Reliability Engineering and System Safety 183 (2019) 143–152

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

A repair and replacement policy for repairable systems based on probability T

and mean of profits
⁎
Fatemeh Safaeia, Jafar Ahmadi ,a, N. Balakrishnanb
a
Department of Statistics, Ferdowsi University of Mashhad, P. O. Box 1159, Mashhad 91775, Iran
b
Department of Mathematics and Statistics, McMaster University, Hamilton, Ontario L8S 4K1, Canada

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

Keywords: In the optimization of maintenance policies, when a repairable system fails, an important question that arises is
Exponential distribution whether the system should be repaired or replaced by a new system. This paper deals with this problem and
Maximal benefit presents a policy to decide between repairing or replacing a system that has failed. Upon taking into account the
Preventive maintenance costs involved, an optimal policy is developed that maximizes the benefit per unit time for repair and re-
Repairable system
placement. It is assumed that preventive maintenance is performed periodically thus increasing the lifetime of
Repair time limit
Symmetric distribution
the system, and as a result delaying its failure. Perfect and imperfect repairs are considered to find the solution.
Numerical computations are given to study the effect of model parameters on the optimal solutions. The pro-
posed plan is used on a real data set containing failure times and repair times.

1. Introduction non-parametric solution for the estimation of the optimal repair time
limit. Love et al. [13] proposed a discrete semi-Markov decision model
Maintenance policies for reliability systems have received con- to determine the optimal repair and replacement policy under general
siderable attention in recent years. Perhaps, the earliest work on op- repairs. Jack et al. [6] proposed a repair and replace strategy based on
timal preventive maintenance policies is due to Barlow and Hunter [2]. usage rate for items sold with a two-dimensional warranty. Their
As mentioned by Garg and Deshmukh [5] and also by Sarkar et al. [18], strategy was based on a specified region of the warranty defined in
several maintenance policies and replacement models have since been terms of age and usage with the first failure in the region rectified by
proposed in the literature. Each maintenance policy has its own char- replacement and all other failures being minimally repaired. Kim and
acteristics, advantages and disadvantages. Some well-known policies Yun [9] considered a repair time limit replacement policy with esti-
are age replacement policy, random age replacement policy, periodic mation error and obtained numerically the optimal repair time limits to
preventive maintenance policy, block replacement policy, failure limit maximize the average profit rate. Whang and Zhang [22] investigated
policy, sequential preventive maintenance policy, repair cost limit optimal repair and replacement policies for a system with two types of
policy, repair time limit policy, repair number counting policy, re- failures. Park et al. [15] proposed an optimal maintenance model which
ference time policy, mixed age policy, preparedness maintenance involves the warranty policy. Their model adopts the repair time
policy, group maintenance policy and opportunistic maintenance threshold along with the warranty period and a certain type of system
policy. In this work, we intend to consider preventive maintenance, maintenance policy during the post-warranty period. Lim et al. [12]
which usually leads to delay the failure time, and on the other hand, proposed age replacement policy based on the imperfect repair with
considering the time limit policy for the decision to repair and replace. random probability and studied the optimal age replacement policy
In fact, we provide a double maintenance policy. that minimizes the expected cost rate per unit time. Zhang and Xie [25]
Koshimae et al. [10] introduced a repair and replacement policy proposed an imperfect-maintenance model and developed two methods
where for a given t0, once the system fails, the repair is started instantly for parameter estimation. Nguyen et al. [14] analyzed the imperfect
and if the repair is completed within time t0 then it is put back into repair model with Weibull failure distribution under the arithmetic
operation and otherwise, a new unit is ordered. Dohi et al. [4] con- reduction of age assumption and implemented an optimal age-based
sidered a generalized repair time limit replacement problem with lead preventive maintenance policy. Park et al. [16] developed a warranted
time and imperfect repair, subject to a time constraint, and proposed a cost model based on the renewable minimal repair-replacement

⁎
Corresponding author.
E-mail addresses: [email protected] (F. Safaei), [email protected] (J. Ahmadi), [email protected] (N. Balakrishnan).

https://doi.org/10.1016/j.ress.2018.11.012
Received 20 October 2017; Received in revised form 28 October 2018; Accepted 7 November 2018
Available online 08 November 2018
0951-8320/ © 2018 Elsevier Ltd. All rights reserved.
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

warranty and the generalized age replacement policy. Zhang and Wang that the random variable Y shows the impact of all preventive main-
[24] considered recently a simple repairable system with delayed repair tenance. When the main failure occurs, the question that is of interest is
and studied the optimal replacement policy based on system age. Zhao “Does the system need to be repaired or it should be replaced by a new
et al. [26] collected recent results and proposed some new models in one?” All aspects including costs and the lifetime of the system need to
age replacement policies. Zhao et al. [27] compared two types of re- be considered while answering this question. In the same vein, what
placement policies, i.e., replacement at some periodic times with that at policies should be considered before making a final decision? These
a pre-fixed number of repairs. Sun et al. [21] studied the optimal in- questions are fundamental in practical situations and we deal with
spection and replacement for condition-based maintenance strategy for these issues here. First, we present some notations that will be used in
a multi-unit system. Sheu et al. [19] proposed age maintenance policies the following sections:
for a system with random working times and considered the expected
cost per unit time as a criterion for determining the optimal policy. Ta: Random variable of repair duration with density function g(.),
In this work, we consider a system such that when the system fails, a distribution function G(.) and parameter t;
decision to repair or replacement needs to be made and then find an X: The lifetime of the original system with density and distribution
optimal repair-time limit policy to maximize the benefit of the system in functions f(.) and F(.), respectively;
a cycle. We perform preventive maintenance before the system failed qj: Degree of imperfect repair for the jth cycle;
and it is assumed that preventive maintenance will result in increasing X˜j = qj − 1 X : The lifetime of the original system for jth cycle;
the lifetime of the system, thus delaying the failure time. For example, Y: Random variable related to preventive maintenance with density
consider an electric generator that will earn revenue per hour when the and distribution functions h(.) and H(.), respectively;
generator is in operation. On the other hand, periodic inspection of Ỹj : Random variable related to preventive maintenance for the jth
each generator should be performed and each period inspection re- cycle;
quires an expert operator, which will incur a cost. The periodic in- dj: Period for preventive maintenance in the jth cycle;
spection will, of course, improve the performance of the generator and Z̃j : The lifetime of system by considering preventive maintenance for
delay its failure. On the other hand, when the generator breaks down the jth cycle;
and fails, the responsible authority must decide whether to repair or k: Penalty cost per unit time when the system is off;
purchase the new system. The repair or purchase of the new system will e0: System income per unit time;
have a cost, and so will the length of time that the generator will be shut e1: Repair cost per unit time;
down. All these will result in a loss to the company. Therefore, the e2: Preventive maintenance cost;
director of the company will be interested in knowing both the best c: The cost of buying a new system;
period of inspection and the best decision after the system has failed. In L: The waiting time for the delivery of a new system;
order to decide on repair or replacement, the mechanic will determine V < stW: It means that the random variable V is stochastically
the cost per hour for the repair, and the length of time the repair will smaller than W, i.e., for all t, P(V > t) ≤ P(W > t).
take. The cost of each unit of work hours for the mechanic and all other
costs, such as the cost of purchasing the new system, are incorporated When a major failure occurs, a choice between repair and replace-
into the problem and the optimal time for repair is then determined. If ment needs to be made. This decision may be made in such a way that
the repair time is less than the optimal repair time, repairing the system maximum profit is achieved. In this work, we consider imperfect repair
would be cost-effective; otherwise, a new system should be ordered; and perfect repair is a special case for this type of imperfect repair. The
see, for example, Yung and Bonnett [23]. We will provide two criteria time interval between two starts of the system is defined as a cycle. We
and an algorithm for optimizing decision making on the length of suppose that preventive maintenance was performed at intervals of
preventive maintenance and deciding whether to repair or replace. We time length d, before the occurrence of the major failure. Here, we then
will consider imperfect repair and also perfect repair to solve the pro- determine the optimal d and t in such a way that the system profit is
posed model. It is notable that all of the previous policies are based on maximized. In what follows the perfect repair means that once the
expectation, but here we will also use a measure of probability. perfect repair is done, the system will work as well as in the original
The rest of the paper is organized as follows. In Section 2, we pre- state. But if an imperfect repair is carried out, then it is logical to as-
sent the model description. We obtain an optimal policy that maximizes sume that the lifetime of the system will decrease. Suppose the lifetime
the profit per unit time for the repair and replacement in Section 3. In of the original system is X, then after the first imperfect repair, it
Section 4, some special cases are considered. In Section 5, some nu- changes to X̃ so that X˜ <st X . Thus we can write X˜ = qX , where
merical computations are presented to illustrate the obtained results. 0 < q ≤ 1 and can be considered as the degree of imperfect repair. For
Moreover, the proposed plan is used on a real data set containing failure q = 1, we have perfect repair. This strategy is similar to Kijima model I,
times and repair times. Some concluding remarks are presented in see Kijima et al. [8] and Kijima [7] in which the authors studied it in
Section 6. terms of the age of the system. In this proposed model q can be a
constant value or a random variable between 0 and 1, but in a real
2. Model description situation, it should be determined by an expert and in consultation the
mechanic performing the repair. Therefore for the jth cycle, we have
In many practical problems, when an operating system stops X˜j = qj − 1 X , where 0 < qj ≤ 1 for j = 1, 2, 3, ⋯ and q0 = 1.
working due to failure, we should decide to repair or replace the system
based on the nature of the problem and its current situation. On the 3. Main results
other hand, before the failure, some inspections could be done to im-
Y˜j
prove the system performance. Here, it is assumed that performing For the jth cycle, let Z˜j = X˜j + be the lifetime of system by con-
dj
preventive maintenance before the occurrence of a failure will increase
sidering preventive maintenance, where Ỹj is the random variable re-
the lifetime of the system and consequently prevent more severe fail-
lated to preventive maintenance and dj is the period for preventive
ures. Usually, preventive maintenance is less expensive than repair or
maintenance. In fact, it is assumed that for all preventive maintenance,
replacement. In this work, we assume that the preventive maintenance Y˜j
is performed periodically during the time period of d and it results in the lifetime increases to X̃j + dj
; see Fig. 1. For a more elaborate ex-
the failure time being delayed, that is, the lifetime of the original X planation, let X̃j be the lifetime of the original system at the time of start
increases to Z that depends on X, d and Y, i.e., Z = ν (X , Y , d ), where Y of the jth cycle and after the first preventive maintenance, it increases
is a non-negative random variable independent of X. It can be assumed to X1 = X˜j + Y1, and after the second preventive maintenance, it

144
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 1. The proposed model for inspection and decision for jth cycle.

increases to X2 = X1 + Y2 = X˜j + Y1 + Y2, and so on, and after the ith If the decision is taken to repair the system, then from (1) and (3), the
i
preventive maintenance, it increases to Xi = Xi − 1 + Yi = X˜j + ∑r = 1 Yr , profit per unit of time for the jth cycle is given by
i Y˜j
Y˜j
where X0 = X̃j . We suppose ∑r = 1 Yr = dj
; it is clear that the number of
preventive maintenances (i) depends on time between consecutive TPrepair (t , d ) ≡
j j
(e − ) (X˜ + ) − (e + k) T .
0
e2
dj j dj 1 a

Y˜j
preventive maintenances (dj). Actually, for the jth cycle it is equal to X˜j + + Ta
Y˜ j dj (5)
X˜j +
dj
[ dj
], where [t] denotes the greatest integer less than or equal to t. It Similarly, if the decision is to replace the system, from (2) and (4), the
is further clear that small dj will cause more preventive maintenances. profit per unit of time for the jth cycle is given by
By assumptions stated in the preceding section, for the jth cycle, the Y˜j
decision should be made as follows: (e 0 −
e2
d ) (X˜ +
j dj ) − (kL + c) .
TPreplace (dj ) ≡
Y˜j
(i) When a repair decision is taken, the system must be sent for repair X˜j + +L
dj (6)
and a random time Ta is taken to do the repair and then return the
system. So, in this case, the duration of the system for jth cycle is Now, the optimal threshold value tj and the optimal dj, say tj, opt and dj,
opt, respectively, need to be determined. The optimal threshold value tj
Y˜j is compared with the required repair time, say tR. If tR is less than or
X˜j + + Ta;
dj (1) equal to tj, opt, the decision is made to repair, and otherwise, the system
should be replaced by a new one.
According to (5) and (6), the probability that the repair profit is
(ii) If we decide to replace the system, then a purchase order needs to more than the replacement profit can be expressed as
be done and then it takes a time of L (constant) for a new system to
PTP (t j, dj ) ≡ P {TPrepair (t j, dj ) > TPreplace (dj )}. (7)
be delivered. So, in this case, the duration of the system for the jth
cycle is The probability in (7) is greater than 0.5 means that the repair profit is
Y˜j greater than the replacement profit more often. So, it seems logical that
X˜j + + L. the optimal tj should be such that PTP(tj, dj) > 0.5. The expression PTP(tj,
dj (2) dj) > 0.5 is equivalent to saying that in more than 50 percent of the
cases, the repair profit is greater than the replacement profit, so that
logically we have to decide to repair the system. So, any value of PTP(tj,
For more illustration, the following graphs have been drawn up in dj) greater than 0.5 indicates that the repair is more profitable. For
which a decision is made to repair or replace the system. In Fig. 1, PM example, if PTP (t , d ) = 0.7, a repair against replace is more worth rather
stands for preventive maintenance. than PTP (t j, dj ) = 0.6. It is noteworthy that criterion P{TPrepair(tj,
From (1), the profit for the jth cycle, when the decision is taken to dj) > TPreplace(dj)} > 0.5 represents an ordering known as probability
repair, can be expressed as order (A random variable V is said to be greater than W in probability
order if P(V > W) ≥ P(W > V), i.e., P(V > W) ≥ 0.5). On the other
Y˜j X˜j Y˜j hand, to choose an optimal dj, it is logical that the quantity
e0 ⎛⎜X˜j + ⎟⎞ − e2 ⎡
⎢ + 2⎤ ⎥ − (e1 + k ) Ta.
⎝ d j⎠ d
⎣ j d j ⎦ (3) ETP (t j, dj ) ≡ P {TPrepair (t j, dj ) > TPreplace (d )} E (TPrepair (t j, dj ))
The first term in (3) is the income of the system in a cycle, the ex- + (1 − P {TPrepair (t j, dj ) > TPreplace (dj )}) E (TPreplace (dj ))
X˜j Y˜j = PTP (t j, dj ) E (TPrepair (t j, dj ))
pression ⎡ d + d2 ⎤ in the second term is the number of preventive
⎣ j j⎦
+ (1 − PTP (t j, dj )) E (TPreplace (dj ))
maintenances in a cycle so that the second term is the cost of preventive
maintenances in a cycle; and the last term in (3) is the repair and (8)
penalty cost when the system is being repaired. Hence, (3) represents be maximized. From (8), it is clear that ETP(tj, dj) is a weighted ex-
the profit of a given cycle. Similarly, in the case of a decision of re- pectation. Therefore, by (7) and (8), when the underlying distributions
placement, from (2), profit for the jth cycle is given by and the parameters of the problem are specified, the expressions for
PTP(tj, dj) and ETP(tj, dj) get simplified as a function of only tj and dj. We
Y˜j X˜j Y˜j
e0 ⎛⎜X˜j + ⎟⎞ − e2 ⎡
⎢ + 2⎤ − (kL + c ). remind that the goal of the problem is to find the optimal values of tj
⎝ dj ⎠ ⎣ dj dj ⎥⎦ (4) and dj, say tj, opt and dj, opt. With this in mind, we propose the following

145
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

algorithm: (12), we get

Algorithm 1. Once PTP(tj, dj) and ETP(tj, dj) in (7) and (8) are specified, λ1, j λ2, j dj ⎧ ∞ −h2 z ⎞ −λ1, j z
for the jth cycle the following steps should be adopted: PTP (t j, dj ) =
λ2, j dj − λ1, j ⎨
∫0 FTa ⎛⎜ e ⎟ dz
⎩ ⎝ z + h3 ⎠
h1

∞ −h2 z ⎞ −λ2, j dj z ⎫
(i) Consider the inequality PTP(tj, dj) > 0.5. In this inequality, topt(dj) is − ∫0 FTa ⎛ ⎜ e dz .
⎟

⎬
calculated in terms of dj. Since PTP(tj, dj) is decreasing in tj, find the ⎝ h1 z + h3 ⎠ ⎭ (13)
greatest tj such that PTP(tj, dj) > 0.5, say tj, opt(dj), i.e.,
Usually, the repair time tends to be symmetric around a time specified
t j, opt (dj ) = sup{t j: PTP (t j, dj ) > 0.5} ;
to complete the repair and so symmetric distributions are suitable for
(ii) Substitute topt(dj) in ETP(tj, dj), then ETP(topt(dj), dj) is a function of
Ta. It should be mentioned that Kim and Yun [9] also supposed that the
only dj;
repair time follows a normal distribution. With this in mind, in the next
(iii) Maximize ETP(topt(dj), dj) and find optimal d, say dj, opt;
two subsections, we suppose Ta has uniform and normal distributions.
(v) Substitute dj, opt in tj, opt(dj), i.e., t j, opt = t j, opt (dj, opt ) . Finally, we
obtain (tj, opt, dj, opt).
4.1. Ta has uniform distribution
The operator has to inspect the system and perform preventive
maintenance in lengths of dj, opt. When the system failed, tj, opt is com- In reality, there are some systems whose repair times will not be
pared with the required repair time, say tR. If tR is less than or equal to tj, greater or less than some positive real values. Hence, in this case, it is
opt, the decision is made to repair, and otherwise, the system should be suitable to consider a finite support for repair time. Suppose for the jth
replaced by a new one. It should be noted that the costs have to be set cycle, Ta has a uniform distribution over (t j − θ , t j + θ), where θ > 0
by the experts such as maintenance staff and if the cost of system re- and tj > θ. Then, we have
placement is expensive, then the proposed policy will decide on the
−h2 z
optimal decision to repair the system. ⎧ 0, h1 z + h3
< t j − θ,
⎪
⎪
From the above algorithm, it is clear that the solution depends on − h z
the distributions of X̃j , Ỹj, Ta and also on some other parameters. We
P ⎛Ta <
⎜

⎝ h1 z +
2
h
⎞= 1
3⎠ ⎨ 2θ
⎟
( −h2 z
h1 z + h3
+ θ − tj ,) tj − θ <
−h2 z
h1 z + h3
< t j + θ,
shall now discuss some special cases in detail to illustrate the results ⎪ −h2 z
⎪1, h1 z + h3
> t j + θ.
established here. ⎩
(14)
4. Special cases From (14), the expression in (13) can be written as

1 ⎛ −h2 z + θ − t j ⎞ f (z ) dz
b (t j, dj )
In this section, we suppose for the jth cycle the distribution of Ỹj has PTP (t j, dj ) = 1 +
2θ
∫a (t ,d ) ⎜
Z ⎟

proportional relation with the distribution of X̃j (the lifetime of the ⎝ h1 z + h3

j j
⎠
original system for the jth cycle), i.e., H¯ j (t ) = (F¯j (t )) η , where η > 0 is − FZ˜j (b (t j, dj ))
the proportionality parameter and j = 1, 2, …. In the reliability litera- λ1, j λ2, j dj
ture, it is known as proportional hazard rate model; see for example, = 1+ (J1 (λ1, j ) − J1 (λ2, j dj )) − FZ˜j (b (t j, dj ))
2θ (λ2, j dj − λ1, j )
Kumar and Klefsjö [11]. To do this, we suppose that X has an ex-
ponential distribution with mean 1 and Y has an exponential dis- θ − tj
λ1 +
2θ
{FZ˜j (b (t j, dj)) − FZ˜j (a (t j, dj)) },
1
tribution with mean . So X˜j = qj − 1 X has exponential distribution with
ηλ1 (15)
qj − 1 qj
mean and Ỹj has exponential distribution with mean . Let us now
λ1 ηλ1 where
ηλ1
denote λ1 by λ1, j and by λ2, j. Also, suppose Ta is independent of X̃j
qj − 1 qj − 1 (t j − θ) h3 −(t j + θ) h3
and Ỹj, c < e1L and that a (t j, dj ) = and b (t j, dj ) = ;
(t j − θ) h1 − h2 (t j + θ) h1 + h2
c
dj > e 0 + k + . also,
L (9)
Y˜j
⎛ −h2 z ⎞ e−uzdz.
b (t j, dj )
Upon taking Z˜j = X˜j + dj
, it is easy to show that the density function of J1 (u) = ∫a (t ,d ) ⎜ ⎟

Z̃j is given by
j j ⎝ h1 z + h3 ⎠

λ1, j λ2, j dj By using the exponential integral function [see [1],(p. 228)], J1(u) can
fZ˜j (z ) =
λ2, j dj − λ1, j
(e−λ1,j z − e−λ2,j dj z ), z > 0, λ1, j ≠ λ2, j dj . be expressed in terms of Ei function as
(10)
h2 h3 ⎧ ⎛ ⎛ h h
Upon substituting (5) and (6) into (7), we get J1 (u) = Ei ⎜u a (t j, dj ) + 3 ⎞ ⎟⎞ − Ei ⎜⎛u ⎛b (t j, dj ) + 3 ⎞ ⎟⎞ ⎫ exp
⎜ ⎟ ⎜ ⎟

h12 ⎨⎩ ⎝ ⎝ h1 ⎠ ⎠ ⎝ ⎝ h1 ⎠ ⎠ ⎬
⎭
PTP (t j, dj ) = P (h1 Z˜j Ta + h2 Z˜j + h3 Ta > 0), (11)
⎛ h3 u ⎞
⎜ ⎟

where ⎝ h1 ⎠
e e2 L
h1 = d2 − e1 − k − e0, h2 = e0 L − + kL + c and h3 = c − e1 L. h
j dj + 2 (e au − e bu )exp(−(a + b) u),
From (9) and the condition c < e1L, it is readily deduced that h1 and h1 u
h3 are negative and h2 is positive. Using conditional argument, we can (16)
then write (11) as
where Ei(.) is given by
−h2 Z˜j ⎞ ∞ e−y
PTP (t j, dj ) = P ⎛⎜Ta < ⎟ Ei (x ) = ∫x dy.
⎝ h1 Z˜j + h3 ⎠ y
∞ ⎛ −h2 Z˜j ⎞ Upon substituting (16) in (15), we obtain an exact expression for
= ∫0 P ⎜Ta <
h1 Z˜j + h3
Z˜j = z ⎟ fZ˜j (z ) dz.
PTP(tj, dj) with which, for given values of parameters, the exact value of
⎝ ⎠ (12)
PTP(tj, dj) can be computed. We have done this for some special cases
From the assumption that Z̃j is independent of Ta and using (10) and and the results are presented in the next section.

146
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 2. Ta has the uniform distribution with θ = 1.

To use Algorithm 1 for finding the optimal choice of tj and dj, the λ1, j λ2, j dj
PTP (t j, dj ) = ∫0
−h2 / h1 ⎛ 1 ⎛
ϕ (q − t j ) ⎜ ⎜1 − exp
⎧ −λ1, j h3 σq ⎫ ⎟⎞
required expectations need to be calculated first. The details of calcu- λ2, j dj − λ1, j λ1, j ⎝ ⎨ h2 + h1 σq ⎬
⎝ ⎩ ⎭⎠
lations for obtaining of E(TPrepair, j) for the jth cycle are given in Ap-
pendix A. Upon (A.1) for E(TPrepair, j) we have
−
1 ⎛
⎜1 − exp
⎧ −λ2, j dj h3 σq ⎫ ⎟⎞ ⎞ dq.
⎨ ⎟
(λ1, j λ2, j dj )(e1 + k ) ⎛ 1
λ2, j dj ⎝ ⎩ h2 + h1 σq ⎬ ⎭⎠⎠
1 ⎞
E (TPrepair , j ) = C1 {J2 (λ1, j ) − J2 (λ2, j dj )} − ⎜ − ⎟,
(19)
λ2, j dj − λ1, j ⎝ λ1, j λ2, j dj ⎠
From (19), for given parameters, the value of PTP(tj, dj) can be com-
(17)
puted. We have obtained the numerical values of PTP(tj, dj) in (19) for
λ1, j λ2, j dj (e0 + e1 + k − e2 / dj )
where C1 = 2θ (λ2, j dj − λ1, j )
, and some cases and these results are given in the next section.
From (6), the distribution of TPreplace is seen not to depend on the
1 ⎛ u (t j − θ) distribution of Ta, and so E(TPreplace) for the normal case is the same as
J2 (u) = e (u (t j − θ) − 1) Ei (u (t j − θ))
u2 ⎜ for the uniform distribution given in (18).
⎝
tj + θ ⎞⎞
− eu (t j + θ) (u (t j + θ) − 1) Ei (u (t j + θ)) + ln ⎜⎛ ⎟ .
⎟ 5. Numerical computations
⎝ tj − θ ⎠⎠
In order to illustrate the obtained results, first let us take j = 1, i.e.,
Using (10) and (14), the expectations of TPreplace is given by 1 1
q0 = 1 and λ = 10 and λ = 2, and also L = 10, e0 = 50, e1 = 10,
1,1 2,1

E (TPreplace, j ) =
λ1, j λ2, j dj
∫0
+∞ (e − ) z − (kL + c) (e
0
e2
dj
−λ1, j z
e2 = 2, k = 2 and c = 20 . By using the Algorithm 1, we then found the
optimal values of t1 and d1, denoted by topt and dopt, respectively. It
λ2, j dj − λ1, j z+L should be mentioned that we used “DEoptim” in the R Package to ob-
− e−λ2, j dj z ) dz tain these numerical results.
λ1, j λ2, j dj
= (J3 (λ1, j ) − J3 (λ2, j dj )), 5.1. Ta has uniform distribution
λ2, j dj − λ1, j (18)

where Let us take θ = 1. By using Eq. (15), we obtain PTP(t1, d1) and then
calculate ETP(t1, d1) by using (A.1) and (18). Then, with the use of
2(e2 − dj e0) e LuEi (3, Lu) + e0 dj − e2 Algorithm 1, we found dopt = 0.2487 and topt = 7.6047 . To evaluate the
J3 (u) = − (kL + c ) Ei (Lu).
dj Lu2 effects of parameters on topt and dopt, in Figs. 2–5, we display the graphs
of topt and dopt for different parameter values when Ta follows the uni-
The expressions in (17) and (18) facilitate the numerical computation form distribution.
of these quantities as shown in the next section. In the next subsection, Fig. 2a shows dopt versus system income per unit time, e0, based on
we suppose that the random variable Ta has a normal distribution and the preventive maintenance cost e2 = 2, 3, 4 and 5 values. Clearly, in
then derive the required results. Fig. 2a, dopt is seen to be decreasing in e0 which means that as the
system income takes higher values, the preventive maintenance should
4.2. Ta has normal distribution be done in shorter periods. Also, dopt increases as e2 increases, i.e., when
the preventive maintenance cost is higher, one needs to perform the
Suppose Ta for the jth cycle has a normal distribution with mean preventive maintenances in longer periods.
tj > 0 and variance σ2, where tj is considerably larger than σ so as to According to Fig. 2b, topt is increasing in e0 which means longer
guarantee that P(Ta > 0) ≃ 1. The details of calculations for obtaining of repair time must be assigned for a repair while the system income in-
PTP(tj, dj) are given in Appendix A. Upon substituting (A.3) into (A.2), creases. Furthermore, an increase in e2 reduces the value of topt, from
we arrive at which we conclude that the more expensive the preventive

147
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 3. Ta has the uniform distribution with θ = 1.

maintenance is, we need to allow a shorter optimal repair time. where λ1 = 1/ E (X ) and λ2 = 1/ E (Y ) . It is seen that dopt and topt are both
In Fig. 3, dopt and topt graphs are plotted against the repair cost per decreasing in terms of λ1 (i.e., increasing in terms of E(X)). Therefore,
time unit, for the cost values of 15, 20, 23 and 25. dopt is decreasing in for an original system with larger mean lifetime, the preventive main-
terms of e1 and c. Therefore, if the repair cost or the purchase cost takes tenance can be done in a longer period and also more time can be given
higher values, the preventive maintenance must be performed earlier. for the repair. For λ1 ≥ 1, dopt and topt are both decreasing in terms of λ2
Besides, topt is decreasing in e1 and increasing in c. Thus, if the repair (i.e., increasing in terms of E(Y)). When the mean of Y is increased, it
cost increases, we need to assign a shorter time for the repair and if the means that the preventive maintenance has a good performance, and so
cost of purchase increases, a longer repair time can be allocated. it needs to be performed in longer time and more time is given for
In Fig. 4, topt and dopt are plotted against the penalty cost per unit repair. These are all clearly observed in Figs. 2–5.
while the system is off, for the waiting time for delivery values 10, 12,
13, and 13.5. As observed in Fig. 4a, dopt is decreasing in terms of k and
L. Then, as the penalty for the system being off or the waiting time for 5.2. Ta has normal distribution
delivery of a new system increases, the preventive maintenance should
be performed in shorter time intervals. Additionally, topt is an increasing When Ta has a normal distribution, for the given the parameters as
function in terms of k and L. Thus, as k and L increase, a longer repair explained earlier in this section, upon using (19) and Algorithm 1, for
time must be assigned. σ = 1, we found dopt = 0.2500 and topt = 7.5878. Figs. 6–9 contain the
In Fig. 5, the graphs of dopt and topt are plotted against λ1 and λ2, numerical results when Ta has the normal distribution displaying the
behavior of topt and dopt against other parameters. It is observed that the

Fig. 4. Ta has the uniform distribution with θ = 1.

148
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 5. Ta has the uniform distribution with θ = 1.

behaviors are almost the same as in the case when Ta has the uniform should be done in a shorter period and also less time should be
distribution. provided for the repair.
• When Ta has a normal distribution, we set σ = 1. By utilizing (19)
5.3. Numerical results for the case q < 1 (imperfect repair) and Algorithm 1, we found dj, opt = 0.2142 and t j, opt = 7.3054 . For
j = 1 (q0 = 1, X˜1 = X , perfect repair) as discussed in Section 5.2, the
qj qj
Let us take qj = 0.7, so that λ1, j = λ = 7 and λ2, j = λ = 1.4, also optimal values were determined to be dopt = 0.2500 and topt = 7.5878.
1 2
L = 10, e0 = 50, e1 = 10, e2 = 2, k = 2 and c = 20 . Then, by Therefore, when imperfect repair is done for the jth cycle in which
using Algorithm 1, we determined the optimal values of tj and dj. repair decision is made with qj = 0.7 for the degree of imperfect
repair, here again, the preventive maintenance should be done in a
• When T a has a uniform distribution ( − 1, 1), by using Eq. (15), we
shorter period and less time should be given for the repair.
obtained PTP, j(tj, dj) and then calculated ETP, j(tj, dj) by using (17)
and (18). Then, by employing Algorithm 1, we found dj, opt = 0.2131
5.4. Application to real data set
and t j, opt = 7.3351. For j = 1 (q0 = 1, X˜1 = X , perfect repair) as dis-
cussed in Section 5.1, the optimal values were found to be
In this subsection, an application of the proposed method to real
dopt = 0.2487 and topt = 7.6046. Therefore, when imperfect repair is
data is illustrated. A data set that includes failure times and repair times
done for the jth cycle in which repair decision is made with qj = 0.7
observed over the period 1972-76 on eleven nominally identical coal
for the degree of imperfect repair, the preventive maintenance
fired boilers are taken from Tables 1 and 2 of Baxter [3]. First, we check

Fig. 6. Ta has the normal distribution with σ = 1.

149
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 7. Ta has the normal distribution with σ = 1.

to see whether the exponential distribution provides a suitable fit to the fired boiler fails, topt is compared with the required repair time, say tR. If
data of failure times or not. We used the Kolmogorov–Smirnov (K-S) tR is less than or equal to topt = 7.4617931, the coal fired boiler has to be
goodness-of-fit test to check whether the distribution is adequate. The repaired, and otherwise, it should be replaced by a new one.
K-S statistic between the empirical distribution function and the fitted
exponential distribution function is 0.0501 and the corresponding p-
6. Concluding remarks
value is 0.1968. This indicates the exponential distribution fit is ade-
quate for failure times. Next, we try to fit the normal distribution to
In this paper, a repairable system has been considered. An optimal
repair times. The corresponding K-S statistic is 0.0767 and the corre-
preventive maintenance, repair-replacement policy has been proposed
sponding p-value is 0.023 and this shows the normal distribution fit is
and explained how to inspect and maintain the system before the main
good for repair times. In order to find the optimal values of dopt and topt,
failure. Also when the system fails, the problem of repairing or repla-
i.e. for the first cycle, for this coal fired boiler data, we estimate the
cing by a new system has been discussed. The average long-run cost rate
unknown parameters by using maximum likelihood method. We ob-
2 1 has been calculated and an algorithm has developed to solve the pro-
tained σ^ = 2.879026 and ^ = 8.012818, and let us take other para-
λ1 blem. It has been found the optimal period for preventive maintenance
1
meters λ = 2 and also L = 10, e0 = 50, e1 = 10, e2 = 2, k = 2 and c = 20, and the best decision for repair or replacement in terms of a measure of
2
as given in the previous sections. By using Algorithm 1, we then find probability closeness and weighted expectations of the profits. Upon
topt = 7.4617931 and dopt = 0.2257855. Therefore, based on these in- considering symmetric distribution for repair times, the optimal policy
formation, the operator has to inspect the coal fired boiler and perform has been discussed that maximizes the profit per unit time for repair
preventive maintenance in lengths of dopt = 0.2257855. When the coal and replacement. In order to illustrate the obtained studies, uniform

Fig. 8. Ta has the normal distribution with σ = 1.

150
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

Fig. 9. Ta has the normal distribution with σ = 1.

and normal distributions have been considered and for given values of • In the literature, there are several repair/replacement strategies and
parameters, the optimal period for preventive maintenance and time for many models have been developed to treat repair-replacement
repair have been investigated. Some future works in this direction may policy. Each policy has its own characteristics, advantages and dis-
focus on the following issues: advantages. Also, several authors studied the optimal policy pro-
blems based on age-replacement and periodic replacement, see for
• As mentioned by Zhang and Xie [25], there are many maintenance example, Safaei et al. [17]. Thus, this work can be extended in this
models that are built on stochastic processes and multi-state sys- direction, for example one can consider age-dependent repair.
tems. So, one may extend this study to a multi-stage failure process. • Also, in this work, it has been assumed that the waiting time for
• It is known that in queuing theory, service time is defines as the time replacing the failed system with a new system is a constant, one can
required to serve a customer and the exponential distribution is assume that the delivery time of a new system is also a random
often used to model the service times, see for example, Shortle et al. variable.
[20]. As the repair time is the same as service time, one may assume
an exponential distribution for repair time. Acknowledgment
• In this work, it has been assumed that the original system lifetime
follows an exponential distribution which is a popular life dis- The authors are grateful to the Editor, AE and four anonymous re-
tribution widely used for analyzing lifetime data. Some other life- ferees for making many helpful comments and suggestions on an earlier
time distributions with decreasing or increasing hazard rate func- version which led to this improved version of the paper. This research
tion, for example Weibull or gamma distributions, can be considered was supported by a grant from Ferdowsi University of Mashhad; No 2/
for original system lifetime. 47515.

Appendix A. Some computations

In this appendix, we present explicit expression for E(TPrepair) for the jth cycle when Ta has a uniform distribution, also for PTP(tj, dj) when Ta has a
normal distribution.

• Suppose T a has a uniform distribution, using (10) and (14), the expectation of TPrepair for the jth cycle can be expressed as follows:

E (TPrepair , j ) =
λ1, j λ2, j dj +∞
∫ ∫t −θ
t j+θ (e − ) z − (e + k) t (e
0
e2
dj 1 a
−λ1, j z − e−λ2, j dj z ) dta dz
2θ (λ2, j dj − λ1, j ) 0 j z + ta
λ1, j λ2, j dj ⎧ ∞⎛ z + tj + θ ⎞ ⎞
= ∫ z ln ⎜⎛ ⎟ (e0 + e1 + k − e2 / dj ) − 2θ (e1 + k )
2θ (λ2, j dj − λ1, j ) ⎨ 0 ⎜ ⎝ z + t j − θ ⎠
⎟
⎩ ⎝ ⎠
× (e−λ1, j z − e−λ2, j dj z ) dz }

⎧ ∞ z + t j + 1 ⎞ −λ z
z ln ⎜⎛ 1, j − e −λ2, j dj z dz ⎫
= C1
⎨ 0
∫ z + tj − 1 ⎠
⎟ (e ) ⎬
⎩ ⎝ ⎭
(λ1, j λ2, j dj )(e1 + k ) ⎛ 1 1 ⎞
− ⎜ − ⎟
λ2, j dj − λ1, j ⎝ λ1, j λ2, j dj ⎠
(λ1, j λ2, j dj )(e1 + k ) ⎛ 1 1 ⎞
= C1 {J2 (λ1, j ) − J2 (λ2, j dj )} − ⎜ − ⎟,
λ2, j dj − λ1, j ⎝ λ1, j λ2, j dj ⎠ (A.1)

151
F. Safaei et al. Reliability Engineering and System Safety 183 (2019) 143–152

λ1, j λ2, j dj (e0 + e1 + k − e2 / dj )

where C1 = 2θ (λ2, j dj − λ1, j )
, and

1 ⎛ u (t j − θ)
J2 (u) = e (u (t j − θ) − 1) Ei (u (t j − θ))
u2 ⎜
⎝
tj + θ ⎞⎞
− eu (t j + θ) (u (t j + θ) − 1) Ei (u (t j + θ)) + ln ⎜⎛ ⎟ .
⎟
⎝ tj − θ ⎠⎠

• Suppose T a has a normal distribution. In this case Eq. (13) can be written as

λ1, j λ2, j dj ⎧ ∞ −h2 z

PTP (t j, dj ) =
λ2, j dj − λ1, j ⎨
∫0 Φ⎛ ⎜ − t j / σ ⎞ e−λ1, j zdz
⎟

⎩ ⎝ σ (h1 z + h3) ⎠
∞ −h2 z
− ∫0 Φ⎛⎜ − t j / σ ⎞ e−λ2, j dj zdz ⎫
⎟

⎝ σ (h1 z + h3) ⎠ ⎬
⎭
λ1, j λ2, j dj
= {J4 (λ1, j ) − J4 (λ2, j dj )},
λ2, j dj − λ1, j (A.2)

where J4 (u) = ∫ 0
∞ −h z
1 (
Φ σ (h z 2+ h )
3
− t j/σ ) e−uzdz and Φ(.) is the cdf of the standard normal distribution. By setting a (z ) =
−h2 z
σ (h1 z + h3)
− t j / σ , J4(u) can be
expressed as
∞ a (z )
J4 (u) = ∫0 ∫−∞ ϕ (m) e−uzdmdz
h
1 − 2 −t j/σ ⎛ −uh3 (mσ + t j ) ⎫ ⎞
= ∫−t /σhσ 1 ϕ (m) ⎜1 − exp ⎧ ⎟ dm
u j ⎨
⎩ 2 + h1 (mσ + t j ) ⎬
h ⎭⎠
⎝
h
1 − 2 −uh3 σq ⎫ ⎞
= ∫0 σh1 ϕ (q − t j / σ ) ⎜⎛1 − exp ⎧ ⎟ dq ,
u ⎝ ⎨
⎩ 2 + h1 σq ⎬
h ⎭⎠ (A.3)
where ϕ(.) is the pdf of the standard normal distribution.

Supplementary material

Supplementary material associated with this article can be found, in the online version, at 10.1016/j.ress.2018.11.012.

References 2017;256:439–53.
[15] Park M, Jung KM, Park DH. Optimal post-warranty maintenance policy with repair
time threshold for minimal repair. Reliab Eng Syst Saf 2013;111:147–53.
[1] Abramowitz M, Stegun IA. Handbook of mathematical functions with formulas, [16] Park M, Jung KM, Park DH. Two-dimensional maintenance with repair time
graphs, and mathematical tables. New York: Dover Publications; 1972. threshold and generalized age replacement policy. In: Nakamura S, Qian CH,
[2] Barlow RE, Hunter LC. Optimum preventive maintenance policies. Oper Res Nakagawa T, editors. Reliability Modeling with Computer and Maintenance
1960;8:90–100. Applications. Singapore: World Scientific Publishing; 2017. 295–311
[3] Baxter LA. A case study in availability forecasting. Microelectron Reliab [17] Safaei F, Ahmadi J, Gildeh BS. An optimal planned replacement time based on
1985;25:927–42. availability and cost functions for a system subject to three types of failures. Comput
[4] Dohi T, Matsushima N, Kaio N, Osaki S. Nonparametric repair-limit replacement Ind Eng 2018;124:77–87.
policies with imperfect repair. Eur J Oper Res 1997;96:260–73. [18] Sarkar A, Panja SC, Sarkar B. Survey of maintenance policies for the last 50 years.
[5] Garg A, Deshmukh SG. Maintenance management: literature review and directions. Int J Softw Eng Appl 2011;3:130–48.
J Qual Maint Eng 2006;12:205–38. [19] Sheu SH, Liu TH, Zhang ZG, Tsai HN. The generalized age maintenance policies
[6] Jack N, Iskandar BP, Murthy DNP. A repair–replace strategy based on usage rate for with random working times. Reliab Eng Syst Saf 1 2018;69:503–14.
items sold with a two-dimensional warranty. Reliab Eng Syst Saf 2009;94:611–7. [20] Shortle JF, Thompson JM, Gross D, Harris CM. Fundamentals of queueing theory.
[7] Kijima M. Some results for repairable systems with general repair. J Appl Probab 5th edition New York: John Wiley & Sons; 2018.
1989;26:89–102. [21] Sun Q, Ye Z, Chen N. Optimal inspection and replacement policies for multi-unit
[8] Kijima M, Morimura H, Suzuki Y. Periodical replacement problem without assuming systems subject to degradation. IEEE Trans Reliab 2018;67:401–13.
minimal repair. Eur J Oper Res 1988;37:194–203. [22] Whang GJ, Zhang YL. Optimal repair-replacement policies for a system with two
[9] Kim HG, Yun WY. A repair-time limit replacement policy with estimation error. types of failures. Eur J Oper Res 2013;226:500–6.
Commun Stat-Theory Methods 2010;39:2365–78. [23] Yung C, Bonnett AH. Repair or replace? IEEE Ind Appl Mag 2004;10:48–58.
[10] Koshimae H, Dohi T, Kaio N. Graphical statistical approach to repair limit re- [24] Zhang YL, Wang GJ. An optimal age-replacement policy for a simple repairable
placement problem. J Oper Res Soc Jpn 1996;39:230–46. system with delayed repair. Commun Stat-Theory Methods 2017;46:2837–50.
[11] Kumar D, Klefsjö B. Proportional hazards model: a review. Reliab Eng Syst Saf [25] Zhang M, Xie M. An ameliorated improvement factor model for imperfect main-
1994;44:177–88. tenance and its goodness of fit. Technometrics 2017;59:237–46.
[12] Lim JH, Qu J, Zuo MJ. Age replacement policy based on imperfect repair with [26] Zhao X, Al-Khalifa KN, Hamouda AM, Nakagawa T. Age replacement models: a
random probability. Reliab Eng Syst Saf 2016;149:24–33. summary with new perspectives and methods. Reliab Eng Syst Saf
[13] Love CE, Zhang ZG, Zitron MA, Guo R. A discrete semi-markov decision model to 2017;161:95–105.
determine the optimal repair/replacement policy under general repairs. Eur J Oper [27] Zhao X, Qian C, Nakagawa T. Comparisons of replacement policies with periodic
Res 2000;125:398–409. times and repair numbers. Reliab Eng Syst Saf 2017;168:161–70.
[14] Nguyen DT, Dijoux Y, Fouladirad M. Analytical properties of an imperfect repair
model and application in preventive maintenance scheduling. Eur J Oper Res

152