The current issue and full text archive of this journal is available at

www.emeraldinsight.com/1355-2511.htm

METHODOLOGY AND THEORY Optimal

availability
Optimal availability of
failure-prone systems under
395
imperfect maintenance actions
Slah Samet and Anis Chelbi
Centre de Recherche en Productique (CEREP),
Ecole Supérieure des Sciences et Techniques de Tunis, Tunis, Tunisia, and
Fayçal Ben Hmida
Unité de Recherche en Commande, Surveillance et Sûreté de Fonctionnement,
Ecole Supérieure des Sciences et Techniques de Tunis, Tunis, Tunisia

Abstract
Purpose – The purpose of this paper is to study the evolution of a system stationary availability and
determine the optimal preventive maintenance period, which maximises it in a context where
preventive and corrective maintenance actions are imperfect and have non-negligible durations.
Design/methodology/approach – The quasi-renewal process approach and a ( p, q) rule are
respectively used to model corrective and preventive maintenance. Considering the durations of the
preventive and corrective maintenance actions as well as their respective efficiency extents, a
mathematical model and a numerical algorithm are developed in order to compute the system
stationary availability.
Findings – It has been proven that for any given situation regarding the system, the repair and
preventive maintenance efficiency extents, and the downtime durations for preventive and corrective
maintenance, there is necessarily a finite optimal period T * of preventive maintenance which
maximises the system stationary availability. A sufficient condition for the uniqueness of the optimal
solution has also been derived. Numerical examples illustrated how preventive and corrective
maintenance efficiency extents affect simultaneously the system optimal availability.
Practical implications – The study considers a general industrial framework where preventive and
corrective maintenance actions are imperfect. In fact, neither the best-qualified technicians nor the
most suitable tools or spare parts are found to carry out maintenance actions. In such a context for a
large variety of technical systems, when implementing preventive maintenance policies one should
take into account the efficiency extents of maintenance actions as well as their durations in order to
evaluate and optimise the system availability. The paper provides maintenance managers with a
decision model allowing not only the computation and optimisation of system availability, but also the
investigation of how preventive and corrective maintenance efficiency extents affect simultaneously
the system optimal availability. The proposed model also allows one to find to what extent corrective
actions ineffectiveness should be tolerated without having an important availability loss.
Originality/value – The paper proposes a modified formulation of the quasi-renewal process taking
into account the non-negligible duration of corrective maintenance actions and periodic preventive
maintenance. A new numerical algorithm is also developed in this context to compute the Journal of Quality in Maintenance
quasi-renewal function that it is impossible to find in closed form. This allowed the computation and Engineering
optimisation of system stationary availability. Vol. 16 No. 4, 2010
pp. 395-412
Keywords Maintenance, Maintenance reliability, Preventive maintenance, Engineer availability process q Emerald Group Publishing Limited
1355-2511
Paper type Research paper DOI 10.1108/13552511011084544
JQME 1. Introduction
16,4 In the current context of an increasingly wild competition, the constraints of time,
quality and cost imposed on the industrial companies put them in front of the
obligation to ensure a maximum availability of their production equipment at the
lowest cost. Under these conditions, the implementation of preventive maintenance
strategies proves to be inevitable.
396 A maintenance strategy is defined as a decision rule that establishes the sequence of
actions to be undertaken according to the state of the system. With each maintenance
action one associates, a cost, a duration and a certain quality or efficiency. The
performance of a strategy is then generally evaluated in terms of the average total cost
on a given horizon or in terms of the system stationary availability. Other performance
criteria are proposed in the literature.
A great number of publications propose several maintenance policies, implying
different types of preventive and corrective actions, such as inspections, replacements
by new or used identical equipment, overhauls or repair to the as good as new state, the
minimal repair which consists in bringing the system back to the same operating state
as just before failure (as bad as old state), etc.
Basic maintenance strategies have been proposed in Barlow and Proschan (1965). A
great number of publications on the subject followed with maintenance policies based
on complex mathematical models using the renewal theory and many other stochastic
processes. There are several reviews of the literature on the subject summarising the
various models and gathering them in different classes. Among these reviews, we find
those of Valdez-Flores and Feldman (1989) and Wang (2002).
Various maintenance models consider that maintenance actions are perfectly
executed. Actually, the effectiveness of maintenance actions is generally between the
two extreme limits (“as good as new” and “as bad as old”), what is generally called
imperfect maintenance. Several models of imperfect maintenance were proposed in the
literature. They can be classified in two categories: models based on an arithmetic
reduction of the age of the system (Malik, 1979; Kijima et al., 1988; Kijima, 1989), and
models of reduction of the intensity function (Chan and Shaw, 1993; Shin et al., 1996).
With regard to the reduction of age, the improvement of the state of the system,
following a maintenance action, is equivalent to an arithmetic reduction of its age. For
the models of reduction of the intensity of failures, the improvement of the state of the
system, following a maintenance action, is equivalent to a reduction of its failure rate of
a quantity proportional to its value right before maintenance. A literature review on
imperfect maintenance can be found in Pham and Wang (1996) and Wang and Pham
(2006).
In this paper, we consider a general industrial framework where preventive and
corrective maintenance actions are imperfect. In fact, we do not always find the
best-qualified technicians, nor the most suitable tools or spare parts to carry out
maintenance actions. We consider in this context a periodic preventive maintenance
strategy applied to a single-unit system prone to random failures. A similar
maintenance policy has been studied by Wang and Pham (2006) considering cost
optimisation in a context of maintenance actions with negligible durations. Here, we
model and optimise the system stationary availability for this maintenance policy
taking into account the durations of maintenance actions. Moreover, due to the
computation complexity of the quasi-renewal function used in the mathematical model,
we propose a new numerical algorithm allowing the computation of this key quantity Optimal
for systems whose times to first failure follow any given distribution. availability
In next section we define the strategy and establish the expression of the system
stationary availability. We also prove that there always exists an optimal preventive
maintenance period, T *, maximising the system stationary availability for any given
set of parameters regarding the maintenance actions duration and the system failure
distribution. A sufficient condition for the uniqueness of an optimal solution is also 397
derived. The system asymptotic average availability calls upon the quasi-renewal
function whose analytical expression is very difficult, even impossible, to obtain in a
closed form. In section 3 we develop a computational procedure of this function for any
given quasi-renewal process with non-negligible constant repair time. Numerical
examples are presented and corresponding results are discussed in section 4. The
paper is concluded in section 5.

2. Strategy definition and mathematical model

We consider a randomly failing system with a known continuous lifetime probability
distribution. An imperfect preventive maintenance action is periodically performed
(every T time units). It follows a ( p, q) rule, which means that it restores the system to
the “as good as new” state with probability p and it keeps it in the “as bad as old” state
with probability q (q ¼ 1 2 p). An estimation method of the probability p in practice
can be found in (Whitaker and Samaniego, 1989). If the system fails between successive
preventive maintenance actions, it undergoes an imperfect repair action of
non-negligible duration after which the system failure inter-arrival time reduces to a
fraction of its immediate previous one (all successive lifetimes being independent).
Lifetimes follow a decreasing quasi-renewal process (Wang and Pham, 1996, 1997)
with parameter a (0 , a , 1).
The preventive and corrective maintenance actions have respectively constant
(deterministic) durations Tp and Tc during which the system is unavailable.
We develop, in what follows, the expression of the system stationary availability
under these conditions. This expression will allow finding the optimal period T * for
undertaking a preventive maintenance so as to maximise the system stationary
availability. We will first define the quasi-renewal process in the context of our study.

2.1 The quasi-renewal process with non-negligible repair duration

Wang and Pham (1996, 1997) defined the quasi-renewal process and the quasi-renewal
function in the case of negligible corrective maintenance action duration. This process
(called also geometric process) has been efficiently used to resolve gearboxes
maintenance problems in Kowloon Motor Bus Company, in Hongkong (Leung and
Fong, 2000). It has also been recently applied to the repair process of industrial
refrigerators in a Turkish beverage company (Samath-Paç and Taner, 2009). In both
cases, long-run average cost per time unit was considered as the performance criterion.
The basic formulation of the quasi-renewal process cannot be used as it is to
compute the quasi-renewal function in order to study the systems stationary
availability considering non-negligible repair time and preventive maintenance. We
propose in what follows a more general definition of the quasi-renewal process and the
quasi-renewal function considering constant repair duration.
JQME Observe the sequence of non-negative random variables {T 1 ; T 2 ; :::; T n }, the
counting process {N ðtÞ; t . 0} is said to be a quasi-renewal process with parameter a
16,4 and the first inter-arrival time T1, if T 1 ¼ Z 1 ; T 2 ¼ aZ 2 ; T 3 ¼ a 2 Z 3 ; . . .; T n ¼ a n21 Z n ,
where the Zi are independent and identically distributed and a . 0 is a constant.
Let Tc be the duration of the corrective maintenance actions following each failure
(Figure 1), and F T 1 ðtÞbe the system time to first failure distribution.
398 The total number N(t) of imperfect repairs that has occurred up to time t and the
arrival time of the nth renewal, SSn, satisfy the following expression:
N ðtÞ $ n , SS n # t
That is, N(t) is at least n if and only if the nth renewal occurs prior to or at time t. It is
easily seen that:
!
X n
SS n ¼ T i þ ðn 2 1ÞT c ; n $ 1 ð1Þ
i¼1

Take SS 0 ¼ 0. SS n can be written as follows:

SS n ¼ W n þ ðn 2 1ÞT c for n ¼ 1; 2; 3; :::
Pn
with W n ¼ i¼1 T i .
Let F SSn ðtÞ be the probability distribution function associated with SSn.
F SSn ðt Þ ¼ Pr fSS n , t g ¼ Pr fW n , t 2 ðn 2 1ÞT c g ¼ F W n ðt 2 ðn 2 1ÞT c Þ
F W n ðt 2 ðn 2 1ÞT c Þ is given by the convolution product of the distribution functions
F T i ðt 2 ðn 2 1ÞT c Þ for i ¼ 1; 2; :::; n, which we write as follows: Yni¼ 1 F T i
(t 2 (n 2 1)Tc):
Hence, we have:
n
F SSn ðt Þ ¼ Y F T i ðt 2 ðn 2 1ÞTcÞ ð2Þ
i¼1

Figure 1.
The quasi-renewal process
model with non-negligible
corrective maintenance
duration
The quasi-renewal function Q(t), which corresponds to the mean number of system Optimal
restarts (failure-repair) within the time interval [0, t ] can be obtained as follows:
availability
X
1 X
1
Qðt Þ ¼ E ½ N ðtÞ
¼ Pr f N ðtÞ $ ng ¼ Pr fSS n # t g
n¼1 n¼1
( ! )
X
1 X
n
399
¼ Pr Ti þ ðn 2 1ÞT c # t
n¼1 i¼1
( )
X
1 X
n X
1
¼ Pr T i # t 2 ðn 2 1ÞT c ¼ F SSn ðtÞ:
n¼1 i¼1 n¼1

Thus, considering F Ti ð:Þ as the probability distribution of the system times to the ith
failure
Ti (i ¼ 1, 2, . . . , n), and considering Tc as the duration of the corrective maintenance
actions following each failure, we express the quasi-renewal function as follows:
8 1
>
> X
< F SSn ðtÞ for t . 0
QðtÞ ¼ n¼1 ð3Þ
>
>
: 0 elsewhere

with F SSn ðtÞ ¼ Yni¼1 F T i ðt 2 ðn 2 1ÞT c Þ for i ¼ 1; 2; :::; n.

And as shown by Wang and Pham (1996, 1997):


1
F T i ðt Þ ¼ F T 1 i21 t for i ¼ 1; 2; :::; n and 0 , a , 1: ð4Þ
a
Note that putting T c ¼ 0 in the above development, one finds back the results obtained
by Wang and Pham (1996, 1997) for a quasi-renewal process with negligible repair time.

2.2 The system stationary availability model

As shown in Figure 2, since preventive maintenance action is imperfect, not every
preventive maintenance action is a renewal point. Only those instants, when preventive
maintenance actions restore the system back to new, are the renewal points. This implies
that there can be a random number of preventive maintenance actions in a renewal cycle.

Figure 2.
The renewal cycles
according to the proposed
maintenance strategy
JQME Note that a local clock is considered (t ¼ 0 corresponds to a renewal point).
Figure 3 illustrates one renewal cycle showing corrective maintenance actions in
16,4 one sub-cycle which follows a decreasing quasi-renewal process.
The preventive maintenance period T is the decision variable. Let SA(T) stand for
the system stationary availability. Its expression is developed in what follows based on
the renewal reward theory. Let’s consider DðTÞ as the average duration of a renewal
400 cycle (period between successive perfect preventive maintenance actions), and I ðTÞ as
the average period during which the system is unavailable (submitted to maintenance
actions) during a renewal cycle.
Based on the classical renewal reward theory (Barlow and Proschan, 1965), the
system stationary availability, SA(T), can be expressed as follows:
I ðT Þ
SAðT Þ ¼ 1 2 ð5Þ
DðT Þ
with:
X
1  
D ðT Þ ¼ pq ði21Þ i T þ T p ð6Þ
i¼1

X
1  
I ðT Þ ¼ pq ði21Þ iT p þ T c QðiT Þ ð7Þ
i¼1

Figure 3.
A renewal cycle according
to the proposed strategy
where Q(T) stands for the quasi-renewal function given by equation (3) and which can Optimal
be determined knowing the system time to first failure distribution F T 1 ð:Þ, the repairs
duration Tc, and the parameter a. Statistical estimation of parameter a is discussed in
availability
Wang and Pham (2006). P
Combining equations (5) to (7), and considering that 1
i¼1 q
ði21Þ
i ¼ ð1=1 2 qÞ2 ¼ 1=p 2 ,
we have:
X
1
401
2 ði21Þ
Tp þ p Tc q QðiT Þ
i¼1
1 2 SAðT Þ ¼   :
T þ Tp
Hence, for any given set of known parameters Tp, Tc, a, and F T 1 , the system stationary
availability can be expressed as follows:
8 !
>
> X1
>
> T p þp 2 T c QðTÞþ q ð i21 Þ
QðiT Þ
<
i¼2
SAðT Þ ¼ 1 2 ðTþT p Þ for T $ T p ð8Þ
>
>
>
>
: SAðT Þ ¼ 0 for T # T p

Theorem. The considered maintenance policy admits, for any given set of input
parameters, a finite optimal period T * of preventive maintenance which maximises the
system stationary availability. The optimal availability is given by:
" #
    X 1  
* 2 * ði21Þ *
SA T ¼ 1 2 T c p m T þ q im iT
i¼2

where m(t) stands for the quasi-renewal density (or quasi-renewal rate): mðtÞ ¼
dQðtÞ=dt
If the quasi-renewal density is strictly increasing with time, then T * is unique.
Proof. Note that SA(T) is a continuous function for 0 , T , 1 since Q(T) is
continuous-given that the system time to first failure distribution F T 1 is assumed to be
continuous.
It is easy to see from equation (8) that:
SAð0Þ ¼ 0:
On the other hand, we show in Appendix 1 that: lim SAðTÞ ¼ 0.
T!1
Consequently, we can conclude that there exists a finite period T * which maximises
SA(T) for any given set of parameters: F T 1 , a, Tp and Tc.
The expression of SA(T *) is developed in Appendix 2 as well as the proof of the
sufficient condition for the uniqueness of T * (increasing quasi-renewal rate).
Curve 1 in Figure 4 shows the general behaviour of the system stationary
availability according to the preventive maintenance period T. Curves 2 and 3, in the
same figure, show both possible behaviours of the system stationary availability in the
case of perfect corrective and preventive maintenance actions (the block replacement
policy (Barlow and Proschan, 1965)). It is interesting to notice that if an infinite optimal
JQME
16,4

402
Figure 4.
General behaviours of
system stationary
availability function
SA(T) in the situation of
perfect maintenance
(curves 2 and 3) and
imperfect maintenance
(curve 1)

solution (no corrective maintenance – curve 2) is possible in the case of perfect

maintenance, it cannot be envisaged in an imperfect maintenance situation.
Most of imperfect maintenance policies based on the quasi-renewal process,
including the one considered in this paper, are modelled using the quasi-renewal
function Q(.). It is not possible to obtain this function expressed by equation (3) in
closed form. One approximation in the case of the Normal distribution is given in
Wang and Pham (1996) and Rehmert (2000) provides numerical approximations of
the quasi-renewal function in the cases of exponential and Gamma distributions.
These approximations are very difficult to obtain, they require the inversion of a
truncated infinite sum in the Laplace transform space. In next section, we develop a
new general numerical algorithm to compute the quasi-renewal function for any
quasi-renewal process with systems whose times to first failure follow any given
distribution.

3. Computational procedure of the quasi-renewal function

We propose a numerical algorithm which is a generalisation of the algorithm
developed in Chelbi and Ait-Kadi (2000) designed to compute the n-fold convolutions of
a given distribution function to evaluate the probabilities of shock occurrence within
the classical renewal theory framework. The new algorithm computes the convolution
products used in equation (2) to express the quasi-renewal function:

n
F SSn ðtÞ ¼ Y F T i ðt 2 ðn 2 1ÞT c Þ for i ¼ 1; 2; :::; n: ð2Þ
i¼1

Recall that T1 is the random variable associated to the time to first failure of the system
distributed according to any probability distribution F T 1 ð:Þ.
The expression of the quasi-renewal function, proposed in equation (3), can be
developed as follows:
 QðtÞ ¼ F 1 ðt Þ þ F 1 ðt 2 T c Þ*F 2 ðt 2 T c Þ
|ffl{zffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Optimal
FSS 1 ðt Þ FSS 2 ðt Þ availability
þ F 1 ðt 2 2T c Þ*F 2 ðt 2 2T c Þ*F 3 ðt 2 2T ÞÞ þ ::: ð9Þ
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
FSS3 ðt Þ

where * represents the convolution product operator.

403
The proposed algorithm recursively generates continuous numerical
approximations to each member of the family of functions F SSn ð:Þ derived from a
distribution function F T 1 ð:Þ and its associated density function f T 1 ð:Þaccording to the
following recursive scheme:
8
< F SS1 ðtÞ ¼ F T 1 ðtÞ
Rt ð10Þ
: F SSnþ1 ðtÞ ¼ 0 F SSn ðt 2 xÞf T nþ1 ðxÞdx for n $ 1

with f T n ðtÞ ¼ ð1=a n21 Þf T 1 ðt=a n21 Þ for n $ 1 (Wang and Pham, 1996, 1997)
In order to evaluate F SSn ðtÞ, we use the common numerical representation of a
function as a set of m þ 1 tabulated values of the independent variable at known
equally spaced intervals of length h, with a cubic spline interpolation to supply
immediate values. The choice of a cubic spline interpolation is motivated by the fact
that it preserves positivity and monotonicity (Cleroux and McConalogue, 1976,
Schoenberg and Shisha, 1967), which is essential when approximating distribution
functions.
The calculation of the tabulated values for F SS1 ðtÞ is based on the analytical
identities:
8
< F SS1 ð0Þ ¼ 0
R jhþh ð11Þ
: F SS1 ð jh þ hÞ ¼ F SS1 ð jhÞ þ jh f T 1 ðxÞdx for j ¼ 0; 1; :::; m 2 1

which can be expressed as:

k21 Z
X jhþh
F SS1 ðkhÞ ¼ f T 1 ðxÞdx for k ¼ 1; 2; :::; m ð12Þ
j¼0 jh

We use the Gauss-Legendre quadrature formula to evaluate these partial range

integrals. Equation (12) is used to calculate F SSnþ1 ðtÞ where the integrand is the product
of f T nþ1 and the cubic spline approximation to F SSn ðtÞ.
8
>
> F SS nþ1 ð0Þ ¼ 0 for n $ 0
>
<
k21 Z jhþh
X
>
> F SS ð kh Þ ¼ F SSn ðxÞf T nþ1 ðkh 2 xÞdx for k ¼ 1; 2; :::; m
>
:
nþ1
j¼0 jh
JQME In this formulation, F SSn ðtÞ is defined at points jh and (jh þ h). The intermediate
values required for the Gauss-Legendre formula are supplied by the cubic spline
16,4 interpolation.
Hence, in order to evaluate the quasi-renewal function (equation (9)), one has to
evaluate F SSn ðtÞ for n ¼ 1, 2, 3, . . .
F SS1 ðtÞ ¼ F 1 ðtÞ
404
F SS2 ðt Þ ¼ F 1 ðt 2 T ÞÞ*F 2 ðt 2 T c Þ
where F 1 ðt 2 T c Þ is given by:
8
> F 1 ðt 2 T c Þ ¼ 0 for 0 # t # T c
>
<
X
k21 Z jhþh
>
> F 1 ðkhÞ ¼ f 1 ðx 2 T c Þ dx for k ¼ 1; 2; :::; m and for t . T c
: j¼0 jh

Hence:
8
>
> F SS2 ðt 2 T c Þ ¼ 0 for 0 # t # T c
>
<
k21 Z
X jhþh
>
> F SS2 ðkhÞ ¼ F 1 ðx 2 T c Þ f 2 ðkh 2 x 2 T c Þ dx for k ¼ 1; 2; :::; m and for t . T c
>
: j¼0 jh

The same procedure used above to evaluate F SS2 ðtÞ for n ¼ 1 will be used to evaluate
the successive terms F SSnþ1 ðtÞ for n ¼ 2; ::: determining successively: F 1 ðt 2 nT c Þ,
F 2 ðt 2 nT c Þ, . . . , F n ðt 2 nT c Þ.
8
> F 1 ðt 2 nT c Þ ¼ 0 for 0 # t # nT c
>
<
X
k21 Z jhþh
>
> F 1 ðkhÞ ¼ f 1 ðx 2 nT ÞÞdx for k ¼ 1; 2; :::; m and for t . nT c
: jh
j¼0

8
> F 2 ðt 2 nT c Þ ¼ 0 for 0 # t # nT c
>
< k21 Z
X jhþh
>
> F 2 ðkhÞ ¼ F 1 ðx 2 nT c Þf 2 ðkh 2 x 2 nT c Þdx for k ¼ 1; 2; :::; m and for t . nT c
: j¼0 jh

...
8
> F n ðt 2 nT c Þ ¼ 0 for 0 # t # nT c
>
< k21 Z
X jhþh
>
> F n ðkhÞ ¼ F ðn21Þ ðx 2 nT Þf n ðkh 2 x 2 nT c Þdx for k ¼ 1; 2; :::; m and for t . nT c
: j¼0 jh
Thus: Optimal
8
> F SSnþ1 ðt 2 nT c Þ ¼ 0 for 0 # t # nT c availability
>
>
<
k21 Z jhþh
X
>
> F SSnþ1 ðkhÞ ¼ F n ðx 2 nT c Þf ðnþ1Þ ðkh 2 x 2 nT c Þdx for k ¼ 1; 2; :::; m and for t . nT c
>
: j¼0 jh
405

4. Numerical examples
For illustration purpose of the considered maintenance policy, let’s consider randomly
failing systems whose times to first failure follow a Gamma distribution with shape
parameter a ¼ 3 and scale parameter b ¼ 3.
The following durations are considered:
.
Duration of the corrective maintenance actions: T c ¼ 0:35 time unit.
.
Duration of the preventive maintenance actions: T p ¼ 0:15time unit.
The corrective maintenance actions are characterised by a constant efficiency factor
a ¼ 0:8
For the considered systems and for different values p of the probability to have
perfect preventive maintenance actions, Table I shows the optimal preventive
maintenance periods T * which maximise the system availability SA(T).
Figure 5 shows the evolution of the stationary availability SA(T).
Notice that, as it should be expected, as preventive maintenance actions get more
efficient (increasing p), the optimal values of T * get larger (PM actions should be
performed less frequently). Moreover, efficient PM actions allow more flexibility when
implementing the optimal policy in practice. Indeed, as it is clearly shown by the
availability curves, a given deviation from the optimal solution (in terms of PM period
T) yields significantly less availability loss as PM actions get more efficient.
Considering the same input parameters, and the system with the Gamma time to
failure distribution for illustration purpose, Figure 6 shows that for a given efficiency
degree p of preventive maintenance, as the corrective maintenance actions become
more and more efficient (increasing a), the optimal preventive maintenance period T *
becomes larger and the corresponding system availability SA(T *) increases. It is also
interesting to notice that beyond a certain level of the repair efficiency factor a, the
optimal availability variations tend to stabilise. This allows concluding that it is
possible to tolerate a certain degree of ineffectiveness of the corrective actions without
having a notable availability loss. This flexibility increases as preventive maintenance

p T* SA(T *)
Table I.
1 8.6 0.957 Optimal policies for
0.9 8.1 0.954 different preventive
0.8 7.3 0.951 maintenance efficiency
0.7 6.5 0.947 extents: the Gamma
0.6 5.6 0.942 distribution case
JQME
16,4

406

Figure 5.
Evolution of the system
stationary availability
SA(T) for different
preventive maintenance
factors p: the Gamma
distribution case

Figure 6.
T * and SA(T *) evolution
according to the repairs
efficiency factor, with
T c ¼ 0:35 and Tp ¼ 0:15:
case of the Gamma
distribution

Figure 7.
T * and SA(T *) evolution
according to the corrective
repair efficiency factor, for
various ratios Tc/Tp and a
fixed preventive
maintenance factor
p ¼ 0:7: case of the
Gamma distribution
gets more efficient (increasing p). This figure clearly shows how preventive and Optimal
corrective maintenance efficiency extents affect simultaneously the system optimal availability
availability.
For the same system and for p ¼ 0:7, Figure 7 displays optimal strategies in terms
of the repairs efficiency factor a for different ratios Tc/Tp. We notice that for any given
repair efficiency degree a, as the ratio Tc/Tp increases, the optimal period of preventive
maintenance T * and the corresponding optimal stationary availability decrease. 407

5. Conclusion
In this work, we studied a preventive maintenance policy for randomly failing systems
evolving in a context where preventive and corrective maintenance actions are
imperfect. The imperfection of repair actions has been modelled by a decreasing
quasi-renewal process based on a deterministic repair efficiency factor a. Preventive
maintenance actions follow a ( p, q) rule. We modelled the system stationary
availability and studied its evolution under the proposed maintenance policy. It has
been proven that for any given situation regarding the system, the repair and
preventive maintenance efficiency extents, and the downtime durations for preventive
and corrective maintenance, there is necessarily a finite optimal period T * of
preventive maintenance which maximises the system stationary availability. A
sufficient condition for the uniqueness of the optimal solution has been derived. This
condition is that the quasi-renewal density (quasi-renewal rate) should be increasing
with time.
The expression of the system stationary availability uses the quasi-renewal
function that is impossible to express in a closed form. We proposed a generalised
formulation of the quasi-renewal process considering non-negligible repair durations,
and a numerical algorithm allowing computing the quasi-renewal function for any
quasi-renewal process with systems whose times to first failure follow any given
distribution. The quasi-renewal function being a key quantity for a great number of
imperfect maintenance models based on the quasi-renewal process, the developed
computational procedure can be very useful for such models.
Obtained numerical results, corresponding to a set of particular cases, illustrated
how preventive and corrective maintenance efficiency extents affect simultaneously
the system optimal availability. It allowed concluding that as preventive maintenance
gets more efficient, it is possible to tolerate a certain degree of ineffectiveness of the
corrective actions without having an important availability loss. The knowledge of the
interval of tolerance of the repairs efficiency degree not causing significant availability
loss represents an important source of flexibility for the maintenance management.
Several extensions of this work considering more complex imperfect maintenance
policies are currently under consideration.

References
Barlow, R. and Proschan, F. (1965), Mathematical Theory of Reliability, John Wiley & Sons,
New York, NY.
Chan, J.K. and Shaw, L. (1993), “Modeling repairable systems with failure rates that depend on
age and maintenance”, IEEE Transactions on Reliability, Vol. 42 No. 4, pp. 566-71.
JQME Chelbi, A. and Ait-Kadi, D. (2000), “Generalized inspection strategy for randomly failing systems
subjected to random shocks”, International Journal of Production Economics, Vol. 64
16,4 Nos 1-3, pp. 379-84.
Cleroux, R. and McConalogue, D.J. (1976), “A numerical algorithm for recursively-defined
convolution integrals involving distribution function”, Journal of Management Science,
Vol. 22 No. 10, pp. 1138-46.
408 Kijima, M. (1989), “Some results for repairable systems with general repair”, Journal of Applied
Probability, Vol. 26, pp. 89-102.
Kijima, M., Morimura, H. and Suzuki, Y. (1988), “Periodical replacement problem without
assuming minimal repair”, European Journal of Operational Research, Vol. 37 No. 2,
pp. 194-203.
Leung, K.N.F and Fong, C.Y. (2000), “A repair-replacement study for gearboxes using geometric
process”, International Journal of Quality & Reliability Management, Vol. 17 No. 3,
pp. 285-304.
Malik, M.A.K. (1979), “Reliable preventive maintenance policies”, AIIE Transactions, Vol. 11
No. 3, pp. 221-8.
Pham, H. and Wang, H. (1996), “Imperfect maintenance”, European Journal of Operational
Research, Vol. 94 No. 3, pp. 425-38.
Rehmert, I.-J. (2000), “Availability analysis for the quasi-renewal process”, PhD thesis in
Industrial and Systems Engineering, Virginia Polytechnic Institute and State University,
Blacksburg, VA, October.
Samath-Paç, G. and Taner, M.R. (2009), “The role of repair strategy in warranty cost
minimization: an investigation via quasi-renewal process”, European Journal of
Operational Research, Vol. 197 No. 2, pp. 632-41.
Schoenberg, J.J. and Shisha, O. (Eds) (1967), In Inequalities, Academic Press, New York, NY,
pp. 255-91.
Shin, I., Lim, T.J. and Lie, C.H. (1996), “Estimating parameters of intensity function and
maintenance effect for repairable unit”, Reliability Engineering and System Safety, Vol. 54
No. 1, pp. 1-10.
Valdez-Flores, C. and Feldman, R.M. (1989), “A survey of preventive maintenance models for
stochastically deteriorating single-unit systems”, Naval Research Logistics, Vol. 36 No. 4,
pp. 419-46.
Wang, H. (2002), “A survey of maintenance policies of deteriorating systems”, European Journal
of Operational Research, Vol. 139 No. 3, pp. 469-89.
Wang, H. and Pham, H. (1996), “A quasi-renewal process and its applications in imperfect
maintenance”, International Journal of Systems Science, Vol. 27 No. 10, pp. 1055-62.
Wang, H. and Pham, H. (1997), “Correspondence changes to: “A quasi-renewal process and its
applications in imperfect maintenance”, International Journal of Systems Science, Vol. 28
No. 12, pp. 13-29.
Wang, H. and Pham, H. (2006), Reliability and Optimal Maintenance, Springer, New York, NY.
Whitaker, L.P. and Samaniego, F.J. (1989), “Estimating the reliability of systems subject to
imperfect repair”, Journal of American Maintenance, Vol. 84 No. 405, pp. 301-9.

Further reading
Yañez, M., Joglar, F. and Modarres, M. (2002), “Generalized renewal process for analysis of
repairable systems with limited failure experience”, Reliability Engineering and System
Safety, Vol. 77 No. 2, pp. 167-80.
Appendix 1 Optimal
Proof of lim SAðTÞ ¼ 0:
T!1
!
availability
X
1
2 ði21Þ
T p þ p T c QðT Þ þ q QðiT Þ
i¼2
lim SAðTÞ ¼ lim 1 2  
T!1 T!1 T þ Tp
409
X
1
q ði21Þ QðiT Þ
Tp QðT Þ i¼2
lim SAðT Þ ¼ 1 2 lim   2 p 2 T c lim   2 p 2 T c lim  
T!1 T!1 T þ T p T!1 T þ T p T!1 T þ Tp

" #
QðT Þ X1
QðiT Þ
2 ð i21 Þ
lim SAðT Þ ¼ 1 2 T c p lim  þ q lim   : ðA1Þ
T!1 T!1 T þ T p T!1 T þ T p
i¼2

It is easy to see that: lim QðiTÞ=ðT þ T p Þ ¼ lim QðiTÞ=Tð1 þ T p =TÞ ¼ lim QðiTÞ=T.
T!1 T!1 T!1
Hence, from equation (A1) above:
" #
QðT Þ X 1
QðiT Þ
lim SAðT Þ ¼ 1 2 T c p lim 2
þ q ði21Þ lim
T!1 T!1 T T!1 T
i¼2

" #
QðT Þ X
2
1
ði21Þ QðiT Þ
lim SAðT Þ ¼ 1 2 T c p lim þ q i lim :
T!1 T!1 T T!1 iT
i¼2

The expression QðTÞ=T represents the average number of restarts per time unit in the time
interval [0, T ]. We can also consider this term as the inverse of the average period between
successive restarts over the same interval [0, T ]. Thus, according to the quasi-renewal process
defined in section 2.1, and considering Figure 3:
Ti: the system time to the ith failure following a given probability distribution.
Y i: a random variable defined as the sum of Ti and Tc.
We have:
2 3 2 3
6 7 6 7
QðT Þ 6 1 7 6 n 7
lim lim 6
¼ n!1 7 ¼ lim 6 7
T!1 T 6X n 7
Y i5 n!1 6X n 7
4 4 5
ðT i þ T c Þ
i¼1
n i¼1

2 3
6 7
QðT Þ 6 n 7
lim lim 6
¼ n!1 6Xn
7
7
T!1 T
4 5
T i þ nT c
i¼1
 2 3
JQME
6 7
16,4 QðT Þ 6 n 7
lim lim 6
¼ n!1 6Xn
7
7
T!1 T
4 5
a i21 Z i þ nT c
i¼1
Pn i21 n21
410 Note that i¼1 a Z i is finite because n!1
lim a Z n ¼ 0 for 0 , a , 1.
Hence:

QðT Þ 1
lim ¼ :
T!1 T Tc

Let us define j ¼ iT; when T ! 1, iT ! 1; then lim QðjÞ=j ¼ 1=T c

j!1
" #
X 1
2 QðT Þ ði21Þ QðiT Þ
lim SAðT Þ ¼ 1 2 T c p þ q i lim
T!1 T i¼2
T!1 iT

" #
1 X
1
2 ði21Þ 1
lim SAðT Þ ¼ 1 2 T c p þ q i :
T!1 T c i¼2 Tc
P1
Given that i¼1 q
ði21Þ
i ¼ ð1=1 2 qÞ2 ¼ 1=p 2 we obtain:

1
lim SAðT Þ ¼ 1 2 T c p2 ¼ 0; for 0 , a , 1:
T!1 p 2T c

Hence, we have:

lim SAðT Þ ¼ 0:
T!1

Appendix 2
The optimal strategy T * necessarily verifies the following equation:

dSAðTÞ
¼ 0: ðA2Þ
dT jT¼T *

Equations (8) and (A2) lead to:

! !
    X 1     X 1  
T * þ Tp p 2Tc m T * þ q ði21Þ m iT * Tp þ p Tc 2
Q T* þ q ði21Þ Q iT *
i¼2 i¼2
 2 2  2 ¼0
T * þ Tp T * þ Tp

! !
    X 1     X 1   Tp
T * þ Tp p 2 m T * þ q ði21Þ m iT * 2p 2
Q T* þ q ði21Þ Q iT * ¼ ðA3Þ
i¼2 i¼2
Tc
 !
  X 1       X 1   Tp Optimal
Q T* þ q ði21Þ Q iT * ¼ T * þ T p m T * þ q ði21Þ m iT * 2 :
i¼2 i¼2
p 2Tc availability
Substituting in equation (8), we obtain:
" ! #
    X 1
ði21Þ
  Tp
Tp þ Tc p 2 *
T þ Tp m T þ * q m iT * 2 p 2Tc
  411
SA T * ¼ 1 2   i¼2
T * þ Tp

!
    X 1  
T cp 2 T * þ T p m T * þ q ði21Þ m iT *
  Tp
SA T * ¼ 1 2  2  i¼2
T * þ Tp T * þ Tp
 
T c p 2 2 pT2 Tp c
2  
T * þ Tp

!
  Tp   X 1   Tp
SA T * ¼ 1 2   2 Tcp 2 m T * þ q ði21Þ m iT * þ :
*
T þ Tp T * þ Tp
i¼2

Hence, it is proven that the expression of the optimal availability SA(T *) is:
!
  X 1
ði21Þ
 
* 2
SAðT Þ ¼ 1 2 T c p m T þ * q m iT * :
i¼2

To establish the sufficient condition for the uniqueness of the optimal strategy T *, let FðTÞ be
the left-hand side of equation (A3), hence:
!
    X
1  
ði21Þ
FðT Þ ¼ T * þ T p p 2 m T * þ q m iT *
i¼2
!
  X
1  
ði21Þ
2p 2
Q T* þ q Q iT * :
i¼2

It is easy to see that Fð0Þ ¼ 0.

Hence, the ratio T p =T c being positive, if FðTÞ is an increasing function to infinity then there
exists a unique optimal strategy T *.
!
dFðT Þ   2 dmðT Þ X 1
ði21Þ 2 dmðiT Þ
¼ T þ Tp p þ q i
dT dT i¼2
dT
! !
X
1 X
1
2 ði21Þ 2 ði21Þ
þp mðTÞ þ q i mðiTÞ 2p mðT Þ þ q imðiT Þ
i¼2 i¼2
 !
JQME dFðT Þ   2
X1
ði21Þ 2 dmðiT Þ
16,4 ¼ T þ Tp £ p q i :
dT i¼1
dT

Consequently, a sufficient condition to have dFðTÞ=dTstrictly positive is to have dmðTÞ=dT

strictly positive which means to have an increasing quasi-renewal density function.
412
About the authors
Slah Samet received the BS (2002) and the Master’s (2005) degrees in Industrial Engineering from
École Supérieure des Sciences et Techniques de Tunis, Tunisia. Since 2005 he has been preparing
his doctoral thesis in the Centre de Recherche en Productique (CEREP) at the Higher School of
Science and Technology of Tunis (ESSTT).
Anis Chelbi graduated from Laval University (Canada) and Kyoto University ( Japan). His
PhD degree is in Mechanical Engineering (1996). His main expertise and research interests
include reliability and maintainability-based design, manufacturing systems reliability and
availability optimisation, and CMMS design and implementation strategies. He worked as
Assistant Professor in Mechanical Engineering at Laval University from 1996 to 1998. In 1998,
he joined the University of Tunis at the Higher School of Science and Technology of Tunis
(ESSTT) as a professor, where so far he has been teaching and directing a research team. Besides
his scientific contributions and publications, he has a rich experience as a consultant for
industrial companies in various fields like aeronautics, chemical industry, industrial software
development, ceramics, packaging and food industries. He was a recipient of the Outstanding
Paper Award Winner at the Literati Network Awards for Excellence 2009. Anis Chelbi is the
corresponding author and can be contacted at: [email protected]
Fayçal Ben Hmida received the BS in Electrical Engineering from the Ecole Normale
Supérieure de l’Enseignement Technique de Tunis (ENSET), in 1991. In 1992, he obtained the
Master’s degree in Automatics and Informatics at the Aix-Marseille III University, France, and a
PhD in Productique and Informatics at the same university in 1997. In 1998 he joined the Higher
School of Science and Technology of Tunis (ESSTT) at Tunis University as an Assistant
Professor. He is a member of a research unit on control, monitoring and safety of systems at the
Higher School of Science and Technology of Tunis (ESSTT). His main research interests are fault
detection and isolation (FDI), inference statistics, reliability and maintenance.

To purchase reprints of this article please e-mail: [email protected]

Or visit our web site for further details: www.emeraldinsight.com/reprints
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.