© I F.-\C Sto,ha sti, Control
Yilniu s, Lithuanian SSR, L'SSR, 1'186
EVALUATION OF SYSTEM'S RELIABILITY
WITH A FUZZY PRIOR INFORMATION
O. Hryniewicz
SYsll'lIIS Rnmnh /11.11111111', .\'l'1I'I'/ska 6, \I'a/sOlI', Po/alld
Abstract: The system with a series reliability structure and elements
having exponentially distributed life times is considered.The data from
life tests for each of the system's elements are known.Moreover, it is
assumed that there exists a linear relationship between hazard rates in
test and real conditions but the coefficients of proportionality are
unknown.However, it is also assumed that their values can be evaluated
by experts in the form of fuzzy sets. The method for merging prior in-
formation from these experts' assessments and from the life tests into
the prior distribution of the whole system's hazard rate is proposed.
Closed formulas are obtained for the case of piece-wise linear
membership functions of the fuzzy sets representing the experts'
evaluations.
Keywords. Reliability theorYi statistics, Bayes methodsi Prediction.
INTRODUCTION
One of the most important problems of re -
liability engineering is reliability asse-
ssment of systems during the initial stage
of their exploitation.The r.wuber of systems
to be tested is,as a rule,hardly sufficient
for precise estimation of reliability cha-
racteristics.Consider for example a system
of several elements connected
in series whose life-times are exponentia-
lly distributed. In this case the system'S
life-time is also exponentially distribu-
ted.Now,assume that its mean time to failu-
re equals b.Then,the variance of the stan-
dard maximum likelihood estimator of b is
gi.ven by
where r is the number of observed failures.
Therefore,for reliable system's and small
samples the precision of estimation is in-
sufficient from the practical point of
view. There are two ways which allow us to
obtain more accurate estimators of relia-
bility: either to produce more failures
by applying a more severe stress,increa-
sing the amount of total test time etc.,
or to use prior information to obtain mo-
re accurate estimators of the reliability
to be evaluated,and from the practical
point of view the latter is far more in-
teresting than the former.This evaluation
is based on the fact that even newly de-
signed systems consist of elements whose
reliability can be evaluated from previo-
us experiments.However,there is one seve-
re restriction-these experiments are per-
formed in different conditions lelectrical
or mechanical stresses,enviroment etc.1
and it does not exist any generally accep-
ted method for the transformation of test
-l51
results from one condition of test
to another, Th.e simplust ,and there-
fore widely accepted;method con _
sists in multiplication of certain
reliability characteristics le.g.
the hazard rate I by some coefficie
nts whose values are stress-depen-
dent.These coefficients are tabu-
lated le.g., in the American
Mil i tary Handbook 217 I but t..'1eir
application in practice seems to
be res tr icted. However ,wey can be
evaluated in a more fuzzy form, and
then utilized in the procedures of
reliability prediction.
In the paper we consider the pro-
blem of reliability estimation
with data from three sources: the
tests of the considered system's ,
the tests of elements the system
consists of,and the evaluation of
coefficients used for the transfor
mation of tests results.We utilize
these data to construct the Baye-
sian estimator of the system's ha-
zard rate which gives a more accu-
rate prediction of reliability.
GENERAL ASSUMPTIONS
suppose that the considered system
has a series reliability structure
and its N elements have exponentia
lly distributed times to failures.
Let >-. i' i=l ,2, ..• , N be the hazard
rate of the i-th element.
Suppose now that there are two ty-
pes of tests.First,the whole sy-
stem is tested in normal