Emulator Model–Based Analytical Solution

for Reliability Sensitivity Analysis

Leigang Zhang 1; Zhenzhou Lu 2; Lei Cheng 3; and Zhangchun Tang 4

Abstract: Sensitivity analysis is frequently considered an essential component in engineering design. In the design process of engineered
structures, the output is implicitly related with the input variables. The Kriging model, one of the most commonly used emulator models, is
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

sometimes used for structure analysis. In order to efficiently estimate the sensitivities of failure probability or statistical moments of per-
formance function with respect to distribution parameters of input variables, the analytical solutions are derived based on the Kriging model.
Generally, the Kriging model can be expressed as a tensor product basis function, thus the multivariate integrals can be decomposed into the
sum of univariate integrals, which makes it possible to solve the sensitivity of statistical moments with respect to distribution parameters of
normal input variables by the properties of kernel functions. Next, the fourth-moment reliability sensitivity method is applied to compute the
sensitivity of failure probability analytically. Numerical and engineering examples are introduced to demonstrate the accuracy and efficiency
of the derived analytical solution of sensitivity of failure probability. DOI: 10.1061/(ASCE)EM.1943-7889.0000897. © 2015 American
Society of Civil Engineers.
Author keywords: Sensitivity analysis; Kernel function; Kriging model; Statistical moment; Distribution parameters.

Introduction dimensionality, so it is only suitable for low-dimensional problems.

Monte Carlo simulation (MCS) based reliability sensitivity
Sensitivity analysis, both local and global, has been widely used in method (Lataillade et al. 2002; Melchers and Ahammed 2004;
engineering design. Local sensitivity is often defined as the partial Roger et al. 1999), the representation of numerical simulation
derivatives of failure probability or statistical moments of perfor- method, is widely used. However, it is well known that the signifi-
mance function with respect to distribution parameters of the input cant drawback of MCS is seldom affordable in computational cost,
variables (Zhang et al. 2014), which shows how the distribution even though it is highly accurate. It is also unsuitable for reliability
parameters of input variables affect the failure probability or stat- sensitivity analysis of engineered structures with small real failure
istical characteristics of structures. In this paper, the authors focus probabilities, since in these cases, large numbers of samples are
on the study of local sensitivity. In the past few decades, many sen- needed to get the convergent estimates of failure probability and
sitivity analysis methods have been developed. Generally, these sensitivities. The idea of the surrogate model–based method (Guan
methods can be divided into three groups: the approximate analyti- and Melchers 2001; Kaymaz and McMahon 2005) is to approxi-
cal method, numerical simulation method, and surrogate model mate the implicit performance function by an explicit function, on
based method. The approximate analytical method includes the ad- which the sensitivity of failure probability can be implemented. In
vanced first order and second moment (AFOSM) reliability sensi- the existing methods, the sensitivity of failure probability is esti-
tivity method (Karamchandani and Cornell 1991) and the point mated by other numerical methods after the explicit performance
estimation method (Rosenblueth 1975; Rosenblueth 1981; Zhao function is constructed by the surrogate model. In this paper, mean-
and Ono 2000). AFOSM is not useful for highly nonlinear perfor- while, the authors try to derive the analytical sensitivity results
mance functions, and it depends on the derivative information based on the surrogate model.
of the performance function. The computational cost of the point As a matter of fact, the Kriging model is widely used in engi-
estimation method is exponential growth with the growth of input neering design. For example, Lucifredi et al. (2000) compared the
dynamic Kriging model with the neural network model, and used it
1 to predict maintenance and computation of hydroelectric power
Master Student, School of Aeronautics, Northwestern Polytechnical
systems. Sakata et al. (2003) implemented the structural optimiza-
Univ., Shaanxi, Xi’an 710072, China. E-mail: leigang_zhang@163
.com tion using Kriging approximation. Wang et al. (2013) applied the
2
Professor, School of Aeronautics, Northwestern Polytechnical Kriging method in global sensitivity analysis, etc. This paper at-
Univ., Shaanxi, Xi’an 710072, China (corresponding author). E-mail: tempts to derive the analytical solution of sensitivity of failure prob-
[email protected] ability based on the Kriging model for problems with independent
3
Ph.D. Candidate, School of Aeronautics, Northwestern Polytechnical normal input variables, so that the reliability-based design optimi-
Univ., Shaanxi, Xi’an 710072, China. E-mail: [email protected] zation of engineered structures can be facilitated. Millwater (2009)
.edu.cn introduced a kernel function and derived some universal properties
4
Lecturer, School of Mechatronics Engineering, Electronic Science of it, and then applied it to the computation of reliability sensitivity.
and Technology Univ., Sichuan, Chengdu 611731, China. E-mail:
It is worth mentioning that the Kriging model can be expressed as a
[email protected]
Note. This manuscript was submitted on April 24, 2014; approved on
tensor product basis function (TPBF) (Chen et al. 2005), so a multi-
October 24, 2014; published online on April 23, 2015. Discussion period variate integral can be decomposed into the sum of some univariate
open until September 23, 2015; separate discussions must be submitted for integrals, which makes it possible to solve the sensitivity of failure
individual papers. This paper is part of the Journal of Engineering Me- probability analytically by properties of kernel functions. Besides,
chanics, © ASCE, ISSN 0733-9399/04015016(10)/$25.00. the fourth-moment method shows high accuracy in the computation

© ASCE 04015016-1 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 of failure probability, and has been widely applied in engineering Analysis of the Relation between Sensitivity of
design (Zhao and Ono 2001). Based on the fourth-moment method, Failure Probability and Sensitivity of Statistical
Lu and Song (2010) proposed the reliability sensitivity analysis Moments of the Performance Function
method. In this paper, the authors have chosen the fourth-moment
based sensitivity analysis method as a basic approach for comput- Analysis of the Sensitivity of Failure Probability
ing the sensitivity of failure probability analytically.
The remainder of this paper is organized as follows. First, a brief According to the moment-based reliability assessment method in
introduction of the kernel functions for sensitivity analysis is pre- Zhao and Ang (2003) and Zhao and Ono (2004), the second
sented. Second, the fourth-moment based method is employed to and fourth moment reliability indices, denoted by β 2M and β 4M re-
construct the relation between the sensitivity of failure probability spectively, and the failure probability Pf, can be calculated by
and that of statistical moments of the performance function. Third, Eqs. (5)–(7), respectively
for the widely used Kriging model that can be expressed as a TPBF,
α1g
the corresponding integral expressions for computing the sensitiv- β 2M ¼ ð5Þ
ity of statistical moments of the performance function are derived. α2g
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

Numerical and engineering examples are introduced to illustrate the

correctness and efficiency of the proposed analytical solution of
sensitivity. Conclusions are provided in the end. 3ðα4g − 1Þβ 2M þ α3g ðβ 22M − 1Þ
β 4M ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð6Þ
ð5α23g − 9α4g þ 9Þð1 − α4g Þ
Sensitivity of Failure Probability and the
Corresponding Kernel Functions
Pf ¼ Φð−β 4M Þ ð7Þ
For the performance function Z ¼ gðxÞ related with the indepen-
dent n-dimensional random input vector x ¼ fx1 ; x2 ; : : : ; xn g, where Φð·Þ = cumulative distribution function (CDF) of the stan-
the failure probability of the structures or systems, Pf , can be dard normal random variable, and αmg ðm ¼ 1; : : : ; 4Þ = first four
defined as moments of the performance function, the expressions of which are
given as
Z
Pf ¼ f X ðxÞdx ð1Þ Z
F α1g ¼ gðxÞf X ðxÞdx ¼ E½gðxÞ
Rn
Qn Z 1=2
where fX ðxÞ ¼ l¼1 fXl ðxl Þ = joint probability density function α2g ¼ ½gðxÞ − α1g 2 f X ðxÞdx ¼ fE½gðxÞ − α1g 2 g1=2
(PDF) of x, fXl ðxl Þ = marginal PDF of xl , and F = failure Rn
domain defined by the performance function Z ¼ gðxÞ as F ¼ Z
1 1
fx∶gðxÞ ≤ 0g α3g ¼ 3 ½gðxÞ − α1g 3 fX ðxÞdx ¼ 3 Ef½gðxÞ − α1g 3 g
α2g Rn α2g
Then the sensitivity of failure probability Pf with respect to the Z
kth distribution parameter of the input xl can be derived as 1 1
α4g ¼ 4 ½gðxÞ − α1g 4 fX ðxÞdx ¼ 4 Ef½gðxÞ − α1g 4 g ð8Þ
α2g Rn α2g
Z Z
∂Pf ∂ ∂f X ðxÞ
ðkÞ
¼ ðkÞ
· f X ðxÞdx ¼ ðkÞ
dx ð2Þ
∂θxl ∂θxl F F ∂θxl By using the relation between Pf and β 4M in Eq. (7) and the
ðkÞ
differential chain rule, the sensitivity of Pf with respect to θxl
ðkÞ can be derived as (Lu and Song 2010)
where θxl = kth distribution parameters of the input xl , such as
mean, standard deviation, shape, and scale factor.   
Millwater (2009) expressed the sensitivity of failure probability ∂Pf ∂Pf ∂β 4M ∂Pf ∂β 4M ∂β 2M ∂α1g ∂β 2M ∂α2g
¼ ¼ þ
with respect to distribution parameters of the input as follows: ðkÞ
∂θxi ∂β 4M ∂θðkÞ
xi
∂β 4M ∂β 2M ∂α1g ∂θðkÞ
xi
∂α2g ∂θðkÞ
xi

Z Z ∂β ∂α3g ∂β ∂α4g
∂Pf ∂fX ðxÞ 1 þ 4M ðkÞ þ 4M ðkÞ ð9Þ
¼ · · f ðxÞdx ¼ kθðkÞ · fX ðxÞdx ð3Þ ∂α3g ∂θx ∂α4g ∂θx
ðkÞ
∂θxl F
ðkÞ
∂θxl f X ðxÞ X F xl i i

ðkÞ ∂αmg
where kθðkÞ is the kernel function of the distribution parameter θxl where ðkÞ ðm ¼ 1; : : : ; 4Þ = sensitivities of statistical moments of
x ∂θxi
defined asl the performance function; its analytical solution will be derived in
the next subsection. Eq. (9) gives the relation between sensitivity of
∂f X ðxÞ 1 ∂f Xl ðxl Þ 1 failure probability and that of output moments. The other formu-
kθðkÞ ¼ · ¼ · ð4Þ
xl ðkÞ
∂θxl fX ðxÞ ∂θx
ðkÞ f Xl ðxl Þ lations in Eq. (9) can be given as (Lu and Song 2010)
l

 2 
Given the distribution type of the input variable xl , the ∂Pf 1 β
¼ − pffiffiffiffiffiffi exp − 4M ð10Þ
explicit expression of the kernel function kθðkÞ can be derived ∂β 4M 2π 2
ðkÞ xl
by Eq. (4) for the parameter θxl . Because normal variables are
mostly studied in the field of reliability sensitivity analysis,
the proposed method is also derived based on the normally 3α4g þ 2α3g β 2M − 3
∂β 4M
distributed variables in this paper. The detailed expressions of ¼ ð11Þ
the kernel function for normal variables will be given later. ∂β 2M ½ð9α4g − 5α23g − 9Þðα4g − 1Þ12

© ASCE 04015016-2 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 ∂β 4M β 22M − 1 E½gðxÞm · ðkθðkÞ Þn ðm ¼ 1; : : : ; 4; n ¼ 0; 1Þ. Meanwhile, from
¼ xl
∂α3g 1
½ð9α4g − 5α23g − 9Þðα4g − 1Þ2 Eq. (8) it can be seen that αmg can be transformed into the form
5½3ðα4g − 1Þβ 2M þ α3g πðβ 22M − 1Þðα4g − 1Þα3g of expected value of a certain function indirectly, thereby compu-
þ 3 ð12Þ tations of αmg ðm ¼ 1; : : : ; 4Þ are approximately equivalent to com-
2½ð9α4g − 5α23g − 9Þðα4g − 1Þ2 putations of E½gðxÞm · ðkθðkÞ Þn ðm ¼ 1; : : : ; 4; n ¼ 0; 1Þ. From the
xl

above discussion, it can be concluded that if the approximate ana-

∂β 4M 3β 2M lytical results of E½gðxÞm · ðkθðkÞ Þn  are available, the analytical so-
¼
∂α4g 1 xl
½ð9α4g − 5α23g − 9Þðα4g − 1Þ2 ∂αmg
lutions of ðkÞ ðm ¼ 1; : : : ; 4Þ can be obtained accordingly.
∂θxl
½3ðα4g − 1Þβ 2M þ α3g ðβ 22M − 1Þð18α4g − 18 − 5α23g Þ
− 3
It seems impossible to obtain the analytical expressions of
2½ð9α4g − 5α23g − 9Þðα4g − 1Þ2 E½gðxÞm · ðkθðkÞ Þn  for general performance functions. However,
xl

ð13Þ for the performance function expressed by a TPBF, it is possible

Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

to obtain the analytical results of E½gðxÞm · ðkθðkÞ Þn . Note that

xl

∂β 2M 1 the proposed analytical method is under the assumption that the

¼ ð14Þ
∂α1g α2g random input variables are independent normally distributed,
and for the normal random variables, if substituting the expressions
of kernel functions into E½gðxÞm · ðkθðkÞ Þn , the work is changed
∂β 2M α1g xl
¼− 2 ð15Þ into the computation of E½gðxÞm · ðkθðkÞ Þn ðm ¼ 1; : : : ; 4; p ¼
∂α2g α2g xl

0; 1; 2Þ. Appendix II gives the detailed process of transforming

It can be seen from Eq. (9) that if the analytical solution of E½gðxÞm · ðkθðkÞ Þn  into E½gðxÞm · xpl  for the normal random varia-
xl
∂αmg
ðkÞ ðm ¼ 1; : : : ; 4Þ can be derived, the analytical solution bles. The next section will feature a discussion the solution of
∂θxi
∂Pf E½gðxÞm · xpl  in detail, under the condition that gðxÞ is expressed
of ðkÞ will be obtained correspondingly. as a TPBF.
∂θxi
In summary, the computation of sensitivity of failure probability
is transformed into the computation of E½gðxÞm · xpl ðm ¼
Analysis of the Sensitivity of Statistical Moments 1; : : : ; 4; p ¼ 0; 1; 2Þ, and if the analytical solution of
Using Kernel Functions E½gðxÞm · xpl  is obtained, the analytical solution of the sensitivity
On the basis of the definition expressions of αmg ðm ¼ 1; : : : ; 4Þ of failure probability will be obtained accordingly.
in Eq. (8), the following expressions of the sensitivity of output
moments using kernel functions can be derived
Analytical Solution of E
gx m · x pl  for the
∂α1g TPBF-Based Emulator Model
ðkÞ
¼ E½gðxÞ · kθðkÞ  ð16Þ
∂θxl xl As the performance function can be expressed as a TPBF, the ana-
lytical results of univariate integrals can be used to evaluate the
multivariate integrals. The Kriging model can be expressed as a
∂α2g 1
¼ fE½gðxÞ2 · kθðkÞ  − 2α1g E½gðxÞ · kθðkÞ g ð17Þ TPBF (Chen et al. 2005), so that the analytical expressions of
∂θxl
ðkÞ 2α2g xl xl
E½gðxÞm · xpl ðm ¼ 1; : : : ; 4; p ¼ 0; 1; 2Þ can be derived according
to the properties of TPBF.
  As is well known, the Kriging model has high precision and is
∂α3g 1 3
3α3g 3α1g
¼ 3 E½gðxÞ · kθðkÞ  − þ 3 E½gðxÞ2 · kθðkÞ  sometimes used in engineering problems for structural design op-
∂θxl
ðkÞ α2g xl 2α22g α2g xl
timization and reliability analysis (Kaymaz 2005; Sakata et al.
  2003). Based on this, the analytical solution of E½gðxÞm · xpl ðm ¼
3α3g α1g 3α21g 3ðE½gðxÞ2  − α21g Þ
þ þ 3 − E½gðxÞ · kθðkÞ  1; : : : ; 4; p ¼ 0; 1; 2Þ is derived for the Kriging model in this sec-
α22g α2g α32g xl
tion. This section is organized as follows. First, a brief introduction
ð18Þ of the Kriging model is presented. The concept of TPBF is then
introduced and E½gðxÞm · xpl  is derived further based on a generic
∂α4g 1 4α1g TPBF-expressed function. Finally, for the TPBF-expressed Kriging
ðkÞ
¼ 4
E½gðxÞ4 · kθðkÞ  − 4 E½gðxÞ3 · kθðkÞ  model, the final analytical solution of E½gðxÞm · xpl  is derived.
∂θxl α2g x l α2g xl

 2 
6α1g 2α4g
þ − E½gðxÞ2 · kθðkÞ  A Gentle Introduction to Kriging
α42g α22g xl

  The Kriging method is a semiparametric interpolation technique,

4α1g α4g 4α31g 4ðE½gðxÞ3  − 3α1g E½gðxÞ2  þ 2α31g Þ and Sacks et al. (1989) expressed the deterministic output gðxÞ
þ − 4 −
α22g α2g α42g as two parts: the linear regression part and the nonparametric part.
In this paper, only a constant β is applied in the linear regression
× E½gðxÞ · kθðkÞ  ð19Þ
xl part, which was also done by Chen et al. (2005); in addition, this
formula contributes to a convenient TPBF representation, and gðxÞ
The detailed derivation process of Eqs. (16)–(19) are presented is expressed as
∂αmg
in Appendix I. It can be seen from Eqs. (16)–(19) that ðkÞ
∂θxl
ðm ¼ 1; : : : ; 4Þ are composed of αmg ðm ¼ 1; : : : ; 4Þ and gðxÞ ¼ β þ ZðxÞ ð20Þ

© ASCE 04015016-3 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 where the term ZðxÞ is used to model the deviation from β and is N 
X Y
n 
assumed to be a Gaussian stochastic process with zero mean and σ2 E½ĝðxÞ · xj  ¼ a0 · μxj þ ai · D1;il · C1;il ð29Þ
variance, and the covariance Cov½Zðxi Þ; Zðxj Þ ¼ σ2 Rðxi ; xj Þ, i¼1 l¼1;l≠j

where Rðxi ; xj Þ = a correlation function and is defined as N  

X Y
n

Y
n E½ĝðxÞ · x2j  ¼ a0 · ðμ2xj þ σ2xj Þ þ ai · E1;il · C1;il
Rðxi ; xj Þ ¼ Rl ðxil − xjl Þ ð21Þ i¼1 l¼1;l≠j
l¼1 ð30Þ
where xil and xjl = lth component of vector xi and xj respectively. where j ¼ 1; : : : ; n. For a specific function, the expressions
Many different correlation functions could be used (Currin et al. of C1;il , D1;il , and E1;il can be derived analytically, then
1991; Sacks et al. 1989). In practice, the Gaussian correlation func- E½gðxÞm · xpl  can be obtained analytically. In the following subsec-
tion has become the most widely used one; it is beneficial for the tion, the analytical expressions of C1;il , D1;il , and E1;il of the
following TPBF-based derivation in this paper, and it can be given Kriging model are derived in detail.
as
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

Rl ðxil − xjl Þ ¼ exp½−τ l ðxil − xjl Þ2  ð22Þ Analytical Expressions of C 1;il , D 1;il , and E 1;il for the
Kriging Model
where τ l = correlation parameter.
A Kriging model can be written based on the TPBF as following
The Kriging model can predict the output results exactly; for the
(Chen et al. 2005):
detailed description of Kriging one can refer to Currin et al. (1991);
Kaymaz (2005); Sacks et al. (1989). N  Y
X n 
ĝðxÞ ¼ β þ λi Ril ðxl Þ ð31Þ
i¼1 l¼1
Tensor Product Basis Function
where λi ði ¼ 1; : : : ; NÞ =constant coefficients. Compared with
A multivariate basis function Bi ðxÞ is defined as a product of n Eq. (24), β and λi ði ¼ 1; : : : ; NÞ are equivalent to a0 and
univariate basis functions hil ðxl Þ, i.e., ai ði ¼ 1; : : : ; NÞ respectively. Ril ðxl Þ is equivalent to the univari-
Y
n ate basis function hil ðxl Þ, and based on the definition in Eq. (22),
Bi ðxÞ ¼ hil ðxl Þ; i ¼ 1; 2; : : : ; N ð23Þ Ril ðxl Þ is given as
l¼1
Ril ðxl Þ ¼ exp½−τ l ðxl − xil Þ2  ð32Þ
For example, BðxÞ ¼ x21 x2
can be rewritten as BðxÞ ¼ where xil is lth component of the experimental point xi .
h1 ðx1 Þh2 ðx2 Þ, where h1 ðx1 Þ ¼ x21 and h2 ðx2 Þ ¼ x2 . In addition, For normal distribution, C1;il has been derived as (Chen et al.
the univariate basis function hil ðxl Þ could ¼ 1 to represent a con- 2005)
stant term.  
Then a TPBF can be defined as a linear expansion of 1 τ
C1;il ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − 2 l ðμxl − xil Þ2 ð33Þ
these multivariate basis functions. Here E½gðxÞm · xpl ðm ¼ 2σxl τ l þ 1
2σ2xl τ l þ 1
1; : : : ; 4; p ¼ 0; 1; 2Þ is used as an example. For each m in
E½gðxÞm · xpl , ĝðxÞ denotes the corresponding gðxÞm , and it can
where μxl and σxl = mean and standard deviation of xl . Further, the
be written as
expressions of D1;il and E1;il can be derived as
X
N N  Y
X n   
ĝðxÞ ¼ a0 þ ai Bi ðxÞ ¼ a0 þ ai hil ðxl Þ ð24Þ 1 τl 2
2σ2xl τ l xil þ μxl
D1;il ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − 2 ðμxl − xil Þ ·
i¼1 i¼1 l¼1
2σ2xl τ l þ 1 2σxl τ l þ 1 2σ2xl τ l þ 1

where ai ði ¼ 0; 1; : : : ; NÞ = constant coefficients, N = sample size, 2σ2xl τ l xil þ μxl

and xl = input variable. ¼ C1;il · ð34Þ
2σ2xl τ l þ 1
Here, let C1;il = mean of the univariate basis function hil ðxl Þ
(Chen et al. 2005)  
1 τl
Z ∞ E1;il ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − 2 ðμx − xil Þ2

2σ2xl τ l þ 1 2σxl τ l þ 1 l
C1;il ¼ E½hil ðxl Þ ¼ hil ðxl ÞfXl ðxl Þdxl ð25Þ
−∞  2  
2σxl τ l xil þ μxl 2 σ2xl
Similarly, D1;il and E1;il are defined as follows · þ 2
2σ2xl τ l þ 1 2σxl τ l þ 1
Z ∞  2 2 
2σxl τ l xil þ μxl σ2xl
D1;il ¼ E½xl hil ðxl Þ ¼ xl hil ðxl ÞfXl ðxl Þdxl ð26Þ ¼ C1;il · þ ð35Þ
−∞ 2σ2xl τ l þ 1 2σ2xl τ l þ 1

Z ∞ The detailed derivation processes of D1;il and E1;il are shown in

E1;il ¼ E½x2l hil ðxl Þ ¼ x2l hil ðxl Þf Xl ðxl Þdxl ð27Þ Appendix III. When a Kriging model is available, Eqs. (33)–(35)
−∞
can be computed using the parameters obtained. The Kriging model
has high precision to represent the implicit input–output relation,
For Eq. (24), a function expressed as a TPBF, it can derived that
and in order to control the error propagation and prevent the error
N  Y
X n  expansion, the Kriging model ĝðxÞ is constructed for each gðxÞm
E½ĝðxÞ ¼ a0 þ ai C1;il ð28Þ and then E½ĝðxÞ · xpj ðp ¼ 0; 1; 2Þ is computed to obtain a solution
i¼1 l¼1 of E½gðxÞm · xpi .

© ASCE 04015016-4 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 All in all, in this section, the expressions of C1;il , D1;il , and E1;il 5x1 x32
gðxÞ ¼ x24 − ð36Þ
of the Kriging-based emulator model are obtained, which makes it x23
possible to analytically compute sensitivities of statistical moments
of the performance function.
Results of all the methods are listed in Table 1.
This is a simple numerical example with explicit performance
Method Implementation function; from the sensitivity analysis of this example, it is possible
to demonstrate the correctness and precision of the analytical
In this paper, the authors propose a Kriging model–based analytical method proposed in this paper, for the sake of popularizing this
method for reliability sensitivity analysis, and the analytical solu- method in engineering applications. It can be seen from Table 1,
tions are derived under the assumption that the input random var- for 106 evaluations of the performance function, the variation co-
iables are independent normally distributed. The process of efficients of the results obtained by the MCS method are small
obtaining the analytical solution of the reliability sensitivity can enough, and the results can be set as references. For this nonlinear
be briefly described as follows: performance function, it can be seen that the results obtained by
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

1. Construct the Kriging model ĝðxÞ [Eq. (31)] for each m of AFOSM have lower precision. Compared with MCS, IS is an ef-
gðxÞm . Because of the low discrepancy property of experimen- ficient numerical simulation method but its computational cost is
tal points generated by the Latin hypercube sampling (LHS) also large. According to a convergence analysis, it only needs to
method, the LHS method is used to generate the experimental generate 80 experimental points by the LHS method to construct
points to construct the Kriging model in this paper. a reasonable Kriging model and then compute the indices directly
2. Use the parameters of the available Kriging model to compute without any subsequent methods or any evaluations of the perfor-
C1;il , D1;il , and E1;il by Eqs. (33)–(35). Then compute mance function; these results are also convergent to the MCS re-
E½ĝðxÞ · xpl ðp ¼ 0; 1; 2Þ using Eqs. (28)–(30), where a0 sults. Therefore, it shows that the proposed method is reasonable
and ai ði ¼ 1; : : : ; NÞ in Eqs. (28)–(30) = β and λi ði ¼ and is of high precision and efficiency in calculation.
1; : : : ; NÞ in Eq. (34) respectively. In other words, E½gðxÞm · The results in Table 1 show that the absolute values of sensitiv-
ðkθðkÞ Þn ðm ¼ 1; : : : ; 4; n ¼ 0; 1Þ have been obtained. ities of x2 and x3 are much larger than those of x1 and x4 , which
xl indicates that x2 and x3 have larger impact on the failure probability,
3. Substitute E½gðxÞm · ðkθðkÞ Þn  obtained in step (2) into whereas the failure probability is not sensitive to the distribution
xl
Eqs. (16)–(19) to compute the sensitivity of the statistical mo- parameters of x1 and x4 . It can also be noted that an increase of the
ments of the performance function. mean μx3 would decrease the failure probability since the corre-
4. Compute the sensitivity of failure probability by its relation sponding sensitivity is negative, whereas an increase of the mean
with the sensitivity of the statistical moments of the perfor- μx2 would increase the failure probability. An increase of the stan-
mance function in Eq. (9). dard deviations of these two variables would also increase the failure
probability since the corresponding sensitivities are positive. There-
fore, more attention should obviously be paid to the parameters of
Illustrative Examples variables x2 and x3 to change the failure probability of this problem.
In this section, a numerical example and two engineering examples
are employed to illustrate the precision of the results obtained by Automobile Front Axle
the proposed analytical method (AM). The AFOSM method and
the important sampling (IS) method (Melchers 1989) are applied In automobile engineering, the front axle beam (Zhang et al. 2015)
for comparison. Additionally, MCS method is employed as a refer- in Fig. 1 is used to carry the weight of the front part of the vehicle.
ence. Results of the sensitivities and the number of the performance As the complete front part of the body rests on the body front axle
function evaluations N call are presented for comparison. beam, it must be robust in construction. Note that the I-beam struc-
tures are widely used in the design of front axle due to its high bend
strength and light weight. As shown in Fig. 1(a), the dangerous
Numerical Example
cross section happens in the I-beam part, and Fig. 1(b) gives the
Considering a nonlinear performance function gðxÞ, the in- cross section of the I-beam. The maximum normal stress and shear
cluded four variables xi ði ¼ 1; : : : ; 4Þ are independent normally stress = σ ¼ M=W x and τ ¼ T=W ρ respectively, where M and
distributed, where x1 ∼ Nð460; 72 Þ, x2 ∼ Nð20; 2.42 Þ, x3 ∼ Nð19; T = bending moment and torque, and W x and W ρ = section factor
0.82 Þ, and x4 ∼ Nð342; 31.42 Þ, and gðxÞ is given by and polar section factor, which can be written as

Table 1. Sensitivity Results of the Numerical Example

Results MCS AFOSM IS AM
Pf 0.0170(0.0076) 0.0202 0.0165 0.0174
∂Pf =∂μx1 2.636 × 10−4 ð0.0111Þ 2.001 × 10−4 2.227 × 10−4 2.389 × 10−4
∂Pf =∂μx2 0.0141(0.0080) 0.0096 0.0144 0.0137
∂Pf =∂μx3 −0.0121ð0.0153Þ −0.0098 −0.0121 −0.0118
∂Pf =∂μx4 −7.465 × 10−4 ð0.0090Þ −8.457 × 10−4 −7.569 × 10−4 −7.228 × 10−4
∂Pf =∂σx1 2.738 × 10−5 ð0.0097Þ 2.344 × 10−5 2.128 × 10−5 2.072 × 10−5
∂Pf =∂σx2 0.0238(0.0098) 0.0307 0.0243 0.0234
∂Pf =∂σx3 0.0061(0.0147) 0.0053 0.0071 0.0072
∂Pf =∂σx4 8.993 × 10−4 ð0.0124Þ 1.646 × 10−3 8.781 × 10−4 8.508 × 10−4
N call 106 15 5,000 80
Note: Numbers in brackets are variation coefficients of the results of MCS method.

© ASCE 04015016-5 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 Table 2. Distribution Parameters of Input Variables of the Automobile
Front Axle
Variables Mean Standard deviation
a=mm 12 0.06
b=mm 65 0.325
t=mm 14 0.07
h=mm 85 0.425
M=ðN · mmÞ 3.5 × 106 1.75 × 105
T=ðN · mmÞ 3.1 × 106 1.55 × 105

Table 3. Failure Probability of the Automobile Front Axle

Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

Method Pf N call
MCS 0.0193(0.0092) 106
AFOSM 0.0238 22
IS 0.0197 6,000
AM 0.0190 120
Note: Numbers in brackets are variation coefficients of the results of MCS
method.

Fig. 1. The front axle of an automobile: (a) schematic diagram of the

front axle; (b) cross section of the front axle

aðh − 2tÞ3 b
Wx ¼ þ ½h3 − ðh − 2tÞ3  ð37Þ
6h 6h

a3 ðh − 2tÞ Fig. 2. Sensitivities of Pf with respect to μxl of the automobile front

W ρ ¼ 0.8bt2 þ 0.4 ð38Þ axle
t

To check the static strength of the front axle, the performance

function can be given as
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
g ¼ σs − σ2 þ 3τ 2 ð39Þ

where σs = limit stress of yielding. According to the material prop-

erty of the front axle, the limit stress of yielding σs ¼ 460 MPa. The
geometry variables of I-beam a, b, t, h, and the loads M and T are
independent normal variables with distribution parameters listed
in Table 2. Results of the failure probability are given in Table 3.
Sensitivity results of all the methods are shown in Figs. (2) and (3).
This engineering example is applied to further demonstrate the
advantages of the proposed method. In order to get enough precise
results, 106 samples are also used to make the results as references,
with the variation coefficient of Pf ¼ 0.0092. Those of the sensi-
tivities are all small enough and for simplicity, they are not shown
here. From Table 3 and Figs. 2 and 3 it can be seen that the results
obtained by other three methods agree with the results of MCS on
the whole. While actually, for highly nonlinear performance func-
Fig. 3. Sensitivities of Pf with respect to σxl of the automobile front
tion, AFOSM is not competent. The results of IS method show high
axle
precision, while 6,000 calls of the performance function are large

© ASCE 04015016-6 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 Table 5. Sensitivity Results of the Aircraft Inside Flap
Results IS AM
Pf 0.0922 0.0905
∂Pf =∂μx1 ð×10−5 Þ −1.1857 −1.0583
∂Pf =∂μx2 ð×10−5 Þ −1.4040 −1.0322
∂Pf =∂μx3 1.6105 1.6173
∂Pf =∂σx1 ð×10−6 Þ 4.0916 3.8194
Fig. 4. The finite element model of an aircraft inside flap ∂Pf =∂σx2 ð×10−6 Þ 5.9236 6.2400
∂Pf =∂σx3 2.1310 2.1569
N call 2 × 104 500

all the same. For 120 evaluations of the performance function, the
proposed method can obtain convergent results. This further shows
the rationality of the method proposed, and the computational bur- corresponding IS results very well. Additionally, the number of
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

den can be largely reduced. FEM evaluations of IS is 2 × 104 , whereas the analytical method
Analyzing the results in Figs. 2 and 3, the sensitivities of dis- of this paper just needs to call the FEM 500 times to get the con-
tribution parameters of the geometry variables a and t of the I-beam vergent results, thereby fully demonstrating the engineering appli-
are much larger than those of other variables, so the failure prob- cation of the proposed method.
ability is much more sensitive to the distribution parameters of the The reliability model in this example sets the elastic modulus E,
length of a and t, and the change of the distribution parameters of shear modulus G, and concentrated load F as random variables to
other variables have small effect on the failure probability. Besides, study how the material properties of the flap and the concentrated
it can be noted that the sensitivities with respect to μa and μt are load F affect the failure probability of the whole structure. From the
negative, meaning that an increase of these two parameters would results in Table 5, it can be seen that the absolute values of sensi-
decrease the failure probability. An increase of σa and σt would tivities with respect to distribution parameters of x3 are much larger
increase the failure probability since their corresponding sensitiv- than those of the other two variables, which indicates that the dis-
ities are positive. Therefore, in the preliminary design step of the tribution parameters of elastic modulus E and shear modulus G
I-beam of the automobile front axle, under the guidance of the have little influence on the failure probability, while a small change
sensitivity information obtained, it is possible to get the design op- of the distribution parameters of Fλ can lead to a large change of
timization efficiently based on reliability. the failure probability. It can also be seen from the last column of
Table 5 that the sensitivities are all positive, so an increase of the
An Aircraft Inside Flap distribution parameters of x3 would increase the failure probability.
From the discussion above, it could be concluded that the concen-
An aircraft inside flap is considered whose finite element model is trated load F, which is transformed from the aerodynamic load, has
constructed in Patran and is shown in Fig. 4. The structure suffers great impact on the failure probability of this aircraft inside flap.
from aerodynamic load, which is transformed into concentrated Besides, an increase of the load F through an increase of the mean
load F applied to the nodes of the finite element model (FEM). of Fλ , and an increase of uncertainty of F through an increase of the
Taking the most dangerous working condition into consideration, standard deviation of Fλ , would both increase the risk of the struc-
the failure is defined as the maximum of the strain of all the nodes ture. The above conclusions are consistent with real situations, so
not exceeding an admissible maximal strain. The performance more attention should be paid to the aerodynamic load in the
function is given by reliability-based design optimization of the aircraft inside flap.
gðxÞ ¼ 32.5 − fðE; G; FÞ ð40Þ

where E and G =elastic modulus and shear modulus, respectively. Conclusions

As an instrumental random variable Fλ represents the randomness
of the load applied to the nodes, the value of F is F ¼ ð1 þ Fλ ÞF0 , Since the Kriging model is sometimes used to analyze structures in
where F0 = constant nominal value. All variables are mutually in- engineering, this paper aims to analytically obtain the sensitivity of
dependent normal variables whose distribution information is given failure probability with respect to distribution parameters of input
in Table 4. variables on this emulator. For a TPBF-based Kriging model, ker-
This is an example with implicit performance function, which is nel function is used to get the sensitivity of statistical moments of
expressed by a FEM. It is a time-consuming process to analyze the the performance function with respect to distribution parameters
FEM using the MCS method and the computational burden is un- of inputs analytically. Then by employing fourth-moment based
affordable. Therefore, the IS method is only applied for compari- reliability sensitivity analysis method, the sensitivity of failure
son. The sensitivity results are listed in Table 5. probability can be computed approximately.
Similar to what has been seen in the previous examples, the The Kriging model–based analytical method for reliability
results computed by the proposed method can agree with their sensitivity analysis is illustrated by numerical and engineering
examples, and the results can well verify the correctness and
high precision of the proposed method. Furthermore, for engi-
Table 4. Distribution Information of Input Variables of the Aircraft Inside neering problems involving implicit performance functions, the
Flap Kriging model is sometimes selected as an emulator. Now the pro-
posed analytical method can be used to compute sensitivities
Variable Sign Mean Standard deviation
directly when the emulator model is available, which is computa-
EðKN=m2 Þ x1 72,450 1,449 tionally cheap and timesaving, and can provide important informa-
GðKN=m2 Þ x2 27,236.8 544.736 tion for reliability optimization conveniently. The advantages of the
Fλ x3 0 0.1
proposed method show its good engineering applicability.

© ASCE 04015016-7 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 Appendix I. Derivation of the Sensitivity of Statistical Moments of Performance Function

In Eq. (8), the definition expression of α1g is α1g ¼ ∫ Rn gðxÞX ðxÞdx; based on it, the sensitivity of the first moment of the performance
function with respect to distribution parameters of random variables can be derived as follows:
Z
∂α1g ∂
ðkÞ
¼ ðkÞ
gðxÞfX ðxÞdx
∂θxl ∂θxl Rn
Z
∂f ðxÞ
¼ gðxÞ XðkÞ dx
Rn ∂θxl
Z
∂f X ðxÞ 1
¼ gðxÞ · fX ðxÞdx
X ðxÞ
ðkÞ
R n
∂θxl f
Z
¼ gðxÞ · kθðkÞ · fX ðxÞdx
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

xl
Rn
¼ E½gðxÞ · kθðkÞ  ð41Þ
xl

Assuming that V g = variance of performance function, then V g ¼ α22g . On the basis of α2g ¼ ½∫ Rn ðgðxÞ − α1g Þ2 f X ðxÞdx1=2 , it follows that
Z
∂α2g 1 ∂V g 1 ∂
¼ ¼ ðgðxÞ − α1g Þ2 fX ðxÞdx
∂θkxl α2g ∂θkxl α2g ∂θkxl Rn
Z Z 
1 2 ∂fX ðxÞ ∂ðgðxÞ − α1g Þ2
¼ ðgðxÞ − α1g Þ dx þ fX ðxÞdx
α2g Rn ∂θkxl Rn ∂θkxl
Z Z 
1 2
¼ ½kθðkÞ ðgðxÞ − α1g Þ fX ðxÞdx − 2 ðgðxÞ − α1g ÞE½gðxÞ · kθðkÞ fX ðxÞdx
2α2g xl xl

1
¼ fE½gðxÞ2 · kθðkÞ  − 2α1g E½gðxÞ · kθðkÞ g ð42Þ
2α2g xl xl

Similarly, α3g ¼ α13 ∫ Rn ðgðxÞ − α1g Þ3 fX ðxÞdx, and it follows that

2g
 Z 
∂α3g ∂ 1 3
ðkÞ
¼ ðkÞ 3
ðgðxÞ − α1g Þ f X ðxÞdx
∂θxl ∂θxl α2g Rn
  Z Z
∂ 1 3 1 ∂
¼ ðkÞ 3
· ðgðxÞ − α 1g Þ f X ðxÞdx þ · ðgðxÞ − α1g Þ3 fX ðxÞdx
∂θxl α2g Rn α32g ∂θðkÞ xl R n

Z Z   Z 
3 ∂α2g 3 ∂f X ðxÞ 3
∂α1g 2
¼− 4 x ðgðxÞ − α1g Þ f X ðxÞdx ðgðxÞ − α1g Þ dx−3 ðkÞ ðgðxÞ − α1g Þ f X ðxÞdx
α4g ∂θðkÞ
xl Rn
ðkÞ
∂θxl ∂θxl
   
1 3
3α3g 3α1g 2
3α3g α1g 3α1g 3ðE½gðxÞ2  − α21g Þ
2
¼ 3 E½gðxÞ · kθðkÞ  − þ 3 E½gðxÞ · kθðkÞ  þ þ 3 − E½gðxÞ · kθðkÞ  ð43Þ
α2g xl 2α22g α2g xl α22g α2g α32g xl

Additionally, α4g ¼ α14 ∫ Rn ðgðxÞ − α1g Þ4 f X ðxÞdx, and thus

2g
 Z 
∂α4g ∂ 1 4
ðkÞ
¼ ðkÞ 4
ðgðxÞ − α 1g Þ f X ðxÞdx
∂θxl ∂θxl α2g Rn
  Z Z
∂ 1 4 f ðxÞdx þ 1 · ∂
¼ ðkÞ 4
· ðgðxÞ − α 1g Þ X ðgðxÞ − α1g Þ4 fX ðxÞdx
∂θxl α2g Rn α42g ∂θðkÞxl Rn
Z Z   Z 
4 ∂α2g 4 1 ∂fX ðxÞ 4
∂α1g 3
¼− 5 ðgðxÞ − α 1g Þ f X ðxÞdx þ ðgðxÞ − α 1g Þ dx − 4 ðgðxÞ − α1g Þ f X ðxÞdx
α2g ∂θðkÞxl
α42g ∂θxl
ðkÞ ðkÞ
∂θxl
 2 
1 4α1g 6α1g 2α4g
¼ 4 E½gðxÞ4 · kθðkÞ  − 4 E½gðxÞ3 · kθðkÞ  þ − E½gðxÞ2 · kθðkÞ 
α2g xl α2g xl α42g α22g xl

 
4α1g α4g 4α31g 4ðE½gðxÞ3  − 3α1g E½gðxÞ2  þ 2α31g Þ
þ − 4 − E½gðxÞ · kθðkÞ  ð44Þ
α22g α2g α42g xl

Appendix II. Derivation Process from E
gx m · k θk  n  to E
gxm · x kl 

xl

For normal distributions, the distribution parameters of the input random variable xl are its mean μxl and standard deviation σxl , so according
to Eq. (4), the expressions of kernel functions corresponding to μxl and σxl , denoted kμx and kσx , can be derived as
l l

© ASCE 04015016-8 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

    
1 xl − μ x l 1 xl − μ x l 2
kμx ¼ · ð45Þ E½gðxÞm · kσx  ¼ E gðxÞm · · −1
l σ xl σ xl l σ xl σ xl
1 2μx
¼ 3
E½gðxÞm · x2l  − 3 l E½gðxÞm · xl 
   σ xl σxl
1 x l − μx l 2
kσ x ¼ · −1 ð46Þ μ2xl − σ2xl
l σ xl σ xl þ E½gðxÞm  ð48Þ
σ3xl

In the condition of n ¼ 1, substituting into E½gðxÞm · ðkθðkÞ Þn 

xl
Appendix III. Derivation of D 1;il and E 1;il for
the expressions of kernel functions derived above [Eqs. (45) and Kriging Model
(46)], E½gðxÞm · ðkθðkÞ Þn  is written as follows:
xl
The Gaussian correlation function used in this paper is Ril ðxl Þ ¼
  exp½−τ l ðxl − xil Þ2 , and the PDF of normal random variable xl is
1 xl − μxl
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

E½gðxÞm · kμx  ¼ E gðxÞm · · fXl ðxl Þ ¼ pffiffiffiffi

1
exp½−ðxl − μxl Þ2 =2σ2xl . For normal distributions,
l σxl σxl 2πσxl
substituting the Gaussian correlation function and the PDF into
1
¼ ðE½gðxÞm · xl  − μxl E½gðxÞm Þ ð47Þ the definition expressions of D1;il and E1;il [Eqs. (26) and (27)],
σ2xl the expressions of D1;il and E1;il can be written as follows:

Z ∞
D1;il ¼ xl Ril ðxl Þf Xl ðxl Þdxl
−∞
Z ∞
1
¼ pffiffiffiffiffiffi xl exp½−τ l ðxl − xil Þ2  exp½−ðxl − μxl Þ2 =ð2σ2xl Þdxl
2πσxl −∞
  sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Z    
1 τl 2 1 2σ2xl τ l þ 1 ∞ 2σ2xl τ l þ 1 2σ2xl τ l xil þ μxl 2
¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp − 2 ðμx − xil Þ × p ffiffiffiffiffi
ffi xl exp − xl − dxl
2σ2xl τ l þ 1 2σxl τ l þ 1 l 2π σ2xl −∞ 2τ 2l 2σ2xl τ l þ 1
 
1 τl 2
2σ2xl τ l xil þ μxl
¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp − 2 ðμxl − xil Þ ·
2σ2xl τ l þ 1 2σxl τ l þ 1 2σ2xl τ l þ 1

2σ2xl τ l xil þ μxl

¼ C1;il · ð49Þ
2σ2xl τ l þ 1

Z ∞
E1;il ¼ xl Ril ðxl ÞfXl ðxl Þdxl
−∞
Z ∞
1
¼ pffiffiffiffiffiffi x2l exp½−τ l ðxl − xil Þ2  exp½−ðxl − μxl Þ2 =ð2σ2xl Þdxl
2πσxl −∞
  sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Z    
1 τl 2 1 2σ2xl τ l þ 1 ∞ 2 2σ2xl τ l þ 1 2σ2xl τ l xil þ μxl 2
¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp − 2 ðμ − xil Þ × p ffiffiffiffiffi
ffi xl exp − xl − dxl
2σ2xl τ l þ 1 2σxl τ l þ 1 xl 2π σ2xl −∞ 2σ2xl 2σ2xl τ l þ 1
   2  
1 τl 2
2σxl τ l xil þ μxl 2 σ2xl
¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp − 2 ðμx − xil Þ · þ 2
2σ2xl τ l þ 1 2σxl τ l þ 1 l 2σ2xl τ l þ 1 2σxl τ l þ 1
 2  
2σxl τ l xil þ μxl 2 σ2xl
¼ C1;il · þ 2 ð50Þ
2σ2xl τ l þ 1 2σxl τ l þ 1

Acknowledgments References

This work was supported by the Natural Science Foundation of Chen, W., Ruichen, J., and Agus, S. (2005). “Analytical variance-based
China (Grant 51175425) and the Research Found for the Doctoral global sensitivity analysis in simulation-based design under uncer-
Program of Higher Education of China (Grant 20116102110003). tainty.” J. Mech. Des., 127(5), 875–886.

© ASCE 04015016-9 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016

 Currin, C., Mitchell, T., Morris, M. D., and Ylvisaker, D. (1991). “Bayesian illustrated with dike ring reliability calculations.” Comput. Phys. Com-
prediction of deterministic functions, with applications to the design mun., 117(1-2), 86–98.
and analysis of computer experiments.” J. Am. Stat. Assoc., 86(416), Rosenblueth, E. (1975). “Point estimation for probability moments.” Proc.
953–963. Nat. Acad. Sci., 72(10), 3812–3814.
Guan, X. L., and Melchers, R. E. (2001). “Effect of response surface param- Rosenblueth, E. (1981). “Two-point estimates in probability.” Appl. Math.
eter variation on structural reliability estimates.” Struct. Saf., 23(4), Modell., 5(5), 329–335.
429–444. Sacks, J., Schiller, S. B., and Welch, W. J. (1989). “Design for computer
Karamchandani, A., and Cornell, C. A. (1991). “Sensitivity estimation experiment.” Technometrics, 31(1), 41–47.
within first and second order reliability methods.” Struct. Saf., 11(2), Sakata, S., Ashida, F., and Zako, M. (2003). “Structural optimization using
95–107. kriging approximation.” Comput. Methods Appl. Mech. Eng., 192(7-8),
Kaymaz, I. (2005). “Application of Kriging method to structural reliability 923–939.
problems.” Struct. Saf., 27(2), 133–151. Wang, P., Lu, Z. Z., and Tang, Z. C. (2013). “An application of the Kriging
Kaymaz, I., and McMahon, C. A. (2005). “A response surface method method in global sensitivity analysis with parameter uncertainty.” Appl.
based on weighted regression for structural reliability analysis.”
Math. Modell., 37(9), 6543–6555.
Probab. Eng. Mech., 20(1), 11–17.
Zhang, L. G., Lu, Z. Z., Cheng, L., and Fan, C. Q. (2014). “A new method
Downloaded from ascelibrary.org by Guangzhou University on 12/08/24. Copyright ASCE. For personal use only; all rights reserved.

Lataillade, A. D, et al. (2002). “Monte carlo method and sensitivity

for evaluating Borgonovo moment-independent importance measure
estimations.” J. Quant. Spectrosc. Radiat., 75(5), 529–538.
with its application in an aircraft structure.” Reliab. Eng. Syst. Saf.,
Lu, Z. Z., and Song, J. (2010). “Reliability sensitivity by method of mo-
132(2), 163–175.
ments.” Appl. Math. Modell., 34(10), 2860–2871.
Lucifredi, A., Mazzieri, C., and Rossi, M. (2000). “Application of multi- Zhang, L. G., Lu, Z. Z, and Wang, P. (2015). “Efficient structural reliability
regressive linear models, dynamic Kriging models and neural network analysis method based on advanced Kriging model.” Appl. Math.
models to predictive maintenance of hydroelectric power systems.” Modell., 39(2), 781–793.
Mech. Syst. Signal Process., 14(3), 471–494. Zhao, Y. G., and Ang, A. H. (2003). “System reliability assessment by
Melchers, R. E. (1989). “Importance sampling in structural system.” Struct. method of moments.” J. Struct. Eng., 10.1061/(ASCE)0733-9445
Saf., 6(1), 3–10. (2003)129:10(1341), 1341–1349.
Melchers, R. E., and Ahammed, M. (2004). “A fast approximate method for Zhao, Y. G., and Ono, T. (2001). “Moment methods for structural reliabil-
parameter sensitivity estimation in Monte Carlo structural reliability.” ity.” Struct. Saf., 23(1), 47–75.
Comput. Struct., 82(1), 55–61. Zhao, Y. G., and Ono, T. (2000). “New point estimates for probability mo-
Millwater, H. (2009). “Universal properties of kernel functions for prob- ments.” J. Eng. Mech., 10.1061/(ASCE)0733-9399(2000)126:4(433),
abilistic sensitivity analysis.” Probab. Eng. Mech., 24(1), 89–99. 433–436.
Roger, M., Jan, C. M., and Noortwijk, V. (1999). “Local probabilistic sen- Zhao, Y. G., and Ono, T. (2004). “On the problems of the fourth moment
sitivity measures for comparing FORM and Monte Carlo calculations method.” Struct. Saf., 26(3), 343–347.

© ASCE 04015016-10 J. Eng. Mech.

J. Eng. Mech., 2015, 141(8): 04015016