Structural and Multidisciplinary Optimization (2022) 65:195

https://doi.org/10.1007/s00158-022-03286-x

RESEARCH PAPER

Importance sampling‑based algorithms for efficiently estimating

failure chance index under two‑fold random uncertainty
Fen Li1 · Zhenzhou Lu1 · Kaixuan Feng1 · Xia Jiang1

Received: 12 December 2021 / Revised: 24 May 2022 / Accepted: 24 May 2022 / Published online: 30 June 2022
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022

Abstract
Under condition of two-fold random uncertainty, a key for evaluating failure chance index (FCI) of structure system is estimat-
ing failure probability function (FPF) with respect to arbitrary realization of random distribution parameters. However, the
estimation of FPF in small failure probability problem remains a challenge. For addressing this heavy computational burden
of estimating FCI, importance sampling-based algorithms are proposed in the paper. In the proposed algorithm, by standard
normal transformation of random inputs in performance function, the random distribution parameters are firstly transformed
into the performance function as arguments. And an augmented performance function is established in space spanned by
random inputs and their random distribution parameters. Secondly, by using differential approximation and conditional
probability theory, the proposed algorithms obtain the whole FPF by one numerical simulation of estimating the failure
probability of the augmented performance function, which improves the efficiency of estimating FPF and FCI. Subsequently,
combined with adaptive Kriging model and reduction of candidate sample pool, the proposed algorithms design importance
sampling-based strategies for estimating FPF, which further improves the efficiency of estimating FPF and FCI. Finally, the
accuracy and efficiency of the proposed algorithm are verified by numerical and engineering application examples.

Keywords Random distribution parameters · Failure probability function · Failure chance · Importance sampling · Adaptive
kriging · Reduction of candidate sample pool

Abbreviations Y Model output

FCI Failure chance index Pf (𝜽) Failure probability function
FPF Failure probability function Chf (𝛼) Failure chance index
PDF Probability density function Pr{⋅} Probability operator
FPC Failure primitive chance ux Standard normal input vector
MCS Monte Carlo simulation u𝜽 Standard normal variables corre-
X Random input vector sponding to 𝜽
fX (x) Joint probability density function of u = (ux , u𝜽 ) Expanded standard normal variables
X h(⋅) Importance sampling probability
𝜽 Distribution parameter vector density function
f𝜽 (𝜽) Joint probability density function of I(⋅) Indicator function of the failure
𝜽 domain
𝜽k (k = 1, 2, ⋯ , N𝜽 ) 
kth realization of distribution param- var(⋅) Variance
eter vector cov(⋅) Variation coefficient
U(⋅)  U learning function
𝜋(⋅) Probability classification function
Responsible Editor: Byeng D Youn Pf 𝜀 Normalization coefficient
𝛼corr Correction factor
* Zhenzhou Lu
[email protected]
1
Present Address: School of Aeronautics, Northwestern
Polytechnical University, P.O. Box 120, Xi’an City,
Shaanxi Province 710072, People’s Republic of China

13
Vol.:(0123456789)

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 2 of 18 F. Li et al.

1 Introduction the number of interpolation points. The Bayesian inference

method expresses the FPF as three parts, i.e., the augmented
For engineering structures, the input variability of their geo- failure probability, the conditional PDF of 𝜽 on the failure
metric dimensions, material properties, load environment, domain and the unconditional PDF of 𝜽 . Among them, the
etc. leads to uncertainty in their output performance (stress, augmented failure probability is the failure probability of
strain, life, etc.) (Liu 2002; Xu et al. 2020; Zhou et al. 2021), the structure under simultaneously considering the random-
thus it is necessary to carry out uncertainty analysis on the ness of the inputs and the distribution parameters, and it
structure for obtaining the law of the output performance. can be obtained by one Monte Carlo simulation (Au 2005).
In structural uncertainty analysis, random variables (August The conditional PDF of the distribution parameters on the
et al. 1984) are usually used to describe variability of struc- failure domain can be fitted as a byproduct (Park and Mar-
tural inputs, and the uncertainty of random variables can ron 1990; Ormoneit and White 1999) of the distribution
be completely described by its probability density function parameter failure samples obtained in estimating the aug-
(PDF) (Bayrak and Acar 2018). At this time, the output per- mented failure probability, and the unconditional PDF of 𝜽
formance of the structure also possesses random uncertainty, is preassigned. Therefore, the Bayesian inference method
and the failure probability (Bourinet et al. 2011) can be used requires only one Monte Carlo numerical simulation of the
to quantitatively characterize the reliability degree of the augmented failure probability for solving the FPF, which is
structure. When the sample size of the structure input is significantly more efficient than the interpolation method.
small, the distribution parameter of the PDF of random input However, commonly used probability density fitting meth-
will no longer be a fixed value, but a variable with uncer- ods generally may encounter the difficulty in fitting the tail
tainty. And the PDF can also be used to describe uncertainty characteristics and high-dimensional problems. For alle-
of the distribution parameters. For this situation that inputs viating this difficulty, Ref. (Feng et al. 2019) proposed an
and their distribution parameters are random, it is called improved Bayesian inference method based on differential
two-fold random uncertainty (Peng and Liu 2007; Li and approximation to estimate the FPF. This method transforms
Lu 2015). fitting the conditional PDF in the original Bayesian inference
When the structure has two-fold random uncertainty, its method into estimating the conditional probability, which
failure probability will change as the distribution param- reduces the computational complexity of solving the FPF.
eters. That is, the failure probability of the structure can be At the same time, this improved Bayesian inference method
expressed as a function of the distribution parameter vector is suitable for estimating the FPF under multi-dimensional
𝜽 , which is called the failure probability function (FPF) (Au distribution parameters, and it expands the application range
2005). On the basis of extracting the characteristics of the of the original Bayesian inference method. For the complex
FPF, Ref. (Peng and Liu 2007) proposed a failure primi- structure in actual engineering, it should be noted that the
tive chance (FPC) to quantitatively characterize the reliabil- computational cost required for one Monte Carlo simulation
ity degree of the structural system under two-fold random (MCS) is unaffordable. When solving small failure probabil-
uncertainty. In order to overcome the problem that the tradi- ity problems, even if efficient surrogate models such as Krig-
tional FPC may lead to contradictory conclusions due to its ing are embedded in MCS (AK-MCS), it is also quite time-
lack of self-duality property, Ref. (Li et al. 2021) proposed consuming to update Kriging model by traversing candidate
an improved failure chance index (FCI). From the definition sample pool of MCS. Therefore, it is necessary to study a
of the FCI, it is known that accurately solving the FPF of 𝜽 more efficient method to solve the FPF, so as to reduce the
at any realization is the key to estimate the FCI. computational cost of estimating the FCI.
For solving the FPF, commonly used methods include For improving the efficiency of estimating the FCI, effi-
interpolation method (Jensen 2005; Gasser and Schueller cient algorithms are established in this paper based on the
1997) and Bayesian inference method (Au 2005). In the importance sampling strategy (Au and Beck 1999). First
interpolation method, the failure probability is estimated first of all, the distribution parameters of the random inputs are
at multiple interpolation points of the random distribution transformed into arguments of the performance function by
parameter, and then the expression of the FPF in the entire standardly normalizing the random inputs. Then an aug-
distribution domain of the random distribution parameter mented performance function, where both the random inputs
is obtained through the interpolation formula. This method and their random distribution parameters are arguments, is
needs estimating the failure probability of the structure established. Secondly, based on the augmented performance
under different realizations of random distribution param- function, solving the FPF is transformed into one numerical
eters multiple times, and it results in large computational simulation for estimating the augmented failure probability
cost. In addition, the computational accuracy of this method by differential approximation and conditional probability
is also affected by the selected interpolation formula and theory, on which the FCI is solved. Finally, the importance
sampling and meta-model importance sampling strategies

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 3 of 18 195

are designed for solving the FPF. Moreover, the adaptive |

Pf (𝛼) = {P∗f |Pr{Pf (𝜽) > P∗f } = 𝛼} (3)
Kriging model as well as the candidate sample pool reduc- |
tion strategy is embedded to achieve the high computational
efficiency of estimating the FPF and the FCI. In general, the |
Pf (𝛼) = {P∗f |Pr{Pf (𝜽) < P∗f } = 𝛼} (4)
contributions of this paper are listed as follows: |
where fX (x|𝜽) is the joint PDF of X under the condition
(1) By standard normalization of the random inputs, an of 𝜽 , and Pr{⋅} is probability operator.
augmented performance function including the dis- By definition of the FCI of the structure under two-fold
tribution parameters is constructed for avoiding the random uncertainty, the key of estimating Chf (𝛼) is to solve
double-layer analysis of estimating the FCI. the upper α fractile Pf (𝛼) and lower α fractile Pf (𝛼) of Pf (𝜽).
(2) Two kinds of importance sampling-based algorithms For efficiently solving Pf (𝛼) and Pf (𝛼), the probability distri-
are proposed for efficiently estimating FCI.
bution of Pf (𝜽) is required to be efficiently estimated. Gener-
(3) In order to further improve the efficiency of solving
ally, there are two ways to estimate the probability distribution
the FCI, the candidate sample pool reduction strategy
of Pf (𝜽). One way is to solve the analytical expression of
is inserted in the proposed methods.
Pf (𝜽), and then the probability distribution of Pf (𝜽) can be
deduced by the joint PDF of 𝜽 and the expression of Pf (𝜽). The
The structure of the paper is as follows. Section 2 gives
other way is to deduce the probability distribution of Pf (𝜽) by
a brief description of the analyzed problem. Section 3 pre-
the samples of Pf (𝜽), i.e., estimate the sample realization of
sents the ideas of efficiently estimating the FCI. Section 4
Pf (𝜽) corresponding to each realization of 𝜽. Obviously, the
explains the implementation of the efficient methods based
main difficulty of the former way is that it is almost impossible
on importance sampling strategies for estimating the FCI. In
to accurately obtain the analytical expression of Pf (𝜽) for com-
Sect. 5, the accuracy and efficiency of the proposed methods
plex problems, while the latter way needs to solve the failure
are verified by some examples. Finally, conclusions are sum-
probability samples at each realization of 𝜽 one by one, which
marized in Sect. 6.
is time-consuming. Thus, it is necessary to develop efficient
algorithm for obtaining the probability distribution of Pf (𝜽) in
order to efficiently solve Chf (𝛼) and evaluate the reliability
2 Problem description
degree of structures under two-fold random uncertainty.
Denote performance function as Y = g(X), in which both an
n-dimension input vector X = {X1 , X2 , ⋯ , Xn }T and an n𝜽
3 Ideas for efficiently solving Chf (˛)
-dimension distr ibution parameter vector
𝜽 = {𝜃1 , 𝜃2 , ⋯ , 𝜃n𝜽 }T are random. The random uncertainties
From the analysis of the problem description in Sect. 2, it is
of X and 𝜽 are respectively described by the joint PDF
∏n n𝜽
∏ shown that the key to solve Chf (𝛼) under two-fold random
fX (x) = fXi (xi ) and f𝜽 (𝜽) = f𝜃j (𝜃j ), in which fXi (xi ) and uncertainty is to obtain the probability distribution of Pf (𝜽),
i=1 j=1
f𝜃j (𝜃j ) are marginal PDF of Xi (i = 1, 2, ⋯ , n) and that is, the solution of Chf (𝛼) is transformed into the solu-
tion of the probability distribution of Pf (𝜽). Due to it is very
𝜃j (j = 1, 2, ⋯ , n𝜽 ), respectively. Then, the FCI denoted by
difficult to analytically infer the probability distribution of
Chf (𝛼) of the structure under α confidence can reasonably
Pf (𝜽), this paper constructs an efficient method for estimating
measure the reliability degree of the structure under that the
the probability distribution of Pf (𝜽) from the perspective of
distribution parameters of inputs are all random. The defini-
quickly calculating the sample values of Pf (𝜽). That is, the
tion of Chf (𝛼) is shown as Eq. (1), and the properties of
probability distribution of Pf (𝜽) can be obtained by quickly
Chf (𝛼) refer to Ref. (Li et al. 2021).
solving the sample values of Pf (𝜽), on which the upper and
1 lower fractiles can be obtained from the probability distribu-
Chf (𝛼) = [P (𝛼) + Pf (𝛼)] (1)
2 f tion of Pf (𝜽), and then Chf (𝛼) can be obtained. In fact, Bayes-
ian formula is proposed in Ref. (Li and Lu 2015) to quickly
where Pf (𝛼) and Pf (𝛼) are respectively the upper α fractile
solve Pf (𝜽) under different realizations of 𝜽, and this method
and lower α fractile of the FPF Pf (𝜽), and the definitions of can solve the Pf (𝜽) corresponding to different realizations of
Pf (𝜽), Pf (𝛼) and Pf (𝛼) are shown as follows respectively. 𝜽 only by one Monte Carlo simulation (MCS), but the MCS
method is still time-consuming. In addition, this method must
Pf (𝜽) = Pr{g(x) ≤ 0}=
�g(x)≤0 (2) also approximately solve the PDF of 𝜽 conditional on the fail-
fX (x|𝜽)dx
ure domain, and the computational complexity of estimating
the PDF is higher than that of estimating the probability.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 4 of 18 F. Li et al.

In order to avoid solving the problem of PDF with high standard normal space UX as T x (ux , 𝜽) . By T x (ux , 𝜽) , the
computational complexity and the large computational cost of performance function y = g(x) in X space can be trans-
MCS method, this paper designs a more efficient importance formed into the augmented performance function in the
sampling strategy combined with differential approximation expanded space containing the random distribution param-
and conditional probability theory to solve the samples of eters 𝜽 and the standard normal inputs ux corresponding to
Pf (𝜽) under different realizations of 𝜽 , on which the prob- x, and the expression is shown below.
ability distribution of Pf (𝜽) and Chf (𝛼) of the structure under
two-fold random uncertainty can be efficiently obtained.
y = g(x) = g(Tx (ux , 𝜽)) = gux (ux , 𝜽) (5)
In general, the algorithm in this paper can be divided into
when n-dimensional random inputs are independently
three steps. The first step is to carry out standard normaliza-
normal, i.e., Xi ∼ N(𝜇Xi , 𝜎X2 ) (i = 1, 2, ⋯ , n, 𝜇Xi and 𝜎Xi are
tion of random inputs in order to convert random distribution i
respectively the mean and standard deviation of Xi , and 𝜇Xi
parameters into input arguments of the augmented performance
and 𝜎Xi are the components of 𝜽 ), the transformation in the
function, which constructs the foundation for introducing
above Eq. (5) can be expressed by the following Eq. (6).
importance sampling strategies to estimate failure probability
of the augmented performance function. In the second step, dif- xi = 𝜇Xi + 𝜎Xi uxi (i = 1, 2, ⋯ , n) (6)
ferential approximation transforms the failure probability con-
ditional on a certain realization of random distribution param- In Eq. (6), uxi (i = 1, 2, ⋯ , n) indicates the i - th dimen-
eters as the failure probability conditional on the distribution sional component of the standard normal input vector ux.
parameters falling into the differential region in the vicinity of If the n-dimensional random inputs are not normally
the given realization. Then, the conditional probability theory distributed, T x (ux , 𝜽) can be obtained by Rosenblatt (Rack-
is used to convert the failure probability conditional on the dif- witz and Fiessler 1978) or Nataf (Regis and Anne 2009)
ferential region into the ratio of the intersection probability of transformation. The standard normal transformation is also
failure and distribution parameter falling into the differential implemented for 𝜽 in the augmented performance function
region to the probability of the distribution parameter falling gux (ux , 𝜽). Denote the standard normal variables correspond-
into the differential region. In determining the differential ing to 𝜽 as u𝜽 , and denote the transformation from 𝜽 to u𝜽
region, it is necessary to satisfy the validity of the differential as T 𝜽 (u𝜽 ), then gux (ux , 𝜽) can be transformed into the fol-
approximation and the convergence of estimating the probabil- lowing form.
ity. After determining a rational differential region, the third
step is to introduce the importance sampling strategies into the
y = gux (ux , 𝜽) = gux (ux , T𝜽 (u𝜽 )) = gux ,u𝜽 (ux , u𝜽 ). (7)
augmented performance function, so that the intersection prob- In Eq. (7), y = gux ,u𝜽 (ux , u𝜽 ) is the augmented perfor-
ability of failure and distribution parameter falling into the dif- mance function in standard normal space. To simplify the
ferential region can be solved efficiently. In the third step, two expression, denote u = (ux , u𝜽 ), then the augmented perfor-
importance sampling strategies, i.e., the importance sampling mance function in standard normal space can be simplified
strategy based on design point (Echard et al. 2013) and the as follows.
importance sampling strategy of meta-model (Zhu et al. 2020),
are introduced to estimate the converted probability. Moreo- y = gu (u) (8)
ver, in order to further improve the efficiency of solving the
converted probability, the adaptive Kriging surrogate model
and the candidate sample pool reduction strategy are embedded 4.2 Differential approximation of the FPF
in those two strategies, which results the improvement of the at realization of 
efficiency of solving Pf (𝜽) and Chf (𝛼).
From the analysis in Sect. 3, the FPF Pf (𝜽k ) at different reali-
zation 𝜽k (k = 1, 2, ⋯ , N𝜽 ) of 𝜽 can be expressed as Eq. (9)
4 Implementation of solving Chf (˛) based by the differential approximation proposed in Ref. (Feng
on importance sampling strategies et al. 2019).

Pf (𝜽k ) = Pr{Fx |𝜽 = 𝜽k } = lim Pr{Fx |𝜽 ∈ 𝜽D

k
} (9)
4.1 Standard normalization of random inputs D
Δ𝜽k →0
and augmented performance function
where Fx is the failure domain defined by the performance
function y = g(x), that is, Fx = {g(x) ≤ 0}, 𝜽D
under two‑fold random uncertainty
k is differential
For an n-dimensional random inputs X with joint PDF region in the vicinity of 𝜽k , Δ𝜽D
k is space volume of 𝜽D
k.
n
∏
fX (x) = fXi (xi ), denote the transformation from X to the
i=𝟷

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 5 of 18 195

Under the condition of the space volume Δ𝜽D k → 0 , the

4.3 The importance sampling strategy for solving
FPF Pf (𝜽k ) can be approximated as follows by using the Pf ()
conditional probability theory.
Equation (11) shows that solving the failure probability
Pr{Fx ∩ 𝜽 ∈ 𝜽D
k} Pr{Fu } with respect to y = gu (u) by numerical simula-
Pf (𝜽k ) ≈ Pr{Fx |𝜽 ∈ 𝜽D
k}= (10)
Pr{𝜽 ∈ 𝜽D tion is the key for estimating the complete FPF Pf (𝜽k )
k}
(k = 1, 2, ⋯ , N𝜽 ). After Pf (𝜽k ) is obtained, the probability
After the standard normal transformation is performed distribution of Pf (𝜽) and the upper and lower fractiles at
on the inputs in the performance function y = g(x), and the α confidence level of Pf (𝜽) can be obtained. Furthermore,
corresponding standard normal transformation is also per- the Chf (𝛼) under two-fold random uncertainty can be easily
formed on 𝜽 , Pf (𝜽k ) in Eq. (10) can be equivalently trans- obtained. Obviously, the failure probability Pr{Fu } corre-
formed into Eq. (11). sponding to the augmented performance function y = gu (u)
can be solved by the importance sampling strategy. The fol-
Pf (𝜽k ) = Pr{Fu ||u𝜽 = u𝜽k } ≈ Pr{Fu ||u𝜽 ∈ uD
𝜽
} lowing will give two importance sampling strategies to solve
k

Pr{Fu ∩ u𝜽 ∈ uD }
𝜽 ( ) (11) Pr{Fu } and Pf (𝜽k )(k = 1, 2, ⋯ , N𝜽 ) by the differential region
determined by the failure sample points obtained in impor-
k
= k = 1, 2, ⋯ , N𝜽 ,
Pr{u𝜽 ∈ uD }
𝜽k tance sampling for Pr{Fu }.
where Fu is the failure domain in the expanded standard
normal space u = (ux , u𝜽 ), that is, Fu = {gu (u) ≤ 0}. And
4.3.1 The importance sampling strategy based on design
point for solving Pf ()
u𝜽k is the sample of standard normal distribution parameter
space corresponding to the realization of 𝜽k , that is,
By introducing the importance sampling (IS) PDF hU (u) for
u𝜽k = T −1 (𝜽k ), in which T −1 (⋅) is the inverse transformation
𝜽 𝜽 solving Pr{Fu }, the failure probability Pr{Fu } corresponding
of the standardized normalization transformation T 𝜽 (⋅) from
to the augmented performance function y = gu (u) can be
𝜽 to u𝜽 , uD is the differential region in the vicinity of u𝜽k.
𝜽k expressed Eq. (12) (Melchers 1989).
It can be seen from Eq. (11) that the denominator of Eq.
(11) can be obtained analytically after determining fU (u)
∫ ∫
Pr{Fu } = IFu (u)fU (u)du = IFu (u) h (u)du
u𝜽k = T −1𝜽
(𝜽k ) with respect to 𝜽k (k = 1, 2, ⋯ , N𝜽 ) and the hU (u) U
differential region uD 𝜽k
in the vicinity of u𝜽k . Therefore, it can (12)
be considered that by Eq. (11), the solution of the FPF sam- In Eq. (12), fU (u) is the joint PDF of the standard normal
ple corresponding to the random distribution parameter sam- vector u, IFu (u) is the indicator function of the failure domain
ple is transformed into solving the intersection probability Fu. When u ∈ Fu, IFu (u) = 1, otherwise IFu (u) = 0.
of failure event Fu and event u𝜽 ∈ uD 𝜽k
in the expanded stand- In the IS method, the basic principle of constructing the
ard normal space. IS PDF is to make that the samples with larger contribution
If numerical simulation method is used to solve Pr{Fu }, to the failure probability appear with a greater probability,
then Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
} can be solved by the failure sample thereby the variance of the failure probability estimation
obtained in solving Pr{Fu }. And for arbitrary distribution can be reduced. Since the design point (Au and Beck 2002)
parameter realization 𝜽k (k = 1, 2, ⋯ , N𝜽 ), after converting contributes most to the failure probability, the IS PDF can
𝜽k into the corresponding sample u𝜽k = T −1 (𝜽k ) in the stand- be constructed with the design point as the density center.
𝜽
ard normal space and determining the differential region uD Generate an N-size sample set Su = {u1 , u2 , ⋯ , uN } by
the IS PDF hU (u), and the failure probability Pr{Fu } defined
𝜽k
in the vicinity of u𝜽k , Pr{Fu ∩ u𝜽 ∈ uD } can be obtained as
𝜽k by Eq. (12) can be estimated by Pr{F ̂ u } as follows.
a byproduct by the failure sample obtained during the
numerical simulation for solving Pr{Fu }. Thus, as long as N { }
1 ∑ fU (ul )
the numerical simulation method is used to solve Pr{Fu } and ̂
Pr{Fu } = IFu (ul ) (13)
N l=1 hU (ul )
the differential region uD 𝜽k
is selected by differential approxi-
mation, the complete FPF Pf (𝜽k ) can be obtained as the After using the IS method to obtain the Pr{F
̂ u } of Pr{Fu },
byproduct. the failure sample set denoted as SF = {uF1 , uF2 , ⋯ , uFN }
F
(where NF represents the size of the failure sample set) of
the augmented performance function y = gu (u) can be
obtained at the same time. Next, we will discuss how to use
SF to estimate the probability Pr{Fu ∩ u𝜽 ∈ uD𝜽k
} of the inter-
section that event Fu and the event u𝜽 ∈ uD
𝜽
occur simultane-
k

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 6 of 18 F. Li et al.

ously. Denote FA = Fu ∩ u𝜽 ∈ uD 𝜽k
, i.e., FA represents the w h e n p ∈ [1, 2, ⋯ , m] , d u e t o u[p] ∈ uD 𝜽k
and
intersection of Fu and u𝜽 ∈ u𝜽 , and then the expression of
D
FA = Fu ∩ u𝜽 ∈ u𝜽 , thus IFA (u[p] ) = IFu (u[p] ) holds. But
D
k k
Pr{Fu ∩ u𝜽 ∈ uD 𝜽
} can be rewritten as follows. w h e n q ∈ [m + 1, m + 2, ⋯ , N] , d u e t o u[q] ∉ uD 𝜽k
,
k
IFA (u[q] ) = 0. Based on this, Eq. (16) can be re-expressed as
∫
Pr{Fu ∩ u𝜽 ∈ uD
𝜽 } = Pr{FA } =
k
IFA (u)fU (u)du (14) Eq. (17).
m { }
where IFA (u) is the indicator function of FA, when u ∈ FA, ̂ D 1 ∑ fU (u[p] )
Pr{F u ∩ u 𝜽 ∈ u𝜽k
} = I (u )
N p=1 Fu [p] hU (u[p] ) (17)
IFA (u) = 1, otherwise IFA (u) = 0.
Introduce the IS PDF hU (u) into Eq. (14), and
Pr{FA } = Pr{Fu ∩ u𝜽 ∈ uD } can be solved by Eq. (15). It is shown from Eq. (17) that the solution of
u ∩ u𝜽 ∈ u𝜽k } only needs the indicator function IFu (u[p] )
𝜽 k ̂
Pr{F D

fU (u) of the sample in SuD , and it doesn’t need to call the aug-
∫ ∫
Pr{Fu ∩ u𝜽 ∈ uD
𝜽 }= IFA (u)fU (u)du = IFA (u) h (u)du.
k hU (u) U 𝜽k

m e n t e d p e r fo r m a n c e f u n c t i o n . T h e va r i a n c e
(15)
u ∩ u𝜽 ∈ u𝜽k }) a n d va r i a t i o n c o e f f i c i e n t
̂
Var(Pr{F D
For estimating Eq. (15), it is firstly necessary to determine
u ∩ u𝜽 ∈ u𝜽k }) of Pr{Fu ∩ u𝜽 ∈ u𝜽k } can be
̂
Cov(Pr{F D ̂ D
the size of the differential region uD in the vicinity of u𝜽k . If
𝜽 k
obtained by Eqs. (18) and (19), respectively.
{ m
{ } }
1 1∑ fU2 (u[p] )
̂
Var(Pr{F D
u ∩ u𝜽 ∈ u𝜽k }) = IFu (u[p] ) ̂
− (Pr{F D 2
u ∩ u𝜽 ∈ u𝜽 })
(18)
N−1 N p=1 h2U (u[p] ) k

√
the size of uD is too large, the accuracy of Pr{Fu ||u𝜽 ∈ uD } ̂ D
𝜽k
|
𝜽k Var(Pr{Fu ∩ u𝜽 ∈ u𝜽 })
approximating Pf (𝜽k ) = Pr{Fu |u𝜽k } in Eq. (11) may be low. ̂
Cov(Pr{Fu ∩ u𝜽 ∈ uD
𝜽k }) =
k

| ̂ D
If the size of uD is too small, the size NFA of the sample set Pr{Fu ∩ u𝜽 ∈ u𝜽 }
k
𝜽k
SFA = {u1 A , u2 A , ⋯ , uNA } in FA = Fu ∩ u𝜽 ∈ uD
F F F
will be (19)
𝜽
Theoretically, the larger the sample size m of SuD is, the
FA k

small, which may lead to large variation coefficient of esti- 𝜽k

mating Pr{FA } = Pr{Fu ∩ u𝜽 ∈ uD } by SFA. Therefore, this smaller the variation coefficient Cov(Pr{F
̂ u ∩ u 𝜽 ∈ uD 𝜽k
}) of
𝜽k
section employs the idea in Ref. (Feng et al. 2019) to propose ̂Pr{Fu ∩ u𝜽 ∈ uD } is. Therefore, the value of m can be
𝜽k
the following criteria for adaptively determining the size of increased adaptively to make Cov(Pr{F
̂ u ∩ u𝜽 ∈ u }) less
D
𝜽k
uD𝜽k
. The proposed criteria firstly guarantees that the variation than the required threshold C∗, and it is recommended that
coefficient of the estimated value of Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
} is less C∗ is better to be less than or equal to 0.1.
than the given threshold, and then it makes u𝜽 as small as
D According to the above strategy, the size of uD 𝜽k
can be
k
possible to ensure the accuracy of Pr{Fu ||u𝜽 ∈ uD } approxi- determined adaptively, on which Pr{Fu ∩ u𝜽 ∈ u𝜽 } can be
D
𝜽k k
|
mating Pf (𝜽k ) = Pr{Fu |u𝜽k }. The specific adaptive strategy estimated. Then, the denominator Pr{u𝜽 ∈ uD 𝜽k
} in Eq. (11)
|
is given as follows for determining uD . can be simply obtained by the numerical integration method
𝜽k
First of all, compute the Euclidean distance in Eq. (20). Finally, by Eqs. (11) and (1), Pf (𝜽) and Chf (𝛼)
{d1 , d2 , ⋯ , dN } between each sample in Su = {u1 , u2 , ⋯ , uN } can be solved in turn.
and u𝜽k = T −1 (𝜽k ), denote the distance {d1 , d2 , ⋯ , dN } sorted
∫u𝜽 ∈uD
𝜽
in ascending order as {d[1] , d[2] , ⋯ , d[N] }, and the sample set Pr{u𝜽 ∈ uD
𝜽 }= fU𝜽 (u𝜽 )du𝜽 (20)
k

corresponding to Su is denoted as S[u] = {u[1] , u[2] , ⋯ , u[N] }. 𝜽k

Next collect the first m-size sample in S[u] into the sample Comparing with traditional Monte Carlo simulation
set SuD = {u[1] , u[2] , ⋯ , u[m] } to construct uD 𝜽k
. Based on Eq. method, the IS method can greatly reduce the computational
𝜽k

(15), t he estimated value Pr{F ̂ u ∩ u𝜽 ∈ u } of

D
𝜽k cost for estimating the failure probability Pr{Fu } of the aug-
Pr{Fu ∩ u𝜽 ∈ uD } can be obtained by S[u] shown in Eq. (16). mented performance function gu (u). Combining IS based on
𝜽 k

{ m { } N { }}
1 ∑ fU (u[p] ) ∑ fU (u[q] )
̂
Pr{F D
u ∩ u𝜽 ∈ u𝜽k } = IFA (u[p] ) + IFA (u[q] ) (16)
N p=1
hU (u[p] ) q=m+1
hU (u[q] )

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 7 of 18 195

design point with the adaptive Kriging model (Echard et al. second-moment method (AFOSM) (Du and Chen 2001) or
2013) (AK-IS) can further reduce the computational cost for other optimization methods, and construct IS PDF hU (u) by
obtaining Pr{Fu }. Comparing with adaptive Kriging model taking u∗ as the density center.
combined with the Monte Carlo simulation method (AK- Step 2 Generate N-size candidate sample pool
MCS) (Echard et al. 2011), the size of the candidate sample Su = {u1 , u2 , ⋯ , uN } by hU (u), and the input–output sam-
pool of AK-IS is smaller than that of AK-MCS, thereby the ples in the process of solving u∗ are collected into the initial
training time of the Kriging model in AK-IS is shortened training set T u.
and the efficiency is improved. Step 3 Construct the Kriging model ĝ u (u) of gu (u) by
The following context will give specific steps of the training set T u.
method AK-IS to estimate Chf (𝛼) , and its flow chart is Step 4 Compute U-learning function Uĝ u (u) at each sam-
shown in Fig. 1. ple u in Su by Eq. (21) (Echard et al. 2013), and select new
Step 1 Solve the design point u∗ of the augmented per- training point unew by Eq. (22).
formance function y = gu (u) by the improved first-order

Fig. 1  The flow chart of AK-IS

for estimating Chf (𝛼)

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 8 of 18 F. Li et al.

|
|𝜇ĝ u (u)|
| 4.3.2 The importance sampling strategy based
Uĝ u (u) = | | (21) on meta‑model for solving the Pf ()
𝜎ĝ u (u)
The IS method based on the design point is easy to realize,
u new
= arg min Uĝ u (ul ) but it depends on the solution of the design point. For com-
(22)
ul ∈Su
plex multiple design points and multiple failure domains,
In Eq. (21), 𝜇ĝ u (u) and 𝜎ĝ u (u) respectively represent the the IS method based on the design point will be difficult to
prediction mean and prediction standard deviation of the implement. For addressing this difficulty, the importance
Kriging model ĝ u (u). sampling method based on meta-model (Meta-IS) developed
Step 5 If Uĝ u (unew ) ≥ 2, stop the adaptively learning pro- in Ref. (Dubourg et al. 2013) is employed to solve the distri-
cess and execute the next step 6, otherwise compute gu (unew ) bution of Pf (𝜽) and Chf (𝛼).
by calling augmented performance function at unew, then add In the Meta-IS method, the importance sampling density
{unew gu (unew )} into the training set T u and return to step 3. function hU (u) corresponding to the augmented performance
Step 6 Calculate ÎFu (ul ) (ul ∈ Su ) by Eq. (23), and calcu- function gu (u) is shown as follows.
late the estimated value Pr{F ̂ u } of Pr{Fu } by Eq. (24). 𝜋(u) ⋅ fU (u)
(25)
ĝ u (ul ) ≤ 0
{ hU (u) =
Pf𝜀
1
̂IF (ul ) = ul ∈ Su (23)
ĝ u (ul ) > 0 where 𝜋(u) and Pf 𝜀 are respectively the probability clas-
u
0
sification function of the Kriging model ĝ u (u) and normali-
N { } zation coefficient, and their definitions are shown as follows.
1 ∑ f (u )
̂
Pr{F } = ÎFu (ul ) U l (24) ( )
𝜋(u) = Pr{̂gu (u) ≤ 0} = Φ −
u
N l=1 hU (ul ) 𝜇ĝ u (u)
(26)
𝜎ĝ u (u)
Step 7 Generate N𝜽 samples 𝜽k (k = 1, 2, ⋯ , N𝜽 ) by f𝜽 (𝜽),
and use u𝜽k = T −1 𝜽
(𝜽k ) to obtain the standard normal sample
corresponding to 𝜽k.
∫ (27)
Pf 𝜀 = 𝜋(u)fU (u)du
Step 8 Calculate the Euclidean distance di (i = 1, 2, ⋯ , N)
between each sample point ui ∈ Su and u𝜽k , denote these
distances sorted in ascending order as {d[1] , d[2] , ⋯ , d[N] }, where Φ(⋅) expresses the cumulative distribution function
and denote t he cor responding sample set as of standard normal variable. /
S[u] = {u[1] , u[2] , ⋯ , u[N] }. By taking hU (u) = 𝜋(u)fU (u) Pf 𝜀 as the importance sam-
Step 9 Set the initial number m of importance sampling pling density function, the failure probability Pr{Fu } of the
samples for estimating Pr{Fu ∩ u𝜽 ∈ uD } as m = 1. augmented performance function can be derived as follows
(Dubourg et al. 2013).
𝜽k
Step 10 Calculate ̂
Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
},
̂ D
and ̂ D
respec- fU (u)
∫ ∫
Var[Pr{F u ∩ u 𝜽 ∈ u 𝜽k
}] Cov[Pr{F u ∩ u 𝜽 ∈ u𝜽k
}]
Pr{Fu } = IFu (u)fU (u)du = IFu (u)
h (u)du
tively by{ Eqs. (17–19), in which IFu (u) is estimated by hU (u) U
ĝ u (u) ≤ 0
u ∩ u𝜽 ∈ u }] ≤ C ,
1
. If Cov[Pr{F
̂ D ∗ Pf 𝜀 fU (u) IFu (u)
∫ ∫ 𝜋(u) U
Î (u) =
Fu 𝜽k = IFu (u) hU (u)du = Pf 𝜀 h (u)du
0 ĝ u (u) > 0 𝜋(u)fU (u)
execute the next step 11. Otherwise, let m = m + 1, repeat
u ∩ u𝜽 ∈ u𝜽k }] ≤ C .
= Pf 𝜀 𝛼corr
this step until Cov[Pr{F
̂ D ∗
(28)
Step 11 Estimate Pr{u𝜽 ∈ uD } and Pf (𝜽k ) by Eqs. (11)
𝜽k
where 𝛼corr is named as correction factor, and its expres-
and (20) respectively, then complete estimation of Chf (𝛼)
sion is shown in Eq. (29).
defined in Eq. (1) by the numerical simulation method (Li
et al. 2021) based on Pf (𝜽k ). IFu (u)
∫
In order to further improve the computational efficiency, 𝛼corr = hU (u)du (29)
𝜋(u)
the candidate sample pool reduction strategy (Li et al. 2021)
can be embedded in the above solution step, that is, the Equation (28) shows that the Meta-IS method can convert
sample points that satisfies Uĝ u (u) ≥ 2 in step 4 are deleted the solution of the failure probability Pr{Fu } into the prod-
from Su to reduce the size of the candidate sample pool, uct of the normalization coefficient Pf 𝜀 and the correction
thereby the speed of updating the Kriging surrogate model factor 𝛼corr . Using numerical simulation method, Pf 𝜀 and
is improved. 𝛼corr can be estimated by the following Eqs. (30) and (31),
respectively,

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 9 of 18 195

N𝜀 meta-model importance sampling (Meta-IS-AK), in which

1 ∑ f
P̂ f 𝜀 = 𝜋(ui ), (30) the existing Kriging model for generating the IS sample is
N𝜀 i=1
further updated to accurately identify the state of uhj and
estimate IFu (uhj ). Thus, the computational efficiency of Meta-
1 ∑ IFu (uhj )
Ncorr
IS-AK is higher than that of the Meta-IS method.
𝛼̂ corr = (31)
Ncorr j=1 𝜋(uhj ) After Pr{F
̂ u } is solved by Meta-IS-AK, the probability
Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
} can be estimated as byproduct without
where ui is the i-th (i = 1, 2, ⋯ , N𝜀 ) sample point gener-
f
additional computational cost of calling actual augmented
ated by fU (u), uhj is the j-th (j = 1, 2, ⋯ , Ncorr ) sample point performance function, and the specific method is given as
generated by hU (u) in Eq. (25). follows.
At the same time, the variances and variation coef- According to the fundamental of the Meta-IS method and
ficients of P̂ f 𝜀 and 𝛼̂ corr can be obtained by Eqs. (32–35), the definition of Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
} in Eq. (14),
respectively. Pr{Fu ∩ u𝜽 ∈ u𝜽 } can be rewritten as Eq. (38).
D
k

∫ FA
Pr{Fu ∩ u𝜽 ∈ uD
𝜽 } = Pr{FA } = I (u)fU (u)du
k

f (u) Pf 𝜀 fU (u)
∫ ∫
= IFA (u) U hU (u)du = IFA (u) h (u)du (38)
hU (u) 𝜋(u)fU (u) U
IFA (u)
∫ 𝜋(u) U
= Pf 𝜀 h (u)du = Pf 𝜀 𝛼A

[N ] where the definition of 𝛼A is shown in Eq. (39).

1 ∑ 𝜀
1 f
Var(P̂ f 𝜀 ) = 𝜋 (ui ) − P̂ f 𝜀 (32)
2 2
N𝜀 − 1 i=1 N𝜀 IFA (u)
∫
𝛼A = hU (u)du (39)
𝜋(u)
√
Var(P̂ f 𝜀 ) Equation (38) transfor ms the estimation of
Cov(P̂ f 𝜀 ) = (33) Pr{Fu ∩ u𝜽 ∈ uD } to those of Pf 𝜀 and 𝛼A , and Pf 𝜀 can be
𝜽k
P̂ f 𝜀
obtained as a byproduct of solving Pr{Fu } by the IS samples.
For estimating 𝛼A, the differential region uD in the vicinity
[ ] 𝜽k
1 1 ∑ IF (uhj )
Ncorr
of u𝜽k need to be determined first by the principle demon-
(34)
2
Var(𝛼̂ corr ) = − 𝛼̂ corr strated in Sect. 4.3.1, that is, on the premise of ensuring that
Ncorr − 1 Ncorr j=1 𝜋 2 (uhj )
the variation coefficient of estimating Pr{Fu ∩ u𝜽 ∈ uD 𝜽k
} is
√ less than the given threshold, make uD 𝜽k
as small as possible.
Var(𝛼̂ corr )
Cov(𝛼̂ corr ) = (35) The specific determination process is similar to that in
𝛼̂ corr Sect. 4.3.1, which will be briefly explained below.
First of all, denote the sorted Euclidean distance
Finally, the estimated value Pr{F
̂ u } and its variation coef-
{d1 , d2 , ⋯ , dNcorr } between uhj (j = 1, 2, ⋯ , Ncorr ) and u𝜽k in
ficient Cov(Pr{Fu }) can be estimated by Eqs. (36) and (37),
̂
respectively. ascending order as {d[1] , d[2] , ⋯ , d[Ncorr ] }, and the correspond-
ing samples are denoted as {uh[1] , uh[2] , ⋯ , uh[N ] }. Then, take
corr
̂
Pr{F ̂ ̂ corr
u }=Pf 𝜀 𝛼 (36) the first m samples of {uh[1] , uh[2] , ⋯ , uh[N ] } to estimate 𝛼A by
corr
𝛼̂ A shown as follows.
√
{m }
̂
Cov(Pr{Fu }) = Cov2 (P̂ f 𝜀 ) + Cov2 (𝛼̂ corr ) (37) 1 ∑ IFA (uh[p] ) Ncorr I (uh )
∑ FA [q]
𝛼̂ A = + (40)
Ncorr p=1 𝜋(uh[p] ) q=m+1 𝜋(u[q] )
h
In the traditional Meta-IS, the indicator function value
IFu (uhj ) of importance sample uhj (j = 1, 2, ⋯ , Ncorr ) is
Since IFA (uh[p] ) = IFu (uh[p] ) (p = 1, 2, ⋯ , m) and
obtained directly by calling the actual augmented perfor-
IFA (uh[q] ) = 0 (q = m + 1, m + 2, ⋯ , Ncorr ) , Eq. (40) can be
mance function, which requires a large amount of computa-
simplified as Eq. (41).
tional cost. Therefore, Ref. (Zhu et al. 2020) proposed a
method of adaptive Kriging model combined with

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 10 of 18 F. Li et al.

1
m I (uh )
∑ Fu [p] U-learning function of ĝ (2) u (u) and update T u as
𝛼̂ A = (41) Tu = (Tu ∪ {unew/ , gu (unew )}), use T u to update ĝ (2) u (u) until
(uj ) ≥ 2.
Ncorr 𝜋(uh[p] ) |
p=1 h | h
min |𝜇ĝ (2) (uj )| 𝜎ĝ (2)
uhj ∈Sh | u | u
And the corresponding variance Var(𝛼̂ A ) and the variation
Step 8 Estimate 𝛼̂ corr and Cov(𝛼̂ corr ) by Eqs. (31) and (35)
coefficient Cov(𝛼̂ A ) of the estimation 𝛼̂ A can be obtained by
based on ĝ (2)u (u). Then estimate Pr{Fu } and Cov(Pr{Fu }) by
̂ ̂
following Eqs. (42) and (43), respectively.
Eqs. (36) and (37).
[ m h
] Step 9 Generate N𝜽 samples 𝜽k (k = 1, 2, ⋯ , N𝜽 ) by f𝜽 (𝜽),
1 1 ∑ IF (u[p] )
Var(𝛼̂ A ) = − 𝛼̂ A2 (42) and obtain u𝜽k by u𝜽k =T −1 𝜽
(𝜽k ).
Ncorr − 1 Ncorr p=1 𝜋 2 (uh[p] ) Step 10 Sort the Euclidean distance between uhj ∈ Sh and
u𝜽k in ascending order. Record the sorted distances as
√
Var(𝛼̂ A ) {d[1] , d[2] , ⋯ , d[Ncorr ] }, and the corresponding sample is
Cov(𝛼̂ A ) = (43) recorded as S[h] = {uh[1] , uh[2] , ⋯ , uh[N ] }.
𝛼̂ A corr
S t e p 1 1 C a l c u l a t e Pr{F ̂ D
u ∩ u𝜽 ∈ u𝜽k } and
The estimation Pr{F
̂ u ∩ u𝜽 ∈ u𝜽k } of Pr{Fu ∩ u𝜽 ∈ u𝜽k }
D D
Cov[Pr{Fu ∩ u𝜽 ∈ u𝜽 }] by Eqs. (44) and (45) based on set-
̂ D

u ∩ u𝜽 ∈ u𝜽k }] ≤ C , execute the

and the variation coefficient Cov(Pr{Fu ∩ u𝜽 ∈ uD }) of
k
𝜽k ting m = 1. If Cov[Pr{F ̂ D ∗
̂
Pr{F ∩ u ∈ uD
} can be estimated by Eqs. (44) and (45). next step 12, otherwise let m = m + 1 and repeat this step
u ∩ u𝜽 ∈ u𝜽k }] ≤ C .
u 𝜽 𝜽 k

̂ D ̂ ̂A until Cov[Pr{F ̂ D ∗
Pr{Fu ∩ u𝜽 ∈ u𝜽 } = Pf 𝜀 𝛼 (44)
k
Step 12 Estimate Pf (𝜽k ) by Eq. (11). Then estimate Chf (𝛼)
√ defined in Eq. (1) by numerical simulation method.
Cov(Pr{Fu ∩ u𝜽 ∈ uD
𝜽
}) = Cov2 (P̂ f 𝜀 ) + Cov2 (𝛼̂ A ) (45)
k

The method of determining the value m is to adaptively 5 Examples

increase the value of m until Cov[Pr{F ̂ u ∩ u𝜽 ∈ u𝜽k }] is less
D

than the given threshold C∗, and then the similar steps in In this section, the results of the double-layer Monte Carlo
Sect. 4.3.1 can be executed to obtain the Pf (𝜽) and Chf (𝛼) in simulation method (DMCS) are used as the reference solu-
turn. tion, and the computational results of two examples are used
A specific implementation of the Meta-IS-AK method to verify the feasibility, efficiency, and accuracy of solv-
is demonstrated as follows to estimate Chf (𝛼), and its flow ing Chf (𝛼) by the proposed methods in this paper. The four
chart is shown in Fig. 2. methods including adaptive Kriging combined with Monte
Step 1 Generate initial training samples by fU (u) and Carlo simulation (AK-MCS), adaptive Kriging combined
calculate the corresponding value of the augmented perfor- with design-point-based importance sampling (AK-IS),
mance function, then construct the training sample set T u. candidate sample pool reduction strategy embedded AK-IS
Step 2 Construct Kriging model ĝ (1) (AK-IS-CSPR) and the adaptive Kriging combined with the
u (u) by T u.
Step 3 Generate sample set Sh = {uh1 , uh2 , ⋯ , uhN } of the meta-model importance sampling (Meta-IS-AK) are com-
corr / pared in each example.
meta-model ĝ (1)u (u) based IS PDF hU (u) = 𝜋(u)fU (u) Pf𝜀 by
Markov Chain Monte Carlo method (MCMC).
5.1 Roof truss structure
Step 4 Execute K-means cluster analysis on Sh . After
obtaining K centroids, these K centroids and their corre-
In the roof truss structure (Zhang et al. 2015) shown in
sponding augmented performance function values are added
Fig. 3, the upper chord and other compression rods of the
to T u , then the enriched T u is used to update the Kriging
roof truss are made of reinforced concrete rods, and the
model ĝ (1)
u (u). lower chord and other tie rods are made of steel rods.
Step 5 Calculate the leave-one-out estimation 𝛼̂ corrLOO of
Assuming that the roof truss bears the action of uniformly
the correction factor 𝛼corr . If 0.1 < 𝛼̂ corrLOO < 10 , execute
distributed load q, after the uniformly distributed load q is
step 6, otherwise return to step 3.
transformed into nodal load, there will be P = ql∕4. Struc-
Step 6 Estimate P̂ f 𝜀 and Cov(P̂ f 𝜀 ) by Eqs. (30) and (33)
tural mechanics analysis
( can)obtain the displacement of
based on ĝ (1)
u (u). ql2 3.81
point C ΔC = 2 A E + A E along the vertical ground
1.13
Step 7 Take Sh as the candidate sample pool and take c c s s

u (u) as the initial model(of g u (u). After

direction, where Ac, As, Ec, Es and l are the cross-sectional
ĝ (1) ̂ (2) / select)new sample
point by u = arg min
new | h |
|𝜇ĝ (2) (uj )| 𝜎ĝ (2) (uhj ) from Sh by area, elastic modulus and length of concrete and steel rods,
uhj ∈Sh | u | u
respectively. Considering the safety and applicability of the

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 11 of 18 195

Fig. 2  The flow chat of the

Meta-IS-AK method for esti-
mating Chf (𝛼)

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 12 of 18 F. Li et al.

D F

A B
E G
qN/m

P
P C P
Ac Ac

12
F

l
D 0.7
Ac 5 Ac As As
0.7 5A Ac

12
A B

l
c

3 As E 2 As G 3 As

0.278l 0.222l 0.25l 0.25l

Fig. 3  Schematic diagram of roof truss structure

Fig. 4  Relationship curve graph between Chf (𝛼) and α in case (1) of
the Example 5.1 (𝜇l ∼ U(11, 13))

Table 1  The distribution of the random inputs of roof truss structure

Inputs Distribution type Mean Standard deviation

q(N∕m) Normal 20,000 1400

l(m) Normal 𝜇l 0.12
Ac (m2 ) Normal 𝜇 AC 4.8 × 10−3
As (m2 ) Normal 9.82 × 10−4 5.892 × 10−5
Es (N∕m ) 2 Normal 2 × 10 10
1.2 × 109
Ec (N∕m ) 2 Normal 1 × 10 11
6 × 109

roof truss, with the vertical downward deflection of the top

C of the roof truss not greater than 3.3 cm as the constraint
condition, the performance function of the structure can be
established as g = 0.033 − ΔC . Assume that all input vari-
ables obey the normal distribution and are independent, and
the distribution parameters are shown in Table 1.
Fig. 5  Relationship curve graph between Chf (𝛼) and α in case (2) of
The following two cases are employed to verify the
the Example 5.1 (𝜇l ∼ U(11, 13), 𝜇Ac ∼ U(0.03, 0.05))
accuracy and efficiency of the proposed methods. Case
(1): consider one of the distribution parameters is random,
that is, 𝜇l ∼ U(11, 13) and 𝜇AC = 0.04m2 . Case (2): con- Kriging surrogate model (AK-MCS, AK-IS, AK-IS-CSPR,
sider two of the distribution parameters are random, that is, Meta-IS-AK) have no obvious advantages in efficiency over
𝜇l ∼ U(11, 13) and 𝜇AC ∼ U(0.03, 0.05). the DMCS method, especially the AK-MCS method, whose
The results of Chf (𝛼) under case (1) and case (2) are computational time is even more than that of DMCS. Com-
shown in Figs. 4 and 5 respectively, and the comparison of paring three methods of AK-MCS, AK-IS and Meta-IS-AK
computational efficiency of four methods as well as refer- in Tables 2 and 3, it can be seen that the size of the candi-
ence DMCS are listed in the Tables 2 and 3, respectively. date sample pool drawn by AK-IS and Meta-IS-AK using
It can be seen from the curves of Chf (𝛼) vs confidence importance sampling strategy is much smaller than that of
level α in Figs. 4 and 5 that the results of the five meth- the AK-MCS method. In addition, among the three meth-
ods are basically consistent, which verifies the accuracy of ods, the Meta-IS-AK based on the meta-model has the least
the proposed methods in this paper. In this example, the amount of computational time. In the methods based on the
performance function is a simple explicit function and the importance sampling strategy proposed in this paper, the
time to call the performance function is shorter than the distribution parameters are converted into the arguments of
time to call the Kriging surrogate model of the performance the performance function, and the increase in the dimen-
function, thus four comparison methods with embedded sion of the inputs will increase the calculation amount of

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 13 of 18 195

Table 2  The results in case Methods DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
(1) of the Example 5.1
(𝜇l ∼ U(11, 13)) Sample pool size 4
10 × ­10 6 6
2 × ­10 2 × ­10 5
2 × 10 5
3000
Computational capacity 104 × ­106 141 177 125 128
Computational time 1427 s 3182 s 394 s 171 s 353 s

The bold values are best results of the examples in this paper

Table 3  The results in case Method DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
(2) of the Example 5.1
(𝜇l ∼ U(11, 13), 𝜇A ∼ U(0.03, 0.05))
c Sample pool size 104 × 2 × ­106 2 × ­106 4 × ­105 4 × 105 4000
Computational capacity 104 × 2 × ­106 197 408 180 183
Computational time 3037 s 6249 s 7722 s 239 s 1075 s

The bold values are best results of the examples in this paper

5.2 Cantilever tube structure

The cantilever tube structure (Jiang et al. 2011) shown in

Fig. 6 has four independent input random variables, which
are the applied loads F1 and F2, the tube wall thickness t and
the outer diameter d. The distribution types and parameters
are shown in Table 4. In practical applications, an output that
engineers often pay attention to is the equivalent stress at
the center of the top surface, which is expressed as Eq. (46),
Fig. 6  Schematic diagram of the force of the cantilever tube
√
𝜎max = 𝜎x2 + 3𝜏zx 2
(46)

Table 4  The distribution of the random inputs of cantilever tube

structure w h e r e 𝜎x = (F1 sin 𝜃1 + F2 sin 𝜃2 + P)∕A + Md∕(2I) ,
𝜏zx = Td∕(4I) , M = F1 cos 𝜃1 L1 + F2 cos 𝜃2 L2 ,
Inputs Distribution type Parameter 1 Parameter 2
A = 𝜋[d − (d − 2h) ]∕4 ,
2 2
I = 𝜋[d4 − (d − 2h)4 ]∕64 ,
t(mm) Normal 𝜇t 𝜎t = 0.1 P = 12000N , T = 1800N , 𝜃1 = 𝜋∕18 , 𝜃2 = 𝜋∕36 ,
d(mm) Normal 𝜇d = 42 𝜎d = 0.5 L1 = 60mm , L2 = 120mm.
F1 (N) Weibull kF1 = 150 𝜆F1 = 5000 When the equivalent stress is greater than the yield
F2 (N) Weibull kF2 = 120 𝜆F2 = 5000 strength of the cantilever tube structure, the structure is
determined to fail, so the limit state function of the structure
can be expressed by Eq. (47),
√
G(X) = R − 𝜎max = R − 𝜎x2 + 3𝜏zx 2
(47)
the importance sampling method. Therefore, in case (2),
because the dimension of inputs is two-dimensional more In Eq. (47), yield strength is set to R = 460MPa . Assum-
than the dimension of the inputs of original performance ing that the mean value of the thickness of the cantilever
function, the computational cost of the AK-IS method in obeys a uniform distribution, i.e., 𝜇t ∼ U(4, 6), the methods
Table 3 is more than that of AK-MCS. But if the candi- proposed in this paper are used to solve Chf (𝛼), and detailed
date sample pool reduction strategy is adopted in the AK-IS results are shown in Fig. 7 and Table 5.
method, it can significantly reduce the number of function In this example, two variables follow the normal dis-
calls and computational time for solving Chf (𝛼), which can tribution, and the other two variables follow the Weibull
be clearly observed from the comparison of the results of distribution. Comparing the results obtained by AK-MCS
AK-IS and AK-IS-CSPR in Tables 2 and 3. and AK-IS-CSPR in Table 5, it can be found that by using
importance sampling and candidate sample pool reduction
strategies, the size of the candidate sample pool is smaller
and the computational time is shorter. Although the compu-
tational cost of Meta-IS-AK method is slightly higher than

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 14 of 18 F. Li et al.

0.08

0.07 DMCS
AK-MCS
AK-IS
0.06 AK-IS-CSPR
Meta-IS-AK

0.05

0.04

0.03

0.02

0.01

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig.8  The geometry of the rudder surface structure
Fig. 7  Relationship curve graph between Chf (𝛼) and α of the Exam-
ple 5.2 (𝜇t ∼ U(4, 6))

g(E, 𝜈, P0 , tsk , trs ) = 0.855 − Δmax (E, 𝜈, P0 , tsk , trs ) (48)

that of AK-MCS, the calculation time is reduced by nearly
half. And the curves of Chf (𝛼) vs 𝛼 of the five methods in where E is the Young modulus, υ indicates Poisson ratio,
Fig. 7 are basically consistent, which verifies the accuracy P0 represents the initial aerodynamic force, tsk represents
of the proposed methods. Moreover, this example verifies the thickness of the ribs and stringers, and trs represents the
that the method proposed in this paper is also applicable to thickness of the skin.
input variables with Weibull distribution. Two cases are considered to verify the accuracy and effi-
ciency of the proposed methods. Case (1): consider one of
5.3 Rudder surface structure the distribution parameters is random, that is, 𝜇trs ∼ U(2, 4),
and 𝜇tsk = 0.7mm . Case (2): consider two of the distribu-
The rudder surface (Feng et al. 2020) is a control device that tion parameters are random, that is, 𝜇trs ∼ U(2, 4) and
generates balance and control force by deflection to manipu- 𝜇tsk ∼ U(0.6, 0.8). Chf (𝛼) results of these two cases are dis-
late the rocket. In the early design stage of the rocket, it is played in Figs. 10 and 11 respectively, while the comparison
very important to control the reliability of the rudder surface of efficiency of five methods is shown in Tables 7 and 8
to estimate the failure probability of the rudder surface under respectively.
the condition of two-fold random uncertainty. Therefore, the In this example, the curves of Chf (𝛼) vs α shown in
methods proposed in this paper are used to solve Chf (𝛼) of Figs. 10 and 11 are basically consistent, which verifies
the rudder surface structure under the condition of two-fold the accuracy of the AK-IS, AK-IS-CSPR and Meta-IS-AK
random uncertainty. The geometry of the rudder surface is proposed in this paper. It can be seen from Tables 7 and 8
shown in Fig. 8, and the parameters of all random inputs are that the computational time of the DMCS method is much
shown in Table 6. The finite element analysis of the rudder more than those four compared methods. For the rudder
surface structure is shown in Fig. 9. Taking the maximum surface, such a complex structure with long computational
displacement of the rudder surface structure not larger than time of single performance function, the obvious advantage
8.55 cm as the constraint, the performance function of the is reflected in computational efficiency of the four com-
rudder surface structure can be established as shown in Eq. pared method embedded with Kriging model. It is shown in
(48). Tables 7 and 8 that the size of candidate sample pool and the

Table 5  The results of the Method DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
Example 5.2 (𝜇t ∼ U(4, 6))
3 5 5 4 4
Sample pool size 5 × ­10 × 9 × ­10 9 × ­10 9 × ­10 9 × 10 1000
Computational capacity 5 × ­103 × 9 × ­105 32 40 32 36
Computational time 2846 s 141 s 174 s 108 s 91 s

The bold values are best results of the examples in this paper

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 15 of 18 195

Table 6  The distribution of random inputs of the rudder surface

structure
Inputs Distribution Mean Coefficient
of variation

E(MPa) Normal 117,600 0.01

𝜈 Normal 0.3 0.01
P0(KN) Normal 0.16 0.05
tsk(mm) Normal 𝜇tsk 0.02
trs(mm) Normal 𝜇trs 0.02

computational time in AK-IS method are all less than those

in AK-MCS method. Comparing the results of AK-MCS
and Meta-IS-AK, although the performance function evalu-
ation number of the Meta-IS-AK method is slightly higher
than that of the AK-MCS method, the candidate sample pool Fig. 10  Relationship curve graph between Chf (𝛼) and α in case (1) of
the Example 5.3 (𝜇trs ∼ U(2, 4))
size of the Meta-IS-AK is much smaller than that of the
AK-MCS method. In addition, small candidate sample pool
size can shorten the time for adaptively updating the Krig-
ing model. In addition, from the comparison of the results
of the AK-IS-CSPR and AK-IS methods in Tables 7 and 8,
it can be clearly seen that the candidate sample pool reduc-
tion strategy can significantly improve the computational
efficiency of the Chf (𝛼) without reducing the computational
accuracy.

5.4 Flap structure

An aircraft flap structure (Ling et al. 2020) is shown in

Fig. 12, the flap structure is a device to increase the lift
during the flight of the aircraft. The structure is installed on
the trailing edge or leading edge of the wing, and consists
of four parts: fixed wing, movable wing, control mechanism
and beam composition. The implicit performance function
of the structure shown in Eq. (49) is established under the Fig. 11  Relationship curve graph between Chf (𝛼) and α in case (2) of
the Example 5.3 (𝜇tsk ∼ U(0.6, 0.8), 𝜇trs ∼ U(2, 4))
constraint that the maximum displacement of the flap cannot
exceed the critical value Δ.

Fig.9  The result of the finite

element analysis of the rudder
surface structure

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 16 of 18 F. Li et al.

Table 7  The results in case Method DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
(1) of the Example 5.3
(𝜇trs ∼ U(2, 4)) Sample pool size 3
5 × ­10 × 6 × ­10 5
2 × ­10 6
4 × ­10 5
4 × 10 5
2000
Computational capacity 5 × ­103 × 6 × ­105 46 69 44 48
Computational time 38.6 h 352 s 138 s 24 s 98 s

The bold values are best results of the examples in this paper

Table 8  The results in case Method DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
(2) of the Example 5.3
(𝜇tsk ∼ U(0.6, 0.8), 𝜇trs ∼ U(2, 4)) Sample pool size 5 × ­103 × 8 × ­105 5 × ­106 6 × ­105 6 × 105 4000
Computational capacity 5 × ­103 × 8 × ­105 90 151 89 96
Computational time 118 h 5082 s 1464 s 530 s 366 s

The bold values are best results of the examples in this paper

Fig. 12  Finite-element model of

the flap structure of aircraft

Table 9  The distribution of random inputs of the flap structure

Variable Distribution Mean Coefficient
of variation

t1(cm) Lognormal 2 0.01

t2(cm) Lognormal 2 0.01
t3(cm) Lognormal 4 0.01
ES(MPa) Normal 70,380 0.01
GS(MPa) Normal 26,458.6 0.01
Ef(MPa) Normal 72,750 0.01
Gf(MPa) Normal 27,236.8 0.01
F(kN) Normal 0.32 0.07

g(t1 , t2 , t3 , ES , GS , Ef , Gf , F) = Δ − Δmax (t1 , t2 , t3 , ES , GS , Ef , Gf , F)

(49)
Fig. 13  Relationship curve graph between Chf (𝛼) and α of the Exam-
where Δ = 20cm , and the maximum displace- ple 5.4
ment Δmax can be obtained by finite element method.
{t1 , t2 , t3 , ES , GS , Ef , Gf , F} are independent random
Assume that the mean of the distribution parameters
inputs variables, in which t1 , t2 , t3 respectively represent
of t2 is random and obeys a uniform distribution, i.e.,
the thickness of beam I-III, ES and GS are respectively the
𝜇t2 ∼ U(0.5, 3.5). Chf (𝛼) estimated by different methods is
Young’s modulus and shear modulus of the skin, Ef and Gf
plotted in Fig. 13, and the comparison of computational effi-
are respectively the Young’s modulus and shear modulus of
ciency of five methods is listed in Table 10.
the frame, and F is the aerodynamic load of the wing. The
From the curves in Fig. 13, it can be seen that the results
first three variables {t1 , t2 , t3 } follow a lognormal distribu-
obtained by these five methods are basically consistent,
tion, and the last five variables {ES , GS , Ef , Gf , F} follow
which verifies the accuracy of the proposed importance
a normal distribution. The distribution parameters and dis-
sampling-based algorithms. About the computational cost
tribution types of the eight random variables are shown in
comparison of the five methods, it is shown in Table 10 that
Table 9.
the AK-IS-CSPR method is the most efficient method among

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Importance sampling‑based algorithms for efficiently estimating failure chance index under… Page 17 of 18 195

Table 10  The results of the Method DMCS AK-MCS AK-IS AK-IS-CSPR Meta-IS-AK
Example 5.4 (𝜇t2 ∼ U(0.5, 3.5))
3 6 6 5 5
Sample pool size 5 × ­10 × 2 × ­10 2 × ­10 3 × ­10 3 × 10 2500
Computational capacity 5 × ­103 × 2 × ­106 163 167 73 136
Computational time 41.7 h 4470 s 681 s 33 s 121 s

The bold values are best results of the examples in this paper

those five compared methods, which verifies the effect of Supplementary Information The online version contains supplemen-
candidate sample pool on improvement of the computational tary material available at https://d​ oi.o​ rg/1​ 0.1​ 007/s​ 00158-0​ 22-0​ 3286-x.
efficiency. Furthermore, the example contains random inputs
Acknowledgements This work was supported by the National Natural
variables mixed with non-normal variables and normal vari- Science Foundation of China (Grant No. NSFC 52075442), and the
ables, which proves that the proposed methods can also be National Science and Technology Major Project (2017-IV-0009-0046).
used in the case of mixed variables.
Declarations

6 Conclusion Conflict of interest On behalf of all authors, the corresponding author

states that there is no conflict of interest.
The failure chance index (FCI) can quantitatively charac- Replication of results The MATLAB codes used to generate results are
terize the reliability degree of the structure under two-fold available in the Supplementary material.
random uncertainty, and the key for solving the FCI is to
accurately and efficiently estimate the probability density
distribution of the failure probability function (FPF) of the
structure. In order to efficiently solve the FPF and FCI, two
References
kinds of analysis method based on importance sampling Au SK (2005) Reliability-based design sensitivity by efficient simula-
strategy are proposed in the paper. Firstly, by standard tion. Comput Struct 83(14):1048–1061
normal transformation of random inputs, the distribution Au SK, Beck JL (1999) A new adaptive important sampling scheme.
parameters are transformed into the performance function as Struct Saf 21(2):135–158
Au SK, Beck JL (2002) Important sampling in high dimensions. Struct
arguments, and an augmented performance function contain- Saf 25(2):139–163
ing inputs and distribution parameters is established. Then, August G, Baratta A, Casciat F (1984) Probabilistic methods in struc-
based on the augmented performance function, and by using ture engineering. Chapman and Hall, London
differential approximation and conditional probability the- Bayrak G, Acar E (2018) Reliability estimation using markov chain
Monte Carlo-based tail modeling. AIAA J 56(3):1211–1224
ory, the FPF and FC can be solved by only one simulation Bourinet JM, Deheeger F, Lemaire M (2011) Assessing small failure
estimation of the failure probability of the augmented perfor- probabilities by combined subset simulation and support vector
mance function. When estimating the failure probability of machines. Struct Saf 33(6):343–353
the augmented performance function, the adaptive Kriging Du XP, Chen W (2001) A most probable point-based method for effi-
cient uncertainty analysis. Design Manufact 4(1):47–66
model combined with the design-point-based importance Dubourg V, Sudret B, Deheeger F (2013) Metamodel-based impor-
sampling method and the adaptive Kriging model combined tance sampling for structural reliability analysis. Probab Eng
with the meta-model based importance sampling method Mech 33:47–57
are used respectively. Moreover, the first method is mainly Echard B, Gayton N, Lemaire M (2011) AK-MCS: an active learning
reliability method combining Kriging and Monte Carlo Simula-
suitable for single design point, while the second method tion. Struct Saf 33:145–154
can be applied to multiple design points and multiple failure Echard B, Gayton N, Lemaire M, Relun N (2013) A combined impor-
domains. These two methods can all reduce the size of can- tance sampling and Kriging reliability method for small failure
didate sample pool of the traditional Monte Carlo method, probabilities with time-demanding numerical models. Reliab Eng
Syst Saf 111:232–240
the size of training sample pool and the time for training the Feng KX, Lu ZZ, Ling CY, Yun WY (2019) An innovative estimation
Kriging model, which will improve the efficiency of esti- of failure probability function based on conditional probability of
mating the FCI. In addition, combined with the candidate parameter interval and augmented failure probability. Mech Syst
sample pool reduction strategy, the above methods can fur- Signal Process 123:606–625
Feng K, Lu ZZ, Wei N, Li GJ (2020) Adaptive kriging model for
ther improve the efficiency of updating the Kriging model. fuzzy safety degree analysis to time-dependent structure. AIAA
Finally, the analysis results of the roof truss structure, beam J 59(9):1–11
structure and the rudder surface structure verify the accuracy Gasser M, Schueller GI (1997) Reliability-based optimization of struc-
and efficiency of the methods proposed in this paper. tural systems. Math Methods Oper Res 46(3):287–307

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

195 Page 18 of 18 F. Li et al.

Jensen HA (2005) Structural optimization of linear dynami- Peng J, Liu BD (2007) Birandom variables and birandom program-
cal systems under stochastic excitation: a moving reli- ming. Comput Ind Eng 53(3):433–453
ability database approach. Comput Methods Appl Mech Eng Rackwitz R, Fiessler B (1978) Structural reliability under combined
194(12–16):1757–1778 random load sequences. Comput Struct 9:489–494
Jiang C, Li WX, Han X, Liu LX, Le PH (2011) Structural reliability Regis L, Anne D (2009) An innovating analysis of the Nataf trans-
analysis based on random distributions with interval parameters. formation from the copula viewpoint. Probab Eng Mech
Comput Struct 89(23–24):2292–2302 24(3):312–320
Li F, Lu ZZ, Feng KX (2021) Improved chance index and its solutions Xu C, Liu Z, Zhu P, Li MS (2020) Sensitivity based adaptive sequen-
for quantifying the structural safety degree under twofold random tial sampling for metamodel uncertainty reduction in multilevel
uncertainty. Reliab Eng Syst Saf 212:107635 systems. Struct Multidisc Optim 62(11):1473–1496
Li LY, Lu ZZ (2015) A new algorithm for importance analysis of the Zhang LG, Lu ZZ, Pan W (2015) Efficient structural reliability analysis
inputs with distribution parameter uncertainty. Int J Syst Sci 1–13 method based on advanced Kriging model. Appl Math Model
Ling CY, Lu ZZ, Zhang XB, Zhou ST (2020) Safety analysis for the 39(2):781–793
posfust reliability model under possibilistic input and fuzzy state. Zhou ST, Lu ZZ, Ling CY, Yun WY (2021) Global reliability sensitiv-
Aerospace Sci Technol 99(Apr.): 105739.1–105739.19. ity analysis index and its efficient numerical simulation solution
Liu BD (2002) Uncertainty Theory, 2th edition. Springer, 2002, New in presence of both random and interval hybrid uncertainty. Struct
York. Multidisc Optim 63(2):551–573
Melchers RE (1989) Importance sampling in structural system. Struct Zhu XM, Lu ZZ, Yun WY (2020) An efficient method for estimating
Saf 6:3–10 failure probability of the structure with multiple implicit failure
Ormoneit D, White H (1999) An efficient algorithm to compute maxi- domains by combining Meta-IS with IS-AK. Reliab Eng Syst Saf
mum entropy densities. Economet Rev 18:127–140 193:106644
Park BU, Marron JS (1990) Comparison of deta-driven bandwidth
selectors. J Am Stat Assoc 85:66–72 Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]