Reliability Engineering and System Safety 144 (2015) 83–94

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

Multi-level predictive maintenance for multi-component systems

Kim-Anh Nguyen a, Phuc Do b,c,n, Antoine Grall a
a
ICD, ROSAS, LM2S, Université de Technologie de Troyes, UMR 6281, CNRS, Troyes, France
b
Université de Lorraine, CRAN, UMR 7039, Campus Sciences, BP 70239, Vandoeuvre-les-Nancy 54506, France
c
CNRS, CRAN, UMR 7039, France

art ic l e i nf o a b s t r a c t

Article history: In this paper, a novel predictive maintenance policy with multi-level decision-making is proposed for
Received 11 April 2015 multi-component system with complex structure. The main idea is to propose a decision-making process
Received in revised form considered on two levels: system level and component one. The goal of the decision rules at the system
16 July 2015
level is to address if preventive maintenance actions are needed regarding the predictive reliability of the
Accepted 17 July 2015
system. At component level the decision rules aim at identifying optimally a group of several
Available online 28 July 2015
components to be preventively maintained when preventive maintenance is trigged due to the system
Keywords: level decision. Selecting optimal components is based on a cost-based group improvement factor taking
Multi-component system into account the predictive reliability of the components, the economic dependencies as well as the
Condition-based maintenance
location of the components in the system. Moreover, a cost model is developed to find the optimal
Opportunistic
maintenance decision variables. A 14-component system is finally introduced to illustrate the use and
Economic dependencies
Prognostic the performance of the proposed predictive maintenance policy. Different sensitivity analysis are also
Importance measure investigated and discussed. Indeed, the proposed policy provides more flexibility in maintenance
decision-making for complex structure systems, hence leading to significant profits in terms of
maintenance cost when compared with existing policies.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction decision-making process relies on the diagnostic/prognostic of the

The maintenance process includes preventive and corrective
actions carried out to retain a system or restore it to an operating
condition. It is a complex process of the utmost importance especially
for the manufacturing firms. Maintenance costs may take from 15 to
70% of total production costs [1]. Optimal maintenance policies aim to
provide optimal system reliability/availability and safety performance
at the lowest possible maintenance cost [21]. In the literature, there
are mainly two types of maintenance techniques including time-based
maintenance (TBM) and condition-based maintenance (CBM). For
TBM maintenance, preventive maintenance decision is based on the
system age and on the knowledge of the statistical information on the
system lifetime [6,9]. As a consequence, the realistic operating
conditions of the system over time are not taken into account. Unlike
TBM maintenance, CBM consists in advanced maintenance technique.
It is popularly reported in the literature and the maintenance
Abbreviation: CBM, condition-based maintenance; CM & OM & PM, corrective and
opportunistic and preventive maintenance; MCS, minimal cut set; NED & PED,
negative and positive economic dependence; RBD, reliability block diagram; RUL,
remaining useful life
n
Corresponding author.
E-mail address: phuc.do@univ-lorraine.fr (P. Do).
system condition over time. It has been recently introduced and
becomes nowadays an interesting approach for maintenance optimi-
zation [2,3,7,8,12].
A large number of CBM policies has been investigated, developed
and successfully applied to mono-component system, e.g., [7,12,30]
using the current equipment condition (deterioration level) and
[15,31,8] taking the future equipment health state for making main-
tenance decision. However, these policies cannot be adapted directly
to multi-component systems [4], in which interdependencies such as
stochastic, structural and economic dependencies may exist between
components [25]. In fact, the stochastic dependence implies that the
deterioration process or failure of a component may affect the lifetime
distribution of other ones. The structural dependence exists if com-
ponents structurally form a part, so that maintenance of a failed
component implies maintenance of working components, or at least
dismantling them. Finally, the economic dependence involves
(i) positive economic dependence (PED) implying that joint main-
tenance of a group of components is cheaper that performing main-
tenance on components separately, and (ii) negative economic depen-
dence (NED) which exist if combining maintenance activities is more
expensive than performing maintenance on components separately
[6]. It is important to note that taking into account all these
dependencies in maintenance optimization is necessary, but makes

http://dx.doi.org/10.1016/j.ress.2015.07.017
0951-8320/& 2015 Elsevier Ltd. All rights reserved.
84 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94
Nomenclature
Notations
n total number of components of considered system
i index for components of the system, with i ¼ 1; 2; …; n
s(t) system state at time t
si(t) state of component i at time t
X i;t ¼ xi;t degradation level of component i measured at time t
X 1:n;t ¼ x1:n;t degradation level of whole system measured at
time t
ðX i;t Þt Z 0 stochastic process describing the degradation of com-
ponent i over time t
αi ; β i shape and scale parameters of Gamma distribution for
component i
Zi fixed failure threshold of component i
cp;i ; cc;i specific PM cost and specific CM cost of component i,
with cc;i Z cp;i
cinsp inspection cost for a component
cs maintenance set-up cost
cpsd planned shutdown cost
cusd unplanned shutdown cost
cd unplanned downtime cost rate
T; T n inter-inspection interval (or inspection cycle) and
optimal value of T
Tk k-th inspection time, with T k ¼ k T and k A N
T k ; T kþ time just before/after the inspection date Tk
Gl l-th group of components
j Gl j cardinal number of the group Gl, 1 r j Gl j r n
Ri ðT k þ 1 j xi;T k Þ predictive reliability of component i at the next
inspection time T k þ T given its degradation level, xi;T k ,
at time Tk
the maintenance models too complicated to solve or to analyze. This is
specially true in CBM framework. In the literature most existing CBM
maintenance policies for multi-component systems take advantage of
PED to minimize the average maintenance cost, see for instance
[2,3,13,26,16]. Nevertheless, these existing CBM policies addressing
multi-component systems can be applied on several specific struc-
tures (i.e., series, parallel, series-parallel, parallel-series, k-out-of-n).
However, actual system structures are more complex with a large
number of components and very complex inter-connections between
components which can be described as mixtures of basic connections
[11,28,33]. It is shown in [29] that taking into account the complex
structure in a maintenance model may lead to consider both PED and
NED dependencies. However, the maintenance policy proposed in [29]
is based on TBM technique. Recently, a CBM policy has been
introduced for complex structure systems using Birnbaum importance
measure which can help to take into account effectively the system
structure in maintenance decision-making [19]. Each component is
individually selected to be preventively maintained at discrete times
but the maintenance grouping impacts (PED and/or NED between
components) are not really highlighted.
To face this issue, in the present paper, a predictive maintenance
policy with multi-level decision-making process is proposed for
complex structure systems. The main idea is to jointly consider
prognostic information (i.e., predictive reliability or RUL) of compo-
nents and system, PED and NED dependencies as well as system
structure in the maintenance decision-making process. More precisely,
the proposed maintenance policy is divided into two levels:

At system level and at each inspection time, the reliability of
components is firstly predicted. Based on the predictive relia-
bility of components, the predictive reliability of the system is
secondly determined and compared with a prefixed reliability
RðT k þ 1 j x1:n;T k Þ predictive reliability of system at the next
inspection time T k þ T given the degradation level of
n components, x1:n;T k , at time Tk
RGl ðT k þ 1 j x1:n;T k Þ system predictive reliability at the next inspec-
tion time T k þT when all components of group Gl are
replaced preventively at time Tk
C GIF ðtÞ cost-based group improvement factor computed
Gl
when all components of group Gl are replaced at time
t
R0 ; Rn0 PM threshold and optimal value of R0 defined for the
policy (T; R0 )
Rp ; Rnp PM threshold and optimal value of Rp defined for the
policy (T; Rp ; Rop )
Rop ; Rnop OM threshold and optimal value of Rop defined for the
policy (T; Rp ; Rop )
Nb total number of inspection times within ½0; T rep 
Ck consists of the inspection cost, the running cost and
the maintenance costs of the k-th inspection cycle (i.e.
ðT k  1 ; T k , with k Z1)
C 1 ðT; R0 Þ long run maintenance cost per time unit of the policy
(T; R0 )
C 1 ðT; Rp ; Rop Þ long run maintenance cost per time unit of the
policy (T; Rp ; Rop )
Trep time at which the system is completely replaced
I indicator function: I fg ¼ 1 if the condition fg is true
and 0 otherwise
Ω set of all components of the system
Policy (T; R0 ) Multi-level predictive maintenance policy
Policy (T; Rp ; Rop ) opportunistic predictive maintenance policy
threshold. Preventive maintenance is needed if the predictive
reliability of the system is lower than the prefixed threshold.

At component level: the main idea is to find an optimal
grouping of several components to be preventively maintained
when maintenance decision at system level is triggered. Deci-
sion making is made on the basis of a “cost-based group
improvement factor” taking into account the predictive relia-
bility of components, the economic dependencies as well as the
location of components in the system. The later is the first
original contribution of the paper.
The proposed maintenance policy has two decision variables
including inter-inspection interval and PM threshold at system level.
These decision variables need to be optimized. The second contribu-
tion of the paper is to develop a cost model in order to find the opti-
mal maintenance decision variables. In addition, Monte Carlo simula-
tion technique is used to evaluate the maintenance cost rate.
The rest of this paper is organized as follows. Section 2 is devoted
to the description of the general model of degradation and failure of a
multi-component system. The complex structure of the system is also
investigated. Section 3 focuses on the predictive reliability/RUL at both
component and system level for a complex structure system. Addi-
tionally, the cost-based group improvement factor is also introduced.
Section 4 describes the maintenance costs structure and the proposed
multi-level predictive maintenance policy. The global cost model is
also developed here. To illustrate the use and advantage of the
proposed policy, a 14-component system with complex structure is
introduced. A comparison on the performance of the proposed policy
and an existing policy is investigated through different sensitivity
analyses. Finally, the present paper ends with some conclusions and
outlines for possible future works.
 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94
2. System degradation modeling and assumptions
2.1. General degradation process description
In this work, we consider a complex multi-component engineering
system comprised of n deteriorating components. The components
can be identical or not. They are inter-connected according to a
complex structure/configuration with components interdependencies
described, e.g., in term of reliability block diagram (RBD) [24]. The
configuration is denominated as “complex” because it can be a
mixture of several commonly basic connections (e.g. series, parallel,
k-out-of-n, bridge).
Each component in the system is assumed to be subject to an
underlying degradation process which can cause random failures.
Such a degradation processes may be a physical degradation processes
such as accumulative wear, corrosion, fatigue and crack growth which
are major causes of wear-out failures [5]; or it may be an artificial
process describing the phenomenon of the health status or the
performance of a component worsening with usage or age [32].
To model the degradation process of the considered system, in this
paper, a time-dependent stochastic process is considered. It has more
advantageous in describing intermediate degradation states than
rough lifetime models. To this end, the following general assumptions
related to degradation of component i (i ¼ 1; …; n) are given:

The degradation evolution of each component is gradual, with
stochastically independent increments over time. The degradation
increments, Δxi , are supposed to be stationary and non-negative.

(
The components in the system are structurally dependent
considering the viewpoint of the RBD model and economically
dependent from the maintenance viewpoint.

The degradation process of each component is stochastically
independent from the others.

The condition (degradation level) of component i at time t can
be measured and described as a scalar random aging variable
X i;t . The degradation trajectory is a path of the time-dependent
stochastic process ðX i;t Þt Z 0 .

The initial degradation level of component i is X i;0 ¼ 0. The
component i is considered initially as new. The higher X i;t , the
closer component i is to failure.

If X i;t ¼ xi;t exceeds a predefined critical threshold Zi, compo-
nent i is considered as failed, even if it is still operating. Zi can
be seen as a limit degradation level defined according to
economical, technical or security reasons. The failure of an
item is assumed to be self-announcing and a failed component
is stopped as soon as possible.

The stoppage of a component does not affect the production
performance if the system is still functioning.
For a complex structure system, failure of one or several compo-
nents does not necessarily lead to a failure of the system, but if the
failed components are not immediately replaced, they may con-
duct some other components to idle state (see [19] for a complete
description). During the idle state, components may deteriorate in
a different way than in operating state. In this paper the deteriora-
tion of a component is supposed to stop as far as the component
remains in idle state. This phenomenon introduces dependencies
between the components deterioration processes.
2.2. Complex structure: definitions and analysis
To tackle with the complexity of the system configuration the
following concepts are recalled, [22]:

Critical component: a component is called “critical” if its failure or
stoppage, while the other components are in operating state,
85
leads the system to failure or shutdown. Oppositely, a compo-
nent is“non-critical” if the system can be still operating when it
fails or stops. It is important to note that in a complex structure a
“non-critical” component may become critical in a period of
time when other components are failed [19]. In such a case it is
called a “temporary critical” component.

Cut set: a cut set is a set of components for which the system
fails if all components in cut set fails.

Minimal cut set (MCS): a cut set is said to be minimal if it cannot
be reduced without loosing its status as a cut set.

Critical group: a group is “critical” if the system fails when all
components of the group fail. According to the concept of MCS,
a critical group contains at least one MCS. Oppositely, a group is
defined as “non-critical” if the system can still operate when all
the components of the group have failed.

Path set: a path set is a set of components whose operation
ensures that the system is functioning.

Minimal path set: a path set is said to be minimal if it cannot be
reduced without loosing its status as a path set.

Structure function is a binary function, denoted ϕðsðtÞÞ, which
allows to determine the state of the system from the given state
of all components in the system. We have
(
1 if the system is functioning at time t
ϕðsðtÞ ¼ ϕðs1 ; s2 ; …; sn Þ ¼
0 if the system is in a failed state at time t
where sðtÞ ¼ ðs1 ðtÞ; s2 ðtÞ; …; sn ðtÞÞ is a state vector; si ðtÞ
ði ¼ 1; …; nÞ is binary variable representing the state of the
component i
1 if component i is functioning at time t
si ðtÞ ¼
0 if component i is in a failed state at time t
It is important to note that si ðtÞði ¼ 1; …; nÞ are independent
since all components are stochastically independent.

Structure function of the complex structures can be obtained by
using the minimal cut sets concept or the minimal path sets
one
○ Using the minimal path sets
np
!
ϕðsðtÞÞ ¼ 1  ∏ 1  ∏ si ðtÞ ð1Þ
j¼1 i A Pj
where Pj is j-th minimal path set, and np is the total number
of minimal path sets of the system;
○ Using the minimal cut sets
" #
nc
ϕðsðtÞÞ ¼ ∏ 1  ∏ ð1  si ðtÞÞ ð2Þ
j¼1 i A Cj
where Cj is j-th minimal cut set, and nc is the total number of
minimal cut sets of the system.
As an example, let us consider a complex structure system whose
reliability block diagram (RBD) is represented in Fig. 1. It should be
noted that this system is a part of a natural gas plant as described
in [11, p. 178]. The system is composed by 8 minimal path sets
including P 1 ¼ f1; 3; 4; 5; 6; 7; 8; 9; 10; 11g, P 2 ¼ f2; 3; 4; 5; 6; 7; 8; 9;
10; 11g, P 3 ¼ f1; 3; 4; 5; 6; 14g, P 4 ¼ f2; 3; 4; 5; 6; 14g, P 5 ¼ f1; 3; 4; 13g,
P 6 ¼ f2; 3; 4; 13g, P 7 ¼ f1; 3; 12g, P 8 ¼ f2; 3; 12g and 10 minimal cut
sets including C 1 ¼ f1; 2g, C 2 ¼ f3g, C 3 ¼ f4; 12g, C 4 ¼ f5; 13; 12g,
C 5 ¼ f6; 13; 12g, C 6 ¼ f7; 14; 13; 12g, C 7 ¼ f8; 14; 13; 12g, C 8 ¼ f9; 14;
13; 12g, C 9 ¼ f10; 14; 13; 12g, C 10 ¼ f11; 14; 13; 12g.
The structure function of the system in this case is represented by

using the minimal path sets
" #
8
ϕðsðtÞÞ ¼ 1  ∏ 1  ∏ si ðtÞ
j¼1 i A Pj
86 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94

4 5 6 7 8 9 10 11
1

3 14
Start End
13
2
12

Fig. 1. Reliability block diagram of the 14-component system.

¼ 1  ½1 s1 ðtÞs3 ðtÞs4 ðtÞs5 ðtÞs6

[t 1 ; t 2 ] (with t 1 o t 2 ) is written as
ðtÞs7 ðtÞs8 ðtÞs9 ðtÞs10 ðtÞs11 ðtÞ
! !
½1  s2 ðtÞs3 ðtÞs4 ðtÞs5 ðtÞs6 ðtÞs7 ðtÞs8 np np

ðtÞs9 ðtÞs10 ðtÞs11 ðtÞ½1 s1 ðtÞs3 ðtÞs4 ðtÞs5 Rðt 2 j t 1 Þ ¼ 1  ∏ 1  ∏ Ri ðt 2 j t 1 Þ ¼ 1 ∏ 1  ∏ Ri ðt 2 j t 1 Þ ;

j¼1 i A Pj j¼1 i A Pj
ðtÞs6 ðtÞs14 ðtÞ½1  s2 ðtÞ
ð7Þ
s3 ðtÞs4 ðtÞs5 ðtÞs6 ðtÞs14 ðtÞ½1  s1 ðtÞs3 ðtÞs4 ðtÞs13
  ðtÞ½1  s2 ðtÞs3 ðtÞs4 ðtÞs13 ðtÞ½1  s1 ðtÞs3 ðtÞs12 ðtÞ where Ri ðt 2 j t 1 Þ is the conditional reliability of component i given it
½1  s2 ðtÞs3 ðtÞs12 ðtÞ; ð3Þ is functioning at time t1. If the deterioration level of the compo-
nent is measured at time t1 such that X i;t1 ¼ xi;t 1 , Ri ðt 2 j t 1 Þ can be

using the minimal cut sets stated as follows:
" #
10 Ri ðt 2 j xi;t 1 Þ ¼ PðX i;t2 o Z i j X i;t 1 ¼ xi;t1 Þ: ð8Þ
ϕðsðtÞÞ ¼ ∏ 1  ∏ ð1 si ðtÞÞ
j¼1 i A Cj
If the deterioration process of component i is non decreasing,
¼ ½1  ð1  s1 ðtÞÞð1  s2 ðtÞÞ½1 ð1 s3 ðtÞÞ Ri ðt 2 j xi;t 1 Þ can then be estimated by
½1  ð1 q4 ðtÞÞð1  s12 ðtÞÞ
Z Z i  xi;t1
½1  ð1 s5 ðtÞÞð1  s12 ðtÞÞð1  s13 ðtÞÞ
Ri ðt 2 j xi;t 1 Þ ¼ PðX i;ðt2  t 1 Þ o Z i  xi;t1 Þ ¼ f i ðuÞ du: ð9Þ
½1  ð1 s6 ðtÞÞð1  s12 ðtÞÞð1  s13 ðtÞÞ 0
½1  ð1 s7 ðtÞÞð1  s12 ðtÞÞð1  s13 ðtÞÞ
Finally, the degradation-based conditional reliability of the system
ð1  s14 ðtÞÞ½1  ð1  s8 ðtÞÞð1  s12 ðtÞÞ
comprised of n components, denoted Rðt 2 j x1:n;t1 Þ, can be obtained
ð1  s13 ðtÞÞð1  s14 ðtÞÞ½1  ð1  s9 ðtÞÞ
by introducing Eq. (9) into Eq. (7).
ð1  s12 ðtÞÞð1  s13 ðtÞÞð1  s14 ðtÞÞ
At each inspection time the predictive system reliability will be
½1  ð1 s10 ðtÞÞð1  s12 ðtÞÞð1  s13 ðtÞÞ
used as an essential indicator for PM decision making for the
ð1  s14 ðtÞÞ½1  ð1  s11 ðtÞÞð1  s12 ðtÞÞ proposed predictive maintenance policy. This point will be dis-
ð1  s13 ðtÞÞð1  s14 ðtÞÞ: ð4Þ cussed in Section 4.

3.2. Cost-based group improvement factor

3. Reliability prediction and cost-based group improvement
factor In order to improve the reliability of the system, the reliability
improvement importance measure, defined as the reliability
3.1. Reliability prediction difference of system when a component is replaced by a new
one [22], seems to be an interesting indicator. The latter can help
Mathematically, for non repairable components the reliability to identify the importance ranking of components relating to the
of the system, denoted R(t), is expressed as improvement ability on the system reliability. It must however be
noted on one hand that this importance measure does not take
RðtÞ ¼ E½ϕðsðtÞÞ; into account the maintenance cost. From a practical point of view,
a high improvement ability with very high maintenance cost may
where E½ indicates the mathematical expectation. As mentioned above, not be a proper choice for maintenance optimization. A “reason-
ϕðsðtÞÞ can be determined by using the minimal path set concept or the able” improvement ability with lower maintenance cost may be
minimal cut set one. Without loss of generality, we assume that ϕðsðtÞÞ is more appropriate. On the other hand, due to economical and/or
given by Eq. (1), the reliability of the system is expressed then technical reasons, it is preferable that several components are
preventively maintained at each maintenance time [9,10]. Never-
! !
np np theless, the existing reliability improvement measure does not
RðtÞ ¼ 1  ∏ 1  ∏ Eðsi ðtÞÞ ¼ 1 ∏ 1  ∏ Ri ðtÞ ; ð5Þ take into account the simultaneous replacement impacts of several
j¼1 i A Pj j¼1 i A Pj
components together, e.g. economic dependence between compo-
nents or joint improvement impacts. Therefore, in this paper an
where Ri ðtÞ ¼ Eðsi ðtÞÞ is the reliability of component i. Since component i extension of this importance measure, namely “cost-based group
gradually deteriorates, its reliability can be expressed as follows: improvement factor” is proposed for group-based maintenance
Z Zi
decision-making.
Ri ðtÞ ¼ 1  PðX i;t Z Z i Þ ¼ PðX i;t oZ i Þ ¼ f i ðuÞ du; ð6Þ The cost-based group improvement factor for a group Gl is
0 denoted C GIF
Gl . It is defined as follows:

where f i ðÞ is the probability density function describing the deterioration RGl ðt 2 j t 1 Þ  Rðt 2 j t 1 Þ
evolution of component i. C GIF
Gl ðt 1 Þ ¼ ; ð10Þ
C p;Gl
Let us now assume that the system is known to be functioning
at time t1, the reliability of the system within the interval horizon where:
 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94
Predict R(Tk 1 | x1:n,Tk ) with assuming that
all (n - m) failed components are already replaced
No PM No
R(Tk
action is 1 | x1:n,Tk ) R0 ?
needed
Yes
Gl
Assess R (Tk 1 | x1:n,Tk ) with l 1,...,2 m-1
Gl is not No Gl
R (Tk 1 | x1:n,Tk ) R0 ?
considered
Yes
Gl is admissible group
Calculate CGGIF (Tk ) for all admissible groups
l
Define the optimal g roup Gk* such that
CGGIF
* (Tk ) = max CGGIF
l
(Tk )
k l 1,...,|S adm |
Remplace preventively the optimal group
Fig. 2. Procedure of multi-level preventive maintenance decision-making.

RG ðt 2 j t 1 Þ indicates that the system predictive reliability calcu-
l
lated at time t2 when all the components of the group Gl are
preventively replaced at time t1.

C p;Gl is the cost related to PM operation performed on the group
Gl and described in detail in Section 4.
By definition, C GIF
Gl ðt 1 Þ
takes jointly into account various informa-
tion related to: the predictive reliability behavior of all components
of the system, the costs (e.g. PM cost of maintained components,
positive/negative economic dependencies), the system structure
(degradation dependencies between components). The use of the
proposed C GIF
Gl factor for selecting an optimal group of components
to be preventively maintained will be described in the next section.
4. Multi-level predictive maintenance policy
4.1. Assumptions associated with inspection, maintenance and costs
Inspection operation: We assume that the degradation level of
functioning components can only be known by periodic inspection
operations at times T k ¼ kT, where T is the inter-inspection
interval and k A N. The inspection operations are assumed to be
instantaneous, perfect and non-destructive. For each functioning
component, an inspection action incurs a cost cinsp.
Maintenance operations: CM activity is applied to failed com-
ponents and PM activity is carried out on running components. In
this work, both maintenance activities can completely restore a
component to “as good as new” state and the duration related to
these activities is assumed to be negligible. We assume also that
necessary maintenance resources (such as spare parts and crews,
and maintenance tools) to execute the maintenance actions are
always available at planned inspection dates.
A PM intervention on a specific component/group of compo-
nents may lead the system to a planned shutdown. Although the
maintenance durations are assumed to be negligible, a planned
shutdown usually incurs a planned shutdown cost cpsd. If the
system has stopped due to the failure of a component, an
87
unplanned shutdown cost cusd, is incurred (generally, cusd Z cpsd ).
Moreover, when the system remains in its failure state, it is
System level
necessary to pay an additional cost (unavailability cost rate) cd
for each time unit. In that way, the cost structures which applied in
this paper are given hereafter.
Within the paper I fg denotes the indicator function with I fg ¼ 1
if fg is true; otherwise I fg ¼ 0.

When replacing running component i at time t, a PM cost C p;i , is
incurred such that
C p;i ¼ cp;i þ cs þ cpsd I fComponent i is critical or temporary criticalg ð11Þ
|fflfflfflffl{zfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
unitary PM cost planned shutdown cost
Component level
where cp;i is the specific PM cost; cs the maintenance set-
up cost.

When replacing failed component i at time t given that the
component has failed at time tiF (with 0 ot iF r t), a corrective
cost C c;i , is incurred such that
C c;i ¼ cc;i þ cs þ ½ðt  T F Þcd þ cusd I fti ¼ T F g ð12Þ
|fflfflfflffl{zfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
F
unitary CM cost unavailability cost
where cc;i is the unitary specific CM cost for component i; The
unavailability cost is incurred if the system failure is due to the
Gk* failure of component i i.e. if t iF ¼ T F with TF indicating the failure
time of the system.

Assume the system has not failed and all the components in Gl
are working but have to be preventively replaced. By replacing
preventively group Gl of elements j Gl j (1 r j Gl j r n) at the same
time t, a PM group cost C p;Gl , is incurred
X
C p;Gl ¼ C p;i  cs ðj Gl j  1Þ
i A Gl
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
0
saved setupcost:PED
1
X
 cpsd @ I fcomponent i is critical or temporary criticalg  I fGl is a critical groupg
A
i A Gl
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
savedplannedshutdowncost:PEDorNED
ð13Þ
The second part on the right-hand side of Eq. (13) represents the
saved set-up cost due to positive economic dependence (PED)
between components. The third part could be positive or negative.
In fact, it is positive (i.e., PED) if the group Gl contains more than
one critical and/or temporary critical component. If Gl is a critical
group but its components are non-critical one, the third part is
negative and represents negative economic dependence (NED). It is
important to note that the third part does not exist in multi-
component series systems since all components of a series system
are critical one.
4.2. Description of multi-level predictive maintenance policy
The surviving components of the system are regularly inspected at
times T k ¼ k T (k A N). The inspection operation performed on com-
ponent i at time Tk returns its current condition X i;T k ¼ xi;T k . The
corrective replacement of failed components and the preventive
replacement of running components are only carried out at inspection
times Tk. T is the first decision variable.
The main intention of the proposed predictive maintenance
policy is to jointly take into account in the decision process the
predictive reliability of the components and the system, the
system structure as well as the economic dependencies between
components. For this purpose a two-level PM decision process is
proposed at both system and component level as shown in Fig. 2.

At the system level, the reliability of the system before the next
inspection time, RðT k þ 1 j x1:n;T k Þ, is firstly predicted considering
that all failed components during ðT k  1 ; T k  have been already
88 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94

replaced. It is then compared with a reliability threshold R0 The cost-rate can be generally defined as
(0 o R0 r1) to trigger PM actions, where R0 is a decision Cðut Þ
variable. The decision rule at system level is shown in Fig. 2. C 1 ¼ lim : ð16Þ
t-1 ut
If PM action is necessary, the remaining issue is to find an
“optimal” group of components to be preventively maintained. On the basis of the renewal theory [23,27], the cost-rate in Eq.
The latter can be done at component level. (16) can be revaluated by

At the component level, the PM decision rules are based on the E½CðH 1 Þ
cost-based group improvement factor which allows to select C1 ¼ ; ð17Þ
E½H 1 
the group of components to be maintained. If at time Tk, m
components are working, the number of possible groups is where H1 is the usage length of the first renewal cycle, i.e. of the
then 2m  1. A group is said to be admissible if the replacement first life cycle of the system. To illustrate how to compute the
of all the components of the group can upgrade the system mean maintenance cost rate, one supposes that the system is
predictive reliability above R0. Based on the set of all the completely replaced at H 1 ¼ T rep , which means that all compo-
admissible groups, denoted Sadm, an optimal group is deter- nents of the system are replaced at the same time Trep. Since, the
mined as the one having the highest value of the cost-based maintenance cost of the whole system till time Trep, CðT rep Þ, is the
group improvement factor C GIF n sum of all partial costs of N b inspection cycles within time interval
 ðT k Þ. Gk denotes the optimal P b
group that is obtained at inspection date Tk. It is then such that ½0; T rep , i.e. CðT rep Þ ¼ N C . It is noted that the time considered
k¼1 k
n o in the denominator of Eq. (17) is total operating time of the system
C GIF
Gn ðT k Þ ¼ maxm
C GIF Gl
Gl ðT k Þj R ðT k þ 1 j x1:n;T k Þ 4 R0 : ð14Þ [20]. So, the cost-rate can be rewritten as
k l ¼ 1;‥2  1
hP i
Nb
Gl
R ðT k þ 1 j x1:n;T k Þ represents the predictive reliability of the E k ¼ 1
Ck
C1 ¼ ; ð18Þ
system after PM on group Gl. Due to the existence of PED E½T rep  DðT rep Þ
between components in the system, the set-up cost cs is
assigned only once for all maintenance activities (including where DðT rep Þ is total downtime of the system during its mission
PM and CM) performed at the same time Tk. Therefore, cs is not (i.e. total unavailability time) and T rep  DðT rep Þ is the usage dura-
added to C p;Gl if there is at least one failed component j that tion until time Trep.
must be correctively replaced at Tk. T F;k is a random variable Ck is the inspection cost, the running cost and the maintenance
describing the failure time of the system in the k-th inspection costs of the k-th inspection cycle (i.e. in time interval ðT k  1 ; T k )
cycle. If the failure of the system occurs at time T F;k r T k and gathered at inspection time Tk
after inspection time T k  1 , i.e. T F;k 4 T k  1 , then cpsd is not also C k ¼ C insp þ C corr þ C prev ð19Þ
k k k
added to C p;Gl . As a result, the PM cost related to the group Gl
can be expanded from Eq. (13) as follows: where
0 1
X
@ P A

Cinsp
k is the cost related to inspection operations on the surviv-
C p;Gl ¼ C p;i  cs j Gl j  I f I fx Z Z g ¼ 0g
i A Gl j A Ω⧹Gl
j;T k j ing components at time Tk.
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Ccorr
k is the cost related to the replacement of failed components
saved setup cost: PED
2 3 at time Tk (i.e. it includes the set-up costs and the system's
X unplanned shutdown cost).
 cpsd 4 Iipsd  I fGl is a critical groupg ∏ I ft j a T F;k g
5; ð15Þ
Cprev is the cost related to the preventive replacement of the
k
i A Gl j A Ω⧹Gl F

|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} optimal group of running components Gnk at time Tk (i.e. it includes

saved planned shutdown cost: PED or NED the set-up costs and the system's planned shutdown cost).
where Ω⧹Gl is the set of all the components outside the group
Gl and Since inspection operation is only performed on the surviving
○ I fP I fx Z Z g ¼ 0g
¼ 0 if there is not any failed component components up to time Tk, thus, the inspection cost at Tk is
j A Ω⧹Gl j;T k j
calculated as
0 1
in Ω⧹Gl , 0 otherwise
X
○ ∏j A Ω⧹Gl I ft j a T g ¼ 1 if the system does not fail in the k-th insp @
C k ¼ cinsp n  I fxj;T Z Z j g A: ð20Þ
F F;k k
j A Ω⧹Gnk
inspection cycle is due to the failure of a component from
For the corrective replacement cost Ccorr k , if there is no failed
Ω⧹Gl , 0 otherwise.
component between (T k  1 ; T k ], then C corr
k ¼ 0. Otherwise, on the
basis of Eq. (12), Ccorr is calculated by
Finally, the factor C GIF
Gl ðT k Þ can be obtained by including Eq. (15) X
k

into Eq. (10). C corr

k ¼ ðcc;j þcs ÞI fxj;T Z Z j g
k
The maintenance policy described previously will be hereafter j A Ω⧹Gnk
referred as “policy ðT; R0 Þ”. It is important to note that the inter- 0 1
X  
inspection interval T and the PM threshold R0 are two decision  cs @ I fxj;T A þ ðT k  T F;k Þcd þcusd I j
Z Zj g  1 ft F ¼ T F;k g
;
k
variables supporting the model. They have to be sought in order to j A Ω⧹Gk n |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
reach the optimum value of the objective function which is repre- |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} unavailability cost in ðT k  1 ;T k 
saved setup cost at T k : PED
sented by the long run maintenance cost per unit time. Therefore the
ð21Þ
next subsection is devoted to the cost model development.
In the same manner, the PM cost Cprev k is determined as follows:
4.3. Optimization for the proposed maintenance policy if no PM action is required at time Tk, then C prev k
¼ 0. Otherwise, on
the basis of Eq. (11), Cprev k is computed as
We now develop a mathematical expression of long run X  
C prev ¼ C p;i  cs j Gnk j I fNo failed component at T k g
maintenance cost per unit time (or cost-rate). Let Cðut Þ be the k
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
i A Gn k
cumulative cost function of the whole system till usage ut at time t. saved setup cost at T k : PED
 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94 89

2 3
X   5. Numerical example
cpsd 4 Iipsd  1  I ft j ¼ T F;k g
I fGnk is a critical groupg
5: ð22Þ
F
i A Gnk In this section, we show how the proposed predictive main-
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
saved planned shutdown cost at T k : PED or NED tenance policy can be used in PM optimization of complex systems
through a 14-component system whose structure is sketched in
Finally, from Eqs. (18), (20)–(22), the cost-rate of the system Fig. 1. This structure could have been any combination of series,
can be calculated as follows: parallel, series-parallel and parallel-series configuration, or even
1 more generally any structure for which the system reliability can
C 1 ðT; R0 Þ ¼ h P b i
E T rep  N ðT  T F;k Þ be computed from the components reliability.
k¼1 k
8 0 0 1 In order to highlight the effectiveness of the proposed multi-
<X Nb X X level maintenance policy (i.e. policy ðT; R0 Þ), a comparison is
E @cinsp @n  I fxj;T Z Z j g A þ C p;i
: k proposed with another opportunistic maintenance policy. The
k¼1 j A Ω⧹Gnk i A Gnk
0 1 policy (T; Rp ; Rop ), which has been recently introduced for series
X X structure systems in [26] and for k-out-of-n systems in [14] is
þ ðcc;j þ cs ÞI fxj;T Z Z j g  cs @ I fxj;T Z Z j g þ j Gk j  1A
n

j A Ω⧹Gnk
k
j A Ω⧹Gnk
k
considered. The performance for the two maintenance policies is
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} numerically compared for the 14-component system. The policy
saved setup cost at T k : PED
(T; Rp ; Rop ) has three decision variables: the inter-inspection inter-
2 3
X val T, a PM threshold Rp fixed for all components of the system,
 cpsd 4 Iipsd  ð1  I ft j ¼ T g ÞI fGnk is a critical groupg 5 and an opportunistic maintenance (OM) threshold Rop (0 r Rp o
F F;k
i A Gnk
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Rop r 1). The description of this maintenance policy is reported in
saved planned shutdown cost at T k : PED or NED Appendices A and B. It is herein implemented and applied to the
1 14-component system.
  C To compare the performance of the two policies, a relative cost-
þ ðT k  T F;k Þcd þ cusd I ftj ¼ T g C g: ð23Þ
F;k A gain in the optimal cost-rate of the proposed policy (T; R0 )
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
F

unavailability cost in ðT k  1 ;T k  compared to the policy (T; Rp ; Rop ) is used as a criterion. It is

denoted CG and defined as follows:
Stochastic Monte Carlo simulation is used to evaluate this cost
criterion. For each value of the decision variables fT; R0 g, the C 1 ðT n ; Rnp ; Rnop Þ  C 1 ðT nTR ; Rn0 Þ
CG ¼  100 ð25Þ
corresponding cost-rate is calculated. To ensure that the conver- C 1 ðT n ; Rnp ; Rnop Þ
gence of cost-rate is reached, a large number of simulation
realizations must be done. where Tn, Rnp and Rnop (respectively T nTR and Rn0 ) are the optimal
By considering different values of fT; R0 g on a mesh grid, the decision variables of the policy (T; Rp ; Rop ) (respectively (T; R0 )).
minimum cost-rate can be obtained. The latter addresses the According to the definition, CG 4 0 means that the policy (T; R0 ) is
optimal value of the decision variables fT; R0 g, i.e., more effective than the policy (T; Rp ; Rop ) and less effective else.
1
In this section the degradation behavior of each component i is
C 1 ðT n ; Rn0 Þ ¼ min C ðT; R0 Þ : ð24Þ
T 4 0;R0 A ð0;1 assumed to be described by homogeneous gamma stochastic
processes ðX i;t Þt Z 0 (i ¼ 1; …; 14). The degradation increment
Table 1 between two times Tk and T k þ 1 follows a gamma probability
Data of a 14-component system. density function (pdf) with a shape parameter αi ðT k þ 1  T k Þ A R þ n
and a scale parameter β i A R þ n as shown in Appendix C. In this
cinsp ¼ 5, cs ¼ 30, cd ¼325, cusd ¼ cpsd ¼ 160
way, the conditional reliability of component i is predicted by
Component αi βi Zi cp;i cc;i Component αi βi Zi cp;i cc;i using Eq. (8)

1 0.5 1 30 130 195 8 1.1 1.5 35 140 210

Ri ðT k þ 1 j xi;T k Þ ¼ 1  PðX i;T k þ 1  T k ZZ i  xi;T k Þ
2 0.5 1 30 130 195 9 1.4 1.7 42 130 195 Z þ1
3 0.4 1.5 32 250 375 10 0.85 1.2 38 100 150 ¼ 1 f αi ðT k þ 1  T k Þ;βi ðxÞ dx
4 0.55 0.9 34 180 270 11 1 1 40 100 150 Z i  xi;T
k
5 0.6 1.1 27 150 275 12 0.65 1.4 28 200 300  
6 0.7 1.1 27 150 275 13 0.8 1.5 31 210 315 Γ αi ðT k þ 1 T k Þ; βi ðZ i  xi;T k Þ
¼ 1   ð26Þ
7 0.3 1.2 20 120 180 14 0.7 1.25 30 180 270 Γ αi ðT k þ 1  T k Þ

Fig. 3. Maintenance cost rate as a function of iterations numbers. (a) Policy (T; R0 ). (b) Policy (T; Rp ; Rop ).
90 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94
and the predictive reliability of the system RðT k þ 1 j x1:14;T k Þ can be
obtained from Eqs. (7) and (26). Table 1 reports the data of the 14-
component system.
5.1. Optimization of the proposed policy and the policy (T; Rp ; Rop )
cost-rate of the proposed policy (T; R0 ) and the policy
(T; Rp ; Rop ) are evaluated with different values of fT; R0 g and
fT; Rp ; Rop g using Eqs. (23) and (27) respectively. The obtained
results are the following:

For the policy (T; R0 ): the minimum cost-rate C 1 ðT n ; Rn0 Þ ¼
35:7216 has been obtained with the optimal inter-inspection
interval T n ¼ 43, the optimal PM threshold Rn0 ¼ 0:77.

For the policy (T; Rp ; Rop ): the minimum cost-rate C 1 ðT n ; Rnp ;
Rnop Þ ¼ 50:5536 corresponds to the optimal inter-inspection
interval T n ¼ 53, the optimal PM threshold Rnp ¼ 0:43 and the
optimal OM threshold Rnop ¼ 0:83.
Fig. 4. Cost-rate of the proposed maintenance policy as a function of T and R0.
Fig. 5. System predictive reliability relation before and after improvement
operations.
Fig. 6. Sensitivity analysis to: (a) set-up cost cs and (b) planned shutdown cost cpsd.
The convergence of the mean maintenance cost with respect to
the number of iterations is illustrated in Fig. 3.
Where, Fig. 3(a) shows that the convergence of the optimal cost-rate
(i.e. C 1 ðT n ; Rn0 Þ) of the policy (T; R0 ) is stable from Nb Z 18; 000
interactions, while that of the policy (T; Rp ; Rop ), in Fig. 3(b), is stable
from Nb Z 26; 000 interactions. This is because the policy (T; R0 ) is quite
simple to be implemented since only two decision variables T and R0
have to be considered while the policy ðT; Rp ; Rop Þ requires the optimiza-
tion of three decision variables. Thus, the convergence of the policy
ðT; Rp ; Rop Þ is slower than that of the proposed maintenance policy.
According to the obtained results, the cost-gain CG is about
29.34%. This means that the proposed maintenance policy (T; R0 ) is
significantly more profitable than the policy (T; Rp ; Rop ). This can be
explained by the fact that the proposed maintenance policy takes
into account the system structure and grouping effects, which can
lead to a reduction on maintenance cost.
For the policy ðT; R0 Þ, the cost-rate with different values of T and
R0 is sketched in Fig. 4. It is clear that the cost-rate surface is
convexly shaped and the global optimal solution exists. Addition-
ally, the maintenance impacts on the system reliability is illustrated
in Fig. 5 in the optimal case T n ¼ 43 and Rn0 ¼ 0:77. The results show
that whenever the system reliability drops below the threshold R0
(at inspection time T k ), the system improvement is triggered. Then
an optimal group of components Gnk ) is selected to be replaced
preventively. As a result, the system predictive reliability after
maintenance is raised to values higher than the threshold R0.
A short comparison of the two maintenance policies (T; Rp ; Rop )
and (T; R0 ) will be proposed for a series structure in Section 5.3.
The cost-gain between two policies may depend on the system
parameters and mainly on unitary costs such as inspection, set-up,
planned/unplanned shutdown costs as well as unplanned down-
time cost rate. The next Subsection is devoted to different
sensitivity analyses for the two maintenance policies.
5.2. Sensitivity analyses of two maintenance policies
In order to observe the trend of the cost-gain between the two
maintenance policies, sensitivity analysis with respect to the cost
parameters is herein investigated. The performance of the proposed
policy are compared to that of the policy (T; Rp ; Rop ). The sensitivity
analysis is performed separately on each considered cost parameter
by varying it when the others remain fixed. The obtained numerical
results are shown in Fig. 6 for the set-up cost and the planned
shutdown cost and Fig. 7 for the inspection cost and the downtime
cost rate. For these figures, each point on the cost-gain curve is
obtained from the respective optimal maintenance cost rates of the
two policies. The sensitivity analysis is summarized as follows:
 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94

Fig. 6(a) is obtained by varying the set-up cost cs from 0 to 90
with increments of 15 cost units and assigning the other
parameters of the system as in Table 1.

Fig. 6(b) is obtained by varying the planned shutdown cost cpsd from
0 to 640 with a step of 80 cost units (noted herein that cusd ¼ cpsd )
and assigning the other parameters of the system as in Table 1.

Fig. 7(a) is obtained by varying the inspection cost cinsp from
1 to 19 in increment of 2 cost units and assigning the other
parameters of the system as in Table 1.

Fig. 7(b) is obtained by varying the downtime cost rate cd from
0 to 1000 with a step of 100 cost units and assigning the other
parameters of the system as in Table 1.
5.2.1. Sensitivity analysis to the set-up cost cs and the planned
shutdown cost cpsd
As shown in Fig. 6, both cost-gains increase as the set-up cost
and the planned shutdown cost increase, respectively.
Sensitivity analysis to the set-up cost cs: Fig. 6(a) shows the influence
of the set-up cost on the maintenance cost of both policies. When the
set-up cost increases, the cost-rate of the two policies increases.
However, the maintenance cost of the policy ðT; Rp ; Rop Þ increases
faster than that of the policy ðT; R0 Þ. In other words, the higher cs, the
higher cost-gain is. Indeed the policy ðT; R0 Þ can avoid unnecessary
operations on the system thanks to the decision at system level. What
is more it has more opportunities than to replace preventively some
groups of components in order to save set-up-costs (due to PED). This
proves the effectiveness of the grouping policy of the proposed ðT; R0 Þ
maintenance decision rule.
Sensitivity analysis to the planned shutdown cost cpsd: Fig. 6
(b) shows the influence of the planned shutdown cost on the
maintenance cost of both policies. The impact of cpsd is similar to
that of cs. Because cpsd is taken into account to select the optimal
group of components for PM through the cost-based improvement
factor. This allows the policy ðT; R0 Þ to avoid actively unnecessary
shutdowns (or reducing the NED) due to the PM intervention on
system MCSs. As a result, the policy ðT; R0 Þ is more profitable than
the policy ðT; Rp ; Rop Þ when cpsd increases. Indeed, one can see
again on Fig. 6(b) that the cost-gain increases rapidly with cpsd.
This is an outstanding advantage of the proposed policy ðT; R0 Þ
when the planned shutdown cost cpsd is expensive.
Through the two previous cases (i.e. the influence of the set-up
cost and of the planned shutdown cost on both maintenance
policies) the proposed policy ðT; R0 Þ is shown to better drive its
decision variables in order to exploit the interactions among
components (by enhancing the PED and reducing the NED among
components) better than other policies for multi-component
systems with the complex structure.
Fig. 7. Sensitivity analysis to: (a) inspection cost cinsp and (b) downtime cost rate cd.
91
5.2.2. Sensitivity analysis to the inspection cost cinsp and the
downtime cost rate cd
Fig. 7 shows the impact of the inspection cost and the down-
time cost rate that has an opposite effect to that of the set-up cost
and the planned shutdown cost. The cost-gain is a decreasing
function of cinsp and of cd, respectively.
Sensitivity analysis to the inspection cost cinsp: The results shown
in Fig. 7(a) can be explained by the inspection cost. As it increases,
the optimal inter-inspection interval of both policies is set to
higher values to avoid frequent inspections. The results obtained
after optimization of the two policies show that the optimal inter-
inspection interval of the policy ðT; Rp ; Rop Þ (e.g. T n ¼ 53) is always
much longer that of the policy ðT; R0 Þ (e.g. T n ¼ 43). The number of
inspection times of the policy ðT; R0 Þ is greater than that of the
policy ðT; Rp ; Rop Þ during the system mission. As a consequence the
maintenance cost of the policy ðT; R0 Þ increases faster and
as a result the cost-gain function is a decreasing one as shown
in Fig. 7(a).
Sensitivity analysis to the downtime cost rate cd: The results
obtained about the sensitivity analysis of the downtime cost rate
cd show that the two maintenance costs increase sharply with cd.
When cd increases, the PM thresholds (i.e. Rp, Rop, R0) rise and the
inter-inspection intervals decrease for both policies. This means
that the components of the system are replaced earlier to prevent
penalty costs due to the system failure. The cost-gain in this case is
a decreasing function of cd as shown in Fig. 7(b). As mentioned
previously, this can be explained by the fact that the policy ðT; R0 Þ
tends to keep failed components (i.e. less critical components) till
inspection time is greater. The system sustainability is not ensured
as with the policy ðT; Rp ; Rop Þ. Consequently, the increasing of
C 1 ðT; R0 Þ is faster than that of C 1 ðT; Rp ; Rop Þ as cd increases.
5.3. Performance evaluation of the two policies performed on a
subsystem consisting of multi-component in series
In order to study the influence of the system structure on the
both policies, a series subsystem of 5 components (i.e. components
7, 8, 9, 10, 11) is extracted from the given system in Fig. 1. Its
production loss is lower than the whole complex system. In this
subsection, the following considered cost parameters are lower
than previously: cd ¼90, cusd ¼ cpsd ¼ 45. The other parameters are
unchanged as given in Table 1. The optimization process and the
sensitivity analysis are similar as in Sections 5.1 and 5.2.
Optimization of the two policies: The cost-rate of the policy
(T; R0 ) and (T; Rp ; Rop ) are respectively evaluated with different
values of fT; R0 g and fT; Rp ; Rop g using Eqs. (23) and (27). The
obtained results are the following:
92 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94

For the policy (T; R0 ): the minimum cost-rate is
C 1 ðT n ; Rn0 Þ ¼ 20:551. The optimal inter-inspection interval is
T n ¼ 34, the optimal PM threshold is Rn0 ¼ 0:116.

For the policy (T; Rp ; Rop ): the minimum cost-rate is
C 1 ðT n ; Rnp ; Rnop Þ ¼ 20:458. The optimal inter-inspection interval
is T n ¼ 34, the optimal PM threshold and the optimal OM
threshold are respectively Rnp ¼ 0:55 and Rnop ¼ 0:85.
The obtained cost-gain in this case is about  0.454% which shows
the policy (T; R0 ) seems to be less profitable than the policy
(T; Rp ; Rop ). The influences of cost parameters of the system (i.e.
inspection cost, set-up cost, planned shutdown cost as well as
unplanned downtime cost rate) on the cost-gain are investigated
hereafter through the different sensitivity analysis.
Sensitivity analysis to the maintenance costs: The sensitivity
analysis for each considered cost parameter is performed by varying
it while the other remain unchanged. The obtained numerical
results of the four parameters are shown in Fig. 8 in which:

Fig. 8(a) is obtained by varying the set-up cost cs from 0 to 90 in
increment of 15 cost units and assigning cinsp ¼ 5, cd ¼90,
cusd ¼ cpsd ¼ 45;

Fig. 8(b) is obtained by varying the planned shutdown cost cpsd
from 0 to 180 with a step of 22.5 cost units (noted herein that
cusd ¼ cpsd );

Fig. 8(c) is obtained by varying the inspection cost cinsp from
1 to 19 in increment of 2 cost units and assigning cs ¼30,
cd ¼90, cusd ¼ cpsd ¼ 45;

Fig. 8(d) is obtained by varying the downtime cost rate cd from
0 to 300 with a step of 30 cost units assigning cinsp ¼ 5, cs ¼30,
cusd ¼ cpsd ¼ 45.
All the other parameters are assigned as in Table 1.
Fig. 8 shows almost all the cost-gains are negative. This confirms
again that the policy (T; R0 ) is slightly less profitable than the policy
Fig. 8. Sensitivity analysis to (a) inspection cost cinsp, (b) set-up cost cs, (c) downtime cost rate cd, (d) planned shutdown cost cpsd.
(T; Rp ; Rop ) in the case of series system. This can be explained
because a series multi-component system is never stronger than
the weakest of its link/node. Thus, one should not invest too much
maintenance resource in order to improve the reliability of most of
components in the system if only one component remains at a very
low reliability level. The policy (T; Rp ; Rop ) makes maintenance
decisions at component level in combination with opportunistic
maintenance. All the components of system are considered at the
same priority level for the reliability improvement. It is more
appropriate for series structures than policy (T; R0 ). Finally, for series
multi-component systems, using the cost-based group improvement
factor for maintenance decision-making at system level is useless.
The maintenance policy to consider should decide maintenance
actions at component level and combine with the opportunistic
maintenance. This conclusion was also drawn by Li [17].
Let us summarize the comparison of the two policies from the
numerical results that have been obtained from various parameter
sets. In the case of series systems the cost-gain CG is slightly
negative (around  0.5%) for all the tested cases. In contrast, in the
general case of complex structures the cost-gain is highly positive
(more than 20%) in the vast majority of tested cases. The policy
ðT; R0 Þ, based on PM decision-making at system level has proved
its effectiveness in incorporating information about the system
structure into the decision process. The grouping policy is efficient
from the PED viewpoint and avoids actively unnecessary system
shutdowns from the NED viewpoint. The higher cs and cpsd is, the
more the policy ðT; R0 Þ is effective. The components selection for
PM activities is not only performed on the basis of their reliability,
but also on their incurred costs with respect to the ability to
improve the system reliability. This allows the policy ðT; R0 Þ to
preventively replace more important components while retaining
less important components (i.e. for example components with low
reliability improvement ability but hight replacement cost). So, the
policy ðT; R0 Þ allows more savings in maintenance cost compared
to the policy ðT; Rp ; Rop Þ, mainly if cinsp and cd are low.
 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94 93

It is important to note that the proposed maintenance policy is represented in the policy by a PM threshold Rp, 0 r Rp r 1. In
quite simple to be implemented since only two decision variables addition, in order to increase the effectiveness of grouping main-
T and R0 have to be considered while the policy ðT; Rp ; Rop Þ requires tenance operation, the opportunistic maintenance (OM) activities
the optimization of three decision variables. are handled through an OM threshold Rop. The PM actions have to
be implemented on more deteriorated components than the OM
ones, thus it is reasonable to assume that 0 r Rp o Rop r1. The NED
6. Conclusions is considered that if the grouping of PM activities lead the system
to a shutdown, a penalty cost cpsd is incurred. To do this end, at
In this work, a novel predictive maintenance policy with multi- inspection time Tk, the current state of surviving component i
level decision approach is proposed for multi-component systems (i ¼ 1; …; n) is firstly considered. The decision rules are placed for
with complex structure. Predictive reliability, positive and nega- component i is as follows:
tive dependencies as well as the structure of the system are taken
into account in the decision-making process. This process is
If component i has been failed, it has to be correctively
composed of two levels: a system level and a component one. replaced.
The first level aims at triggering maintenance interventions, while
If component i is surviving, but Ri ðT k þ 1 j xi;T Þ r Rp , component i k
the second level addresses the optimal selection of a group of is then preventively replaced.
several components to be preventively maintained using the cost-
If the system has been failed or if the system is functioning,
based group improvement factor. Thanks to the two level decision but the PM on the selected components lead the system to a
process, the proposed maintenance policy is more flexible and shutdown. This is considered as an opportunity to main-
efficient for complex structure system when compared with an tain other functioning ones. More precisely, component j
existing effective maintenance policy. In the cases of series (j a i) will also be opportunistically maintained if
structure systems, some numerical results show that the perfor- Rp o Rj ðT k þ 1 j xj;T k Þ r Rop .
mance of the proposed maintenance policy is slightly lower than
If all above decision rules above are not met, no maintenance
the existing one. Nevertheless the number of decision variables of action is then carried out on the component i at time Tk.
the proposed policy is lower. This point can be considered as an
advantage for the optimization process. Monte-Carlo simulation
technique is used to evaluate the maintenance cost-rate. This
method is robust but the computing time can intensively increase Appendix B. Cost model of the policy ðT; Rp ; Rop Þ
when the number of components is high.
This paper is the development of our research in the framework Let us denote by Gk the group of selected running components
of predictive maintenance approaches for multi-component sys- for the preventive replacement at time Tk. The same manner as
tems presented partially in [18]. Our future works will focus on the presented in Section 4.3, the long-run maintenance cost rate of the
development of the proposed maintenance policy with considera- system under the policy ðT; Rp ; Rop Þ can be formulated as follows:
tion of uncertainty on the deterioration parameters of components.

8 0 0 1
1 <X Nb X
1
C ðT; Rp ; Rop Þ ¼ h iE @cinsp @n  I fxi;T Z Z i g A
P b
E T rep  N ðT  T F;k Þ :k ¼ 1
k
k¼1 k i A Ω⧹Gk
X X X
þ ðcc;i þ cs ÞI fxi;T Z Z i g : þ C p;i I fRi ðT k þ 1 j xi;T Þ r Rp g þ ðcp;j þ cs ÞI fRp o Rj ðT k þ 1 j xj;T Þ r Rop g I fsystem is shutdowng
k k k
i A Ω⧹Gk i A Gk j A Gk ;j a i
0 1
X X X
 cs @ I fxi;T Z Zi g þ I fRi ðT k þ 1 j xi;T Þ r Rp g þ I fRp o Ri ðT k þ 1 j xi;T Þ r Rop g I fsystem is shutdowng 1
A
k k k
i A Ω⧹Gk i A Gk j A Gk ;j a i
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
saved setup cost at T k : PED
1
2 3 C
X   C
C
 cpsd 4 Iipsd  ð1  I fti ¼ T F;k g ÞI fG is a critical groupg
5 þ ðT k  T F;k Þcd þ cusd I i
ft F ¼ T F;k g Cg; ð27Þ
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} C
F k
i A Gk A
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} unplanned shutdown cost in ðT k  1 ;T k 
saved planned shutdown cost at T k : PED or NED

Actually these parameters are frequently estimated with little data

available and it may be difficult to ensure the estimators consis- The optimal decision variables fT n ; Rnp ; Rnop g of the policy
tency. Furthermore, we will focus also on mathematical develop- ðT; Rp ; Rop Þ can be obtained by minimizing the mean maintenance
ment for an analytical calculation of the maintenance cost-rate in cost rate function C 1 ðT; Rp ; Rop Þ, i.e.,
order to reduce the computing time. 1

C 1 ðT n ; Rnp ; Rnop Þ ¼ min C ðT; Rp ; Rop Þ : ð28Þ

T 4 0;0 r Rp o Rop r 1

Appendix A. Maintenance decision rules of the policy

ðT; Rp ; Rop Þ Appendix C. The pdf function is expressed as
1
The policy ðT; Rp ; Rop Þ in this study is extended from the f αi ðT k þ 1  T k Þ;βi ðxÞ ¼ βαi ðT k þ 1  T k Þ xαi ðT k þ 1  T k Þ  1 e  βi x I fx Z 0g ; ð29Þ
Γ ½αi ðT k þ 1  T k Þ i
ðΔT; Rp ; Ru Þ policy reported in [14] in which the NED and structural
R þ1
dependency between components of the system are not taken into where Γ ðtÞ ¼ 0 ut  1 expð  uÞ du denotes the Euler gamma
account in maintenance decision-making. The PM activities are function. The couple of parameters (αi ; βi ) is supposed to be given
94 K.-A. Nguyen et al. / Reliability Engineering and System Safety 144 (2015) 83–94
but can be estimated from monitoring data. The mean degradation
rate and the variance are given by αi =βi and αi =βi , respectively.
2 management conference, Milan, Italy; 2013.
References
[1] Bevilacqua M, Braglia M. The analytic hierarchy process applied to main-
tenance strategy selection. Reliab Eng Syst Saf 2000;70:71–83.
[2] Bouvard K, Artus S, Berenguer C, Cocquempot V. Condition-based dynamic
maintenance operations planning & grouping. application to commercial
heavy vehicles. Reliab Eng Syst Saf 2011;96(6):601–10.
[3] Castanier B, Grall A, Bérenguer C. A condition-based maintenance policy with
non-periodic inspections for a two-unit series system. Reliab Eng Syst Saf
2005;87(1):109–20.
[4] Cho DI, Parlar M. A survey of maintenance models for multi-unit systems. Eur J
Oper Res 1991;51(1):1–23.
[5] Dasgupta A, Pecht M. Material failure mechanisms and damage models. IEEE
Trans Reliab 1991;40(5):531–6.
[6] Dekker R, Wildeman RE, van der Duyn Schouten FA. A review of multi-
component maintenance models with economic dependence. Math Methods
Oper Res 1997;45(3):411–35.
[7] Dieulle L, Bérenguer C, Grall A, Roussignol M. Sequential condition-based main-
tenance scheduling for a deteriorating system. Eur J Oper Res 2003;150(2):451–61.
[8] Do P, Voisin A, Levrat E, Iung B. A proactive condition-based maintenance
strategy with both perfect and imperfect maintenance actions. Reliab Eng Syst
Saf 2015;133:22–32.
[9] Do Van P, Barros A, Bérenguer C, Bouvard K, Brissaud F. Dynamic grouping
maintenance strategy with time limited opportunities. Reliab Eng Syst Saf
2013;120:51–9.
[10] Do Van P, Vu H, Barros A, Berenguer C. Grouping maintenance strategy with
availability constraint under limited repairmen. In: The 8th IFAC international
symposium on fault detection, supervision and safety for technical processes,
SAFEPROCESS-2012, August 2012, Mexico city, Mexico; 2012. p. 486–91.
[11] Faulin J, Juan AA, Martorell S, Ramírez-Márquez J-E. Simulation methods for
reliability and availability of complex systems. Springer-Verlag, London; 2010.
[12] Grall A, Dieulle L, Bérenguer C, Roussignol M. Continuous-time predictive-main-
tenance scheduling for a deteriorating system. IEEE Trans Reliab 2002;51(2):141–50.
[13] Horenbeek AV, Pintelon L. A dynamic predictive maintenance policy for
complex multi-component systems. Reliab Eng Syst Saf 2013;120:39–50.
[14] Huynh KT, Barros A, Berénguer C. Multi-level decision-making for the
predictive maintenance of k-out-of-n: F deteriorating systems. IEEE Trans
Reliab, 64(1), 2015, 94–117.
[15] Huynh KT, Castro IT, Barros A, Bérenguer C. On the use of mean residual life as
a condition index for condition-based maintenance decision-making. IEEE
Trans Syst Man Cybern: Syst 2014;44(7):877–93.
[16] Lesobre R, Bouvard K, Berenguer C, Barros A, Cocquempot V. A maintenance
free operating period policy for a multi-component system with different
information levels on the components state. In: Prognostic and system health
[17] Li L, Ambani S, Ni J. Plant-level maintenance decision support system for
throughput improvement. Int J Prod Res 2009;47(24):7047–61.
[18] Nguyen KA, Do P, Grall A. Predictive grouping maintenance strategy for
complex structure systems using importance measure. In: International
conference on quality, reliability, risk, maintenance, and safety engineering
(QR2MSE) 2013, international conference on quality, reliability, risk, main-
tenance, and safety engineering (QR2MSE) 2013, Emeishan, Sichuan, China,
IEEE, Piscataway, New Jersey; 2013. p. 582–88.
[19] Nguyen K-A, Do P, Grall A. Condition-based maintenance for multi-component
systems using importance measure and predictive information. Int J Syst Sci:
Oper Logist 2014;1(4):228–45.
[20] Nguyen K-A, Do P, Grall A. A joint predictive maintenance and spare parts
provisioning policy for multi-component systems using RUL prediction and
importance measure. In: 2nd european conference of the prognostics and
health management society, PHME 2014, vol. 5, no. 14; 2014. p. 500–11.
[21] Pham H, Wang H. Imperfect maintenance. Eur J Oper Res 1996;94(3):425–38.
[22] Rausand M, Høyland A. System reliability theory: models, statistical methods
and applications. 2nd ed.John Wiley & Sons, Hoboken, New Jersey; 2004.
[23] Ross S. Stochastic processes. John Wiley & Sons, Inc., Chichester, New York;
1996.
[24] Shaikh A, Mettas A. Application of reliability, availability, and maintainability
simulation to process industries: a case study. In: Simulation methods for
reliability and availability of complex systems. Springer, London; 2010.
[25] Thomas L. A survey of maintenance and replacement models for maintain-
ability and reliability of multi-item systems. Reliab Eng 1986;16(4):297–309.
[26] Tian Z, Liao H. Condition based maintenance optimization for multi-
component systems using proportional hazards model. Reliab Eng Syst Saf
2011;96(5):581–9.
[27] Tijms HC. A first course in stochastic models. New York: John Wiley & Sons;
2003.
[28] VU HC, DO P, Barros A. Bérenguer C. Maintenance grouping strategy for multi-
component systems with dynamic contexts. Reliab Eng Sys Saf 2014;132:
233–49.
[29] VU HC, DO P, Barros A, Bérenguer C. Maintenance planning and dynamic
grouping for multi-component systems with positive and negative depen-
dencies. IMA j Manag Math 2015;23(2):145–70.
[30] Wang H. A survey of maintenance policies of deteriorating systems. Eur J Oper
Res 2002;139(3):469–89.
[31] Wang W, Zhang W. A model to predict the residual life of aircraft engines
based upon oil analysis data. Naval Res Logist (NRL) 2005;52(3):276–84.
[32] Xu Z, Ji Y, Zhou D. A new real-time reliability prediction method for dynamic
systems based on-line fault prediction. IEEE Trans Reliab 2009;58(3):523–38.
[33] Yun WY, Moon I, Kim G. Simulation-based maintenance support system for
multi-functional complex systems. Prod Plan Control 2008;19(4):365–78.