Application of Rejection Sampling based methodology to variance based parametric sensitivity analysis
Application of Rejection Sampling based methodology to variance based parametric sensitivity analysis
art ic l e i nf o a b s t r a c t
Article history: For estimating the effect of uncertain distribution parameter on the variance of failure probability
Received 11 November 2014 function (FPF), the map from distribution parameters to FPF is built and the high efficient approximation
Received in revised form form is extended to solve the parametric variance-based sensitivity index. Then the parametric variance-
2 April 2015
based sensitivity index can be firstly expressed as the moments of the FPF, and the FPF is approximated
Accepted 25 April 2015
by a product of the univariate functions of the distribution parameters, on which the moments of the FPF
Available online 5 May 2015
approximated by the univariate functions can be easily evaluated by the Gaussian integration using the
Keywords: values of the FPF at the Gaussian nodes. Thus the primary task of evaluating the parametric variance-
Failure probability function (FPF) based sensitivity is transformed to calculate the FPF at Gaussian nodes of the univariate functions, for
Efficient approximation
which Monte Carlo (MC), Extended Monte Carlo (EMC) and Rejection Sampling (RS) are employed and
Parametric variance-based sensitivity index
compared here. Only one set of samples of inputs are needed in either EMC or RS. Several numerical and
Extended Monte Carlo (EMC)
Rejection Sampling (RS) engineering examples are presented to verify the accuracy and efficiency of the proposed approximate
methods. Additionally, the results also reveal the virtue of RS which can be more accurate and more
unlimited than EMC.
& 2015 Elsevier Ltd. All rights reserved.
1. Introduction model [3,8], and the fuzzy set model [9], among which the random
model is especially of interest in engineering applications [10,11],
In contemporary engineering design, an increasing number of namely epistemic uncertainty is also described by probability
model evaluations, especially for finite-element models, have been theory. Then both the two kinds of uncertainties can be character-
established to simulate the behavior of the structures. Whereas, ized by probability theory. If sufficient data is available, the alea-
the large number of unknown uncertainties often preclude the tory model inputs can be considered as random variables with
designers and analysts from identifying useful information and constant distribution parameters, whereas if only sparse data is
making reasonable decisions of model behavior. presented, then the aleatory model inputs can also be considered
Commonly, two kinds of uncertainties are involved in engi- as random variables but with uncertain distribution parameters.
neering reality, i.e. aleatory uncertainty and epistemic uncertainty Note that the distribution parameters can be represented by either
[1–3]. The former one mainly describes the inherent variability confidence intervals in the presence of classical statistic method,
associated with a structural system, which is referred to as irre- or probability density functions (PDFs) under the condition that
ducible, objective uncertainty [3]. While the latter one mostly the Bayesian statistic method is employed [12]. In this paper, the
results from lack of knowledge of fundamental phenomena and is input distribution type is assumed to be known and the distribu-
related to our ability to understand, measure and describe the tion parameters are considered as random variables and repre-
system under study [4], which can be considered as reducible by sented by PDFs.
collecting more information [5]. Employing the probability theory, Sensitivity analysis (SA) is widely used in engineering reliability
the aleatory uncertainty can be modeled as random variable with design, which can be classified into three groups: local SA, global
probability density function (PDF). Whereas when dealing with SA (GSA) and regional SA. Among them, GSA is the most potential
the epistemic uncertainty, more than one theory has been devel- in engineering applications, and the indicators can also be called
oped, such as the random model [6–8], non-probability convex set uncertainty importance measures [13] which have attracted lots of
research interest in the literature. At present, a number of SA
techniques with only aleatory uncertainty, such as a variance-
n
Corresponding author.
based importance method [14–19] for single output or multiple
E-mail addresses: [email protected] (L. Cheng), output under a statistic or stochastic process, transformation
[email protected] (Z. Lu), [email protected] (L. Zhang). invariance property and kinds of the moment-independent importance
http://dx.doi.org/10.1016/j.ress.2015.04.020
0951-8320/& 2015 Elsevier Ltd. All rights reserved.
10 L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18
n
measures [20–28], the elementary effect method and the correspond- probability density function (PDF) f X ðxj θ Þ [41] and the output Y is
ing applications [29,30], the derivative-based method [31,32], the a function of X. Consequently, the statistic characteristics of Y,
regional analysis techniques and their evolutions [33–36] and the such as expectation, variance and failure probability and so on,
parametric method [37–39], have been developed by different rese- only depend on θ and can be regarded as functions of θ. Suppose
archers. However, it is also essential for analysts to measure the the relationship between failure probability and distribution
contribution of distribution parameters to the probabilistic response parameters can be represented as P f ¼ ψ ðθÞ, the square integrable
such as the expectation function, standard deviation function and function ψ ðθÞ defined in the hypercube parameters space H n can
failure probability function (FPF) of the model output. be obtained following the high dimensional model representation
In the past decades, plenty of methods have been developed for (HDMR) by Sobol [15]:
estimating the variance-based sensitivity indices including the
X
n X
n
Monte Carlo simulation (MC) [40], which always involves a ‘triple- ψ ðθÞ ¼ ψ 0 þ ψ i ðθ i Þ þ ψ ij ðθi ; θj Þ þ ⋯ þ ψ 1;2;:::;p ðθ1 ; θ2 ; :::; θn Þ
loop’ evaluation; approximate methods by Krzykaca-Hausmann i¼1 1riojrn
[41], which is a single-loop procedure with the computational cost ð1Þ
increasing with the aleatory dimensionality; Fourier amplitude
sensitivity test [42,43]; the metamodel method [44–46], where where
Wang [44] applied the Kriging method to compute the sensitivity ψ 0 ¼ EðP f Þ
index and is a double-loop method; Extended Monte Carlo
ψ i ¼ EðP f j θi Þ EðP f Þ ð2Þ
method (EMC), where Wei [47] used EMC to estimate the prob-
ψ ij ¼ EðP f j θi ; θj Þ ψ i ψ j EðP f Þ
abilistic response function; separation method, where the aleatory
and epistemic uncertainties in a system is separated by introd-
EðP f Þ and EðP f j dÞ are the unconditional and conditional expec-
ucing a standard uniformly auxiliary variable [48], and so on.
tations of failure probability function ψ ðθÞ respectively. High order
However, all those methods still involves large computational
items can be obtained analogously.
burden. In this paper, the idea of Zhang [49] is introduced to
According to the basic idea of Sobol's measure, the total
parametric sensitivity analysis, where we treat the failure prob-
variance of failure probability can be decomposed into:
ability function (FPF) and the distribution parameters as the
corresponding output–input, thus the map is constructed simi- X
n X
n
larly. Then the FPF is approximated by a product of low dimen- V¼ Vi þ V ij þ ⋯ þ V 1;2;…;n ð3Þ
i¼1 1riojrn
sional functions of distribution parameters [50] and the par-
ametric sensitivity index can be expressed as a form of the where V is the total variance of the failure probability function,
moments of the univariate functions of distribution parameters. and
For symmetric PDFs, based on N-point Gaussian quadrature, only
ðn 1ÞN þ 1 failure probability evaluations (FPEs) are essential (n is V i ¼ Vðψ i Þ ¼ V½EðP f j θi Þ
ð4Þ
the dimensionality of distribution parameters). Then the primary V ij ¼ Vðψ ij Þ ¼ V½EðP f j θi ; θj Þ V ½EðP f j θi Þ V½EðP f j θj Þ
aim turns to estimating the FPF at Gaussian nodes corresponding
to univariate functions, for which three methods are introduced in are the first order and the second order variance contributions of
this paper: Monte Carlo simulation (MC); Extended Monte Carlo the distribution parameters to the failure probability function
technique (EMC); Rejection Sampling Technique (RS). EMC tech- respectively.
nique shares the similar idea with the importance sampling (IS) Thus the variance-based sensitivity index can be defined as
procedure [47,51], whereas its efficiency is highly affected by the V θi ðEθ i ½P f j θi Þ
sampling PDF. Then Rejection Sampling (RS) [52,53] is introduced Sθ i ¼ ð5Þ
VðP f Þ
to estimate the FPF with only one set of model evaluations. RS
reuses the original generated samples and selects proper points at Although the variance decomposition of P f is conceptually
given distributions. Comparing with MC, EMC and RS can further simple, its computation in a general setting is a challenging task
decrease the computational effort for they only need one set of as it involves a ‘triple-loop’ high dimensional integrations. Thus
samples. Once the FPF is obtained, the parametric sensitivity index the computational cost can be a large burden for traditional Monte
can be evaluated based on the approximate estimation form, i.e. Carlo (MC) simulation, especially in case of a complex model. Thus
MCA, EMCA and RSA methods are proposed in this paper. simple approximations to substitute and simplify the evaluation of
This paper is organized as follows: Section 2 reviews the Sθi are developed.
definition of the variance-based parametric sensitivity index based
on FPF. Then the idea of Zhang is introduced to construct an
approximation of the sensitivity index in Section 3. Section 4
3. An approximation for the parametric sensitivity
applies the MC, EMC and RS techniques to estimate the FPF
respectively for sake of calculating the sensitivity index. The three
The most popular approximation is called Cut-HDMR, in which
methods are introduced and compared in Section 4. Then the
the corresponding deterministic model can be written as a form of
proposed methods are applied to two numerical and two engi-
summation of univariate functions of input variables. Subse-
neering examples in Section 5. Section 6 gives conclusions.
quently, Zhang proposed an improvement on Cut-HDMR, i.e. mul-
tiplicative form of approximation of the original function (M-DRM)
[50]. Similarly, since the failure probability function is always
2. Review of definition of importance measure
positive, P f ¼ ψ ðθÞ can be transformed into a logarithmic form:
Suppose the distribution parameter uncertainty is involved in φðθÞ ¼ lnðP f Þ ¼ ln ½ψ ðθÞ ð6Þ
structure systems, then the performance function of a model can
be given as Y ¼ gðX; θÞ, where X ¼ ðX 1 ; X 2 ; …; X m Þ is the vector of Then, the Cut-HDMR expansion of φðθÞ is expressed as:
input variables and θ ¼ ðθ1 ; θ2 ; :::; θn Þ is the vector of parameters. X
n X
Holding the parameters θ at a reference value θ , then the
n
φðθÞ ¼ φ0 þ φi ðθi Þ þ φij ðθi ; θj Þ þ… ð7Þ
uncertainty of input variables can be described by conditional i¼1 1riojrn
L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18 11
where and uniform PDFs, thus the number of FPEs can be written as:
8
>
>
φ0 ¼ ln ½ψ ðcÞ X
n
>
< φ ðθ Þ
i i ¼ ln ½ψ ðc1 ; …; ci 1 ; θi ; ci þ 1 ; …; cn Þ φ0 NFPEs ¼ 1 þ ðN k 1Þ ð17Þ
> φij ðθ i ; θ j Þ
> ¼ ln ½ψ ðc1 ; …; ci 1 ; θi ; ci þ 1 ; …; cj 1 ; θj ; cj þ 1 ; …; cn Þ φi ðθi Þ φj ðθj Þ φ0 k¼1
>
:
…
One can find that once the ρk and ωk are obtained, no extra
ð8Þ
FPEs are essential to complete the sensitivity analysis. Then the
c ¼ ½c1 ; c2 ; …; cn T is the mean vector of the uncertain distribution FPEs for the whole sensitivity analysis based on the approximate
parameters. Here, the approximate representation of FPF is estimation in Eq. (11) is N FPEs .
obtained by retaining only up to the first order components of
Eq. (7):
X
n
4. Methods for estimation of the sensitivity index
φðθÞ ln ½ψ ðc1 ; …ci 1 ; θi ; ci þ 1 ; …; cn Þ ðn 1Þln ½ψ ðcÞ
i¼1
Thus the primary task for evaluating Sθi is transformed to
X
n
calculate the FPF P f ðθ Þ ¼ ψ ðθk ; c k Þ corresponding to the set of
G
¼ ln ½ψ ðθi ; c i Þ ðn 1Þln ½ψ ðcÞ ð9Þ
distribution parameters θ determined by Gaussian Integration
G
i¼1
Grid, and P f ðθ Þ is expressed as the product of the univariate
G
where c i ¼ ½c1 ; …; ci 1 ; ci þ 1 ; …; cn T .
functions. Three methods are introduced here to estimate the FPF
By inverting the transformation (Eq. (6)), the original FPF can
in this section, i.e. Monte Carlo simulation (Section 4.1, MC),
be written as:
Extended Monte Carlo simulation (Section 4.2, EMC) and the
1 n n Rejection Sampling technique (Section 4.3, SA).
P f ¼ ψ ðθÞ ¼ exp½φðθÞ ψ ðcÞ ∏ ψ ðθ i ; c i Þ ð10Þ
i¼1
Thus, the FPF is approximated by a product of the univariate 4.1. Monte Carlo simulation
functions of distribution parameters. Based on M-DRM, an approx-
imate estimation of the variance-based sensitivity index of the By definition, P f ðθ Þ can be expressed as:
G
input variables on output response is given in [49]. Analogous to Z
[49], the estimation of parametric sensitivity index in Eq. (5) can P f ðθ Þ ¼
G
I F ðxÞf ðxj θ Þdx ¼ Ef ðxj θG Þ ðI F ðxÞÞ
G
ð18Þ
be approximated as
(
Vi ωi =ρ2i 1 1 gðxÞ r 0
Sθ i ¼ ¼ ð11Þ where I F ðxÞ ¼ is the failure indicator function.
V ð∏nk ¼ 1 ωk =ρ2k Þ 1 0 gðxÞ 4 0
where ρk and ωk are denoted as the mean and mean square of the Eq. (18) indicates that P f ðθ Þ can be expressed as the expecta-
G
X
Nk
ωk wkl ½ψ ðc1 ; …; ck 1 ; θkl ; ck þ 1 ; …; cn Þ2 ð15Þ 4.2. Extended Monte Carlo simulation
l¼1
where θkl and wkl are the lth Gaussian Abscissa and its corre- However, we want to further reduce the computational burden
sponding weight determined by the probability distribution of θk . of MC. Thus the basic idea of importance sampling (IS) procedure
And N k is the number of Gaussian nodes corresponding to θk . is introduced to the calculation of FPF. Suppose now we have a
n
In the proposed method, we assume that for symmetric distr- joint PDF hx ðxj θ Þ of model input vector with constant distribution
n n n n
ibutions, the middle nodes in n Gaussian quadrature coincide with parameter vector θ ¼ ðθ1 ; θ2 ; …; θn Þ. Then the FPF can be derived
the mean value, thus all the n Gaussian formulas share one as:
common node, i.e. c ¼ ½c1 ; c2 ; …; cn T . Then in a N k -node Gauss Z !
I F ðxÞf x ðxj θ Þ I F ðxÞf x ðxj θ Þ
G G
n
P f ðθ Þ ¼ θ
G
quadrature, the total number of failure probability evaluations n h x ðxj Þdx ¼ E h n ð20Þ
hx ðxj θ Þ hx ðxj θ Þ
(FPEs) for calculating ρk and ωk is merely
n
X
s X
n where hx ðxj θ Þ is the selected importance sampling density of X
N FPEs ¼ 1 þ ðN k 1Þ þ Nk ð16Þ n
with constant distribution parameters θ . Eq. (20) indicates that
k¼1 k ¼ sþ1
failure probability function P f ðθ Þ can be expressed as the expec-
G
n
tation of the function I F ðxÞf x ðxj θ Þ=hx ðxj θ Þ. Given a set of model
G
where s denotes the number of variables with the symmetric
ðjÞ ðjÞ ðjÞ ðjÞ
distribution, and n s is the number of other distributions. In our input vector samples x ¼ ðx1 ; x2 ; :::; xm Þðj ¼ 1; 2; …; MÞ generated
n
by the joint PDF hx ðxj θ Þ, the failure probability function P f ðθ Þ
G
work, we mainly deal with symmetric PDFs, such as normal PDFs
12 L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18
can be estimated by: xð2Þ ; …; xðMÞ g and pick out the corresponding I F ðxðkÞG Þ. Then
P f ðθ Þ can be evaluated.
G
1 X I F ðxðjÞ Þf x ðxðjÞ j θ Þ
M G
Step 3: Iterate Step 2 for NFPEs times.
P^ f ðθ Þ ¼
G
n ð21Þ
Mj¼1 hx ðxðjÞ j θ Þ
From the above steps one can find that at each given θ , we
G
As we can see from Eq. (21), once the M samples are generated,
n only need to select proper samples xG from the original samples x
I F ðxÞ and hx ðxj θ Þ can be calculated. Then given N FPEs different
and their corresponding failure indicator value I F ðxG Þ. Thus in fact,
distribution parameters, we only need to calculate the values of
in the process of estimating the FPF P f ðθ Þ using Rejection
G
f x ðxj θ Þ at each different θ . Comparing Eq. (21) with Eq. (19), it is
G G
Sampling, only one set of samples of input variables are essentially
not hard to find that EMC can avoid repeatedly sampling at each
generated and only M model evaluations are needed.
θG . Then only M samples are generated and only M model
Comparing Section 4.2 with Section 4.3, one can see that only
evaluations are needed to compute the FPF with different θ for
G
M samples are used and only M model evaluations are essential
EMC. Because of the high efficiency of the EMC, it has been widely
needed for both EMC and RS. However, at each given θ , EMC uses
G
used for random parameters to estimate the probabilistic response
all the M original samples to estimate the failure probability;
function. And its efficiency and accuracy mainly depends on the
n whereas RS picks some proper samples out of the M original
selection of the optimal sampling density hx ðxj θ Þ. Thus Rejection
samples each time to calculate the failure probability.
Sampling is introduced in Section 4.3.
Once the FPF is evaluated by the three methods, i.e. MC, EMC
and RS, the parametric sensitivity index Sθi can then be esti-
mated using the approximation in Eq. (11), i.e. the constructed
4.3. Rejection Sampling technique
MCA, EMCA and RSA method. Detailed procedures involve three
steps:
Rejection Sampling is a technique for generating samples of a
desired distribution by sampling from a different distribution.
a. Estimate the FPF P f ðθ Þ ¼ ψ ðθk ; c k Þ using MC (or EMC or RS);
G
The technique has been developed for both random samples and
b. Evaluate the first and second order moment (i.e.ρk and ωk ) of
quasi-random samples such as low discrepancy sequences [52,53].
the FPF.
Then Rejection Sampling [53] is used to compute the FPF.
c. Employ ρk and ωk to calculate the parametric sensitivity index
Following DeGroot and Schervish [52], let f P j θn ðpÞ be a PDF of a
Sθi using Eq. (11).
desired distribution θ for a random variable P, and let f X j θ ðxÞ be
n
some other PDF for a random variable X, with the property that
there exists a constant r, such that rf X j θ ðxÞ Z f P j θn ðxÞ for all x, To clearly demonstrate the main content in this work, an
where x is a realization of X. The rejection technique can then be illustration is given in Fig. 1. It can be seen that the computational
used to generate J samples of f P j θn ðpÞ as shown in the Algorithm cost for evaluating the parametric sensitivity index Sθi using MCA,
following Wang [53]. The algorithm is rigorous techniques for EMCA and RSA methods are N FPEs M, M and M respectively,
performing Rejection Sampling on nested uniform distributions. which can be a great improvement comparing with the direct
However, in practice, we may use the bounds of the desired new ‘triple-loop’ MC method (M M M k) and the Sobol's method
interval to do Rejection Sampling directly by rejecting all points (M M k).
outside the bounds.
(i.e. the Algorithm) to generate xðkÞG of f X j θG ðxÞ from fxð1Þ ; Fig. 1. Illustration of MCA, EMCA and RSA methods.
L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18 13
Table 1
Numerical integration grids of distribution parameters for calculating FPF.
a1 a2 a3 b1 b2 b3 MC EMC RS
uncertainty analysis techniques [54]. It can be written as follows: here (corresponding to Table 1, line 4, a1 ¼ 1:9333; b1 ¼ 2:6180).
2 In Fig. 2, the gray dots represent the original 1000 sample points of
Y ¼ sin X 1 þ 5 sin X 2 þ 0:1X 43 sin X 1
the original distribution U½ π ; π and the black star points
where input variables X 1 , X 2 and X 3 are independent and uni- represent the accepted points corresponding to the new distribu-
formly distributed, i.e. X i Uðai ; bi Þði ¼ 1; 2; 3Þ. Assume that the tion U½ 1:9333; 2:6180. That is to say, when given a new
lower and upper bounds are epistemic parameters following distribution, one can select the proper points by reusing the
another uniform distributions ai U½ π ; π =2ði ¼ 1; 2; 3Þ, bi original samples of the original distribution. Samples of other
U½2π =3; π ði ¼ 1; 2; 3Þ. distributions of inputs can also be generated by the similar way.
Table 1 gives the numerical integration grids for calculating FPF. Once the FPF is obtained as can be seen in Table 1, the
n
Then we want to select the hx ðxj θ Þ selected for EMCA and f X j θ ðxÞ parametric sensitivity index can be computed by the four methods
for RSA methods of each input variable. As we can see, the inputs involved in this paper, which is shown in Fig. 3. One can find that
are obeying uniform distributions in Example 1. For uniform the ranking of uniform parameters (ai and bi ) computed by the
distributions, we often choose the minimum interval that could four method is a1 4 b3 4 a3 4b1 4b2 4 a2 . The results verified the
include all the desired distributions, which can not only provide all accuracy and reasonality of the proposed approximate methods.
the possible points but also avoid wasting unnecessary samples. We have stated that for uniform distributions, we often choose
n
Here both hx ðxj θ Þ and f X j θ ðxÞ are chosen as U½ π ; π . the original PDF as the minimum interval which can include all the
Obviously, 6 symmetric parameters are given and 5 Gaussian desired distributions. And we choose U½ π ; π in Example 1.
nodes are used in Example 1. Then the corresponding FPF by MCA, However, we want to explore the effect of other original PDFs on
EMCA and RSA methods can be computed. Clearly, MCA estimates the relative errors of EMCA and RSA, and we demonstrate the
FPF by traditional Monte Carlo simulation at each given distribution, results in Table 2. As we can see, the RSA often shows lower error
i.e. numbers of cycles are involved. Whereas EMCA evaluates FPF by than the EMCA for arbitrary selected initial PDF around U½ π ; π .
n
Eq. (21) with only one set of samples x generated from hx ðxj θ Þ, That is to say, comparing with EMCA, RSA shows less restricted by
n n
where the ratio values f x ðxj θ Þ=hx ðxj θ Þ at generated inputs x are the selection of the original PDF for uniform input variables.
computed. Similarly, RSA computes FPF by reusing the samples
generated from the original distribution f X j θ ðxÞ. RSA selects the Example 2. Consider a computational model: gðXÞ ¼ 40 18X 1 þ
proper points from the original samples depending on the desired X 22 þ X 2 þ X 23 þ 5X 3 , where X 1 , X 2 and X 3 are three independent
distribution and only uses the chosen samples to compute the FPF. normally distributed variables with parameter uncertainty,
The main difference between EMCA and RSA is that EMCA uses all namely X i Nðμi ; σ 2i Þði ¼ 1; 2; 3Þ, where μi and σ i are also rep-
the samples generated from the original distribution while RSA picks resented by normal distributions μi Nð4; 12 Þði ¼ 1; 2; 3Þ and
the proper samples out of the original samples and rejects other σ i Nð2; 0:52 Þði ¼ 1; 2; 3Þ, respectively. Similarly, the numerical
improper samples. One can clearly see that the FPEs of our proposed integration grids for calculating FPF are given in Table 3. Note that
n
methods are NFPEs ¼ 25. Thus the model evaluations of MCA, EMCA the hx ðxj θ Þ for EMCA and f X j θ ðxÞ for RSA methods of each input
and RSA method are 25 104 , 104 and 104 respectively. variable are both chosen as N½4; 4:52 and N FPEs ¼ 25.
In order to show how Rejection Sampling is used for input In order to show how Rejection Sampling is used for input
variables with uniform distribution, a sequence of 103 samples are variables with normal distribution, a sequence of 103 samples are
shown in Fig. 2. One can clearly see that the original f X ðxÞ is generated with the original PDF N½4; 4:52 , which is obtained by
U½ π ; π , and we take the desired PDF f P ðpÞ as U ½ 1:9333; 2:6180 obeying the 3σ -rule as the EMC. Figs. 4 and 5 show the
14 L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18
Fig. 2. Rejection Sampling for uniform distribution with different upper and lower bound.
Table 3
Numerical integration grid of distribution parameters for calculating FPF.
μ1 μ2 μ3 σ1 σ2 σ3 MC EMC RS
1
0.9
0.8
0.7
0.6
CDF
0.5
0.4
0.3
0.2
0.1
0
-15 -10 -5 0 5 10 15 20 25
x
1
0.9
Table 4
0.8
ΔP f of EMCA and RSA under different selected PDFs.
0.7
0.6
N ¼ 10,000 N½4; 32 N½4; 3:52 N½4; 42 N½4; 4:52 N½4; 52 N½4; 5:52 N½4; 62
CDF
0.5
0.4 EMCA 0.0034 0.0032 0.0056 0.0042 0.0070 0.0114 0.0123
0.3 RSA 0.0068 0.0033 0.0036 0.0032 0.0030 0.0035 0.0024
0.2
0.1
0 where D0 and h1 are the rivet diameter and rivet length in state B,
-15 -10 -5 0 5 10 15 20 25
respectively.
x
After stage II, we assume the top and bottom heads of the
Fig. 5. Rejection Sampling for different standard deviations. formed rivet in state C have the same dimensions, then the volume
Vol2 of the rivet in state C can be written as:
where d and h are the rivet diameter and rivet length in state A,
respectively. π π
Vol2 ¼ D20 t þ 2 D21 H ð24Þ
At the end of stage I, the volume Vol1 of the rivet in state B can 4 4
be calculated as follows:
where t is the whole thickness of the two sheets, the parameters
π D1 and H are the diameter and the height of the driven rivet head
Vol1 ¼ D20 h1 ð23Þ in state C, respectively.
4
16 L. Cheng et al. / Reliability Engineering and System Safety 142 (2015) 9–18
σ max ¼ Kðεy ÞnSHE ð25Þ Input variable d (mm) h (mm) D0 (mm) t (mm) K (MPa)
Mean μd μh μD0 μt μK
where K is the strength coefficient, the parameter nSHE is the strain
Standard deviation 0.5 2 0.051 0.05 10.944
hardening exponent of the rivet material, and εy is the true strain
in the y-direction of the rivet head in its formation. And in our
model, the true strain εy is composed of two parts: the strain εy1
caused in stage I and the strain εy2 caused in stage IIstrain εy can
be expressed as follows: Table 6
Distribution parameters of the mean values of the input variables.
εy ¼ εy1 þ εy2 ð26Þ
Parameter μd μh μD0 μt μK
where εy1 ¼ ln hh1 and εy2 ¼ ln h12H t.
Mean 5 20 5.1 5 547.2
Combining the Eqs. (22)–(26) and assuming that the volume
Standard deviation 0.1 0.5 0.051 0.05 5.472
change of the rivet can be neglected, one can obtain the maximum
squeeze stress for a certain riveting process as follows:
!nSHE
2
d h D20 t
σ max ¼ K ln 2
ð27Þ
2Hd
In this paper, the material of the rivet is 2017-T4 and its strain
hardening exponent is nSHE ¼ 0:15, its ultimate squeeze strength is
σ sq ¼ 580 MPa. Then we can obtain the limit state function as
follows:
g ¼ σ sq σ max ð28Þ
Fig. 9. A planar ten-bar structure: (a) schematic representation; (b) finite element 6. Conclusions
model built in Ansys.
In this paper, we extend the M-DMR based high efficient
approximate form of the variance based sensitivity into the
Table 7 parametric sensitivity analysis. Then the parametric sensitivity
Distribution parameters of the means of the input variables. index can be estimated by three approximate methods, i.e. MCA,
EMCA and RSA.
Parameter μL μAi μE μP 1 μP 2 μL
Firstly, the map of distribution parameters to FPF is con-
Mean 1 0.001 100 80 10 10 structed, then the parametric sensitivity index is similarly
Cov 0.01 expressed as the form of the moments of the FPF, and the FPF
can be approximated by a product of the univariate functions of
the distribution parameter.
Secondly, the Gaussian integration technique is used to calcu-
late the moments of the approximate FPF.
At last, three methods, i.e. MC, EMC and RS, are introduced to
compute the FPF at Gaussian nodes of the univariate functions.
Comparing with the ‘triple-loop’ direct Monte Carlo simulation,
the proposed MCA method only requires NFPEs M model evalua-
tions. And the EMCA and RSA methods can further improve the
computational efficiency of analyzing the parametric importance
with only M model evaluations: EMCA calculates the values of PDF
at the set of samples; RSA selects points from the original samples.
Several examples are used to demonstrate the rationality and
efficiency of the proposed three methods. Besides, the results also
show that comparing with the EMCA method, RSA method can
provide more accurate results with fewer limitations from the
Fig. 10. MCS for sensitivity of all distribution parameters. selection of PDF.
51475370), and the Innovation Foundation for Doctor Dissertation [27] Baucells M, Borgonovo E. Invariant probabilistic sensitivity analysis. Manage
of Northwestern Polytechnical University (CX201509). Sci 2013;59(11):2536–49.
[28] Zhang LG, Lu ZZ, Cheng L, Fan CQ. A new method for evaluating Borgonovo
moment-independent importance measure with its application in an aircraft
References structure. Reliab Eng Syst Saf 2014;132:163–75.
[29] Campolongo F, Cariboni J, Saltelli A. An effective screening design for
sensitivity analysis of large models. Environ Model Softw 2007;22
[1] Der Kiureghuan A, Ditlevsen O. Aleatory or epistemic? Does it Matter? Reliab (10):1509–18.
Eng Syst Saf 2009;31(3):105–12. [30] Ruano MV, Ribes J, Seco A, Ferrer J. An improved sampling strategy design for
[2] Helton JC. Alternative representations of epistemic uncertainty. Reliab Eng application of Morris method to systems with many input factors. Environ
Syst Saf 2004;85:1–10. Model Softw 2012;37(10):103–9.
[3] Guo J, Du XP. Sensitivity analysis with mixture epistemic and aleatory [31] Lambonia M, Iooss B, Popelin A-L, Gamboa F. Derivative-based global sensi-
uncertainties. AIAA J 2007;45(9):2337–49. tivity measures: general links with Sobol' indices and numerical tests. Math
[4] Sun S, Fu GT, Djordjevic S, Khu ST. Separating aleatory and epistemic
Comput Simul 2013;87:45–54.
uncertainties: probabilistic sewer flooding evaluation using probability box. J [32] Sobol IM, Kucherenko S. Derivative based global sensitivity measures and
Hydrol 2012;520-421(14):360–72.
their link with global sensitivity indices. Math Comput Simul 2009;79
[5] Yao W, Chen X, Luo W, Tooren M, Guo J. Review of uncertainty-based
(10):3009–17.
multidisciplinary design optimization methods for aerospace vehicles. Prog
[33] Bolado-Lavin R, Castaings W, Tarantola S. Contribution to the sample mean
Aerosp Sci 2011;47(6):450–79.
plot for graphical and numerical sensitivity analysis. Reliab Eng Syst Saf
[6] Wang P, Lu ZZ, Tang ZC. Importance measure analysis with epistemic
2009;94(6):1041–9.
uncertainty and its moving least squares solution. Comput Math Appl
[34] Tarantola S, Kopustinskas V, Bolado-Lavin R, Kaliatka A, Uspuras E,
2013;66(4):460–71.
Vaisnoras M. Sensitivity analysis using contribution to sample variance plot:
[7] Li LY, Lu ZZ, Li W. State dependent parameter method for importance analysis
application to a water hammer model. Reliab Eng Syst Saf 2012;99:62–73.
in the presence of epistemic and aleatory uncertainties. Sci China Technol Sci
[35] Li LY, Lu ZZ. Regional importance effect analysis of the input variables on
2012;55:1608–17.
failure probability. Comput Struct 2013;12:74–85.
[8] Cheng L, Lu ZZ, Li LY. Regional importance measures based on failure
[36] Wei PF, Lu ZZ, Ruan WB, Song JW. Regional sensitivity analysis using revised
probability in the presence of epistemic and aleatory uncertainties. Proc Inst
mean and variance ratio functions. Reliab Eng Syst Saf 2014;121:121–35.
Mech Eng C: J Mech 2014;228(6):1029–47.
[37] Wu YT, Mohanty S. Variable screening and ranking using sampling-based
[9] Qiu ZP. Convex models and interval analysis method to predict the effect of
sensitivity measures. Reliab Eng Syst Saf 2006;91(6):634–47.
uncertain-but-bounded parameters on the buckling of composite structures.
[38] Lu ZZ, Song SF, Yue ZF, Wang J. Reliability sensitivity analysis by line sampling.
Comput Methods Appl Mech Eng 2005;194:2175–89.
Struct Saf 2008;30(6):517–32.
[10] Moore R, Lodwick W. Interval analysis and fuzzy set theory. Fuzzy Sets Syst
[39] Melchers RE, Ahammed M. A fast approximate method for parameter
2003;135(1):5–9.
sensitivity estimation in Monte Carlo structural reliability. Comput Struct
[11] Lei Cheng, Zhenzhou Lu, Luyi Li. Global sensitivity analysis of fuzzy distribu-
tion parameter on failure probability and its single-loop estimation. J Appl 2004;82(1):55–61.
Math 2014;2014:11 490718. [40] Saltelli A, Annoni P, Azzini I, Campolongo F, Ratto M, Tarantola S. Variance
[12] Hofer E, Kloos M, Krzykacz-Hausmann B, Peschke J, Woltereck M. An based sensitivity analysis of model output design and estimator for the total
approximate epistemic uncertainty analysis approach in the presence of sensitivity indices. Comput Phys Commun 2010;181(2):259–70.
epistemic and aleatory uncertainties. Reliab Eng Syst Saf 2002;77(3):229–38. [41] Krzykacz-Hausmann B. An approximate sensitivity analysis of results from
[13] Sankaraman S, Mahadevan S. Likelihood-based representation of epistemic complex computer models in the presence of epistemic and aleatory uncer-
uncertainty due to sparse point data and/or interval data. Reliab Eng Syst Saf tainties. Reliab Eng Syst Saf 2006;91(10–11):1210–8.
2011;96(7):814–24. [42] Xu C, Gertner GZ. A general first-order global sensitivity analysis method.
[14] Kuo W, Zhu X. Importance measures in reliability, risk, and optimization: Reliab Eng Syst Saf 2008;93(7):1060–71.
principles and applications. Chichester, UK: John Wiley & Sons; 2012. p. 472. [43] Tarantola S, Gatelli D, Mara TA. Random balance designs for the estimation of
[15] Sobol IM. Global sensitivity indices for nonlinear mathematical models and first order global sensitivity indices. Reliab Eng Syst Saf 2006;91(6):717–27.
their Monte Carlo estimates. Math Comput Simul 2001;55(1–3):271–80. [44] Wang P, Lu ZZ, Tang ZC. An application of the Kriging method in global
[16] Saltelli A. Sensitivity analysis for importance assessment. Risk Anal 2002;22 sensitivity analysis with parameter uncertainty. Appl Math Model 2013;37
(3):579–90. (9):6543–55.
[17] Garcia-Cabrejo O, Valocchi A. Global sensitivity analysis for multivariate [45] Ratto M, Pagano A, Young PC. Non-parametric estimation of conditional
output using polynomial chaos expansion. Reliab Eng Syst Saf 2014;126: moments for sensitivity analysis. Reliab Eng Syst Saf 2009;94(2):237–43.
25–36. [46] Buzzard GT. Global sensitivity analysis using sparse grid interpolation and
[18] Cao J, Du FR, Ding ST. Global sensitivity analysis for dynamic systems with polynomial chaos. Reliab Eng Syst Saf 2012;107:82–9.
stochastic input processes. Reliab Eng Syst Saf 2013;118:106–17. [47] Wei PF, Lu ZZ, Song JW. Extended Monte Carlo simulation for parametric
[19] Rocco Sanseverino CM, Ramirez-Marquez JE. Uncertainty propagation and global sensitivity analysis optimization. AIAA J 2014;52(4):867–78.
sensitivity analysis in system reliability assessment via unscented transforma- [48] Sankararaman S, Mahadevan S. Separating the contributions of variability and
tion. Reliab Eng Syst Saf 2014;132:176–85. parameter uncertainty in probability distributions. Reliab Eng Syst Saf
[20] Zhou CC, Lu ZZ, Li LY, Feng J, Wang BT. A new algorithm for variance based 2013;112:187–99.
importance analysis of models with correlated inputs. Appl Math Model [49] Zhang XF, Pandey MD. An effective approximation for variance-based global
2013;37(3):864–75. sensitivity analysis. Reliab Eng Syst Saf 2014;121:164–74.
[21] Borgonovo E. A new uncertainty importance measure. Reliab Eng Syst Saf [50] Zhang XF, Pandey MD. Structural reliability analysis based on the concepts of
2007;92(6):771–84. entropy, fractional moment and dimensional reduction method. Struct Saf
[22] Anderson B, Borgonovo E, Galeotti M, Roson R. Uncertainty in climate change 2013;43(4):751–69.
modeling: can global sensitivity analysis be of help? Risk Anal 2014;34 [51] Norouzi M, Nikolaidis E. Integrating subset simulation with probabilistic re-
(2):271–93. analysis to estimate reliability of dynamic systems. Struct Multidiscip Optim
[23] Song SF, Lu ZZ, Li W, Cui LJ. The uncertainty importance measures of the 2013;48(3):533–48.
structural system in view of mixed uncertain variables. Fuzzy Sets Syst [52] DeGroot M, Schervish M. Probability and statistics. 3rd ed.. Boston, MA:
2014;243:25–35. Addison-Wesley; 2002.
[24] Plischke E, Borgonovo E, Smith CL. Global sensitivity measures from given [53] Wang X. Improving the rejection sampling method in quasi-Monte Carlo
data. Eur J Oper Res 2013;226(3):536–50. methods. J Comput Appl Math 2000;114:231–46.
[25] Luo XP, Lu ZZ, Xu X. A fast computational method for moment-independent [54] Storlie CB, Swiler LP, Helton JC, Sallaberry CJ. Implementation and evaluation
uncertainty importance measure. Comput Phys Commun 2014;185(1):19–27. of nonparametric regression procedures for sensitivity analysis of computa-
[26] Borgonovo E, Tarantola S, Plischke E, Morris MD. Transformation and invar- tionally demanding models. Reliab Eng Syst Saf 2009;94:1735–63.
iance in the sensitivity analysis of computer experiment. J R Stat Soc Ser B [55] Cai ZY. Precision design of roll-forging die and its application in the forming of
2014. automobile front axles. J Mater Process Tech 2005;168(1):95–101.