You are on page 1of 15

International Journal of Automation and Computing 6(4), November 2009, 364-378

DOI: 10.1007/s11633-009-0364-4

On Problems of Multicomponent System Maintenance


Modelling
Tomasz Nowakowski Sylwia Werbińka∗
Department of Logistics and Transportation Systems, Wroclaw University of Technology, Wroclaw, Poland

Abstract: We present an overview of some recent developments in the area of mathematical modeling of maintenance decisions for
multi-unit systems. The emphasis is on three main groups of multicomponent maintenance optimization models: the block replacement
models, group maintenance models, and opportunistic maintenance models. Moreover, an example of a two-unit system maintenance
process is provided in order to compare various maintenance policies.

Keywords: Multicomponent systems, corrective and preventive maintenance, maintenance free operating periods (MFOP) concepts.

1 Introduction 1) inspection maintenance,


2) preventive maintenance for single-unit and multi-unit
The prime maintenance objective is to ensure that a sys- systems,
tem performs its intended functions. Thus, generally main- 3) condition-based maintenance,
tenance can be divided into two main types: corrective 4) maintenance information management.
maintenance and preventive maintenance. The first category consists of those studies that consider
Corrective maintenance (CM) is any maintenance action the problem of optimally scheduling of inspections for sys-
that occurs when a system has been already failed; so, there tems that deteriorate or age.
is no possibility to optimize its performance with respect to The second category of research work, which is particu-
a given economic or reliability criteria. larly relevant to this paper, includes studies that examine
In the situation, when it is necessary to avoid system the problems of execution of maintenance actions for single-
failures during operation, especially when such an event is unit or multi-unit systems.
costly or/and dangerous, it is important to perform planned If the deterioration of a system or a control parameter,
maintenance actions. Preventive maintenance (PM), ac- strongly correlated with the state of the system, can be
cording to MIL-STD-721B, means all actions performed in directly measured, and if the system is subject to failure
an attempt to retain an item in a specified condition by only if it deteriorates beyond a given threshold level, it is
providing systematic inspection, detection, and prevention more appropriate to base the maintenance decision on the
of incipient failures. actual deterioration state of the system. This leads to the
The interest in development and investigation of mainte- choice of the third category – condition-based maintenance
nance problems has been extensively discussed in the liter- policy.
ature since the early 1960s. Thus, there are many possible The last category takes into account all those studies that
ways to classify the literature related to maintenance opti- discuss the proper organization of maintenance information
mization models based on the following[1−4] : management processes necessary to effective performance of
1) information availability, a system.
2) single-unit versus multi-unit systems, In the investigated literature, most of the PM models
3) time-dependent/action relationship, consist in optimizing the execution of maintenance pro-
4) model types, cesses of single-unit systems and consider a single decision
5) optimization criterion, variable. The well-known maintenance models for single-
6) methods of solution, unit systems are the age and the block replacement models.
7) planning time horizon (finite/infinite), The basic references in this area are [3, 4]. Aven et al.[13]
8) maintenance effect (perfect, minimal, imperfect). and Frostig[14] give a comparison of those maintenance poli-
The excellent basic review in the area of maintenance cies for stochastically failing equipment.
models for proper scheduling maintenance actions was pre- However, for a complex multi-component system, it may
pared by Pierskalla and Voelker[2] , where authors inves- not be advisable to replace the entire system just because
tigated discrete time vs. continuous time maintenance of the failure of one component. In fact, the system comes
models, which was later updated by Valdez-Flores and back into operation after repair or replacement of the failed
Feldman[5] . For other surveys see [1, 4, 6–11]. component by an operative one. As a result, there is an
A review of the current literature in maintenance mod- increasing interest in analyzing the maintenance models for
eling problems indicates that the existing models can be multi-unit systems.
classified under four main categories[12] : One reason for that is that the problem of proper main-
Manuscript received July 24, 2008; revised February 13, 2009 tenance scheduling is much more difficult to analyze. On
*Corresponding author. the one hand, interactions between components complicate
E-mail address: sylwia.werbinska@pwr.wroc.pl
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 365

the maintenance modeling and optimization process. On 2 Multicomponent maintenance mod-


the other hand, the components dependencies also offer els
the opportunity to group maintenance actions, which may
save costs of system performance. The general classification scheme for multicomponent
The other reason is that there is now more awareness maintenance models is presented in Fig. 1.
of the importance of developing and using more realistic Nicolai and Dekker[7] surveyed and summarized the re-
models. Maintenance and replacement models for single- search in the area of multicomponent maintenance opti-
unit systems are usually too simple when compared to the mization models. Their classification scheme mostly bases
on the type of dependency between system components.
complex systems where the applications occur. Moreover,
Thus, the focus of this overview is on the other defined
improvements in analytical techniques and the availability
group of multicomponent maintenance models.
of fast computers have allowed more complex systems to be
As it has been mentioned, the two standard maintenance
investigated. policies, the age and the block replacement models, are gen-
The maintenance optimization surveys by Cho and erally used for optimization of maintenance action perfor-
Parlar[1] , Nicolai and Dekker[7] , Thomas[5] , Wang[4] , review mance of single-unit systems or such multi-unit systems,
the growing body of literature on multi-unit systems. More where neither economic dependence, failure dependence,
comprehensive discussion in maintenance from application nor structural dependence exists. Fig. 1 shows Classifica-
point of view can be found in [12, 15]. tion scheme of maintenance optimization models for multi-
The definition of multi-component maintenance models unit systems.
is given by Cho and Parlar[1] as: multi component main-
tenance models are concerned with optimal maintenance
policies for a system consisting of several units of machines
or many pieces of equipment, which may or may not depend
on each other.
Interactions between components can be classified into
three different types (see e.g., [5]): economic dependence,
structural dependence, and stochastic dependence. In this
paper, an economic dependence implies that an opportunity
for a group replacement of several components costs less
than separate replacements of the individual components.
Stochastic dependence, also named as failure or probabilis-
tic dependence, occurs if the condition of components influ-
ences the lifetime distribution of other components. Finally,
structural dependence means that components structurally Fig. 1 Classification scheme of maintenance optimization mod-
form a part, so that maintenance of a failed component els for multi-unit systems[1,4,7]
implies maintenance of working components.
If all units in a system are economically or stochasti- The most common maintenance policy is the block re-
cally independent of one another, maintenance policies for placement policy. Under such a policy, an operating unit
single-unit models can be applied to the multi-unit main- is preventively replaced by new ones at times kT (k =
1, 2, 3, · · · ) independently on the age and state of the sys-
tenance problems analysis. However, if there can be de-
tem. Such a policy is rather wasteful, since sometimes al-
fined components interactions, then the optimal mainte-
most new systems are replaced. To overcome this unde-
nance policy is not the one of considering each subsystem
sirable feature, various modifications have been developed.
separately, and maintenance decisions will not be indepen- The age replacement policy can also be used. Under this
dent. policy, a unit is always replaced at its deterioration age T
Following the introduction, this paper is focused on or failure, which ever occurs first, where T is a constant.
multi-unit systems. Moreover, authors present the review If there exists any dependence between components (eco-
of maintenance modeling devoted to the case of binary nomic/stochastic/structural) to optimize maintenance deci-
states. Consequently, the paper is organized as follows. In sions, we may use one of the three groups of maintenance
Section 2, we present an overview of the most often applied policies:
multicomponent maintenance models. We do not aim to 1) group maintenance policies,
give a list of all papers that have appeared. Instead, we 2) opportunity-based replacement policies,
want to investigate the main ways of the multicomponent 3) cannibalization maintenance policies.
maintenance models development, presented in the recent First, the group maintenance policies may be used. Un-
literature. Moreover, in Section 2.5, we mention about the der such a policy, a group of items is replaced at the same
problem of multi-state deteriorating system maintenance time to take advantage of economies of scale. The main
modeling, investigating the main research directions in this problems connected with analyzing the group maintenance
area. Later, there is an example provided of a two-unit policies for multicomponent systems have been defined as:
system to compare the obtained results for various mainte- 1) definition of categories of units that should be replaced
nance policies, and a briefly summary. when failure occurs,
366 International Journal of Automation and Computing 6(4), November 2009

2) cost reduction by including redundant parts into sys- Scarf and Deara[17] investigated various block replace-
tem design, ment policies under general type of costs structure for two-
3) maintenance scheduling for systems of independently unit system in series. Taking into account the following
operating machines. assumptions,
Opportunity-based replacement models base on the rule, 1) perfect repair policy,
that replacement is performed at the time when an oppor- 2) failures detected immediately,
tunity arrives, like scheduled downtime, planned shutdown 3) negligible replacement times of components and sys-
of the machines, or failure of a system in close proximity tem,
to the item of interest. The analyzed problems are likely there are proposed two simple maintenance policies. First,
the same either for opportunity replacement models or for independent block replacement policy, which assumes that
group replacement models. the replacement of system components is performed on sys-
In the situation, when one machine is inoperative due tem failure and at fixed intervals Ti (i= 1, 2). The total
to, e.g., lack of components and, at the same time, one long-run cost per unit time for defined policy over [0, T1 ] is
or more other machines are inoperative due to the lack of given by
different components, maintenance personnel may “canni-   T2
balize” operative components from one or more machines to 1
C(T1 ) = c12
w1 H(T2 ) + [1 + H(T2 − t)] u(t)dt+
repair the other(s). This practice is common in systems that T
2
 T12  12 0
are composed of sufficiently identical component parts (see, cw2 + 0 cw1 H(T1 − T2 )+
e.g., [16]). Thus, cannibalization is performed to increase  T1 −T2  
system availability when spare parts or repair facilities are 0
[1 + Hy (T1 − T2 − y)]uy (t)dt dFy (t) + c12
w2
not available. As a result, the main problems that are in- (1)
vestigated in this area mostly regard to inventory planning where c12
w1 is cost of failure replacement of both components,
problem, spares allocation problem, or supply cannibaliza- H(t) is system renewal function, Hy (t) is system renewal
tion issue. function as a function of time y, Ti is time between PM
Following these considerations, in the next sections, we actions of unit i, u(t) is supportive function dependent on
examine various types of maintenance policies for multi-unit maintenance costs and distribution function of time to sys-
systems, which are the most commonly used. tem renewal[17] , and c2w2 is cost of preventive replacement
of component 2.
2.1 Block replacement models The second interesting maintenance policy is combined
block replacement policy. Under this policy, replacement
Taking into account the basic assumptions of simple
of both components simultaneously is performed at fixed
block replacement policies, each unit is replaced at failure.
intervals or on failure of the system, whether units failed
Moreover, all units in the system are replaced at periodic
or not. When system satisfies the same assumptions, the
intervals regardless of their individual age. The mainte-
analytical function of the long-run cost per unit time for
nance problem is usually aimed at finding the optimal cycle
this policy is given as
length T in order, either to minimize total maintenance and
operational costs or to maximize system availability.   T
1
The classification of block replacement policies depen- C(T ) = 2 · c12
w1 H(T ) + [1 + H(T − t)] u1 (t)dt +
T  0
dent on the types of considered maintenance problems of
deteriorating systems is presented in Fig. 2. c2w2 + c12w2
(2)
where c12w2 is cost of preventive replacement of both compo-
nents, and u1 (t) is supportive function dependent on main-
tenance costs and distribution function of time to system
renewal[17] .
The solutions of presented replacement policies were ob-
tained with the use of simulation process, by approximating
the time to failure distribution of the system by a Gamma
distribution.
The problems of periodic replacement of failed compo-
nents, named as modified block policies, were extended by
other researchers. Lai and Yuan[18] considered a parallel
redundant system that consists of n identical components
and fails when all components have failed. In the presented
model, there are taken into account two types of failures:
1) independent failures of one component in a system,
2) failures of many components of the system at the same
time, not necessary independent.
The second type of failure event may be synchronized
and is named as a common cause shock failure. This kind
of event is classified depending on its effect into two kinds:
Fig. 2 Block replacement policies for deteriorating systems 1) Non-lethal. When each component is assumed to fail
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 367

independently with probability p; the number of failed com- formulated the maintenance scheduling problem to main-
ponents is then a random variable, tain a set of machines for a given determined T . The study
2) Lethal. When every components in the system fails. presents the completely deterministic approach to decide
The defined maintenance actions performed in a system for each period t ∈ T , which machine to service (if any)
include replacement when system fails (lethal shock) or at such that total servicing costs and operating costs are min-
scheduled times kT (k = 1, 2, 3, · · · ), and minimal repair in imized. The solution is obtained with the use of branch and
case of non-lethal shock occurrence. price algorithm.
Taking into account the following assumptions: The block replacement policy is usually the basis for
1) independent failure process of any component, group maintenance policies presented in the next section.
2) continuous monitoring and detection of component
failures, 2.2 Group maintenance models
3) negligible replacement and minimal repair times, Maintenance activity carried out on technical systems
the long-run expected cost per unit time is obtained as usually regards to the group of components. A group main-
tenance is performed either when a fixed time interval is ex-
C(T ) =
pired or when a fixed number of units are failed, whichever

n−1 T T comes first. There are four main classes of group replace-
cmr 0
λk (u)F (u)du + c0 0 λn (u)F (u)du + cw2
k=1 ment policies (see Fig. 3). A T -age policy, which assumes
T (3)
that system replacement is performed after every T units of
0
F (u)du
time. An m-failure policy calls for replacing the system at
where cw2 is cost of periodic replacement of a system, c0 is the time of the m-th failure. The last policy, named (m, T )-
additional penalty cost, cmr is cost of minimal repair of a policy, combines features of both of the described classes.
component, F (t) is common survival function of a system, Under such a policy, system replacement is performed at
F (t) = 1−F (t), λn (t) is intensity function of the unplanned the time of the m-th failure or at time T , whichever occurs
replacement process, and λk (t) is intensity function of the first. The T -policy refers to the assumptions of the block
independent failures of components process. replacement.
Most studies on the periodic replacement policy focus
on the expected maintenance and operational costs func-
tion. However, there are also studies, which also take into
account the reliability criteria. One of those is a simple
block replacement model for series component system de-
veloped by Duarte et al.[19] The problem is to determine
the optimum frequency to perform preventive maintenance
in equipment in order to ensure its availability. For the Fig. 3 Group replacement policies classification
simplified assumptions:
1) constant repair rate of components, Group replacement policies are very popular due to the
2) increasing hazard rate of the components, ease with which they can be implemented in a real pro-
3) perfect repair policy, duction setting. As systems become more complicated and
the objective function (defined as a cost function per unit require new technologies or methodologies, various modifi-
time) is evaluated by cations of basic group maintenance model have been devel-
n  i oped to solve maintenance problems.

cw2 · ai · Ai ai · ciw1
C(A1 , A2 , · · · , An ) = + (4) Types of group replacement models are presented in
i=1
2μ(1 − Ai ) μ(1 − Ai ) Fig. 4.
A number of replacement models have been proposed for
subject to two-unit systems. Scarf and Deara[17] considered group

n
Ai  A (5) maintenance policies for two-unit system with failure de-
i=1
pendence. In their paper, the failure dependence is assumed
that whenever one component fails, it can induce the failure
where ciw2 is cost of preventive maintenance of component of the other components.
i, ciw1 is cost of corrective maintenance of component i, The simple replacement policy assumes that the system
ai is linear coefficient of hazard rate of component i, Ai is replaced when either both components fail or at fixed
is component iavailability, and μ is constant repair rate of intervals, whichever occur first. Taking into account the
components. assumptions presented in Section 2.1, where it described
The algorithm calculates the interval of time between PM block replacement policies investigated, the long-run cost
actions for each component, minimizing the costs in a cer- per unit time can be estimated as
tain period of time when times to failure are increasingly   T
and repair times are non-negligible. 1 12
C(T ) = cw1 H(T ) + [1 + H(T − t)] u1 (t)dt + c12
w2 .
However, presented models are aimed at optimization of T 0
the cycle length T between preventive maintenance actions (6)
performance. There is also a number of research works that More recently, Zequeira and Berenguer[21] introduced a sim-
deal with the problem of cyclically scheduling maintenance ple group replacement policy with periodic testing and in-
activities assuming a fixed cycle length. Grigoriev et al.[20] spections for two-unit standby parallel system with failure
368 International Journal of Automation and Computing 6(4), November 2009

Fig. 4 Static models for deteriorating systems

dependence. Under such a policy, components failures oc- with failure dependence and minimal repair. In the pre-
cur randomly but are detected only by periodic testing or sented model, whenever unit 1 fails, it causes an increase
inspections. If the component is found failed during the in the failure rate of unit 2 by a certain degree. Moreover,
inspection, then the corrective maintenance of the whole each unit 2 failure induces unit 1 into instantaneous failure.
system is performed. As a result, there is a system failure occurrence.
Besides periodic inspections, preventive maintenance ac- Investigated maintenance policy bases on system replace-
tions are scheduled for the system at a fixed time T since ment at age T or at failure, whichever occurs first. Before
the end of the last maintenance action (corrective or pre- the complete replacement of the two-unit system, each unit
ventive). Taking into account the following assumptions: 1 failure is assumed to be reconditioned by minimal repairs.
1) maintenance actions render both components to “as For simplified assumptions:
good as new” condition, 1) continuously monitoring of the system,
2) maintenance actions have constant durations, 2) negligible repair and replacement times,
3) lifetimes of both components after system maintenance the long-run expected cost per unit time in the steady state
action are s-independent up to the failure of a component, are evaluated by
4) inspections or tests are reliable, T
the expected costs per unit time due to testing, mainte- cw2 F 2 (t) + cw1 F 2 (t) + 0 F 2 (y)cmr (y)r1 (y)dy
C(T ) = T
nance, and accident consequences are given by 0
F 2 (z)dz
1 (8)
C(T ) = [cins Nins (T, M ) + cw1 (1 − Pop (M T ))+ where r1 is failure rate of unit 1, F2 (t) is lifetime distribu-
Tc (7)
cw2 Pop (M T ) + κcdw E(Tdw )] tion function of unit 2 and F 2 (t) = 1 − F2 (t), and cmr (y)
is cost of minimal repair of a unit (dependent on the occur-
where cins is mean cost of testing or inspection, cw1 is mean rence time of its failure).
cost of CM or replacement of the system when at least one Possible extension of the presented maintenance mod-
components has failed, cw2 is mean cost of renewal of the els for a two-unit system is to develop multi-unit systems.
system when neither component has failed, cdw is mean cost Haurie and L Ecuyer[23] considered a simple group replace-
of downtime in a cycle, Nins (T, M ) is expected number of ment model for a multicomponent system having identical
components tests or inspections in a cycle, Pop (t) is prob- elements. The system is comprised of n elements working
ability that neither component is failed at time t, Tdw is independently under the same conditions. During the op-
random variable of downtime during a cycle, κ is frequency erational cycle, if the element fails it has to be replaced
of true demands calling for the system to start up or func- by a new one. Simultaneously, there are performed pre-
tion, and E(Tdw ) mean downtime in a cycle. ventive actions, when a repairman can replace any number
The solution of presented model is provided for a simple of working elements. The maintenance problem is solved
case with numerical examples. by using the dynamic programming equation in the frame-
In another study, Lai and Chen[22] presented an eco- work of the theory of optimal control of jump processes. In
nomic periodic replacement model for a two-unit system another study, Yasui et al.[24] summarized the main basic
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 369

group replacement policies for an n-unit parallel redundant where T c is mean of the cycle length.
system. The first investigated replacement policy assumes Finally, the expected cost per unit time for (m, T ) policy
that the system is replaced when all components fail or at can by estimated as
the determined time T1 . The expected cost rate is derived 1
from C(T ) = [cw2 + ncs + (cw1 − cs )E(N ) + cdw E(Tdw ) ]
Tc
nc1 + cw2 F (T1 )n
C(T1 ) =  T1 (9) (14)
0
[1 − F (t)n ] dt where N is total number of components that fail during a
cycle.
where c1 is acquisition cost of one unit, and F (t) is probabil- However, in order to adapt the presented basic group re-
ity density function (pdf) of time to failure of a component. placement policies for practical use, the minimal repair of
Next, suppose that system is replaced only at scheduled failed components in a system before the scheduled preven-
moments kT (k = 1, 2, 3, · · · ) if the total number of unit tive maintenance action performance is introduced.
failures exceeds N2 until this moment. Expected cost ratio Sandve and Aven[26] considered the optimal replacement
may be obtained as problem of a multicomponent system, where components
are minimally repaired at failures.
C(N ) =
 2 ∞ N −1 Taking into account the following assumptions:
1 
cdw (n i (k+1)T
i )[F (kT )] kT [F (t) − F (kT )]n−i dt + 1) random and independent components failures,
k=0 i=0
 ∞ N −1 −1 2) repair times of component i are independent and iden-
1
tically distributed,
nc1 } × T n i
(i )[F (kT )] [F (kT )]n−i
k=0 i=0 3) maintenance actions restores system to “as good as
(10) new” condition,
where cdw is downtime cost per unit per unit of time. two main replacement policies under the cost constraint are
To modify this policy, assume that the replacement of investigated.
the system is performed at system failure or at periodic First, the standard T -policy is analyzed. The expression
times kT (k = 1, 2, 3, · · ·) if the total number of unit fail- for the long-run average cost per unit time is given by
ures exceeds N1 until moment of preventive replacement.   T
1
Expected cost ratio is given as C(T ) = cw2 + E [a(t)] dt (15)
T 0
C(N ) = where a(t) is the expected cost per unit of time at tdue to
 1 ∞ N −1
1
minimal repairs and downtime costs[26] .
cw2 i )[F (kT )] [F ((k + 1)T ) − F (kT )]
(n i n−i
+
k=0 i=0 An extension of the presented policy is the (T, S)-policy.
n
∞ N 1 −1
Under such a policy, the system is replaced at time S or at
nc1 } × (i )[F (kT )]i [F (kT )]n−i ·
the first failure after time T , whichever comes first. Now,
 k=0 i=0
 n−i  −1
 (k+1)T F (t) the average cost per unit time has the following form:
1− 1− dt .   
kT
F (kT ) min{YT ,S}

(11) C(T ) = cw2 + E a(t)dt ·


0
The expected cost ratio for the presented group replacement   −1
S
policies is also obtained for a “k-out-of-n” system. T+ P (YT > t)dt (16)
Later, Popova and Wilson[25] presented a comparison of T
closed-form results for expected cost function and variance where YT is time to the first component failure after T .
per unit time, derived for the three major classes of group Unfortunately, the computation of E[a(t)] is possible only
replacement policy (T -age, m-failure and (m, T )). to obtain with an approximation under some simplified
There is a system investigated comprising of n indepen- assumptions[26] .
dent components working in parallel structure. Failure time Another possible extension of the described replacement
of the components has a phase distribution. The assump- policy is to propose a two-phase maintenance policy for a
tions taken into account are group of identical repairable units with two types of com-
1) variability of costs from cycle to cycle, ponent failure (minor/catastrophic), presented in Sheu and
2) maintenance actions render system components to “as Jhang[27] .
good as new” condition, Under such a policy, there is a defined time-interval (0, T ]
3) negligible repair time. as the first phase, and the time interval (T, T + W ] as the
Expected cost per unit time for T -age policy equals to second phase. Individual units have two types of failures.
  T Type I failures (minor) are removed by minimal repairs (in
1
C(T ) = cw2 +ncs +(cw1 −cs )nF (T )+ncdw F (t)dt both phases), whereas type II failures (catastrophic) are re-
T 0 moved by replacements in the first phase, or are left idle in
(12)
the second phase.
where cs is cost of a functioning component maintenance.
The group maintenance in a system is performed at time
Expected cost per unit time for m-failure policy is given
T + W or upon the k-th idle, whichever comes first. At an
by
inspection, all idle units are replaced with new units and
1 all functioning units are overhauled so that they become as
C(T ) = [cw2 + mcw1 + (n − m)cs + cdw E(Tdw ) ] (13) good as new.
Tc
370 International Journal of Automation and Computing 6(4), November 2009

The long-run average cost per unit time for a generalized The presented classes of maintenance models are based
group maintenance policy is given by on the assumption, that the failure distribution of the sys-
tem is known with certainty. However, the failure distri-
Cop + Cmr1 + Cdw1 + Con + Cw1 + Cins1 bution of a system is usually unknown or known with un-
C(T, W, k) =
T + E(b) certain parameters in practice (see Fig. 5). In this case,
(17) the Bayesian group replacement policies are proposed. In
where Cop is expected total operational costs over the time particular, Popova[29] presented the optimal structure of
interval (0, T ], Cmr1 is expected total minimal repair costs Bayesian group replacement policies for a parallel system
over the time interval (T, T + τ ], Cdw1 is expected total of n items with exponential failure times and random fail-
downtime costs over the time interval (T, T + τ ], Con is ure parameter. Each time a replacement decision is made,
expected total overhaul costs at the time T + τ , Cw1 is ex- all n items are replaced.
pected total replacement costs at the time T + τ , Cins1 is
expected total inspections costs at the time T + τ , and b is
minimum of the order statistics[27] .
Most of the multicomponent maintenance models are
aimed at optimizing the expected maintenance and oper-
ational cost per unit time. In the article, Chelbi et al.[28]
proposed a preventive maintenance strategy for a serial sys-
tem consisting of ncomponents, which are not necessarily
identical. For the following assumptions:
1) random and stochastically independent components
failures,
2) instantaneous detection of failures,
3) availability at any given time of resources required to
undertake the replacements,
there are two replacement policies investigated under relia-
bility constraint.
The first proposed strategy is defined as follows: a pre-
ventive replacement of a system is undertaken at the same
Fig. 5 Types of group replacement models
moments kT (k = 1, 2, 3, · · · ). When any component fails
between consecutive PM actions, it is replaced by a new From the renewal theory, the expected cost per unit time
one. equals to
Taking into account, that the maintenance of every com-
ponents in a system is made separately, the PM of every cw2 − csal E(Nsal ) + cdw E(Tdw )
C(T ) = (21)
component i is undertaken at moments kT i . The station- T
ary availability of every component i would be given by
where csal is salvage cost of a component, and Nsal is num-
Tw1i Hi (Ti ) + Tw2i ber of working items salvaged.
Ai (Ti ) = 1 − (18) For the defined cost constraint, there is a discussion pre-
Ti
sented about optimality results for group replacement poli-
where Hi (t) is component i renewal function, Tw1i is aver- cies for system withn machines.
age time of corrective replacement of component i, and Tw2i Taking a further step, Sheu et al.[30] proposed an adap-
is average time of preventive replacement of component i. tive preventive maintenance model with minimal repair for
The penalty function that allows measuring the differ- repairable system and developed a Bayesian technique to
ence between group PM or separately replacement of every derive optimal maintenance policy.
component for any given period T ∗ follows: In the discussed model, a planned maintenance is car-
ried out as soon as T time units have elapsed since the last

n
(n − 1)TF maintenance action. When the system fails before age T ,
Dp (T ) = (Ai (T ∗ ) − Ai (T )) − (19)
i=1
T it is either correctively maintained (or replaced after (N -
1) maintenances) or minimally repaired, depending on the
where TF is fixed average time duration of PM operation. random repair cost at failure. At the N -th maintenance,
For the situation, when there is a considered minimal the system is replaced rather than maintained. There are
repair performance instead of replacement of failed com- also defined two types of system failures defined: minor fail-
ponents, the stationary availability of the system is given ure, when the minimal repair is performed, and catastrophic
by failure, where corrective maintenance takes place.
n T n For the following assumptions:
Tw1i 0 ri (t)dt + TF + Tiv 1) random repair cost for system failures,
i=1 i=1
A(T ) = (20) 2) increasing failure rate of a system,
T
3) negligible times of any maintenance actions,
where ri (t) is failure rate of component i, and Tiv is variable 4) infinite time span,
average time duration of a preventive replacement relating the objective is to determine the optimal maintenance plan
to the component i. that minimizes expected cost per unit of time, which is
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 371

given by the following equation: tenance models can be classified as stationary or dynamic.
In stationary models, there is assumed a long-term stable
C(N,
 T ) =   situation when the rules for maintenance do not change
δ(cw2 +cwN )
1
N over the planning horizon. The presented models in this
cdw + xhc (x)dx Fi,pII (T ) +
pII 0 i=1
overview mostly regard to this type. However, stationary
 N 
−1 models cannot incorporate dynamically changing informa-
T
(N − 1)cw2 + cwN } × 0
F i,pII (z)dz tion during operational process performance, such as a vary-
i=1
ing deterioration of components or unexpected opportuni-
(22)
ties.
where cwN is planned replacement cost at the N -th main-
To take such short-term circumstances into account,
tenance, hc (t) is density function of random repair cost cr ,
there are dynamic models proposed which can adapt the
pII is probability of catastrophic failure occurrence, and
long-term plan according to information becoming available
Fi,pII (t) is survival distribution of the time between the
on the short term. This yields a dynamic grouping policy.
(i − 1)-th and i-th unplanned maintenance.
Wildeman et al.[32] described a rolling-horizon approach
The maintenance of deteriorating systems is also fre-
that takes a long-term tentative plan as a basis for a sub-
quently modeled using Markov decision theory. Gurler and
sequent adaptation according to information that becomes
Kaya[31] considered a multicomponent system where the
available on the short term.
lifetime of each component is described by several stages,
In this paper, there is a multicomponent system consid-
which are further classified as good, doubtful, PM due, and
ered with n components. On each component i, a PM ac-
down. The system is composed of n identical and indepen-
tivity i can be carried out. The approach presented in this
dently operating components that are connected in series.
work enables interactive planning, taking into account op-
The maintenance policy assumes that system is replaced
portunities and a varying use of components during opera-
when a component enters a PM due or down state, and
tional processes performance. There is proposed a dynamic-
the number of components in doubtful states at that mo-
programming algorithm.
ment is at least N . The maintenance time is assumed to be
During the performance processes of a multi-unit sys-
negligible.
tem, some maintenance opportunities may occur due to,
The proposed maintenance policy is described by a multi
e.g., breakdowns of a unit in a series configuration. In most
dimensional Markov process. For the model under study,
cases, opportunities cannot be predicted in advance, and
there is an expression derived for the long-run average cost
because of their random occurrence, there can be used op-
per unit time given by
portunistic maintenance models to make effective mainte-
C(K, N ) = nance planning.
cw1 (E(N1 ) + (1 − p)E(N2 )) + cw2 pE(N2 ) + cw3 2.3 Opportunistic maintenance models
(23)
E(Tr ) + P (Tr = 0)E(Tr1 )
In a system with several stochastically failing parts, it
where cw3 is system replacement cost, N1 is number of tran- may be advantageous to follow an opportunistic mainte-
sition from node i to itself before a system replacement per- nance policy, in which the maintenance action to be taken
formance, N2 is number of transition from node i to (i-1) on a given part at a given time depends on the rest of the
before a system replacement performance, p is probability system. Types of group maintenance policies considered in
of preventive replacement performance, Tr is time to system the paper are presented in Fig. 6.
replacement, Tr1 is time to system replacement given that
there are N +1 doubtful components at t= 0, and P (T2 = 0)
is probability of system termination when there are more
than N doubtful components.
Optimization of the maintenance parameters is carried
out with the use of numerical methods.
Other model of group maintenance policy based on
marginal cost considerations, formulated as a Markov deci-
sion chains, is given by Dekker and Roelvink[15] . The group
age replacement problem is considered as a replacement de-
cision which is based on sufficient information about the
history of the process, being a vector containing all compo-
nent ages.
Although, lots of studies can be found that investigate
the group maintenance of multicomponent systems as a
Fig. 6 Types of opportunistic maintenance models
Markov decision process, the state space in such prob-
lems grows exponentially with the number of components. One of the earliest treatments of the opportunistic re-
Therefore, the Markov decision modeling is not tractable placement policy is the study of Radner and Jorgensen[33] .
for systems with more than three components. There are There is an opportunistic replacement of a single unin-
also a limited number of problems that can be solved based spected part considered in the presence of several moni-
on assumptions of Markov decision theory. tored parts. The policy has the (ni , N ) structure such that
Taking into account the planning aspect, the group main- an uninspected part is replaced alone on its failure or the
372 International Journal of Automation and Computing 6(4), November 2009

arrival of its preventive replacement age N , and replaced by replacements. Both types of the failures are age depen-
opportunistically with a failed part i, if its age has reached dent. The system is replaced at type II failure or at the
a critical age ni . opportunity after age T , whichever occurs first.
The presented approach is continued by van der Duyn Taking into account the following assumptions:
Schouten and Vanneste[34] , where authors investigated the 1) cost of minimal repair depends on the random and
maintenance problem of a two-component series system, deterministic part,
taking into account the possibility of (n, N )-strategies ap- 2) instantaneous detection and reparation of failures, and
plication. 3) infinite horizon planning,
In another study, Epstein and Wilamowsky[35] presented the total expected long-run cost per unit time is given as
the deterministic approach to investigate an opportunistic  ∞
replacement problem of a two-component system. The con- C(T ) = 0 [cwT + (cwII − cwT )Fp (T + ω)+ 
T 
sidered problem is defined as finding the optimal mainte- u2 (z)F p (z)pI (z)r(z)dz gw (ω)dω ×
0 (26)
nance plan when the exact time of both failure and main-  −1
T
tenance opportunities are known at the outset. 0
F p (T − z)g p (z)dz
More recently, Fard and Zheng[36] discussed an oppor-
tunistic failure rate replacement policy for a non-repairable where cwT is cost of replacement at the opportunity after
multi-unit system. T , cwII is cost of replacement at type II failure, Fp (t) is
The considered maintenance policy assumes that a unit survival distribution of the time between successive type II
is replaced when it fails or when its failure rate reaches a failures, u2 (z) is expectation with respect to random re-
given limit L. When a failed unit is replaced or its hazard pair costs, pI (t) is probability of type I failure occurrence,
rate exceeds limit L, all operating units with their failure r(t) is system failure rate, and gw (t) is pdf of time between
rates falling in (L − u, L) are also replaced. successive opportunities.
Taking into account the following simplified assumptions: In another study, Pham and Wang[39] discussed the op-
1) increasing in cycle time hazard rate of units, portunistic maintenance of a k-out-of-n: G system with
2) negligible replacement time of units, imperfect PM and allowable partial failure occurrence.
3) infinite planning horizon, The following maintenance policy was designed: each
4) s-independent failure events, failure of a system component in the time interval (0, τ )
from the renewal theory, the expected system cost rate in is immediately removed by a minimal repair. Components
the steady state is given: that fail in the time interval (τ, T ) can be lying idle. The
system is replaced when the total operating time reaches

n1
1
C(L, u) = ni [(cw2 + cw5 )P (fi ) + cw2 P (pi ) + T or with CM and PM actions when there is exactly m
i=1 T ci components idle, whichever occurs first. That is, if m com-
(cw2 + cw5 )P (ai )] ponents fail in the time interval (τ, T ), CM combined with
(24) PM is undertaken; if less than m components fail in the
where cw5 is fixed replacement cost, fi , pi , ai are event of time interval (τ, T ), then PM is carried out at time T .
corrective, preventive, or active replacement of type i unit, For the following supplementary assumptions:
n1 is number of types of units in a system, and ni is number 1) s-independent failure events,
of type i units. 2) negligible time of minimal repair,
The presented model has been modified by Zheng and 3) age-dependent and number of minimal repairs-
Fard[37] . In this paper, there is a repairable multi-unit sys- dependent minimal repair costs,
tem considered operating under the same replacement pol- 4) increasing failure rate of every component,
icy with one exception. When a unit fails with the hazard 5) perfect preventive maintenance,
rate in (0,L − u), then it is minimally repaired, with known the limiting average system availability for an infinite plan-
repair rate, instead of replacing it. ning horizon are obtained:
For the same assumptions presented above, the expected
  T −τ 
system cost rate is evaluated as A(τ, T ) = τ + 0 F m (t)dt ×

T −τ
C(L, u) =  0
F m (t)dt + τ + Tw2 + Fm (T − τ )(Tf − Tw2 )}−1
n1
1 rmr1 (27)
ni (cmr1 + )H(Di ) + cw2 P (pi)+
i=1 T ci μ (25) where Fm (t) is cumulative distribution function (cdf) of
cw5 time between maintenance, Tf is time to perform CM to-
(cw1 + cw5 )P (f i) + (cw2 + )P (ai)
1 + mi gether with PM, and Tw2 is time to perform PM alone the
long-run. The expected system maintenance cost per unit
where cmr1 is fixed repair cost, rmr1 is repair cost rate, μ
time is
is repair rate, mi is mean value of number of type i units
which are actively replaced at the end of cycle j given an  τ
C(τ, T ) = n 0φ(y)rc (y)dy+Fm (T − τ )(cf − cw2 )+
active replacement on unit l of type i at the end of cycle j,  T −τ
hi (t) is hazard rate of a type i unit, and Di is h−1 cw2 } × 0 F m (t)dt + Fm (T − τ )(Tf − Tw2 )+
i (L − u).
Taking a further step, Jhang and Sheu[38] investigated τ + Tw2 }−1
a multi-unit system that has two types of failures. Type (28)
I failures (minor failures) are removed by minimal repairs, where φ(t) is expectation with respect to minimal repair
whereas type II failures (catastrophic failures) are removed costs and number of performed minimal repairs, rc (t) is
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 373

failure rate of a component, and cf is cost of CM combined 2) MTBF can only be accurately calculated after a con-
with PM. siderable period of time,
When taking into account, that PM actions are imper- 3) MTBF relies on probabilistic.
fect, the expressions for expected system maintenance cost As a result, the high costs driven by the consequences
per unit time and limiting system stationary availability of the systems unreliability have been the starting point
also have been developed. for searching new measures to describe the reliability of a
However, a key conclusion from the literature on multi- system — especially in a field of aircraft maintenance.
component maintenance models is that the optimal main- The new reliability measure, having been investigated
tenance policies are difficult to compute. In addition, be- since mid 1990s, is maintenance free operating periods
cause of their complex form, it is very difficult to use them (MFOP). This measure is defined as a period of operation
in practice. For this reason, some other methods have been during which the equipment must be able to carry out all
developed, which give the opportunity to obtain models de- its assigned missions without any maintenance action and
signed to yield practical and easy to implement policies. For without the operators being restricted in any way due to
example, Hopp and Kuo[40] developed the three heuristics system faults or limitations[45,46] . The main classification
and a lower bound for a system with all non-safety-critical scheme is presented in Fig. 7, and the comparison of both
components. First, hierarchical approach for scheduling re- the measures, MTBF and MFOP, is given in [44].
placement epochs for n components is defined. Later, the The first MFOP mathematical models are developed in
obtained results are compared to the other heuristics: se- [45]. In the presented paper, MFOP is defined as a period
quential approach and base interval approach. In another during which the equipment shall operate without failure
study, Haque et al.[41] applied genetic algorithm with fuzzy and without the need for any maintenance (however, mi-
logic controller to get a near optimal decision for oppor- nor, planned maintenance is permissible). The probability
tunistic replacement of a multi-unit system. of not having any unscheduled maintenance for a period of
tmf life units is given as maintenance-free operation period
2.4 Maintenance modeling using MFOP survivability (MFOPS).
concepts Authors developed two mathematical models to predict
MFOPS:
Maintenance and reliability problems of technical sys- 1) MFOP prediction is the mission reliability approach.
tems have received considerable attention by many re- 2) MFOP prediction is the alternating renewal theory.
searchers and practitioners. The main models that address Let us consider a system with n components connected
maintenance strategy have been presented above. in series. Let us also assume that
The presented literature review should be also extended 1) system time to failure and time to repair follows arbi-
on the reliability measure that considers reliability require- trary distributions,
ments to be based on operational requirements. Moreover, 2) time to failure distributions of various items of a sys-
the investigated measure depends on the planned and un- tem are independent.
planned maintenance action performance. The probability that the system will survive the i-th cy-
Reliability can be measured and investigated in a dif- cle of MFOP, given that it survives (i − 1) cycles is given
ferent ways. Generally, overall multi-system performance as
depends on individual system availability, which is quan- n
Rk (i · tmf )
M F OP S(tmf , i) = (29)
tifiable by the measures time to failure (TTFi ) or time to Rk ([i − 1] · tmf )
k=1
restoration (TTRi ) (see, e.g., [42]). Example of such a real
system can be found in [43]. where Rk (tmf ) is reliability of the k-th component for (the
However, the most common reliability measure is main- first) tmf life units.
tenance time between failures (MTBF), which has become In the second model, there is used alternating renewal
a way of defining the acceptability of failures by counting theory approach. MFOPS is found during a stated period
the number of failures per hour[44,45] . of T along with the maintenance recovery period.
However, the presented measure has limitations, e.g., Let us consider a repairable system, assuming that
[44]: 1) time to failure distribution of an item follows arbitrary
1) MTBF assumes a constant failure rate, distribution with density function f (t),

Fig. 7 MFOP modelling concepts


374 International Journal of Automation and Computing 6(4), November 2009

2) maintenance recovery time of the item follows arbi- of standby components, subject to factors such as mainte-
trary distribution with density function g(t), nance capability, or system availability maximization, cost
3) the item can be in one of two states {1,0}, where “1” of standby items, or operational cost minimization. Many
is up state and “0” is down state. of maintenance models which investigate such problems are
The probability which is the system will operate for at presented below. However, real systems and its components
least tmf life units before it fails during T hours of operation can function in degraded states.
is given by Systems that can exist in more than two states are called
 T multi-state systems. For a stochastically failure system in
P1 (T ) = R (tmf ) + f (μ |tmf ) P0 (T − μ) dμ (30) which deterioration takes place only from state to state
0 (state-dependent deterioration), replacement policies and
inspection policies using Markov processes were investi-
and  T gated. An interesting example of such a model is presented
P0 (T ) = g (ν) P1 (T − ν) dν (31) in [51], where an optimal inspection time interval for a pe-
0
riodically inspected 4-state Markov system is determined.
where f (μ |tmf ) is probability that system fails at time μ. However, under the Markovian formulation, the deteriora-
There are also some numerical solutions for various fail- tion of a system is indicated only by the changes of states.
ure distributions provided in the paper. Thus, there is an assumption made that performance of a
Another way of MFOP concepts model is presented in system within each state does not age. In practice, this
[47]. Authors describe the ultra reliable aircraft model assumption may be invalid.
(URAM): an aircraft reliability and maintenance discrete To overcome this simplification, more general system
event simulation model that is designed to investigate configurations, in which random shocks could occur or
MFOP concepts. There are also identified key factors de- deterioration could take place within a state (state-age-
termining MFOP achievement. deterioration) were examined with the use of semi-Markov
The extension of MFOP model is given by Todinov[48−50] . processes.
The author proposed the new reliability measure: minimum State-age-dependent replacement policies for a multi-
failure-free operating period (MFFOP), which can be de- state deteriorating system is considered, e.g., in [52]. The
fined as a combination of specified minimum intervals before optimization criterion in this model is to minimize the ex-
random variables in a finite time interval, whose existence pected long-run cost rate.
is guaranteed with a minimum probability PM F F OP [49] . Some practical application frameworks for multi-state
Assuming that a system is composed of a non-repairable systems maintenance modeling also have been developed.
component, and taking into account the following assump- For example, see [53], where a flow transmission water pipe
tions: system is analyzed. In that paper, a statistical model for
1) “as good as new” replacement of a component, condition-based maintenance policy of multi-state systems
2) “critical repair” to be a critical event that leads to a has been developed by combining the universal generating
system halt or degeneration of the required, function below function and Markov chain analysis theories.
a minimum acceptable level and to require an immediate Another interesting problem in the area of multi-state
intervention for repair, system maintenance modeling is selective maintenance op-
3) random failures following a homogeneous Poisson pro- timization for multi-state systems (see, e.g., [54]), where
cess, the problem is what maintenance activities should be per-
minimum probability PM F F OP is given by formed during the limited amount of time. To solve the

r  k integer, non-linear programming problem is used the short-
(λa)k exp(−λa) ks
PM F F OP = × 1− (32) est path method.
k! a
k=0

3 An example
where (λa) exp(−λa)/k! is probability of exactly k failures
k

in the finite time interval a, and p(S |k) = (1 − ks/a)k is According to the replacement policies, described in the
conditional probability that given k random failures, before presented overview, a system is maintained to different op-
each failure there will be a failure-free gap of length at least eration levels with various time intervals. Moreover, every
s. maintenance strategy takes into consideration various in-
Author also provide some application examples, and put and output parameters (e.g., cost constraints) or model
presents simulation algorithms for evaluating the probabil- assumptions (e.g., non-/negligible maintenance time, ran-
ity of MFFOP existence. dom/constant repair cost, different dependencies between
2.5 Multi-state system maintenance mod- components) in order to obtain the optimal maintenance
policy. Consequently, the issue of this section is to study
eling
a described below system performance under the different
The plethora of existing studies in reliability and main- maintenance policies. The presented comparison is limited
tenance modeling has been devoted to the case of binary only to those policies that can be evaluated analytically.
states. That is, a system and its components could, at any We consider a system that is composed of two identi-
point in time, presume only two operational states: func- cal components (denoted as units 1 and 2), whose failures
tion or failure. For this issue, maintenance policy problem are random and stochastically independent. This kind of a
is usually connected with determination of optimal number system with no redundancy will often be better to model
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 375

as one component or to model each component separately. According to the presented model assumptions, to eval-
Despite this, the various maintenance policies, described in uate the long-run cost per unit time, there must be defined
Section 2, can be compared on the basis of economic cost cost of minimal repair of failed unit 1 denoted by cm (t).
and system reliability. Suppose that cm (t) is a constant equal to 60 per failure, for
The problem of proper maintenance schedule planning the exponential time to failure of both components the ex-
for a two-unit system can be considered for the following pected cost per unit time given by (8) are easily obtainable.
main cases: Fig. 10 shows plots of the expected cost per unit time as
1) identical/various probability distributions of time to a function of T for both age policies: T -age policy and pe-
failure of a given components, riodic replacement with minimal repair. Moreover, Fig. 11
2) identical/various failure rates of components, illustrates the long-run cost per unit time as a function of
3) minimal repair performance instead of replacement of T for various failure rates of system components.
failed components,
4) failure dependence occurrence.
First, to obtain the maintenance scheduling plan for
the presented system, the simple group replacement model
is proposed by Scarf and Deara[17] . Under such a pol-
icy, failed components are replaced immediately. The sys-
tem itself is only renewed when both components are re-
placed simultaneously at failure or at scheduled moment
kT (k = 1, 2, 3, · · ·) (see Fig. 8).

Fig. 9 The expected cost per unit time as a function of T for


considered group replacement model
Fig. 8 System maintenance plan

The exponential distribution F (t) = 1 − exp(−λt) is used


to model the lifetimes of both components. The parameters
are λ1 = λ2 = 0.001, and λ1 = 0.1 and λ2 = 0.001.
Considered costs are as follows:
1) cost of failure replacement of both components c12
w1 =
1500,
2) cost of corrective replacement of component 2 c2w1 =
1000,
3) cost of preventive replacement of both components c12
w2
= 100.
According to the defined model assumptions, the long-
run cost per unit time described by formula (6), when p = 1
and q = 0 is given by Fig. 10 The expected cost per unit time as a function of T for
T -age policy and periodic replacement with minimal repair
C(T ) =
T
c12 2
w1 H1 (T ) + cw1 0
[1 + H1 (T − t)] F 1 (t)dH2 (t) + c12
w2
T
(33)

Fig. 9 presents the long-run cost per unit time as a func-


tion of time interval T for various values of components
failure rates.
In the second approach, the T -age policy proposed by
Popova[29] is investigated.
To obtain the long-run cost per unit time for this model,
according to (13), there is a need to estimate the down-
time cost of cdw per unit time until the failed component
is replaced and the cost of either replacing or repairing a Fig. 11 The expected cost per unit time as a function of T for
functioning component to as good as new, denoted by cs . periodic replacement with minimal repair
Thus, the long-run cost per unit time is evaluated for the
same simplified assumptions, when cdw is equal to 250 and To complete the analysis, the expected cost per unit time
cs equals 70. Taking a further step, a more difficult model, as a function of failure rate λ for the investigated replace-
defined by Lai and Chen[22] , can be considered. ment policies (see Figs. 12–14) is also presented.
376 International Journal of Automation and Computing 6(4), November 2009

Presented examples are given to illustrate the methods,


and let us make the following conclusions:
1) It can be seen that the cost C(T ) can vary for a dif-
ferent level of parameter λi . When the failure intensity for
both components is the same, the lower long-run cost per
unit time as a function of T can be obtained for age re-
placement policy with minimal repairs. However, taking
into account the failure dependence between components
the decision which policy should be chosen is less obvious.
2) It is also clear, that for small values of parameter λ,
there will be more failures in the system. Thus, the choice
between the policies in terms of system performance should
be confirmed also by the reliability analysis.
Moreover, the implementation of replacement policies re- Fig. 14 The expected cost per unit time as a function of λ2
quires some level of maintenance reporting and control to when T = 2000 time units and λ1 = 0.001
gain successful results. The principal advantage of the block
policies is that component age does not need to be moni-
4 Conclusions
tored. Thus, the requirements of the maintenance manage-
ment becomes simpler. The incorporation of minimal re- In this paper, we have reviewed the literature on the most
pair performance instead of failed component replacement commonly used optimal multicomponent maintenance mod-
requires the definition and evaluation of any costs which are els. This let us draw following conclusions:
involved in this maintenance action. 1) The main mathematical methods used for analyzing
This demonstrates once again, that the decision whether maintenance scheduling problems include applied probabil-
to consider independent or grouped block replacement poli- ity theory, renewal reward processes, and Markov decision
cies or an age policy depends on the relative mean times to theory. However, there are a lot of maintenance problems,
failure of the two components working in a system. where the functional relationship between the system s in-
A possible extension of this work is to model the system put and output parameters cannot be described analyti-
performance with the use of other probability distributions cally. Thus, in practice, various maintenance models have
of components lifetimes, and to propose a simulation model been developed, which apply linear and nonlinear program-
that let us obtain the optimal policy parameters. ming, dynamic programming, simulation processes, and
heuristic approaches. These were only mentioned in the
presented overview.
2) Most maintenance models, to obtain the optimal main-
tenance parameters, take into account only the cost con-
straint. However, maintenance actions are aimed at improv-
ing system dependability. For complex systems, where var-
ious types of components have different maintenance cost
and different reliability importance in the system, it is more
appropriate to analyze the optimal maintenance policy un-
der cost and reliability constraints simultaneously.
3) Many maintenance models consider the grouping of
maintenance activities on a long-term basis with an infini-
tive horizon. In practice, planning horizons are usually fi-
nite for a number of reasons: information is only available
Fig. 12 The expected cost per unit time as a function of λ when over the short term, a modification of the system changes
T = 2000 time units the maintenance problem completely, and some events are
unpredictable.
4) In most existing literature on maintenance theory, the
maintenance time is assumed to be negligible. This as-
sumption makes, e.g., availability modeling impossible or
unrealistic. Obtained results are not traceable to practical
situations.
5) Most of the maintenance models for complex system
are based on the following assumptions: infinite system
planning horizon, steady-state conditions, perfect repair
policy, etc. The models resulting from these assumptions
are often an oversimplified version of the real-world system
behavior.
6) Most maintenance models assume that whenever a sys-
Fig. 13 The expected cost per unit time as a function of λ1 tem component is to be replaced, a new component is imme-
when T = 2000 time units and λ2 = 0.001 diately available. This implies that either the components
T. Nowakowski and S. Werbińka / On Problems of Multicomponent System Maintenance Modelling 377

are highly standardized so that they can be immediately [9] Y. S. Sherif. Reliability Analysis: Optimal Inspection &
delivered from suppliers, or that they are so inexpensive, Maintenance Schedules of Failing Equipment. Microelec-
that there can be stored large amount of spares as a protec- tronics and Reliability, vol. 22, no. 1, pp. 59–115, 1982.
tion against system failures. Taking into account the “real [10] R. Dekker. Applications of Maintenance Optimization Mod-
els: A Review and Analysis. Reliability Engineering and
life situations”, the number of spare parts is usually lim-
System Safety, vol. 51, no. 3, pp. 229–240, 1996.
ited, and the procurement lead time is non-negligible. This
[11] R. P. Nicolai, R. Dekker. A Review of Multi-component
implies, that the maintenance policy and spare provisioning
Maintenance Models. In Proceedings of European Safety
policy must be closely coupled, because separate treatment and Reliability Conference, pp. 289–296, 2007.
of them will not result in the system optimal maintenance [12] P. A. Scarf. On the Application of Mathematical Models in
policy. Maintenance. European Journal of Operational Research,
7) Maintenance modelling development during the last vol. 99, no. 3, pp. 493–506, 1997.
decades has taken into account the application of imper- [13] T. Aven, R. Dekker. A Useful Framework for Optimal
fect PM, system performance under uncertainty (e.g., lack Replacement Models. Reliability Engineering and System
of information, unknown distribution functions of compo- Safety, vol. 58, no. 1, pp. 61–67, 1997.
nents), different type of system failures occurrence, dynamic [14] E. Frostig. Comparison of Maintenance Policies with Mono-
grouping, inspection maintenance, etc. However, the more tone Failure Rate Distributions. Applied Stochastic Models
extended model is considered with various maintenance pa- in Business and Industry, vol. 19, no. 1, pp. 51–65, 2003.
rameters and more complex system behavior, to obtain the [15] R. Dekker, I. F. K. Roelvink. Marginal Cost Criteria for
Preventive Replacement of a Group of Components. Eu-
the robust optimal solution.
ropean Journal of Operational Research, vol. 84, no. 2,
From the theoretical point of view, much of the mainte- pp. 467–480, 1995.
nance work is of mathematical interest only exploring the [16] W. W. Fisher. Markov Process Modelling of a Maintenance
modelling methods. That is one of the reasons, why the System with Spares, Repair, Cannibalization and Man-
application of maintenance models has been rather limited power Constraints. Mathematical Computer Modelling,
in practice. The difficulty with application of maintenance vol. 13, no. 7, pp. 119–125, 1990.
models lies in making the models simple enough to be both [17] P. A. Scarf, M. Deara. Block Replacement Policies for a
tractable and accessible to practitioners. Two-component System with Failure Dependence. Naval
8) Moreover, maintenance and replacement decisions Research Logistics, vol. 50, no. 1, pp. 70–87, 2003.
base on the information, e.g., failure data of the equipment [18] M. T. Lai, J. Yuan. Periodic Replacement Model for a Par-
under consideration, maintenance performance times, and allel System Subject to Independent and Common Cause
Shock Failures. Reliability Engineering and System Safety,
type and number of necessary support resources. Sufficient vol. 31, no. 3, pp. 355–367, 1991.
data rarely exist for estimating parameters in a complex
[19] J. A. C. Duarte, J. C. T. A. Craveiro, T. P. Trigo. Op-
model, and if data do exist, they are often unreliable. This timization of the Preventive Maintenance Plan of a Series
makes the application of mathematical models to support Components System. International Journal of Pressure Ves-
maintenance and replacement decisions less obvious. sels and Piping, vol. 83, no. 4, pp. 244–248, 2006.
[20] A. Grigoriev, J. van de Klundert, F. C. R. Spieksma. Mod-
References eling and Solving the Periodic Maintenance Problem. Eu-
ropean Journal of Operational Research, vol. 172, no. 3,
[1] D. I. Cho, M. Parlar. A Survey of Maintenance Models for pp. 783–797, 2006.
Multi-unit Systems. European Journal of Operational Re- [21] R. I. Zequeira, C. Berenguer. Maintenance Cost Analysis
search, vol. 51, no. 1, pp. 1–23, 1991. of a Two-component Parallel System with Failure Interac-
[2] W. P. Pierskalla, J. A. Voelker. A Survey of Maintenance tion. In Proceedings of Annual Symposium on Reliability
Models: The Control and Surveillance of Deteriorating Sys- and Maintainability, IEEE Press, pp. 220–225, 2004.
tems. Naval Research Logistics Quarterly, vol. 23, no. 3, [22] M. T. Lai, Y. C. Chen. Optimal Periodic Replacement Pol-
pp. 353–388, 1976. icy for a Two-unit System with Failure Rate Interaction.
[3] C. Valdez-Flores, R. M. Feldman. A Survey of Preven- International Journal of Advanced Manufacturing Technol-
tive Maintenance Models for Stochastically Deteriorating ogy, vol. 29, no. 3, pp. 367–371, 2006.
Single-unit Systems. Naval Research Logistics, vol. 36, [23] A. Haurie, P. L Ecuyer. A Stochastic Control Approach to
no. 4, pp. 419–446, 1989. Group Preventive Replacement in a Multicomponent Sys-
[4] H. Wang. A Survey of Maintenance Policies of Deteriorat- tem. IEEE Transactions on Automatic Control, vol. 27,
ing Systems. European Journal of Operational Research, no. 2, pp. 387–393, 1982.
vol. 139, no. 3, pp. 469–489, 2002. [24] K. Yasui, T. Nakagawa, S. Osaki. A Summary of Optimum
[5] L. C. Thomas. A Survey of Maintenance and Replacement Replacement Policies for a Parallel Redundant System. Mi-
Models for Maintainability and Reliability of Multi-item croelectronics and Reliability, vol. 28, no. 4, pp. 635–641,
Systems. Reliability Engineering, vol. 16, no. 4, pp. 297– 1988.
309, 1986. [25] E. Popova, J. G. Wilson. Group Replacement Policies
[6] T. Nakagawa. A Summary of Discrete Replacement Poli- for Parallel Systems Whose Components have Phase Dis-
cies. European Journal of Operational Research, vol. 17, tributed Failure Times. Annals of Operations Research,
no. 3, pp. 382–392, 1984. vol. 91, pp. 163–189, 1999.
[7] R. P. Nicolai, R. Dekker. Optimal Maintenance of Multi- [26] K. Sandve, T. Aven. Cost Optimal Replacement of Mono-
component Systems: A Review, Economic Institute Report tone, Repairable Systems. European Journal of Operational
2006–29, Erasmus University Rotterdam, Holand, 2006. Research, vol. 116, no. 2, pp. 235–248, 1999.
[8] H. Pham, H. Wang. Imperfect Maintenance. European [27] S. H. Sheu, J. P. Jhang. A Generalized Group Maintenance
Journal of Operational Research, vol. 94, no. 3, pp. 425– Policy. European Journal of Operational Research, vol. 96,
438, 1996. no. 2, pp. 232–247, 1996.
378 International Journal of Automation and Computing 6(4), November 2009

[28] A. Chelbi, D. Ait-Kadi, H. Aloui. Availability Optimiza- [45] U. D. Kumar, J. Knezevic, J. Crocker. Maintenance Free
tion for Multi-component Systems Subjected to Periodic Operating Period – An Alternative Measure to MTBF and
Replacement. In Proceedings of European Safety and Reli- Failure Rate for Specifying Reliability? Reliability Engi-
ability Conference, pp. 1109–1114, 2007. neering and System Safety, vol. 64, no. 1, pp. 127–131, 1999.
[29] E. Popova. Basic Optimality Results for Bayesian Group [46] H. Wu, Y. Liu, Y. Ding, J. Liu. Methods to Reduce Di-
Replacement Policies. Operations Research Letters, vol. 32, rect Maintenance Costs for Commercial Aircraft. Aircraft
no. 3, pp. 283–287, 2004. Engineering and Aerospace Technology, vol. 76, no. 1,
[30] S. H. Sheu, R. H. Yeh, Y. B. Lin, M. G. Yuang. A Bayesian pp. 15–18, 2004.
Approach to an Adaptive Preventive Maintenance Model. [47] L. Warrington, J. A. Jones, N. Davis. Modelling of Mainte-
Reliability Engineering and System Safety, vol. 71, no. 1, nance, within Discrete Event Simulation. In Proceedings of
pp. 33–44, 2001. Annual Reliability and Maintainability Symposium, IEEE
[31] Ü. Gürler, A. Kaya. A Maintenance Policy for a System Press, Seattle, WA, USA, pp. 260–265, 2002.
with Multi-state Components: An Approximate Solution. [48] M. T. Todinov. Reliability and Risk Models. Setting Reli-
Reliability Engineering and System Safety, vol. 76, no. 2, ability Requirements, John Wiley & Sons Ltd., England,
pp. 117–127, 2002. 2005.
[32] R. E. Wildeman, R. Dekker, A. C. J. M. Smit. A Dy- [49] M. T. Todinov. A New Reliability Measure Based on Spec-
namic Policy for Grouping Maintenance Activities. Eu- ified Minimum Distances before the Locations of Random
ropean Journal of Operational Research, vol. 99, no. 3, Variables in a Finite Interval. Reliability Engineering and
pp. 530–551, 1997. System Safety, vol. 86, no. 1, pp. 95–103, 2004.
[33] R. Radner, D. W. Jorgenson. Opportunistic Replacement of [50] M. T. Todinov. Minimum Failure-free Operating Inter-
a Single Part in the Presence of Several Monitored Parts. vals Associated with Random Failures of Non-repairable
Management Science, vol. 10, no. 1, pp. 70–84, 1963. Components. Computers & Industrial Engineering, vol. 45,
[34] F. A. van der Duyn Schouten, S. G. Vanneste. Analysis no. 3, pp. 475–491, 2003.
and Computation of (n, N )-strategies for Maintenance of a [51] J. Karpiński. A Multistate System under an Inspection and
Two-component System. European Journal of Operational Repair Policy. IEEE Transactions of Reliability, vol. 35,
Research, vol. 48, no. 2, pp. 260–274, 1990. no. 1, pp. 76–77, 1986.
[35] S. Epstein, Y. Wilamowsky. Opportunistic Replacement in [52] C. T. Lam, R. H. Yeh. Optimal Replacement Policies
a Deterministic Environment. Computers and Operational for Multi-state Deteriorating Systems, Technical Report
Research, vol. 12, no. 3, pp. 311–322, 1985. No. 92–41, DEEPBLUE, 1993.
[36] N. Fard, X. Zheng. An Approximate Method for Non- [53] C. M. Tan, N. Raghavan. A Framework to Practical Pre-
repairable Systems Based on Opportunistic Replacement dictive Maintenance Modeling for Multi-state Systems. Re-
Policy. Reliability Engineering and System Safety, vol. 33, liability Engineering and System Safety, vol. 93, no. 8,
no. 2, pp. 277–288, 1991. pp. 1138–1150, 2008.
[37] X. Zheng, N. Fard. A Maintenance Policy for Repairable [54] C. Chen, M. Q. H. Meng, M. J. Zuo. Selective Mainte-
Systems Based on Opportunistic Failure-rate Tolerance. nance Optimization for Multi-state Systems. In Proceedings
IEEE Transactions on Reliability, vol. 40, no. 2, pp. 237– of IEEE Canadian Conference on Electrical and Computer
244, 1991. Engineering, IEEE Press, Canada, vol. 3, pp. 1477–1482,
1999.
[38] J. P. Jhang, S. H. Sheu. Opportunity-based Age Replace-
ment Policy with Minimal Repair. Reliability Engineering
and System Safety, vol. 64, no. 3, pp. 339–344, 1999.
Tomasz Nowakowski is a professor of
[39] H. Pham, H. Wang. Optimal (τ ,T) Opportunistic Mainte- Wroclaw University of Technology, Wro-
nance of a k-out-of-n: G System with Imperfect PM and claw, Poland. He is a head of Division
Partial Failure. Naval Research Logistics, vol. 47, no. 3, of Logistics and Transportation Systems at
pp. 223–239, 1999. Mechanical Engineering Faculty. He is also
[40] W. J. Hopp, Y. L. Kuo. Heuristics for Multicomponent a president of Polish Logistic Association,
Joint Replacement: Applications to Aircraft Engine Main- member of teams of Polish Academy of Sci-
tenance. Naval Research Logistics, vol. 45, no. 5, pp. 435– ences, vice-president of Polish Maintenance
458, 1998. Society. He is editor of Terotechnology sec-
tion of quarterly Scientific Problems of Ma-
[41] S. A. Haque, A. B. M. Z. Kabir, R. A. Sarker. Optimization
chines Operation and Maintenance.
Model for Opportunistic Replacement Policy Using Genetic
His research interests include reliability, maintainability,
Algorithm with Fuzzy Logic Controller. In Proceedings of
safety of technical (generally transportation and logistic) sys-
the Congress on Evolutionary Computation, IEEE Press,
tems, computer aided operation, knowledge based (expert) sys-
vol. 4, pp. 2837–2843, 2003.
tems, uncertainty of operation, and maintenance data.
[42] U. K. Rakowsky. Modelling Reliability-adaptive Multi-
system Operation. International Journal of Automation and
Computing, vol. 3, no. 2, pp. 192–198, 2006. Sylwia Werbińka received her M. Sc. in
mechanical engineering (specialty: logis-
[43] C. Dietl, U. K. Rakowsky. An Operating Strategy for tics) from Wroclaw University of Technol-
High-availability Multi-station Transfer Lines. Interna- ogy, Poland in 2004. She is currently a
tional Journal of Automation and Computing, vol. 3, no. 2, Ph. D. candidate in Division of Logistics
pp. 125–130, 2006. and Transportation Systems at Wroclaw
[44] M. A. Brown, C. J. Hockley. Cost of Specifying Mainte- University of Technology.
nance/Failure Free Operating Periods for Royal Air Force Her research interests include logistic
Aircraft. In Proceedings of Annual Reliability and Main- support systems modelling, systems relia-
tainability Symposium, IEEE Press, Philadelphia, PA, bility, and maintenance processes design-
USA, pp. 425–432, 2001. ing.

You might also like