Computers & Industrial Engineering 62 (2012) 1011–1024

Contents lists available at SciVerse ScienceDirect

Computers & Industrial Engineering

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

Three m-failure group maintenance models for M/M/N unreliable queuing service
systems
Gia-Shie Liu ⇑
Department of Information Management, Lunghwa University of Science and Technology, No. 300, Sec. 1, Wanshou Rd., Guishan Shiang, Taoyuan County 33306, Taiwan

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

Article history: This paper considers group maintenance problems for an unreliable service system with N independent
Received 15 September 2009 operating servers and a Markovian queue. A specific class of group maintenance policies is developed
Received in revised form 6 November 2011 where the repair is started as soon as the number of failed servers reaches a predetermined threshold.
Accepted 19 December 2011
This is actually a Quasi Birth-and-Death Process with two dimensions, the level for the arrival/service
Available online 5 January 2012
process and the phase for the failure/repair process. Two models with positive repair time and another
with instantaneous repair are considered. The matrix geometric approach is applied to calculate the
Keywords:
steady state distribution and the expected average cost for all three models. For the theoretical analysis,
Group maintenance
m-Failure
this paper proves that there exists an optimal group maintenance parameter m, which can find the min-
Matrix-geometric imal average cost for all three models. Additionally, some mathematical properties and sensitivity anal-
Queuing system yses are numerically demonstrated based on various parameters. Finally, the comparisons of these three
Continuous time Quasi Birth-and-Death proposed models in many aspects are also discussed.
Processes Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction operating normally even some service units are malfunctioning.

It is very important to keep service systems operating normally
in this service oriented era. Falling this, a lot of cost would occur
due to the loss of customers and the delay of the production of
goods. For examples, most telephone companies or mobile phone
companies will lose their customers because of frequent communi-
cation traffic jam problems caused by failures of their service sys-
tems; the failure of the related power systems operated by the
power company will cause the production delay of some manufac-
turing companies; moreover, for some systems such as nuclear
power systems, military weapon systems, airplanes, and subma-
rines, it is exceptionally essential to keep away from breakdowns
during their operation since it can be falled in a very dangerous
and disastrous condition. To reach this goal, besides inherent reli-
ability, an appropriate maintenance policy can be applied to avoid
failures of such operating systems.
This paper will focus on maintenance problems of some specific
large service or operating systems such as telecommunication sys-
tems, internet service systems, electric power systems, nuclear
power systems, military submarines, aircraft carriers, space sta-
tions, satellite systems, and automated manufacturing systems.
Since the services or functions provided by these systems are so
important, most of them are multi-unit systems, which can apply
parallel or redundant system designs to keep the whole system
⇑ Tel.: +886 2 26788221 (H), +886 2 82093211x6328 (O); fax: +886 2 26788221.
E-mail addresses: liugtw@yahoo.com.tw, liug@mail.lhu.edu.tw
Of course, the multi-unit system can also provide more sufficient
services to meet all kinds of requests or missions given by custom-
ers. The simplest maintenance policy for the multi-unit system is
to treat all units independently and follow the same maintenance
policy separately for each unit in the system. Nevertheless, for
those specific systems focused by this paper, the replacement
activity is not possible or difficult to be implemented immediately
upon every single unit failure because the missions or operations
cannot be canceled immediately for safety or it will cause a lot of
setup cost, production delaying cost, customer holding cost or cus-
tomer loss cost to stop the service/production operations. In that
case, it is obvious that group maintenance policies can play a key
role to keep those multi-unit systems operating normally.
In recent years, a large amount of researches has been devoted
to finding optimal maintenance policies under various assump-
tions. The maintenance models for multi-unit systems are gener-
ally based on those for single-unit systems. For single-unit
systems, various maintenance policies, such as block replacement
policies (Beichelt, 1993; Chien & Chen, 2007; Kennee, Gharbi, &
Beit, 2007; Scarf, Dwight, & Al-Musrati, 2005; Sheu, 1997, 1998;
Sheu & Griffith, 2002), age-replacement policies (Beichelt, 1993;
Scarf et al., 2005; Sheu, 1998; Shen & Chien, 2004), periodic pre-
ventive maintenance policies (Jung & Park, 2003; Jung, Park, &
Park, 2010; Sheu, Lin, & Liao, 2006; Vaughan, 2005; Yeh & Lo,
2001), failure limit maintenance policies (Arunraj & Maiti, 2010;
Carazas & Souza, 2010; Chan & Asgarpoor, 2006; Das, Lashkari, &
Sengupta, 2007; Grall, Berenguer, & Dieulle, 2002; Lapa, Pereira,

0360-8352/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cie.2011.12.028
1012 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024

Nomenclature

c number of repairmen k vector of arrival rate in each state of number of operat-

e [1,1, . . . , 1]0 ing servers
f failure rate for each server D(k) diag (k0, k1, . . . , kN)
m the threshold of number of failed servers to initiate l service rate of one server for each customer
repair process lw service rate when w servers are operational
m the optimal m l vector of service rate in each state of number of operat-
N total number of servers in the system ing servers
n the repaired number of failed servers Dðl Þ diag (0, l, 2l, . . . , Nl)
Q infinitesimal generator for failure/repair process Dðk þ l
 Þ diag (k0, k1 + l, k2 + 2l, . . . , kN + Nl)
Q infinitesimal generator for M/M/N continuous time Mar- pw the stationary probability of w servers are operational
kov Process p [p0, p1, . . . ,pN]
P the related transition probability matrix derived from Q S fixed repair cost
r repair rate with one repairman fixing one failed server r_cost variable repair cost per failed server
rw repair rate when w servers are operational h holding cost per customer per unit time
T the threshold of scheduled time to initiate repair A0 DðkÞ
process A2 DðlÞ
s another threshold of scheduled time to initiate repair A2 Q  DðkÞ  Dðl Þ
process T00 Q  DðkÞ
T the optimal T T01 DðkÞ
w number of operational servers in the system Tx2 DðkÞ for x ¼ 1; . . . ; N  1
x number of customers in the system Tx0 diag{lmin(x, w), 0 6 w 6 N}, for 1 6 x 6 N  1,
yxw probability of x customers in system and w servers are Tx1 Q  DðkÞ  T x0 for 1 6 x 6 N  1
operational tij the expected transition time from state i to j
yx [yx0, yx1, . . . , yxN] Mw the expected umber of customer in system when w

y [y0, y1, y2, . . .] servers working
k customer arrival rate M [M0, M1, . . . , MN]
kw customer arrival rate when w servers are operational

& de Barros, 2006; Love & Guo, 1996; Monga, Zuo, & Toogood,
1997; Pham & Wang, 1996; Saassouh, Dieulle, & Grall, 2007; Wu
& Clements-Croome, 2005), are proposed combined with minimal
repairs, unscheduled replacements and other options based on dif-
ferent situations. It can be noted that most maintenance models for
single-unit systems mentioned above usually assume that all fail-
ures are instantly detected and repaired (Arunraj & Maiti, 2010;
Beichelt, 1993; Carazas & Souza, 2010; Kennee et al., 2007; Scarf
et al., 2005; Sheu, 1997, 1998; Sheu & Chien, 2004; Sheu & Griffith,
2002). In the real world, this is not always true; what usually hap-
pened is that the single-unit system must stop its service and
maybe lose the customers waiting in the line. Therefore, this paper
will consider the random repair time and multi-unit service
systems.
For multi-unit systems, a lot of review papers have been con-
ducted recently (Cho & Parlar, 1991; Dekker, Wildeman, Schouten,
& Frank, 1997; Wang, 2002). According to the observation of Wang
(2002), maintenance policies for multi-unit systems can be imple-
mented by applying single-unit system maintenance policies for
each separate unit respectively if there exists economic indepen-
dence, failure independence, and structure independence between
units within those systems (Aghezzaf & Najid, 2008). The multi-
unit systems discussed in this paper sre apparently economic
dependent. It means that carrying out maintenance operations
on several units simultaneously costs less money or time than on
each unit individually. There are various types of maintenance pol-
icies for multi-unit systems, such as block replacement policies,
opportunistic maintenance policies, or group maintenance policies.
For classical block replacement policies, all units in the system are
replaced simultaneously at periodic intervals, and each unit failed
in between will also be replaced immediately (Barlow & Proschan,
1965; Sun, Xi, Du, & Pan, 2010). Since the classical block mainte-
nance policies are possible to force the system to replace nearly
new unfailed units, many studies have developed various modified
block maintenance policies to evade this kind of unnecessary
waste (Anisimov, 2005; Archibald & Dekker, 1996; Scarf & Cavalc-
ante, 2010; Sun et al., 2010). For opportunistic maintenance poli-
cies, upon a unit failure, besides replacing the failed one, other

Fig. 1. Transition flow for group maintenance process.

 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024
unfailed units are also possible to be replaced depending on some
predetermined control threshold strategies such as the control lim-
it for age (Berg, 1976, 1978; Pham & Wang, 2000), the control limit
for failure rate (Zheng & Fard, 1991). For group maintenance poli-
cies, failed units in the system will be replaced together while
reaching some predetermined thresholds of criteria including time,
cost, number of failed units etc. Since this paper focuses on some
specific systems in which immediate replacement action is difficult
to be carried out while every single unit failure is occured, group
maintenance policies are more appropriate to be applied in our
study.
According to our observation, there are three main types of
group maintenance policies which have been studied in the litera-
ture. The first type is T-age group maintenance policy. The general
idea in this policy is that no failed server is repaired until a sched-
uled time T, and then all failed servers in the system are fixed
simultaneously. Barlow and Proschan (1965) show that the opti-
mal scheduled time for preventive maintenance is nonrandom
and there exists a unique optimal policy if the distribution of time
to failure has an increasing failure rate. Okumoto and Eslayed
(1983) show that, in the case of the exponential distributions, a
closed form expression for T is developed; for general underlying
failure distributions, bounds for T are derived. Dekker and Roelv-
ink (1995) determine the optimum time for preventively replacing
a fixed group of components by introducing marginal cost criteria;
the marginal cost is interpreted as the extra cost caused by defer-
ring preventive replacement for an additional time unit and both of
block and group age replacement are developed and compared. Li
and Xu (2004) develop an age group replacement model to choose
the set of components randomly for replacement and study the ef-
fect of dependency in this selection; it is showed that preventive
maintenance only makes sense for components whose lifetimes
are new-better-than-used. Sun et al. (2010) simulate and analyze
the effect of three different multi-component maintenance policies
on the system performance, including age replacement, block
replacement, and block replacement with minimal repair. The sec-
ond type of group maintenance policy is m-failure. The general
idea for this policy is that no repairs are performed until m servers
have failed, and then all failed servers in the system are fixed
simultaneously. Assaf and Shanthikumar (1987) consider two
models with exponential failure times and a more general mainte-
nance policy, under which when the number of failed servers
reaches m, n servers are repaired. They show that the optimal re-
pair policy is either not to repair any failed server or to repair all
failed servers (n = m), and this is effectively a m-failure group
maintenance policy. By extending the repair and replacement
model developed by Assaf and Shanthikumar (1987), optimal m-
failure policies with nonnegative random repair time are discussed
in Wilson and Benmerzouga (1990). Pham and Wang (2000) fur-
ther extend two (s, T) opportunistic maintenance policies to the
group maintenance policy, which triggers the corrective replace-
ment depending on the number of failed component. Liu (2004)
develops a specific class of m-failure group replacement policy
for an M/M/N production/service system. A matrix-geometric
method is applied to perform the steady state analysis and to ob-
tain the expected average number of customers and the expected
average cost. Mathematical analysis and numerical examples are
given to demonstrate the properties of the optimal policy for var-
ious sets of parameter values. The last type is (m, T) group mainte-
nance policy. The general idea is that no failed machine is repaired
until a scheduled time T or upon m servers have failed, whichever
comes first, and then all failed machines in the system are fixed
simultaneously. Nakagawa (1983) considers the optimal number
m to minimize the mean cost rate when the scheduled replace-
ment time T is specified. Ritchken and Wilson (1990) consider a
generalization of the combined (m, T) group replacement model
1013
which requires inspection at either the scheduled time T or the
time when exactly m machines have failed, whichever comes first.
At an inspection, all failed machines are replaced with new units
while operating machines are serviced so that they become as good
as new. Sheu and Jhang (1996) propose a two-phase maintenance
policy for a group of identical repairable units to select two time
interval and m to minimize the expected cost per unit time for
an infinite time span. Besides the above three typical types of
group maintenance policies Popova (2004), develops the optimal
structure of Bayesian group replacement policies for a parallel sys-
tem of n units with exponential failure times and random failure
parameters. This paper has showed that it is optimal to inspect
the system merely at failure times and obtained the exact form
of the optimal Bayesian group replacement for a two-unit system.
Assuming that machine failure times conform to a Weibull distri-
bution Das et al. (2007), optimizes maintenance costs by imple-
menting a group policy subject to desirable machine reliability
threshold.
Most group maintenance models mentioned above assume that
the failed machines take negligible repair time and the operating
machines produce output at a constant rate (Barlow & Proschan,
1965; Assaf & Shanthikumar, 1987; Das et al., 2007; Dekker & Roe-
lvink, 1995; Li & Xu, 2004; Nakagawa, 1983; Okumoto & Eslayed,
1983; Pham & Wang, 2000; Popova, 2004; Ritchken & Wilson,
1990; Sheu & Jhang, 1996; Sun et al., 2010). For the repair time
of failed units, in fact, it can be quite different depending on differ-
ent types of multi-unit systems. In the real world, we believe that it
is impossible for any physical system to take zero repair time to re-
place its failed units. But, some small or simple multi-unit systems,
such as internet service system or telecommunication service sys-
tem, may be assumed to take negligible repair time for replacing
failed servers or failed switchboards due to their modularizing de-
signs. Moreover, it is more reasonable to assume the positive repair
time for most repair processes of systems such as automated man-
ufacturing system, electric power system, nuclear power system,
and submarine’s navigation system. Therefore, besides one model
for instantaneous repair, assuming the positive repair time for
other two models in this paper adds options to be more credible
and better reflecting the real repair process. For these two models
with positive repair time, one model allows other unit failures dur-
ing maintenance process, and the other model assumes that other
unit failures will not occur because all normal operating units are
also checked to make sure they will stay in normal condition dur-
ing maintenance process. For presuming constant production rate,
it means that when a failed machine is left un-repaired, there is a
production loss cost incurred at a constant rate. Although the
assumption of continuous inflow is appropriate for high-volume
automatic production processes, in the real world, it may be more
realistic to assume that jobs arrive according to a random arrival
stream. In this paper, the production loss cost rate is not constant,
but instead it depends on the number of jobs waiting for service at
any given time. It can be obtained more precise production loss
cost than the traditional group maintenance models in which con-
stant production job arrivals are assumed.
Therefore, the group maintenance problems considered in this
paper fall in the classification of queuing systems with unreliable
servers, in which a matrix-geometric method can be employed
for steady state analysis of a certain class of continuous time Mar-
kov Processes. The earliest results on matrix-geometric solutions
are contained in the paper of Evans (1967) and the Ph.D. thesis
of Wallace (1969) for block-Jacobi generators of continuous-
parameter Markov Processes of the GI/M/1 type, called Quasi
Birth-and-Death (QBD) Processes. The general theory, largely
developed by Neuts (1981), was motivated by the problem of ana-
lyzing the steady state behavior of non-Markovian (M/G/1 and
GI/M/1) queues operating in a random environment that can be
1014 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024
in one of N appropriately defined states. Besides, Latouche and
Ramaswami (1999) further introduce the more up-to-date devel-
opments of these celebrated matrix-geometric methods and call
them matrix analytic methods subsequently. They mainly present
the basic mathematical concepts and the algorithms of these ma-
trix analytic theories. In fact, these matrix analytic methods can
be applied in several areas, In recent years, many matrix analytic
methods are continuously developed or improved to solve the
QDB problems for different queuing systems, such as M/M/1 queue
in a random environment (Neuts, 1978a, 1978b), Phase-Type
queues (Latouche & Ramaswami, 1999; Neuts, 1981), Tandem
queues with blocking (Latouche & Neuts, 1980), a Markovian
queue with server breakdowns and repairs (Neuts & Lucantoni,
1979), Markovian queues with marked transition (He & Neuts,
1998), Multiserver retrial queues (Artalejo, Gomez-Corral, & Neuts,
2001), Fluid queues (Dzial, Breuer, da Silva Soares, Latouche, &
Remiche, 2005; Silva Soares & Latouche, 2006), and even the most
notable telecommunication systems at the present time (Conti,
Gregori, Lenzini, & Neuts, 1994; Li, Widjaja, & Neuts, 1998). How-
ever, most of them focus on evaluating the average performance
characteristics for some certain situations of systems, such as find-
ing the steady-state probability distribution, little effort has been
spent on the control policy or aspect, with the exception of the re-
lated problems of server vacations (Alfa & Frigui, 1996). In this pa-
per we consider the group maintenance policies for M/M/N
unreliable queuing service systems. As we mentioned before, the
production loss cost is not constant, but depending on the number
of job waiting in the queue line. This is actually a QBD process with
two dimensions, the level for the arrival/service process and the
phase for the failure/repair process. Consequently, it is very appro-
priate to develop the group maintenance policies by implementing
the matrix analytic methods to handle the failure/repair process of
servers and the arrival/service process of customers simulta-
neously, and obtain the steady state probability vector for number
of customers waiting in the system.
In the sequel, Section 2 will discuss the problem description and
formulation. Section 3 will do the modeling and steady state anal-
ysis. Section 4 will derive the related cost function. Section 5 will
demonstrate the properties of cost functions. Section 6 will illus-
trate the numerical results and related comparisons. Finally, Sec-
tion 7 will give some conclusions and discussions.
2. Problem description and formulation
This paper considers a unreliable queuing system with N iden-
tical servers and a single queue. The applied system can be internet
service system, telecommunication service system, electric power
system, nuclear power system or automatic manufacturing system,
etc. depending on different group maintenance models developed
in this paper. The customer arrivals of system follow a Poisson pro-
cess with arrival rate kw generally depending on the number of
working servers w. Each customer entering the system can only
be served by one server; the service time for each customer follows
an exponential distribution with service rate l. Each server in sys-
tem operates for a random period of time, which follows exponen-
tial distribution with failure rate f. The failed servers are repaired
by a crew of c repair persons. The repair rate for each repair person
to fix one failed server is r. It is assumed that the repair crew will
do the jobs together and devote their efforts proportionately to all
servers being repaired at a given time. Therefore, the whole system
repair rate during maintenance process will depend on the number
of servers being repaired and equal rw = rc/(N  w) while w servers
are operational (and N  w under repair). A class of group mainte-
nance policies is developed, called m-failure, under which the
maintenance process is started when the number of failed servers
reaches a predetermined level m. During the period of mainte-
nance process, all failed servers are replaced with new ones or re-
paired to become as good as new at the same time.
There is a fixed cost S associated with the initiation of the main-
tenance process, independent of the number of failed servers.
Additionally, there is a variable cost r_cost for each failed server
undergoing repair. Finally, there is a delay cost h per unit time
for each customer waiting in the system. This cost structure im-
plies the following tradeoff. For small values of m, the maintenance
process starts when only few servers are failed; therefore, the
number of working servers will be usually high and the delay cost
will be relatively low. On the other hand, the fixed cost S is paid
more often, therefore the average repair cost will be relatively
high. In contrast, for large values of m, the repair process is trig-
gered less often, hence it will keep the average repair costs low
but at the same time increase the delay cost. The objective of this
paper is to determine the optimal value of m that minimizes the
expected average repair and delay cost.
Depending on different maintenance process assumptions,
there are totally three group maintenance models developed in
this paper. The first group maintenance model assumes positive re-
pair time and allows other functional servers failed during mainte-
nance process. The second group maintenance model assumes
positive repair time and does not allow normal operating servers
failed during maintenance process because all of them will also
be checked to make sure they are operational, when those failed
units are replaced. The third group maintenance model assumes
that the repair time for replacing failed units is negligible due to
their modularizing designs.
2.1. Model 1: Positive repair time and server failures allowed during
maintenance
This group maintenance model assumes positive repair time
and allows failures of other normally operating servers during
maintenance process. In practical engineering application, this
model can be applied in multi-unit systems such as typical auto-
mated manufacturing system, electric power system, nuclear
power system, and submarine’s navigation system. The time of
maintenance process follows an exponential distribution with sys-
tem repair rate rw = rc/(N  w) while w servers are operational.
While the repairs are still in process, any server that fails in be-
tween will also enter the repair facility, and the system repair rate
is adjusted accordingly.
The above group maintenance model is related to that in Neuts
and Lucantoni (1979), in which the failed servers start repair at the
failure instant if any repairman is available. This model is actually a
QBD process with two dimensions, the level for the customer arri-
val/service process and the phase for the server failure/repair pro-
cess. The infinitesimal generator Q for the customer arrival/service
process combined with the server failure/repair process takes the
following form:
2 3
T 00 T 01 0 0 0 0 0 0 0 .. .
6T T T 0 0 0 0 0 0 .. . 7
6 10 11 12 7
6 7
6 . .. . . . . .. .. . . .. .. . . .. . .. .. . .. . 7
6 7
6 7
6 0 0 0 0 T N2;0 T N2;1 T N2;2 0 0 .. . 7
6 7
Q ¼6
6 0 0 0 0 0 T N1;0 T N1;1 T N1;2 0 .. . 7 7 ð1Þ
6 7
6 0 0 0 0 0 0 A2 A1 A0 .. . 7
6 7
6 0 0 0 0 0 0 0 A2 A1 .. . 7
6 7
6 7
4 0 0 0 0 0 0 0 0 A2 .. . 5
: : : : : : : : : :
The elements of Q ,the square blocks of dimension
(N + 1)  (N + 1), are defined as zero matrices except that:
 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024 1015

DðkÞ ¼ diagðk0 ; k1 ; . . . ; kN Þ; Dðl

 Þ ¼ diagð0; l; 2l; . . . ; NlÞ; T 00 ¼ Q  DðkÞ; T 01 ¼ T 12 ¼ T 22 ¼ . . . ¼ T N1;2 ¼ A0 ¼ DðkÞ
T x0 ¼ diagfl minðx; wÞ; 0 6 w 6 Ng; for 1 6 x 6 N  1; T x1 ¼ Q  DðkÞ   T x0 for 1 6 x 6 N  1; A2 ¼ Dðl  Þ;
A1 ¼ Q  DðkÞ  A2 ð2Þ

The matrix Q is the generator which describes the failure/repair The stationary probability vector p of Q in (5) is found by solv-
process. All elements in Q are equal to zero except that ing the equations p
 Q ¼ 0; p
 e ¼ 1, and it is given explicitly by
!1
Q 00 ¼ r 0 ; Q w;w1 ¼ wf for 1 6 w 6 N; Q w;N N Nf X
N1
p0 ¼ p1 ¼ p2 ¼ .. . ¼ pNm1 ¼ 0; pN ¼ þ þ1 ;
¼ rw for 0 6 w 6 N  m; Q w;w r Nm j¼Nmþ1 j
¼ Q w;w1  Q w;N ; for 1 6 w 6 N ð3Þ Nf
In summary, the generator Q can be described as follows:
pNm ¼ pN ;
r Nm
N
2 3 pj ¼ pN ; for j ¼ N  m þ 1;. . .; N  1
r 0 0 : 0 0 0 : 0 0 r0 j
6 f f  r 1 : 0 0 0 : 0 0 r1 7
6 7
6 : : : : : : : : : : 7
6 7
6 0 0 : ðN  mÞf r Nm  ðN  mÞf 0 : 0 0 r Nm 7
Q ¼6 7
6 0 0 : 0 ðN  m þ 1Þf ðN  m þ 1Þf : 0 0 0 7
6 :
6 : : : : : : : : : 77
2.3. Model 3: Instantaneous repair
4 0 0 : 0 0 0 : ðN  1Þf ðN  1Þf 0 5
0 0 : 0 0 0 : 0 Nf Nf
This group maintenance model assumes that the repair time for
ð4Þ
replacing failed servers is negligible due to their standardization
The stationary probability vector p  of Q in (4) is obtained by and modular designs. It also means that the replacement of all
solving the equations p
 Q ¼ 0; p
 e ¼ 1, and it is given explicitly by failed servers will be finished immediately after the repair process

" ! ! !#1
X r0 Y
r 0 Nmþ1
j
ðk  1Þf þ rk1 X N
r0 Nmþ1Y ðk  1Þf þ r k1 Y
j
k1
p0 ¼ 1 þ þ þ
f j¼2
f k¼2 kf j¼Nmþ2
f k¼2
kf k¼Nmþ2
k
!
r Y ðk  1Þf þ r k1
j
r
p1 ¼ 0 p0 ; pj ¼ 0 p0 ; for j ¼ 2; . . . ; N  m þ 1
f f k¼2 kf
! !
Y ðk  1Þf þ r k1
r 0 Nmþ1 Y j
k1
pj ¼ p0 ; for j ¼ N  m þ 2; . . . ; N
f k¼2
kf k¼Nmþ2
k

2.2. Model 2: Positive repair time and no server failures allowed during
maintenance
This group maintenance model assumes positive repair time
and does not allow normally operating servers failed during main-
tenance process because all of them will be checked to make sure
they are operational when those failed servers are replaced. In
practical engineering application, this model can be applied in typ-
ical automated manufacturing system, electric power system, nu-
clear power system, submarine’s navigation system, in which
their normally operating servers are checked to make sure they
will stay in function during maintenance process. In this model,
the basic model is similar to model 1 described above; the only dif-
ference is that there are no server failures during the period of
maintenance. Therefore, the infinitesimal generator Q for this
model takes the form in (1) and its elements have almost the same
definition in (2) but with modified generator of the failure/repair
process Q. Only few elements in Q need to be changed as follows:
is initiated. In the real engineering application, it is impossible for
any physical system to take zero repair time to replace its failed
units. However, some completely modularizing-design multi-unit
systems, such as internet service system or telecommunication ser-
vice system, may be assumed to take negligible repair time for
replacing failed servers or failed switchboards. Consequently, there
are no server failures, customer arrivals and customer services during
the period of maintenance. The basic structure of this model is similar
to model 2. The main difference between this model and model 2 is
that there are no customer arrivals and customer services during
the period of maintenance. To keep the same structure of group
maintenance models developed in this paper, we need to apply an
approximation approach to build this model by assuming that r is
much larger than other transition rates. This model has exactly the
same generator matrix Q as in (5) of model 2, but has different cus-
tomer arrival and service rate diagonal matrix as compared with (2):
DðkÞ ¼ diagð0; . . . ; 0; kNmþ1 ; . . . ; kN Þ; Dð l
Þ

Q w;w1 ¼ wf for N  m þ 1 6 w 6 N; and Q w;w1 ¼ 0 for 1 6 w ¼ diagð0; . . . ; 0; N  m þ 1l; . . . ; NlÞ

6Nm By assuming a very large r, the stationary probability vector p of

Q in (5) is obtained by solving the equations p Q ¼ 0; p
 e ¼ 1, and it
In summary, the modified generator Q can be described as is given explicitly by
follows:
2 3
r0 0 : 0 0 0 : 0 0 r0 p0 ¼ p1 ¼ p2 ¼ . . . ¼ pNm1 ¼ pNm ¼ 0; pN
6 0 r1 : 0 0 0 : 0 0 r1 7 !1
6 7
6
6 : : : : : : : : : : 77 X
N1
N N
6 0 0 : 0 rNm 0 : 0 0 rNm 7 ¼ þ1 ; pj ¼ pN ; for j ¼ N  m þ 1; . . . ; N  1
Q ¼6
6
7 j j
6 0 0 : 0 ðN  m þ 1Þf ðN  m þ 1Þf : 0 0 0 77 j¼Nmþ1
6 : : : : : : : : : : 7
6 7
4 0 0 : 0 0 0 : ðN  1Þf ðN  1Þf 0 5 It is clear that all infinitesimal generators developed in models
0 0 : 0 0 0 : 0 Nf Nf
1–3 are following the general form of the infinitesimal generator
ð5Þ described in Neuts (1981). Therefore, this paper will develop the
1016 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024

related matrix geometric method to compute the steady state prob- (2) Model 2:
ability distribution for all three proposed models. The details of
modeling and steady state analysis are presented in the next section.
p A2 e > p A0 e ) ½0; 0; .. .; pNm ; .. .; pN 
 diagð0; l;. .. ;N  mÞl; .. .; NlÞe > ½0; 0;. .. ; pNm ;. .. ; pN 
3. Modeling and steady state analysis X
N X
N
 diagðk0 ; . .. ;kN Þe ) kw pw < l wpw
This section models the problem described in Section 2 as a con- w¼0 w¼Nm
tinuous Markov reward process, and applies the matrix geometric (3) Model 3:
method to compute the steady state distribution. To this end, let xt,
wt denote the number of customers in the system and the number p A2 e > p A0 e ) ½0; 0; . . . ; pNmþ1 . . . ; pN 
of operating servers at time t, respectively. Then {wt, t P 0} is a  diagð0; 0; . . . ; ðN  m þ 1Þl; . . . ; NlÞe
continuous time Markov Process with infinitesimal generator Q > ½0; 0; . . . ; pNmþ1 ; . . . ; pN diagð0; 0; . . . ; kNmþ1 ; . . . ; kN Þ
that depends on the specific assumptions made on the repair pro-
X
N X
N
cess. Furthermore, {(wt, xt), t P 0} is also a continuous time Markov e) kw pw < l wpw
Process with infinitesimal generator Q . Let p  denote the stationary w¼Nmþ1 w¼Nmþ1
 the stationary distribution of Q .
distribution of Q and y
h
The physical meaning of stability condition, p 
k<p l
 assumed in
3.1. The related properties for steady state analysis
this model is that the weighted mean arrival rate should be less than
The steady state distribution is computed for all three models the weighted mean service rate. It also means that the service capa-
based on the different repair assumptions. The first thing need to bility of operating servers can handle all customers waiting in this
be considered is whether the queue is stable in these three models. system so that the number of customers in the queue line will not
From the stability condition of QBD processes for M/M/1 Queue in a go to the infinity. If the stability condition does not hold, the queuing
Markovian environment given in Latouche and Neuts (1980, 1981), line will go to the infinity and is not possible to obtain the stationary
distribution. In addition, stability condition p 
k<p l can also be de-
it follows that for the model described above, the queue is stable if 
and only if scribed as stability index pplk < 1 . According to the result of Lemma 3,
when stability condition holds, the system is stable and all irreduc-
p k < p l ; ð6Þ ible Markov Process Q of three proposed models are positive recur-
rent. This property is the most important part required to be able to
where p  is the stationary probability vector of Q in (4) or (5),
 continue the steady-state analysis of our continuous QBD processes.
k ¼ ½k0 ; k1 ; k2 ; . . . ; kN  is the vector of arrival rates, and l
 ¼ ½l0 ; l1 ;
The following Theorem 3.1, which is a particular analogue of Theo-
l2 ; . . . ; lN  is the vector of service rates. This is commonly referred
rem 1.7.1 and Theorem 3.1.1 given in Neuts (1981), is provided to
to as the ‘negative drift’ condition for stability in matrix-geometric
set up the required condition to obtain the stationary probability
solutions. (Neuts, 1981; Latouche & Ramaswami, 1999). In the con-
vector of this proposed Q . We shall subsequently assume that Lem-
tinuous QBD of our three proposed models, the similar stability
ma 3 holds, Theorem 3.1 then holds.
property holds, using the similar type of argument as the proof of
Throrem 7.2.4 given in Latouche and Ramaswami (1999).
Theorem 3.1. if the irreducible Markov Process Q is positive
recurrent, then
Lemma 3. The irreducible Markov Process Q is positive recurrent or
the queue is stable if and only if the following stability condition holds:
X
N XN XN XN (a) The minimal nonnegative matrix R satisfies the matrix-quaratic
ð1Þ Model 1 : k w pw < l wpw ð2Þ Model 2 : kw pw < l wpw equation
w¼0 w¼1 w¼Nm w¼Nm
X
N X
N  ÞÞ þ R2 Dðl
DðkÞ þ RðQ  Dðk þ l Þ ¼ 0
ð3Þ Model 3 : kw pw < l wpw
w¼Nmþ1 w¼Nmþ1 (b) The eigenvalues of R lie inside the unit disk.
P k
(c) B[R]e = 0, where the matrix B½R ¼ 1 k¼0 R Q k0
Proof. According to Theorem 7.2.4 given in Latouche and Ramasw- (d) There exists a positive vector y0 satisfied with y0 B[R] = 0, and
ami (1999), if the number of states N is finite and that the matrix normalized by y0(I  R)1e = 1
A = A0 + A1 + A2 is irreducible, the continuous time QBD is positive (e) Partitioning the stationary probability vector y  _of Q into the
recurrent if and only if (N + 1)-vectors y0, y1, y2, . . ., the N(N + 1)-vector (y0, y1,
p A2 e > p A0 e: ð7Þ y2, . . . , yN1) are first obtained by solving the system
It is clear that all infinitesimal generators developed in our three
y0 T 00 þ y1 T 10 ¼ 0
models, which have the form of Q in (1), are the continuous time
QBD processes. y DðkÞ þ y T x;1 þ y
x1 x xþ1 T xþ1;0 ¼ 0 for 1 6 x 6 N  2
Furthermore, the matrix Q in (4) and (5), which describe the yN2 DðkÞ þ yN1 ½T N1;1 þ RDðl
 Þ ¼ 0
failure/repair processes of our three models, are irreducible and the ð8Þ
X
N2
1
related number of states N is also finite. Therefore, depending on yx e þ yN1 ð1  RÞ e ¼ 1
different forms of Q , Q, 
k; l
 in model 1, model 2, and model 3 (7) x¼0
can be modified as below: then yx ¼ yN1 RxNþ1 for x P N  1; N > 2

(1) Model 1:
Proof.
p A2 e > p A0 e ) ½p0 ; p1 ; . . . ; pN Dðl Þe > ½p0 ; p1 ; . . . ; pN DðkÞe
) ½p0 ; p1 ; . . . ; pN diagð0; l; . . . ; NlÞe > ½p0 ; p1 ; . . . ; pN  (a) The proof of this first statement follows immediately from
X
N X
N the results of Theorem 1.7.1 and Theorem 3.1.1 given in
 diagðk0 ; . . . ; kN Þe ) kw pw < l wpw Neuts (1981) by using Dð kÞ to substitute for A0 ; Q  Dð
kþ
w¼0 w¼1 l Þ to substitute for A1 ; Dðl Þ to substitute for A2.
 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024 1017

(b) The expected number of transitions before the first return to X

1 X
1 X v 1
1 X

the level i, given that the process starts in the state (i, j), is A0 e ¼ ðR  Rv ÞAv e R k ) A0 e ¼ R k Av e ) A0 e
v ¼2 k¼0 v ¼2 k¼1
finite if and only if the Markov Process Q is positive recurrent.
P k X
1 X
1
It is given by the jth component of the vector 1 k¼1 R e, which ¼ Rk Av e
P1 k
is finite if and only if the matrix k¼0 R ¼ ðI  RÞ1 is finite, or k¼1 v ¼kþ1
equivalently the eigenvalues of R lie inside the unit disk
(c) It can be easily verified that (b) According to the definition of Q in (1), it is trivial to show
that
X
1 X
1 X
k X
1 X
1 X
1 X
1
B½Re ¼ Rk Q k0 e ¼  Rk Av e ¼  R k Av e A0 e ¼ Rk Av e ) k ¼ RA2 e ) k ¼ Rl

k¼0 k¼0 v ¼0 v ¼0 k¼v k¼1 v ¼kþ1
! !
X
1
k
X
1
v 1
X
1
v h
¼ R R Av e ¼ ðI  RÞ R Av e ¼ 0
3.2. The analytic and algorithmic solutions
k¼0 v ¼0 v ¼0
P1 P P1 Following the results of Lemma 3, Theorem 3.1 and Lemma 3.2,
(d) Clearly y0 B½R ¼ y0 k¼0 Rk Q k0 ¼ k¼0 y0 Rk Q k0 ¼ 1
k¼0 yk Q k0
P1 the vector of steady state probabilities corresponding to x custom-
Since yQ ¼ 0; k¼0 yk Q k0 ¼ 0 can be easily verified. The nor-
ers in the system can be expressed as yx = yN1RxN+1, for x = N,
malizing equation can be derived as follows:
N + 1, . . ., where matrix R is the solution to the following system
X
1 X
1 X
1
of equations:
yx e ¼ 1 ) y 0 Rx e ¼ 1 ) y 0 Rx e ¼ 1 ) y0 ðI  RÞ1 e
x¼0 x¼0 x¼0 X
1
Rk Ak ¼ 0;  ¼ k
Rl ð9Þ
¼1
k¼0
Finally, It can be concluded that the vector y0 is strictly positive
Furthermore, vectors y0, y1, . . . ,yN1 are computed using appro-
whenever Q is positive recurrent.
PN2 1
priate initial conditions. Therefore, the computation of the station-
(e) From yQ ¼ 0 and x¼0 yx e þ yN1 ðI  RÞ e ¼ 1; y0 ; y1 ; . . . ; ary distribution y entails, on the one hand, the solution of the
yN1 can be easily obtained. above system of Eq. (9), and on the other hand, the computation
of the initial vectors.
We next need to prove that yx = yN1RxN+1 for x P N  1, N > 2. In the following, the special case N = m = 1 is considered first for
By conditioning on the time and the state of the last visit to the set all three models. Although this case is trivial from the point of view
N  1, if there is such a visit, the relation follows that of policy design, it is sufficiently simple to admit analytic solutions
ðnÞ ðnÞ P Pn ðrÞ
P x;j;x;j ¼ N1 Px;j;x;j þ m nr
v ¼1 r¼0 P x;j;N1;v N1 PN1;v ;x;j . Adding these and provides useful intuition. Afterwards the case of general N and
equations for n ranging from 1 to n and dividing the resulting m will be computed in the form of some specific algorithm depend-
sums by n, as n ? 1, the left-hand side tends to be yx,j. Since the ing different models.
P ðnÞ
sum 1 n¼1 N1 P x;j;x;j is finite, the first term on the right-hand side
1. In the special case that N = 1 and m = 1 with analytic solutions
tends to be zero. The second term
The same analytic solutions can be derived for all three models.
X n X n  r
m
1X n
ðrÞ nr
Xm
1X ðrÞ
X
n
n The transition rate matrix Q in (4) and (5) take the same form
P x;j;N1;v N1 P N1;v ;x;j ¼ P x;j;N1;v N1 P N1;v ;x;j  
v ¼1 n n¼1 r¼0 v ¼1 n r¼0
 
n¼0 cr cr
Q¼ ð10Þ
Pm P  ðrÞ f f
tends to v ¼1 yN1;v R
xNþ1
, since n1 nr¼0 Px;j;N1;v tends to yN1,v and
Pn r n xN+1
for 1 6 v 6 m. This completes The stationary distribution of Q is found by solving pQ = 0 and
n¼0 N1 P N1;v ;x;j has the limit R
the proof of this statement. h pe = 1, thus
An additional accuracy check for obtaining R is provided by the f rc
following lemma: p0 ¼ ; p1 ¼ ð11Þ
rc þ f rc þ f

Lemma 3.2. If eigen-values of R lie inside the unit disk and Ae = 0 The queue is stable (Q is positive recurrent) if and only if
P
where A ¼ 1k¼0 Ak , then p k < p l ) f k0 þ crk1 < crl ð12Þ
P1 k
X
1 X
1 Matrix R is the minimal solution of equation k¼0 R Ak ¼ 0,
ðaÞ A0 e ¼ Rk  ¼ k
Av e ðbÞ Rl which takes the following form:
k¼1 v ¼kþ1
 Þ þ R2 Dðl
DðkÞ þ R½Q  Dðk þ l Þ ¼ 0
Proof. 8
> R00 ðcr þ k0 Þ  fR01  k0 ¼ 0
>
>
< lðR00 R01 þ R01 R11 Þ þ crR00  R01 ðf þ k1 þ lÞ ¼ 0
(a) From Theorem 3.1, there exists a minimal nonnegative matrix
P ) ð13Þ
R satisfies the matrix-quadratic equation 1 Rk Ak ¼ 0 and >
> R10 ðcr þ k0 Þ  fR11 ¼ 0
P1 k¼0k >
:
lies inside the unit disk. Clearly, k¼0 R A k ¼ 0 ) A0 þ lðR10 R01 þ R211 Þ þ crR10  R11 ðf þ k1 þ lÞ þ k1 ¼ 0
P P1 v
RðA  A0  1 v ¼2 Av Þ þ v ¼2 R Av ¼ 0 By multiplying e to both
sides of equation, it can be shown that ¼
In conjunction with equation Rl k, the following explicit
X
1
!
X
1 X
1 X
1
solution for R is obtained:
A0 e þ R A  A0  Av e þ Rv Av e ¼ 0 ) A0 e þ RAe  RA0 e  RAv e þ Rv Av e ¼ 0 Specifically,
v ¼2 v ¼2 v ¼2 v ¼2
X
1 X
1 2 3
k0 ðf þlÞ k0
) A0 e  RA0 e ¼ ðR  Rv ÞAv e ) ðI  RÞA0 e ¼ ðR  Rv ÞAv e
v ¼2 v ¼2 R¼ 4 lðcrþk0 Þ l 5 ð14Þ
f k1 k1
Since R lies inside the unit disk, the above equation can be derived lðcrþk0 Þ l
as follows:
1018 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024

The matrix geometric solution can now be expressed as yx = y0Rx, this system equation. To do so, first let D0 and Dx, 1 6 x 6
P1
for x = 1, 2, . . . . Furthermore, it can be shown that x¼0 yxw ¼ N  1, denote the matrices diag(T00), and diag(Tx1),
P1
pw ; w ¼ 0; . . . ; N, or equivalently in vector form, x¼0 yx ¼ p . Substi- 1 6 x 6 N  1, respectively.
tuting yx = y0Rx into the previous equation to obtain y0 ðI  RÞ1 ¼ p , Then the system of Eq. (8) may be written into the form
from which it follows that
y0 ¼ p
 ðI  RÞ ð15Þ y0 ¼ ½y0 ðT 00 þ D0 Þ þ y1 T 10 D1
0 ;

By substituting (11) and (14) into (15), we obtain the explicit yx ¼ ½yx1 DðkÞ þ yx ðT x1 þ Dx Þ þ yxþ1 T xþ1;0 D1
x ; for 1 6 x 6 N  2;
result yN1 ¼ fyN2 DðkÞ þ yN1 ½T N1;1 þ DN1 þ RDðlÞgD1
h i N1 ;
f ðlcrk0 f k1 crÞ lcrk0 f k1 cr
y0 ¼ p
 ðI  RÞ ¼ lðcrþk0 Þðcrþf Þ lðcrþf Þ ð16Þ X
N2
yx e þ yN1 ðI  RÞ1 e ¼ 1: ð20Þ
According to the stability condition (12), it seems reasonable to x¼0
conclude that y0 is positive.
The first three equations above can be solved with successive
iteration method. The last equation, which is known as normal
2. In the general case that N P 2 with algorithmic solutions
equation, may be used at each step to keep the successive iterations
within a compact set that contains the unique solution vector.
It is not easy to obtain an analytic expression in the general case
by solving the system of equations in (8) and (9) when the number 4. The computation of cost function
of servers in the system N or the order of the system is getting
large. This section considers the long-run expected average cost
For obtaining matrix R from system of equations in (9), the com- regarding three different models:
puting steps are as follows:
First, the equation A0 + RA1 + R2A2 = 0 will be modified as succes-
2 4.1. Model 1: Positive repair time and server failures allowed during
sive form R ¼ A0 A 
1  R A2 A1 . Then, the iterative method is
maintenance
applied to approximate the optimal R. Besides, another equation,
Rl ¼ k, is used to make sure the successive iterations within a com-
In the following the long-run expected average cost of this spe-
pact set which contains the unique solution matrix.
cific model can be computed by the following intermediate steps:
For computing the initial vectors y0, y1, . . . , yN1, it can be
implemented by applying the specific algorithm depending on
1. Expected number of customers in the system: First, denote the
different proposed models:
mean vector M ¼ ½M 0 ; M 1 ; . . . . . . ; M N , where Mw = Lwpw and
Lw = E [number of customer jw servers working], 0 6 w 6 N.
(1) Positive Repair (model 1 and model 2)
From the results of stability and steady state analysis, M can be
In some applications of these two models, the orders of the
expressed as follows:
systems may be very high, but their particular structures
allow efficient solution. Since the vector  k defined in these X
1 X
N2 X
1
M¼ xyx ¼ xyx þ xyx ; and
models are positive, the equations in (8) can be solved recur- x¼1 x¼1 x¼N1

sively, starting with the penultimate one. This leads to X

1 X
1 X
1 X
1
j
X
1
xyx ¼ yN1 xRxNþ1 ¼ yN1 ðj þ N  1ÞRj ¼ yN1 jR þ yN1 ðN  1ÞRj
x¼N1 x¼N1 j¼0 j¼0 j¼0
yi ¼ yN1 Hi ; for 0 6 i 6 N  2: ð17Þ X
1
j1 1 2 1
¼ yN1 R jR þ yN1 ðN  1ÞðI  RÞ ¼ yN1 RðI  RÞ þ yN1 ðN  1ÞðI  RÞ
j¼1
2
The matrices Hi are defined as ¼ yN1 ðI  RÞ ½ðN  1ÞI  RðN  2Þ

HN1 ¼ I;
HN2 ¼ ðT N1;1 þ RA2 ÞD1 ðkÞ; ð18Þ Therefore,
X
N2

HNm ¼ ðHNmþ1 T Nmþ1;1 þ HNmþ2 T Nmþ2;0 ÞD ðkÞ; 1 M¼ xyx þ yN1 ðI  RÞ2 ½ðN  1ÞI  ðN  2ÞR ð21Þ
x¼1
for 3 6 m 6 N:

From the above results, the expected number of customers in

The vector yN1 is then obtained by solving the system system is then given by Me.
2. Expected holding cost:
yN1 ðH0 T 0;0 þ H1 T 1;0 Þ ¼ 0;
" # Let h be the holding cost per unit time for each customer wait-
XN2
1 ing in system, then the expected holding cost is given by hMe.
yN1 Hm e þ ðI  RÞ e ¼ 1: ð19Þ
3. Expected cycle time:
m¼0
Let tij be the expected transition time from state i to j, and tii the
expected recurrence time of state i. In this model the cycle is the
The computations in (17)–(19) can be implemented very efficiently time between two successive repairs. Because the group main-
in terms of computer memory. The recursive Eq. (18) can be com- tenance is initiated when there are exactly N-m operating serv-
puted by storing only two of the matrices Hi at a time, while the ers in the system, the cycle time can be defined as tNm, Nm.
P
vector N2m¼0 H m e is computed successively in each step. The system From matrix Q in (4) and the theory from Markov renewal pro-
Eq. (19) can then be solved easily after the recursive computations cesses (Ross, 1970), then
are completed.
1 X Q jk
(2) Instantaneous repair (model 3) t j;Nm ¼  þ tk;Nm ; for j
Since some elements of vector  k are zero in this model, it is Q jj k–Nmk–j Q jj
impossible to implement the same approach used in (17)–
¼ 0; 1; . . . ; N: ð22Þ
(19) to solve the system Eq. (8). A more time-consuming,
but safe, iterative method (Neuts, 1981) is presented to solve The above equations can be solved to obtain tNm, Nm.
 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024 1019

4. Expected variable cost of repair: 4.3. Model 3: Instantaneous repair

This part only considers the transition mechanism for the num-
ber of operating servers w from starting the repair process to This model has the similar computing procedure to that in
actually finishing it. First group maintenance is started when model 1, but with a few differences as follows:
state w = N  m. The next transition will be either to state N if
maintenance is finished or to state N  m  1 if there is another 1. Expected cycle time:
working server failed, whichever happens first. If maintenance Because of instantaneous replacement, the time between two
is finished first, then this cycle ends and a new cycle begins. successive repairs will only include the time spending on the
Otherwise, the current cycle will continue until the group main- failure processes from w = N to w = N + 1  m. Therefore
tenance is finished. Therefore, during a transition cycle the state X
N
1
w varies between 0 and N  m. From any state w between 1 and t Nm;Nm ¼ ð26Þ
if
N  m, the next transition can be either to N with transition rate i¼Nþ1m

cr/(N  w), or to w  1 with transition rate wf. From state w = 0, 2. Expected variable cost of repair
the only possible transition is to state N with transition rate First the group maintenance is activated when state w is equal
cr/N. When w = N, a new cycle begins. The corresponding tran- to N-m. Because of the instantaneous repair, the group mainte-
sition flow diagram is shown in Fig. 1. nance must be finished before any other transition going on.
From the above transition mechanism, the probability of the Hence the number of repaired servers for each group replace-
number of repaired servers for each group maintenance cycle ment must be m. Then the expected variable cost of repair
is derived as follows: can be obtained as follows: var_rep_cost(m) = mr_cost.
cr=m
pðm servers repairedÞ ¼ ; 5. The properties of the cost function
cr=m þ ðN  mÞf
" #
Yi1
ðN  jÞf cr=i
pði servers repairedÞ ¼  for i ¼ m þ 1;...;N In this section, to facilitate the proof process, it is assumed that
j¼m
ðcr=jÞ þ ðN  jÞf cr=i þ ðN  iÞf
k0 = k1 = . . . = kN for model 1 and model 2, and k0 = k1 = . . . = kN,
ð23Þ kNm+1 = . . . = kN is assumed for model 3.

Then the expected variable cost of repair can be obtained as Lemma 5.1. If m = 0, then tNm, Nm = 0.
P
follows:v ar rep costðmÞ ¼ r cost  Ni¼m ip½i ser v ers repaired, where
Proof.
r_cost is the variable repair cost per failed server.
5. Expected average cost:
(1) Model 1
In this model, several costs are incurred within each cycle.
From Eq. (22), it directly follows that t Nm;Nm ¼
These are the fixed repair cost, variable repair cost, and holding P Q ½Nm;k
1
cost.  Q ½Nm;Nm þ k Nm Q ½Nm;Nm t k;Nm : By substituting m = 0
By applying the above steps, the expected average cost can be in the previous equation, it is trivial to obtain
expressed as a function of parameter m as follows: 1 X Q N;k
t N;N ¼  þ tk;N :
Q N;N k–N
Q N;N
S  v ar rep costðmÞ
av g cðmÞ ¼ þ h Me þ ð24Þ
tNm;Nm t Nm;Nm
According to Q in (4), it then follows:
This equation can be used to determine the optimal group replace-
1 Nf
ment parameter m. t N;N ¼ þ tN1;N :
Nf þ r N Nf þ r N

4.2. Model 2: Positive repair time and no server failures allowed during
Since r N ¼ cr0 ¼ 1, it can be easily verified that tN, N = 0.
maintenance
(2) Model 2
By substituting m = 0 to the Eq. (25), it follows
This model has the similar computing procedure to that in P
t N;N ¼ r1N þ Ni¼Nþ1 if1
model 1, but with a few differences as follows:
According to Q in (5), it then follows that t N;N ¼ r1N . Since
rN ¼ cr0 ¼ 1, it can be verified that tN, N = 0.
1. Expected cycle time:
(3) Model 3
Because there is no server failure during the period of mainte-
From Eq. (26), it can be shown that
nance, the time between two successive repairs will include
the time spending on repair and the failure processes from X
N
1
w = N to w = N  m + 1. Therefore
t Nm;Nm ¼ :
i¼Nþ1m
if

1 X
N
1
tNm;Nm ¼ þ ð25Þ It is obvious that tNm, Nm = 0 when m = 0. This completes the
rNm if
i¼Nþ1m proof. h
2. Expected variable repair cost:
First the group maintenance is started when state w is N  m.
Since there are no server failures during the period of mainte- Corollary 5.2. If m = 0, then the expected average cost avg_c(m) = 1.
nance in this model, the number of repaired servers for each
group maintenance action must be m. Then the expected vari- Proof. Substituting m = 0 in (24) for all three models, it is easy to
able repair cost can be obtained as var_rep_cost(m) = mr_cost show that avg_c(0) = 1 according to the results of Lemma 5.1. h
1020 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024

Lemma 5.3. tNm, Nm is increasing in m. Example 1. N = 8, c = 4, kw = 5, w = 0, . . . ,8, l = 3, f = 0.1, r = 1,

Proof. S = 2500, r_cost = 1000, h = 200

(1) Model 1 It suffices to show that tNm1, Nm1  tNm,

pk/pl Average tNm, Expected
Nm P 0. From Eq. (22) and the transition flow for group
number of Nm average
maintenance process in Fig. 1, it follows that
customers cost

1 ðN  mÞf Model m=1 0.2158 1.6699 1.549 2720
tNm;Nm ¼ þ t Nm1;Nm 1
ðN  mÞf þ cr
m
ðN  mÞf þ cr
m
cr  m=2 0.2343 1.6767 3.258 1814
m m=3 0.2561 1.6934 5.1896 1471
þ cr t N;Nm
ðN  mÞf þ m m=4 0.2829 1.7334 7.4413 1272
m=5 0.3175 1.8311 10.1815 1137
1
and t Nm1;Nm1 ¼ t Nm1;Nm þ t Nm;Nm1 P tNm1;Nm þ ðNmÞf þcr
. m = 6 0.3656 2.0928 13.7449 1056
m
Therefore, m=7 0.4422 3.09 18.9658 1127
m=8 0.6079 14.4461 29.1786 3249
t Nm1;Nm1  tNm;Nm P tNm1;Nm
cr Model m=1 0.2128 1.6668 1.5 2667
1 m
þ cr  t Nm;Nm ¼  t Nm1;Nm 2
ðN  mÞf þ m ðN  mÞf þ cr
m m=2 0.2303 1.6674 3.1786 1749
cr
m 1 1 m=3 0.2516 1.6699 5.0952 1413
 cr  t N;Nm þ 
ðN  mÞf þ m ðN  mÞf ðN  mÞf þ cr
m m=4 0.2782 1.6803 7.3452 1221
m=5 0.313 1.723 10.0952 1088
It is clear that tNm1, Nm1  tNm, Nm P 0. m = 6 0.3619 1.9106 13.6786 1004
P
1
(2) Model 2 From (25), it follows tNm;Nm ¼ rNm þ Ni¼Nþ1m if1 . It m=7 0.4397 2.9137 18.9286 1085
is clear that tNm, Nm is increasing in m. m=8 0.6079 14.4461 29.1786 3249
P
(3) Model 3 From (26), it follows tNm;Nm ¼ Ni¼Nþ1m if1 .
Model m=1 0.2083 1.6634 1.2525 3127
3
It is obvious that tNm, Nm is increasing in m. The proof is
m=2 0.2232 1.6639 2.6836 2010
completed. h
m=3 0.2414 1.6651 4.3527 1597
m=4 0.2644 1.6696 6.3552 1357
Lemma 5.4. The stability index is increasing in m. m=5 0.2948 1.688 8.8577 1184
m=6 0.3383 1.7725 12.1936 1052
 m = 7 0.409 2.3457 17.1961 1022
Proof. By definition, the stability index is equal to PNp k .
l w¼1
wpw m=8 0.5662 11.8507 27.1986 2756

According to the structure of generator Q in (4), when m
increases, the system spends more time staying in states with
P
lower w. Hence,l N 
w¼1 wpw becomes smaller while the value of pk
is still the same and equal to k. Therefore, the stability index is
increasing in m for all three models. h
Theorem 5.5. There exists an optimal group maintenance parameter
1 6 m 6 N, for which the expected average cost is minimized.
Proof. From the result of Corollary 5.2, it given by the average cost
function avg_c(m) = 1 where m = 0. On the other hand, if m > N
(which means that the group maintenance is never implemented),
the average number of customers in system Me tends to infinity.
Then the average holding cost h  Me ¼ 1, and certainly the
expected average cost avg_c (m) goes to infinite as well. From
(22), it is trivial that tN, N = 0 < tNm, Nm where 1 6 m 6 N. Hence,
avg_c(m) 6 avg_c(0) = avg_c(N + 1) = 1 where 1 6 m 6 N where
1 6 m 6 N. Therefore, for all three models, there exists an optimal
group maintenance parameter 1 6 m 6 N, which minimizes the
average cost. h
6. Numerical examples and result analysis
According to the results of the above example, the stability rate
is increasing in m, the amount of customers in system is increasing
in m, the transition time of a cycle is increasing in m, and the ex-
pected average cost is convex in m for all three models.
To compare the three proposed models, the following results
can be concluded:
First, the optimal threshold of number of the failed servers to
initiate repair process is 6 for model 1, 6 for model 2, and 7 for
model 3. Secondly, for every specific threshold m, model 3 has
the lowest stability rate, then model 2, then model 1. Thirdly, for
every specific threshold m, model 3 has the least amount of cus-
tomers in system, then model 2, then model 1. Furthermore, for
every specific threshold m, model 3 has the shortest transition
time of a cycle, then model 2, then model 1. Finally, the optimal ex-
pected average cost is 1056 for model 1, 1004 for model 2, and
1022 for model 3.
The foregoing results are anticipated, because when no server
failures, customer arrivals, and customer services are allowed dur-
ing the period of maintenance, the system spends more time stay-
ing in states with more servers working, which means a lower
stability rate, fewer customers waiting in system, a shorter cycle
transition time, and later group maintenance initiation.
In the following, the numerical analysis is performed by chang-
ing value in some specific parameter based on Example 1.

This section will present numerical examples and the related
result analysis for three proposed group maintenance models,
which can be applied to different types of multi-unit systems.
Example 1 is the basic example to illustrate all three models in this
section:
6.1. Change arrival rate k only
The following example examine the sensitivity of the model to
the arrival rate, and try to find some monotonic properties on dif-
ferent performance measures for all three models.
 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024 1021

Example 2. N = 8, c = 4, kw, w = 0, . . . ,8, l = 3, f = 0.1, r = 1, S = 2500, Example 3 (continued)

r_cost = 1000, h = 200
pk/pl Average tNm, Expected
number of Nm average
pk/pl Average tNm, Expected
customers cost
number of Nm average
customers cost S = 6500, 0.4422 3.09 18.9858 1338
m = 7
Model kw = 4, 0.3537 2.0603 18.9658 921
1 m = 7 Model S = 2000, 0.3619 1.9106 13.6786 967
kw = 5, 0.3656 2.0928 13.7449 1056 2 m = 6
m = 6 S = 2500, 0.3619 1.9106 13.6786 1004
kw = 6, 0.4388 2.7417 13.7449 1186 m = 6
m = 6 S = 7000, 0.4397 2.9137 18.9286 1322
kw = 7, 0.4445 2.8084 10.1815 1332 m = 7
m = 5
Model S = 0, m = 6 0.3383 1.7725 12.1936 847
Model kw = 4, 0.3518 1.9231 18.9286 887 3
2 m = 7 S = 100, 0.3383 1.7725 12.1936 855
kw = 5, 0.3619 1.9106 13.6786 1004 m = 6
m = 6 S = 2000, 0.409 2.3457 17.1961 993
kw = 6, 0.4342 2.4904 13.6786 1119 m = 7
m = 6 S = 2500, 0.409 2.3457 17.1961 1022
kw = 7, 0.4382 2.5718 10.0952 1257 m = 7
m = 5 S = 7000, 0.409 2.3457 17.1961 1458
m = 7
Model kw = 3, 0.2454 1.0959 17.1961 772
S = 100000, 0.5662 11.8507 27.1986 6341
3 m = 7
m = 8
kw = 4, 0.3272 1.6082 17.1961 874
m = 7
kw = 5, 0.409 2.3457 17.1961 1022 From the results in Example 3, it is evident that the optimal ex-
m = 7 pected average cost is increasing in S and the optimal group
kw = 6, 0.406 2.2344 12.1936 1144 replacement parameter m is increasing in S for all three models.
m = 6 This is intuitively expected, because when S is high, it is not eco-
nomical to perform the group maintenance too early. Accordingly,
it can be advised that higher setup cost needs later group mainte-
According to the results in Example 2, the optimal group main-
nance operation for all three proposed models.
tenance parameter m is decreasing in k and the optimal expected
average cost is increasing in k for all three models. This is antici-
pated as well, because when the arrival rate is high, the queue size 6.3. Change variable repair cost r_cost only
will be higher on average, which means earlier maintenance is nec-
essary. Therefore, it can be suggested that higher arrival rate needs The following example examines the sensitivity of three pro-
earlier group maintenance operation for all three proposed group posed models to the variable repair cost, and tries to find some
maintenance models developed in this paper. There also exist the monotonic properties on different performance measures.
following outcomes which are not showed in the above table:
the stability rate, the number of customers in system and the ex- Example 4. N = 8, c = 4, kw = 5, w = 0, . . . ,8, l = 3, f = 0.1, r = 1,
pected average cost are increasing in arrival rate k for every specific S = 2500, h = 200
m, but the transition time of a cycle is the same in all k for every
specific m. pk/pl Average tNm, Expected
number of Nm average
customers cost
6.2. Change setup cost S only
Model r_cost = 500, 0.3656 2.0928 13.7449 828
The next example examines the sensitivity of the model to the 1 m = 6
setup cost, and tries to find some monotonic properties on differ- r_cost = 1000, 0.3656 2.0928 13.7449 1056
ent performance measures for all three models. m = 6
r_cost = 3000, 0.4422 3.09 18.9858 1881
Example 3. N = 8, c = 4, kw = 5, w = 0, . . . ,8, l = 3, f = 0.1, r = 1, m = 7
r_cost = 1000, h = 200
Model r_cost = 500, 0.3619 1.9106 13.6786 784
2 m = 6
pk/pl Average tNm, Expected
r_cost = 1000, 0.3619 1.9106 13.6786 1004
number of average
Nm
m = 6
customers cost
r_cost = 3000, 0.4397 2.9137 18.9286 1824
Model S = 2000, 0.3656 2.0928 13.7449 1020 m = 7
1 m = 6
Model r_cost = 100, 0.3383 1.7725 12.1936 609
S = 2500, 0.3656 2.0928 13.7449 1056
3 m = 6
m = 6
(continued on next page)
1022 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024

Example 4 (continued) convex in m for all three models. Combining theoretical results in
Section 5 and numerical results in this section, it can be concluded
pk/pl Average tNm, Expected
that there exists a unique m to minimize the expected average cost
number of Nm average
for all three proposed group maintenance models developed in this
customers cost
paper.
r_cost = 1000, 0.409 2.3457 17.1961 1022 For comparisons of three proposed models, Example 1 clearly
m = 7 shows the trend of decreasing stability rate, amount of customers
r_cost = 30000, 0.5662 11.8507 29.1986 11286 in system, and cycle transition time in the order of model 1, 2,
m = 8 and then 3. Of course, the applied maintenance policy should be de-
signed appropriately according to the characteristic of system. But
in some cases that systems can carry out all repair processes of
According to the results of the previous table, it is very clear three proposed group maintenance models, from the aspect of run-
that the optimal expected average cost is increasing in r_cost and ning the service/production process more smoothly, it can be pro-
the optimal group maintenance parameter m is also increasing in posed that group maintenance model 1 can be the first priority
r_cost for all three cases. It is intuitive that when r_cost is high, implemented to run the repair process; if the amount of customers
the average cost to perform the group maintenance becomes high in system is always getting too high, by additionally checking those
as well. Therefore, it is more economic to perform the group main- functional servers during maintenance process, group maintenance
tenance after more server failures. It can be concluded that higher model 1 can be turned into model 2 to improve this situation;
variable repair cost needs later group maintenance operation for all moreover, by greatly increasing the repair rate of maintenance pro-
three proposed models. cess, group maintenance model 2 can be transformed into model 3
to further decrease the amount of customers in system. In addition,
6.4. Change holding cost h only all 15 examples also show that the group maintenance threshold for
model 3 is great than or equal to those for other two models; it can
The final example performs sensitivity analysis to the holding be concluded that instantaneous repair model can perform later
cost and tries to find some monotonic properties on different per- group maintenance than other two positive repair models due to
formance measures for three various proposed models. very quick, even instantaneous repair process.

Example 5. N = 8, c = 4, kw = 5, w = 0, . . . ,8, l = 3, f = 0.1, r = 1,

S = 2500, r_cost = 1000 7. Conclusion

pk/pl Average tNm, Expected This paper first formulates the group maintenance problem
number of average which can handle the failure/repair process of servers and the arri-
Nm
customers cost val/service process of customers simultaneously for M/M/N unreli-
able service systems. There are three variations in our proposed
Model h = 100, 0.4422 3.09 18.9658 818 group maintenance models. The first one is the basic model with
1 m = 7 positive repair time and allows server failures during maintenance.
h = 200, 0.3656 2.0928 13.7449 1056 The second one is a modified model with positive repair time, but
m = 6 does not allow server failures during maintenance. The last one is a
h = 600, 0.3175 1.8311 10.1815 1869 modified model with instantaneous repair. Compared with the tra-
m = 5 ditional group replacement models, these three group mainte-
Model h = 100, 0.4397 2.9137 18.9286 793 nance models developed in this paper provide the related
2 m = 7 decision makers with a better tool to design a appropriate mainte-
h = 200, 0.3619 1.9106 13.6786 1004 nance policy according to the characteristics of service system and
m = 6 the random environment in the real world.
h = 700, 0.313 1.723 10.0952 1949 According to the theoretical results of Theorem 3.1 and Lemma
m = 5 3.2, the matrix geometric solutions of the steady state probability
distribution of system are calculated by applying two specific algo-
Model h = 10, 0.5662 11.8507 27.1986 505 rithms to solve the related linear equations depending on three dif-
3 m = 8 ferent proposed models. Furthermore, the expected number of
h = 200, 0.409 2.3457 17.1961 1022 customers in the system and the expected average cost functions
m = 7 depending on the group maintenance policy parameter m are
h = 400, 0.3383 1.7725 12.1936 1406 developed based on the above matrix geometric solutions of three
m = 5 proposed models.
Three m-failure group maintenance policies are built where the
repair is initiated as soon as the number of failed servers reaches a
By observing the above table, it is very clear that the optimal ex-
predetermined threshold. For the theoretical analysis, this paper
pected average cost is increasing in h, and the optimal group main-
first shows that the transition time of a cycle and the stability rate
tenance parameter m is decreasing in h for all three cases. It is
are both increasing in m, and then proves that there exists an opti-
intuitive that when h is higher, the average holding cost is also
mal group maintenance parameter m, which can be used to find
higher. Therefore, it is more economic to perform the group main-
the minimal average cost for all three models.
tenance earlier. It can be suggested that higher holding cost needs
All numerical results obtained so far indicate that the stability
earlier group maintenance for all three proposed models.
rate, the amount of customers in system, and the transition time
Finally, these numerical results of all 15 examples completely of a cycle are increasing in m, and the expected average cost is con-
demonstrate the properties of the theoretical results proved in Sec- vex in m for all three models. Therefore, there exists a unique m to
tion 5, which include the stability rate increasing in m, the transi- minimize the expected average cost. These numerical results com-
tion time of a cycle increasing in m, and the expected average cost pletely demonstrate the properties of the above theoretical results.
 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024
This paper also studies the sensitivity of the optimal m value
and average cost value with respect to system parameters such
as arrival rate, setup cost, variable repair cost, and holding cost.
For comparison of three proposed models, the results clearly show
the trend of decreasing stability rate, amount of customers in sys-
tem, and cycle transition time in the order of model 1, 2, and then
3. It is also showed that the group maintenance threshold for
instantaneous repair model is great than or equal to those for other
two positive repair time models. The results shown above provide
decision makers greater insight by implementing the different
group maintenance policies based on different levels of system
parameters.
In practical engineering application, according to the results de-
rived by mathematical analysis and numerical examples in this
paper, it is clear that model 1 and model 2 can be applied in mul-
ti-unit systems such as typical automated manufacturing system,
electric power system, nuclear power system, submarine’s naviga-
tion system depending on if their normally operating servers are
checked to make sure they will stay in operation during mainte-
nance process, and model 3 can be applied in some completely
modularizing-design multi-unit systems, such as internet service
system or telecommunication service system, in which it is as-
sumed to take negligible repair time for replacing failed servers
or failed switchboards.
References
Aghezzaf, El-Houssaine, & Najid, Najib M. (2008). Integrated production planning
and preventive maintenance in deteriorating production systems. Information
Science, 178, 3382–3392.
Alfa, Attahiru Sule, & Frigui, I. (1996). Discrete NT-policy single server queue with
Markovian arrival process and phase type service. European Journal of
Operations Research, 88, 599–613.
Anisimov, Vladimir V. (2005). Asymptotic analysis of stochastic block replacement
policies for multicomponent systems in a Markov environment. Operations
Research Letters, 33, 26–34.
Archibald, Thomas W., & Dekker, Rommert (1996). Modified block-replacement for
multiple-component systems. IEEE Transactions on Reliability, 45(1), 75–83.
Artalejo, J. R., Gomez-Corral, A., & Neuts, M. F. (2001). Analysis of multiserver
queues with constant retrial rate. European Journal of Operational Research, 135,
569–581.
Arunraj, N. S., & Maiti, J. (2010). Risk-based maintenance policy selection using AHP
and goal programming. Safety Science, 48, 238–247.
Assaf, D., & Shanthikumar, G. (1987). Optimal group maintenance policies with
continuous and periodic inspections. Management Science, 33(11), 1440–1452.
Barlow, R. E., & Proschan, F. (1965). Mathematical theory of reliability. New York:
Wiley.
Beichelt, F. (1993). A unifying treatment of replacement policies with minimal
repair. Naval Research Logistics, 40, 51–67.
Berg, M. (1976). Optimal replacement policies for two-unit machines with
increasing running costs – I. Stochastic Processes and Applications, 5, 89–106.
Berg, M. (1978). General trigger-off replacement procedures for two-unit systems.
Naval Research Logistics, 25, 15–29.
Carazas, F. G., & Souza, G. F. M. (2010). Risk-based decision making method for
maintenance policy selection of thermal power plant equipment. Energy, 35,
964–975.
Chan, G. K., & Asgarpoor, S. (2006). Optimum maintenance policy with Markov
processes. Electric Power Systems Research, 76, 452–456.
Chien, Yu-Hung, & Chen, Jih-An (2007). Optimal age-replacement policy for
repairable products under renewing free-replacement warranty. International
Journal of Systems Science, 38(9), 759–769.
Cho, DI., & Parlar, M. (1991). A survey of maintenance models for multi-unit
systems. European Journal of Operational Research, 51, 1–23.
Conti, M., Gregori, E., Lenzini, L., & Neuts, M. F. (1994). An M/G/1 type approach to
the approximation of the slot-occupancy pattern in a DQDB network.
Performance Evaluation, 21, 59–80.
Das, K., Lashkari, R. S., & Sengupta, S. (2007). Machine reliability and preventive
maintenance planning for cellular manufacturing systems. European Journal of
Operational Research, 183, 162–180.
Dekker, Rommert, & Roelvink, Ingrid F. K. (1995). Marginal cost criteria for
preventive replacement of a group of components. European Journal of
Operations Research, 84, 467–480.
Dekker, Rommert, Wildeman, Ralph E., Schouten, Van Der Duyn, & Frank, A. (1997).
A review of multi-component maintenance models with economic dependence.
Mathematical Methods of Operations Research, 45, 411–435.
1023
Dzial, Tessa, Breuer, L., da Silva Soares, A., Latouche, G., & Remiche, Marie-Ange
(2005). Fluid queues to solve jump processes. Performance Evaluation, 62,
132–146.
Evans, R. V. (1967). Geometric distribution in some two-dimensional queuing
systems. Operations Research, 15, 830–846.
Grall, A., Berenguer, C., & Dieulle, L. (2002). A condition-based maintenance policy
for stochastically deteriorating systems. Reliability Engineering and System
Safety, 76, 167–180.
He, Qi-Ming., & Neuts, M. F. (1998). Markov chains with marked transitions.
Stochastic Processes and their Applications, 74, 37–52.
Jung, C. M., & Park, D. H. (2003). Optimal maintenance plicies during the post
warranty period. Reliability Engineering and System Safety, 82, 173–185.
Jung, Ki Mun, Park, Minjae, & Park, Dong Ho (2010). System maintenance cost
dependent on life cycle under renewing warranty policy. Reliability Engineering
and System Safety, 95, 816–821.
Kennee, J.-P., Gharbi, A., & Beit, M. (2007). Age-dependent production planning and
maintenance strategies in unreliable manufacturing systems with lost sale.
European Journal of Operational Research, 178, 408–420.
Lapa, Celso Marcelo F., Pereira, Claudio Marcio N. A., & de Barros, Marcio Paes
(2006). A model for preventive maintenance planning by genetic algorithms
based in cost and reliability. Reliability Engineering and System Safety, 91,
233–240.
Latouche, G., & Neuts, M. F. (1980). Efficient algorithmic solutions to exponential
tandem queues with blocking. SIAM Journal of Algebraic and Discrete Methods, 1,
93–106.
Latouche, G., & Ramaswami, V. (1999). Introduction to matrix analytic methods in
stochastic modeling. Philadelphia: ASA-SIAM.
Liu, Gia-Shie (2004). M-failure group replacement model for queueing systems with
unreliable and repairable servers. Journal of the Chinese Institute of Industrial
Engineers, 21(5), 423–431.
Li, Jian-Min, Widjaja, I., & Neuts, M. F. (1998). Congestion detection in ATM
networks. Performance Evaluation, 34, 147–168.
Li, Haijun, & Xu, S. H. (2004). On the coordinated random group replacement policy
in multivariate repairable systems. Operations Research, 52(3), 464–477.
Love, C. F., & Guo, R. (1996). Utilizing Weibull failure rates in repair limit analysis for
equipment replacement/preventive maintenance decisions. Journal of
Operational Research Society, 47(11), 1366–1376.
Monga, A., Zuo, M. J., & Toogood, R. (1997). Reliability based design of systems
considering preventive maintenance and minimal repair. International Journal of
Reliability, Quality and Safety Engineering, 4(1), 55–71.
Nakagawa, T. (1983). Optimal number of failures before replacement time. IEEE
Transaction on Reliability, 32(1), 115–116.
Neuts, M. F. (1978a). The M/M/1 queue with randomly varying arrival and service
rates. Opsearch, 15(4), 139–157.
Neuts, M. F. (1978b). Further results on the M/M/1 queue with randomly varying
rates. Opsearch, 15(4), 158–168.
Neuts, M. F. (1981). Matrix-geometric solutions in stochastic models – an algorithmic
approach. New York: Dover.
Neuts, M. F., & Lucantoni, D. M. (1979). A Markovian queue with N servers subject to
breakdowns and repairs. Management Science, 25, 849–861.
Okumoto, K., & Eslayed, E. A. (1983). Optimum group maintenance policy. Naval
Research Logistics Quarterly, 30, 667–674.
Pham, H., & Wang, H. (1996). Imperfect maintenance. European Journal of
Operational Research, 94, 425–438.
Pham, H., & Wang, H. (2000). Optimal (s, T) opportunistic maintenance of a k-out-of-
n: G system with imperfect PM and partial failure. Naval Research Logistics, 47,
223–239.
Popova, Elmira (2004). Basic optimality results for Bayesian group replacement
policies. Operations Research Letters, 32, 283–287.
Ritchken, P., & Wilson, J. (1990). (m, T) Group maintenance policies. Management
Science, 36, 632–641.
Ross, S. M. (1970). Applied probability models with optimization applications. New
York: Dover.
Saassouh, B., Dieulle, L., & Grall, A. (2007). Online maintenance policy for a
deteriorating system with random change of mode. Reliability Engineering and
System Safety, 92, 1677–1685.
Scarf, P. A., & Cavalcante, Cristiano A. V. (2010). Hybrid block replacement and
inspection policies for a multi-component system with heterogeneous
component lives. European Journal of Operational Research, 206, 384–394.
Scarf, P. A., Dwight, R., & Al-Musrati, A. (2005). On reliability criterion and the
implied cost of failure for a maintained component. Reliability Engineering and
System Safety, 89, 199–207.
Sheu, Shey-Huei (1997). Extended block replacement policy of a system subject to
shocks. IEEE Transactions on Reliability, 46(3), 375–382.
Sheu, Shey-Huei (1998). A generalized age and block replacement of a system
subject to shocks. European Journal of Operations Research, 108, 345–362.
Sheu, Shey-Huei, & Chien, Yu-Hung (2004). Optimal age-replacement policy of
system subject to shocks with random lead-time. European Journal of Operations
Research, 159, 132–144.
Sheu, Shey-Huei, & Griffith, W. S. (2002). Extended block replacement policy with
shock models and used items. European Journal of Operational Research, 140,
50–60.
Sheu, Shey-Huei, & Jhang, Jhy-Ping (1996). A generalized group maintenance policy.
European Journal of Operations Research, 96, 232–247.
1024 G.-S. Liu / Computers & Industrial Engineering 62 (2012) 1011–1024
Sheu, Shey-Huei, Lin, Yuh-Bin, & Liao, Gwo-Liang (2006). Optimum policies for a
system with general imperfect maintenance. Reliability Engineering and System
Safety, 91, 362–369.
Silva Soares, Ana da, & Latouche, G. (2006). Matrix-analytic methods for fluid
queues with finite buffers. Performance Evaluation, 63, 295–314.
Sun, Ji-Wen, Xi, Li-feng, Du, Shi-Chang, & Pan, Er-Shun (2010). Tool maintenance
optimization for multi-station machining systems with economic consideration
of quality loss and obsolescence. Robotics and Computer-Integrated
Manufacturing, 26, 145–155.
Vaughan, Timothy S. (2005). Failure replacement and preventive maintenance spare
parts ordering policy. European Journal of Operational Research, 161, 183–190.
Wallace, V. (1969). The solution of quasi birth and death processes arising from
multiple access computer systems. Ph.D. Thesis. Systems Engineering
Laboratory, University of Michigan, Tech. Rept. No. 07742-6-T.
Wang, Hongzhou (2002). A survey of maintenance policies of deteriorating systems.
European Journal of Operations Research, 139, 469–489.
Wilson, J. G., & Benmerzouga, A. (1990). Optimal m-failure policies with random
repair time. Operations Research Letters, 9, 203–209.
Wu, Shaomin, & Clements-Croome, Derek (2005). Preventive maintenance models
with random maintenance quality. Reliability Engineering and System Safety, 90,
99–105.
Yeh, Ruey Huei, & Lo, Hui-Chiung (2001). Optimal preventive-maintenance
warranty policy for repairable products. European Journal of Operational
Research, 134, 59–69.
Zheng, X., & Fard, N. (1991). A maintenance policy for repairable systems based
on opportunistic failure rate tolerence. IEEE Transactions on Reliability, 40,
237–244.