Advanced Maintenance Modelling

E-M algorithm to identify missing causes of failures

for repairable systems with a burn-in period

Isha Dewan
Indian Statistical Institute
7, S. J. S. Sansanwal Marg, New Delhi, India

Yann Dijoux
Université de Technologie de Troyes
10010, Troyes, France

Abstract

In this paper, we consider complex repairable systems presenting a burn-in period.

Different actions of maintenances are considered : corrective maintenances (CM) are
performed after a failure which can be due to the degradation of the system or due to
manufacturing defects and we also consider preventive maintenances (PM). The mainte-
nance process is described using the competing risks framework. We assume that imperfect
maintenances are carried out. The most usual imperfect maintenance models are the vir-
tual age models. More specifically, we consider asymmetric virtual age models for each
kind of maintenance. Moreover, we assume that some of the causes of failure may not be
necessarily recorded.
The model that we present assumes that the degradation process is independent of
the burn-in process. Preventive maintenance are considered deterministic. Maintenance
efficiencies are considered using a classical virtual age model for preventive and corrective
maintenance due to the ageing, with different improvement factors. Maintenance after a
corrective maintenance due to burn-in defects are assumed minimal. To identify missing
causes of failure, we propose an adapted version of the E-M algorithm. Properties of the
model are presented and numerical estimations are carried out.

Keywords: Imperfect maintenance, burn-in period, E-M algorithm, asymmetric virtual

ages, competing risks.

1 The maintenance process

1.1 Introduction
All along their life, complex industrial systems are subject to failures. After a failure, a
corrective maintenance(CM), also called repair, is carried out in order to put the system
functional again for the longest time possible. Preventive maintenance (PM) can also be
carried out when the system is operating and intends to slow down the wear process and
reduce the frequency of occurrence of system failures. There are two well-spread assump-
tions on maintenance efficiency defined as minimal repair or As Bad As Old (ABAO)
and perfect repair or As Good As New (AGAN). Obviously, reality is between these two
extreme cases: standard maintenance reduces failure intensity but does not leave the

1
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

system as good as new. Then, imperfect maintenance models have been proposed. The
most popular of them are the virtual age models, defined by Brown-Proschan [1] and
Kijima et al [7]. These models are adapted to wearing-out systems, which are described
by reliability models with increasing initial failure intensity. However, if we consider sys-
tems presenting a burn-in period, the risk of observing a failure soon after the startup of
system is important. It’s not reasonable to model these systems simply by a continuous
degradation.
In this paper, we propose to dissociate or separately identify both causes of failure.
To that purpose, we use the competing risks framework to describe a risk due to the
degradation, an other due to burn-in defects. Preventive maintenances are assumed to
be deterministic, then the PM times correspond to censoring times of the CM process.
It seems realistic that the actions performed and their efficiency differ depending on the
kind of maintenance. Then we consider different virtual age assumptions for each risk,
which corresponds to use asymmetric virtual ages as defined in [6]. It seems reasonable
to assume a minimal repair for the risk of burn-in failure and a classical virtual age
model for the risk related to the degradation and preventive maintenances. If we consider
that the kind of maintenance is recorded, we can easily derive the classical results of the
competing risks models associated to imperfect maintenance. Unfortunately, the kind
of maintenance performed is not always observed or recorded and in practice, very few
dataset provide this information. Then we present an adaptation of the E-M algorithm
in order to estimate the parameters of the model. Finally, simulations and numerical
estimations are presented.

1.2 Notations
The PM-CM process is the sequence of PM times and CM times. Maintenance durations
are assumed to be negligible or not taken into account. Then, we introduce the following
notations.

• {Ck }k≥1 the maintenance times (CM and PM), with C0 = 0.

• {Wk }k≥1 the times between maintenances, Wk = Ck − Ck−1 .

• K = {Kt }t≥0 the counting maintenance (CM and PM) process.

• The indicators of kind of maintenance Uk :




 1 if the k th maintenance is preventive (PM)
Uk =  0 if the k th maintenance is corrective, due to the degradation (1)
 −1 if the k th maintenance is corrective, due to a burn-in defect


The maintenance process is completely given by three maintenance intensities defined

as:
1
∀t ≥ 0, λut i = limP (Kt+∆t − Kt− = 1, Ui = ui |Ht− ) ui ∈ {−1, 0, 1} (2)
∆t→0 ∆t

where Ht− is the past of the maintenance process at time t, i.e. the set of all events that
have occurred before t. Without external variable, it usually corresponds to the times

2
 Advanced Maintenance Modelling

and kind of maintenances occurred before t: (Kt , C1 , U1 , . . . , CKt− , UKt− ). Moreover, if

we consider deterministic PM, it is not necessary to define the intensity λ1t , as we can see
the PM process as a censoring of the corrective maintenances process.

1.3 The competing risks framework

A simple way of modelling multiple potential causes of an event is the competing risks
framework, first developed in the context of maintenance by Cooke and Paulsen [2]. When
the system restarts after the (i−1)th maintenance, it is not known if the next maintenance
could be a PM, CM due to the degradation or a CM due to design defects. Let Yi be
the potential time to next PM, Zi be the potential time to next degradation failure and
Ri the potential time to next burn-in failure. Then, the real time to next maintenance
is Wi = min(Yi , Zi , Ri ) and we also observe the maintenance indicator Ui . Ui equals to 1
if we carry out a PM, 0 if the maintenance is corrective and due to the degradation and
equals to -1 if it is a CM due to manufacturing defects.

In the classical competing risks framework, it is assumed that the variables Yi (respec-
tively Zi and Ri ) are independent and identically distributed. It means that the effect
of each maintenance is to restore the system as good as new (AGAN). In the perfect
maintenance case, it is only necessary to define the joint distribution for a new system.
For example, we can express the three-dimensional survival function:

S1 (y, z, r) = P (Y1 > y, Z1 > z, R1 > r)

In many situations, the initial risks are assumed to be independent, which implies
that the survival function is product of three marginal survival functions SY , SZ , SR .
These functions are the survival functions of Y, Z and R, respectively. But if we consider
condition-based preventive maintenance, the PM and CM processes can not be assumed
independent. This lead to several competing risks models such as the Random sign models,
the Alert-Delay model.

Another very limiting assumption is the iid hypothesis. Doyen and Gaudoin [6]
proposed a general framework call the Generalized Competing Risks where the couples
{(Yk , Zk , Rk )}k≥1 are not assumed to be iid. The couples {(Wk , Uk )}k≥1 are therefore also
not iid. Thus, the effect of every PM and CM can be imperfect. The usual compet-
ing risks objects are naturally generalized by introducing a conditioning on the past of
the PM-CM process. For example, we can introduce the CM-PM conditional generalized
survival function as:

Sk+1 (y, z, r; W1 , U1 , . . . , Wk , Uk ) = P (Yk+1 > y, Zk+1 > z, Rk+1 > r | W1 , U1 , . . . , Wk , Uk )

(3)
It has been shown in [6] that the intensities and likelihood function can be expressed
in terms of the PM-CM survival functions. Among the first GCR models, we can name
the Langseth and Lindqvist model [8] and the CIGCR models [4]. In the following, as the
PM are assumed to be deterministic, we will mainly work on a GCR based on the couples
{(Zk , Rk )}k≥1 and consider that the {Yk }k≥1 are censoring times.

3
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

1.4 Models of imperfect maintenance

1.4.1 The basic models
Let’s first consider one kind of maintenance, e.g. only CM. We present the basic models of
maintenance efficiency. The maintenance process is defined by a unique failure intensity
λt . We assume that before the first failure, the failure intensity is a deterministic and
continuous function of time λ(t), called the initial intensity. λ is in fact the failure rate
of the first failure time C1 = W1 . The most usual initial intensity has the following
expression: λ(t) = αβtβ−1 , α > 0, β > 0.
The basic assumptions on CM efficiency are known as minimal repair or As Bad As
Old (ABAO) and perfect repair or As Good As New (AGAN). In the ABAO case, each
maintenance leaves the system in the same state as it was before the failure. The failure
process is then a Non Homogeneous Poisson Processes (NHPP) and its failure intensity is
a continuous function of time: λt = λ(t). In the AGAN case, each maintenance perfectly
repairs the system and leaves it as if it were new. The failure process is a Renewal
Process (RP) and its failure intensity is of the form: λt = λ(t − CKt ). Reality is naturally
between these two extreme cases: standard maintenance reduces failure intensity but does
not leave the system as good as new. Then, imperfect maintenance models have been
proposed. We present the most popular of them called the virtual age models.

1.4.2 Virtual age models

A virtual age [7] model is characterized by a sequence {Ai }i≥1 of positive random variables,
called effective ages, such that A0 = 0 and the conditional distributions of interfailure
times are given by:

∀i ≥ 1, ∀x ≥ 0, P (Wi+1 > x|Ai , W1 , . . . , Wi ) = P (Y > Ai + x|Y > Ai , Ai )

where Y is a random variable with the same distribution as the first failure time W1 . This
means that after the ith CM, the system behaves as a new one having survived until Ai .
Then, it is easy to prove that the failure intensity is [5]: λt = λ(t − CKt− + AKt− ).
A virtual age model can be defined by a particular value of the effective ages Ai and
by an initial intensity. For instance, we have:

• ABAO: Ai = Ci .

• AGAN: Ai = 0.

• The Brown-Proschan (BP, [1]) model: each maintenance is AGAN with probability
p and ABAO
 with probability
 1 − p. It corresponds to a virtual age model for which
i
X i
Y
Ai =  (1 − Bk ) Wj , where Bk is a random variable independent of Wk and of
j=1 k=j
{Wj , Bj }1≤j<k , and its distribution is Bernoulli with parameter p. The Bk represent
the efficiencies of successive maintenances.

• The Arithmetic Reduction of Age model with memory ∞ (ARA∞ , [5]): the effect of
ith maintenance is to reduce the virtual age of the system of an amount proportional

4
 Advanced Maintenance Modelling

to its age just before the failure. Then, the effective age is: Ai = (1 − ρ)(Ai−1 + Wi ),
Kt− −1
(1 − ρ)j CKt− −j )
X
and we prove that the failure intensity is: λt = λ(t − ρ
j=0

• The Arithmetic Reduction of Age model with memory 1 (ARA1 , [5]): the effect of
ith maintenance is to reduce the virtual age just before maintenance, Ai−1 + Wi , of
a quantity proportional to the time elapsed since last maintenance ρWi . Then, the
effective age is: Ai = Ai−1 + (1 − ρ)Wi = (1 − ρ)Ci , and the failure intensity is:
λt = λ(t − ρCKt− )

1.4.3 Asymmetrical Virtual age models AVA

The concept of Generalized Virtual age models (GVA) has been introduced in [6]. Let’s
consider two competing risks {(Zk , Rk )}k≥1 . The idea of the model is to assume that there
exist a sequence of effective ages {Ak }k≥1 , with A0 = 0, such that after k th maintenance,
the risk variables Rk+1 and Zk+1 behave as the risk variables of a new system having
survived until Ak . Then, we obtain:

P (Rk+1 > r, Zk+1 > z | Ak , w1 , . . . , uk ) = P (R > Ak +r, Z > Ak +z|R > Ak , Z > Ak ) (4)

where (R, Z) is a random couple with the same distribution as (R1 , Z1 ).

In this case, the effect of maintenance is symmetrical since the same effective age
Ak influences both risks Rk+1 and Zk+1 . It is then possible to generalize the concept
by defining Asymmetrical virtual ages models (AVA) {(AR,k , AZ,k )}k≥1 , leading to the
following equation [6]:

P (Rk+1 > r, Zk+1 > z | AR,k , AZ,k , w1 , . . . , uk ) = P (R > AR,k +r, Z > AZ,k +z|R > AR,k , Z > AZ,k )
(5)
One of the major issues in the competing risks framework is the identifiability of the
model. There is identifiability problem in the classical approach or in the generalized
virtual age approach. However, considering Asymmetrical Virtual ages allows, under
certain conditions, to obtain a complete identifiability of the model. The reason is that in
the classic or GVA approach, the age related to each risk is systematically identical. This
leads to compute the joint survival function only on the diagonal. It is then impossible
to identify the joint survival function on the whole bidimension space. The AVA models
present more flexibility and require to have the expression on the joint survival function
at every point.

2 The model
2.1 Motivations
The majority of the studies considering systems presenting a burn-in period are based
on a bathtub shaped intensity. This classic shapes allows to take into account a burn-in
period when the function is decreasing, a potential useful life when the function is constant
and the useful life when the intensity is finally increasing. However, it seems relevant to
consider a modelling based on the competing risks framework for multiple reasons:

5
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

• The bathtub shaped intensity is the addition of two very different processes, one
linked to the degradation of the system and the other linked to failures due to design
defects. It seems then relevant to differentiate these two causes of failures.

• When considering non increasing function such as bathtub shaped functions, it is

quite complex to fit a model of imperfect maintenance such as virtual ages models
[3]. The reason is quite simple: virtual age models assume that after a repair, the
system behaves as a younger system. But a younger system does not necessarily
mean a more reliable system since there could be manufacturing defect present.

• It would be interesting to present a framework where the information about the

causes of failure are available. This additional information allows to have a more
precise analysis of the system. In this context, the competing risks framework seems
completely appropriate. However, we will present how to use this framework when
the causes of failure are not recorded and focus our last study in this particular case.

• This model is flexible in the sense that other failure modes or other kinds of preven-
tive maintenance can be added with a minimal impact on the general modelling.

2.2 The assumptions

The assumptions of our model are as follows:

• The PM times are deterministic and can be considered as a censoring of the com-
peting risks process.

• The risks Z1 and R1 for a new system are independent. This assumption seems
realistic as the nature of the risks are very different.

• The marginal distribution of Z1 is a Weibull distribution: λZ (t) = αZ βZ tβZ −1 , with

parameter αZ > 0 and βZ > 1.

• The marginal distribution of R1 is a Weibull distribution: λR (t) = αR βR tβR −1 , with

parameter αR > 0 and βR < 1.

• The maintenance efficiency after a CM due to a burn-in failure is minimal (ABAO).

This kind of maintenance simply restarts the system in the same state as before the
burn-in failure. It does not make sense to restore the system in a younger state for
this risk, granted that it will correspond to insert new manufacturing defects and
then to carry out a harmful maintenance. Moreover, we assume that there is no
impact on the degradation process.
(
AR,n = Cn
Un = −1 ⇒
AZ,n = AZ,n−1 + Wn

• The maintenance efficiency after a CM due to a degradation failure is of ARA∞

type with parameter ρ0 . This choice is made as an example and should be strongly
dependent on the kind of systems we study. There is no impact on the burn-in
process.

6
 Advanced Maintenance Modelling

(
AR,n = Cn
Un = 0 ⇒
AZ,n = (1 − ρ0 )(AZ,n−1 + Wn )

• The maintenance efficiency after a PM is of ARA∞ type with parameter ρ1 . There

is no impact on the burn-in process.
(
AR,n = Cn
Un = +1 ⇒
AZ,n = (1 − ρ1 )(AZ,n−1 + Wn )

It is possible to have an explicit expression of the effective ages AR,n and AZ,n . To
simplify the notations, let’s introduce the parameter ρ−1 = 0, which corresponds to assume
that the CM after a burn-in failure is of ARA∞ type with parameter ρ−1 = 0, which is
equivalent to a minimal maintenance. Naturally, as maintenance after a birth failure is
minimal, AR,n is always equal to the real age of the system. The effective ages AZ,n need
simple and straightforward calculations to have an explicit expression. Finally we obtain
the following expressions:

∀n > 0, AR,n = Cn

n Y
X n
∀n > 0, AZ,n = [ (1 − ρuk )]Wj
j=1 k=j

Figure 1 presents an example of the evolution of the maintenance intensities.

Figure 1: Example of trajectory

3 Two considerations on the model

3.1 Considering recorded kinds of maintenance
In a first part, we consider that both times and kinds of maintenance are recorded. It
is then possible from the dataset to distinguish both causes of failures. Considering
independent risks, the knowledge of the initial marginal distributions and the effective
ages are enough to completely define the process. In particular, we can easily calculate

7
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

the maintenance intensities and the likelihood function associated to the observation of
couples {(Wj , Uj )}1≤j≤n . The general expression of the likelihood function is as follows:

n
SZ (ai−1 + wi )
[λR (ci )]1I{ui =−1} [λZ (ai−1 + wi )]1I{ui =0}
Y
L(w1 , u1 , .., wn , un ) = SR (cn )
i=1 SZ (ai−1 )

with ai = AZ,i = ij=1 [ ik=j (1 − ρuk )]Wj . The expression remains correct for other kind
P Q

of virtual ages for AZ,i . The Maximum Likelihood estimation leads to have some explicit
estimators (αR , βR , αZ ), and MLEs of other estimators can be derived by simple numerical
methods. In particular, we obtain:
n
X
ui (ui − 1)
i=1
α̂R = ˆ
2cβnR
n
X
ui (ui − 1)
i=1
β̂R = X
n
cn
ui (ui − 1)log( )
i=1 ci
n
X
2n − ui (ui − 1)
i=1
α̂Z = n
ˆ ˆ
[(âi−1 + wi )βZ − (âi−1 )βZ ]
X
2
i=1

n n
(βˆZ , ρˆ0 , ρˆ1 ) = arg max (n− [(ai−1 +wi )βZ −(ai−1 )βZ ])+log(βZ ))
X X
ui (ui −1)/2)(−log(
βZ ,ρ0 ,ρ1
i=1 i=1

n
X
+βZ ui (ui − 1)/2log(ai−1 + wi )
i=1
Pi Qi
with ai = ai (ρ0 , ρ1 ) = j=1 [ k=j (1 − ρuk )]Wj .

3.2 Considering unrecorded kinds of maintenance

3.2.1 Presentation
In many dataset, the exact kinds of corrective maintenances are generally not recorded.
For this model, it corresponds to only observe the maintenance times Wi and the kinds
of maintenance with no distinction between the 2 causes of failures. As the effective ages
AZ,n strongly depend on the kinds of the previous maintenances, it becomes far more
complex to express the intensities and likelihood function in a statistical approach. We
present some ideas to overcome this issue and we will focus on the latest method in the
next section:

• As the Ui are not observed, it is necessary to condition on types of past maintenances.

This will lead to computation of a ”proper” intensity and proper likelihood function.
This conditioning is extremely expensive in time granted that we need to evaluate

8
 Advanced Maintenance Modelling

if a maintenance is due do birth defects or due to the degradation. If the dataset

contains n observations, the cost of the conditioning is O(2n ) which is difficult to
compute for large values of n.

• In practice, the manufacturing defects occur during the burn-in period. By defini-
tion, this period takes place at the startup of the system. A reasonable compromise
is then to use an expert opinion to judge the end of the burn-in period. Then we
can compute the conditioning for the first maintenances and assume that the other
maintenances are due to the degradation. The cost of the conditioning is then O(2p )
with hopefully p << n. An even simpler possibility should be to assume that all
the first maintenances are due to burn-in defects and the other to the degradation.

• The basic idea of the Expectation-Maximization method is to find the maximum-

likelihood estimates of the parameters of an underlying distribution from a dataset
when the data is incomplete or has missing values. The method leads to have a
sequence of parameter estimators converging to the actual estimates.

3.2.2 The EM algorithm

The algorithm that we propose is a simplified version of the actual EM algorithm. The rea-
son is that the EM algorithm implies the computation of P (ui = −1|w1 , . . . , wn , u1 , . . . , ui−1 )
that we modify by P (ui = −1|w1 , . . . , wi , u1 , . . . , ui−1 ). The first quantity is the proba-
bility to observe a burn-in failure given all previous kinds of maintenance and all the
maintenance times. The second quantity is the probability to observe a burn-in failure
given all the previous times and kinds of maintenances including the current maintenance
time. These two probabilities are different granted that the information on the future
maintenance times can affect the probability of observing a burn-in failure. However, it is
reasonable to consider that the impact of the future times of maintenance is limited and
in practice, these two probabilities are extremely close.
The main steps of the algorithm are:

• Initialization part: We set initial parameters n = 0, θn = (αR

n
, βRn , αZn , βZn , ρn0 , ρn1 ).

• E-step: For each value i ranged from 1 to k, number of observations:

– if the ith maintenance is a PM, then Ui = +1.

– if the ith maintenance is corrective, compute the following probability:

λR (ci )
pi = P (ui = −1|w1 , . . . , wi , u1 , . . . , ui−1 ) =
λR (ci ) + λZ (âi−1 + wi )
– If the previous probability is greater than 0.5, we consider than the maintenance
is due to a burn-in failure otherwise due to a degradation failure:

Ûi = −1I{pi >0.5}

– Estimate the virtual age Âi = (1 − ρÛi )(Âi−1 + Wi ).

• M-step: Maximize the complete likelihood function L(w1 , û1 , .., wk , ûk ) as pre-
sented in the previous section. Then, we obtain a new set of estimators θn+1 =
n+1
(αR , βRn+1 , αZn+1 , βZn+1 , ρn+1
0 , ρ1n+1 ).

9
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

• Stopping condition: If there is convergence of θn+1 , we stop the algorithm. Oth-

erwise we use the parameters θn+1 back in the E-step. Obviously, we have to create
a stopping criteria in case of no convergence.
We can add some general remarks about the implementation of the algorithm:
• The EM algorithm is generally very sensitive to the choice of the initial estima-
tor. In particular, by choosing initial parameters absolutely not consistent with the
observations, we have a non negligible risk of having no convergence.
• The expression of the probability pi presented during the E-step can be proved.
First let’s consider a system with two new and independent components in series R
and Z with failure rates λR and λZ . Then we have:

λR (w)
P (R caused the failure|time to failure=w) =
λR (w) + λZ (w)
After the (i-1)th failure, the system R is restored as bad as old and the system Z
is restored with an age ai−1 . The two components are still functioning indepen-
dently, the only differences from a new system are the new failure rates of R and Z,
respectively λR (ci−1 + t) and λZ (ai−1 + t). Then we finally obtain:

λR (ci−1 + wi )
P (R caused the ith failure|ith inter failure time=wi ) =
λR (ci−1 + wi ) + λZ (ai−1 + wi )

This proof can directly be adapted in our case to compute pi .

• There is different possibilities for the stopping condition. In our case, we chose
a Cauchy criteria such as |θn+1 − θn | < . We compute a maximum number of
iterations in case of no convergence.

3.2.3 Statistical study of the model with unrecorded kinds of maintenances

We present a general overview of the performance of the EM algorithm considering our
model:
• As it is quite difficult to simulate realistic PM times granted they are assumed
to be deterministic, we first made simulations considering only CM. The previous
methodology and the EM algorithm remain correct. Then, there is no ρ1 parameter
to estimate.
• The following estimators have been chosen during the initialization part:
1
αR = √ βR = 0.5
W1
n 1
ρ0 = 0.5 αZ = or αZ = βZ = 1.5
cβnz wnβz
For the shape parameters and the maintenance efficiency, the initialization choice
seems natural. The EM algorithm is more sensitive to the initialization of the
pseudo-scale parameters αZ and αR . Their initialization is chosen considering basic
properties of the NHPP or Renewal Process.

10
 Advanced Maintenance Modelling

• Performances of the E-step are extremely good. In few iterations, the types of CM
are identified with the E-step. Table 1 presents one example of estimation. General
validity of the estimation of the kinds of maintenance based on simulations will be
presented during the oral talk.

Ci Wi Ui Ûi
0.84 0.84 -1.0 -1.0
1.72 0.87 -1.0 -1.0
3.63 1.91 -1.0 -1.0
7.27 3.63 -1.0 -1.0
15.02 7.74 0 -1.0
30.73 15.70 -1.0 -1.0
51.49 20.76 0 0
89.47 37.97 0 0
105.89 16.42 0 0
131.03 25.14 0 0
154.59 23.55 0 0
170.90 16.30 0 0

Table 1: Example of estimation of the kinds of maintenance

(αR = 3, βR = 0.7, αZ = 0.01, βZ = 2, ρ0 = 0.7).

• Logically, when the burn-in period represents the major part of the observations,
the estimations of αR and βR are better. On the other hand, when the major part
of the observations consists in the wear-out period, the estimations of αZ , βZ and
ρ0 tends to be better.

• The period of time when we can observe consecutively both kind of CM is relatively
short. This means that a sequence of Ui such as 0, 0, −1, 0, −1, 0 never exceeds four
or five observations. This may be a limit of the Weibull intensities. Meanwhile, it
allows to have a quite precise idea of the end of the burn-in period. This information
can be very useful for the maintenance team.

First results of estimations are very positive on the efficiency of the method. A further
and more precise study need to be carried out to validate the method of estimation. The
next step will be then to apply this methodology a real dataset. These two will be
developed during the oral presentation.

Conclusion
In this paper, we have proposed a general modelling of imperfect maintenance for
systems presenting a burn-in period and subject to corrective and preventive maintenance.
The model allows to dissociate the causes and effects of both failure modes by using the
competing risks framework and the virtual age models. Considering the two failure modes
involved, it seems realistic not to have the same assumption on the virtual ages. The first
contribution of the paper is to propose asymmetrical virtual ages. The complexity of
the model increases when we assume that the causes of failure may not be recorded. To
overcome this issue, we have proposed different methods. The second contribution of
the paper is to detail an estimation method using the EM algorithm. First simulations

11
 Proceedings of the 38th ESReDA Seminar, Pécs, May 4-5, 2010

results are very promising. A validation of the model need to be carried out to judge the
robustness of the algorithm but we observe that we can obtain very relevant information
from the estimation such as the end of the burn-in period.

References
[1] Brown, M., Proschan, F. (1983) Imperfect repair, Journal of Applied Probability, 20,
pp 851–859.

[2] Cooke, R.M., Paulsen, J. (1997) Concepts for measuring maintenance performance and
methods for analysing competing failure modes, Reliability Engineering and System
Safety, 55, pp 135–141.

[3] Dijoux, Y. (2009) A virtual age model based on a bathtub shaped initial intensity,
Reliability Engineering and System Safety, 94(5), pp 982–989

[4] Dijoux, Y., Doyen, L., Gaudoin, O. (2008) Conditionally independent generalized
competing risks for maintenance analysis, in Advances in Mathematical Modeling
for Reliability, T. Bedford, J. Quigley, L. Walls, B. Alkali, A. Daneshkhah and G.
Hardman eds., IOS Press, Amsterdam, pp 88-95, 2008.

[5] Doyen, L., Gaudoin, O. (2004) Classes of imperfect repair models based on reduction
of failure intensity or virtual age, Reliability Engineering and System Safety, 84(1),
45-56.

[6] Doyen, L., Gaudoin, O. (2006) Imperfect maintenance in a generalized competing risks
framework, Journal of Applied Probability, 43(3), 825–839.

[7] Kijima, M., Morimura, H. and Suzuki, Y. (1988) Periodical replacement problem
without assuming minimal repair. European Journal of Operational Research, 37, 194-
203.

[8] Langseth, H., Lindqvist, B.H. (2003) A maintenance model for components exposed to
several failure mechanisms and imperfect repair. In Doksum, K. and Lindqvist, B.H.,
editors, Mathematical and Statistical methods in Reliability, Quality, Reliability and
Engineering Statistics, World Scientific Publishing Co., 355–370.

[9] Lindqvist, B.H. (2006) On the Statistical Modeling and Analysis of Repairable Sys-
tems, Statistical Science, 21(4), 532-551.

12