432

Application of Bayesian Methods for Age-dependent Reliability Analysis

Robertas Alzbutas
1,2
, Tomas Iemantas
1,2
1
Lithuanian Energy Institute, Breslaujos 3, LT-44403 Kaunas, Lithuania
2
Kaunas University of Technology, Student 50, LT-51368 Kaunas, Lithuania
Abstract
In this paper authors present a general methodology for age-dependent
reliability analysis of degrading or ageing systems, structures and
components. The methodology is based on Bayesian methods and models
its ability to incorporate prior information and on idea that ageing can be
thought as age-dependent change of beliefs about system reliability
parameters (mainly failure rate); beliefs changes not just due to new data or
other information which becomes available in time, but also it changes due to
flow of time.
The main objective of this paper is to present the clear methodology of how
Bayesian methods can be applied to deal with reliability of real ageing
systems. Methodology describes step-by-step failure rate analysis of ageing
systems: from the model building to its verification and generalization with
Bayesian model averaging.
The proposed methodology is able to deal with sparse and rare failure events
as is the case in electrical components, piping systems and various other
systems with high reliability. In a case study of electrical Instrumentation and
Control components the proposed methodology was applied to analyse age-
dependent failure rate together with treatment of uncertainty due to age-
dependent model selection.
1. Introduction
Safety of energy facilities, chemical factories, oil companies, etc. in many
cases depends on their components reliability, which is mainly age-
dependent. Unnoticed on time, ageing effect can cause failures or multiple
damages at given non-standard operating conditions or breakdown situations.
System ageing is mainly caused by two impacts: operating conditions and
technical inspection actions. The ageing itself can be analysed and controlled
depending on the type of the system.
Systems can be distinguished into two main categories: active and passive.
The term passive system identifies a system, which is composed entirely of
passive components and structures, or a system, which uses active
components in a very limited way to initiate subsequent passive operation [6].
The passive components (e.g. heat exchangers, pipes, vessels, electrical
cables, structures, etc.) are usually neglected or not modelled implicitly in the
probabilistic reliability and risk assessment models of complex systems as
having very low failure probability, but they could have an increasing
contribution due to ageing effects. While safety importance of active systems
433
and their ageing were recognised widely, in this paper the focus is on more
complicated reliability data analysis of passive systems.
Age-dependent reliability study requires more data and more extended
models than a usual reliability analysis. With regard to data, one basic issue is
scattering of failure histories for passive components and systems. Because
of this scattering of failure data, reliability and risk model parameters, which
are estimated from the raw data, have large associated uncertainty.
Usually, passive systems/components do not provide large samples of failure
data and because of lack of data it is extremely difficult to deal with
degradation of passive systems/components by using classical statistical
methods which requires considerable amount of data.
However, uncertainty, related to data scattering could be reduced and
assessment of age-dependent reliability of passive systems could be
improved by considering, so called, prior information experience of other
similar facilities, subjective expert insights. Then, by the use of available
statistical data, prior knowledge can be revised by Bayes formula [21]. When
data arrives, Bayes theorem tells how to move from prior beliefs to new
conditional probabilities for the quantities of interest.
The main advantage of the Bayesian approach is that it can rely on multiple
sources of evidence including: warranty data, customer research surveys,
proving ground test data, etc. It also has the potential to systematically
quantify and process soft evidence such as expert knowledge [2].
In addition to the ability to deal with sparse data, Bayesian methods are
appropriate for use in PRA (Probabilistic Risk Assessment) because they are
derived from the framework of subjective probability [9]. Further, as Siu and
Kelly noticed in their paper [9], practical advantage of the subjective
probability framework in PRA applications is that propagation of uncertainty
through complex models is relatively simple. On the other hand, it is very
difficult, and intractable in real problems, to propagate classical statistical
confidence intervals through PRA models to estimate a confidence interval for
a composite result of interest.
Despite of the advantages offered by Bayesian methods, applicability of it was
very limited and generally confined by the use of so called conjugate prior
distributions, which provides analytically tractable problem solutions just for
quite unsophisticated applications.
The advent of Markov Chain Monte Carlo (MCMC) sampling has proliferated
Bayesian inference throughout the world, across a wide array of disciplines
[12]. MCMC methods are a class of algorithms for sampling from probability
distributions based on constructing a Markov chain that has the desired
distribution as its equilibrium distribution [22, 23]. The state of the chain after a
large number of steps is then used as a sample from the desired distribution.
MCMC algorithms enabled development and application of highly complex
Bayesian models. The freely available software package known as Bayesian
434
inference Using Gibbs Sampling (WinBUGS) has been in the vanguard of this
proliferation since the mid-1990s [13].
In spite of this, still there is no agreed Bayesian methodology for analysis of
age-dependent failures of systems. Thus, authors suppose that this paper will
be a suitable contribution to advancement of reliability analysis. The
methodology presented in the paper will help to assess more precisely the
age-dependent reliability and to understand and express uncertainty of result.
2. Theoretical issues of age-dependent failure rate
2.1 Dynamic models for failure rate analysis
If the evidences show possible trend in statistical failure data, then one can
consider trend model for failure rate. Several examples for failure rate trend
can be:
- piecewise constant
0
( ) t = ; (1)
- linear
1 2
( ) t t u u = + ; (2)
- exponential (log-linear)
1 2
ln ( ) t t u u = + ; (3)
- power-law (Weibull)
2
1
( )
t
t e
u
u = ; (4)
- Xie and Lai model ( ) ( ) ( )
2 4
1 1
1 2 1 3 4 3
t t t
u u
u u u u u u

= + , [24]; (5)
- generalized Makeham ( )
2 3
1
4
1
t
t e
t
u
u
u
u
= +
+
, [3]. (6)
Linear ageing is simple, an obvious natural way to give a first-order
approximation to changes in the failure rate, but it does seem to have a
practical disadvantage. Wolford et al. (1992) [4] analyzed two data sets using
several functional forms for ( ) t ; one such analysis is reported by Atwood
(1992) [5]. They found that a Bayesian posterior distribution for ( ) t was
approximately lognormal when a log-linear or power-law model was used for
( ) t , but this was not a case when a linear model was used. Apparently, the
approximate log normality required a much larger data set when linear ageing
was assumed in comparison to the case when power-law or exponential
ageing were assumed.
Usually failure rate can be divided into three distinct periods: burn-in period,
useful life, wear-out period. For such general trend the linear, power-law or
exponential distribution cant provide desirable fit to the data. Due to this
reason, models that have ability to shape-up whole bathtub curve are needed
and Xie & Lai or generalized Makeham trend models can be applied [3].
Notwithstanding flexibility of these models, it can be quite difficult to apply
them by using frequentist framework, because due to number of parameters
and lack of data. Classical statistical methods are ill-suited for this situation,
leading in such cases to excessively wide confidence intervals.
Some authors [6, 7, 8] introduce a threshold of age at which ageing is
assumed to begin. Then ( ) t is assumed to be a constant before the
threshold of age is attained, and to increase according to one of the above
435
formulas afterwards. The threshold is generally unknown, and must be
estimated from the data. Thresholds cause difficulty in classical statistics,
because the assumptions for the asymptotic theory of maximum likelihood
estimation are typically violated. Therefore, it is difficult to quantify the
uncertainty in the estimate of the threshold. However, Bayesian estimation,
using simulation package such as BUGS [28], is still possible.
2.2 Choice of prior distribution
As noted in introduction section, Bayesian methods are capable to joint
various sources of information: statistical data, expert opinions, historical
information, experience in quantification of uncertainty in similar systems or
components, etc. These sources can be classified into subjective (e.g. expert
opinions, etc.) and objective (statistical data).
To quantify subjective information, the need of subjective probability
framework arises. The theory of subjective probability has been created to
enable one to talk about probabilities when the frequency viewpoint does not
apply. The main idea of subjective probability is to let the probability of an
event reflect the personal belief in the chance of the occurrence of the event
[21]. In Bayesian theory, such subjective probability is expressed in terms of
subjective prior distribution. These distributions have their own advantages
and disadvantages [25]:
1. Subjective prior distributions are proper (they integrate or sum to 1);
2. They are generally well behaved analytically. The effect of a subjective
prior distribution on the posterior is as if there were additional replications
of the data;
3. Subjective prior distributions may be used to introduce the informed
understanding, beliefs, and experience of a scientist or decision maker into
the Bayesian analysis of a problem to take advantage of this additional
information about the phenomenon under study;
4. Often there is insufficient information in a problem to solve it by objective
Bayesian or frequentist methods. The subjective Bayesian approach may
be the only way to find an acceptable solution since it brings additional
information to bear on the problem.
Disadvantages:
1. It is not always easy to assess subjective prior distributions because it is
not always easy to translate the prior knowledge into a meaningful
probability distribution.
2. Results of a Bayesian analysis that used a subjective prior distribution are
meaningful to the particular analyst whose prior distribution was used, but
not necessarily to other researchers, because their priors might have been
different.
However, the situations when no useful prior information is available are quite
frequent. In that case Bayesian analysis can be carried out by using a weak or
non-informative prior, that expresses prior ignorance about quantities of
interest.
436
In general, the use of non-informative prior distributions causes some
problems in applications: such distributions are improper (i.e. they do not
integrate to 1) and sometimes can yield improper posterior distribution; also
Bayes factors, that are common quantities in models comparisons, cannot be
calculated [25].
Non-informative prior distributions also cannot be used in WinBUGS software,
because the Markov chain generated by MCMC [20] never converge (since
the chain cannot be recurrent) [29]. Appropriate choice for WinBUGS
simulations can be diffuse prior distributions. These distributions express as
little information as possible, but they are still proper.
Several suggestions can be made for choice of prior:
1. For positive values gamma distribution with parameters 0.01, 0.01 a b = =
can be used, because it gives approximately equal weights for positive
values (more diffuse gamma distribution cause difficulties in WinBUGS
sampling).
2. If quantity of interest is restricted to the interval [0,1] , then appropriate
diffuse distribution is ( ) 1,1 beta or ( ) 0.5, 0.5 beta .
3. When parameter is expected to be somewhere around 0, then diffuse
normal distribution (e.g. normal distribution with zero mean and very small
precision) is helpful.
For all these cases uniform distribution is appropriate, but range should not be
too narrow.
2.3 Application of Bayesian methods for age-dependent modelling
Ageing can be thought as age-dependent change of beliefs about systems
reliability parameters (mainly failure rate). Beliefs changes not just due to new
data or other information (mentioned above) which becomes available in time,
but also it changes due to flow of time.
One of the difficulties of Bayesian inference is inability to deal with changes of
age-dependant parameter as a continuous process. This problem partially can
be overcome by considering ageing (or degradation) as step-wise process,
which is constant in some period of time and has value jump in other period.
Mathematically this can be expressed as a jump process:
( ) ( )
1
1
{ }
1
1
N
i
i
t t t
i i
d t d t

=
< <
+
=

; (7)
where ( ) d t is any model of characteristic (or reliability parameter) under
consideration and constant ( )
i
d t is value of characteristics at each time
period
i
t ; N number of time intervals.
437
Model of characteristic ( ) d t can have any functional form. It can be linear,
Weibull, or some other form. Depending on adopted formula, ( ) d t will be
based on vector of parameters { }
1
,...,
m
u u O = :
( ) ( ) , d t d t = O . (8)
If analysis considers more than one model, then indexation is used for
different models, i.e. ( ) ,
i i
d t O , where
i
d denotes i
th
model with
i
O vector of
parameters.
Modelling conception introduced above allows interpreting distribution of
parameters as age-dependent. If prior knowledge and beliefs about system
reliability parameter is represented by probability density distribution ( ) t O
and statistical observations has likelihood ( ) ( )
| f y d t , then, according to
Bayes theorem, age-dependent beliefs about system reliability parameter or
failure rate is expressed as posterior distribution:
( )
( ) ( ) ( )
( ) ( ) ( )
| ,
| ,
| ,
f y d t
y t
f y d t d
t
t
t
O
O O
O =
O O O
}
. (9)
Assume that parameters
1
,...,
m
u u are a priori independent, then, according to
definition of independent random variables, prior distribution of O can be
expressed as:
( ) ( )
1
m
i i
i
t t u
=
O =
[
; (10)
where ( ), 1,
i i
i m t u = are priors for components of vector O.
If data set contains n statistical observations, then posterior distribution is
represented as:
( )
( ) ( )
( ) ( )
1 1
1 1
| ,
| ,
| ,
m n
i i j
i j
m n
i i j
i j
f y d t
y t
f y d t d
t u
t
t u
= =
= =
O
O
O =
O O
[ [
[ [
}
. (11)
Usually it is the case when several trend models fits data almost equally well,
i.e. possible set of good models can be represented as
( ) ( ) ( )
1 1
, ,..., ,
r r
d t d t M = O O , (12)
438
where ( ) , , 1,
i i
d t i r O = are models which were considered as having good fit.
In such circumstances uncertainty of modelling cannot be handled
appropriately within classical statistical framework.
As noticed in [10], standard statistical practice ignores model uncertainty.
Data analysts typically select a model from some class of models and then
proceed as if the selected model had generated the data. This approach
ignores the uncertainty in model selection, leading to over-confident
inferences and decisions that are more risky that one thinks they are.
According to Hoeting [10], Bayesian averaging advantages include better
average predictive performance than any single model that could be selected.
Model averaging is more correct because it takes into account a source of
uncertainty that analyses based on model selection ignore [11].
Denote ( ) D t as failure rate averaged over set of models M . Considering our
notation, posterior probability of averaged age-dependent failure rate can be
represented as:
( ) ( ) ( ) ( ) ( )
1
| | , , , |
r
j j j j
j
p D t y p D t y d t p d t y
=
= O O

. (13)
Posterior probability distribution for model
j
M is given by
( )
( ) ( )
( ) ( ) ( ) ( )
1
| , ,
|
| , ,
j j j j
j r
l l l l
l
p y d t p d t
p M y
p y d t p d t
=
O O
=
O O

; (14)
where
( )
,
j j
p d t O is prior probability distribution of models. In the case of
non-informative prior distribution, equal discrete probabilities can be assigned
for each model
( )
1
,
j j
p d t
r
O = and posterior probability distribution for model
( )
,
j j
d t O becomes:
( )
( )
( ) ( )
( )
( ) ( )
1 1
1
| ,
| ,
, | , 1,
1
| , | ,
j j
j j
j j r r
l l l l
l l
p y d t
p y d t
r
p d t y j r
p y d t p y d t
r
= =
O
O
O = = =
O O

. (15)
Even though Bayesian Model Averaging (BMA) seems to have advantages
over one-model-fitting, little work has been done in the engineering field to
address this for model uncertainty. Alvin et al. [18] used BMA to predict the
vibration frequencies of a bracket component, Zhang and Mahadevan [19]
applied it in fatigue reliability analysis on the butt welds of a steel bridge, and
most recent work was done by Inseok Park et al. [27]. Authors analyzed
439
uncertainty of 4 finite elements models for laser peening process. However, all
these works used relatively simple models and probabilistic approaches and
there was no need to adopt advanced probability sampling techniques such
as Markov Chain Monte Carlo methods [20].
2.4 Model selection
There are various techniques for model validation in Bayesian framework [14,
15, 16]. One of possible approaches to analyse model fitness is to use tail-
area probability or as it is sometimes known, the posterior predictive p-value:
( ) ( )
( ) ( )
( ) ( )
, ,
, , | | |
rep
rep rep rep
D y D y
p P D y D y y I p y p y dy d
u u
u u u u u
(
>

= > =
} }
; (16)
where
rep
y is the replicated data that could have been observed, or, to think
predictively, as the data that would appear if the experiment that produced y
were replicated in future with the same model [15]. Posterior p-value
expresses the differences between statistical data and replicated. Rule of
thumb is p-values close to 0.5 [14, 26].
( ) , D y u is discrepancy measure and can have any functional form, e.g.:
( ) ( ) ( )
2
1
; D y E Y E Y u = and ( )
| | ( )
| |
2
2
|
;
|
t t
t
y E y
D y
Var y
u
u
u

. (17)
Chi-square statistics ( )
2
; D y u is quite popular among researchers; however
as will be showed the use of just one discrepancy measure can be very
misleading.
The use of discrepancy measures can be used to assess fitness of each
model individually, i.e. rejection and acceptance of one model does not
depend on other models.
Another possible way to analyse fitness of models is to use Deviance
Information Criterion (DIC), which is already implemented in WinBUGS as
inner function. DIC can be used to compare different models with each other.
Spiegelhalter et al. [17] suggest the following rule of thumb: that models with
DIC difference within the minimum value lower than two (2) deserve to be
considered as equally well, while models with values ranging within 2-7 have
considerably less support. DIC of i
th
model is defined as:
( ) ( )
2ln | , 2
i D
DIC L y i p = O + ; (18)
where
D
p is the effective number of parameters [17].
More information about Bayesian model selection can be found in [14, 16].
440
3. Case study
3.1 Data representation
Data set represents the failure and replacement dates of electrical
instrumentation and control (I&C) components. The considered data is quite
similar to the real (data were encoded and places where it was collected cant
be identified) operating experience data collected in French or German
nuclear power plants. In particular, it is a large sample that represents one
technological group of continuously operating components. The data set
contains records from type T reactors, which are operated by a single utility
with a single management philosophy. The components and composition of
them in all reactors are similar (design, manufacturer, technology, etc.). In all
reactors the components of type A are subjected for ageing effect as their
operate in the environment with more stressful pressure and temperature. The
scope of maintenance is the same for all components.
All data were collected during eleven years, from January 1, 1990 through
December 31, 2000. The components in the sample do not all have same
date of being put into service, and as a consequence do not have the same
ages at the beginning and end of observation. The failure counts were taken
from a review of the maintenance data, so any reported date of failure is
actually the date of the periodic test. A critical failure is one that causes the
component to lose its safety function modelled for PRA.
There were 20 reactor units of type T, each with 20 components of type A.
So, each year there would be 400 component-year except for the fact that
some of the reactor units were commissioned before and after the start of the
data collection (Table 1).
Failure rates, presented in Table 1, gives the first impression about failure
behaviour over time: failure rate increases in time showing system ageing
effect. Also several statistical tests were performed for the ageing verification
and were presented in the report of JRC Institute for Energy [4].
3.2 Bayesian model for non-homogeneous Poisson count data
In this analysis, failure rates are considered as constant values in each year,
but at every year this value jumps at the value which can be calculated from
linear, Weibull or other model.
Consider as the model for the failure rate ( ) { }
; 0 t t > a jump process
structure described above:
( )
{ }
1
1
1
N
i
i
t t t
i i
t
=
< s
+
=

. (19)
In each year period failures occurs as non-homogeneous Poisson process but
with different failure rate parameter , 1, 2,...,15
i
i = . In every time period (which
in this case is equal to one year) equipment was in operation for
i
t time
441
(operating time). Denote number of failure that occurred in one year as
i
N .
Probability of failure can be expressed as:
( )
( )
!
i i
k
i i
i
e
P N k
k
t
t

= = . (20)
Age,
Years
Number of
component failures
Component operating
time, Years
Failure rate,
1/Year
1 1 126.60 0.0079
2 1 171.62 0.0058
3 3 231.36 0.0130
4 1 314.80 0.0032
5 10 396.60 0.0252
6 8 400.00 0.0200
7 16 396.76 0.0403
8 11 380.00 0.0289
9 12 363.34 0.0330
10 8 336.73 0.0238
11 16 281.68 0.0568
12 9 273.42 0.0329
13 10 288.44 0.0347
14 16 168.58 0.0949
15 15 85.16 0.1761
Table 1. Failure data of I&C components under consideration
Likelihood function, that contains all information obtained from data, is:
( ) ( ) ( ) { }
( ) ( )
1
,
| exp ,
!
i
N
n
i i
i i
i i
t
L P y t
N
t
t
=
O
O = O = O
[
. (21)
Since in data source [4] there is no available information about which
particular I&C components were under observation, prior distribution for
parameters of failure trend function is chosen as diffuse distribution. In
WinBUGS implementation diffuse prior gamma distribution were assigned for
all parameters and in all models except for Xie and Lai model for one
parameter beta distribution where assigned. So for Xie and Lai failure rate
trend model joint prior distribution can be expressed as
( )
( ) ( )
( )
3
1
1
4 4
1
1
1
1
,
a
i
b
a
i
b
i
a B
e |
o
u
t u u
o |
u

O =
I
[
; (22)
When for other models join prior distribution is:
( )
( )
1
1 a
m
i
b
a
i
b
i
a
e
u
t
u
=

O =
I
[
. (23)
442
Stated model can be graphically represented within Doodle menu in WinBUGs
software (Figure 1), where oval nodes represents stochastic and logic
variables, arrows show how variables are related and nodes standard and
chi represents discrepancy measures.
Figure 1. Graphical representation of age-dependent stochastic model
In further analysis 5 trend models of failure rate were considered. Linear,
exponential and power models represent class of trends which is common in
ageing analysis and Makehan and Xie & Lai models represents more flexible
bathtub trend class. We excluded constant failure rate model because ageing
effect of considered data [4] has been already validated in other analysis [5].
3.3 Model fitting and screening
Estimated posterior p-values for different failure rate models are in Table 2:
Linear Exponential Power law Gen. Makeham Xie & Lai
p(D
1
) 0.5458 0.6333 0.7134 0.6178 0.7006
p(D
2
) 0.0042 0.0278 0.0102 0.0306 0.0110
Table 2. Posterior p-values for different failure rate models
As can be seen from posterior p-values p(D
2
) presented in Table 1, none of
proposed trend models of failure rate gives good enough fit and all models
should be rejected. However, p-values p(D
1
) shows satisfactory discrimination
abilities linear and generalized Makeham trend models can be interpreted
as better fit than exponential and power low failure rate trend models.
Acceptance by using standard deviation discrepancy measure D
1
can be
justified by simple graphical assessment of fitted models (Figure 1).
443
Figure 2. Comparative representations of accepted trend models
It is well known that more complex curves will fit data more precisely, but
fitness of very complex models can lead to over fitting (e.g. perfect fitness can
be achieved by splines, but this apparently leads to nonsensical inference).
Nevertheless, this obscurity can be solved by using DIC measure. This
criterion naturally adopts Occams razor principle, because it incorporates
penalty - the effective number of parameters: more complex models will be
penalized more severely. DIC values for all models under consideration are
presented in Table 3.
Model Linear Exponential Power law Gen. Makeham Xie & Lai
DIC 90.416 93.869 95.871 87.824 91.730
Table 3. Values of Deviance Information Criterion
As can be seen from DIC values, generalized Makeham model shows best fit.
Also, linear and Xie & Lai model can be accepted.
Two measures of fitness discrepancy measure and DIC shows slightly
different results and unambiguous answer cannot be given. Preference to one
model over another can lead to too pessimistic or optimistic predictions of
ageing phenomena behaviour. Such uncertainty related to the selection of
model for further use has to be quantified to make sure that applications of
model will not be influenced on incorrect choice of trend. Such quantification
will be demonstrated in further analysis where Bayesian model averaging
(BMA) will be applied.
3.4 Bayesian averaging for age-dependent failures
As was concluded previously, discrepancy measure and DIC gave quite
ambiguous results; these criterions together distinguished three models as
Statistical data
Linear trend
Gen. Makeham trend
Xie & Lai trend
444
having best fit linear, generalized Makeham and Xie & Lai trends. In
practice, usual decision is to adopt just one model, but as mentioned in
theoretical part, this could lead to overoptimistic results if model uncertainty is
not incorporated into modelling process.
In this part of paper application of Bayesian model averaging to analyze age-
dependent failures will be demonstrated. We will perform averaging procedure
for three models: linear, generalized Makeham and Xie & Lai. This selection
was made after evaluating posterior p-values, deviance information criterion
and visual assessment of trends.
The results of Bayesian model averaging are presented in Table 4.
Statistical data Averaged
failure rate
Linear
model
Gen. Makeham
model
Xie & Lai
model Age Failure rate
1 0.007901 0.01155 0.005689 0.009791 0.005329
2 0.005827 0.01174 0.009651 0.010970 0.006956
3 0.012967 0.01325 0.013610 0.012600 0.009997
4 0.003177 0.01543 0.017570 0.014650 0.013640
5 0.025214 0.01813 0.021530 0.017120 0.017690
6 0.020000 0.02130 0.025500 0.020080 0.022080
7 0.040327 0.02497 0.029460 0.023620 0.026740
8 0.028947 0.02918 0.033420 0.027830 0.031660
9 0.033027 0.03401 0.037380 0.032840 0.036810
10 0.023758 0.03955 0.041340 0.038800 0.042190
11 0.056802 0.04593 0.045300 0.045890 0.047770
12 0.032916 0.05331 0.049260 0.054330 0.053550
13 0.034669 0.06187 0.053220 0.064370 0.059520
14 0.094910 0.07185 0.057180 0.076340 0.065680
15 0.176100 0.08354 0.061150 0.090620 0.072010
Table 4. Bayesian model averaging and failure rates of various models
Posterior probabilities of linear, generalized Makeham and Xie & Lai models
(0.1186, 0.7338, 0.2662 accordingly) shows that accelerating ageing (i.e.
Makeham and Xie & Lai) is more probable. These results shows that if just p-
values had been used for validation of model fitness and linear trend is
chosen, then it is highly probable that the use of this model would lead to
overoptimistic forecasting.
Posterior averaged failure rate values are better calibrated, because not only
uncertainty in parameters accounted, but also uncertainty regarding model
selection. Of course, accounting of model uncertainty cannot be fully
performed, because there is infinite number of possible models.
4. Conclusions and final remarks
In this paper authors presented general methodology of Bayesian methods
application for age-dependant analysis. It was showed that this methodology
is able to deal with disperse and small data amount along with multiple
parameter set (Makeham and Xie & Lai trend models). As an illustrative
445
example, the proposed methodology was applied for ageing analysis of
electrical I&C components. This application was carried in terms of non-
homogeneous Poisson model with several failure trends.
For fitting and screening of various trend models, it was noticed that none of
model selection approaches can give unambiguous answer. P-values can be
quite misleading and can either show no discriminatory abilities (as in case of
chi-square p-value) or can suggest more than one model as having good fit
(as in case of standard deviation p-value). Deviance information criteria can
also suggest more than one model (and not necessarily the same one as p-
value criteria). Thats why there is a high chance to omit model which can also
lead to satisfactory results. Thus, model selection should be performed very
carefully. It is worth to mention, that other model selection and validation
criteria (such as Bayesian information criteria, Bayesian factors, etc. not
described or used in this paper) can also suffer from such shortcomings.
To evade the vices of model selection and validation approaches, Bayesian
model averaging were performed for the best three (according to assessment
of p-values and DIC) trend models of failure rate. Such averaging over set of
selected trends finally results to better predictive performance, because
averaged future failure rates will not be such over-optimistic as would be with
linear trend and will not be too pessimistic as would be with Makeham trend.
Notwithstanding all the advantages of Bayesian model averaging, this
approach also undergoes some problems: BMA cannot deal with infinite set of
models and when one chooses finite set of possible models the best one can
be not included in this set. Moreover, the set can contain no satisfactory
model and averaging will not result to better performance.
This paper and its results can be used as groundwork for further assessment
of ageing systems, structures and components. Its generality and idea, that
ageing or degradation can be thought as age-dependent change of beliefs
about system reliability parameters, allows analysis of wide spectrum of
problems - it can be stochastic behaviour of crack growth (in this case
characteristic ( ) d t of interest would be crack growth rate), it can be
degradation modelling as transitions through Markovian states ( ( ) d t could be
transition rates, time-homogeneous or time-inhomogeneous, between
degradation states), etc.
Though advantages of Bayesian methods for the particular real case analysis
were demonstrated, but to draw more general conclusions on the impact of
Bayesian inference for solving problem of statistical data rareness and its
scatter in age-dependent reliability analysis, there is a need of separate
investigation of artificial data. As future work authors planned to analyse
Bayesian inference abilities on generated samples of various sizes,
dispersions and trend functions. This will allow to make more general insights
both on advantages and shortcomings of specific Bayesian methods.
446
Acknowledgement
This research was funded by a grant (No. ATE-10/2010) and student research
fellowship award from the Research Council of Lithuania.
References
1. Aritomi M., Crijns J., Dastidar P., Dennielou Y., Fischer J., Forsberg C.
IAEA TEC-DOC-626. Safety related terms for advanced nuclear power
plants, (1991).
2. Krivtso V., Wasiloff J., Classical vs. Bayes reliability growth in theory and
practice. Annual Quality Congress, Indianapolis, IN, Vol. 54, No. 0, pp 311-
316 (2000).
3. Lai C.D., Xie M., Stochastic Ageing and Dependence for reliability.
Springer, New York, (2006).
4. Wolford A.J., Atwood C.L., Roesener W.S., Ageing Data Analysis and Risk
Assessment Development and Demonstration Study. EG&G Idaho, Inc.,
Idaho Falls, (1992).
5. Atwood C.L., Parametric estimation of time-dependent failure rates for
probabilistic risk assessment, Reliability Engineering & System Safety,
Volume 37, Issue 3, 181-194 (1992).
6. Rodionov A., Application of Statistical Methods for Identification of Ageing
Trends // IRSN/DSR Report No. 47, (2005).
7. Okazaki T., Aldemir T., Ageing Effects on Time-Dependent Nuclear Plant
Unavailability with Changing Surveillance Interval. Proceedings of
International Conference PSAM 7 ESREL04. pp 15591566 (2004).
8. Radulovich R.D., Vesely W.E., Aldemir T., Ageing Effects on Time-
Dependent Nuclear Plant Component Unavailability: an Investigation of
Variations From Static Calculations // Nuclear Technology. Vol. 112(1). pp
2141 (1995).
9. Siu N.O., Kelly D.L., Bayesian parameter estimation in probabilistic risk
assessment, Reliability Engineering & System Safety, Volume 62, Issues
1-2, 89-116 (1998).
10. Hoeting J., Madigan D., Raftery A.E., Volinsky C., Bayesian Model
Averageing: A Tutorial. Statistical Science, 14:382-417 (1999).
11. Bijak J., Bayesian model averageing in forecasting international migration.
European Population Conference, Liverpool, (2006).
12. Kelly D.L., Smith C.L., Bayesian inference in probabilistic risk assessment
The current state of the art. Reliability engineering and system safety. 94
628-643 (2009).
13. Lunn D.J., et al., WinBUGS a Bayesian modelling framework: concepts,
structure, and extensibility. StatComput: 32537 (2000).
14. Ntzoufras I., Bayesian Modelling Using WinBUGS. Wiley, (2009).
15. Gelman A., Meng X., Stern H., Posterior predictive assessment of model
fitness via realized discrepancies. Statistica Sinica 6, 733-807 (1996).
16. Dei D.K., Rao C.R., Bayesian Thinking: Modelling and Computation.
Handbook of statistics (2005).
17. Spiegelhalter D., Best N., Carlin B. and van der Linde A., Bayesian
measures of model complexity and fit (with discussion), Journal of the
Royal Statistical Society B 64, 583-639 (2002).
447
18. Alvin K.F., Oberkampf W.L., Diegert K.V., Rutherford B.M., Uncertainty
quantication in computational structural dynamics: a new paradigm for
model validation. In: Proceedings of the 16th international modal analysis
conference, Santa Barbara, CA; pp 11918 (1998).
19. Zhang R., Mahadevan S., Model uncertainty and Bayesian updating in
reliability-based inspection. Structural Safety; 22(2):14560 (2000).
20. Gamerman D., Markov Chain Monte Carlo Stochastic Simulation for
Bayesian Inference. London: Chapman and Hall, (2006).
21. Berger J.O., Statistical Decision Theory and Bayesian Analysis, Second
Edition. Springer, (1985).
22. Gilks W.R., Richardson S. and Spiegelhalter D.J., "Markov Chain Monte
Carlo in Practice". Chapman & Hall/CRC, (1996).
23. Ching W.K., Michael K.N., Markov Chains. Models, Algorithms and
Applications. Springer, (2006).
24. Xie M., Lai C.D., Reliability analysis using an additive Weibull model with
bathtub-shaped failure rate function. Reliability Engineering and System
Safety; 52:87-93 (1995).
25. Chib S., Clyde M., Woodworth G., Zaslavsky A., Subjective and objective
Bayesian statistics: principles, models, and applications. Wiley, (2003).
26. Gelman A., and Meng, X.L., Model checking and model improvement, in
W. Gilks, S. Richardson, and D. Spiegelhalter, eds., Markov Chain Monte
Carlo in Practice, Chapman & Hall, Suffolk, UK, pp. 189-201 (1996).
27. Park I., Amarchinta H.K., Grandhi R.V., A Bayesian approach for
quantification of model uncertainty, Reliability Engineering & System
Safety, Volume 95, Issue 7, July 2010, pp 777-785 (2010).
28. Lunn D.J., Thomas A., Best N., and Spiegelhalter D., WinBUGS -- a
Bayesian modelling framework: concepts, structure, and extensibility.
Statistics and Computing, 10:325337 (2000).
29. Robert C., The Bayesian Choice, Second edition. Springer, (2001).