Professional Documents
Culture Documents
k k is a
reliability index which geometrically represents a
minimum distance from the origin to the limit-state
surface in u space. Then the failure probability can be
computed approximately:
p
f
1 R % / b 7
Due to its simplicity and efciency, the HLRF
algorithm is often used to solve the optimization in
Eq. 6:
b
k
G U
k
_ _
rG U
k
_ _ _ _
T
U
k
rG U
k
_ _ _
_
_
_
U
k1
b
k
rG U
k
_ _
rG U
k
_ _ _
_
_
_
8
where rG denotes the gradient vector of G with
respect to the random variables U.
3 Hybrid reliability analysis with both random
and interval variables
In many practical engineering problems, the precise
probability distributions can not been obtained for
some uncertain parameters due to limited information,
and generally we can only achieve their variation
intervals. When both random and interval uncertain-
ties exist in a structure, the random variables can be
treated by probability distributions while the interval
variables can be treated by their bounds. Then we can
A new reliability analysis method 171
123
create a limit-state function g X; Y , where Y repre-
sent an m-dimensional interval vector:
Y 2 Y
L
; Y
R
_
; Y
i
2 Y
L
i
; Y
R
i
_
; i 1; 2; . . .; m 9
where the superscripts L and R represent the lower and
upper bounds of an interval, respectively. Thus the
failure probability can be computed by:
p
f
Pr g X; Y \0 f g 10
Due to the interval variables Y, generally the failure
probability of the above problem will belong to a
following interval instead of a single value:
p
f
2 p
L
f
; p
R
f
_ _
11
where the lower bound p
f
L
and upper bound p
f
R
can be
given by (Du 2007):
p
L
f
Pr max
Y
g X; Y \0
_ _
12
p
R
f
Pr min
Y
g X; Y \0
_ _
13
In the above two equations, max
Y
g X; Y and
min
Y
g X; Y represent the minimum and maximum
values of the limit-state function over the interval
uncertainty of Y. Geometrically, they represent two
bound limit-state surfaces which have the nearest and
furthest distances to the origin among all the limit-
state functions with different values of Y, as shown in
Fig. 1 for a two-dimensional problem. For a multi-
dimensional problem, we will also obtain a strip for
the limit state, except that the two bounds will be
hyper-surfaces rather than the two-dimensional
curves. And similarly, the interval variables Y will
also have different values on these two hyper-surface
bounds.
Through the above analysis, we know that the
failure probability of a structure with both random and
interval uncertainties has a variation interval. Never-
theless, the upper bound of this interval, namely
maximal failure probability is practically our most
concern, as a strict reliability requirement can be just
satised only after the worst case is considered.
Therefore, in this paper the maximum failure proba-
bility p
f
R
is used to represent the reliability degree of a
structure with the mixed uncertainties, in other words,
Eq. 13 is only required to be solved to obtain a worst-
case reliability over the ranges of interval variables.
By using the FORM to solve Eq. 13, the optimiza-
tion in Eq. 6 becomes:
b
L
min
U
U k k
s:t: min
Y2Y
I
G U; Y 0
_
_
_
14
where b
L
represents the minimal reliability index,
based on which the maximum failure probability p
R
f
/ b
L
_ _
can be approximately obtained. Equation 14
is a two-layer nesting optimization, which can be
decomposed into:
Outer-layer optimization for probability reliability
analysis
b
L
min
U
U k k
s:t: G U; Y
0
_
15
Inner-layer optimization for interval analysis
G U; Y
min
Y
G U; Y
s:t: Y
L
i
Y
i
Y
R
i
; i 1; 2; . . .; m
_
16
Obviously, Eq. 14 is a very complex optimization
problem. If we directly solve this problem by using the
existing nonlinear programming methods such as SQP
and SLP, the nesting optimization composed by
Eqs. 15 and 16 will be caused inevitably. In most
practical applications, the time-consuming simulation
models are involved and the nesting optimization
based on an actual simulation model will lead to a
very heavy computational burden. Therefore, the
0
1
X
2
X
( ) max , 0 g =
Y
X Y
( ) min , 0 g =
Y
X Y
Fig. 1 Bound limit-state surfaces caused by the intervals Y
172 C. Jiang et al.
123
optimization efciency is hard to satisfy the engineer-
ing requirements.
Recently, a kind of decoupling strategy has been
developed to solve the above problem (Du 2007; Guo
and Du 2009; Luo et al. 2009), and its owchart is
illustrated in Fig. 2. It can be found that an efcient
sequential single-loop (SSL) procedure has been
created, in which the probability analysis in terms of
randomvariables has been decoupled fromthe interval
analysis in terms of interval variables. The outer-layer
and inner-layer optimizations have relatively inde-
pendent iterations, which also provide intermediate
values of X or Y for each other in each step. Through
the above treatment, the computational efciency of
the nesting optimization can be greatly improved.
However, a rigorous mathematical foundation is in
absence for this decoupling strategy, and it seems
difcult to ensure a robust convergence performance
for this method. Especially for some engineering
problems with high dimensions and nonlinearity, it
seems a challenging job to make the above SSL
procedure converge to a stable result.
4 A new method to solve the hybrid reliability
4.1 Creation of the equivalent model
In our problem, there exist two types of uncertain
variables, namely random variables X and interval
variables Y. Here, a new problem is rstly created, in
which the random variables X are maintained while
the interval variables Y are all changed to independent
randomvariables following uniform distributions over
their respective intervals:
Y
i
$u Y
L
i
; Y
R
i
_ _
; i 1; 2; . . .; m 17
Thus a traditional reliability problem with only
random variables is obtained, and the number of the
random variables becomes m ? n. Here, for ease of
expression a same symbol Y is used to represent the
two different types of variables in the original and
newly created problems.
Also using the FORM to solve the above problem,
an optimization problem like Eq. 6 can be created:
b min
U; V
U k k
2
V k k
2
_
s:t: G
0
U; V 0
_
_
_
18
where Uand Vrepresent the standard normal variables
transformed by X and Y, respectively. G
0
denotes the
transformed limit-state function in uv space.
4.2 Equivalence and its proof
In this paper, the above two reliability analysis models
are called original model and equivalent model,
respectively. In practice, between these two problems,
there exists a following equivalence:
Equivalence Only by changing the interval variables
to corresponding uniform distributions, a hybrid
reliability problem with random variables X and
interval variables Y can be changed to a conventional
reliability problem with only random variables. By
using the FORM to compute the reliability of these two
problems, mathematically they will have a same
solution of X and Y.
Proof Now, the equivalent problem created in Sect.
4.1 is rstly considered. By solving the problem in
Eq. 18, an MPP U
; V
; Y
Y
U Y
Y Y Y
Interval analysis
Probability analysis
( ) ( )
( )
( ) ( )
( )
( ) ( )
( ) ( )
( )
( ) ( )
( )
1
1
1
1
1
,
,
,
k
k k
k k
k k
k k
k k
G
G
G
= +
U
U Y
U Y
U
U Y
( ) k
Y
( ) k
U
Converge?
Yes
k=k+1
No
Initial U
and Y
k=1
Fig. 2 SSL procedure for
computation of p
f
R
(Guo and
Du 2009)
A new reliability analysis method 173
123
(Breitung 1994; Madsen et al. 1986; Madsen 1985),
and hence X
; Y
_
21
Equation 20 is a constraint optimization problem, and
based on the KarushKuhnTucker necessary condi-
tion (Nocedal and Wright 1999) its optimum X
; Y
must satisfy:
r
X
i
f
X
X k
1
r
X
i
g X; Y 0; i 1; 2; . . .; n
k
1
r
Y
i
g X; Y k
3i
k
2i
0; i 1; 2; . . .; m
k
2i
Y
L
i
Y
i
_ _
0; i 1; 2; . . .; m
k
3i
Y
i
Y
R
i
_ _
0; i 1; 2; . . .; m
g X; Y 0
Y
L
i
Y Y
R
i
; i 1; 2; . . .; m
_
_
22
where k
1
, k
2i
and k
3i
denote the Lagrange multipliers.
Secondly, the original hybrid reliability problem is
considered. As analyzed above, to compute the
maximum failure probability through the FORM, an
optimization problem in Eq. 14 needs to be solved to
obtain an MPP U* and corresponding X*. Also based
on the property of the MPP, X* can be sure to be an
optimum of the following problem:
max
X
f
X
X
s.t. min
Y2Y
I
g X; Y 0
_
23
where Yare interval variables. In the above equation, a
sub-optimization problem min
Y2Y
I
g X; Y in terms of
interval variables Y exists in the constraint, and its
KarushKuhnTucker necessary condition can be
given:
r
Y
i
g X; Y k
3i
k
2i
0; i 1; 2; . . .; m
k
2i
Y
L
i
Y
i
_ _
0; i 1; 2; . . .; m
k
3i
Y
i
Y
R
i
_ _
0; i 1; 2; . . .; m
Y
L
i
Y Y
R
i
; i 1; 2; . . .; m
_
_
24
Substituting Eq. 24 into Eq. 23 yields:
max
X
f
X
X
s.t. g X; Y 0
r
Y
i
g X; Y k
3i
k
2i
0; i 1; 2; . . .; m
k
2i
Y
L
i
Y
i
_ _
0; i 1; 2; . . .; m
k
3i
Y
i
Y
R
i
_ _
0; i 1; 2; . . .; m
Y
L
i
Y Y
R
i
; i 1; 2; . . .; m
_
_
25
Then the KarushKuhnTucker necessary condition
of Eq. 25 can be further obtained, which can be found to
have an exactly same expression as Eq. 22. Therefore, it
has been proved that mathematically the original prob-
lem and the equivalent problem have a same solution
when using the FORM to compute their reliability.
4.3 Computational procedure
Based on the above equivalence analysis, a new
method can then be constructed to deal with the
uncertain structures with both random variables and
interval variables, in which the maximum failure
probability is evaluated through solving an equivalent
model rather than the original problem. As we can see,
a complex nesting optimization will be involved for
the original model, and the existing techniques will
encounter severe troubles in computational efciency
and convergence performance when solving such a
nesting optimization. Nevertheless, the equivalent
model is a conventional reliability problem with only
random variables, which can be solved by many well-
established algorithms with high efciency and robust
convergence performance.
After solving the equivalent model in Eq. 18, an
optima U
; V
and corresponding X
; Y
in
original space can be obtained. Considering that this
174 C. Jiang et al.
123
X
; Y
; V
can be
obtained after solving this problem.
(4) Based on the U
; V
, a minimal reliability
index b
L
U
rG
0
Z
k
_ _
Z
k
_ _
T
G
0
Z
k
_ _
rG
0
Z
k
_ _ _
_
_
_
2
rG
0
Z
k
_ _
Z
k
27
where rG
0
represents a gradient vector with respect to
the variables Z.
In Eq. 26, a denotes a step size which can be
determined by minimizingthe following merit function:
m Z
1
2
Z k k c G
0
Z
28
where the constant c should satisfy
c [
Z k k
rG
0
Z k k
29
Here, an inexact one-dimensional search technique
(Yuan and Sun 2005) is employed to solve the merit
function for computational reduction. The above
iteration will be repeated until an optimum is reached.
5 Numerical examples and discussion
The present method will be rstly applied to two
simple problems, and subsequently to two relatively
complex problems with more variables. In each
numerical example, the sequential single-loop proce-
dure (Du 2007; Guo and Du 2009) (termed as SSL
method) will be also used, and the analysis results
from the two different methods under exactly the same
convergence criteria will be contrasted.
5.1 A cantilever beam
A cantilever beam modied from the numerical
example in reference (Liang et al. 2004) is considered,
as shown in Fig. 3. The Youngs Modulus E of the
material is 2:10 10
5
MPa. Subjected to a horizontal
force P
x
and a vertical force P
y
, the maximum stress at
the xed end of the beam should be less than a yield
strength S = 320 MPa. The limit-state function can be
created as follows:
g b; h; P
x
; P
y
_ _
S
6P
x
L
b
2
h
6P
y
L
bh
2
30
where b and h denote the width and height of the cross
section, respectively; L denotes the length of the beam.
In this problem, b, h and L are treated as random
variables, while P
x
and P
y
interval variables. All the
uncertain variables are given in Table 1.
L
h
b
x
P
y
P
Fig. 3 A cantilever beam (Liang et al. 2004)
A new reliability analysis method 175
123
The reliability analysis results are listed in Table 2.
It can be found that the present method and the SSL
method can both converge for this problem. Very close
results are obtained for these two methods, as their
largest deviation of the variables is only 7.11% which
occurs in the vertical force P
y
. It can be also found that
the interval variables have relatively larger deviations
than the random variables between the results of these
two methods. It is because that here the actual values
of the two interval variables just locate on their
bounds. In our method, the interval variables need to
be changed to uniform distributions, and their bounds
will lead to plus innity or minus innity when
transformed to the normal distributions by using the
FORM. Therefore, it seems difcult to numerically
obtain precise values for interval variables under this
case. It is just the reason why the SSL method gives a
slight better b
L
than the present method, as a smaller
value has been observed for its b
L
. However, these
deviations are relatively small, which can be com-
pletely allowed by practical engineering requirements.
On the other hand, the present method needs a total of
128 functional evaluations for the reliability analysis
while the SSL method needs 198 evaluations, which
indicates that our method has relatively higher
efciency.
5.2 A cantilever tube
A cantilever tube modied from the numerical exam-
ple in reference (Du 2008) is considered, as shown in
Fig. 4. Three external forces F
1
, F
2
and P, and a
torsion T are applied to the tube, and the maximumvon
Mises stress r
max
on the top surface of the tube at the
origin should be less than a yield strength S
y
. Then a
following limit-state function can be created:
g X S
y
r
max
31
where r
max
can be computed by:
r
max
r
2
x
3s
2
zx
_
r
x
P F
1
sin h
1
F
2
sin h
2
A
M
I
32
where r
x
denotes the normal stress. The area A,
bending moment M, moment of inertia I and torsional
stress s
zx
can be given by:
A
p
4
d
2
d 2t
2
_ _
M F
1
L
1
cos h
1
F
2
L
2
cos h
2
I
p
64
d
4
d 2t
4
_ _
s
zx
Td
4I
33
Table 1 Uncertain variables for the cantilever beam
Random
variable
Distribution
parameter 1
Distribution
parameter 2
Variable
type
b (mm) 100 15 Normal
h (mm) 200 20 Normal
L (mm) 1000 100 Normal
P
x
(N) 47000 53000 Interval
P
y
(N) 23000 27000 Interval
For the normal distribution, the parameters 1 and 2 denote the
mean and standard deviation, respectively
For the interval variable, the parameters 1 and 2 denote the
lower and upper bounds, respectively
Table 2 Comparison of the analysis results from the present method and the SSL method for the cantilever beam
Method Number of
iterations
Number of
functional
evaluations
b
L
Random variables
(b, h, L) (mm)
Interval variables
(P
x
, P
y
) (N)
SSL 18 198 1.773 0.0757 0.1888 1.0447 53000.00 27000.00
Our method
(deviation)
16 128 1.906
(7.5%)
0.0737
(2.71%)
0.1883
(0.26%)
1.0468
(0.20%)
50454.08
(4.80%)
25080.79
(7.11%)
1
L
2
L
T
t
P
x
d
1
F
1
2
F
2
y
z
Fig. 4 A cantilever tube (Du 2008)
176 C. Jiang et al.
123
In this problem, the nine parameters t, d, F
1
, F
2
, P,
T, S
y
, h
1
and h
2
are treated as randomvariables, and the
two dimensions L
1
and L
2
are treated as interval
variables. All the uncertain variables are given in
Table 3.
The reliability analysis results are listed in Table 4.
It can be found that the present method and the SSL
method can both converge for the cantilever tube
problem. Also, very close results are achieved for the
uncertain variables by using these two methods.
Especially for the 9 random variable the deviations
of the two methods are very small, while for the 2
interval variables the deviations are relatively larger. It
is also because that here the worst case of the intervals
locate on their bounds just like the rst numerical
example. Due to the deviations of the uncertain
variable, a small deviation also occurs in the minimum
reliability index b
L
, which seems completely allow-
able for practical applications. In the respect of
efciency, the present method needs a total of 336
functional evaluations while the SSL method needs
480 evaluations, and hence our method also works
more efciently in this example.
5.3 A 10-bar aluminum truss
A well-known 10-bar aluminum truss (Au et al. 2003;
Jiang et al. 2007; Kang and Luo 2010; Guo et al. 2009)
is investigated, as shown in Fig. 5. The Youngs
Modulus E of the truss is 68,948 MPa and the density
q is 2; 768 kg=m
3
. The length L of the horizontal and
vertical bars is 9.144 m. Subjected to two vertical
forces F
1
and F
2
, and a horizontal force F
3
, the vertical
displacement d
y
on the joint 2 needs to be less than an
allowable value d
ymax
, and hence a following limit-
state function can be created:
g X d
y max
d
y
X 34
According to the equilibrium and compatibility
equations, the axial forces N
j
, j = 1, 2,,10 in the
bars can be computed by Au et al. (2003):
Table 3 Uncertain variables for the cantilever tube
Random
variable
Distribution
parameter 1
Distribution
parameter 2
Variable type
t (mm) 5 0.1 Normal
d (mm) 42 0.5 Normal
F
1
(N) 3000 200 Type I extreme
value
F
2
(N) 3000 300 Type I extreme
value
P (N) 12000 1200 Normal
T (N m) 90 9 Normal
S
y
(MPa) 220 22 Normal
h
1
() 0 10 Uniform
h
2
() 5 15 Uniform
L
1
(mm) 110 130 Interval
L
2
(mm) 50 70 Interval
For the type 1 extreme value distribution, the parameters 1 and
2 denote the u and a in its CDF exp exp
xu
a
_ _ _
;
respectively
Table 4 Comparison of the analysis results from the present
method and the SSL method for the cantilever tube
Item SSL Our method
(deviation)
Number of iterations 22 16
Number of functional
evaluations
480 336
b
L
3.416 3.692 (8.08%)
t (mm) 4.97 4.97 (0%)
d (mm) 41.7 41.7 (0%)
F
1
(N) 3436 3473 (1.08%)
F
2
(N) 3468 3472 (0.12%)
P (N) 12310 12338 (0.23%)
T (N m) 90 90 (0%)
S
y
(MPa) 168 162 (3.57%)
h
1
() 4.92 4.93 (0.20%)
h
2
() 9.90 9.91 (0.10%)
L
1
(mm) 130 125 (3.85%)
L
2
(mm) 70 65 (7.14%)
1
F
2
F
3
F
1
2
3
4
5
6
9.144m 9.144m
9
.
1
4
4
m
x
y
Fig. 5 A 10-bar aluminium truss (Au et al. 2003)
A new reliability analysis method 177
123
N
1
F
2
2
p
2
N
8
; N
2
2
p
2
N
10
N
3
F
1
2F
2
F
3
2
p
2
N
8
;
N
4
F
2
F
3
2
p
2
N
10
N
5
F
2
2
p
2
N
8
2
p
2
N
10
; N
6
2
p
2
N
10
N
7
2
p
F
1
F
2
N
8
; N
8
a
22
b
1
a
12
b
2
a
11
a
22
a
12
a
21
N
9
2
p
F
2
N
10
; N
10
a
11
b
2
a
21
b
1
a
11
a
22
a
12
a
21
_
_
35
where
a
11
1
A
1
1
A
3
1
A
5
2
p
A
7
2
p
A
8
_ _
L
2E
a
12
a
21
L
2A
5
E
a
22
1
A
2
1
A
4
1
A
6
2
p
A
9
2
p
A
10
_ _
L
2E
b
1
F
2
A
1
F
1
2F
2
F
3
A
3
F
2
A
5
_
2
p
F
1
F
2
A
7
_
2
p
L
2E
b
2
2
p
F
3
F
2
A
4
2
p
F
2
A
5
4F
2
A
9
_ _
L
2E
_
_
36
Then the maximum displacement d
y
can be com-
puted by:
d
y
6
i1
N
0
i
N
i
A
i
2
p
10
i7
N
0
i
N
i
A
i
_ _
L
E
37
where N
i
0
can be obtained from Eq. 35 with a
substitution F
1
= F
3
= 0 and F
2
= 1.
In this problem, the areas A
i
; i 1; 2; . . .; 10 of the
bars are treated as random variables, and the forces F
1
,
F
2
and F
3
are treated as interval variables. Thus the
total number of the uncertain variables reaches to 13,
and they are given in Table 5. For this relatively
complex problem, the SSL fails to converge while our
method can work well, and the results using our
method are listed in Table 6. It can be found that 36
iterative steps are required for the hybrid reliability
analysis and the total number of the functional
evaluations reaches to a relatively large number 592,
as more uncertain variables are involved in this
problem. The minimum reliability index b
L
of the
vertical displacement d
y
caused by the three interval
forces is 1.682 and its corresponding maximum failure
probability of the truss is 0.0462, which indicate that
the present truss has not a high reliability of its vertical
stiffness and a structural modication seems required.
5.4 A 25-bar steel truss
A 25-bar steel truss (Jiang et al. 2007) as shown in
Fig. 6 is considered. The Youngs Modulus E of the
Table 5 Uncertain variables for the 10-bar aluminum truss
Random
variable
Distribution
parameter 1
Distribution
parameter 2
Variable
type
A
1
A
6
(mm
2
) 4000 50 Normal
A
7
; A
8
(mm
2
) 4000 50 Lognormal
A
9
; A
10
(mm
2
) 4000 200 Type I
extreme
value
F
1
(N) 442800 446800 Interval
F
2
(N) 442800 446800 Interval
F
3
(N) 1709200 1849200 Interval
For the lognormal distribution, the parameters 1 and 2 denote
the l and r in its CDF
1
2p
p
rx
exp
ln xl
2
2r
2
_ _
; respectively
Table 6 Reliability
analysis result for the
10-bar aluminum truss
Item Result
Number of
iterations
36
Number of
functional
evaluations
592
b
L
1.682
A
1
(mm
2
) 3813
A
2
(mm
2
) 3998
A
3
(mm
2
) 4153
A
4
(mm
2
) 4117
A
5
(mm
2
) 3998
A
6
(mm
2
) 3998
A
7
(mm
2
) 3840
A
8
(mm
2
) 3936
A
9
(mm
2
) 3956
A
10
(mm
2
) 4068
F
1
(N) 444928
F
2
(N) 445148
F
3
(N) 1727669
178 C. Jiang et al.
123
truss is 199949.2 MPa and the Poissons ratio m is 0.3.
The length L of each horizontal or vertical bar is
15.24 m. The joint 12 is hinge-supported, and the
joints 6, 8 and 10 are roller-supported. Three vertical
loads F
3
, F
2
and F
1
are applied on the joints 7, 9 and
11, and a horizontal load F
4
is applied on the joint 1.
For this problem, the vertical displacement d
z
of the
joint 9 should be less than an allowable value d
zmax
,
and hence a following limit-state function can be
created:
g X d
z max
d
z
X 38
In this problem, the areas A
i
; i 1; 2; . . .; 25 of the
25 bars are treated as random variables, and the four
forces F
1
, F
2
, F
3
and F
4
are treated as interval
variables. Therefore a total of 29 uncertain variables
are involved, as shown in Table 7. A nite element
analysis (FEA) model is created to compute the
displacement d
z
, in which each bar is treated as a truss
element and the total number of the elements is 25.
The SSL method and the present method are both
used to deal with the above structure. Just like the
preceding 10-bar aluminum truss, our method can
work well for this complex problem while the SSL
method is incapable of converging at a stable solution.
As shown in Table 8, the whole reliability analysis by
using our method needs 16 iterative steps and a total of
512 FEA simulations. Under the variation of the four
forces within their intervals, the possible minimum
reliability index b
L
of the vertical displacement d
z
is
1.376, which corresponds to a maximum structural
failure probability 0.0844. It also indicates that the
vertical stiffness of the 25-bar structure has a relatively
low reliability, and its structure should be modied
when used for practical applications.
5.5 Discussions on the computational results
Synthesizing the computational results of the above 4
numerical examples, the Table 9 can be created for
performance comparison of the present method and
the current SSL method. Firstly, the SSL only works
well for the two simple numerical examples (cantile-
ver beam and cantilever tube), while fails to converge
for the two relatively complex ones (10-bar truss and
25-bar truss) with more uncertain variables. Never-
theless, our method can converge at a stable solution
for all the four numerical examples. Secondly, for the
rst two examples for which the two methods can both
converge, our method exhibits higher efciency than
the SSL method. Compared with the SSL method, our
method can save 35 and 30% computational cost for
the two examples, respectively. Thirdly, the reliability
analysis result of our method seems inferior to the one
of the SSL method as its maximum failure probability
(1) (2) (3) (4)
(10) (9) (8) (7) (6) (5)
(11)
(13) (12) (14) (15)
(16)
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
(25)
1
2
3 4 5
6
7 8
9
10
11
12
1
F
2
F
3
F
4
F
1
5
.
2
4
m
6x15.24m=91.44m
(14)
1
2
Fig. 6 A 25-bar steel truss (Jiang et al. 2007)
Table 7 Uncertain variables for the 25-bar steel truss
Random
variable
Distribution
parameter 1
Distribution
parameter 2
Variable
type
A
1
A
17
(mm
2
) 5200 50 Normal
A
18
A
21
(mm
2
) 5200 50 Lognormal
A
22
A
25
(mm
2
) 5200 250 Type I
extreme
value
F
1
(N) 1.679E6 1.879E6 Interval
F
2
(N) 2.124E6 2.324E6 Interval
F
3
(N) 1.679E6 1.879E6 Interval
F
4
(N) 1.234E6 1..434E6 Interval
A new reliability analysis method 179
123
is slight smaller than the SSL method. It is because that
the worst case of the interval variables often occurs on
their bounds which in our treatment will lead to a
numerical difculty of plus innity or minus innity,
and hence a numerical error will be produced.
However, the numerical examples show that the
computational deviations between the SSL method
and our method are relatively small, which can fully
satisfy the precision requirement of practical applica-
tions. As a result, the numerical examples demonstrate
that the present method has high efciency and
simultaneously an acceptable analysis precision. More
importantly, through creating an equivalent model it
can be ensured to have a robust convergence perfor-
mance, which is a very useful advantage for complex
engineering problems.
6 Conclusions
In this paper, a new method is developed for reliability
analysis of the uncertain structures with both random
and interval variables. Based on the intervals, it can be
made possible to conduct the reliability analysis for
many practical structures with limited information,
and hence the application range of the probability-
based reliability analysis techniques can be greatly
expanded. In the hybrid reliability analysis, the
maximum failure probability caused by the interval
variables needs to be evaluated, which generally lead
to a very complex nesting optimization problem by
using the existing methods. In our method, the hybrid
reliability is evaluated through solving a simple
equivalent model, which is a conventional reliability
problem with only random variables and can be easily
solved by many well-established methods. Thus the
complex nesting optimization can be successfully
avoided and our method can be ensured to be very
efcient and robust.
Table 8 Reliability
analysis result for the
25-bar steel truss
Item Result
Number of
iterations
16
Number
of FEA
simulations
512
b
L
1.376
A
1
(mm
2
) 5212.3
A
2
(mm
2
) 5198.4
A
3
(mm
2
) 5202.1
A
4
(mm
2
) 5198.1
A
5
(mm
2
) 5220.1
A
6
(mm
2
) 5205.9
A
7
(mm
2
) 5197.0
A
8
(mm
2
) 5201.4
A
9
(mm
2
) 5199.5
A
10
(mm
2
) 5208.6
A
11
(mm
2
) 5197.2
A
12
(mm
2
) 5081.9
A
13
(mm
2
) 5022.6
A
14
(mm
2
) 5093.4
A
15
(mm
2
) 5196.4
A
16
(mm
2
) 5198.8
A
17
(mm
2
) 5149.2
A
18
(mm
2
) 5200.7
A
19
(mm
2
) 5067.6
A
20
(mm
2
) 5081.1
A
21
(mm
2
) 5068.0
A
22
(mm
2
) 5147.4
A
23
(mm
2
) 5297.9
A
24
(mm
2
) 5255.1
A
25
(mm
2
) 5322.0
F
1
(N) 1.776E6
F
2
(N) 2.301E6
F
3
(N) 1.776E6
F
4
(N) 1.351F6
Table 9 Performance comparison of the present method and the SSL method
Example Performance
Which outputs better solution? Which is more efcient? Which is convergent?
Numerical example 1 SSL method Present method Both
Numerical example 2 SSL method Present method Both
Numerical example 3 Present method
Numerical example 4 Present method
180 C. Jiang et al.
123
Four numerical examples with different levels of
complexity are investigated, in which the present
method is also compared with the existing SSL
method. The analysis results show that the present
method has high efciency and very robust conver-
gence performance. Additionally, it can be used very
easily and conveniently, because all of the well-
established methods for conventional reliability anal-
ysis can be directly employed to solve the equivalent
model while nearly need not make any modication.
On the other hand, there will be a numerical error in
our method when the worst case of the interval
variables occurs on the bounds. However, this devi-
ation is relatively small and it can fully satisfy the
request of the practical structural analysis.
Acknowledgments This work is supported by the National
Science Foundation of China (10802028), Key Project of
Chinese National Programs for Fundamental Research and
Development (2010CB832700), Hunan Natural Science
Foundation (10JJ4039).
References
Au, F.T.K., Cheng, Y.S., Tham, L.G., Zeng, G.W.: Robust
design of structures using convex models. Comput. Struct.
81, 26112619 (2003)
Bayarri, M.J., Berjer, J.O., Higdon, D., Kennedy, M.C., Kottas,
A., Paulo, R., Sacks, J., Cafeo, J.A., Cavendish, J., Lin,
C.H., Tu, J.: A framework for validation of computer
models. In: Foundations for Verication and Validation in
the 21st Century Workshop. John Hopkins University/
Applied Physics Laboratory, Laurel (2000)
Ben-Haim, Y., Elishakoff, I.: Convex Models of Uncertainties
in Applied Mechanics. Elsevier Science Publisher,
Amsterdam (1990)
Breitung, K.W.: Asymptotic approximation for multinormal
integrals. ASCE J. Eng. Mech. 110(3), 357366 (1984)
Breitung, K.W.: Asymptotic Approximations for Probability
Integrals. Springer, Berlin (1994)
Cheng, Y.S., Zhong, Y.X., Zeng, G.W.: Structural robust design
based on hybrid probabilistic and non-probabilistic mod-
els. Chin. J. Comput. Mech. 22(4), 501505 (2005)
Ditlevsen, O., Madsen, H.O.: Structural Reliability Methods.
Wiley, Chichester (1996)
Du, X.P.: Interval reliability analysis. In: ASME 2007 Design
Engineering Technical Conference and Computers and
Information in Engineering Conference (DETC2007), Las
Vegas, NV, USA (2007)
Du, X.P.: Saddle point approximation for sequential optimiza-
tion and reliability analysis. ASME J. Mech. Des. 130,
01100110110022 (2008)
Du, X.P., Sudjianto, A., Huang, B.Q.: Reliability-based design
with the mixture of random and interval variables. In:
ASME 2003 Design Engineering Technical Conference
and Computers and Information in Engineering Confer-
ence (DETC2003), Chicago, IL, USA (2005)
Elishakoff, I., Colombi, P.: Combination of probabilistic and
convex models of uncertainty when limited knowledge is
present on acoustic excitation parameters. Comput. Meth-
ods Appl. Mech. Eng. 104, 187209 (1993)
Elishakoff, I., Colombi, P.: Ideas of probability and convexity
combinedfor analyzingparameter uncertainty. In: Schueller,
G.I., Shinozuka, M., Yao, J. (eds.) Proceedings of the 6th
International Conference on Structural Safety and Reliabil-
ity, pp. 109113. Balkema Publishers, Rotterdam (1994)
Elishakoff, I., Cai, G.Q., Starnes, J.H.: Non-linear buckling of a
column with initial imperfections via stochastic and non-
stochastic convex models. Int. J. Nonlinear Mech. 29,
7182 (1994)
Guo, J., Du, X.P.: Reliability sensitivity analysis with random
and interval variables. Int. J. Numer. Methods Eng. 78,
15851617 (2009)
Guo, S.X., Lu, Z.Z.: Hybrid probabilistic and non-probabilistic
model of structural reliability. J. Mech. Strength 24(4),
524526 (2002)
Guo, X., Bai, W., Zhang, W.S.: Extremal structural response
analysis of truss structures under load uncertainty via SDP
relaxation. Struct. Saf. 87, 246253 (2009)
Hasofer, A.M., Lind, N.C.: Exact and invariant second-moment
code format. ASME J. Eng. Mech. Div. 100, 111121
(1974)
Jiang, C., Han, X., Li, W.X., Liu, J.: A hybrid reliability
approach based on probability and interval for uncertain
structures. ASME J. Mech. Des. (in press)
Jiang, C., Han, X., Liu, G.R.: Optimization of structures with
uncertain constraints based on convex model and satis-
faction degree of interval. Comput. Methods Appl. Mech.
Eng. 196, 47914890 (2007)
Kang, Z., Luo, Y.J.: Reliability-based structural optimization
with probability and convex set hybrid models. Struct.
Multidiscip. Optim. 42(1), 89102 (2010)
Liang, J.H., Mourelators, Z.P., Tu, J.: A single-loop method for
reliability-based design optimization. In: Proceedings of
DETC04 ASME 2004 Design Engineering Technical
Conferences and Computers and Information in Engi-
neering Conference, Salt Lake City, UT, USA (2004)
Luo, Y.J., Kang, Z., Alex, L.: Structural reliability assessment
based on probability and convex set mixed model. Comput.
Struct. 87, 14081415 (2009)
Madsen, H.O.: First order vs. second order reliability analysis of
series structures. Struct. Saf. 2, 207214 (1985)
Madsen, H.O., Krenk, S., Lind, N.C.: Methods of Structural
Safety. Prentice-Hall, Englewood Cliffs (1986)
Nocedal, J., Wright, S.J.: Numerical Optimization. Springer,
New York (1999)
Qiu, Z.P., Yang, D., Elishakoff, I.: Probabilistic interval reli-
ability of structural systems. Int. J. Solids Struct. 45,
28502860 (2008)
Rackwitz, R., Fiessler, B.: Structural reliability under combined
random load sequences. Comput. Struct. 9(5), 489494
(1978)
Rubinstein, R.Y., Kroese, D.P.: Simulation and the Monte-Carlo
Method, 2nd edn. Wiley, New York (2007)
Thoft-Christensen, P., Murotsu, Y.: Application of Structural
Systems Reliability Theory. Springer, Berlin (1986)
A new reliability analysis method 181
123
Yuan, Y.X., Sun, W.Y.: Optimization Theories and Methods.
Scientic Press, Beijing (2005)
Zhang, Y., der Kiureghian, A.: Two improved algorithms for
reliability analysis. Reliability and optimization of struc-
tural systems, In: Proceedings of the Sixth IFIP WG7.5
Working Conference on Reliability and Optimization of
Structural Systems, Assisi, Italy (1995)
Zhu, L.P., Elishakoff, I.: Hybrid probabilistic and convex
modeling of excitation and response of periodic structures.
Math. Probl. Eng. 2, 143163 (1996)
182 C. Jiang et al.
123