Reliability Engineering and System Safety 204 (2020) 107123

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

Condition-based maintenance with dynamic thresholds for a system using T

the proportional hazards model
Rui Zhenga, , Bingkun Chenb, Liudong Gua
⁎

a
School of Mechanical Engineering, Southeast University, Nanjing 211189, China
b
School of Mechanical and Aerospace Engineering, Jilin University, Changchun 130022, China

ARTICLE INFO ABSTRACT

Keywords: The hazard rate of many practical systems depends not only on age but also on a diagnostic covariate process.
Condition-based maintenance Effective maintenance decisions for such systems need to combine both age information and the covariate in-
Dynamic thresholds formation obtained from condition monitoring. This paper proposes a condition-based maintenance (CBM)
Multiple maintenance actions policy with dynamic thresholds and multiple maintenance actions for such a system subject to periodic in-
Proportional hazards model
spection. The hazard rate is described by the proportional hazards model with a continuous-state covariate
Policy-iteration algorithm
process. At each inspection epoch, appropriate action is selected from no maintenance, imperfect maintenance,
and preventive replacement based on two dynamic thresholds. Over an inspection interval, the system may
experience minor failure or catastrophic failure that can be addressed by minimal repair and corrective re-
placement, respectively. The objective is to determine the optimal thresholds that minimize the long-run average
cost rate. A modified policy-iteration algorithm is developed to solve the optimization problem in the semi-
Markov decision process (SMDP) framework. The effectiveness of the proposed approach is illustrated by a
practical numerical example. The comparison with the other CBM policies confirms the outstanding performance
of the proposed policy.

1. Introduction system is left in operation otherwise [6–8]. Golmakani and Fattahipour

[9] developed an algorithm to simultaneously optimize the replacement
Condition-based maintenance (CBM) has been recognized as an ef- threshold and the inspection interval. Tian et al. [10] used a multi-
fective and efficient method of preventing unexpected failures and re- objective CBM model to find the optimal replacement threshold which
ducing operation and maintenance (O&M) costs [1]. It utilizes updated could make a systematical tradeoff between cost minimization and re-
information collected from condition monitoring for the maintenance liability maximization. Tian and Liao [11], Zhao et al. [12] utilized the
decision-making of deteriorating systems [2]. Cox's proportional ha- replacement policy for the maintenance of each component in a multi-
zards model (PHM) is one of the most popular stochastic models de- component system. Xu et al. [13] proposed a group replacement policy
scribing deteriorating processes in the CBM framework. The model is for a two-unit system where a system-level risk threshold was set to
particularly suitable for the situation where the hazard rate of a system determine whether to perform group replacement or not. Assuming the
depends not only on its age but also on diagnostic processes such as hazard rate of the core component in a series system followed the PHM,
erosion, corrosion, crack growth, and wear [3]. These processes are Jafari et al. [14] optimized the hazard-rate thresholds for opportunistic
treated as covariates and integrated into the hazard function through a replacement and preventive replacement simultaneously. The PHM
link function [4]. A sound CBM decision for a system using the PHM covariates mentioned above were described by Markovian models,
needs to combine both age information and the covariate information whereas a more practical method is to characterize them as continuous-
obtained from condition monitoring [5]. state stochastic models. Tang et al. [15] proposed a CBM policy for a
As a natural combination of age and covariate values, the hazard system subject to dependent soft failure and hard failure. A random-
rate in the PHM has been frequently used for setting CBM rules. coefficient regressive degradation process was integrated into the
Considerable attention has been devoted to the popular replacement models of both failure modes. At each inspection epoch without soft
policy where preventive replacement is performed if the calculated failure, the replacement policy was used to prevent the system from
hazard rate at an inspection epoch exceeds a certain threshold, and the running at a high risk of hard failure. Liu et al. [16] studied the CBM

⁎
Corresponding author.
E-mail address: [email protected] (R. Zheng).

https://doi.org/10.1016/j.ress.2020.107123
Received 10 January 2020; Received in revised form 6 July 2020; Accepted 11 July 2020
Available online 15 July 2020
0951-8320/ © 2020 Elsevier Ltd. All rights reserved.
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

models where the PHM covariate was characterized as a Wiener pro- guild practical maintenance implementation.
cess. The optimal hazard-rate replacement threshold was obtained in The contributions of this paper are summarized as follows: (1)
two scenarios where the parameters of the covariate distribution were Development of a CBM model considering dynamic thresholds and
known and unknown in advance respectively. Duan et al. [17] devel- multiple maintenance actions; (2) Consideration of both stationary and
oped a CBM policy using the PHM with a gamma process as its cov- non-stationary covariate processes of a PHM model; (3) Employment of
ariate. Besides the replacement threshold, an age threshold was also a discretization method for the calculation of SMDP quantities; (4)
considered for selecting maintenance actions. Aforementioned CBM Development of a modified policy-iteration algorithm for the optimi-
policies involved a constant hazard-rate threshold which was con- zation problem.
venient to be optimized in computation. The remainder of this paper is organized as follows. Section 2 de-
However, making a constant decision throughout the lifecycle of a scribes the hazard function and the CBM policy and then gives the
system is not cost-effective in many situations. One way to extend the optimization problem. Section 3 develops the computational algorithm
traditional replacement policy is to consider dynamic thresholds. For a in the SMDP framework. In Section 4, a numerical example is used to
system subject to dependent soft and hard failures, Tang and Yu [18] demonstrate the effectiveness of the proposed approach. Section 5
proposed a CBM policy with a dynamic replacement threshold, which summarizes the conclusions and discusses possible directions in future
was proven to be monotonically non-increasing with the inspection research.
number. Yahyatabar and Najafi [19] studied a CBM policy with a dy-
namic replacement threshold for a single-unit system. A multi-stage 2. Problem description
stochastic programming approach was developed to optimize the
threshold by minimizing the cost rate in a finite horizon. An extension The following notation is adopted in this paper:
of this study to series-parallel systems was presented in [20]. However,
there still remains a gap in the computation algorithm for finding the
t system age
optimal policy which involves more than one dynamic threshold.
ξ system age at failure
Moreover, the replacement policy only involves preventive and q probability of a failure to be catastrophic
corrective replacements which cannot be performed in engineering X covariate process
practice unless a system is severely deteriorated or under a catastrophic I state space of the covariate process
failure. Therefore, another way to extend the traditional policy is to M absorbing covariate state
h(t, Xt) hazard function at age t
consider imperfect maintenance and/or minimal repair. Mousavi et al. h0(∙) baseline hazard function
[21] proposed a CBM policy where imperfect maintenance or replace- Ψ(∙) link function
ment was selected at each inspection epoch according to a constant v(t) shape parameter of a gamma process
hazard-rate threshold. An algorithm was developed to optimize the u scale parameter of a gamma process
δ a sufficiently small time interval
threshold and the imperfect maintenance policy simultaneously. In
Δ inspection interval
Yahyatabar and Najafi [19,20], the proposed policies excluded cor- K discretization level of Δ, i.e., Δ = Kδ
rective replacement, and all failures were addressed by minimal repair (n, i) system state of age nΔ and covariate state i
irrespective of their failure types. However, further efforts are still re- S system state space
quired to make CBM decisions considering multiple maintenance ac- α improvement factor of imperfect maintenance
CI inspection cost
tions and different failure types.
Cmr cost of minimal repair
Motivated by these requirements, this paper investigates the opti- Cpm cost of imperfect maintenance
mization problem of a CBM policy with dynamic thresholds and mul- Cpr cost of preventive replacement
tiple maintenance actions for a system subject to periodic inspections. Ccr cost of corrective replacement
Tmr time of minimal repair
The hazard rate is modeled by the PHM with a continuous-state cov-
Tpm time of imperfect maintenance
ariate process. Unlike the previous PHM that mostly incorporated sta- Tpr time of preventive replacement
tionary covariate processes, this paper also examines the non-stationary Tcr time of corrective replacement
case which is suitable for many practical deterioration processes (e.g., p(n,i),(n , i ) transition probability from state (n , i ) S to state (n , i ) S
[22,23]). At each inspection epoch where the covariate value is avail- τ(n, i) expected sojourn time for state (n , i ) S
able, appropriate action is selected from no maintenance, imperfect c(n, i) expected cost for state (n , i ) S
v(n, i) relative value for (n , i ) S
maintenance, and preventive replacement based on two dynamic
G maintenance policy
thresholds. During an inspection interval, the system is subject to thresholds for imperfect maintenance at the nth inspection
L1n
random failure that triggers a corrective maintenance action im- L2n thresholds for preventive replacement at the nth inspection
mediately. Two failure types are considered in this paper. If a cata- g long-run average cost rate
strophic failure occurs, corrective replacement is carried out to restore
the system to the brand-new state. If a minor failure occurs, minimal
repair is used to restore the system to the operational state. The ob- 2.1. The hazard function
jective is to determine the optimal thresholds that minimize the long-
run average cost rate. The optimization problem is formulated in the This paper proposes a CBM policy for a system whose hazard rate
semi-Markov decision process (SMDP) framework. depends both on its age t and on a covariate process X = {Xt | t > 0}.
The formulation process requires the calculation of conditional re- According to Cox's PHM, the hazard function of the system is expressed
liability between two consecutive inspection epochs, which is quite as
difficult due to the “random jump” of the continuous-state covariate h (t , Xt ) = h 0 (t ) (Xt ) (1)
process [24,25]. To address this uncertainty properly, this paper de-
velops a discretization method which is distinguished from the previous where h0(t) is the baseline hazard function, and Ψ(Xt) is a positive link
methods [17,26] by using a contract transition matrix to reduce the function describing the effect of the covariate process on failure time.
memory requirement in computation. On that basis, the SMDP quan- Its survivor function is given by
tities, i.e., transition probabilities, expected sojourn time, and expected t
cost, are derived. A modified policy-iteration algorithm which is proved Pr( > t|Xs , 0 < s < t ) = exp h 0 (s ) (Xs ) ds
0
(2)
to be effective is developed to find the optimal maintenance policy. The
graphical representation of the obtained policy is then presented to The process X can be either external environmental factors (e.g.,

2
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

temperature, humidity, and work load) [27,28] or internal diagnostic catastrophic and minor are assumed to be q (q ∈ [0, 1]) and (1 – q),
variables (e.g., crack, corrosion, and wears) [29]. In this paper, the respectively. If a catastrophic failure occurs, corrective replacement is
covariate process is assumed to be monotonically increasing and performed immediately to return the system to the state (0, 0). If a
modeled by a gamma process. This assumption is realistic in many minor failure occurs, minimal repair is performed immediately to re-
practical situations such as the crack growth of a mechanical system. store the system state to its previous known system state. This as-
The process starts from zero, i.e., X0 = 0. The increments between time sumption implies that the effect of minimal repair is a little more than
points t and t + Δt are independent and follow a gamma distribution as “as bad as old” since both the system age and the covariate state are
follows non-decreasing over timer. However, this small rejuvenation can be
neglected and, following the assumption, the subsequent computation
u v (t + t ) v (t ) x v (t + t ) v (t ) 1e ux
f (x|t , t ) = Pr(Xt + t Xt = x ) = , x 0 procedure will be significantly simplified.
(v (t + t ) v (t ))
The proposed maintenance policy can then be expressed in the
(3) matrix form as
where v(t) ≥ 0 is the shape parameter; u > 0 is the scale parameter;
L10 L 20
(x ) = 0 s x 1e sds is the gamma function [30]. If v(t) is a linear
function of time t, the process X reduces to a stationary process; L11 L21
G=
otherwise the process is non-stationary.
For the convenience of both computation and the subsequent policy L1N L 2N (6)
description, the covariate process X is discretized into a multi-state
where [L10 , L11, and
...,L1N ]T are the dynamic thresholds
[L 20 , L 21 , ...,L2N ]T
Markov chain with state space I = {0, 1,…, M}, where M is the ab-
for imperfect maintenance and preventive replacement, respectively.
sorbing state. Letting d denote the discretization interval, then Xt is
The objective is to determine the optimal policy that minimizes the
supposed to be in state i {0, ...,M 1} if Xt [id , (i + 1) d) , and in
long-run average cost rate.
state M if Xt [Md , + ) . Similar discretization methods can be found
in [31]. Using the mid-point (i + 0.5)d as the covariate value of state
3. Computational algorithm in the SMDP framework
i ∈ I approximately, the probability that the process X transits from
state i ∈ I to state j ∈ I, where i ≤ j, between time points t and t + Δt is
This section formulates and solves the maintenance optimization
given by
problem in the SMDP framework. First, a method is developed to cal-
P(t , i),(t + t,j) = Pr(Xt + t =j|Xt = i ) culate the conditional reliability between two discrete time points.
ub Second, we derive the SMDP quantities including transition prob-
= f (x|t , t ) dx
lb abilities, expected sojourn time, and expected cost. A modified policy-
(v (t + t ) v (t ), lb ·u ) (v (t + t ) v (t ), ub ·u ) algorithm is later presented to solve the optimization problem.
=
(v (t + t ) v (t ))
(4) 3.1. Conditional reliability

where (a , b) = b x a 1e xdx is the incomplete gamma function; lb and

To proceed with the derivation of SMDP quantities, it is necessary to
ub are the lower and upper bounds given as following
know the reliability over the next inspection interval conditional on the
lb = max(0, (j i 0.5) d) j = i , i + 1, ...,M system state at the current inspection epoch. As mentioned before, the
(j i + 0.5) dj = i , i + 1, ...,M 1 covariate value may jump with any increment at any time point. To
ub = address this uncertainty in computation, the inspection interval Δ is
+ j=M (5) discretized into K equidistant intervals, i.e., δ = Δ/K, where K is an
enough large integer that generates a sufficiently small δ. It is then
reasonable to assume that the covariate process has no transition during
2.2. The CBM policy
the time interval δ. To calculate the conditional reliability function, a
partial transition probability matrix of the joint age and covariate
The system is subject to condition monitoring at discrete equidistant
process between time points t and t + δ is defined as
time epochs nΔ, where n = 1, ...,N , Δ is the inspection interval, and NΔ
is the maximum useful lifetime of the system. In computation, N should H(t ) = [hij (t )|i , j I ](M + 1) × (M + 1)
be set to be great enough so as not to influence the optimal policy [32].
= [Pr ( > t + , Xt + =j| > t ,Xt = i )|i , j I]
The covariate value is available at each inspection epoch and the in-
spection time is insignificant. = P (t , i),(t +
t+
, j )·exp ((i + 0.5) d)· h 0 (s ) ds |i, j I
Define the system state space by t
(7)
S = (0, 0) {(n, i)|n = 1, 2, ...,N ; i I } , where (0, 0) is the “as good as
new” state; n indicates the system age to be nΔ; and i represents the where the elrement hij(t) denotes the probability that the system tran-
covariate state. At the nth inspection epoch, two thresholds L1n and L2n , sition from covariate state i ∈ I at the time point t to state j ∈ I at t + δ;
where L1n L2n , are set to select actions for the state (n, i) as follows: P(t , i),(t + , j) is the transition probability of the process X that can be cal-
culated in Eq. (4). The proof of Eq. (7) is given in Appendix A. It is

• If n < N and i < L , no maintenance is taken.

n worth mentioning that H(t) does not consist of the transition probability

• If n < N and L i < L , imperfect maintenance is carried out. The

1
n n to the failure state. Therefore, the sum of each row is less than 1 with
1 2
system state after imperfect maintenance is restored to (n, ⌊iα⌋), probability one for any t ≥ 0 and i ∈ I.
where ⌊ · ⌋ is the floor function and α ∈ (0, 1) is a predetermined Assume that currently the system age is t and the covariate state is
improvement factor. A similar model has also been considered in i ∈ I. Then the probability that the system survives to the age t + rδ
[33,34]. (r = 0, 1,…) with the covariate state j ∈ I is given by

• If n = N or i L 2n , preventive replacement is performed to return r 1

H (t + k )· eTj
the system state to (0, 0). Pr (Xt + r = j, > t + r Xt = i , > t ) = ei ·
k=0 (8)
During an inspection interval, the system is subject to random where ei is a row vector of (M + 1) with the (i + 1)th element to be 1
failure that is self-announcing. The probabilities of a failure to be and the other elements to be 0; eTj is the transposed vector of ej. The

3
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

proof of Eq. (8) is given in Appendix B.

Based on Eq. (8), the probability that the system will survive to time
point t + rδ (r = 0, 1,…) can be derived by

R (r |t , i ) = Pr ( > t + r |Xt = i, > t)

= Pr (Xt + r = j, > t + r Xt = i, > t)
j I
r 1
= ei · H (t + k )· eTj
j I k=0
r 1
= ei · H (t + k )·1
k=0 (9)

where 1 is a column vector of M + 1 with all the elements equal to 1.

Then we can derive the mean residual life (MRL), which is an im-
portant health measurement in reliability engineering applications.
Using rδ as the residual life if the system fails between t + (r - 1)δ and
t + rδ approximately, the MRL of the system conditional on age t and Fig. 1. Transition diagram for the no-maintenance state.
covariate state i ∈ I is given by
p(n, i),(n, i) = (1 q)·Pr ( < (n + 1) > n , Xn = i )
MRL (t , i) = E { |Xt = i , > t}
= (1 q)[1 R ( |n , i)]
K 1
= r ·Pr (t + (r 1) < < t + r Xt = i , > t)
r=1
= (1 q) 1 ei · H (n + k )· 1
k=0 (12)
= r ·[R ((r 1) t , i) R (r |t , i)]
r=1 The last path denotes the situation where a catastrophic failure
occurs before the next inspection epoch. Corrective replacement is
= · R (r | t , i )
performed to return the system state to (0, 0). The transition probability
r=0
r 1 is calculated as
= · ei· H (t + k )· 1
r=0 k=0 (10) p(n, i),(0,0) = q· Pr ( < (n + 1) > n , Xn = i)
= q·[1 R ( |n , i )]
It is worth mentioning that the proposed method is also applicable K 1
for the situation where the PHM covariate is modeled by multi-state = q· 1 ei · H (n + k )· 1
Markovian process or the other continuous-state stochastic processes. k=0 (13)

In Fig. 1, the left notation below each transition path denotes its
3.2. SMDP quantities sojourn time. If the system survives to the next inspection epoch, the
sojourn time is Δ. Otherwise, the sojourn time is associated with failure
The SMDP is determined by the following quantities [35]: time and corresponding repair time. The expected sojourn time of the
current state is given by
• i′)p ∈ S at=thethenext
(n, i),(n , i ) probability that the system will be in the state (n′,
decision epoch if the current state is (n, i) ∈ S; (n, i)

• τ = the expected sojourn time until the next decision epoch if the
(n, i) = · p(n, i),(n + 1, j) +
0
[(1 q)(t + Tmr ) + q (t + Tcr )] dF (t|n , i)
current state is (n, i) ∈ S; j I

• = the expected cost incurred until the next epoch if the current
c (n, i)
= · R ( |n , i) + [(1 q)(t + Tmr ) + q (t + Tcr )] dF (t|n , i )
state is (n, i) ∈ S. 0

= R (t|n, i) dt + [(1 q) Tmr + qTcr ][1 R ( |n , i)]

These SMDP quantities under a stationary policy G are derived in 0
K m 1
the following three scenarios: = · ei · H (n + k )·1
Scenario 1: The current system state is (n, i), where i < L1n and n < N. m =1 k=0
According to the policy description, no maintenance is taken at the K 1
current state. There are three types of possible transition paths as + [(1 q) Tmr + qTcr ] 1 ei · H (n + k )· 1
shown in Fig. 1. k=0

The notation above each transition path denotes its transition (14)
probability. The first path denotes the situation where the system sur-
where F (t|n , i) = 1 R (t|n , i ) is the probability that the system fails
vives to the next inspection epoch with the covariate state j ∈ I.
before ξ = nΔ + t; Tmr denotes the time of minimal repair; and Tcr
According to Eq. (8), the transition probability is calculated as
represents the time of corrective replacement.
p(n, i ),(n + 1, j) = Pr (X(n+ 1) = j , > (n + 1) Xn = i , >n ) The right notation below a transition path denotes its cost. If the
K 1
system survives to the next inspection epoch, an inspection is required
to reveal the covariate value which incurs inspection cost CI. If cata-
= ei· H (n + k )·eTj
k=0 (11) strophic or minor failure occurs, corresponding repair cost is incurred.
Since the next system state after minimal repair or corrective replace-
The second path denotes the situation where a minor failure occurs ment is known, it is unnecessary to carry out an inspection and thus no
before the next inspection epoch. Minimal repair is then carried out and inspection cost is incurred. The expected cost for the current state can
restores the system state to (n, i). The transition probability is given by be calculated as

4
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

Fig. 2. Transition diagram for the imperfect-maintenance state.

c (n, i ) = CI · p(n, i),(n + 1, j) + Cmr ·p(n, i ),(n, i) + Ccr ·p(n, i),(0,0) 0) after performing corrective replacement. The transition probability is
j I given by
= CI · R ( |n , i) + Cmr (1 q)[1 R ( |n , i )] p(n, i),(0,0) = p(n, i),(n, i ) · p(n, i ),(0,0)
+ Ccr q [1 R ( |n , i)]
= 1·[q· Pr ( < (n + 1) > n , Xn = i )]
= [CI (1 q) Cmr qCcr ]· R ( |n , i) + (1 q) Cmr + qCcr
= q·[1 R ( |n , i )]
K 1
K 1
= [CI (1 q) Cmr qCcr ]· ei · H (n + k )· 1
k=0
= q· 1 ei · H (n + k )· 1
+ (1 q) Cmr + qCcr k=0 (18)

(15) The expected sojourn time of the current state consists of two parts:
(1) the sojourn time between the current state and the intermediate
where CI, Cmr and Ccr are costs associated with inspection, minimal
system state Tpm; and (2) the expected sojourn time between the in-
repair, and corrective replacement, respectively.
termediate state and the possible next states. The calculation of the
Scenario 2: The current system state is (n, i), where L1n i < L2n
second part is similar to Eq. (14). In sum, the expected sojourn time of
and n < N.
the current state is calculated as
For the current state, imperfect maintenance is carried out to restore
the system to state to (n, ⌊iα⌋) which requires no maintenance and thus (n, i) = Tpm + ·R ( n , i )
is treated as an intermediate state. Starting with the intermediate state,
there are three possible transition paths which are similar to those in + [(1 q)(t + Tmr ) + q (t + Tcr )] dF (t n , i )
0
Fig. 1. The transition paths of the current state are presented in Fig. 2.
= Tpm + R (t n, i ) dt
The first path denotes that the system successfully transits to next 0

inspection epoch with covariate state j ∈ I. The transition probability is + [1 R ( n , i )][(1 q) Tmr + qTcr ]
given by K m 1
= Tpm + · ei · H (n + k )· 1
p(n, i),(n+ 1, j ) = p(n, i),(n, i ) ·p(n, i ),(n + 1, j) m=1 k=0
= 1·Pr (X(n+ 1) = j , > (n + 1) Xn = i , >n ) K 1
K 1 + 1 ei · H (n + k )·1 [(1 q) Tmr + qTcr ]
= ei · H (n + k )·eTj k=0 (19)
k=0
The expected cost of the current state consists of two parts: the cost
(16)
of imperfect maintenance Cpm, and the expected cost incurred after the
The second path denotes that a minor failure occurs in next in- intermediate state is similar to Eq. (15). Therefore, the expected cost is
spection interval. Minimal repair is incurred to restore the system state
c (n, i)
to (n, ⌊iα⌋). The transition probability is
= Cpm + CI · p(n, i ),(n + 1, j) + Cmr ·p(n, i ),(n, i ) + Ccr ·p(n, i ),(0,0)
p(n, i),(n, i ) = p(n, i ),(n, i ) ·p (n, i ),(n, i ) j I
= 1·[(1 q)·Pr ( < (n + 1) > n , Xn = i )] = Cpm + [CI (1 q) Cmr qCcr ]· R ( n , i ) + (1 q) Cmr + qCcr
= (1 q)·[1 R ( n , i )] K 1
K 1 = Cpm + [CI (1 q) Cmr qCcr ]· e i · H (n + k )· 1
= (1 q)· 1 ei · H (n + k )· 1 k=0
k=0 (17) + (1 q) Cmr + qCcr

The last path represents the situation where a catastrophic failure (20)
occurs in the next inspection interval. The system state will transit to (0, Scenario 3: The current system state is (n, i), where i L 2n or
n = N.

5
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

algorithm where policy improvement is conducted for each state [35],

in this step, all the system states at the same inspection are treated as a
whole and the sum of their relative values are used as the policy-im-
provement quantity. More specifically, for n = 1, 2, ...,N 1, determine
(L1n , L 2n) yielding the minimum in
Fig. 3. Transition diagram for the preventive-replacement state.
M
min c(n, i) (L1n , L 2n) g (G)· (n, i) (L1n , L 2n) +
(L1n , L2n) i=0

Preventive replacement is carried out and the transition path is

deterministic as shown in Fig. 3. p(n, i ),(n , i ) (L1n , L 2n)· v(n , i ) (G)
The next state after replacement is (0, 0), and the transition prob- (n , i ) S
(28)
ability is
A new policy Ḡ can then be created by replacing where (L1n , L 2n) ,
p(n, i),(0,0) = 1 (21)
n = 1, 2, ...,N 1, in G with the values obtained in Eq. (28). The ef-
The expected sojourn time is equal to the preventive replacement fectiveness of the policy-improvement quantity is proved in
time Tpr, i.e., Appendix C.
Step 4 (convergence test). If G¯ = G , then the algorithm is stopped
(n, i) = Tpr (22)
and the optimal policy is G; Otherwise, go to step 2 with G replaced by
The expected cost is the cost of preventive replacement Cpr. That is, Ḡ . According to Appendix C, Ḡ is uniformly better than G, or just as
good as G. In the latter situation, both G and Ḡ are optimal policies.
c(n, i) = Cpr (23)
Following this step, policy evaluation and policy improvement are re-
peatedly and alternatively performed for monotonically improving
3.3. A modified policy-iteration algorithm policies and relative values until the optimal policy is obtained. Because
a finite SMDP has only a finite number of policies, the algorithm must
Once the SMDP quantities are obtained, according to the renewal converge to the optimal policy in a finite number of iterations.
theory, the long-run average cost rate under the policy G is given by

g (G) = (G)·c (n, i) (G)/ (G )· (G ) 4. Numerical examples

(n, i) (n, i) (n, i )
(n, i) S (n, i) S (24)
In this section, the proposed approach is applied to support the
where π(n, i)(G) is equilibrium (stationary) state probability that can be maintenance decision for a type of feeding system of boring machines.
derived by solving the following set of equations [35]: The wear of this system results in the increase of manufacturing error
(G) = p(n , i ),(n, i) (G)· that can be obtained through quality inspection and then integrated
(n, i) (n , i ) (G), (n , i ) S
(n , i ) S into the hazard function as a covariate. The nondecreasing covariate
process is modeled by a gamma process, where the shape parameter v
(n, i) (G) = 1
(t) = 0.0115t and the scale parameter u = 4.86. The baseline hazard
(n, i) S (25)
function follows the Weibull distribution h 0 (t ) = t 1/ , where
However, it is computationally not feasible to find the optimal η = 2050 and β = 4.63, and the link function has an exponential form
policy by computing average cost rates for all possible policies. Linear (Xt ) = exp(0.281Xt ) [17]. The other parameters are set as follows:
programming and dynamic programming (e.g., policy-iteration algo- M = 7, d = 1, K = 5, Δ = 100, N =20, q = 0.4; α = 0.5, CI = 10,
rithm and value-iteration algorithm) are more versatile methods to Cpr = 3500, Ccr = 5500, Cpm = 250, Cmr = 800; Tpr = 5, Tcr = 10,
handle similar stochastic problems. This paper develops a modified Tpm = 5, Tmr = 5.
policy-iteration algorithm for the optimization problem. The mod- Fig. 4 presents the computational process of the proposed algorithm
ification is made due to the restriction that different system states at the starting from the initialized policy given in Step 1. Any newer policy
same inspection epoch share the same threshold pair. The algorithm is generates lower average cost rate and performs better than the older
presented as follows: policy does, which confirms the effectiveness of the policy-improve-
Step 1 (initialization). For the convenience of computation, set ment quantity. The algorithm converges after just 5 iterations, which
(L1n , L 2n) , where
= {(l1, l2 ) 0 l1 l2 M + 1; l1, l2 N} (26)
Initialize a policy G as follows: (L10 , L20 ) = (M + 1, M + 1) ,
(L1n , L 2n) = (M , M ) for n = 1,2,…,N - 1, and (L1N , L 2N ) = (0, 0) . In this
policy, preventive replacement is performed if the covariate state is M
or at the last epoch, and no maintenance is carried out otherwise.
Step 2 (policy evaluation). Obtain the average cost rate and relative
values of the current policy by solving:

v (n, i ) = c (n, i) (G) g (G)· (n, i) (G) + p(n, i),(n , i ) (G) v (n , i )

(n , i ) S
v (ns, is ) = 0 (27)
where (n, i) ∈ S; (ns, is) ∈ S is an arbitrarily chosen state; g(G) is the
average cost rate of policy G; and v(n,i) is the relative value function of
state (n, i) ∈ S.
Step 3 (policy improvement). It is noted that the threshold pair for
n = 0 and n = N are deterministic and the aim is to improve the
thresholds for n = 1, 2, ...,N 1. Unlike the classical policy-iteration Fig. 4. Long-run average cost rate for different iterative cycles.

6
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

Fig. 5. Optimal maintenance actions for different combinations of age and

covariate state.

emphasizes the high efficiency of the proposed algorithm. The minimal

long-run average cost rate is 272.3 per inspection cycle.
Fig. 5 presents the optimal maintenance policy obtained in the
above process. When the system age is no greater than 4Δ, the two
thresholds L1n = L2n = 8 and no maintenance is performed during this
period. When the system age is between 5Δ and 12Δ, L1n < 8 and
L2n = 8 . During this period, no maintenance or imperfect maintenance
is determined based on the detected covariate state and the threshold
L1n . Preventive replacement is not required no matter what the cov-
ariate state is. When the system is at the 13th or 14th inspection epoch,
all the three preventive maintenance actions are involved. When the
system age is no less than 15Δ, imperfect maintenance is excluded from
maintenance actions. The system is preventively replaced once the
system age reaches 19Δ.
It is interesting to notice that both thresholds are non-increasing
with system age. Imperfect maintenance is performed when the system
age is between 5Δ and 14Δ. Over this period, the threshold for im-
perfect maintenance L1n is no less than 3, because it is not cost-effective
to perform imperfect maintenance for a low covariate state. Preventive
replacement turns up after the age reaches 13Δ. The threshold for
preventive replacement L2n drops sharply due to the rapidly increasing
baseline hazard rate at old ages. As shown in Fig. 5, the system state
space is divided into three disjoint segments by the two dynamic
thresholds. From the graphical presentation of maintenance policy, a
technician can simply obtain the most effective maintenance action

Fig. 7. Illustration of the proposed policy.

after condition monitoring.

Since previous CBM policies usually consider hazard-rate thresh-
olds, it is then of great interest to investigate the hazard rates of the
obtained thresholds. Fig. 6 shows the hazard rates of L1n , L2n , and all
covariate states for different inspection epochs. As can be seen, the
hazard rates of L1n and L2n (i.e., the obtained hazard-rate thresholds) are
not parallel to the x-axis, which distinguishes our policy from the pre-
vious policies with constant hazard-rate thresholds. Similar to Fig. 5,
three disjoint maintenance segments are also presented in Fig. 6, from
which a maintenance technician can easily find the most appropriate
Fig. 6. Hazard thresholds of the proposed policy. maintenance action.

7
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

Table 1
Average cost rate for parameters b, q, and α.
b Average cost rate q Average cost rate α Average cost rate

1.0 272 0.4 272 0.3 268

1.2 356 1.0 302 0.7 274

Fig. 7(a) illustrates the application of the obtained policy to a si-

mulated covariate trajectory. No maintenance is performed when n ≤
6. At the 7th inspection epoch, the system undergoes imperfect main-
tenance that reduces the covariate value. The system is preventively
replaced when its age reaches 15Δ. The same process is reflected in the
hazard rate domain in Fig. 7(b). The imperfect maintenance taken at 7Δ
is effective in reducing the following hazard rates. It is interesting to
notice that the hazard rates of 9~14th inspection epochs where no
maintenance is performed are greater than that of the 7th inspection
Fig. 9. Optimal threshold for varying q.
epoch where imperfect maintenance is conducted. It also illustrates that
the obtained policy is different from the previous level-crossing policy
with constant hazard-rate thresholds. The MRL for this process is pre-
sented in Fig. 7(c). The imperfect maintenance increases the MRL of the
system and the system is replaced when the MRL value is very small.

4.1. Sensitivity analysis

This section examines the effects of several parameters on the op-

timal maintenance policy. First, we consider the situation where the
covariate process is non-stationary. Since the shape parameter is often
proportional to a power law [30], we set v(t) = 0.0115tb. Table 1 lists
the average cost rates for b = 1 (stationary case) and b = 1.2 (non-
stationary case). In the non-stationary case, the covariate value in-
creases faster over time and thus the optimal average cost rate is much
higher than that of the stationary case. The thresholds for the two si-
tuations are given in Fig. 8. Compared with the stationary case, the
effect of imperfect maintenance in the non-stationary case becomes
more unobvious, which is shown as a compressed imperfect main-
tenance segment.
Then we consider the catastrophic failure probability q. When q
increases, it is more possible to perform a costly corrective replacement Fig. 10. Optimal thresholds for varying α.
which results in the increasing of the average cost rate. The average
cost rates for q = 0.4 and q =1.0 are listed in Table 1 and the
thresholds of the two situations are shown in Fig. 9. When q increases,
both thresholds tend to move leftwards to reduce the high risk of cat- The improvement factor of imperfect maintenance α is another in-
astrophic failure. teresting parameter that can influence the optimal maintenance policy.
As shown in the last two columns of Table 1, a smaller α which indicates
a more powerful imperfect maintenance produces a less average cost
rate. Fig. 10 shows the dynamic thresholds for the situations of α = 0.3
and α = 0.7. When α decreases, the segment area of imperfect main-
tenance is enlarged which confirms the cost-effectiveness of the action.

4.2. Comparison with other policies

To demonstrate the performance of the proposed policy, the section

conducts a comparison with the following CBM policies:
Policy 1: L1n = L2n = Ln for n = 0, 1, ...,N . This policy excludes im-
perfect maintenance from the maintenance policy. One dynamic
threshold [L0, L1, ..., LN]T is set to determine whether to perform pre-
ventive replacement or not.
Policy 2: q = 1 and L1n = L2n = Ln for n = 0, 1, ...,N . In this situa-
tion, the proposed policy reduces to the replacement policy with a
dynamic threshold [L0, L1, ..., LN]T. This policy is similar to the policies
considered in [18–20].
Policy 3: q = 1 and h (n , L1n ) = h (n , L2n ) = L for n = 0, 1, ...,N .
Fig. 8. Optimal thresholds for varying b. In this situation, the proposed policy reduces to the popular

8
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

replacement policies where a dynamic threshold and a constant hazard-

rate threshold is considered, respectively. The comparison between the
two policies shows that the policy with a dynamic threshold is more
cost-effective than that with a constant threshold. Among these policies,
the proposed policy has the minimal average cost rate, which confirms
its superiority over the other CBM policies.

5. Conclusions

This paper has proposed a CBM policy with dynamic thresholds and
multiple maintenance actions for a periodically inspected system sub-
ject to both aging and a continuous-state covariate process. Two failure
types are considered and the failure time is modeled by the PHM. The
proposed policy involves various maintenance actions such as no
maintenance, imperfect maintenance, preventive replacement, minimal
repair, and corrective replacement. Two dynamic thresholds are set to
select actions at inspection epochs. The optimization of the thresholds is
conducted in the SMDP framework. The SMDP quantities are derived
Fig. 11. Optimal thresholds of Policy 1 and Policy 2.
based on a discretization method. A modified policy-iteration algorithm
is developed to find the optimal maintenance policy that minimizes the
long-run average cost rate. A practical numerical example is provided
Table 2 to illustrate the proposed method. Some important conclusions are
Comparison of different maintenance policies summarized as follows: (1) The thresholds for imperfect maintenance
Maintenance policy Control limits Average cost rate and preventive replacement are monotonically non-increasing over
system age; (2) It is necessary to consider multiple maintenance actions
Proposed policy See Figs. 4 or 5 272.3
for effective CBM decision-making; (3) The obtained graphical policy
Policy 1 See Fig. 11 273.4
Policy 2 See Fig. 11 301.5 presentation is supportive in practical maintenance activities; (4) The
Policy 3 L = 1.16 × 10−3 302.9 proposed policy outperforms the other existing CBM policies.
There are still some interesting topics that can be pursued in future
research: (1) The improvement factor of imperfect maintenance can be
random rather than deterministic; (2) The parameters of the covariate
replacement policy with a constant hazard-rate threshold L. As men- process can be unknown in advance and updated by condition mon-
tioned above, such a policy has been widely considered; see for instance itoring [36]; (3) Another interesting extension is to jointly consider the
[9,17]. proposed CBM policy and the lot-sizing for a production system [37].
The proposed algorithm can be used for solving the optimization
problem of Policy 1 and Policy 2. The obtained thresholds are presented CRediT authorship contribution statement
in Fig. 11, and average cost rates are listed in Table 2. Policy 3 can be
optimized by a procedure that includes the following three steps: (1) Rui Zheng: Conceptualization, Methodology, Data curation,
Discretization of a suitable hazard rate region (e.g., [0, 0.02] in this Software, Writing - original draft. Bingkun Chen: Investigation,
paper) to generate candidates of L; (2) Computation of the long-run Writing - review & editing. Liudong Gu: Visualization, Software,
average cost rate according to Eq. (27) for each candidate; and (3) Validation.
Section of the candidate with the minimum of average cost rate. In-
terested readers can refer to [15,17] for more information about the
Declaration of Competing Interest
optimization procedure. The obtained threshold and average cost rate
are shown in Table 2.
It is noted that policy 1 is a special case of the proposed policy by There is no conflict of interest. The manuscript has not been sub-
removing imperfect maintenance. The comparison between the pro- mitted to any other journal or a conference.
posed policy and policy 1 proves that imperfect maintenance is effective
in reducing maintenance cost. Similarly, policy 2 acts as a special case Acknowledgments
of policy 1 by excluding minimal repair. Consequently, the comparison
between policy 1 and policy 2 shows that minimal repair should also be This work was supported by China Scholarship Council (CSC) under
considered while making CBM decisions. Policy 2 and policy 3 are Grant No. 201806090017.

Appendix A: Proof of Eq. (7)

Proof. The element of H(t) can be given by

hij (t ) = Pr( > t + , Xt + = j| > t , Xt = i)
= Pr(Xt + =j| > t + ,Xt = i)·Pr ( > t + | > t ,Xt = i )
= P (t , i),(t + , j)·Pr ( > t + | > t ,Xt = i ) (A.1)

9
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

For a sufficiently small time interval δ, it is reasonable to assume that the covariate state does not change between t and t + δ. Thus, according to
Eq. (2), the second component in Eq. (A.1) is given by
Pr( > t + | > t ,Xt = i)
t+
= exp h 0 (s ) (Xt + s ) ds
t

t+
exp h 0 (s ) (Xt ) ds
t

t+
= exp [(i + 0.5) d]· h 0 (s ) ds
t
(A.2)
Substituting Eq. (A.2) into Eq. (A.1), we can obtain that
t+
hij (t )=P(t , i),(t + , j)·exp [(i + 0.5) d]· h 0 (s ) ds
t
(A.3)
which completes the proof.

Appendix B: Proof of Eq. (8)

Proof. Start the proof from considering the transition between t and t + δ. Using Eq. (7), it is easy to obtain the probability of each covariate state at
the time point t + δ through
o1 = ei · H (n ) (B.1)
where the (j+1)th ( j = 0, 1, ...,M ) element in o1 is the transition probability from i ∈ I to j ∈ I.
Using o1 as the initial state of the time interval between t + δ and t + 2δ, the probability of each covariate state at the end of this interval is
o2 = o1·H (t + ) = ei · H (t )· H (t + ) (B.2)
Recursively, we can obtain the probability of each covariate state at the time point of t + rδ (r = 1, 2, ...) as follows
or = or 1·H (t + (r 1) )
= or 2 ·H (t + (r 2) )· H (t + (r 1) )
= ...
r 1
= ei · H (t + k )
k=0 (B.3)
Then we can obtain the probability that the system survivals to state j ∈ I at t + rδ (r = 1,2,…) given the state i ∈ I at t as follows
r 1
Pr(Xt + r = j, > t + r |Xn = i , > t ) = or · eTj = ei · H (t + k )·eTj
k=0 (B.4)
It is noted that if r = 0, the probability equals to 1 if j = i, and equals to 0 otherwise. Eq. (B.4) holds for this situation which completes the proof.

Appendix C: Proof of the effectiveness of Eq. (28)

Proof. First, the classical policy-iteration algorithm is considered. Since the actions of state (0, 0) and (N, i), where i ∈ I, are deterministic, we just
need to consider the action-selection for the other states. The action set for state (n, i) ∈ S, where n = 1, 2, ...,N 1 and i ∈ I, can be defined as
a(n, i) A = {NM , IM , PR} , where NM, IM, and PR denote no maintenance, imperfect maintenance, and preventive replacement, respectively. Then,

v(n, i) (G, a) = c(n, i) (a) g (G)· (n, i) (a ) + p(n, i ),(n , i ) (a)· v(n , i ) (G)
(n , i ) S (C.1)
defines the relative value of selecting a ∈ A for state (n, i) and thereafter following the existing policy G. In the policy-improvement step, the action
for state (n, i) is determined by
min {v(n, i) (G, a (n , i ))}
a(n, i) A (C.2)
According the policy improvement theorem, a new policy constructed with the actions obtained in Eq. (C.2) is not worse than the current policy G
[35].
Second, consider the policy proposed in this paper. Assume that in the current policy G L1n = k (k = 0, 1, ...,M ). Consider the situation where L2n is
fixed and L1n is moved from k to k + 1. It indicates that only the action of state (n, k) is changed from IM to NM, and the actions of the other states
keep the same as G. Therefore, according to Eq. (28), we have

10
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

M c (n, i) (k + 1, L 2n) g (G)· (n, i) (k + 1, L2n )

+ p(n, i),(n , i ) (k + 1, L 2n)·v (n , i ) (G)
i=0 (n , i ) S

M c (n, i) (k, L2n) g (G)· (n, i ) (k , L 2n)

+ p(n, i),(n , i ) (k, L2n)·v (n , i ) (G)
i=0 (n , i ) S

c (n, k ) (k + 1, L2n) g (G)· (n, k ) (k + 1, L2n)

= + p(n, k ),(n , i ) (k + 1, L2n)·v (n , i ) (G)
(n , i ) S

c (n, k ) (k, L2n ) g (G)· (n, k ) (k, L 2n)

+ p(n, k ),(n , i ) (k, L2n )· v (n , i ) (G)
(n , i ) S

c (n, k ) (NM ) g (G)· (n, k ) (NM )

= + p(n, k ),(n , i ) (NM )· v (n , i ) (G)
(n , i ) S

c (n, k ) (IM ) g (G)· (n, k ) (IM )

+ p(n, k ),(n , i ) (IM )·v (n , i ) (G)
(n , i ) S

= v(n, k ) (G, NM ) v(n, k) (G, IM ) (C.3)

It can be seen that, Eq. (C.3) builds a connection between the proposed algorithm and the traditional policy-iteration algorithm. If Eq. (C.3) is less
than 0, then the action NM performs better than the current action IM does, and thus the threshold L1n = k + 1 is better than L1n = k . Otherwise, the
performance of the threshold L1n = k + 1 is not better than that of L1n = k . Therefore, repeatedly perform the above process for two thresholds, one
can determine the threshold pair (L1n , L 2n) that generates the minimal relative values for each state at the nth inspection. That is to say, Eq. (28)
generates a threshold pair that is no worse than the current policy, which confirms the effectiveness of the policy-improvement step of the proposed
algorithm.

References hazards model. Mech Syst Sig Process 2018;111:285–302.

[18] Tang D, Yu J. Optimal replacement policy for a periodically inspected system
subject to the competing soft and sudden failures. Eksploat Niezawodn
[1] Peng Y, Dong M, Zuo MJ. Current status of machine prognostics in condition-based 2015;17(2):228–35.
maintenance: a review. Int J Adv Manuf Tech 2010;50:297–313. [19] Yahyatabar A, Najafi AA. A multi-stage stochastic programming for condition-based
[2] Alaswad S, Xiang Y. A review on condition-based maintenance optimization models maintenance with proportional hazards model. J Ind Syst Eng 2019;12(1):18–38.
for stochastically deteriorating system. Reliab Eng Syst Saf 2017;157:54–63. [20] Yahyatabar A, Najafi AA. Condition based maintenance policy for series-parallel
[3] Zhang Q, Hua C, Xu G. A mixture Weibull proportional hazard model for mechanical systems through Proportional Hazards Model: A multi-stage stochastic program-
system failure prediction utilising lifetime and monitoring data. Mech Syst Signal Pr ming approach. Comput Ind Eng 2018;126:30–46.
2014;43:103–12. [21] Mousavi SM, Shams H, Ahmadi S. Simultaneous optimization of repair and control-
[4] Cox DR. Regression models and life tables (with discussion). J Roy Stat Soc limit policy in condition-based maintenance. J Intell Manuf 2017;28:245–54.
1972;B34:187–220. [22] Nicolai RP, Dekker R, van Noortwijk JM. A comparison of models for measurable
[5] Makis V, Wu J, Gao Y. An application of DPCA to oil data for CBM modeling. Eur J deterioration: Application to coatings on steel structures. Reliab Eng Syst Saf
Oper Res 2006;174:112–23. 2007;92:1635–50.
[6] Makis V, Jardine AKS. Optimal replacement in proportional hazard model. INFOR: [23] Nicolai RP, Frenk JBG, Dekker R. Modelling and optimizing imperfect maintenance
Inform Syst Oper Res 1992;30(1):172–83. of coatings on steel structures. Struct Saf 2009;31:234–44.
[7] Banjevic D, Jardine AKS, Makis V, Ennis M. A control-limit policy and software for [24] Wu X, Ryan S. Value of condition monitoring for optimal replacement in the pro-
condition-based maintenance optimization. INFOR: Inform Syst Oper Res portional hazards model with continuous degradation. IIE Trans 2010;42:553–63.
2001;39(1):32–50. [25] Banjevic D, Jardine AKS. Calculation of reliability function and remaining useful
[8] Volk PJ, Coetzee JL, Banjevic D, Jardine AKS, Makis V. Optimal component re- life for a Markov failure time process. IMA J Manag Math 2006;17:115–30.
placement decisions using vibration monitoring and the proportional-hazards [26] Zhao S, Makis V, Chen S, Li Y. Evaluation of reliability function and mean residual
model. J Oper Res Soc 2002;53(2):193–202. life for degrading system subject to condition monitoring and random failure. IEEE
[9] Golmakani HR, Fattahipour F. Optimal replacement policy and inspection interval Trans Reliab 2018;67(1):13–35.
for condition-based maintenance. Int J Prod Res 2011;49(17):5153–67. [27] Sen PK. HRQoL and concomitant adjusted mean residual life analysis. In: Nikulin
[10] Tian Z, Lin D, Wu B. Condition based maintenance optimization considering mul- MS, Balakrishnan N, Mesbah M, Limnios N, editors. Parametric and semiparametric
tiple objectives. J Intell Manuf 2009;23:333–40. models with applications to reliability, survival analysis and quality of life. Boston:
[11] Tian Z, Liao H. Condition based maintenance optimization for multi-component Birkhäuser; 2004. p. 349–62.
systems using proportional hazards model. Reliab Eng Syst Saf 2011;96:581–9. [28] Lehmann A. On a degradation-failure model for repairable items. In: Nikulin MS,
[12] Zhao H, Xu F, Liang B, Zhang J, Song P. A condition-based opportunistic main- Balakrishnan N, Mesbah M, Limnios N, editors. Parametric and semiparametric
tenance strategy for multi-component system. Struct Health Monit models with applications to reliability, survival analysis and quality of life. Boston:
2019;18(1):270–83. Birkhäuser; 2004. p. 65–79.
[13] Xu M, Jin X, Kamarthi S, Noor-E-Alam Md. A failure-dependency modeling and [29] Zhao S, Makis V, Chen S, Li Y. Health assessment method for electronic components
state discretization approach for condition-based maintenance optimization of subject to condition monitoring and heard failure. IEEE T Instru Meas
multi-component systems. J Manuf Syst 2018;47:141–52. 2019;68(1):138–50.
[14] Jafari L, Naderkhani F, Makis V. Joint optimization of maintenance policy and [30] van Noortwijk JM. A survey of the application of gamma processes in maintenance.
inspection interval for a multi-unit series system using proportional hazards model. Reliab Eng Syst Saf 2009;94(1):2–21.
J Oper Res Soc 2018;69(1):36–48. [31] Wang J, Makis V, Zhao X. Optimal condition-based and age-based opportunistic
[15] Tang D, Yu J, Chen X, Makis V. An optimal condition-based maintenance policy for maintenance policy for a two-unit series system. Comput Ind Eng 2019;134:1–10.
a degrading system subject to the competing risks of soft and hard failure. Comput [32] Elwany AH, Gebraeel NZ, Maillart LM. Structured replacement policies for com-
Ind Eng 2015;83:100–10. ponents with complex degradation processes and dedicated sensors. Oper Res
[16] Liu B, Liang Z, Parlikad AK, Xie M, Kuo W. Condition-based maintenance for sys- 2011;59(3):684–95.
tems with aging and cumulative damage based on proportional hazards model. [33] Liu B, Wu S, Xie M, Kuo W. A condition-based maintenance policy for degrading
Reliab Eng Syst Saf 2017;168:200–9. systems with age- and state-dependent operating cost. Euro J Oper Res
[17] Duan C, Makis V, Deng C. An integrated framework for health measures prediction 2017;263:879–87.
and optimal maintenance policy for mechanical systems using a proportional [34] Liu B, Xie M, Xu Z, Kuo W. An imperfect maintenance policy for mission-oriented

11
R. Zheng, et al. Reliability Engineering and System Safety 204 (2020) 107123

system subject to degradation and external shocks. Comput Ind Eng imperfect information and the proportional hazards model. Int J Prod Res
2016;102:21–32. 2007;45(4):989–1012.
[35] Tijms HC. Stochastic models: an algorithmic approach. New York: John Wiley & [37] Jafari L, Makis V. Joint optimal lot sizing and preventive maintenance policy for a
Sons; 1994. production facility subject to condition monitoring. Int J Prod Econ
[36] Ghasemi A, Yacout S, Ouali MS. Optimal condition based maintenance with 2015;169:156–68.

12