ; NÞ and CðT 1 ; NÞ.

N CðT min ; NÞ CðT 1 ; NÞ N CðT min ; NÞ CðT 1 ; NÞ N CðT min ; NÞ CðT 1 ; NÞ

1 158.21 218.18 11 524.35 664.75 21 503.17 648.94
2 375.79 476.51 12 523.75 665.19 22 500.03 645.87
3 446.62 561.82 13 522.61 664.94 23 496.80 642.64
4 480.24 603.12 14 521.04 664.14 24 493.49 639.27
5 498.88 626.67 15 519.13 662.88 25 490.11 635.77
6 510.01 641.28 16 516.94 661.24 26 486.68 632.14
7 516.82 650.73 17 514.52 659.28 27 483.19 628.41
8 520.93 656.94 18 511.90 657.03 28 479.66 624.56
9 523.23 660.96 19 509.12 654.54 29 476.08 620.62
10 524.25 663.43 20 506.20 651.84 30 472.47 616.59

−100 ticular, with some subsidiary conditions, the optimal maintenance

C(T∞, N) policy ðT  ; N  Þ of this case can be obtained by applying the two the-
C(T
min
, N) orems found. In general, we note that jointly convex property
−200 always exists in our three numerical examples and an optimal
maintenance policy can be found by the algorithm proposed in this
−300 study.
(2) When the probability of repairable failure is p ¼ 0 and the
C (T,N)

preventive repair cycle is T min 6 T 6 T 1 , the established model

−400
−500
−600
−700
0 20 40 60
N of unrepairable systems.
Fig. 6. The plot of CðT min ; NÞ and CðT 1 ; NÞ.
time in the first period are b ¼ 1 Day, g ¼ 8 Days and x ¼ 10 Days
respectively, (4) the purchasing price of new RFU is
c0 = 58,000 USD, the expectation of preventive repair cost, failure
repair cost, replacement cost and working reward per day are
c1 ¼ 180 USD/Day, c2 ¼ 450 USD/Day, c3 ¼ 300 USD/Day and
r ¼ 850 USD/Day, respectively, and (5) the probability of unre-
pairable failure is q ¼ 0:03.
Then, Theorem 1 can be employed to find the optimal policy.
The results are presented in Table 6. Hence, an optimal mainte-
nance policy ðT  ; N  Þ is determined by the following:
CðT  ; N Þ ¼ minfminCðT min ; NÞ; minCð1; NÞg
N N
¼ minfCð5; 11Þ; Cð1; 12Þg
¼Cð1; 12Þ ¼ 665:19 USD=Day
where the optimal preventive repair cycle T  ¼ 1 means preventive
repair is not needed at all for this system. The system should be
replaced when the 8th repairable failure occurs. The plot of
CðT min ; NÞ and CðT 1 ; NÞ are shown in Fig. 6.
6. Discussion
Based on the analytical and numerical results, the established
model can be applied to three different types of maintenance sys-
tems as follows: (1) When the probability of repairable failure is
0 < p < 1 and the preventive repair cycle is T min 6 T 6 T 1 , the
established model is appropriate for systems with preventive
repair and two types of failures. In this case, the analytical result
is a more general form of the model proposed by Lam (2007), in
which the probability of the repairable failure is equal to 1. In par-
simplifies to case II for unrepairable systems with only unre-
pairable failure. In this case, the replacement policy N is not appli-
cable and only the preventive repair cycle T needs to be
determined. This specific case was not discussed in Wang and
Zhang (2013)’s work as their model only considered the replace-
ment policy N. With high modularization and integration, most
equipment and components nowadays may only encounter unre-
pairable failures. Therefore, the established model covers the case
80 100
(3) When the probability of repairable failure is 0 < p < 1 and
the preventive repair cycle is T ¼ 1, the established model applies
to the systems with two types of failures and no preventive repair.
In this case, our analytical results are different from those found in
Wang and Zhang (2013)’s work. In Wang and Zhang (2013), only
the successive operating times were assumed to constitute a GP
model while the consecutive failure repair times were not. In our
model, we assume that a failure repair is not a perfect repair, thus
the successive operating times after a failure repair, the consecu-
tive preventive repair times after a failure repair, and the consecu-
tive failure repair times after a failure repair constitute GP models
with rates a, b1 , and b2 , respectively. This assumption is more rea-
sonable for degenerative systems.
Several practical implications can be drawn regarding the appli-
cation of the model. First, the model of case (1) applies to the main-
tenance policy of systems with two types of failures and preventive
repair. This type of system is usually expensive or critical equip-
ment. For example, aircraft engine, elevator, bus engine and so
on. Second, the model of case (2) applies to the unrepairable sys-
tems. For instance, CPU of sever machines, printed circuit board
of electrical equipment, tyre of mine truck and so on. Finally, the
maintenance problem of short life cycle components or inexpen-
sive equipment, such as seats of public transport, could be handled
by the applying the model of case (3).
7. Conclusions
In this study, a geometric process maintenance model with pre-
ventive repair and two types of failures(repairable failure and
unrepairable failure) has been developed. In the model, a preven-
tive repair will be conducted when the successive operating time
of the system reaches T, and the system will be replaced by a
new one when a repairable failure occurs for N times or an unre-
pairable failure occurs.
112 Z. Zheng et al. / Computers & Industrial Engineering 98 (2016) 102–112
The long-run average cost rate of the system is obtained analyt-
ically, on the basis of which the optimal maintenance policy
ðT  ; N  Þ is achieved to minimize the long-run average cost. With
different parameter settings, the model can be applied to three
types of actual maintenance systems, including systems with pre-
ventive repair and two types of repair, systems with preventive
repair and only unrepairable failure, and systems without preven-
tive repair. An algorithm and two theorems are employed to obtain
the optimal maintenance policy ðT  ; N  Þ.
Based on the numerical results, several managerial insights
regarding how to determine an appropriate maintenance policy
have been obtained: (1) As the system rapidly deteriorates with
the number of repair times, a larger T  and a smaller N  are desir-
able to lower the long-run average cost. (2) As the probability of
unrepairable failure q increases, decreasing T  and increasing N 
are recommended. (3) As the purchase price of the system
increases, decreasing T  and increasing N  are also recommended.
Several extensions to the modeling framework are considered
worthy of further investigation. First, the established model only
considers repairable and unrepairable failures. In fact, the failure
mode can be classified according to various criteria for different
systems. Different failure modes of varying degrees may have
impact on the maintenance policy. Thus, it would be useful to con-
sider different failure modes for different systems. Second, it is
assumed in this study that the probability of repairable failure p,
is constant. In some cases, p may depend on the operating time,
the frequency of preventive repair and the failure times of the sys-
tem. One can relax this assumption by making the probability p a
function of the preventive repair cycle and the failure times of the
system. Finally, extending the model to two-component series or
parallel systems is also an interesting and practical topic.
Acknowledgements
The authors thank the editor and three anonymous referees for
their constructive comments and guidance. This research is sup-
ported by the National Natural Science Foundation of China (NSFC:
71571070, 71571071, 71271089) and the Natural Science Founda-
tion of Guangdong Province (2015A030311032).
References
Barlow, R., & Hunter, L. (1960). Optimum preventive repair policies. Operations
Research, 8(1), 90–100.
Block, H. W., Borges, W. S., & Savits, T. H. (1985). Age-dependent minimal repair.
Journal of Applied Probability, 370–385.
Brown, M., & Proschan, F. (1983). Imperfect repair. Journal of Applied probability,
851–859.
Finkelstein, M. (2010). A note on converging geometric-type processes. Journal of
Applied Probability, 47(2).
Kijima, M. (1989). Some results for repairable systems with general repair. Journal of
Applied probability, 89–102.
Lam, Y. (2007). A geometric process maintenance model with preventive repair.
European Journal of Operational Research, 182(2), 806–819.
Lam, Y., Zhang, Y. L., & Zheng, Y. H. (2002). A geometric process equivalent model for
a multistate degenerative system. European Journal of Operational Research, 142
(1), 21–29.
Lam, Y., Zhu, L. x., Chan, J. S., & Liu, Q. (2004). Analysis of data from a series of events
by a geometric process model. Acta Mathematicae Applicatae Sinica, 20(2),
263–282.
Leung, F. K., & Fong, C. (2000). A repair-replacement study for gearboxes using
geometric processes. International Journal of Quality & Reliability Management,
17(3), 285–304.
Leung, F., & Lee, Y. (1998). Using geometric processes to study maintenance
problems for engines. International Journal of Industrial Engineering-Theory
Applications and Practice, 5(4), 316–323.
Lin, Y. L. Y. (1988). Geometric processes and replacement problem. Acta
Mathematicae Applicatae Sinica, 4, 366–377.
Morimura, H. (1969). On some preventive maintenance policies for IFR. University of
North Carolina, Department of Statistics.
Park, K. S. (1979). Optimal number of minimal repairs before replacement. IEEE
Transactions on Reliability, 28(2), 137–140.
Phelps, R. (1981). Replacement policies under minimal repair. Journal of the
Operational Research Society, 549–554.
Ross, S. M. (1996). Stochastic processes (Vol. 2)New York: John Wiley & Sons.
Swanson, L. (2001). Linking maintenance strategies to performance. International
Journal of Production Economics, 70(3), 237–244.
Tang, Y. Y., & Lam, Y. (2006). A d-shock maintenance model for a deteriorating
system. European Journal of Operational Research, 168(2), 541C556.
Waeyenbergh, G., & Pintelon, L. (2002). A framework for maintenance concept
development. International Journal of Production Economics, 77(3), 299–313.
Wang, G. J., & Zhang, Y. L. (2009). A geometric process repair model for a two-
component system with shock damage interaction. International Journal of
Systems Science, 40(11), 1207–1215.
Wang, G. J., & Zhang, Y. L. (2013). Optimal repair–replacement policies for a system
with two types of failures. European Journal of Operational Research, 226(3),
500–506.
Wang, G. J., & Zhang, Y. L. (2014). Geometric process model for a system with
inspections and preventive repair. Computers & Industrial Engineering, 75,
13–19.
Yeh, L. (1988). A note on the optimal replacement problem. Advances in Applied
Probability, 479–482.
Yeh, L. (1992). Nonparametric inference for geometric processes. Communications in
Statistics-Theory and Methods, 21(7), 2083–2105.
Yeh, L., & Chan, S. K. (1998). Statistical inference for geometric processes with
lognormal distribution. Computational Statistics & Data Analysis, 27(1), 99–112.
Zhang, Y. L. (2002). A geometric-process repair-model with good-as-new preventive
repair. IEEE Transactions on Reliability, 51(2), 223–228.
Zhang, Y. L., & Wang, G. J. (2007). A deteriorating cold standby repairable system
with priority in use. European Journal of Operational Research, 183(1), 278–295.
Zhang, Y. L., & Wang, G. J. (2010). An optimal replacement policy for a multistate
degenerative simple system. Applied Mathematical Modelling, 34(12),
4138–4150.