You are on page 1of 24

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON RELIABILITY 1

Multi-Level Decision-Making for The Predictive


Maintenance of -Out-of- :F Deteriorating Systems
Khac Tuan Huynh, Anne Barros, and Christophe Bérenguer

Abstract—Traditional maintenance decisions in the framework NOTATION


of condition-based maintenance applied to multi-component sys-
tems are performed either at the system level or at the component Gamma function, incomplete gamma
level. These decisions however cannot always assure the best main- function
tenance performance. To remedy this drawback, the present paper
introduces a multi-level decision-making approach that combines
Sum over all permutations
maintenance decisions at the system level and the component level. of for
The effectiveness of such an approach is investigated through an which , and
-component deteriorating system with a -out-of- :F structure, (including
and economic dependence. In fact, based on the degradation and terms)
failure model of the considered -out-of- :F system, two new op-
portunistic predictive maintenance strategies with different types Sum over all permutations
of maintenance decision-making are proposed. In the first one, the of
decisions rely only on the remaining useful lifetime of the com- for which ,
ponents; while in the second one, they are based on both the re- , and
maining useful lifetimes of the system and that of its components. (including
The maintenance cost models of these strategies are developed on
the basis of semi-regenerative theory, optimized, and then com- terms)
pared with each other. The comparison results show that the multi- Indicator function
level decision-making approach allows us to more effectively avoid
inopportune interventions, to better take into account the interac- Shape, scale parameters of gamma
tions among components, and hence to provide more flexible and degradation process of the component
profitable predictive maintenance strategies for multi-component
systems.
Average rate, variance rate of gamma
degradation process of the component
Degradation processes of the maintained
Index Terms—Economic dependence, gamma process,
component
-out-of- :F deteriorating system, mathematical cost model, Degradation processes of the unmaintained
multi-level maintenance decision, opportunistic predictive main- component
tenance strategies, prognostic condition index, semi-regenerative
theory. Degradation failure threshold, failure time
of the component
Gamma density function with parameters
ACRONYMS AND ABBREVIATIONS
, and of the component
CBM Condition-based maintenance Conditional survival function, conditional
reliability of the component
RUL Remaining useful lifetime
Multivariate process describing the
PR, CR Preventive replacement, corrective
degradation evolution of the maintained
replacement
system
OPR, PPR Opportunistic PR, pure PR
Multivariate Markov chain describing states
MC Monte Carlo of the maintained system at inspection times
Invariant probability measurement
(stationary law) of
Conditional transition law
Manuscript received October 27, 2013; revised February 28, 2014; accepted
April 13, 2014. Associate Editor: M. Zuo. Failure time of the system
K. T. Huynh and A. Barros is with ICD, ROSAS, LM2S, Université de Tech-
nologie de Troyes, 10004 Troyes cedex, France (e-mail: tuan.huynh@utt.fr; Conditional RUL of the system at time
anne.barros@utt.fr).
C. Bérenguer is with Univ. Grenoble Alpes, GIPSA-lab, F-38000 Grenoble, Conditional reliability of the system
France. He is also with CNRS, GIPSA-lab, F-38000 Grenoble, France (e-mail: Cost per inspection, cost per setup
christophe.berenguer@grenoble-inp.fr).
Digital Object Identifier 10.1109/TR.2014.2337791
operation, downtime cost rate

0018-9529 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON RELIABILITY

Cost per PR, cost per CR of the component degradation domain corresponding to
replacement of the group of different
Cumulative maintenance cost of the system components
at time degradation domain corresponding to
Long-run maintenance cost rate of the no replacement or OPR of the component
system degradation domain corresponding to
Number of system inspections in no replacement or OPR of the group of
different components
Number of PR of the component in
degradation domain corresponding
Number of CR of the component in to unknown operation and OPR of the
Number of groups of components that are component
replaced at the same time in degradation domain corresponding to
Downtime of the system in unknown operation and OPR of the group
of different components
First replacement time of the system
degradation domain for which
First inspection time of the system interventions are allowed at the component
Times just before, and after the -th level
inspection time complement domain of
Degradation levels at the time just before, is defined as
and after the -th inspection time
Inspection period, optimal value of
, PPR threshold, optimal value of I. INTRODUCTION
OPR threshold, optimal value of
Threshold of the system conditional
reliability below which interventions should
W ITH the fast development of modern industry, mainte-
nance operations have an important role in keeping and
improving system availability and safety, as well as product
be triggered at the component level, optimal
quality. Over the past few decades, one has witnessed the
value of
evolution of maintenance strategies from the naive breakdown
First hitting degradation level of the maintenance, to the blind time-based preventive maintenance
threshold at time for the component [1], and lately towards the sophisticated condition-based
First hitting degradation level of the maintenance (CBM) [2]. The CBM approach bases the deci-
threshold at time for the component sion-making on real-time diagnostic information of impending
failures or prognostics of future system health, so if properly
degradation domain corresponding to
no replacement of the component established, it can efficiently avoid inopportune interventions,
and reduce significantly the system downtime compared to
degradation domain corresponding to other maintenance strategies [3]–[5]. The main drawback of
no replacement of the group of different CBM is the need for monitoring equipment [6], [7]. However,
components the growing dissemination of condition monitoring techniques
degradation domain corresponding to can reduce this drawback [8], and the implementation of a
OPR on the component CBM strategy is encouraged among organizations seeking
degradation domain corresponding to improve their maintenance performance under budget and
to OPR of the group of different resources constraints, and to gain a competitive advantage [9].
components In the literature, models of CBM strategies have been widely
developed and successfully applied to single-component sys-
degradation domain corresponding to
tems (see e.g., [10]–[12]). Theoretically, a strategy for single-
PPR on the component
component systems may be applicable to multi-component sys-
degradation domain corresponding to tems by summarizing their global health state as a 1-D con-
PPR of the group of different components dition index based on which one makes a maintenance deci-
degradation domain corresponding to sion [13]. In this case, the maintenance decision is known as
CR on the component the decision at system level. Such a decision, as shown by Cho
degradation domain corresponding to and Parlar in [14], is especially suitable to systems whose com-
CR of the group of different components ponents are independent of each other. However, a multi-
component system in practice is always subject to inherent in-
degradation domain corresponding to teractions among the components (i.e., economic, structural,
replacement of the component and stochastic dependence [15]). Economic dependence implies
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 3

Fig. 1. Maintenance decision-making approaches in the framework of multi-component systems.

that costs can be saved when several components are jointly according to the -out-of- :F structure. From the maintenance
maintained. Stochastic dependence means that the failure or the viewpoint, only economic dependence is studied because it is
degradation of a component can affect the state of one or more the most common dependence type of multi-component sys-
other components. Finally, structural dependence indicates that tems. Indeed, one can find, in the literature, numerous published
components structurally form a unified part, so that the main- papers dealing with economic dependence, but very few works
tenance of one component implies maintenance of other com- that consider structural and stochastic dependencies in mainte-
ponents. Because of the existence of these types of interac- nance models [1], [14], [15], [32]–[34]. Furthermore, the mod-
tions, the optimal strategy for a multi-component system is not eling of structural dependence, in some way, can come down
simply the juxtaposition of the optimal strategies of its compo- to economic dependence (see e.g., [27]). Although stochastic
nents [16]. As a consequence, strategies for single-component dependence is not considered in this paper to avoid mathemat-
systems cannot be properly applied to multi-component ones. ical complexity, one can see in the following that the proposed
Faced with this situation, a few efforts have been recently devel- maintenance strategies are rather generic, and can be applied
oped using CBM strategies considering component interactions. for such a dependence. Traditionally, the economic dependence
One can cite, for example, [16]–[21] that study strategies with can be taken into account in CBM decision-making through 2
economic dependence, and [22]–[26] that consider strategies main approaches [1]: grouping maintenance, and opportunistic
with integrating stochastic dependence. But considering struc- maintenance. Compared to the former, the latter does not require
tural dependence in CBM decision-making is still very limited. planning maintenance operations in advance [33]; it seems then
As far as we know, only Van Horenbeek and Pintelon incor- more suitable to multi-level decision-making. Thus, we limit
porate structural dependence into their CBM model [27]. All our study to the class of opportunistic maintenance strategies.
the existing CBM strategies applied to multi-component sys- Moreover, unlike most existing CBM strategies whose main-
tems share the common weak point that maintenance decisions tenance decision-making relies directly on a diagnostic condi-
are only made at the component level. The information on the tion index (e.g., current degradation level), we base the deci-
global system (e.g., system health state, system structure) is then sions on a prognostic condition index which is the remaining
not directly incorporated into decision-making, and the highest useful lifetime (RUL) of components or the system. This RUL
maintenance performance is thus not always guaranteed. In this is associated to the conditional reliability at a future time given
context, the present paper aims to remedy this drawback by a current degradation level. Thus the maintenance strategies
introducing a so-called multi-level decision-making that com- in this paper are called reliability-based opportunistic predic-
bines the decisions at both the system level and the component tive maintenance strategies. In fact, two reliability-based oppor-
level (see Fig. 1). This type of decision promises to effectively tunistic predictive maintenance strategies have been proposed
avoid inopportune interventions, exploit better the interactions for the considered -out-of- :F deteriorating system. In the first
among components, and hence provide more flexible and prof- strategy, only the decision-making at the component level is
itable CBM strategies for multi-component systems. used, while in the second strategy, multi-level decision-making
In this work, we apply multi-level CBM decision-making is applied. The mathematical cost models of these maintenance
to an -component deteriorating system with a -out-of- :F strategies are also developed, optimized, and compared with
structure and economic dependence among components. The each other. The comparison results under different configura-
choice of -out-of- :F structure, on the one hand, satisfies a tions of system characteristics and maintenance costs allow us to
wide range of applications in both industrial and engineering give some conclusions on the effectiveness of multi-level CBM
systems [28]. On the other hand, it allows one to obtain a main- decision-making.
tenance cost model with complete probabilistic formulation and As such, the main contributions of this paper are threefold.
numerical integral calculations. Of course, more complex struc- • Introducing the multi-level maintenance decision-making
tures (e.g., structures combining a significant amount of series approach, and assessing its benefits for maintenance per-
and parallel connections [29]–[31]) could be considered at the formance. To the best of our knowledge, this approach has
price of a more difficult (even intractable) analytical formula- not been studied before in the literature on maintenance
tion: one would then have to resort to simulation approaches to modeling.
deal with the problem. This choice also implies the structural • Incorporating a prognostic condition index into mainte-
dependence among the components from the reliability view- nance decision-making. Using RUL as a condition index
point, in the sense that they contribute to the system failure makes the maintenance decisions more robust with respect
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON RELIABILITY

to the estimation errors of degradation and failure model For , the survival function of the conditional system RUL
parameters [35], [36]. In the context of multi-component can be expressed as
systems, this can also provide effective maintenance struc-
tures, and is very easy to optimize [37].
• Proposing two new reliability-based opportunistic predic-
tive maintenance strategies, and developing their full ana-
lytical cost model for -out-of- :F deteriorating systems.
This work helps to provide a sufficiently general frame- (2)
work to study the value of multi-level maintenance deci-
sion-making. where , and is the conditional
reliability function of the system at time given .
Equation (2) shows that RUL and system failure time share the
II. SYSTEM MODELING AND PROGNOSTIC CONDITION INDEX same probability law, and thus assessing the system RUL at a
Consider a system subject to preventive or corrective main- current time can be done by assessing the conditional reliability
tenance operations driven by the knowledge on the future of the system at a future time. This notion is also true for the
system health state (i.e., according to a predictive maintenance case of components (i.e., setting the value of in (2) to 1, one
strategy). The system is built from identical or non-identical obtains the expression for a component). Note that the above
deteriorating components. Each component suffers an under- formulas are suitable for the case of perfect observations (i.e.,
lying degradation process which can cause random failures. the observations can return the underlying degradation level of
Such a process may be a physical deterioration process such a component or system without error). In this case, using the
as cumulative wear, bearing wear, crack growth, erosion, entire degradation history or just the latest degradation level
corrosion, fatigue, etc. [38]; or it may be an artificial process give the same value of RUL because of the Markov property
describing the phenomenon that the health state or the perfor- (i.e., memoryless property) of the degradation and observation
mance of a component worsen with usage and age [39], [40]. model. When the observations are imperfect (i.e., noisy mea-
We assume that the degradation of each component evolves surements, partial observations, etc.), the Markov property is
stochastically, —independently, and gradually over time. The no longer true; to guarantee the accuracy of the RUL prediction,
components in the systems are however structurally dependent we should rely on the entire observation history rather than on
from the reliability viewpoint, and economically dependent the latest observation only [48]. Faced with this case, the above
from the maintenance viewpoint. In the literature, a simple, formulas remain applicable, but a degradation estimation step
classical way to describe such a system is using lifetime models has to be added before computing the RUL. Indeed, the vector
[41]. These models are somewhat disadvantaged in character- of underlying degradation level has to be estimated from the
izing the system behavior because they cannot reflect properly entire history of imperfect observations using elaborate simu-
the intermediates states (i.e., degradation states) of the system. lation techniques such as particle filter [9], or Markov chain
To avoid this drawback, it is recommended to base the system Monte Carlo [49]. Then, we replace in (2) by the estimated
modeling on time-dependent stochastic processes [42]. The degradation vector to obtain the value of RUL. To simplify the
notion of process helps us to describe more finely the behavior problem, we consider only the case of perfect observations in
of components and the system, and hence allowing a more the present paper.
accurate prediction of their RUL. In the rest of this section, we present in detail the degradation
Actually, RUL is a key prognostic index for maintenance de- and failure model of the components and system, and the ana-
cision-making because it can provide an idea about how long a lytical formulation of their conditional reliability. As in [50], we
component or a system at a particular age will still survive. In consider successively the component level and the system level.
the literature, many approaches have been developed for RUL
prediction (see e.g., [43]–[46]). Generally, they can be divided A. Component Level
into non-condition-based approaches, and condition-based ap- Considering the -th component of the maintained system,
proaches [47]. The latter allows introducing the current health , its accumulated degradation level at time
state of the components and system into the prediction, so it can can be summarized by a scalar random aging variable . Be-
lead to a higher accuracy compared to the former approach [45]. tween two consecutive maintenance operations, this aging vari-
Thus the condition-based approach will be used in this paper able evolves as a continuous-time monotonically increasing sto-
to evaluate the RUL of the components and system. As such, chastic process with . can be consid-
let , , be the random variable representing ered as the degradation level at time of the unmaintained -th
the degradation state of the -th component at time , component. In this paper, is assimilated by a homo-
be the random vector characterizing the geneous gamma process with shape parameter , and scale pa-
corresponding global degradation state of the system at this rameter . The choice of the gamma process for the degrada-
time, and be its random failure time. The RUL of the tion has been justified by many practical applications [11]. Of
system at time given can course, other monotone stochastic processes (e.g., any mono-
be defined as the conditional random variable [10] tone process from the Levy family [51]) could have been used
at the price of more difficult mathematical derivations. Thus, for
(1) , the degradation increment of the component
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 5

in the time interval follows a gamma probability law with fault-tolerant systems, and communication systems. Because of
density function the -out-of- :F configuration of the system, its failure time
is equal to the -th smallest value among the set of com-
ponent failure times . Mathematically, it is the
(3) -th order statistic among . Thus, let be the
number of failed components up to time , according to the
notion of order statistic, without any maintenance operation,
where denotes the Euler gamma func-
the conditional reliability function of the system at given the
tion, and denotes the indicator function which equals 1 if the
system state at time , , is computed as [52, page
argument is true and 0 otherwise. Before the first maintenance
96]
operation, the process is the same as . The
behavior of such degradation processes depends closely on the
parameters , and its average degradation rate, and vari-
ance are given by , and respectively.
The parameters and can be given from degradation data
by different classical estimation methods [11].
To describe the failure of component , a threshold-type
model is used. In fact, for economic (e.g., poor products
quality, high consumption of raw material) or safety reasons
(e.g., high risk of hazardous breakdowns), a component is
usually declared as failed when it is no longer able to fulfill its (7)
mission in an acceptable condition even if it is still functioning.
Thus, a high degradation level of a component is unacceptable.
According to this view, we consider that the component fails where , and are given from (5), and (6) re-
as soon as its degradation level exceeds a critical prefixed spectively. The expression (7) can be evaluated by a recursive
threshold . Let be the corresponding random failure time. method represented in [52, pages 105–106]. The system con-
It can be expressed by ditional reliability will be used in Section III as a prognostic
condition-index to decide at the system level whether or not an
(4) intervention should be triggered at the component level.
Without any maintenance operation, the conditional distribution III. RELIABILITY-BASED OPPORTUNISTIC
function of at time given the degradation level at time PREDICTIVE MAINTENANCE STRATEGIES
, , is computed as [4] In this section, we propose two different opportunistic predic-
tive maintenance strategies for the aforementioned -out-of- :F
deteriorating system. In the first one, the global state of the
system health is not directly incorporated into maintenance de-
cision-making, while in the second, this information is used
to decide, at the system level, whether or not an intervention
(5) should be triggered at the component level. Both strategies re-
flect the joint effects of economic dependence and the CBM
where denotes the incomplete decision on the maintenance cost. Moreover, unlike most ex-
gamma function. The associated conditional reliability of the isting CBM strategies whose maintenance decision-making re-
-th component is then lies directly on the degradation level of components (i.e., diag-
(6) nostic condition index), we base the decision on prognostic con-
dition indices which are the conditional RUL of components or
One will see in Section III that the conditional reliability given the system given their degradation level (i.e., their associated
by (6) will be used as a prognostic condition index for preven- conditional reliability from a probabilistic point of view). The
tive and opportunistic maintenance decision-making at the com- proposed strategies are therefore called reliability-based oppor-
ponent level. tunistic predictive maintenance strategies.

B. System Level A. Maintenance Assumptions


The system has a -out-of- :F structure, i.e., it fails iff at We assume that the degradation of each component in the
least of the components fail. This structure is equivalent system is hidden, and the component failure is non-self-an-
to the -out-of- :G structure where the system works nouncing. An inspection on the global system is necessary to
iff at least of the components work. This is a reveal the degradation level and the failure or working state of
more general case of a series system (i.e., 1-out-of- :F struc- all components. These assumptions are practical because the
ture) and a parallel system (i.e., -out-of- :F structure). As ar- degradation of components is usually not observed directly;
gued by Kuo and Zuo in [28], such a structure is applicable moreover, as mentioned above, for economic or safety reasons,
to many industrial systems such as oil and gas supply chains, the component can be considered as failed even if it is still
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON RELIABILITY

running. So no indicator can exhibit the degradation and failure than two replacement operations are performed at the same time.
state of the system except to do an inspection. Note that the Thus, economic dependences among components exist, and one
notion of inspection mentioned here is not simply the data would like performing maintenance jointly on several compo-
collection, but also the feature extraction from the collected nents to save set-up costs.
data, the construction of degradation indicators, and perhaps
more [53]. In other words, this activity includes all the tasks B. Reliability-Based Opportunistic Predictive Maintenance
before the task Maintenance Decision Making in a CBM pro- Strategy Without Considering the Global System Health: the
gram [10]. Such an inspection operation is itself costly, and Strategy
takes time; however, compared to the life cycle of a system, the In this section, we develop an opportunistic maintenance
time for an inspection is negligible. Thus we assume that each strategy integrating both the CBM decision and economic
inspection operation is instantaneous, perfect, non-destructive, dependences. The CBM decision is integrated in the strategy
and incurs a cost . through a pure preventive replacement (PPR) threshold ,
On each component, two maintenance operations are avail- , on the component conditional reliability (review
able: a preventive replacement, and a corrective replacement. A Section II-A), while the economic dependence is handled by
replacement can be either a true physical replacement, an over- an opportunistic preventive replacement (OPR) threshold ,
haul, or a repair such that the components are as-good-as-new . Compared to an OPR, a PPR has to be performed
after the repair. Although perfect maintenance actions are on more deteriorated components. Moreover, the higher the
applied at the component level, the maintenance of the entire degradation level of a component, the smaller the value of
system is still imperfect because the system cannot return to its conditional reliability. Thus, we obtain the relationship
an as-good-as-new state except that all its components are re- .
placed at the same time. Because the considered system can be 1) Strategy: The system is regularly inspected
heterogeneous, its components may incur different costs for the with a period of . Each inspection reveals the degradation
same type of maintenance operation. Furthermore, even though level and the failure or working state of all components of the
both the preventive and the corrective maintenance actions put system. At an inspection time , , we re-
a component back in its as-good-as-new state, their costs are place correctively the failed components (i.e.,
not necessarily identical because the corrective replacement for the component , ), and
is unplanned, and its cost can also comprise different costs preventively replace the surviving components if their condi-
associated with failure, like damage to the environment. The tional reliability at the next inspection time given the actual de-
corrective replacement is thus likely to be more complex and tected degradation level is less than a threshold (i.e.,
more expensive than the preventive replacement. Additionally, ). When a component is
because the replacement actions on a component are usually replaced (either preventively or correctively), this time is the
true, complete physical replacements, and not partial repairs, opportunity to replace other surviving ones. More precisely, if
the associated costs are then fixed for its varied degradation a replacement is scheduled for component at the inspection
levels. To sum up, if , and , , are re- time , we also opportunistically replace component if
spectively the preventive, and corrective replacement cost its conditional reliability at the next inspection time given the
of component , one has the relation . Note that, actual detected degradation level is less than a threshold
contrary to the constant cost incurred for each replacement (i.e., ). If
at the component level, the maintenance costs for the entire none of the above conditions apply, nothing is done at , and
system are generally varied, because they depend closely on the maintenance decision is postponed until the next inspection
the number of replaced components and on the nature (i.e., time . Obviously, , , and are the de-
either preventive or corrective) of replacement actions on cision parameters to be optimized. To enhance the importance
each component. Also, because the components’ states can be of these parameters, we call the strategy .
known only by inspections, and because the time to maintain 2) Illustration and Analysis: For an illustration, we apply
them is usually much smaller than their lifetime, we consider the strategy defined by the parameters ,
that a replacement, whether preventive or corrective, is in- , and for a 2-components parallel system
stantaneous, and can be performed only at an inspection time. (i.e., 2-out-of-2:F system) characterized by the set of parame-
Therefore, a system downtime may appear if at least of the ters , , ;
components fail within two successive inspection times, hence then we sketch the degradation evolution of components,
an additional cost is incurred from the system failure time until and the associated conditional reliability at inspection times.
the next replacement time at a cost rate . Fig. 2(a), and Fig. 2(b) show respectively the numerical results.
Consider also that the components in the system are econom- In Fig. 2(a), the notations , and refer respectively to
ically dependent, i.e., cost can be saved when several compo- the degradation levels of a component at the time just before,
nents are maintained together rather than separately [33]. In fact, and just after the inspection or replacement time . We can
associated with a replacement, operations, such as sending a remark that the strategy allows an essentially
maintenance team to the site, dismantling and reassembly of the instantaneous maintenance operation on a component when-
system, etc., are needed. These operations incur a set-up cost ever it is detected to be in a critical state (i.e., high degradation
which is assumed to be cost-independent of the operation na- level or failure), regardless of the health state of global system.
ture, and such a cost is taken into account only once if more This characteristic commonly ensures a high system reliability
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 7

Fig. 2. Evolution of the behavior of 2-components parallel deteriorating system under the strategy. (a) Components degradation level at actual
inspection or replacement times. (b) System and components conditional reliability at next inspection times.

TABLE I parameters or decision parameters do not lead to a significant


SEQUENCE OF MAINTENANCE OPERATIONS UNDER THE loss in the performance of maintenance strategies). The number
STRATEGY IN THE TIME INTERVAL [0,200]
of decision parameters of the strategy is only
3, and is always invariant for all sizes of systems. Although
this number is the same in the case of degradation-based main-
tenance strategies of mono-threshold type, the
strategy is always more efficient. The same reliability threshold
level (see the evolution of the system conditional reliability at for all components can correspond to different degradation
the bottom of Fig. 2(b)); however, it can also lead to redundant levels of different components. Making a maintenance decision
maintenance operations. Actually, for some configurations of based on a reliability level is therefore more flexible than based
the system (e.g., parallel structure), the fact that a component on assuming the same degradation level for all components. As
falls into a critical state does not always significantly reduce the a result, the strategy is generally more profitable
system reliability. Carrying out instantaneously a maintenance than the ones of mono-threshold type, especially for heteroge-
operation on the component can be unnecessary for the global neous systems [37]. With the same reason, the
system health state. As a result, from an economic viewpoint, strategy can be considered as a degradation-based maintenance
using the strategy in such contexts can be strategy of multi-threshold type. Compared to the original ones
non-optimal. A new reliability-based opportunistic predictive (see e.g., the strategy developed in [16]), the
maintenance strategy considering the global system health strategy may have a lower performance because of an additional
proposed in Section III-C can remedy this drawback. Table I constraint on the components’ reliability. However, for original
presents explicitly the sequence of maintenance operations degradation-based maintenance strategies of multi-threshold
performed on each component in the time interval [0,200] type, the number of decision parameters increases rapidly
corresponding to Fig. 2. with respect to the number of components. Such strategies
Now, we analyze the merit of the strategy seem infeasible in practical applications, because their decision
compared to most existing CBM strategies whose maintenance structure is relatively complex, and optimizing their decision
decision-making is directly based on the degradation level of parameters is intractable. This reason is also why we avoid re-
components. In the framework of multi-component systems, liability-based opportunistic predictive maintenance strategies
the literature of degradation-based maintenance strategies can of multi-threshold type, which are theoretically better than the
be divided into mono-threshold type (see e.g., [54], [55]), and strategy, but practically worse.
multi-threshold type (see e.g., [16], [56]). In the former, only
one degradation threshold is applied for all components for C. Reliability-Based Opportunistic Predictive
each type of PR operation (i.e., either PPR or OPR); while Maintenance Strategy Considering Global System Health
in the latter, various degradation thresholds can be used for Strategy
different components. The strategy relies on Despite its advantages, the strategy presented
the future health state of components (i.e., through their con- in the above section, as well as most of the existing maintenance
ditional reliability) to make a PR decision rather than their strategies for multi-component systems (see e.g., [1], [5], [10]),
actual degradation state. This strategy leads to a more robust is still a semi-blind strategy because the information about the
maintenance decision-making compared to degradation-based global system health is not taken into account in maintenance
maintenance strategies [35], [36] (the robustness here means decisions. The aim of this section is to propose a maintenance
that the estimation errors of degradation and failure model strategy that can overcome this shortcoming. The idea is that we
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON RELIABILITY

Fig. 3. Evolution of the behavior of 2-components parallel deteriorating system under the strategy. (a) Components degradation level at the
actual inspection or replacement times. (b) System and components conditional reliability at the next inspection times.

should not always intervene at the component level when one or TABLE II
more components are in critical states, but only when necessary, SEQUENCE OF MAINTENANCE OPERATIONS UNDER THE
STRATEGY IN THE TIME INTERVAL [0,200]
i.e., when the health state of the global system is critical. Thus,
in addition to an PPR threshold and an OPR threshold as
in the strategy, a new threshold ,
, on the system conditional reliability is introduced to decide,
at the system level, whether or not an intervention should be
made at the component level. This approach proceeds from the
multi-level decision-making approach mentioned in Section I. 2) Illustration and Analysis: To illustrate the
1) Strategy: The system is regularly in- strategy, the same system with the
spected with period . Given the degradation level of each same random degradation steps of components as in the
component at inspection time , , illustration of the strategy is reused (i.e.,
, the conditional reliability of the system at the next inspec- , , and ).
tion time, , is computed according to The parameters of the maintenance strategy are chosen as
(7), and a decision at the system level based on this conditional , , , and . For such a
reliability is made. If is still greater maintained system, the degradation behavior of components,
than a threshold , nothing is done, and the maintenance de- and the associated system’s and components’ conditional
cision is postponed until the inspection time . reliability at inspection times are respectively sketched in
Otherwise, interventions have to be performed at the compo- Fig. 3(a), and Fig. 3(b). In Fig. 3(a), the notations ,
nent level following the rules of the strategy. and refer respectively to the degradation levels of a
Therefore, at the component level, we replace correctively the component at the time just before, and after the inspection
failed components (i.e., for the or replacement time . We can observe in Fig. 3(b) that the
component , ), and preventively the surviving conditional reliability of the system is regularly evaluated at
components if their conditional reliability at the next inspection each inspection time, while the conditional reliability values of
time given the actual detected degradation level is less than a components are computed at scattered times. The reason for
threshold (i.e., ). this difference is that, under the strategy,
When a component is replaced (either preventively or correc- making a maintenance decision for a component closely
tively) at the inspection time , this is the opportunity to re- depends not only on its own health state but also on the state
place other components, and we opportunistically replace the of other components in the system, or more generally on the
component if their conditional reliability at the next in- state of the global system. This fact leads to the fact that a
spection time given the actual detected degradation level is less component may not be replaced even if its degradation level
than a threshold (i.e., exceeds its failure threshold provided that the global system
). If none of the above conditions apply, nothing is done reliability is above the threshold (see Fig. 3(a)). Table II shows
at , and the maintenance decision is postponed until the next the sequence of maintenance operations performed on each
inspection time . Thus, , , , and component in the time interval [0,200] corresponding to Fig. 3.
are the decision parameters to be optimized. To enhance the Compared to the strategy, the
importance of these decision parameters, we call this strategy strategy usually maintains the system
. reliability at a lower level (see the evolutions of the system
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 9

conditional reliability under both strategies at the top of Because the maintained system returns to an as-good-as-new
Fig. 3(b), and the bottom of Fig. 2(b)), but it can better avoid condition after a replacement, the process repre-
unnecessary maintenance operations. This result is due to senting its degradation behavior is a regenerative process. Thus,
the introduction of conditional reliability threshold into in principle, one can apply the classical renewal-reward theorem
maintenance decision process. In fact, the threshold , on the to evaluate the long-run maintenance cost rate (8) as the ratio of
one hand, assures the compromise between the system relia- the expected cost on a renewal cycle over the expected length
bility and maintenance costs, while on the other hand it brings of a cycle [60, page 41]
out additional opportunities to group maintenance actions on
several components to save set-up costs. Thus, introducing (10)
makes maintenance decisions more flexible. Indeed,
the strategy returns to the where denotes the first replacement time of the whole
strategy when . The former is then more general, and system. However, for the present case, the time between two
hence can lead to lower maintenance cost than the latter. consecutive as-good-as-new system replacements is relatively
long, and hence the evolution of the maintained system on a re-
IV. MAINTENANCE COST MODEL AND OPTIMIZATION newal cycle is very complicated. Consequently, computing the
This section aims at developing a mathematical model based expectation quantities in (10) (i.e., and ) is al-
on the long-run maintenance cost rate to assess the perfor- most impossible. To be able to apprehend the system evolution,
mance of the proposed maintenance strategies. As shown in we have to shorten the study horizon. The semi-regenerative
Section III, the strategy is more general theory represented in [57] can be used for this purpose. In fact,
than the strategy because the former returns to according to the assumptions on the considered system, after
the latter when . Thus we focus here on the cost model a system inspection at time , , the evolution of every
formulation of the strategy, from which the component, and hence of the global system, depends only on
strategy can be directly derived. The cost model the revealed degradation levels. The degradation evolution of
is developed on the basis of semi-regenerative theory [57], and the maintained system is then a semi-regenerative
is optimized by a generalized pattern search algorithm [58]. process, and , where , is its
embedded Markov chain with stationary law .
A. Cost Criterion and Notations Hence, applying the semi-regenerative properties of the process
The long-run maintenance cost rate of the whole system , the long run maintenance cost rate given from (8)
is defined as [59] can be rewritten as [59]

(8)

According to the structure of the proposed strategies, can


be computed by

(11)

where is the first inspection time of the whole system,


and is the expectation with respect to the law . Thus, the
semi-regenerative property allows us to shorten significantly the
(9) study horizon (i.e., from a renewal cycle to an inspection cycle),
and analyzing the behavior of the maintained system is now
where is the random number of system inspections in feasible. However, the price to pay for this advantage is that we
. , and are respectively the random have to compute the expectation quantities with respect to the
number of PR (i.e., either OPR or PPR), and CR of the com- stationary law of the maintained system state ((10) v.s. (11)).
ponent in —independently of the operation performed In the following, we focus on the formulation of the stationary
on other components. is the random number of groups law , and the quantities , ,
for which components, , are replaced at the same , , , and
time in , and denotes the random downtime of the corresponding to the strategy.
system in . The cumulative redundant setup cost in (9) is the But first we have to define our model in parts.
amount of cost savings when the economic dependence among • Let , and
components is considered in maintenance decision-making. be the
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON RELIABILITY

minimum values of degradation level at time such that and the others are opportunistically replaced).
the conditional reliability of component at time ,
given is lower than the PPR threshold , and where , , and
the OPR threshold respectively. Under assumptions of are given from the set
the considered system, is a non- of indices for which ,
increasing function of for a fixed , i.e., . After the replacement, all the
, and thus components are as-good-as-new at , and the next in-
because . spection is scheduled after time units. The probability
• Let , , density corresponding to the degradation levels at
, , the next inspection time is given by ,
, , and where is given from (3).
be respec- • Scenario 2:2 The group of among components
tively the degradation domain corresponding to the is replaced at (among the replaced compo-
actions of no replacement, OPR, PPR, CR, unknown nents, components are opportunistically re-
operation, replacement, and no replacement or OPR on placed), and the others are left in their states.
the component at an inspection time .
• Let , , , , ,
, where ,
, and be the corresponding
degradation domain for a group of different compo- , ,
nents, where . is , and
the subset including indices given from the set of are given from the set of indices for which
indices . When , one obtains the , ,
degradation domain; and when , one obtains the and . The next inspection is
degradation domain for all components. scheduled at time units later. The probability density
• Let corresponding to the degradation levels at
be the degradation domain for which maintenance the next inspection time is .
operations are allowed at component level at an inspection The probability density corresponding to the degrada-
time . is the complement domain of . tion levels at the next inspection time is
.
B. Stationary Law of the Maintained System State • Scenario 3:3 No component is replaced at .
The behavior of the maintained system can be characterized , or . The system
by the stationary law of the Markov chain . Let degradation state at is the same as the one at ,
, and be the system health states detected at the begin- and the next inspection is scheduled at time units later,
ning (i.e., ), and the end (i.e., ) of one semi-regenerative hence the probability density corresponding to the degra-
cycle. The stationary law is the solution of the following in- dation levels at the next inspection time is given
variant equation [61, page 97]. by .
As such, the transition law is computed
(12)
as shown in (13) at the bottom of the next page. Thus, from
(12) and (13), the invariant equation of the stationary law
where is the transition law from the state to is rewritten as shown in (14) at the bottom of the next
the state . Note that, because we considered a semi-regener- page. Mathematically, (14) has the form of a homogeneous
ative cycle from to , the integration in (12) is taken over multidimensional integral equation of the second kind [64].
the domain . This condition is slightly different than the in- We have to solve such an integral equation to obtain the
variant equation studied in the series [13], [62], [63] of Dieulle stationary law . From a numerical point of view, due
et al., where the integration domain is just the working zone of to the regularity of the densities , the equation
the system, because these authors considered a semi-regenera- can be solved using an iterative method based on the method
tive cycle taken from to . This difference is due to the of successive approximations. As an example, we present in
fact that maintenance actions can be performed at , and that the Appendix an iterative algorithm which can be applied to
the degradation levels of the system at and can take determine the stationary law of a 2-components system subject
different values. The transition law in (12) can to the strategy. This algorithm is easily
be obtained by an exhaustive analysis of all the possible evo- extended for higher dimensional systems (see e.g., Algorithm 2
lution and maintenance scenarios on a semi-regenerative cycle. represented in the Appendix for a 3-components system). Many
These scenarios are as follows. 2Scenario 2 is composed of sub-scenarios de-
• Scenario 1:1 All components are replaced at (i.e., pending on the value of and , and the permutations of
components are preventively or correctively replaced, .
3Contrary to Scenarios 1 and 2, Scenario 3 cannot be divided into sub-
1Scenario 1 is composed of sub-scenarios depending on the value scenarios.
of , and the permutations of .
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 11

numerical tests were carried out, and they have shown that the Also, on a semi-renewal cycle, one and only one inspection is
algorithm converges quickly to the true stationary law (see also performed. Thus, the expected number of system inspections at
the Appendix). Fig. 4 illustrates the shape of the stationary law steady state with respected to is obtained by
of the 2-component parallel maintained system char-
acterized by the data set , , (16)
, , , ,
Consider now the component , , in the system.
. In Fig. 4(a), the surface of is given by
On a semi-renewal cycle, if it is preventively replaced, it
numerical calculation of (30) with Algorithm 1 presented in
is replaced only once. Furthermore, the replacement can
the Appendix, while in Fig. 4(b) it is returned by simulation
be either OPR or PPR. So, the expected number of PR of
with the kernel density estimation method developed in [65].
the component on a semi-renewal cycle at steady state
The similitude of both surfaces confirms the exactitude of
with respected to is computed as shown in (17) at the
the proposed formulas, as well as the numerical computation
bottom of the next page, where ,
approach.
and are respectively
C. Expectation Quantities With Respected to the Stationary the subsets including indices, and given
Law from the set of indices .
denotes the sum over all permutations
As aforementioned, a semi-renewal cycle corresponds to an
of for which
inspection period. The expected length of a semi-renewal cycle
and , denotes a
at steady state with respect to is then given by
vector of elements among , and
denotes a vector of elements among
(15) . Similarly, the expected number

(13)

(14)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON RELIABILITY

Fig. 4. Shape of the stationary law of the 2-component parallel system under the strategy. (a) Surface of given by
numerical calculation of (30). (b) Surface of given by kernel density estimation method.

of CR of component on one semi-renewal cycle at steady Finally, the expected system downtime on one semi-renewal
state with respected to is given by cycle at steady state with respected to is given by

(18)

For a group of components, its expected replacement (20)


number on one semi-renewal cycle at steady state with
respected to can be expressed as shown in (19) at the where is the conditional reliability of the main-
bottom of the page, where , tained system under the strategy at time
, and are given the system degradation state at time , .
respectively the subsets including indices, indices, Note that does not simply denote the initial time at which the
and indices given from the set of indices . system starts working, but rather indicates the end of the latest

(17)

(19)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 13

semi-renewal cycle. This means that the degradation levels of the strategy, we apply the intervention
components at times and can be different due to main- costs , , , , and
tenance operations. The expression of shares the on the 2-component parallel system characterized by the
same form of (7), but the elementary conditional reliability and set of parameters , , and
conditional distribution functions constituting are . Fig. 5 shows the numerical result. The
different. That is, evolution of the maintenance cost rate as
a function of is sketched in Fig. 5(a). Each point on the
cost rate curve is obtained by assigning a value in the interval
[0.5,1] to , and setting the other parameters at their optimal
values that correspond to the taken value of . Thus, the point
associated with indicates the optimal maintenance cost
(21) rate of the strategy. Fig. 5(a) shows clearly that
one can adjust the value of to obtain a lower maintenance
where the conditional distribution functions cost rate than the optimal one of the strategy.
are computed as Fig. 5(b), Fig. 5(c), and Fig. 5(d) show respectively the
variation of the maintenance cost rate of the
strategy when 2 among 3 decision parameters
vary, while the remaining parameter and are fixed.
if the component is These cost rate curves and cost rate surfaces are found by
numerical computation, and verified by Monte Carlo (MC)
replaced at time 0 (22)
simulation (see also Table III and Table IV in the Appendix).
otherwise. The convexity of the cost rate curves and surfaces shows
the existence of an optimal setting of decision parameters.
Mathematically, the event that the component is replaced at Actually, given the mathematical expression of the cost rate
time 0 can be expressed as , and , many optimization algo-
rithms can be applied for finding these values (see e.g., [66]).
In the present paper, we use the generalized pattern search
algorithm considered in [58]. Fig. 6(a) and Fig. 6(b) represent
respectively the optimal results of the strategy,
and the strategy for the aforementioned
maintained system. In each figure, the bar chart on top shows
(23) the initial values and the optimal values of decision parameters,
while the dot plot at the bottom represents the evolution
The conditional reliability is then obtained
of the maintenance cost rate towards the optimal value. In
by fact, for the chosen system parameters and maintenance
costs, the strategy obtains its optimal cost rate
(24)
at ,
where is given from (22). and ; the strategy
obtains its optimal cost rate at
D. Maintenance Strategies Optimization , , , and
. These results are completely coherent with
Using (14), and introducing (15), (16), (17), (18), (19), (20) the ones given from the cost rate curve and surfaces represented
into (11), one obtains the cost model of the in Fig. 5. Comparing the optimal decision parameters of both
strategy. Optimizing the strategy means maintenance strategies, one can see that, thanks to the decision
finding the optimal values , , , and threshold at system level (i.e., ), the
that minimize the long-run maintenance cost rate strategy can really lead to more savings in maintenance cost
where compared to the strategy.

V. VALUE OF MULTI-LEVEL MAINTENANCE DECISION-MAKING


In previous sections, a multi-level maintenance decision is
achieved by combining the maintenance decision-making at
(25) the system level and component level. Although effectively
avoiding redundant maintenance operations, such a decision
The strategy is a special case of the does not always assure the highest reliability level of the
strategy, so its cost model can be directly system. It is then interesting to find out the configurations for
derived from the latter model by taking or . which this kind of maintenance decision is really efficient. For
For an illustration on the long-run maintenance cost rate of this end, we perform comparison studies on the performance of
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON RELIABILITY

Fig. 5. Shape of the long-run expected maintenance cost rate of a 2-components parallel system under the strategy. (a) as
a function of . (b) as a function of and . (c) as a function of and . (d) as a function of
and .

TABLE III criterion for the performance comparison is the relative gain
VALUES OF EXPECTATION QUANTITIES AND MAINTENANCE COST RATE in the optimal maintenance cost rate of the
OBTAINED FROM NUMERICAL COMPUTATION AND MC SIMULATION
strategy compared to the strategy [67]

(26)

TABLE IV If , the
VALUES OF EXPECTATION QUANTITIES AND MAINTENANCE COST RATE GIVEN
FROM NUMERICAL COMPUTATION AND MC SIMULATION
strategy is more profitable than the
strategy in terms of the maintenance cost;
if ,
both strategies have the same profit. Note that the case
does not
exist because the strategy is more gen-
eral than the strategy. Moreover, for an easier
analysis, we consider that different components incur the
same cost for the same type of maintenance operation (i.e.,
the strategy and the strategy , and , for ). This study
under different situations of system characteristics (i.e., het- allows us to evaluate the value of multi-level maintenance
erogeneity and structure) and maintenance costs. The used decision-making.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 15

Fig. 6. Illustration of optimal results returned by a generalized pattern search algorithm. (a) Optimal solution of strategy. (b) Optimal solution of
strategy.

A. Sensitivity to the System Heterogeneity strategy compared to the


strategy.
The heterogeneity of a multi-component system can be
reflected through the differences in average degradation rate B. Sensitivity to the Maintenance Costs
and variance among their components. To study the impact To analyze the sensitivity of maintenance strategy per-
of the system heterogeneity on the performance of mainte- formance to the maintenance costs, we vary one of the
nance strategies, we consider then a 2-components parallel maintenance costs (i.e., inspection cost , PR cost ,
system, and we investigate the evolution of the relative gain system downtime cost rate , or set-up cost ), fix the
when the average other costs, and investigate the evolution of the relative gain
degradation rate and variance of a component (e.g., component . Without loss of
2) are fixed, and the parameters of the other (i.e., component 1) generality, one can fix the CR cost at , and consider
vary. Thus, two case studies are obtained. the four following cases.
• Difference in average degradation rate: • Varied inspection cost: varies from 1 to 15 with compo-
, , varies from 1 to 4 with nent step 2, , , and .
component step 0.25, and . • Varied PR cost: , varies from 30 to 90 with com-
• Difference in degradation variance: , ponent step 10, , and .
, , and varies from • Varied downtime cost rate: , , varies from
1/3 to 10/3 with component step 1/3. 10 to 100 with component step 10, and
Other parameters and maintenance costs are chosen as • Varied set-up cost: , , , and varies
, , , , and . from 5 to 50 with component step 5.
Fig. 7(a), and Fig. 7(b) show the numerical results for the first The above case studies are applied to a 2-component parallel
case, and second case respectively. Generally, the relative gain system defined by the set of parameters ,
tends to increase with the growth of the ratio of degradation rate , and . Note that
, because it creates a higher variety among components, the choice of a parallel system with 2 identical compo-
and hence introducing an additional decision at system level nents allows an easy and fair study. Numerical results are
becomes more significant. This condition is effectively true illustrated by Fig. 8(a), Fig. 8(b), Fig. 8(c), and Fig. 8(d).
when is not small and not too large (see Fig. 7(a)). Fig. 8(a) and Fig. 8(c) show that the relative gain
In fact, a small ratio (i.e., for the present is a decreasing
case study) implies that the components in the system are function of inspection cost , and system downtime cost
almost similar, while a too high ratio (i.e., for rate . Indeed, when the inspection cost is expensive,
the present case study) indicates that the global system health the inter-inspection interval is set at a high value to avoid
state closely depends on the state of only one component. more frequent inspections; this setting leads to longer
For such configurations, the decision at system level loses its system downtime. As argued in previous sections, the
significance, and consequently, one obtains an almost constant strategy cannot assure effectively a
relative gain as in Fig. 7(a). One can remark from Fig. 7(b) that highest reliability for the system, so it loses its performance
the relative gain when the downtime interval or the downtime cost increases.
is more or less constant with a varied value of the variance ratio When or is too high, the threshold is set to a value
. It seems that the difference in degradation variance close to 1 to allow interventions as soon as possible at the
among components does not influence the performance of the component level. As such, the strategy
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE TRANSACTIONS ON RELIABILITY

Fig. 7. Impact of the system heterogeneity on the relative gain in the optimal maintenance cost rate. (a) Difference in average degradation rate. (b) Difference in
degradation variance.

returns to the strategy, and the relative gain result, parallel structures (i.e., ) are best suitable to the
tends to 0 (see strategy (see Fig. 9).
Fig. 8(a), and Fig. 8(c)). In Fig. 8(b), and Fig. 8(d), the relative
gain increases with the growth of PR cost , and set-up cost D. General Discussions on the Value of Multi-Level CBM
. The reason for this gain is that the Decision-Making
strategy can avoid effectively redundant maintenance opera-
From the above numerical studies, one can see that, thanks
tions, and allows more opportunities to replace components
to additional intervention decisions at system level, the
together. So, the higher the value of and , the more the
strategy manages better the compromise
strategy is profitable.
between maintenance operations and system reliability. Hence,
the strategy allows more savings in mainte-
C. Sensitivity to the System Structure
nance cost compared to the strategy, and returns
For an easier analysis, only one simple system structure (i.e., to the latter in the worst cases. Thus, for any situation, there
a 2-components parallel system) has been considered in the pre- is no risk in using the multi-level decision-making approach
vious sections. This section studies more deeply the impact of instead of traditional approaches (i.e., decision-making at either
the system structure on the performance of the proposed main- the component level or the system level). In fact, this approach
tenance strategies. For this purpose, we consider a -out-of- :F is especially effective if the system is highly heterogeneous,
system whose components are identical, and each of them consists of many functioning components, the PR cost or set-up
(e.g., component ) is characterized by the parameters , cost are expensive, while the inspection cost or the downtime
, and . The maintenance cost is chosen as cost rate of system are not too high. Otherwise, the advantage
, , , , and . Fig. 9 shows of multi-level decision-making is not really that important, and
the relative gain traditional decision-making approaches could be used.
corresponding to diverse -out-of- :F structures. One can
remark that the relative gain is higher when the number of
VI. CONCLUSIONS AND PERSPECTIVES
required functioning components for the system (i.e., ) in-
creases. In fact, for series structures (i.e., ), the reliability Based on the mathematical cost model of reliability-based
of the system is smaller than the reliability of their components. opportunistic predictive maintenance strategies developed for
The strategy applied to such a system has -out-of- :F deteriorating systems, the main aim of this paper
to set the optimal value of at 1 to allow as quickly as is to assess the value of a multi-level decision-making approach
possible interventions at the component level. In other words, for predictive maintenance. Numerous theoretical analyses and
the strategy returns to the numerical experiments show that this approach allows flexible
strategy, and . strategies, and hence can lead to more savings in maintenance
When increases, the system reliability is higher, and more cost compared to traditional approaches (i.e., decision-making
maintenance operations are required. The at either the component level or system level). What is the most
strategy, which avoids effectively redundant maintenance interesting is that there is no risk in using the multi-level deci-
operations but cannot always assure a highest system re- sion-making approach (i.e., it returns to traditional approaches
liability, becomes more significant, and the relative gain in the worst cases). In addition, the results in this paper also
increases. As a confirm the effectiveness of prognostic condition indices for
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 17

Fig. 8. Impact of the maintenance costs on the relative gain in the optimal maintenance cost rate. (a) Varied inspection cost. (b) Varied PR cost. (c) Varied
downtime cost rate. (d) Varied set-up cost.

level can make maintenance strategies not only robust, but also
simpler and easier to be optimized (see also [35], [37]). Thus,
to exploit the highest performance of maintenance operations,
the multi-level decision-making approach and prognostic con-
dition indices should be used together to build a maintenance
strategy.
Despite our interesting results, the present paper is still
a preliminary study of the multi-level decision-making ap-
proach. The advantage of this approach should be further
investigated with more complex systems, or more practical
assumptions. Regarding complex systems, one can think about
-out-of- systems subject to diverse types of interactions
(i.e., stochastic and structural dependence in addition to eco-
nomic dependence), or systems with more complex structures.
Fig. 9. Impact of the system structure on the relative gain in the optimal More practical assumptions can be related to systems with
maintenance cost rate. multiple failure modes, imperfect observations of the state of
systems, imperfect maintenance operations, and non-negligible
maintenance times. Applying the proposed concept to develop
maintenance decision-making in the context of multi-compo- joint maintenance and spare part provisioning strategies is also
nent systems. In fact, the use of RUL instead of degradation a potential perspective.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

18 IEEE TRANSACTIONS ON RELIABILITY

APPENDIX

Explicit Formulation of the Strategy for the


Case of a 2-Component System
We present in this appendix the explicit expressions of
the stationary law , the expectation quantities
, , , ,
as well as the expected cost rate of a 2-component (either series
or parallel) system subjected to the strategy.
An iterative algorithm to derive the stationary law is
also proposed, and numerically evaluated.
Stationary Law of the Maintained System State: Let’s define
the probabilities , , and as follows.
• The marginal probability associated with a pure replace-
ment (i.e., PPR and CR) of component 2

(27) Fig. 10. Numerical illustration of the convergence of , , and .


(a) Evolution of according to iteration number. (b) Evolution of
according to iteration number. (c) Evolution of according to iteration
where is a simplified notation of defined in number.
Section IV-A, which depends closely on the system struc-
ture (either series or parallel).
• The marginal probability associated with a pure replace- Now, we propose an iterative algorithm to solve (30), shown as
ment of component 1 Algorithm 1.

(28)
Algorithm 1 Iterative algorithm to solve (30).

• The marginal probability associated with a joint replace- 1) Initialization: Set any random values in the interval (0,1]
ment (i.e., either pure replacement or OPR) of the 2 com- for , , and .
ponents 2) Step 1: Introduce , and into (30) to
determine , and decide.
• If converges, end of algorithm.
(29) • Otherwise, go to Step 2.
3) Step 2: Compute the new , , and
according to (27), (28), and (29); and go to Step 1.
Let be the complement domain of . The stationary 4) Output: Converged values of , , , and
law of the state of a 2-component maintained system can be .
explicitly expressed as

Applying Algorithm 1 for the 2-component parallel main-


tained system defined by the set of parameters ,
, , , ,
, and , one obtains the shape of the sta-
tionary law as in Fig. 4. Fig. 10 illustrates the evolu-
tion of , , and according to iteration number.
The figure shows clearly that , , and converge
quickly to their real functions or value (i.e., since the 7-th iter-
ation), and thus the proposed algorithm determines efficiently
the stationary law .
Expectation Quantities With Respected to the Stationary
Law: Let’s define the probabilities , as

(31)
(30)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 19

and Thus, considering , and , the


long-run maintenance cost rate of a 2-component system subject
to the strategy can be expressed as
(32)

Under the strategy, the expectation (38)


quantities , , ,
of a 2-component system are given as follows. where , , , and
• The expected number of PR (either opportunistic or not) are given from (33), (34), (35), and (36) respec-
on a semi-renewal cycle at steady state is tively.
Applying (33), (34), (35), (36), and (38) for the 2-component
maintained system defined with the set of parameters
, , , ,
, , , , , ,
, and , one obtains the values of the expecta-
tion quantities , , ,
, and the maintenance cost rate as in Table III.
Almost identical results are also given from MC simulation.
This concordance validates the derived expressions.
(33)
where and are given from (27) and (28). Explicit Formulation of the Strategy for the
• The expected number of CR on one semi-renewal cycle at Case of a 3-Component System
steady state is
We present in this appendix the explicit expressions of
the stationary law , the expectation quantities
, , , ,
as well as the expected cost rate of a 3-component system sub-
(34) jected to the strategy. An iterative algorithm
to derive the stationary law is also proposed and
numerically evaluated.
• The expected number of group replacements on one semi- Stationary Law of the Maintained System State: Let’s define
renewal cycle at steady state is the probabilities , , , and
as follows.
(35) • The marginal probability associated with a pure replace-
ment (i.e., PPR and CR) of component 3 is
where is given from (29).
• The expected downtime on one semi-renewal cycle at
steady state is
(39)

where is a simplified notation of defined in


Section IV-A, which depends closely on the system struc-
(36) ture.
• The marginal probability associated with a pure replace-
where is given from (21). The ex- ment of component 2 is
pression of depends on the system
structure (i.e., series or parallel) and can be obtained as
(40)

if 1-out-of-2:F
structure (series) • The marginal probability associated with a pure replace-
if 2-out-of-2:F ment of component 1 is
structure (parallel)
(37)
(41)
where is given from (24).
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

20 IEEE TRANSACTIONS ON RELIABILITY

• The marginal probability associated with a joint replace- explicitly expressed as (43), shown at the bottom of the page.
ment (i.e., either pure replacement or OPR) of the 3 com- The stationary law can be obtained by solving (43)
ponents is by Algorithm 2.

Algorithm 2 Iterative algorithm to solve (43).

1) Initialization: Set any random values in the interval (0,1]


for , , , and .
2) Step 1: Introduce , , , and
into (43) to determine , and decide.
• If converges, end the algorithm.
• Otherwise, go to Step 2.
(42) 3) Step 2: Compute the new , ,
, and according to (39), (40), (41), and
(42); and go to Step 1.
Let be the complement domain of . The stationary 4) Output: Converged values of , ,
law of the state of a 3-components maintained system can be , , and .

(43)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 21

Under the strategy, the expectation


quantities , , ,
of a 3-components system are given as follows.
• The expected number of PR (either opportunistic or not)
on a semi-renewal cycle at steady state is shown in (47)
at the bottom of the page, where , ,
and are given from (39), (40), and (41).
• The expected number of CR on one semi-renewal cycle at
steady state is

(48)
Fig. 11. Evolution of as a function of iteration number.
• The expected number of group replacements on one semi-
renewal cycle at steady state is
To illustrate the convergence of Algorithm 2, we apply it
for the 2-out-of-3:F system defined by the parameters
, , , ,
, , and ; and we sketch the evolution
of as a function of iteration number. Fig. 11 shows clearly
that converges quickly to its true value. Algorithm 2 also
converges quickly. (49)
Expectation Quantities With Respected to the Stationary where is expressed as
Law: Let’s define the probabilities , , and
as

(44)

(45)

and

(46)
(50)

(47)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

22 IEEE TRANSACTIONS ON RELIABILITY

if 1-out-of-3:F structure

if 2-out-of-3:F structure (54)

if 3-out-of-3:F structure

is computed by , , , and the mainte-


nance cost rate as in Table IV. The identical results given
by both approaches validate the derived expressions.

REFERENCES
[1] H. Wang, “A survey of maintenance policies of deteriorating systems,”
Eur. J. Oper. Res., vol. 139, no. 3, pp. 469–489, 2002.
[2] R. Ahmad and S. Kamaruddin, “A review of condition-based mainte-
nance decision-making,” Eur. J. Ind. Eng., vol. 6, no. 5, pp. 519–541,
2012.
[3] L. Mann, A. Saxena, and G. M. Knapp, “Statistical-based or condition-
based preventive maintenance?,” J. Qual. Maint. Eng., vol. 1, no. 1, pp.
(51) 46–59, 1995.
[4] K. T. Huynh, A. Barros, C. Bérenguer, and I. T. Castro, “A periodic
inspection and replacement policy for systems subject to competing
failure modes due to degradation and traumatic events,” Rel. Eng. Syst.
and is obtained by Safety, vol. 96, no. 04, pp. 497–508, 2011.
[5] R. Ahmad and S. Kamaruddin, “An overview of time-based and condi-
(52) tion-based maintenance in industrial application,” Comput. Ind. Eng.,
vol. 63, no. 1, pp. 135–149, 2012.
[6] R. Kothamasu, S. H. Huang, and W. H. VerDuin, “System health moni-
toring and prognostics—A review of current paradigms and practices,”
where is given from (42). Int. J. Adv. Manuf. Technol., vol. 28, no. 9, pp. 1012–1024, 2006.
[7] W. Wang, “Condition-based maintenance modelling,” in Complex
• The expected downtime on one semi-renewal cycle at System Maintenance Handbook, ser. Springer Series in Reliability
steady state is Engineering. London, U.K.: Springer, 2008, pp. 111–131.
[8] R. Bogue, “Sensors for condition monitoring: A review of technologies
and applications,” Sensor Rev., vol. 33, no. 4, pp. 295–299, 2013.
[9] K. T. Huynh, A. Barros, and C. Bérenguer, “Maintenance deci-
sion-making for systems operating under indirect condition mon-
itoring: Value of online information and impact of measurement
(53) uncertainty,” IEEE Trans. Rel., vol. 61, no. 2, pp. 410–425, 2012.
[10] A. S. K. Jardine, D. Lin, and D. Banjevic, “A review on machinery
diagnostics and prognostics implementing condition-based mainte-
where is given from (21). Specifi- nance,” Mech. Syst. Signal Process., vol. 20, no. 7, pp. 1483–1510,
2006.
cally, depends closely on the system [11] J. M. van Noortwijk, “A survey of the application of gamma processes
structure, and can be expressed as (54) at the top of the in maintenance,” Rel. Eng. Syst. Safety, vol. 94, no. 1, pp. 2–21, 2009.
page, where is given from (24). [12] A. Prajapati, J. Bechtel, and S. Ganesan, “Condition based mainte-
nance: A survey,” J. Qual. Maint. Eng., vol. 18, no. 4, pp. 384–400,
Thus, considering , and 2012.
[13] A. Grall, L. Dieulle, C. Bérenguer, and M. Roussignol, “Continuous-
, the long-run maintenance cost rate of a 3-compo- time predictive-maintenance scheduling for a deteriorating system,”
nents system subject to the strategy can be IEEE Trans. Rel., vol. 51, no. 2, pp. 141–150, 2002.
expressed as [14] D. I. Cho and M. Parlar, “A survey of maintenance models for multi-
unit systems,” Eur. J. Oper. Res., vol. 51, no. 1, pp. 1–23, 1991.
[15] R. P. Nicolai and R. Dekker, “Optimal maintenance of multi-com-
ponent systems: A review,” in Complex System Maintenance Hand-
book, ser. Springer Series in Reliability Engineering. London, U.K.:
(55) Springer, 2008, pp. 263–286.
[16] B. Castanier, A. Grall, and C. Bérenguer, “A condition-based main-
where , , , and tenance policy with non-periodic inspections for a two-unit series
system,” Rel. Eng. Syst. Safety, vol. 87, no. 1, pp. 109–120, 2005.
are given from (47), (48), (49), and (53) respec- [17] K. Bouvard, S. Artus, C. Bérenguer, and V. Cocquempot, “Condi-
tively. tion-based dynamic maintenance operations planning & grouping. ap-
Verifying the above expressions with the 3-component plication to commercial heavy vehicles,” Rel. Eng. Syst. Safety, vol. 96,
no. 6, pp. 601–610, 2011.
system defined by the parameters , [18] Z. Tian and H. Liao, “Condition based maintenance optimization for
, , , , multi-component systems using proportional hazards model,” Rel.
, , , , , , and Eng. Syst. Safety, vol. 96, no. 5, pp. 581–589, 2011.
[19] Z. Tian, T. Jin, B. Wu, and F. Ding, “Condition based maintenance op-
by numerical computation and MC simulation, one timization for wind power generation systems under continuous mon-
obtains the values of expectation quantities , itoring,” Renew. Energy, vol. 36, no. 5, pp. 1502–1509, 2011.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HUYNH et al.: MULTI-LEVEL DECISION-MAKING FOR THE PREDICTIVE MAINTENANCE OF -OUT-OF- :F DETERIORATING SYSTEMS 23

[20] J. Hu, L. Zhang, and W. Liang, “Opportunistic predictive maintenance [44] W. Wang, “Overview of a semi-stochastic filtering approach for
for complex multi-component systems based on dbn-hazop model,” residual life estimation with applications in condition based mainte-
Process Safety Environ. Protect., vol. 90, no. 5, pp. 376–388, 2012. nance,” Proc. Inst. Mech. Eng., Part O: J. Risk Rel., vol. 225, no. 2,
[21] Z. Zhang, Y. F. Zhou, Y. Sun, and L. Ma, “Condition-based main- pp. 185–197, 2011.
tenance optimization without a predetermined strategy structure for a [45] X. S. Si, W. Wang, C. H. Hu, and D. H. Zhou, “Remaining useful life
two-component series system,” Eksploatacja i Niezawodnosc—Maint. estimation—A review on the statistical data driven approaches,” Eur.
Rel., vol. 14, no. 2, pp. 120–129, 2012. J. Oper. Res., vol. 213, no. 1, pp. 1–14, 2011.
[22] A. Barros, C. Bérenguer, and A. Grall, “Optimization of replacement [46] L. Liao and F. Kottig, “Review of hybrid prognostics approaches for
times using imperfect monitoring information,” IEEE Trans. Rel., vol. remaining useful life prediction of engineered systems, and an appli-
52, no. 4, pp. 523–533, 2003. cation to battery life prediction,” IEEE Trans. Rel., vol. 63, no. 1, pp.
[23] A. Gupta and C. Lawsirirat, “Strategically optimum maintenance of 191–207, 2014.
monitoring-enabled multi-component systems using continuous-time [47] D. Banjevic, “Remaining useful life in theory and practice,” Metrika,
jump deterioration models,” J. Qual. Maint. Eng., vol. 12, no. 3, pp. vol. 69, no. 2, pp. 337–349, 2009.
306–329, 2006. [48] E. Myötyri, U. Pulkkinen, and K. Simola, “Application of stochastic
[24] F. Camci, “System maintenance scheduling with prognostics informa- filtering for lifetime prediction,” Rel. Eng. Syst. Safety, vol. 91, no. 2,
tion using genetic algorithm,” IEEE Trans. Rel., vol. 58, no. 3, pp. pp. 200–208, 2006.
539–552, 2009. [49] H. Liao and Z. Tian, “A framework for predicting the remaining useful
[25] L. Bian and N. Gebraeel, “Stochastic modeling and real-time prognos- life of a single unit under time-varying operating conditions,” IIE
tics for multi-component systems with degradation rate interactions,” Trans., vol. 45, no. 9, pp. 964–980, 2013.
IIE Trans., vol. 46, no. 5, pp. 470–482, 2013. [50] K. T. Huynh, A. Barros, and C. Bérenguer, “On the opportunistic pre-
[26] N. Rasmekomen and A. K. Parlikad, “Maintenance optimization for dictive maintenance decision-making for -out-of- :f deteriorating
asset systems with dependent performance degradation,” IEEE Trans. systems,” in Proc. 19th ISSAT Int. Conf. Reliability and Quality in De-
Rel., vol. 62, no. 2, pp. 362–367, 2013. sign, 2013, pp. 265–269.
[27] A. van Horenbeek and L. Pintelon, “A dynamic predictive maintenance [51] M. Abdel-Hameed, “Degradation processes: An overview,” in Ad-
policy for complex multi-component systems,” Rel. Eng. Syst. Safety, vances in Degradation Modeling, ser. Springer series in Statistics for
vol. 120, pp. 39–50, 2013. Industry and Technology. Boston, MA, USA: Birkhäuser, 2010, pp.
[28] W. Kuo and M. J. Zuo, Optimal Reliability Modeling: Principles and 17–25.
Applications. New York, NY, USA: Wiley, 2003. [52] H. A. David and H. N. Nagaraja, Order Statistics, 3rd Edition. New
[29] L. M. Maillart, C. R. Cassady, C. Rainwater, and K. Schneider, “Selec- York, NY, USA: Wiley, 2003.
tive maintenance decision-making over extended planning horizons,” [53] K. Le Son, M. Fouladirad, A. Barros, E. Levrat, and B. Iung, “Re-
IEEE Trans. Rel., vol. 58, no. 3, pp. 462–469, 2009. maining useful life estimation based on stochastic deterioration
[30] T. W. Lin and C. H. Wang, “A hybrid genetic algorithm to minimize models: A comparative study,” Rel. Eng. Syst. Safety, vol. 112, pp.
the periodic preventive maintenance cost in a series-parallel system,” 165–175, 2012.
J. Intell. Manuf., vol. 23, no. 4, pp. 1225–1236, 2012. [54] L. Wang, J. Chu, and W. Mao, “A condition-based replacement and
[31] Y. Zhou, Z. Zhang, T. Ran Lin, and L. Ma, “Maintenance optimisation spare provisioning policy for deteriorating systems with uncertain de-
of a multi-state series-parallel system considering economic depen- terioration to failure,” Eur. J. Oper. Res., vol. 194, no. 1, pp. 184–205,
dence and state-dependent inspection intervals,” Rel. Eng. Syst. Safety, 2009.
vol. 111, pp. 248–259, 2013. [55] S. G. Jalali Naini, M. B. Aryanezhad, A. Jabbarzadeh, and H. Babaei,
[32] L. C. Thomas, “A survey of maintenance and replacement models for “Condition based maintenance for two-component systems with relia-
maintainability and reliability of multi-item systems,” Rel. Eng., vol. bility and cost considerations,” Int. J. Ind. Eng. Prod. Res., vol. 20, no.
16, no. 4, pp. 297–309, 1986. 3, pp. 107–116, 2009.
[33] R. Dekker, R. E. Wildeman, and F. A. van der Duyn Schouten, “A re- [56] M. Gallo, G. Guizzi, and P. Zoppoli, “Production yield improvement in
view of multi-component maintenance models with economic depen- a series system by a multithreshold CBM model,” Int. J. Syst. Applicat.,
dence,” Math. Meth. Oper. Res., vol. 45, no. 3, pp. 411–435, 1997. Eng. Develop., vol. 1, no. 4, pp. 100–111, 2007.
[34] R. E. Wildeman, R. Dekker, and A. C. J. M. Smit, “A dynamic policy [57] S. Asmussen, Applied Probability and Queues, ser. Stochastic Model-
for grouping maintenance activities,” Eur. J. Oper. Res., vol. 99, no. 3, ling and Applied Probability. New York, NY, USA: Springer-Verlag,
pp. 530–551, 1997. 2003.
[35] K. T. Huynh, S. T. Mai, A. Grall, and C. Bérenguer, “Towards a reliable [58] C. Audet, J. Dennis, and E. John, “Analysis of generalized pattern
condition index for condition-based maintenance decision-making,” in searches,” SIAM J. Optimiz., vol. 13, no. 3, pp. 889–903, 2002.
Proc. 2nd Int. Conf. Control and Fault-Tolerant Systems, SysTol’13, [59] C. Bérenguer, “On the mathematical condition-based maintenance
2013, pp. 826–831. modelling for continuously deteriorating systems,” Int. J. Mater.
[36] K. T. Huynh, I. T. Castro, A. Barros, and C. Bérenguer, “On the use Struct. Rel., vol. 6, no. 2, pp. 133–151, 2008.
of mean residual life as a condition index for condition-based mainte- [60] H. C. Tijms, A First Course in Stochastic Models. New York, NY,
nance decision-making,” IEEE Trans. Syst., Man, Cybern.: Syst., 2014, USA: Wiley, 2003.
10.1109/TSMC.2013.2290772. [61] S. Asmussen and P. W. Glynn, Stochastic Simulation: Algorithms and
[37] K. T. Huynh, A. Barros, and C. Bérenguer, “A reliability-based op- Analysis, ser. Stochastic Modelling and Applied Probability. New
portunistic predictive maintenance model for k-out-of-n deteriorating York, NY, USA: Springer, 2007.
systems,” Chem. Eng. Trans., vol. 33, pp. 493–498, 2013. [62] L. Dieulle, C. Bérenguer, A. Grall, and M. Roussignol, “Continuous
[38] A. Grall, C. Bérenguer, and L. Dieulle, “A condition-based mainte- time predictive maintenance scheduling for a deteriorating system,”
nance policy for stochastically deteriorating systems,” Rel. Eng. Syst. in Proc. Annu. Reliability and Maintainability Symp., 2001, pp.
Safety, vol. 76, no. 2, pp. 167–180, 2002. 150–155.
[39] Z. Xu, Y. Ji, and D. Zhou, “Real-time reliability prediction for a dy- [63] L. Dieulle, C. Bérenguer, A. Grall, and M. Roussignol, “Sequential
namic system based on the hidden degradation process identification,” condition-based maintenance scheduling for a deteriorating system,”
IEEE Trans. Rel., vol. 57, no. 2, pp. 230–242, 2008. Eur. J. Oper. Res., vol. 150, no. 2, pp. 451–461, 2003.
[40] Z. Xu, Y. Ji, and D. Zhou, “A new real-time reliability prediction [64] K. E. Atkinson, The Numerical Solution of Integral Equations of the
method for dynamic systems based on on-line fault prediction,” IEEE Second Kind. Cambridge, U.K.: Cambridge Univ. Press, 1997.
Trans. Rel., vol. 58, no. 3, pp. 523–538, 2009. [65] Z. I. Botev, J. F. Grotowski, and D. P. Kroese, “Kernel density es-
[41] D. M. Frangopol, M. J. Kallen, and J. M. van Noortwijk, “Probabilistic timation via diffusion,” Ann. Statist., vol. 38, no. 5, pp. 2916–2957,
models for life-cycle performance of deteriorating structures: Review 2010.
and future directions,” Progr. Struct. Eng. Mater., vol. 6, no. 4, pp. [66] A. D. Belegundu and T. R. Chandrupatla, Optimization Concepts and
197–212, 2004. Applications in Engineering, 2nd Edition. Cambridge, U.K.: Cam-
[42] N. D. Singpurwalla, “Survival in dynamic environments,” Statist. Sci., bridge Univ. Press, 2011.
vol. 10, no. 1, pp. 86–103, 1995. [67] K. T. Huynh, I. T. Castro, A. Barros, and C. Bérenguer, “Modeling age-
[43] J. Z. Sikorskaa, M. Hodkiewicz, and L. Ma, “Prognostic modelling based maintenance strategies with minimal repairs for systems subject
options for remaining useful life estimation by industry,” Mech. Syst. to competing failure modes due to degradation and shocks,” Eur. J.
Signal Process., vol. 25, no. 5, pp. 1803–1836, 2011. Oper. Res., vol. 218, no. 1, pp. 140–151, 2012.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

24 IEEE TRANSACTIONS ON RELIABILITY

Khac Tuan Huynh received his M.S., and Ph.D. degrees in System Safety and Christophe Bérenguer is a Professor of reliability engineering and control
Optimization from the Troyes University of Technology (UTT), France in 2008, systems at Gipsa-Lab and Grenoble Institute of Technology (Grenoble, France).
and 2011 respectively. From 1997 to 2011, he was Professor at Troyes University of Technology
He is an Associate Professor at UTT, and is a member of the Systems Mod- (Troyes, France). He has served as an officer (treasurer) for the European
eling and Dependability Laboratory. His research interests include stochastic Safety and Reliability Association from 2005 to 2010. His research interests
modeling of systems deterioration, prognostic and health management tech- include system health monitoring, stochastic modeling of systems deterioration,
niques, and maintenance decision-making, optimization, and evaluation. He has performance evaluation and optimization of dynamic maintenance policies,
published articles in journals such as IEEE TRANSACTIONS ON RELIABILITY; and probabilistic safety assessment.
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS; Relia- Prof. Bérenguer is member of the Editorial Boards of Reliability Engineering
bility Engineering & System Safety; European Journal of Operational Research; & System Safety and Journal of Risk and Reliability.
Journal of Risk and Reliability; and in many international conferences.
Dr. Huynh has received First Place in the 2010 Tom Fagan Reliability and
Maintainability Symposium Student Paper Competition.

Anne Barros is a Professor at the Troyes University of Technology (UTT),


and Head of the System Safety and Optimization Ph.D. program at the Uni-
versity of Technology of Troyes (France). Her research interests focus on the
construction of probabilistic decision indicators used for safety assessment, and
maintenance optimization. These indicators are based on stochastic modeling
of the system state, its environment, and generally speaking of the whole infor-
mation given by the monitoring device. She has authored papers on stochastic
models in reliability or maintenance optimization in IEEE TRANSACTIONS ON
RELIABILITY; Reliability Engineering & System Safety; Journal of Risk and Re-
liability; Journal of Loss Prevention in the Process Industries; International
Journal of Reliability; and Quality and Safety Engineering.

You might also like