Professional Documents
Culture Documents
net/publication/266508192
CITATIONS READS
44 167
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Krishna B. Misra on 26 December 2018.
1. INTRODUCTION
Fault tree analysis continues to remain an area of research in the broad
discipline of reliability and safety engineering. The primary objective of fault tree
construction and evaluation is to determine whether or not the system design offers
an acceptable level of safety and reliability, that minimizes the probability of
occurrence of top event (an undesired system event) and is within a specified
bounds. If the system design is found inadequate, the design is modified by first
identifying those critical events that significantly contribute to the top occurrence
of event. When all design changes have been incorporated, the fault tree is
re-evaluated to determine if the revised design provides an acceptable level of
safety.
As discussed in [3, 7 and 8], there are inherent weaknesses in the current
probabilistic risk assessment or safety analysis and these mainly stem from the
existence ofa large amount of uncertainty in the results due to the use of imprecise
failure data and secondly due to uncertainty of modelling in case of dependent
failures, common cause failures, human operator's failures, aging of plants etc. To
mathematically treat some of these uncertainties, Tanaka [1] and Misra [2,3]
proposed fuzzy sets' approach. These approaches are based on fault trees with two
state elements and have certain weaknesses (elaborated in section 4) and are actually
computationally intensive. The computations for arithmetic operations can be
considerably simplified ifL-R representation of fuzzy numbers due to Dubois and
114
Prade [4] is used. Mizumoto and Tanaka [5] also derived some algebraic properties
of fuzzy numbers.
The fuzzy numbers with continuous membership functions have been
considered as extensions of the concept of confidence interval by Kaufmann and
Gupta [6]. Instead of considering a confidence interval at a unique level, it is
considered more generally at all levels between 0 and 1. It has been also proved
that arithmetic operations on fuzzy numbers based on the extension principle can
be decomposed into operations on confidence interval at various levels from 0 to
1. Thus the problem of fuzzy field is transformed into a problem in a nonfuzzy
field without loss of information. Kaufmann and Gupta [7] provided simplified
computations for triangular fuzzy numbers. Recently, Soman and Misra [14]
proposed a new method of computing top event's fuzzy probability, which is
computationally efficient and is free from the weaknesses of previous methods. In
the present paper, we extend this approach to systems consisting of multistate
elements. Extended fuzzy numbers, resolution identity, extension principle and
fuzzy probability vectors form the basis of the methodology of this paper. Formulas
for arithmetic operation on extended fuzzy numbers with continuous membership
function have been derived using the operation on bound points of all a-level set
intervals instead of using general set operations. Based on this framework, top
event's fuzzy probability has been evaluated.
2. EXTENDED FUZZY NUMBER
Let x be a continuous real variable restricted by a possibility
distribution function !A(x) E [0,1] such that
10
II
0.0
In Kaufmann and Gupta [6], it was shown that the computational effort
involved with operations on fuzzy numbers based on the extension principle can
be considerably reduced by decomposing the membership functions into a-levels
and conducting mathematical operations directly on the confidence intervals. But
in general, it cannot be used directly for a - level calculations.
For any fuzzy number A (either a general fuzzy number or an extended
fuzzy number), which has the membership function !'A(X), an interval bounded
by two points at each a-level (0 sa s 1) can be obtained by using the a-cut
method. The symbols At)and A~) have been used in this paper to represent the
left-end- point and the right-end-point of this interval (as is shown in Fig. 1). When
A~a) and A~) are considered for all a- levels, due to monotonicity and normality
of the membership functions, we can express a general or an extended fuzzy number
A, using the following form :
A _ [A(a) A(a)] Os a s 1
L' R '
This was defined as a resolution identity by Zadeh [10]. Using the
expression for the membership function at the a-level instead of usual !'(x), the
algebraic properties of fuzzy numbers can be described more conveniently.
2.1 Arithmetic Operations on Extended Fuzzy Numbers
The arithmetic operations on two extended fuzzy numbers can be
represented more compactly using resolution identity. The properties ofthe results
can be analyzed more conveniently based on interval arithmetic [4,5]. Using the
concept of confidence interval, the following arithmetic operations, similar to those
defined for interval arithmetic, on two extended fuzzy numbers A and B can be
easily performed.
C .. A . B - [C La), C~) ]
'"' [m.·n ( A (a) B (a) A (a) B (a) A (a) B (a) A (a) B (a) )
L' L' R' L' L' R' R' R'
B (a) A (a) B (a) A (a) B (a) A (a) B (a»
max ( A (a) ]
L' L' R' L' L' R ' R' R
When A and B are positive,
C-[C~,C~]- [A~.B~,A~.B~]
Finally,C _ A + B _ [C L(a) , C (a) ] [A (a) A (a) ] X [ 1 1]
R - L' R B (a) , B (a)
R L
if 0 f1- [B ~a) , B~)]
116
0 for 0 ~ p ~ (lil
(li2- P
1 - for (l il ~ P ~ (li2
(li2 - (lil
11 (p)- 1 for (l/2 ~ P ~
fJ/2
PI
p-fJ i2
1 - for fl i2 ~
P ~
~ il
fJ il -fJ i2
0 for f:l il ~ P ~ 1
Another membership function that is suitable for representing a fuzzy
probability is shown in Fig. 3, and is given by
'·0 ------
t
rip)
1·0 p
p- p-
In the second method [12], all the Pi are assumed to be non- interactive on
An' the domain of interest. The membership functions of constituent elements are
individually specified and the joint possibility distribution is derived from the
relation:
JL P (Pl'P2 ' •, Pi' ... , po) - min [ JL 1 (P1), . , JL i (Pi) , . , I.t n (Pn) ],
( P1 ' ... , Pn) £ Il. n
These two methods will not always give the same result [9], and this raises
a question of which method to use? One condition that any self consistent method
must satisfy is that the marginal distribution must be obtainable from joint
distribution via supremum [10]. For example, the marginal of Pi is
JL i (p) - sup [ I.t P (P 1 ,···,Pi,···,Pn) ]
where the supremum is over all (PI, ... ,Pi, ... ,Pn) with Pi oeing fixed, such that
(PI' ... ' Pi'···' Pn ) £ Au· Stein [9] has shown that both these methods are not
consistent. According to him, the solution lies in directly specifying the joint
possibility distribution of (PI' P2, .. ,Pi' .. 'Pn). A class of joint possibility distribution
suitable for interactive fuzzy probabilities is a multivariate fuzzy beta distribution.
In this paper, since all the discussion is based on fuzzy beta distribution we give a
brief review of it.
JL p ( P1 ' ... , Pi ' ... , Pn ) - C (fl 1 ' ... , fl i' ... , fl n ) PI 111 •• Pi IIi .. Po lin (2)
for all ( PI' ... , Pi' ... , Pn ) E A nand fl i OJ!: 0, i - 1 , 2 , ... , n .The constant
C is chosen such that JLp is normalized (i.e., the supremum is 1). The value of C
which normalizes the fuzzy beta is given by
(3)
where the sums and products are carried over all i such that fli > O. The derivation
of above formula is given by Stein [9]. If all fli = 0, then C =1.
J.L k ( p) - C (f:l k ' f:l r ) pll. (1 - P )11 , • where f:l r - ~ f:l i . (4)
... k
'WMr
states.
O~~ 0
In~1 ~ XI OUI:UI
Nonnal output
False but safe output signal
False and dangerous output signal
The fault trees leading to system states T2 and T3 are shown in Fig. 6 and
Fig. 7, respectively.
Resistor is modeled as two state device and let its possibility distribution
corresponding to its failure state be a trapezoidal number given by
f!p3(P31) = T(.05, .1, .2, .25)
Now, we can formulate the problem of estimation of a-cuts of top event
fuzzy probability through a nonlinear mathematical optimization. Let Q~~ be the
expression obtained after substituting the left end a-cut of two state - elements'
fuzzy probabilities in Q2' The left end point of the a-cut of Q L is then obtained by
(a)
Q 2L . Q(a) ( ) (6)
- mm 2L Pu ' .... , P23
Subject to the constraints
PH + P12 + P13 = 1 (7)
P +p +p -1 (8)
21 22 23 0.5 1.0 3.5
C(.5,1,3.5)(pp) (p12) g(P13) = a (9)
C(1,I,8)(P21) (P22) (P2J = a (10)
Pij> 0.0; i=I,2. and j=I,2,3 (11)
In this formulation it is assumed that only one state of the two state elements
is appearing in the fault tree. If both states are appearing, then, it will introduce two
constraints as is the case for multistate elements.
Let us now find a-cuts of Q 2 at a=0.5. The left end point of a- cut of fuzzy
number P3 corresponding to the failure state probability of resistor ~ is 0.075.
Substituting P31 = 0.075 in the expression (5) for Q2' we obtain Q i~) as
Minimizing the value of this expression with respect to the constraints given
in (7) to (11) (with a = 0.5), we obtain Q i~S) = 0.185676. Similarly, we can obtain
Q 2R
(0.5)
= 0.723967.
Table 1 a-CUts of 02
a Q(a) Q(a)
2L 2R
Thus for each a, we find these optimal vectors for each multistate
component and substitute them in the top event probability expression along with
a-cuts of two state elements. If some of the state probabilities of a multistate
component are not appearing in expression for fault tree, they can be combined into
a single state and a reduced joint possibility distribution (in dimension) may be
used. If only one state-probability of a multistate component is appearing in the
expression for top event probability, we need not go for optimization as its marginal
distribution offers the required a-cuts.
124
Example: Consider the fault tree shown in Fig. 7 of section 4.1. The top event
probability expression is given by
(13)
Q 3R
(a) _ P (a) p (a) p (a)
(18)
llR 2lR 32R
5. CONCLUSION
Using resolution identity, extension principle, and an optimization
procedure, fuzzy probability of occurrence of the top event of a fault tree with
multistate elements can be evaluated. To reduce computational load, an
approximate method of evaluation can be effectively used. This is evident from
the illustration provided in this paper.
REFERENCES
[1] H. Tanaka, L.T. Fan, F.S. Lai and K. Toguchi, Fault Tree Analysis by Fuzzy
Probability, IEEE Trans. on Reliability, Vol. 32, pp. 455- 457, 1983.
[2] K.B. Misra and G.G. Weber, A New Method for Fuzzy Fault Tree Analysis,
Microelectronics and Reliability, Vol. 29, pp. 195-216, 1989.
[3] K.B. Misra and G.G. Weber, Use ofFuzzy Set Theory for Level-I Studies in
ProbabilisticRiskAssessment. Fuzzy Set Theory and Systems, Vol. 37,
pp. 139-160, 1990.
[4] D. Dubois and H. Prade, Operations on Fuzzy Numbers, International
Journal of System Science, Vol. 9, pp. 613-626, 1978.
[5] M. Mizumoto and K. Tanaka, Some Properties ofFuzzy Numbers, in Advan-
ces in Fuzzy Set Theory and Applications, North Holland, Amsterdam,
pp. 53-164,1979.
[6] A. Kaufmann and M.M. Gupta, Introduction to Fuzzy Arithmetic, North
Holland, Amsterdam, Holland, pp. 153-164, 1979.
125