You are on page 1of 8

 DOI 10.1515/ijnsns-2012-0050 Int. J. Nonlinear Sci. Numer. Simul.

2014; 15(3–4): 189 – 196

Bin Suo*, Chao Zeng, Yong-sheng Cheng and Jun Li

An Evidence Theory-based Algorithm for System


Reliability Evaluation under Mixed Aleatory and
Epistemic Uncertainties
Abstract: For the reliability evaluation of a complex s­econd-moment method (FORM) [1–3], second-order
system, the aleatory uncertainty and epistemic uncer- ­second-moment method (SORM) [5–6], response surface
tainty always exist at the same time because of the inade- method (RSM) [7–8], and so on. In these methods, limit
quate experimental data and the incomplete information. state function is approximated in some manner, and only
To reduce the computational cost of the reliability evalua- aleatory uncertainty, not epistemic uncertainty, can be
tion, an effective method based on evidence theory was handled. The other one is sampling approach, such as
developed. In this method, the mean of the belief measure Monte Carlo simulation (MCS) [8–9], importance sampling
and the plausibility measure was taken as the approxima- (IS) [10], stratified sampling (SS) [11–12], and other
tion of the system reliability. The discretization methods methods. MCS can get a accurate result if the sample size
for uncertain parameters were discussed in the cases of is large enough. However, for a certain standard devia-
only aleatory uncertainty was involved in the system, and tion, to improve the accuracy of 1 each digit in MCS,
mixed aleatory and epistemic uncertainties were involved, sample size needs to increase the 100 times, so the com-
respectively. Algorithms of the evaluations of belief and putational cost is very high. IS and SS are more efficient
plausibility functions were proposed for monotonic and than MCS, but these kind of methods often require supe-
non-monotonic systems. Four numerical examples under rior skills. For example, suitable importance sampling
different conditions were studied. Simulation results density function need to be selected in IS, and appropriate
showed that, the proposed method was much more ef­ stratifying strategy need to be considered in SS. So
fective than Monte Carlo method without sacrificing the these methods are not a general measure for reliability
accuracy of the resulting reliability. It was a general calculation.
method suitable for various systems with different types In reliability modeling of complicated engineering
of information. system, it is always in the situation that experimental data
is inadequate, or information is not completed, thus a
Keywords: reliability calculation, evidence theory, alea- variety of uncertainties are produced, and aleatory and
tory uncertainty, epistemic uncertainty, cost efficient epistemic uncertainties are always coexisting [13]. The
­algorithm complexity of system internal mechanism and the limita-
tion of people’s knowledge are the sources of epistemic
PACS® (2010). 02.60.Gf, 02.50.Ng uncertainty, and the inherent randomicity is the root
cause of aleatory uncertainty [14–15]. In this case, the reli-
ability calculation is more complex.
*Corresponding author: Bin Suo: Institute of Electronic Engineering,
China Academy of Engineering Physics, Mianyang 621900, China.
Evidence theory is the generalization of probability
E-mail: suo.y.y@163.com theory [16–17]. Researches show that evidence theory has
Chao Zeng, Yong-sheng Cheng, Jun Li: Institute of Electronic strong ability on handling aleatory and epistemic uncer-
Engineering, China Academy of Engineering Physics, Mianyang tainties simultaneously [18–20]. In view of the shortages
621900, China of above two approaches, the motivation of this research
work is greatly increasing the computational efficiency of
reliability calculation with evidence theory when there
1 Introduction are alertory and epistemic uncertainties in a system.

Reliability calculation is one of the most important


­problems in system reliability evaluation. There are two
approaches to calculate system reliability. One is
approximate analytical approach, such as first-order
­

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
190 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation

2 R
 eliability calculation with BelZ ( F ) ≤ PZ ( F ) ≤ PlZ ( F ) (6)

evidence theory From Eq. (6), BelZ ( F ) is the lower bound of Pf , PlZ ( F )
is the upper bound, and [ BelZ ( F ), PlZ ( F )] is the interval
Fundamentals of evidence theory include basic probabil-
­estimation of Pf . If information content about F is great
ity assignment (BPA), focal element, belief function, plau-
enough, then BelZ ( F ) will be close to PlZ ( F ) indefinitely,
sibility function, Dempster’s rule, etc., which details can
and | BelZ ( F ) - PlZ ( F )| will tend to be an infinitesimal. So
be found in references [16–17, 21], and will not be intro-
the mean of the two functions is the approximation of Pf :
duced here anymore. In this section, algorithm for reliabil-
ity calculation with evidence theory will be discussed
=Pf ( BelZ ( F ) + PlZ ( F )) / 2 (7)
when there are aleatory and epistemic uncertainties in the
system.
To calculate system reliability with Eq. (7),
X = { x 1 , x 2 ,, x n } need to be discretized to interval focal
elements first. Denote the frame of discernment of X as
2.1 Procedure of the algorithm
Q X , the focal element as c k (k ∈[1, N ]), where N is the
number of focal elements in Q X , then c k is a Cartesian
In reliability analyzing for an engineering system, limit
product
state equation can be constructed according to the de-
signed functions and requirements. In stochastic reliabil-
c1j1 × c 2j2 × cnjn
ck = (8)
ity model, if X = { x 1 , x 2 ,, x n } is the vector of n random
variables for stochastic factors, then the system limit state
Correspondingly, BPA of ck is
equation is

m1 (c1j1 ) × m2 (c 2j2 )× mn (cnjn )


m(c k ) = (9)
= ( X ) g ( x 1 , x 2 ,,=
Z g= xn ) 0 (1)

Denote= Z 0 { z : z < 0, z ∈Q Z } , then according to evidence


According to the two states hypothesis in stochastic
theory, BelZ ( F ) and PlZ ( F ) can be expressed as
reliability theory, failure surface { X | g ( X ) = 0} divides the
basic variable space into two areas. If g ( X ) > 0, then the
system is in reliable state, and safe domain is BelZ ( F ) Bel
= = X [ f ( Z 0 )]
-1
∑ m(c k ) (10)
ck ⊂ f -1 ( Z 0 )

=Ω s { X | g ( X ) > 0} (2)
PlZ ( F ) Pl=
= X [ f ( Z 0 )]
-1
∑ m(c k ) (11)
ck ∩ f -1 ( Z 0 ) ≠∅
Otherwise the system is in failure state, and its failure
domain is
Then system failure probability is obtained by Eq. (7). If it
is nessisary, sytem reliability can also be calculated with
Ω f { X | g ( X ) < 0}
= (3)
Eq. (5).

If f ( X ) is the joint probability density function of X,


then the system failure probability is
2.2 Discretization of variables

f ( X )dX Discretization of variables is the first step in reliability cal-


Pf P=
= (F) ∫Ω f
(4)
culation. For different types of parameters, the discretiza-
tion methods are different too. In this section, discretiza-
where F is the proposition F= { x ∈Ω f } . Correspondingly, tion approaches will be described in two dissimilar cases.
the reliability is

Rs = 1 - Pf (5) 2.2.1 Case 1: only alertory uncertainty in a variable

Let Q Z be the frame of discernment in failure domain, If there is only alertory uncertainty in a variable
and BelZ ( F ) and PlZ ( F ) are belief and plausibility func- x i (1 ≤ i ≤ n), then the information about xi can expressed as
tion of system failure respectively, then a probability density function (PDF) fi ( x i ).

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation 191

If x i is dependent from each other, which is defined on And its BPA is


the interval [ai , bi ], then it can be discretized into m subin-
tervals, denote = as cij [a j , b j ]( j ∈[1, m]). Then cij is a focal mij = M ij (15)
i i
element of x i , and its BPA is
Commonly, the above n subintervals have the same
bij 1
mi (cij ) = ∫ fi ( x i )dx i (12) length, that is, M ij = . Then the corresponding focal
aij n
element and BPA are

However, x i is generally infinite in reliability calcula- -1  1  1  -1  1  1 


tion, that is, x i ∈R (denote as I xi = [ -∞, +∞ ]). In this condi-
cij =
[CDF x ()   j -   ,CDF x ()   j -  ] (16)
 n  2   n  2 
tion, fi ( x i ) need to be truncated as bounded function by
j j 1
estimating an interval = I xi [l, u]( -∞ < l < u < +∞). Denote m
=i M=
i (17)
n
DI =I xi \ I xi , then the estimated interval I xi should meet
the condition
Furthermore, Stress Strength Interference (SSI) is a
common reliability model in structure engineering. Its
P ( x i ∈DI ) ≤ d (d ≥ 0) (13)
limit state equation can be expressed as

where d is determined by the calculation precision of M = g ( R, S ) = R - S = 0 (18)


system reliability. Then xi can be discretized in the interval
[l, u] with Eq. (12). where R and S are generalized strength and stress
For example, if xi obeys normal distribution ­respectively.
N ( m i ,s i ), and d = 0.03 , then I xi =- [ m i 3s i , m i + 3s i ] means As shown in Fig. 1, the system failure probability is
P( x i ∈DI ) ≤ 0.03. So the domain of xi can be truncated as only related with SSI area. So the improved discretization
[ m i - 3s i , m i + 3s i ]. If one wants to improve the calculation method is discretizing R and S only in their SSI area.
accuracy, a wider interval range, such as 4s or 5s interval, As an example, the process of discretizing R will be
can be selected for xi. described to illustrate the improved method (hereinafter
referred to as ‘IADM’). The mean of R is denoted as
m R ∈[ m Rl , m Ru ], and standard deviation as s R ∈[s Rl ,s Ru ]. The
mean of S is denoted as m S ∈[ m Sl , m Su ], and standard devia-
2.2.2 Case 2: both alertory and epistemic uncertainties
tion as s S ∈[s Sl ,s Su ]. If R and S are truncated into the 3s
in a variable
intervals, then the domains of R and S are

If there are both alertory and epistemic uncertainties in a


variable, then its distribution pattern expresses the uncer- DR =-
[ m Rl 3s Ru , m Ru + 3s Ru ] , DS =-
[ m Sl 3s Su , m Su + 3s Su ] (19)
tainty about alertory part, and its distribution parameters,
which always described as interval numbers, express the Denote the SSI area as FR - S , then
uncertainty about epistemic part.
Let the cumulative distribution function (CDF) of xi to FR - S = DR ∩ DS = [ F l , F u ] = [ m Sl - 3s Su , m Ru + 3s Ru ] (20)
be CDFi ( x i ). Its mean is m i ∈[ m l , m iu ] and standard devia-
i
tion is s i ∈[s il ,s iu ]. Denote the lower CDF of xi as CDF i ( x i ),
If the corresponding lower probability of F l in Fig. 1 is
and upper CDF as CDFi ( x i ), then xi can be discretized by
pRstart , and upper probability of F u in Fig. 1 is pRend ( pRend = 1),
averaging discretization method (ADM) [22].
then [ pRstart , pRend ] can be divided into m equal subintervals.
The [0, 1] ordinate intervals of CDF i ( x i ) and CDFi ( x i )
BPA of each subinterval is
are both discretized into n subintervals of length
(
M ij > 0 j = )
1,2, , n , and M i0 = 0. Then the j-th focal
bpa=
_ R ( pRend - pRstart ) / m (21)
element of xi is the subinterval

With IADM, smaller interval can obtained with the


 same subinterval numbers m, thus a more accurate result
-1  j - 1 Mj  -1 
j -1
M j 
CDF  ∑ M s +
cij =  ,CDF  ∑ M s +  (14) about reliability can be getten with the same computa-
  s 0= 2  s 0 2 
=
tional cost. The process of discretizing S is similar.

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
192 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation

Fig. 1: SSI model. PDFs of R and S, vertical axis.

2.3 E
 valuation of belief and plausibility BelZ ( F ) = ∑ m(c k ) (28)
function k ⊂ N _ BelZ

This section will discuss how to calculate BelZ ( F ) and PlZ ( F ) = ∑ m(c k ) (29)
k ⊂ N _ PlZ
PlZ ( F ) in Eq. (7).
If Z = g(X) is monotonic, then vertex method [23] can
The procedure to estimate N _ BelZ and N _ PlZ using
be used. Denote the mapping of ck in QZ as dk, then
sampling method is as follows:
–– Step 1. Randomly generating Ns (Ns → ∞) X in
dk g=
= ( c k ) [ lk , u k ] (22)
c k (c k ⊂ Q X ), and calculate Z with Eq. (1). If Z < 0 hold
for any Z, then k ∈ N_BelZ and k ∈ N_PlZ. If Z < 0 hold
Each ck is a n-dimensional box, and its vertices are labeled
for some Z, and Z ≥ 0 hold for the others, then k ∈ N_
as vj ( j = 1, 2, . . . , 2n). As Z = g(X) is continuous in ck and
PlZ.
there are no extreme points exist in this region except for
–– Step 2. For 1 ≤ k ≤ N , repeate Step 1.
its boundaries, then

Then BelZ ( F ) and PlZ ( F ) can be calculated with Eq.


dk g=
= (c k ) [min{ g ( v=
j ) : j 1,, n },max{ g ( v=
j ) : j 1,, n }]
j j (28) and (29).
(23)

By Eq. (23), all dk (k ∈ mn) in QZ can be calculated. Then


3 Numerical examples
BelZ ( F )
= ∑ m(di )
= ∑ m(dk ) (24) In this section, four numerical examples are given to vali-
dk ⊆ Z 0 sup( dk ) < 0
date the effectiveness of proposed algorithm. The first one
only includes aleatory uncertainty. The second and the
PlZ ( F )
= ∑ m(di )
= ∑ m(dk ) (25)
third include both aleatory and epistemic uncertainties,
dk ∩ Z 0 ≠∅ inf( dk ) < 0
and the former is a general case, and the latter is a special
case about SSI model. The limit state functions in first
If Z = g(X) is non-monotonic, then vertex method is
three examples are all monotonic, and is non-monotonic
unusabel because of large computational errors. In this
in Example 4.
case, sampling method can be used to calculate BelZ ( F )
and PlZ ( F ) . Denote
Example 1: Limit state equation of a system is

N _=
BelZ {k : c k ⊂ f -1 ( Z 0 )} (26)
g ( X ) = x 13 + x 23 - 4 = 0 (30)
N _=
PlZ {k : c k ∩ f -1 ( Z
0 ) ≠ ∅} (27)
where x1 and x2 independently obey normal distribution
Then Eqs. (10) and (11) can be expressed as N(3, 1) and N(2.9, 1) respectively.

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation 193

Table 1: BPAs of x1 and x2 Table 3: Results of Example 1

x1 m1(Ai) x2 m2(Bi) Method Bel(F ) Pl(F ) Pf Rs

A1 = [0.0,0.6] 0.0068 B1 = [-0.1,0.5] 0.0068 Proposed method


A2 = [0.6,1.2] 0.0277 B2 = [0.5,1.1] 0.0277 m = 10 0.0012 0.0129 0.0071 0.9929
A3 = [1.2,1.8] 0.0791 B3 = [1.1,1.7] 0.0791 m = 20 0.0022 0.0076 0.0049 0.9951
A4 = [1.8,2.4] 0.1592 B4 = [1.7,2.3] 0.1592 m = 30 0.0026 0.0059 0.0042 0.9958
A5 = [2.4,3.0] 0.2257 B5 = [2.3,2.9] 0.2257 Monte Carlo (106 samples) – – 0.0041 0.9959
A6 = [3.0,3.6] 0.2257 B6 = [2.9,3.5] 0.2257 FORM – – 0.0081 0.9919
A7 = [3.6,4.2] 0.1592 B7 = [3.5,4.1] 0.1592 SORM – – 0.0038 0.9962
A8 = [4.2,4.8] 0.0791 B8 = [4.1,4.7] 0.0791
A9 = [4.8,5.4] 0.0277 B9 = [4.7,5.3] 0.0277
A10 = [5.4,6.0] 0.0068 B10 = [5.3,5.9] 0.0068 Table 4: Focal elements of x1 and x2

k Focal element
Table 2: Focal elements and BPAs correlating with failure probability
in ΘZ x1 x2

1 [-0.8691,4.4206] [-2.5140,3.7757]
k dk m(dk ) 2 [2.7814,6.8543] [1.7449,6.8178]
3 [4.9531,8.3020] [4.2786,8.6276]
1 [-4.0010, -3.6590] 0.00005
4 [6.6881,9.4587] [6.3027,10.0734]
2 [-3.8750, -2.4530] 0.0002
5 [8.2460,10.4973] [8.1204,11.3717]
3 [-2.6690, 1.1290] 0.0005
6 [9.5026,11.7540] [9.6283,12.8796]
4 [-3.7850, -2.1470] 0.0002
7 [10.5413,13.3119] [10.9266,14.6972]
5 [-3.6590, -0.9410] 0.0008
8 [11.6980,15.0469] [12.3724,16.7214]
6 [-2.4530, 2.6410] 0.0022
9 [13.1457,17.2186] [14.1821,19.2550]
7 [-2.2730, 1.9570] 0.0005
10 [15.5794,20.8691] [17.2243,23.5139]
8 [-2.1470, 3.1630] 0.0022
9 [-0.9410, 6.7450] 0.0063

Carlo method with 106. These mean that the proposed


Truncating the domains of x1 and x2 as [ m - 3s , m + 3s ], method is much more effective than Monte Carlo
and let m = 10, then BPAs of the two variables are calcu- method without sacrificing the accuracy of resulting
lated by Eq. (12). The results are shown in Table 1. ­measurements.
There are mn = 100 focal elements in Q X . Vertex
method is used to calculate the mapping of ck in ΘZ. The Example 2: The limit state function of a system is
results are shown in Table 2. Only the focal elements meet
the condition inf(dk ) < 0 are listed in the table as the others g ( X ) = x 13 + x 23 - 18 = 0 (31)
are irrespective with the failure probability.
According to Table 2, Eqs. (10) and (11), it is ob- where x1 and x2 independently obey normal distribution,
tained that BelZ ( F ) = 0.0012 , and PlZ ( F ) = 0.0129. So which means are m 1 = [9, 11] and m 2 = [9, 12], and standard
0.0012 ≤ Pf ≤ 0.0129. From Eq. (7), Pf = 0.0071. deviation are s 1 = [4, 6] and s 2 = [5, 7], respectively.
Obviously, these results are imprecise. Increases the Firstly, ADM is used to discretize x1 and x2 with m = 10.
number of focal elements, the result will be more accu- The results are shown in Table 4. BPA of each focal element
rate. Let m be 20 and 30 respectively, and calculate the is 1/10.
system failure probability and reliability with above Then with Eqs. (23)–(25), the lower and upper bounds
method. The results are compared with that of Monte of Pf are Bel( F ) = 0 and Pl( F ) = 0.03 respectively. So the
Carlo method, FORM and SORM, which are shown in ­approximate value of Pf is
Table 3.
If the results of Monte Carlo method can be taken as Pf =( Bel( F ) + Pl( F )) / 2 =0.015
the exact value, then it can be seen from Table 3 that the
results of proposed method is very close to the exact value Increasing m to 100, 200, and 300 gradually to calcu-
when m increases to 30. The number of simulations for late the system reliability. The results are listed in Table 5.
proposed method is only (30 + 1) × (30 + 1) = 961 when As indicated in Table 5, when m increase to 200, the
m = 30. It is far less than the simulation number of Monte calculation result of reliability by proposed method is

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
194 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation

Table 5: Results of Example 2 Example 4: The reliability of the dike revetment depends
in part on the density (D) and thickness (D) of revetment,
Method Lower bound Upper bound Pf Rs and its angle of inclination (a). The other factors relate to
of Pf of Pf the reliability of structure are the significant wave height
Proposed method (H), the offshore peak wave steepness (s), and the model
m = 10 0 0.0300 0.0150 0.9850 parameter (M, which is a factor introduced to represent
m = 100 0 0.0381 0.0191 0.9809 the analyst’s uncertainty about the model itself). The limit
m = 200 0.00005 0.0389 0.0195 0.9805
state function of the dike was given by Hussarts [24] and
m = 300 0.0001 0.0390 0.0195 0.9805
Monte Carlo method 0.0002 0.0386 0.0194 0.9806 modified by Hall [25] as
(4 × 106 samples)
H tan(a )
Z =DD - =0 (32)
cos(a ) M s
Table 6: Results of Example 3
If Z > 0, then the dike is inferred to be reliable. Otherwise,
Method Lower bound Upper bound P̃f the structure is not sufficiently strong to withstand the
of Pf of Pf
stresses from the sea.
Proposed method (ADM) The information about the six parameters in Eq. (32) is
m = 10 0 0.0200 0.0100 as follows:
m = 50 0.0004 0.0328 0.0166
m = 100 0.0009 0.0335 0.0172
m = 200 0.0011 0.0338 0.0175 D∈[1.70, 1.78], D∈[0.70, 0.75] meters,
m = 300 0.0011 0.0340 0.0176 a ∈ arctan([0.31, 0.33]) =
[0.302, 0.320] radians
Proposed method (IADM)
m = 10 0 0.0274 0.0137 M ∈[3.0, 5.1] , H ~ Weibull=
( m 0,=
m [10, 12],
= s [1.1, 1.4])
m = 50 0.0008 0.0330 0.0169
m = 100 0.0011 0.0337 0.0174= s ~ N ( m [0.038,
= 0.040], s [0.004, 0.006])
m = 200 0.0012 0.0340 0.0176
m = 300 0.0012 0.0342 0.0177
Monte Carlo method 0.0014 0.0340 0.0177 D, D, M ,a can be expressed as BPA structure in evi-
(4 × 106 samples) dence theory directly, as shown in Table 7.
H and s are discretized by ADM with m = 20, and the
focal elements are shown in Table 8 and Table 9. BPA of
each focal element is 0.05.
agree with that by Monte Carlo method. The simulation Sampling method described in Section 2.3 is used to
number of proposed method is 4.04 × 104, and 4 × 106 for calculate N_BelZ and N_PlZ. There is no focal element in QX
Monte Carlo method. meet the condition in Eq. (26), so the set of N_BelZ is
Comparing this example with example 1, it can be empty. Elements in N_PlZ are listed in Table 10, in which
seen that if system has aleatory and epistemic uncertain- “(i, j)” is the index of the corresponding focal elements in
ties simultaneously, its computational cost is much higher H and s respectively, and “k” is the index of focal element
than that of the system with only aleoteory uncertainty. in QX. Then lower and upper bounds of dike reliability are
Bel(Z < 0) = 0, Pl(Z < 0) = 0.0608. So the approximation of
Example 3: In a mechanical structure, the generalized failure probability is 0.0304, and reliability is 0.9696.
strength R and stress S independently obey normal distri-
bution. Nominal mean, mR, of R is 550 MPa, and largest
deviation is 17.2 MPa. Its standard deviation, sR, is 44.8
MPa, and largest deviation of sR is 1.34 MPa. For S, these Table 7: Focal elements and BPAs of D, D, M, and a
values are 420 MPa, 10.8 MPa, 30.6 MPa, and 0.92 MPa,
respectively. Variable Focal element BPA

IADM is used to discretize R and S in their SSI area. D [1.70, 1.78] 1


The results of failure probability and reliability are listed D [0.70, 0.75] 1
in Table 6. Comparing with ADM, proposed method with M [0.302, 0.320] 1
[3.0, 5.1] 1
IADM is more effective.
a

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation 195

Table 8: Focal elements of H Table 10: Elements in set N_PlZ

i Focal element k (i, j )

1 [0.7616,1.0306] 141 (8, 1)


2 [0.8523,1.1318] 161 (9, 1)
3 [0.8994,1.1837] 181 (10, 1)
4 [0.9328,1.2203] 201 (11, 1)
5 [0.9595,1.2493] 221 (12, 1)
6 [0.9820,1.2737] 241 (13, 1)
7 [1.0019,1.2952] 261 (14, 1)
8 [1.0200,1.3146] 281 (15, 1)
9 [1.0368,1.3326] 282 (15, 2)
10 [1.0527,1.3497] 301 (16, 1)
11 [1.0680,1.3660] 302 (16, 2)
12 [1.0830,1.3819] 321 (17, 1)
13 [1.0979,1.3977] 322 (17, 2)
14 [1.1108,1.4165] 341 (18, 1)
15 [1.1237,1.4362] 342 (18, 2)
16 [1.1373,1.4571] 343 (18, 3)
17 [1.1521,1.4800] 361 (19, 1)
18 [1.1692,1.5063] 362 (19, 2)
19 [1.1908,1.5398] 363 (19, 3)
20 [1.2264,1.5952] 364 (19, 4)
381 (20, 1)
382 (20, 2)
383 (20, 3)
Table 9: Focal elements of s 384 (20, 4)
385 (20, 5)
j Focal element

1 [0.0162,0.0302]
2 [0.0194,0.0322]
3 [0.0211,0.0334] been widely used in data fusion, pattern recognition, deci-
4 [0.0224,0.0343] sion analysis, and other fields. In this paper, a new appli-
5 [0.0235,0.0350] cation of evidence theory for reliability calculation is
6 [0.0244,0.0356] ­developed when system has aleatory and epistemic uncer-
7 [0.0253,0.0362]
tainties. The results of numerical examples show that, the
8 [0.0261,0.0367]
9 [0.0269,0.0372]
computational efficiency of proposed method is far in
10 [0.0276,0.0377] excess of Monte Carlo method without decreasing the
11 [0.0403,0.0404] degree of accuracy. For complicated engineering system,
12 [0.0408,0.0411] the proposed method has strong practical value. And
13 [0.0413,0.0419] more, the proposed method has extensive applicability. It
14 [0.0418,0.0427]
is applicable and effective when information types about
15 [0.0424,0.0436]
16 [0.0430,0.0445] the uncertain parameters in a system are interval numbers,
17 [0.0437,0.0456] probability distributions, probability boxes, experts’ opin-
18 [0.0446,0.0469] ions, and so on.
19 [0.0458,0.0486] In the proposed method, when system limit state
20 [0.0478,0.0518]
function is non-monotonic, the evaluation of belief and
plausibility functions with sampling method has high
computational complexity. So in future work, more effec-
tive algorithm need to be studied. Maybe intelligent opti-
4 Conclusion mization algorithm is a feasible approach.

As evidence theory has a flexible framework for represent- Funding: This work were supported by Advanced Research
ing uncertain information, in which both aleatory and Foundation of the 12th Five-year National Defense of
epistemic uncertainties can be modeled together, it has China (grant no. 426010401) and China Academy of

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM
196 B. Suo et al., An Evidence Theory-based Algorithm for System Reliability Evaluation

­ngineering Physics Foundation for Development of


E [12] Zhang, F., Lu, Z. Z., Cui, L. J. and Song, S. F., Reliability
Science and Technology (Grant no. 2012B0403058). sensitivity algorithm based on stratified importance sampling
method for multiple failure modes systems, Chinese J. of
Aeronautics. 23 (2010), 660–669.
Received: March 12, 2012. Accepted: February 4, 2014.
[13] Yang, X. H., She, D. X. and Yang, Z. F., et al., Chaotic bayesian
method based on multiple criteria decision making (MCDM) for
forecasting nonlinear hydrological time series, Int. J. of
References Nonlinear Sciences and Numerical Simulation. 10 (2009),
1595–1610.
[14] El Naschie, M. S., Probability set particles, Int. J. of Nonlinear
[1] Abdallah, I. H. M., Waleed, F. H. and Fayez, A. A., Uncertainty
Sciences and Numerical Simulation. 8 (2007), 117–120.
and reliability analysis applied to slope stability, Str. Safety.
[15] Oberkampf, W. L., DeLand, S. M., Rutherford, B. M., Diegert,
22 (2000), 161–187.
K. V. and Alvin, K. F., Error and uncertainty in modeling and
[2] Deng, J., Gu, D. S., Li, X. B. and Yue, Z. Q., Structural reliability
simulation, Reliability Engineering & System Safety. 75 (2002),
analysis for implicit performance functions using artificial
333–357.
neural network, Str. Safety. 27 (2005), 25–48.
[16] Dempster, A. P., Upper and lower probabilities induced by a
[3] Xiao, X. B., Wen, Z. F., Jin, X. S. and Sheng, X. Z., Effects of track
multi valued mapping, Annals Math Statist. 38 (1967),
support failures on dynamic response of high speed tracks, Int.
325–339.
J. of Nonlinear Sciences and Numerical Simulation. 8 (2007),
[17] Shafer, G. A., Mathematical Theory of Evidence, Princeton
615–630.
University Press, 1976.
[4] Lin, J. Z., Chan, T. L., Liu, S., Zhou, K., Zhou, Y. and Lee, S. C.,
[18] Yang, C. D., A scientific realization and verification of Yin-Yang
Effects of coherent structures on nanoparticle coagulation and
theory: complex-valued mechanics, Int. J. of Nonlinear Sciences
dispersion in a round jet, Int. J. of Nonlinear Sciences and
and Numerical Simulation. 11 (2010), 135–156.
Numerical Simulation. 8 (2007), 45–54.
[19] Helton, J. C., Johnson, J. D. and Oberkampf, W. L., An
[5] Carbillet, S., Richard, F. and Boubakar, L., Reliability indicator
exploration of alternative approaches to the representation
for layered composites with strongly non-linear behavior,
of uncertainty in model predictions, Reliability Engineering &
Composites Science & Technology. 69 (2009), 81–87.
System Safety. 85 (2004), 39–71.
[6] Mealier, F. D. N., Guillaumat, L. and Arnoux, P., Reliability
[20] Bae, H. R., Grandhi, R. V. and Canfield, R. A., Epistemic
assessment of locking systems. Probabilistic Engineering
uncertainty quantification techniques including evidence
Mechanics, 25 (2010), 67–74.
theory for large-scale structures, Computers and Structures.
[7] Liu, Y., Yuan, J. M., Dong, L., et al., Radiation characteristics
82 (2004), 1101–1112.
of a reflector antenna under shock wave loading, Int. J. of
[21] Sentz, K. and Ferson, S., Combination of evidence in
Nonlinear Sciences and Numerical Simulation. 13 (2012),
Dempster-Shafer theory, Sandia National Laboratories,
75–79.
Albuquerque, SAND2002-0835, 2002.
[8] Kang, S. C., Koh, H. M. and Choo, J. F., An efficient response
[22] Tonon, F., Using random set theory to propagate epistemic
surface method using moving least squares approximation for
uncertainty through a mechanical system, Reliability
structural reliability analysis, Probabilistic Engineering
Engineering & System Safety. 85 (2004), 169–181.
Mechanics. 25 (2010), 365–371.
[23] Dong, W. M. and Shah, H. C., Vertex method for computing
[9] Cheng, Q., Luo, Z. X. and Zhou, H. C., Numerical simulation on
functions of fuzzy variable, Fuzzy Sets & Systems. 24 (1987),
radiative heat transfer through a two-dimensional rectangular
65–78.
domain with inhomogeneous, absorbing and isotropic/
[24] Hussaarts, M., Vrijling, J. K., De Looff, H. and Blonk, C., The
anisotropic scattering media exposed to collimated irradiation,
probabilistic optimization of revetment on the dikes along the
Int. J. of Nonlinear Sciences and Numerical Simulation. 11
Frisian coast, Proceedings of Coastal Structures. ASCE,
(2010), 927–945.
Balkema, Rotterdam, 2000, pp. 325–329.
[10] Luca, C. and Paolo, V., Monte Carlo importance sampling
[25] Hall, J. and Lawry, J., Imprecise probabilities of engineering
optimization for system reliability applications, Annals of
system failure from random and fuzzy set reliability analysis,
Nuclear Energy. 31 (2004), 1005–1025.
Proceedings of the Second International Symposium on
[11] Guido, W. I. and Tony, L., Efficient estimation and stratified
Imprecise Probabilities and Their Applications, Shaker
sampling, J. of Econometrics. 74 (1996), 289–318.
Publishing BV, Maastricht, 2001, pp. 195–204.

Brought to you by | Carleton University OCUL


Authenticated
Download Date | 6/13/15 2:30 PM

You might also like