ARTICLE IN PRESS

Reliability Engineering and System Safety 95 (2010) 108119

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

Impact of few failure data on the opportunistic replacement policy for

multi-component systems
Radouane Laggoune a,, Alaa Chateauneuf b, Djamil Aissani a
a
b

Laboratory of Modelization and Optimisation of Systems (LAMOS), University of Bejaia, Targua-Ouzemour, Bejaia 06000, Algeria
Laboratory of Mechanics and Engineering (LaMI), PolytechClermont-Ferrand, Campus des Ce zeaux, BP 206, 63174 Aubie re cedex, France

a r t i c l e in f o

a b s t r a c t

Article history:
Received 22 May 2008
Received in revised form
13 July 2009
Accepted 27 August 2009
Available online 31 August 2009

In continuous operating units, the production loss is often very large during the system shut down.
Their economic protability is conditioned by the implementation of suitable maintenance policy that
could increase the availability and reduce the operating costs. In this paper, an opportunistic
replacement policy is proposed for multi-component series system in the context of data uncertainty,
where the expected total cost per unit time is minimized under general lifetime distribution. When the
system is down, either correctively or preventively, the opportunity to replace preventively non-failed
components is considered. To deal with the problem of the small size of failure data samples, the
Bootstrap technique is applied, in order to model the uncertainties in parameter estimates. The Weibull
parameters are considered as random variables rather than just deterministic point estimates. A
solution procedure based on Monte Carlo simulations with informative search method is proposed and
applied to the optimization of preventive maintenance plan for a hydrogen compressor in an oil renery.
& 2009 Elsevier Ltd. All rights reserved.

Keywords:
Opportunistic maintenance
Multi-component system
Data uncertainty
Bootstrap
Monte Carlo simulations

1. Introduction
The preventive maintenance (PM) is often carried out to
prevent or to slow down the deterioration processes. PM is a
scheduled downtime, usually periodical, in which a well-dened
set of tasks (e.g., inspection, replacement, cleaning, lubrication,
adjustment and alignment) are performed. In oil rening facilities,
the problems associated with part replacement are more
concerned than other routine maintenance activities such as
cleaning and lubricating, from the PM scheduling point of view.
This is because the direct costs due to part failure and
replacement are usually very high, and the impact of different
replacement intervals on the overall maintenance cost is often
very sensitive and signicant, in addition to the safety requirements.
In series systems, the one-by-one preventive replacement of
components improves the global system reliability on the account
of its availability, which would be largely penalized, because of
frequent shut downs for component replacements. For multicomponent systems (MCS), an optimal maintenance policy must
take account for interactions between the various components of
the system. The interactions are of three types [1]: economic

 Corresponding author. Tel./fax: +213 34 21 5188.

E-mail address: r_laggoune@yahoo.fr (R. Laggoune).

0951-8320/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2009.08.007

dependence, structural dependence and stochastic dependence,

which are dened as following:

 The economic dependence concerns the inuence of component

operation and maintenance actions on the overall system costs.

In other words, the system cost is not simply the sum of the
individual component costs. In this case, saving in costs or
downtime can be achieved when several components are
jointly maintained. In the present work, only economic
dependence is considered.
The structural dependence concerns components which structurally form a part; therefore maintenance on failed component implies actions on other components. For example,
replacing a part in an engine implies the disassembly and reassembly of other parts.
The stochastic dependence arises when the state of a component
inuences the lifetime distributions of other components or
when components are subjected to common-cause failures.
This is often observed for redundant mechanical systems
where the degradation of a component leads to internal force
redistribution and therefore to overload other components.

The implementation of a PM policy requires a perfect knowledge of the real system reliability and lifetime distribution, which
can be only obtained from a large number of failure data. In other
words, if large set of failure time observations is available, the
component and system lifetimes which are random variables can

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

Nomenclature
C0c

corrective common cost related to the system, to be

paid at each repair upon failure
preventive common cost related to the system, to be
C0p
paid at each time the system undergoes a preventive
maintenance.
specic corrective cost, to be paid at each replacement
Cic
upon failure of component i
specic preventive cost, to be paid at each preventive
Cip
replacement of component i
c
expected total corrective cost of the whole system due
Csyst;j
to failure of component j
p
expected total preventive cost of the whole system
Csys
cost rate for component i (objective function of the
Ci(ti)
mono-component policy)
Cmono(t) cost rate for the mono-group policy (objective function)
C(t,k1,y,kq) cost rate for the opportunistic multi-grouping
policy (objective function)
t
basic
preventive
maintenance
interval
(t mini 1;...;q ti )
integer multiplier of component i, dening the
ki
periodicity of preventive replacements

be accurately described by Weibull distribution, for example.

However, for real-world systems, only few failure data are
available, especially for industries where the safety aspect is
important, such as oil reneries and nuclear power plants. In this
paper, we propose a PM policy for a MCS in the context of rare
data. The economic dependence between components is taken
into account by the introduction of an opportunistic policy
allowing for preventive replacement of non-failed components
during the system shut down (preventively or correctively). The
Bootstrap technique is used to circumvent the lack of data. The
methodology is illustrated by an application to a hydrogen
compressor in an oil renery.
In Section 2, the relevant literature is reviewed, particularly
that dealing with MCS, the problem of uncertainties induced by
the rarity of failure data is also discussed and several ways to
address this problem are given. Section 3 provides the principle of
the bootstrap technique; its usefulness is also discussed. In
Section 4, the costs models are formulated for several maintenance policies. Section 5 gives the solution strategy, in particular
the opportunistic grouping rule and the algorithm allowing for
opportunistic grouping replacement including bootstrap estimates. In Section 6, the proposed approach is illustrated by an
industrial application where the maintenance is rst optimized
without considering uncertainties, and then the bootstrap
technique is integrated to examine the effect of uncertainties in
the Weibull parameters on the optimal strategy obtained above.

2. Maintenance policies for multi-component systems

For MCS, when no strong dependence exists between the
different components, such as a transport eet constituted by
many vehicles; the traditional single-unit model developed by
Barlow and Hunter [2] can be independently applied to each unit,
in order to provide optimal replacement schedule. However, the
general case of MCS implies to take account for the interactions
between various components. The common planning approaches
used for multi-component manufacturing systems include the

109

ti

time interval (age) between preventive replacements

of component i (ti = kit)
q
number of system components
K
the least common multiple of all ki
ti
simulated lifetime of component i
time instant of failure of component j
tj
N
total number of lifetime simulations
total number of parameter simulations
NY
IFk;s;l ; IRk;s;l A 0; 1 binary variables indicating the states of failure
or operation, respectively
MCS
multi-component system (a system constituted by
more than one component), otherwise, it is called a
component or a single-component system
group of components to be preventively replaced at
Gpk
the kth scheduled time instant
group of non-failed components to be opportunistiGhh
cally replaced at the kth interval [(k  1)t,kt]
cumulative distribution failure (CDF) of component i
Fi(.)
cumulative distribution failure of the whole system
Fsys(.)
(due to any component failure)
cumulative distribution failure of the system due to
Fsys,j(.)
component j
b
Weibull shape parameter
Z
Weibull scale parameter

group/block replacement models and the opportunistic maintenance models.

In the block/group maintenance policy, an entire group of
components is replaced at periodic intervals. The interval is
decided based on time, cost or both. The concept of opportunistic
maintenance comes from the fact that the cost of simultaneous
maintenance actions on various components would be less than
the sum of the total cost of individual maintenance actions.
Therefore, providing the opportunity to carry out preventive
maintenance on some components along with the replacement of
failed ones, leads to very small additional cost, compared to
separate replacements. Under these conditions, the maintenance
decisions for one component depend on the states (aging) of the
other components in the system [3,4]. The economic dependence
is common in most continuous operating systems, such as oil
reneries, chemical processing facilities, mass-production manufacturing lines and power generators [3,5,6]. For this type of
systems, the single shut down cost is often much higher than the
cost of the components to be replaced. Therefore, there is a great
potential for cost savings by implementing suitable opportunistic
maintenance policy.
The maintenance and replacement policies of MCS are
extensively discussed in the literature [79]; a number of studies
have reviewed the various policies [1,1012]. These reviews show
that most of the authors use simplied assumptions, or deal with
particular systems (special structure is often assumed), in order to
formulate the decision problem with less mathematical difculty
[4,13,14]. From another point of view, most of the decision models
developed are based on dynamic programming or Markovian
approaches [6], which approximate continuous decision variables
by nite discrete state decision variables. These restrictions in
both maintenance policies and model formulations could affect
the optimality of the solutions because of the reduction of the
solution space. In addition, discrete state decision models are
often difcult to apply to systems with large number of
components and different failure distributions, because of the
astronomic number of combinations as the solution space
growths exponentially with the number of components. It is

ARTICLE IN PRESS
110

R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

Given the standard error of the parameter, the a-level

condence limits may be constructed by assuming that each

percentile is normally distributed and computing y^ 7 za SEy^ ,

worth to note that when the maintenance optimization concerns

a large-scale manufacturing system, the simulation-oriented
approaches can be interesting and perform well [15].
It is to note that most of authors assume the abundance of
data. However, for real-world systems, we have generally few and
partial failure data, leading to biased estimation of the lifetime
parameters and consequently to inappropriate decisions on
maintenance policy. For this reason, the statistical uncertainties
(known also as epistemic uncertainties) should be included in the
evaluation of the expected maintenance and emergency costs. In
order to address this problem and to give a useful assessment of
the uncertainties; several ways can be followed [1619].
The Bayesian approach is useful when the experience feedback is
inexistent or rare and where experts opinions are available [20].
Starting from the subjective probabilities, this approach consists in
combining the experts opinion with statistical observations of the
operation feedback. It is worth to note that if a prior is made up on
the basis of the information coming from two experts, the accuracy of
the resulting failure rate will depend on the degree of independence
between the experts. For a simple case, it is shown that the relative
accuracy increases as the experts become less dependent [21] (i.e. the
reference to a common source decreases). Therefore, the difculty
with Bayesian approach concerns the initial knowledge modeling.
Resampling methods create an ensemble of data sets, where each
set is replicated from the original sample. The Jack-knife algorithm,
introduced for estimating bias and standard errors, generates the new
samples by deleting one (or more) specic data points. In contrast, the
Bootstrap algorithm creates new data sets by sampling with
replacement; one or more data may be repeated more than once in
any resampled data set [22]. It is shown in [23], that the standard
Jack-knife may produce highly inconsistent estimates for the standard
error and/or other measures, in particular those describing percentile
estimate. Therefore, the Jack-knife method tends to be less suited for
the assessment of uncertainties in the practical applications,
especially for small sample size [24]. For this reason, the Bootstrap
technique has been used in the present work.

where za is the standardized normal deviate corresponding to the

condence level a (the Gaussian method). Alternatively, if a
Gaussian approximation is not acceptable and the value of B is
large enough, the appropriate condence interval bounds may be
read from the list of the B estimates of the percentile ranked in
ascending order of magnitude (the percentile method). The Bootstrap percentile method represents a basic form of resampling.
More rened procedures for dening the condence intervals,
such as the Bootstrap-t and the bias-corrected and accelerated (BCa)
methods, are available in [25].
Bootstrap method has become more practical with the general
availability of rapid computing and efcient software. Compared
to standard methods of statistical inference, this method is often
simpler, more accurate, requires fewer assumptions, and has large
applicability [2628]. Resampling provides especially clear advantages when the assumptions of traditional parametric tests are
not met, as with small samples from non-normal distributions.
Additionally, resampling can address questions that cannot be
answered with traditional parametric or non-parametric methods,
such as comparisons of medians or ratios.

4. Maintenance model
Let us consider a system composed of a set of q components
arranged in series, the failure of any component leads to the
failure of the whole system. Let t1,t2,y,tq, be the time intervals
between preventive replacements of components 1,2,y,q, respectively (Fig. 1). It is assumed that each replacement restores the
component to the as good as new condition. During the system
downtime, either for preventive or for corrective maintenance, it
is to decide if we can take the opportunity to replace preventively
some of the non-failed components. This decision should be based
of the reliability decrease and the risk undertaken if these
components fail before reaching the following scheduled
preventive time.

3. Bootstrap technique
Consider the procedure for building a Bootstrap estimate of
standard error of a parameter estimate y^ . Given a data vector
X= [x1,x2,y,xN]T, it can be possible to draw B independent samples,
X1 ; X2 ; . . . ; XB , from the original data with replacement (each Xi
contains N values). For each resample, an estimate of the required

parameter y^ , b=1,2,y,B is computed. The standard error of the

4.1. Cost structure

The replacement costs can be divided into two parts:

 The rst part is related to common system costs, especially the

parameter SEy^ is then estimated by the standard deviation of the

B replications:
v
u
B
B
u 1 X




1X
y^ b  y^ 2 with y^
y^
1
SEy^ t
B  1b1
Bb1 b

0
1

2

3

4

5

production loss due to the system shut down and other xed
costs, such as mobilizing repair crew, disassembling machine,
transportation, tools, etc. The common system cost is noted C0c
for corrective replacement and C0p for preventive replacement.
The second part is related to the specic characteristics of the
component to be replaced, such as spare part costs, specic
tools and repair procedures. For the ith component, the specic

6

7

8

Component 1

2
Component 2

3
Component 3

q
Component q

Fig. 1. Scheduled preventive maintenance plan.

9

10

11

12

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

corrective and preventive costs are noted Cic

respectively.

and Cip ,

When a replacement is carried out, the impact on the total

system cost is given by the sum of common and specic costs. For
scheduled preventive maintenance, the system maintenance cost
is given by
X p
p
C0p
Ci
2
Csys
i A Gp

where Gp is the group of components to be preventively replaced

at the scheduled maintenance time (i.e. all components which
reached the optimal age replacement as it will be detailed below).
For corrective maintenance, when the jth component fails, the
system is shut down and the failed component is replaced; the
opportunity can be taken to replace other critical components.
In this case, assuming that the preventive replacement cost and
the opportunistic one are the same, the corrective system cost is
given by
X p
c
C0c Cjc
Ci
3
Csys;j
i A Gh
iaj

c
is the corrective system cost due to the failure of
where Csys;j
component j, Gh is the group of components to be replaced
preventively during this opportunity (Gh is dened according to a
deterioration-based rule, analyzing the cost/benet balance as it
will be detailed later). According to the renewal theory and
assuming innite horizon, the expected cost per unit time is given
by [29]

CT lim

t-1

ECt
Expected cost during one cycle

Et
Expected length of a cycle

4.2. Models formulation for several maintenance policies

In the following cost models, the discount rate is not
considered as the focus is put on production units where the
maintenance intervals are too short (i.e. few months), leading to
negligible effects on the optimal solution. However, for systems
with large maintenance intervals (i.e. several years), the discounted cost should be considered, which can be easily included
in the herein models. In this case, the opportunistic replacement
may be more protable if it occurs as late as possible, due to the
effect of discount rate. Whatever the cost is discounted or not, the
following models and ideas still remain valid without loss of
generality.
4.2.1. Age-based policy for single component
For a given replacement time t, the expected cost per unit time
is written by the sum of the expected corrective and preventive
costs divided by the expected cycle length for the component i [2]
Ci ti

Cic Fi ti Cip 1  Fi ti
R ti
0 1  Fi t dt

where ti is the time (age) for the preventive replacement of

component i and Fi(.) its cumulative distribution function (CDF).
4.2.2. Equivalent mono-component approach
Here, we assume that all the system components are jointly
replaced, either at system failure or after a certain time t,
whichever occurs rst. Knowing that any component failure leads
to system failure, it is assumed that after each replacement the
system becomes as good as new, the replacement time is
negligible compared to cycle length and the components are
stochastically independent (in fact, while the replacement time is

111

neglected in the cycle length, it should be considered in the

preventive and corrective costs due to extremely high production
losses per unit time). The system renewal cycle is given by
R t Qq
1  Fi t dt and the system total cost per unit time is
0
i1
given by
P
P
C c qi 1 Cic Fsys t C0p qi 1 Cip 1  Fsys t
Rt q
Cmono t 0
6
0 Pi 1 1  Fi t dt
This strategy seems to be suitable only for systems composed
by identical components (with similar lifetime distributions).
However, when the component lifetimes are different, a waste of
money is observed when replacing reliable components under the
conditions enforced by other less reliable ones.
4.2.3. Opportunistic multi-grouping approach
For general engineering systems, the failure rates are very
different from one component to another, and hence, a more
realistic cost model has to be considered. The idea proposed in
this work lies on the optimal denition of the preventive
replacement times (ages) and grouping. In other words, the
replacement time t is selected, such as: (1) at each scheduled
replacement, a decision has to be made for each component to
dene whether it should be preventively replaced or not; and (2)
at each system failure, a decision has to be taken for each
component to see whether it should be opportunistically replaced
or left as it is until the next scheduled replacement. The nal goal
is to plan regular preventive replacements where optimal
component grouping is dened (Fig. 1). In this way, the basic
maintenance timet is dened as the minimum replacement time
(relative to the weakest component): t mini 1;...;q ti . In addition,
the maintenance intervals for various components are dened by:
ti = kit, where ki is an integer multiplier satisfying ki Z1 for
i=1,2,y,q (Fig. 1). In this context, the decision variables are
(t,k1,k2,y,kq), where t is a continuous variable and ki are discrete
variables.
The expression of the total cost per unit time requires the
consideration of the costs involved along a cycle. In this case, the
renewal system cycle is given by the expected time span between
the simultaneous replacement of all the components Kt, where
K= lcm{k1,k2,y,kq} (lcm being the operator of least common
multiple). At the beginning of each new cycle, the system is
totally in the state of as good as new. As illustrated in Fig. 1, the life
cycle of the system involves several replacements of the
components, and therefore, decisions concerning opportunistic
and preventive replacements should be included in the cost
model. At each replacement time tk =kt (with k= 1,2,y,K), the
expected cost is calculated by including corrective, preventive and
opportunistic terms. This formulation allows us to write the total
expected cost per unit time as following:
Ct; k1 ; k2 ; . . . ; kq
PK
Pq
P
P
C c Cjc i A Gh Cip Fsys;j kt C0p i A Gp Cip 1  Fsys kt
k 1
i1 0
k
k

Kt

7
where Fsys,j(.) is the CDF of system failure due to the jth
component. For a series system, the failure of any component
leads to the system failure, then Fsys,j(.)= Fj(.). Ghk is the group of
components to be replaced opportunistically when a failure
occurs in the interval [(k 1)t, kt] and Gpk is the group of
components to be replaced preventively at instant kt: k/ki =Integer
for all i (i=1,y,q; k =1,y,k).
The minimization of the expected cost per unit time aims at
nding the best set of the maintenance times ti:
Find : t; k1 ; k2 ; . . . ; kq

ARTICLE IN PRESS
112

R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

that minimize : Ct; k1 ; k2 ; . . . ; kq

subject to : t Z 0; ki Z1 and ki are integers

When the number of failure data is small, there are large

uncertainties related to the dispersion of the parameter estimates.
In this case, the point estimate cannot be accepted and the
parameters should be modeled by random variables Y, in order to
include their uncertainties. The joint density function f(Y) can be
identied by statistical tools, especially the Bootstrap technique
used in the present work. Therefore, the cost function becomes
conditional, regarding the realization of the uncertain parameters;
Eq. (7) takes the form
Ct; k1 ; . . . ; kq jY
PK
Pq
P
P
C c Cjc i A Gh Cip Fsys;j ktjY C0p i A Gp Cip 1  Fsys ktjY
k 1
i1 0
k
k

Kt

9
where Fsys,j(ktjY) and Fsys(ktjY) are the conditional CDF, which
are dened for a specic realization of the parameter vector Y. To
compute the total expected cost, it is necessary to integrate the
conditional cost function over the probability domain of the
parameters, leading to the following optimization problem:
Find : t; k1 ; k2 ; . . . ; kq
that minimize : Ct; k1 ; k2 ; . . . ; kq

Ct; k1 ; k2 ; :::; kq jYf Y dY

subject to : t Z0; ki Z1 and ki are integers

10

This problem is more general than Eq. (8), as it considers the

parameter distribution, according to the available number of
failure data.

5. Solution strategy
As the cost functions in (8) and (10) contain continuous and
discrete variables, a solution strategy has to be developed for
efcient computation of the optimal replacement plan, especially
for large number of components. Although techniques like
simulated annealing and genetic algorithms can be applied for
general purpose solution procedures ensuring global convergence,
the number of required runs is usually high. Knowing that, in each
run (i.e. each evaluation of the expected cost), we have two nested
Monte Carlo simulation loops (i.e. outer loop for sampling the
Weibull parameters and inner loop for sampling the maintenance
cycle scenario), the reduction of the number of runs
becomes mandatory for efcient solution, even though global
minimum cannot be fully guaranteed. This is the reason why

...

specic procedure has been developed in this work. While the

replacement time can be easily optimized by classical algorithms,
the implication of the discrete variables ki leads to a very large
number of possible combinations. It is thus necessary to reduce
the number of considered combinations for practical systems,
without discarding the potentially optimal combinations.
The proposed solution is a search method based on conditional
information concerning the reliability levels of the components.
As the search algorithm requires a starting point, the initial
solution can be dened by optimal times for the individual
components (minimum of costs in Eq. (5)). This solution gives
reasonable initial values for ki, dened by the ratio between the
component optimal replacement time t0i and the minimum
optimal time t0min , which is written: k0i Integert0i =t0min . Due to
economic dependence, the search range for optimal groups,
dened by ki, can vary from the initial groups, dened by k0i , in
the range 71 dened by: k0i  1 r ki rk0i 1 with ki Z1. This
gives convenient bounds for optimal search and reduces strongly
the number of combinations to only three times the number of
components (instead of the factorial). The large size of random
samplings can be chosen to achieve statistically stable results.

5.1. Opportunistic grouping rule

The deterioration-based decision can be included by analyzing
the cost/benet balance of the component to be preventively
replaced. Let us consider the case where the jth component fails at
the time tj between two scheduled maintenance times kt and
(k+ 1)t, as illustrated in Fig. 2. The opportunity of replacing the
component i leads to an expected cost Cip Ri tj , where Ri(.)=1  Fi(.)
is the component reliability. If the ith component is left without
replacement, two cases are possible: either it remains operating
until the following scheduled replacement, which will cost
Cip Ri k 1tjtj , or it fails before, leading to system breakdown,
the corresponding cost is C0c Cic Fi t with a maximum value at
(k+ 1)t. The decision-making criterion for opportunistic
replacement can thus be dened by comparing the two costs. If
the opportunistic replacement cost is less than the corrective one,
it is better to change the component; otherwise it can be left till
the following planned replacement. This rule is written
If : Cip Ri tj  Ri k 1tjtj
o C0c Cic Fi k 1tjtj
Otherwise

) Then, make opportunistic

replacement of component i
at the time tj.
) Leave it as it is till the next
scheduled replacement

It is to be noted that the probability and reliability functions, in

the above expression, are computed for a given set of the Weibull

Failure and corrective

replacement of Ci

(k+1)

Component i

Last
replacement of Cj

Component j
Opportunistic
replacement of Cj
Cost = CjpRj (ti )

Failure of Cj
before (k+1)
Cost = (C0c+ Cjc)Fj (ti )

Fig. 2. Decision cost basis for opportunistic replacement.

Scheduled
replacement of Cj
Cost = (C0p +Cjp)Rj ((k+1) )

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

113

Perform Bootstrap analysis

on available failure data

Define the replacement plan

Sample the Weibull parameters

Modify the maintenance
plan toward the optimimum

Is failure observed
before the next maintenance
time: tik < tm ?

No

Yes
Regenerate
new lifetimes
for replaced
components

Stop production: C 0c
Replace the failed component: Crc

Check the opportunity

of preventive replacements : SCip

No

Conditional preventive
replacement of critical
components: SCi p

Repeat until the end of parameter simulations

Generate component
lifetimes t ik

Is the end
of the life cycle
reached ?

Yes
Compute the expected total cost
of the life cycle and search for
the minimum cost
Fig. 3. Flowchart of the solution algorithm.

parameters. During Monte Carlo simulations, the decision

changes, depending on the parameter realizations.
5.2. Solution procedure
The owchart of the developed algorithm is depicted in Fig. 3.
After introducing the component failure data (probability
distributions and parameters), the cost parameters C0c ; C0p ; Cic ; Cic
and the initial grouping conguration k0i ; the solution procedure
is given as following:
1. Perform a Bootstrap analysis on the failure data, to determine
the joint distribution of Weibull parameters fi(bi,Zi) for the ith
component. The joint parameter distribution can be built from
the marginal probability distributions of shape and scale
parameters, and the statistical correlation between them. It is
to note that any probability model tting the Bootstrap output
can be used for the marginal distributions of the parameters.
2. Generate a random sample of the parameters bi and Zi,
according to the above joint density functions.
3. Generate random samples of component lifetimes ti, according to the failure probability distributions. The system failure
time is dened by: tsys mini 1;q ti and the corresponding
failed component producing the system shut down is
identied. The replacements are scheduled at the times kt,
where k is an integer varying from 1 to K.

4. At the kth replacement, the simulated system failure time tsys

is compared to the scheduled time for preventive maintenance kt. Two possibilities exist:
(a) If no failure is observed before kt, the preventive
maintenance can be carried out at kt, according to
the current grouping rule, as dened in the updated
plan, and a move to the next scheduled time (k+ 1)t is
done.
(b) If failure is observed, the system is down and the failed
component is correctively replaced. On the basis of the
conditional strategy described above, the opportunity of
replacing other components is considered and the related
preventive costs are computed.
5. For the replaced components in step 4, new lifetimes are
generated (as new components are installed); a move to the
next replacement time is performed (k+ 1)t, and so on, until
the end of the system cycle (until the replacement of all
components simultaneously).
6. For the simulated scenario, the life cycle length and the
corresponding total cost are computed.
7. Repeat steps 26, to generate new scenarios by random
sampling, until the prescribed number of simulations is
reached.
8. The expected total cost per unit time is estimated in terms
of the mathematical average of the computed costs and
the cycle span of the simulated scenarios. For all the
sampled scenarios, the total cost in terms of the statistical

ARTICLE IN PRESS
114

R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

2t

3t

Component 1
Component 2
Component 3

Component 4
Component 5
Component 1
Component 2
Component 3

Component 4
Component 5
Component 1
Component 2
Component 3

Component 4
Component 5

Failure
Corrective replacement

Opportunistic replacement

Scheduled
preventive replacement

Fig. 4. Example of corrective/opportunistic/preventive maintenance simulation.

expectations is given by
0
00
1
NY
N
K
X
X p
1 X
1X
c
c
@
@
@
C0 Ci
Ci AIFk;s;l
ECt
NY l 1 N s 1 k 1
i A Gh
0
1
1
X p
p
C AIR A
@C
0

i A Gp

k;s;l

11

where NY is the total number of parameter simulations, N is

the total number of lifetime simulations, IFk;s;l and IRk;s;l are
binary indicators for the states of failure and operation,
respectively. They depend on the replacement interval k, the
lifetime sample s and the parameter sample l. For the kth
replacement interval, these indicators are dened by
(
IFk;s;l 1; IRk;s;l 0 if failure
IFk;s;l 0; IRk;s;l 1

if operation

9. The procedure is repeated for different replacement intervals

t and grouping congurations ki. The search for the
optimum solution allows us to update the scheduled
maintenance plan, by changing t and ki, in order to dene a
better combination of component grouping. The iterative
scheme is stopped when the optimum solution cannot be
improved.

The large number of random simulations ensures the stability of

the cost estimate and guarantees the solution convergence to the
optimal plan. Several numerical tests and applications have
shown that the expected cost estimate becomes stable for more
than 500 Monte Carlo samples. High precision can be reached for
cost estimate with 10,000 samples, where the coefcient of
variation is always less than 0.1%; this number of Monte Carlo

simulations is considered in the presented numerical applications.

As all signicant combinations are considered, the optimal
solution cannot be missed in the proposed procedure.
Fig. 4 gives an illustrative example for a sample of ve
components, where the scheduled preventive replacement has a
span equal to t. The above described algorithm is applied to
generate random scenarios. The component lifetimes are
rstly generated, as shown in Fig. 4a; it is observed that
component 2 fails before the scheduled replacement and should
be correctively replaced. Given the deterioration of component 1,
the opportunity of replacing it preventively is carried out, as
illustrated in Fig. 4b. New lifetimes are then generated for the new
components (i.e. components 1 and 2). The scheduled
replacement at t is now examined; in this example, components
13 are preventively replaced (Fig. 4b). In the next replacement
interval (t ot r2t), components 1 and 3 appear to have high
deterioration rate and are then replaced at 2t (Fig. 4c). Then,
component 5 shows a failure before reaching the scheduled
replacement at 3t, and so on. This illustration shows how a
component can be dynamically considered in function of the
possible opportunities.

6. Industrial application
The proposed methodology is applied to a centrifugal compressor, located at Skikda renery, which is the most important oil
renery in Algeria and among the most important in Africa. The
multiple staging compressor is driven by a steam turbine; it is
essentially constituted by the stator (diaphragms, landings,
tightness subsystem) and the rotor (shaft, wheels, equilibrium
piston, etc.). The compressor aims to recycle the necessary
hydrogen for the different catalytic reforming reactions, it also
participates to the catalyst regeneration. In addition it is necessary
for the unit pre-heating during operation starting after shut down.

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

115

6.1. Failure and cost data

6.2. Maintenance policies

For the ve components to be considered, Table 1 indicates the

sample sizes and the observed number of failures, ranging from 8
to 15 records. Having the failure and cost data collected in this
study for the oil renery, the two-parameter Weibull model is
tted by the maximum likelihood method applied on the
observed failure times; the shape and scale parameters are
depicted in Table 1.
The failure distributions of these components are depicted in
Fig. 5. It can be observed that the components C286 and C285
have comparable distributions, which is also the case for the
components C230 and C260.
Table 2 gives the corrective and preventive costs for different
components, as well as the production losses due to system
downtime; the large ratio of corrective to preventive costs can be
easily observed in these data. As the production losses are related
to the time necessary to carry out the replacement, it is assumed
to be practically independent of the component itself. The
maintenance optimization is carried out according to different
assumptions in order to compare the optimal solutions and to
show the benets of the proposed maintenance plan.

First, we analyze various replacement policies on the basis of

point estimation of Weibull parameters (i.e. no uncertainty is
considered for the two parameters), in order to select the
appropriate one.

Table 1
Failure data and Weibull parameters of the system components.
Component

Sample
size

Observed
failures

Shape
parameter b

Scale
parameter Z

C286
C285
C275
C230
C260

Sheathing
Sheathing
Tightness
Stub bearing
Tightness
ring

24
23
21
21
34

14
15
15
8
14

1.73
1.88
2.43
2.53
2.14

486
507
286
898
905

Probability density

Code

6.2.2. Equivalent mono-component replacement policy

Instead of replacing separately the different components, one
may suggest to make simultaneous replacement of the ve
components, in order to reduce the down time of the system. In
this case, the associated total cost is given by Eq. (6). The expected
cost, depicted in Fig. 6, shows a minimum at 23 days, with an
expected optimal cost of 438.16 h for the whole system. Naturally,
this solution is not optimal as it does not take into account the
specic costs related to the different components.
6.2.3. Opportunistic multi-grouping optimization
The proposed solution is based on optimal grouping of
maintenance operations (Eq. (7)). Table 4 gives the optimal
solutions for different grouping policies. It is shown that the
minimum cost is achieved when two groups are considered, for
which k1 = k2 = k3 =1 and k4 = k5 =4. A group of high failure rate
components (C286, C285 and C275) with periodic replacement
every 27 days and a group of low failure rate components (C230
and C260) with periodic replacement every 108 days. The
expected cost is only 123.75 h/day, which represents 72% of
reduction with respect to the case of equivalent mono-component
policy. This reduction shows clearly the importance of the choice
of the strategy to be applied and the interest of the proposed
model.

Failure distribution

0.004
0.0035
0.003
0.0025
0.002
0.0015
0.001
0.0005
0

6.2.1. Single-component policy

As a rst step, the policy based on separate components is
considered for comparison purpose. The expected cost is computed for each component, independently, and a one-by-one
optimization is applied (Eq. (5)).
Table 3 gives the optimal solutions for the system components.
While component C286 has the lowest optimal replacement time,
it is shown that two other groups can be possible: C285/C275 and
C230/C260, where the replacement times are close. As discussed
above, the ratios of individual optimal times can be used as a
starting point for group maintenance plan. From Table 4, these
ratios are given by: ki = ti/t1, leading to: k1 =l, k2 = l.4, k3 =l.3, k4 = 3.7
and k5 = 3.5 (latter, the nearest integer will be used for ki).

C286
C285
C275
C230
C260

6.3. Effect of data uncertainties on the maintenance policies

0

200

400
600
time (days)

800

1000

Fig. 5. Probability density functions of the system components.

In this subsection, we consider the uncertainties related to

Weibull parameters, due to the small size of the failure records.
These uncertainties are then included in the replacement policy in
order to underline the effect of data size and the scatter on the
optimality of the maintenance policy. It is to note that for the MCS,
we consider the opportunistic policy for the analysis, as it is the

Table 2
Production loss and maintenance costs.
Component

Production losses
Sheathing
Sheathing
Tightness
Stub bearing
Tightness ring

Code

C286
C285
C275
C230
C260

Corrective
cost (h)

Preventive
cost (h)

Cost ratio
Corr./Prev.

35,000.00
11,281.84
30,390.16
33,244.00
43,542.64
51,856.00

400.00
263.89
143.83
339.95
427.71
955.35

87.5
42.8
211.3
97.8
101.8
54.3

Table 3
Optimal solutions for individual components without uncertainties.
i

Component

1
2
3
4
5

Sheathing
Sheathing
Tightness
stub bearing
Tightness ring

MTBF (days) Optimal time t0i (days) Cost (h/day)

C286
C285
C275
C230
C260

483
475
240
787
844

29.8
42.7
38.5
126.1
122.9

52.97
27.25
32.66
10.86
20.75

ARTICLE IN PRESS
116

R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

Table 4
Optimal solutions for different replacement groups without uncertainties.
Replacement groups k1  k2  k3  k4  k5

Expected system life cycle (days)

Optimal replacement times t1/t2/t3/t4/t5 (days)

Minimum expected cost (h/day)

Equivalent mono-component
1-1-1-3-3
1-1-1-4-4
1-2-1-4-4
1-2-2-4-4

23
73.8
97.5
85.8
71.1

23/23/23/23/23
28/28/28/84/84
27/27/27/108/108
24/48/24/96/96
20/40/40/80/80

438.16
128.09
123.75
136.54
155.28

rather negative and tends to be stronger, this is particularly the

case of component C285.

Equivalent macro-component

Expected cost ()

2000

1500

1000

500

0
0

50

100
Time (days)

150

200

Fig. 6. Equivalent mono-component replacement cost function.

most suitable strategy for our case (i.e. continuous production

systems).
6.3.1. Uncertainties in Weibull parameters
Having small sample sizes, it is more reasonable to use
resampling techniques for better uncertainty assessment. The
Bootstrap technique is applied to provide the statistical distributions of the two Weibull parameters: b and Z, and consequently to
estimate the means and the standard deviations of these
parameters. The Bootstrap is applied by repeating 5000 times
the resampling from the original data, each sample has the same
size as the original data set. For each generated sample, the two
parameters are computed by maximum likelihood estimates.
Then for the 5000 samples, statistical analysis is carried out for
the Weibull parameters in order to determine the probability
distributions, the means, the standard deviations and the
correlation coefcients.
As an example, Fig. 7 gives the Bootstrap distributions of the
Weibull parameters for the component C286. It can be observed
that while the scale parameter seems to be normally distributed,
the shape parameter is strongly skewed and can be represented by
lognormal distribution. The scatter of the parameter estimate is
clearly observed in this gure, which points out the meaningless
of the point estimate on the basis of few failure data. The same
observations are veried for the other components of the system.
The results of the Bootstrap estimates for all system components are depicted in Table 5. We can see that the standard
deviation increases largely when the sample size decreases, which
is particularly the case of component C230 with only 8 failure
records. Depending on the collected failure data, the statistical
correlation between the two Weibull parameters is globally small,
except for component C285. This is due to the large scatter of
failure times, where no clear trend can be observed between the
two parameters (especially with low number of data). When the
scatter of the failure times decreases, the correlation becomes

6.3.2. Impact of uncertainties on single-component policy

In order to investigate the effect of parameter uncertainties on
the optimal times, we have performed random sampling of the
Weibull parameters, where each realization leads to a specic
replacement time and cost. The statistical analysis of the 10,000
random samples allows us to plot the histograms for the
replacement time and cost. Fig. 8 shows the distributions of
optimal replacement times of individual components C286, C285
and C275. The optimal time distribution is less skewed for the
component C275 than for other components; this is conrmed by
its small standard deviation. In Fig. 8, it can be seen the highly
extended distribution tails of the optimal replacement times,
which highlight the fact that point estimation is not appropriate
for maintenance optimization, unless for large size of failure data.
On the basis of probabilistic Weibull parameters, the mean and
standard deviation of the optimal replacement times and
corresponding costs for individual components are given in
Table 6.
We observe that, except for component C230, the means of
optimal times are larger than the deterministic times; in addition,
the standard deviations of these times strongly depends on the
scatter of failure data (it is to note that the coefcients of variation
ranges from 21% to 61%). The impact of this scatter on the expected
cost is much more signicant; as an example, the standard
deviation of component C230 (corresponding to the smallest
sample size) is 103 h/day while its mean is only 77 h/day! This
table shows clearly that the decision making on the basis of point
estimation (independently of the data size) leads to high risk of
production losses, as the optimal cost cannot be guaranteed. In
other words, due to parameter uncertainties, the real maintenance
cost may be doubled with respect to the expected one. This high
sensitivity underlines the necessity to use appropriate description
of various levels of uncertainties (i.e. aleatory and epistemic
uncertainties).
6.3.3. Impact of uncertainties on multi-group replacement policies
The parameter data in Table 5 are now considered for
opportunistic multi-grouping maintenance, where the optimization procedure proposed in Section 5 is applied. As an example,
the distribution of the optimal replacement time for the grouping
strategy 1-1-1-4-4 is depicted in Fig. 9. It shows the large scatter
of the optimal replacement time. This distribution provides the
design-maker with information allowing for best selecting of the
replacement age, according to operational conditions.
Table 7 gives the means and standard deviations for the
optimal time and cost, under various opportunistic groupings.
While the parameter uncertainties lead to slight increase of the
mean values for the replacement time, it shows a standard
deviation varying from 4.8 to 6.7 days, which is very signicant,
compared to the deterministic times. The means of the minimum

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

0.45

0.16
0.12

Probability distribution

0.14
Component C286

0.1
0.08
0.06
0.04
0.02

0.4
0.35
Component C286

0.3
0.25
0.2
0.15
0.1
0.05

1.
25
1.
5
1.
75

60
0

55
0

50
0

45
0

40
0

35
0

30
0

25
0

3
3.
25
3.
5
3.
75

2
2.
25
2.
5
2.
75

Frquence

117

Shape parameter

Scale parameter

Fig. 7. Bootstrap Weibull parameter distributions for the component C286.

Table 5
Bootstrap Weibull parameters distributions of the system components.
Code

C286
C285
C275
C230
C260

14
15
15
8
14

Shape parameter b

Scale parameter Z

Mean

Standard deviation

Mean

Standard deviation

1.73
1.88
2.43
2.53
2.14

0.329
0.667
0.454
2.66
1.01

486
507
286
898
905

59
184
37
271
176

0.12

0.02
 0.66
0.17
 0.12
 0.34

0.08

16

64
68

56
60

Optimal replacement interval (days)

Optimal replacement interval (days)

0.1
0.08
Distribution

Component C275

0.06
0.04
0.02

64
68

48
52

40
44

32
36

24
28

0
16
20

16
20

48
52

0
40
44

0
32
36

0.02

24
28

0.02

10
0
11
2
12
4
13
6

0.04

76

0.04

0.06

64

0.06

Component C285

52

Distribution

0.08

88

Component C286

40

0.1
Distribution

Correlation b vs. Z

0.1

28

Sheathing
Sheathing
Tightness
Stub bearing
Tightness ring

observed failures

56
60

Component

Optimal replacement interval (days)

Fig. 8. Distributions of optimal replacement intervals for individual components.

ARTICLE IN PRESS
118

R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

Table 6
Optimal times for individual components with uncertainties.
Component

Code

Deterministic time (days)

Bootstrap optimums
Optimal time (days)

Sheathing
Sheathing
Tightness
Stub bearing
Tightness ring

C286
C285
C275
C230
C260

29.8
42.7
38.5
126.1
122.9

Opportunistic optimal age

0.25

Distribution

0.2
0.15
0.1
0.05
0
20

22

24

26

28

30

32

34

36

38

Optimal cost (h/day)

Mean

Std. deviation

Mean

Std. deviation

32.7
55.4
38.8
123.6
125.3

12.7
34.1
10.7
35.7
26.8

62.6
50.7
37.0
77.7
39.3

30.9
49.5
16.0
103.1
42.2

more robust than 1-1-1-4-4, and can better t the target cost. This
strategy offers a better decision-making tool, regarding data
uncertainties. This comparison allows us to conrm that deterministic approach does not really provide the optimum result, as
the risk of missing the target can be high.
These results show the importance of considering the inuence
of small size of data on the Weibull parameter scatter and
consequently on the dispersion of the optimal cost. As shown in
this example, the optimal strategy considering uncertainties may
be different from the policy based on deterministic parameters.
One important conclusion is that the optimal strategy cannot be
independent of the size of the failure data. It may change with the
amount of input data and only probabilistic parameters can lead
to robust maintenance policy.

Optimal replacement age (days)

Fig. 9. Distribution of replacement interval (strategy 1-1-1-4-4).

Table 7
Opportunistic policy under deterministic and probabilistic parameters.
Strategy

Optimums

Deterministic parameters

Probabilistic
parameters
Mean

Std. deviation

1-1-1-3-3 Optimal time 28

Cost (h/day) 128.09

29
152.05

4.8
82.2

1-1-1-4-4 Optimal time 27

Cost (h/day) 123.75

28.5
135.5

4.9
72.4

1-2-1-4-4 Optimal time 24

Cost (h/day) 136.54

26.1
136.1

5.8
64.3

1-2-2-4-4 Optimal time 20

Cost (h/day) 155.28

20.6
148.6

6.7
76.85

costs are higher for the rst two groups. For all cases, the standard
deviation of the minimum cost is nearly 50% of the mean value.
While for deterministic parameters, the strategy 1-1-1-4-4
leads to lower cost than for strategy 1-2-1-4-4, the consideration
of parameter uncertainties leads to almost the same mean costs
for both strategies, 135.5 and 136.1 h/day, respectively, which is
still better than for the other groupings. Although the grouping
1-1-1-4-4 is more economical in deterministic consideration, the
strategy 1-2-1-4-4 shows a lower standard deviation for the
minimum cost: 64.3 h/day instead of 72.4 h/day for 1-1-1-4-4. In
other words, the policy 1-1-1-4-4 cannot guarantee the desired
minimum cost, given the data scatter. Therefore, the strategy
1-2-1-4-4, although with very slightly higher mean cost, is much

7. Conclusion
The proposed maintenance plan is based on multi-grouping
optimization for multi-component systems. As in many systems,
the production losses are very high and the maintenance policy
intends to increase the system availability through cost reduction.
The optimal multi-grouping is based on the analysis of isolated
component replacement times. The times are rearranged to allow
group replacements, and thus reducing the whole system downtimes as well as the maintenance costs. The proposed optimization algorithm shows effective cost reduction by selecting the
optimal grouping and time interval for preventive replacement.
Regarding the small sample size, the solution in the deterministic approach may miss the optimum. As a matter of fact, taking
into account the parameter uncertainties, by the introduction of
the Bootstrap technique, showed that the optima obtained by the
proposed approach are different from those obtained by deterministic approach. The consideration of parameter uncertainties
allows us to improve the solution optimality; moreover it gives,
additional information about the solution deviation from the
deterministic one, leading to very useful indicators for better
decision making. One important outcome is to underline that
optimal strategy cannot be independent of the size of the failure
data and probabilistic parameters should be considered for robust
maintenance optimization.
References
[1] Thomas LC. A survey of maintenance and replacement models of multi-item
systems. Reliab Eng 1986;16:297309.
[2] Barlow RE, Hunter LC. Optimum preventive maintenance policies. Oper Res
1960;8:90100.
[3] Rao AN, Bhadhury B. Opportunistic maintenance of multi-equipment
systems: a case study. Qual Reliab Eng Int 2000;16(6):487500.
[4] Van der Duyn Schouten FA, Vanneste SG. Analysis and computation of (n, N)strategies for maintenance of a two-component system. Eur J Oper Res
1990;48:26074.

ARTICLE IN PRESS
R. Laggoune et al. / Reliability Engineering and System Safety 95 (2010) 108119

[5] Volkanovskia A, Mavkoa B, Bosevski T, Causevski A, Cepina M. Genetic

algorithm optimisation of the maintenance scheduling of generating units in
a power system. Reliab Eng Syst Saf 2008;93:75767.
[6] Pila CK, Rausandb M, Vatn J. Reliability assessment of reliquefaction systems
on LNG carriers. Reliab Eng Syst Saf 2008;93:134553.
[7] Huang J, Okogbaa OG. A heuristic replacement scheduling approach for
multi-unit systems with economic dependency. Int J Reliab Qual Saf Eng
1996;3:10.
[8] Tsai YT, Wang KS, Tsai LC. A study of availability-centred preventive
maintenance for multi-component systems. Reliab Eng Syst Saf 2004;84:
261270.
[9] Garbatov Y, Guedes Soares C. Cost and reliability based strategies for fatigue
maintenance planning of oating structures. Reliab Eng Syst Saf 2001;73:
293301.
[10] Dekker R, Wildman RE, Van Der Duyn Schouten FA. A review of multicomponent maintenance models with economic dependence. Math Methods
Oper Res 1997;45(3):41135.
[11] Wang H. A survey of maintenance policies of deteriorating systems. Eur J Oper
Res 2002;139:46989.
[12] Dekker R. Applications of maintenance optimization models: a review and
analysis. Reliab Eng Syst Saf 1996;51(3):22940.
[13] Dekker R, Roelvink IFK. Marginal cost criteria for preventive replacement of a
group of components. Eur J Oper Res 1995;84:46780.
[14] Goyal SK, Gunasekaran A. Determining economic maintenance frequency for
a family of machines. Int J Syst Sci 1992;4:6559.
[15] Crockera J, Kumarb UD. Age-related maintenance versus reliability centred
maintenance: a case study on aero-engines. Reliab Eng Syst Saf 2000;67:
11318.
[16] Borgonovo E, Marseguerra M, Zio EA. Monte Carlo methodological approach
to plant availability modeling with maintenance, aging and obsolescence.
Reliab Eng Syst Saf 2000;67:6173.

119

[17] Rocco CM, Miller AJ, Moreno JA, Carrasquero N, Medina M. Sensitivity
and uncertainty analysis in optimization programs using an evolutionary
approach: a maintenance application. Reliab Eng Syst Saf 2000;67(3):
249256.
[18] Eisinger S, Rakowsky UK. Modeling of uncertainties in reliability centered
maintenance-a probabilistic approach. Reliab Eng Syst Saf 2001;71:15964.
[19] Martorell S, Sanchez A, Carlos S. A tolerance interval based approach to
address uncertainty for RAMS+ C optimization. Reliab Eng Syst Saf
2007;92(4):40822.
[20] Apeland S, Aven T. Risk based maintenance optimization: foundational issues.
Reliab Eng Syst Saf 2000;67:28592.
[21] Williams MMR, Thome MC. The estimation of failure rates for low probability
events. Prog Nucl Energy 1997;31(4):373476.
[22] Efron B, Tibshirani R. In: An introduction to the bootstrap. Chapman and Hall;
1993.
[23] Shao J, Tu D. In: The Jack-knife and bootstrap. New York: Springer; 1995.
[24] Hall MJ, van den Boogaard HFP, Fernando RC, Mynett AE. The construction of
condence intervals for frequency analysis using resampling techniques.
Hydrol Earth Syst Sci 2004;8(2):23546.
[25] Davidson AC, Hinkley DV. In: Bootstrap methods and their application.
Cambridge: Cambridge University Press; 1997.
[26] Rocco SCM, Zio E. Bootstrap-based techniques for computing condence
intervals in Monte Carlo system reliability evaluation. In: Proceedings of the
reliability and maintainability symposium, January 2427, 2005. p. 3037.
[27] Chao-Yu Chou, Yu-Chang Lin, Chun-Lang Chang, Chung-Ho Chen. On the
bootstrap condence intervals of the process incapability index Cpp. Reliab
Eng Syst Saf 2006;91(4):4529.
[28] Dargahi-Noubary GR, Razzaghi M. Bootstrap construction of the upper
condence limit for unreliability. Reliab Eng Syst Saf 1992;37(1):16.
[29] Barlow RE, Proschan F. In: Mathematical theory of reliability. New York: John
Wiley & Sons; 1965.