Journal of Reliability

Science and
Engineering

PAPER • OPEN ACCESS You may also like

- Adaptive artificial neural network for
AK-MCS structural reliability analysis method uncertainty propagation
Yan Shi, Lizhi Niu and Michael Beer
based on improved active learning function - Hierarchical Bayesian modeling for
uncertainty quantification and reliability
updating using data
To cite this article: Hongyan Chu et al 2025 J. Reliab. Sci. Eng. 1 025003 Xinyu Jia, Weinan Hou and Costas
Papadimitriou

- Time-variant system reliability analysis via

stochastic process discretization and most
probable point trajectory approximation
View the article online for updates and enhancements. Dequan Zhang, Hongfei Zhang, Pengfei
Zhou et al.

This content was downloaded from IP address [Link] on 18/01/2026 at 18:10

 Journal of Reliability Science and Engineering

J. Reliab. Sci. Eng. 1 (2025) 025003 (19pp) [Link]

AK-MCS structural reliability analysis

method based on improved active
learning function
Hongyan Chu1,2, Wenzhong Li1,2, Dongyang Sun1,2, Zhifeng Liu1,2,3, Weixu Zheng1,2,
Jingjing Xu1,2,∗ and Qiang Cheng1,2
1
Institute of Advanced Manufacturing and Intelligent Technology, Beijing University of Technology,
Beijing, People’s Republic of China
2
Beijing Key Laboratory of Design and Intelligent Machining Technology for High Precision Machine
Tools, Beijing University of Technology, Beijing, People’s Republic of China
3
Key Laboratory of Advanced Manufacturing and Intelligent Technology for High-end CNC Equipment,
Jilin University, Changchun, People’s Republic of China

E-mail: xujj0908@[Link]

Received 18 February 2025, revised 2 April 2025

Accepted for publication 27 April 2025
Published 15 May 2025

Abstract
In the field of engineering structural reliability analysis, there are often implicit and highly
nonlinear performance function issues. When using approximate analytical methods to estimate
failure probability, the resulting calculations typically involve significant errors. Although
numerical simulation methods can provide high-accuracy solutions for failure probabilities, they
are hindered by excessively long computation times. Consequently, reliability analysis methods
based on active learning Kriging models have gained widespread application in the reliability
analysis of engineering structures. This paper, based on the classical U learning function,
comprehensively considers the impact of sample point prediction values and prediction
variances on the accuracy of failure probability estimation. In the sample space, priority is given
to selecting sample points that are closer to the limit state surface, and higher weights are
assigned to the prediction variances of these sample points. Based on this, a WU learning
function is proposed. In constructing the WU learning function, the appropriate adaptive
orientation function is determined by comparing the sample selection effectiveness of the WU
function under various adaptive orientation functions. Furthermore, the AK-MCS-WU
reliability analysis method is introduced in conjunction with the Kriging model. Two numerical
examples and one engineering case demonstrate that the proposed AK-MCS-WU method can
ensure the stable convergence of failure probability, while providing efficient and accurate
predictions of structural failure probability for reliability problems involving complex
performance functions.

∗
Author to whom any correspondence should be addressed.

Original content from this work may be used under the

terms of the Creative Commons Attribution 4.0 licence. Any
further distribution of this work must maintain attribution to the author(s) and
the title of the work, journal citation and DOI.
© 2025 The Author(s). Published by IOP Publishing Ltd on behalf of the
1 Institute of Systems Engineering, China Academy of Engineering Physics
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Keywords: structural reliability analysis, active learning function, Kriging surrogate model,
adaptive orientation function

1. Introduction second-order reliability method [3, 4], among others. In the

context of numerical simulation, using MCS [5] to assess fail-
The primary objective of structural reliability analysis is to ure probability is a key technique in reliability analysis. When
calculate the probability of structural failure. In the analysis, addressing reliability problems involving small failure probab-
the performance function G (x) is used to represent the failure ilities, methods such as importance sampling (IS) [6, 7], subset
state of the structure: when G (x) < 0, the sample points lie simulation (SS) [8, 9], line sampling [10], and sequential space
within the failure domain; when G (x) > 0, the sample points transformation [11] are commonly employed to predict the
are within the safe domain; and when G (x) = 0, the sample structural failure probability. Although numerical approxim-
points lie on the limit state surface (LSS). Therefore, using ation methods are computationally efficient, ensuring solution
the Monte Carlo simulation (MCS) method, the mathematical accuracy becomes challenging when dealing with implicit and
expression for calculating the failure probability of the struc- nonlinear function problems. On the other hand, while numer-
ture is: ical simulation methods can provide highly accurate solutions
ˆ ˆ for structural failure probabilities, they require extensive com-
Pf = fX (x) dx = I [G (x)] fX (x) dx = Ef {I [G (x)]} putational resources, limiting their practical applicability.
G(x)≤0 Rn Surrogate model technology is one of the most widely
(1) used approaches in the current field of structural reliabil-
ity research. Common surrogate models include polynomial
where fX (x) is the joint probability density function of random chaos expansion [12], support vector machine models [13],
input variables, Ef (·) represents the mathematical expectation radial basis function [14], neural network models [15], and
based on fX (x), and I [G (x)] is the failure indicator function Kriging models [16, 17], among others. Among these sur-
of the performance function. The mathematical definition of rogate models, the Kriging model, based on Gaussian pro-
I [G (x)] is: cess regression, stands out. Compared to other surrogate mod-
els, the Kriging model offers unique advantages: it not only
(
1 G (x) ≤ 0 provides predicted values but also calculates the prediction
I [G (x)] = . (2) variance for sample points, and the magnitude of this variance
0 G (x) > 0
reflects the uncertainty of the predicted values.
The Kriging model was first proposed by Krige [18] for
By converting the expectation in equation (1) into a sum-
surveying the distribution of mineral resources. Later, French
mation, the unbiased estimate of Pf is:
mathematician Matheron [19] extended and refined the model,
X
NMCS forming a complete theoretical framework. Lophaven et al
1 Nf
P̂f = I [G (xi )] = (3) [20] developed the DACE toolbox on the MATLAB plat-
NMCS NMCS form for constructing Kriging models, providing significant
i =1
convenience for researchers worldwide and greatly promot-
where NMCS is the number of sample points extracted from the ing the widespread application of Kriging models in vari-
random variables according to fX (x), and Nf is the number of ous fields. Kaymaz [21] was the first to introduce Kriging
failure sample points. models into structural reliability analysis, using the DACE
According to equation (3), the coefficient of variation of the toolbox, and demonstrated the superiority of this approach.
structural failure probability is calculated as: However, to achieve a Kriging model that meets accuracy
s requirements, a large number of initial sample points are
  1 − P̂f typically required, leading to high computational costs. In
Cov P̂f = . (4)
(NMCS − 1) P̂f response, many scholars have proposed a new reliability ana-
lysis method that combines surrogate models and sampling
For practical engineering problems, directly calculating the techniques through active learning. Notably, the active learn-
structural failure probability through multiple integrals often ing reliability analysis framework AK-MCS (Active learning
presents significant challenges. To effectively compute the reliability method combining Kriging and MCS) proposed by
structural failure probability, many researchers have proposed Echard et al [22] has received considerable attention. This
a range of methods, primarily including approximate analyt- method initially constructs a Kriging model using a small
ical methods, numerical simulation methods, and surrogate number of sample points and then selects the best iterative
model methods. The approximate analytical methods mainly sample points through the active learning function, continu-
consist of the first-order reliability method [1, 2] and the ously refining the Kriging model until it meets the desired

2
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

accuracy, ultimately enabling precise failure probability the limit state function than the original best points. Meng et al
prediction. [34] introduced an improved U function, which sets accept-
Recent improvements to the AK-MCS method focus ance and rejection domains for sampling points, effectively
primarily on two areas. First, various strategies and methods avoiding calculation waste caused by local sample point clus-
are integrated into the sampling process to enhance analysis tering, enhancing both efficiency and robustness in solving
efficiency. In particular, when dealing with small failure prob- complex reliability problems. Cao et al [35] proposed the VF
ability problems, the AK-MCS method often requires a sig- learning function by selecting candidate samples near the LSS
nificant number of iterations, leading to a sharp increase in and multiplying the Kriging model’s prediction standard devi-
computational effort, which can be impractical in engineer- ation by the probability density function.
ing applications. To address this, Huang et al [23] replaced In the process of selecting the best sample point, the U
the MCS with SS in the AK-MCS method, resulting in learning function defines it as the ratio of predicted mean to
the AK-SS method for small failure probability problems. predicted variance. However, it does not consider the differ-
Echard et al [24] introduced IS into the AK-MCS framework, ing contributions of the predicted mean and variance to active
forming the AK-IS reliability analysis method. Compared learning. For instance, a sample point with a small predicted
to AK-MCS, AK-IS improves computational efficiency by mean may not necessarily have a large predicted variance. To
generating better candidate sample points through sampling address this, this paper improves the traditional U function
techniques. Second, in adaptive structural reliability analysis by proposing the WU active learning function, which assigns
based on Kriging models, the sample point selection strategy different weights to prediction variance based on Kriging’s
plays a crucial role in both model accuracy and compu- predicted mean. Compared to the U learning function, WU
tational efficiency. Several researchers have studied differ- enhances failure probability prediction accuracy and reduces
ent point-selection methods, proposing various learning func- the number of performance function evaluations, thus improv-
tions. Notably, the U function proposed by Echard et al [22], ing analysis efficiency.
which considers the probability of misestimating the sign of The structure of the remainder of this paper is as fol-
sample points, and the expected feasibility function (EFF) pro- lows: section 2 briefly introduces the AK-MCS method, the
posed by Bichon et al [25], which accounts for sample points Kriging model, and active learning functions. Section 3 elab-
near the LSS with high variance, are representative examples. orates on the construction process of the WU learning func-
Additionally, Lv et al [26] proposed the H function, which tion, an improvement on the U function, and provides an ana-
uses information entropy to measure uncertainty around the lysis. Section 4 validates the advantages of the WU learning
LSS. The H learning function quantifies the Kriging model’s function through two numerical examples and one engineer-
fitting uncertainty near the limit state and selects sample points ing example. Section 5 presents the conclusions of the study.
with the highest information entropy, thus enhancing com-
putational efficiency. Zhu and Du [5] proposed a D learning
2. Structural reliability analysis method based on
function that considers the contribution of individual sample
Kriging model
points to failure probability prediction errors and the correla-
tion among all points, improving reliability analysis efficiency.
2.1. AK-MCS method
Other common learning functions include: Zhang et al [27]
and Shi et al [28], who developed the reliability-based expec- The core concept of the AK-MCS method lies in lever-
ted improvement function and folded normal-based expec- aging the Kriging model’s ability to effectively approximate
ted improvement function, respectively, based on folded nor- black-box functions. In engineering reliability problems, the
mal distributions; Yang et al [29], inspired by the ‘Improved Kriging surrogate model replaces the actual response func-
Expected Function’, who proposed the risk expectation func- tion to perform reliability analysis. The implementation pro-
tion, which measures the sign prediction accuracy of Kriging cess involves combining the active learning function with the
models; and Sun et al [30], who introduced a minimum Kriging model. The active learning function is used to con-
improvement function that balances Kriging variance and joint tinuously select the optimal sample points for constructing
probability density functions in failure probability estimation, the Kriging model, thereby iteratively improving the model.
effectively avoiding sample wastage in areas with small prob- Once a Kriging model that meets the accuracy requirements is
ability density values. The termination of the active learn- obtained, the MCS method is applied to predict the structural
ing process is determined by a convergence criterion. In this failure probability.
regard, Wang and Shafieezadeh [31] proposed an error-based
convergence criterion, while Zhang et al [32] developed a new
2.2. Kriging model
stopping criterion based on the expected upper limit of relative
error in failure probability. The Kriging model is an exact semi-parametric interpolation
As a classical active learning function, many researchers method. It uses spatial estimation techniques to perform a
have proposed improvements to the U function. Peijuan et al weighted linear combination of the surrounding information
[33] added a line search step to the traditional U function, of the sample point in order to predict the value at that point.
allowing the improved version to find training points closer to Not only does it provide the predicted mean for the point,

3
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

but it also offers the local prediction variance. Notably, the where:
Kriging model’s prediction error has a mean of zero, and its
mean squared error is minimized, making it an optimal linear
unbiased estimator. r(x) = [R (θ, x1 , x) , R (θ, x2 , x) , · · · , R (θ, xe , x)] (12)
The specific expression of the Kriging model is: u (x) = F R T −1
r (x) − f (x) . (13)
T
G(x) = f (x)β + z(x) (5)

where x represents any input variable, G is the correspond- The Kriging variance σ̂ 2 characterizes the prediction error
ing predicted response value, f (x) = [f1 (x) , f2 (x) , · · · , fm (x)]T at the point x, and its magnitude reflects the degree of uncer-
is the vector of basis functions for the polynomial, m is the tainty in the predicted value. This is a key basis for active learn-
number of polynomials, and β = [β1 , β2 , · · · , βm ]T is the vector ing methods.
of regression coefficients. fT (x) β represents a parameterized
polynomial model that captures the trend in the Kriging model.
2.3. Active learning function
The stochastic

process z (x) follows a Gaussian distribution
N 0, σ 2 and provides a local approximation of Kriging’s sim- (1) U learning function
ulation. The covariance equation between any two sample Echard et al [22] proposed using the U learning function to
points xi and xj in the sample set can be expressed as: select updated sample points. The U function is defined as fol-
lows:
Cov [z (xi ) , z (xj )] = σ 2 R (θ, xi , xj ) (6)

where R (θ, xi , xj ) is the correlation function that represents the µĜ (x)
spatial correlation between the sample points xi and xj , and θ = U (x) = (14)
σĜ (x)
T
[θ1 , θ2 , · · · , θn ] is vector of correlation parameters. The typical
correlation function used is the Gaussian correlation function,
and its specific form is: where µĜ (x) and σĜ (x) represent the predicted mean and
predicted standard deviation of Kriging model at the sample
Y
M h
i points, respectively.
m 2
R (θ, xi , xj ) = exp −θm xm
i − xj (7)
When the absolute value of the predicted mean µĜ (x) is
m =1
decreases, the corresponding U value also becomes smaller,
where xm m
i and xj represent the mth element of xi and xj respect-
indicating that the candidate sample point is closer to the limit
ively, and θm is the correlation parameter corresponding to the state function; Moreover, a smaller U value corresponds to a
mth directional component. This parameter represents the rela- larger Kriging prediction variance, which suggests higher pre-
tionship between the distance and the degree of correlation diction uncertainty at that sample point. In other words, this
among design sample points, and its magnitude directly influ- sample point is not only near the LSS but also has a higher
ences the prediction results of the Kriging model. Kriging prediction uncertainty.
given e initial sample points X = [x1 , x2 , · · · , xe ] and the Based on the above analysis, the optimal sample point
corresponding true limit state function response values G = should be the one corresponding to the minimum value of the
[G (x1 ) , G (x2 ) , · · · , G (xe )], for the given correlation parameter formula. The specific point-selection method is as follows:
T
θ = [θ1 , θ2 , · · · , θn ] , the estimated values of the regression
coefficients β and the variance of the Gaussian process are
respectively: xnew = arg minx∈S U (x) (15)

−1
β̂ = FT R−1 F FT R−1 G (8)
1 T   where xnew represents the optimal sample point, and S is the
σ̂ 2 = G − Fβ̂ R−1 Y − Fβ̂ (9) set of candidate points for the optimal sample.
e The stopping criterion for the U learning function is
where βb is the estimated polynomial parameter vector, σ̂ 2 is min {U (x) ≥ 2}. At this point, the probability that the signs
the estimate of variance σ 2 , and R is the correlation coefficient of the sample points in S are predicted correctly is at least
matrix, with R = [Rij ]e×e , where Rij = R (θ, xi , xj ) .. Φ (2) = 97.7%, indicating that the Kriging model is suffi-
The mean and variance of the Kriging prediction at a certain ciently accurate.
point x are:
  (2) EFF learning function
µĜ (x) = Ĝ (x) = fT (x) β̂ + rT (x) R−1 G − Fβ̂ (10) Bichon et al [25] proposed the EFF, which is a function of both
h
−1 i the predicted mean and predicted variance. The EFF function
σ̂ 2Ĝ (x) = σ̂ 2 1 + uT (x) FT R−1 F u (x) − rT (x) R−1 r (x) reflects how far the true response value is from the LSS G (x) =
(11) b. Its specific expression is as follows:

4
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

[ ( ) ( ) ( )]
( ) b − µGb (x) (b − ε) − µGb (x) (b + ε) − µGb (x)
EFF (x) = µGb (x) − b 2Φ −Φ −Φ
σGb (x) σGb (x) σGb (x)
[ ( ) ( ) ( )]
b − µGb (x) (b − ε) − µGb (x) (b + ε) − µGb (x)
− σGb (x) 2ϕ −ϕ −ϕ
σGb (x) σGb (x) σGb (x)
[ ( ) ( )]
(b + ε) − µGb (x) (b − ε) − µGb (x)
+ε Φ −Φ (16)
σGb (x) σGb (x)

where Φ (·) is the standard normal distribution function;

ϕ (·) is the standard normal probability density function; ε =
The convergence criterion of the EFF function is:
2σĜ2 (x); b = 0 represents the LSS.
The expression for the EFF function to select the best max {EFF (x)} ≤ 10−3 (18)
sample point is:
(3) H learning function
Based on information entropy theory, Lv et al [26] proposed
the H learning function, with its specific expression given as
xnew = arg maxx∈S EFF (x) . (17) follows:

     
√ 1 2σGb (x) − µGb (x) −2σGb (x) − µGb (x)
ln 2π σGb (x) + Φ −Φ
2 σGb (x) σGb (x)
 
2σGb (x) − µGb (x) 2σGb (x) − µGb (x)
H (x) = − ϕ . (19)
2 σGb (x)
 
−2σGb (x) − µGb (x) −2σGb (x) − µGb (x)
+ ϕ
2 σGb (x)

The expression for the H function to select the best sample 3.1. WU learning function
point is:
According to the point-selection strategy of the U learning
function, when selecting the best sample point, those with
xnew = arg maxx∈S H (x) . (20) smaller predicted means and higher prediction variances are
preferred. While the U learning function considers both the
In this paper, the convergence criterion of the H function is predicted values and prediction variances of sample points, for
set as: points in the sample set, even if a sample point is closer to the
LSS (i.e. its predicted value is smaller), the prediction variance
at that point may not necessarily be larger. Since both the pre-
max {H (x)} ≤ 0.5. (21) dicted value and prediction variance of sample points influence
the accuracy of failure probability estimation, it is preferable
to prioritize sample points closer to the LSS. Additionally, a
3. Structural reliability analysis based on
higher weight is assigned to the prediction variance of sample
AK-MCS-WU method
points near the LSS. Based on this, the WU learning function
is proposed:
The active learning function plays a crucial role in the AK-
MCS method, with its performance directly influencing the |µGb (x) |
fitting accuracy of the Kriging model and the number of eval- WU (x) = (22)
W (x)
uations of the performance function. In this section, a new WU
learning function is proposed based on the U learning function. where µĜ (x) is the absolute value of the Kriging model’s pre-
Compared to the U learning function, the WU learning func- dicted mean, W (x) is the adaptive weight function, and its
tion not only enhances the accuracy of the Kriging model but expression is:
also reduces the number of performance function calls, thereby
improving analysis efficiency. W (x) = ρ (x) σGb (x) (23)

5
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

where ρ (x) is the adaptive orientation function, and σĜ (x)

is the standard deviation of the Kriging model’s sample
predictions.
Building on the above analysis, the adaptive orientation
function ρ (x) should yield a higher value as the independent
variable approaches 0. Based on the properties of the expo-
nential function and the normal distribution probability dens-
ity function, this paper provides two forms of ρ (x) to address
this requirement.
 
5 |µb (x) |2
ρ1 (x) = 1 + √ exp − G (24)
2 2π 2


ρ2 (x) = 1 + exp − µĜ (x) . (25)

The data characteristics of the two adaptive orientation

functions are analyzed individually, and the results are presen-
ted in figure 1. It can be observed from the figure that as
Figure 1. The adaptive orient function.
|µĜ (x) | approaches zero and becomes smaller, both adapt-
ive orientation functions tend to converge towards a value of
2. Conversely, as |µĜ (x) | increases, the values of these func- more sample points near the LSS, resulting in a lower over-
tions decrease. In comparison to ρ1 (x), when |µĜ (x) | is close all number of selected sample points and, consequently, fewer
to zero, function ρ2 (x) exhibits a more pronounced change calls to the performance function.
in its values, indicating a difference in steepness between the By further comparing figures 2(b) and (c), it is evident
two functions. According to the weight distribution method that the effectiveness of the WU learning function in select-
of the WU learning function and considering the impact of ing sample points varies depending on the choice of the adapt-
sample point prediction variance, greater weights are assigned ive orientation function. When using the adaptive orientation
to sample points that are closer to the LSS. function ρ2 (x), more sample points are located near the LSS,
Similar to the U function, the sample point corresponding indicating that ρ2 (x) leads to a more effective sample selection
to the minimum value of WU (x) is considered the optimal process for the WU learning function.
sample point in the current Kriging model. As analyzed earlier, compared to ρ1 (x), ρ2 (x) exhibits a
steeper curve, with more pronounced changes in function val-
xnew = arg min [WU (x)] . (26) ues when |µĜ (x) | is small and approaches zero. To further
determine the parameters of the adaptive orientation func-
When the WU learning function stops active learning, the tion ρ2 (x), two parameterized forms, ρ2−1 (x) and ρ2−2 (x) are
value of |µĜ (x) | approaches 0, at which point ρ (x) = 2. introduced. The corresponding function curves are shown in
Therefore, the WU learning function is related to the U learn- figure 3. When using ρ2−1 (x) and ρ2−2 (x) as the adaptive ori-
ing function as follows: entation functions, the stopping criterion for the WU function
is set as min {WU (x) ≥ 2/3}.
µĜ (x) µ (x) 1
WU (x) = = Ĝ = U (x) . (27)

ρ (x) σĜ (x) 2σĜ (x) 2 ρ2−1 (x) = 2 + exp −|µGb (x) | (29)


ρ2−2 (x) = 1 + 2 exp − µĜ (x) (30)
The stopping criterion for the U learning function is
min {U (x) ≥ 2}. Based on this, the stopping criterion for the The effectiveness of the WU learning function correspond-
WU learning function can be derived as follows: ing to adaptive orientation functions with three different para-
meter settings in selecting sample points is compared, and the
min {WU (x) ≥ 1} . (28) results are illustrated in figure 4. As shown in the figure, the
sample points selected by the WU learning functions asso-
The effectiveness of the U learning function, WU learning ciated with these three parameterized adaptive orientation
function, and WU learning function under different adaptive functions are all concentrated near the LSS, suggesting that
orientation functions for sample point selection was compared. their selection performance is comparable. Based on this ana-
The simple function G (x) = x21 + x22 − 16 was analyzed and lysis, ρ2 (x) is ultimately chosen as the adaptive orientation
calculated to obtain the distribution of sample points selected function, leading to the final expression of the WU learning
by various learning functions, as shown in figure 2. function.
Comparing figure 2(a) with figures 2(b) and (c), the U
learning function tends to select sample points that are farther |µGb (x) |
WU (x) =

. (31)
from the LSS. In contrast, the WU learning function selects 1 + exp −|µGb (x) | σGb (x)

6
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 2. Sample points selected by U and WU.

Figure 3. ρ2 (x) with different parameters.

Figure 4. Sample points selected by different WU.

7
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

3.2. Structural reliability analysis process based on 4.1. Example 1: two-dimensional four-branch series system
AK-MCS-WU method
The random variables x1 and x2 are independent and follow a
The present paper proposes a structural reliability analysis standard normal distribution. The constant k is used to control
method based on the AK-MCS-WU approach. The key steps the failure probability of the series failure system, with values
are outlined as follows. of 6 or 7. The corresponding limit state function is defined as
Step 1: The MCS sample set NMCS was established by follows.
employing Latin hypercube sampling to extract input samples
 
of group m, based on the data characteristics of random input  (x1 − x2 )
2
(x1 + x2 ) 

3 + − √ 

variables. 
 


 10 2 

Step 2: The design of the experiment (DoE) was conducted. 
 


 (x1 − x2 )
2
(x1 + x2 ) 

n initial sample points were selected from the sample set NMCS , 
3 + + √ 

and the true response of these sample points was calculated G (X) = min 10 2 . (33)
and denoted as G. 
 k 


 (x1 − x2 ) + √ 

Step 3: Utilizing the initial sample point and corresponding 
 


 2 

response value G, the DACE toolbox in MATLAB is employed 
 


 k 

to develop or refine the Kriging agent model. (x2 − x1 ) + √ 
2
Step 4: Based on the Kriging model, the predicted values
for the NMCS sample points are computed.
In the sample space [−5, +5] of the basic variables x1 and
Step 5: Based on the WU learning function, the active learn-
x2 , use the Latin hypercube sampling to extract N = 12 initial
ing function is calculated and the optimal sample point Xnew is
sample points for DoE, and the number of initial sample points
selected.
for MCS is NMCS = 106 . The modeling results of the AK-
Step 6: Refine the learning function of WU and verify
MCS method based on the WU learning function proposed
its convergence. If min {WU (x ≥ 1)}, proceed to Step 7.
in this paper and several other common learning functions
Otherwise, compute the response value Gnew for Xnew , update
are shown in figures 6 and 7. The figures include the initial
the DoE sample point set, and then advance to Step 3.
sample points, the updated sample points selected by the active
Step 7: Refine the coefficient of variation for the failure
learning function, the true limit state function, and the Kriging
probability. Should the Cov (Pf ) ≥ 0.05, it is imperative to
model.
generate a novel initial candidate sample point and to augment
It can be seen from figures 6 and 7 that the AK-MCS meth-
the existing array of initial sample points. Subsequently, pro-
ods using different learning functions are all capable of fitting
ceed to Step 2 anew. Conversely, if Cov (Pf ) < 0.05, it will be
the true limit state function. Among them, the AK-MCS-EFF
concluded that the estimated failure probability is satisfactory,
method, which employs the EFF learning function, selects
thereby achieving the final Kriging model and obtaining the
the fewest best sample points, thereby minimizing the num-
structural failure probability.
ber of calls to the performance function. In comparison to the
The flow chart illustrating the AK-MCS-WU method is
AK-MCS-U method, the AK-MCS-WU method has a distinct
depicted in figure 5.
advantage in reducing the number of performance function
evaluations. Additionally, the best sample points selected by
4. The experimental example verifies the WU learning function are more concentrated near the limit
state curve, which is conducive to improving the accuracy of
In this section, the effectiveness of the proposed AK-MCS- the Kriging model and thus improving the prediction accur-
WU method is validated through three case studies. All cal- acy of the failure probability. In a certain test, the convergence
culation results presented in this paper represent the average curves of the failure probability and the relative—error curves
values obtained from ten independent random [Link] the of the evaluation results during the active learning process of
failure probability computed by the MCS method as a bench- various methods are shown in figures 8 and 9.
mark, the accuracy of each method is assessed, and the relative It can be seen from figures 8 and 9 that all methods can meet
error for each method is calculated as follows: certain evaluation results. During the entire calculation pro-
cess, the methods are sorted in ascending order of the number
of iterations as AK-MCS-EFF, AK-MCS-H, AK-MCS-WU,
PMCS
f − P̂f and AK-MCS-U. Although the first two methods have fewer
εerror = × 100% (32) iterations, their relative errors in predicting the failure prob-
PMCS
f
ability are relatively large. It can be seen from figures 8(b)
and 9(b) that compared with the AK-MCS-U method, the AK-
where PMCS
f is the failure probability calculated by MCS MCS-WU method has a larger prediction error in the initial
method, and P̂f is the failure probability predicted by Kriging stage of iteration. However, as the accuracy of the Kriging
model. model continuously improves, the prediction error becomes

8
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 5. AK-MCS-WU method flow chart.

smaller and smaller, and finally, a Kriging model that meets and AK-MCS-H methods result in 12 + 14.2 and 12 + 32.8
the accuracy requirements is obtained. calls to the performance function, respectively, with predic-
The specific reliability calculation results of each method tion relative errors of 4.7% and 4.5%, and average solution
are shown in table 1. In the table, Ncall represents the number times of 638.89 s and 688.21 s, respectively. When k = 7,
of calls to the performance function, Pf represents the failure all methods have similar effects on the estimation of the fail-
probability, CovPf represents the coefficient of variation, εerror ure probability and the number of calls to the performance
is the relative error between the estimated value of the fail- function.
ure probability and the standard value, and T represents the In summary, although the AK-MCS-EFF method and the
solution time in seconds. AK-MCS-H method can significantly reduce the number of
It can be seen from table 1 that when k = 6, the failure prob- calls to the performance function, their prediction accuracy
ability calculated by the MCS method is 4.473 × 10−3 , with is relatively low, and the average time required for each pro-
the number of calls to the performance function is 1 × 106 . gram run is quite long, resulting in high computational costs.
The failure probability calculated by the AK-MCS-U method Compared with the AK-MCS-U method, the AK-MCS-WU
is 4.437 × 10−3 , with a relative error of 0.80%. Compared method can improve the accuracy of the failure probability
with the MCS method, the number of calls to the perform- estimation to a certain extent, reduce the number of calls to the
ance function of the AK-MCS-U method is significantly performance function, and improve the efficiency of structural
reduced, totaling 12 + 59.3 calls, where 12 represents the reliability analysis. In addition, the box plots of the number of
number of initial sample points and 59.3 represents the aver- calls to the real performance function by various methods are
age number of performance function calls during the Kriging shown in figure 10.
model reconstruction. The failure probability calculated by the It can be seen from figure 10 that compared with the
AK-MCS-WU method is 4.500 × 10−3 , with a relative error AK-MCS-U method, the AK-MCS-WU method has a lower
of 0.60%, and the number of calls to the performance func- median number of calls to the performance function, and
tion is only 12 + 51.1 times. The average time required for its box plot has a smaller box, which means that the
the program to solve once is 143.83 s. The AK-MCS-EFF AK-MCS-WU method has higher stability in the number

9
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 6. Modeling results of various methods when k = 6.

of calls to the performance function during the calculation A Kriging model is constructed in the sample space [−2,
process. +2]. The Latin hypercube sampling is used to extract N = 12
initial sample points for DoE, and the number of initial sample
4.2. Example 2: nonlinear oscillator system
points for MCS is NMCS = 106 . The reliability calculation
results of the AK-MCS method based on the WU learning
The figure 11 illustrates a widely utilized nonlinear oscillator function proposed in this paper and several other common
for validating the reliability analysis method, with its corres- learning functions are shown in table 3.
ponding performance function as fellow. The calculation results show that the AK-MCS-EFF
ω t  method requires the fewest calls to the performance func-
2F1 0 1
G (x) = 3r − | sin | (34) tion but has the largest relative error. The number of calls
mω02 2 to the performance function and the relative errors for the
p
where ω0 = (c1 + c2 ) /m。 other three methods are similar. However, the AK-MCS-
where c1 and c2 represents the spring stiffness, m denotes the H method has the longest average solution time per run,
car mass, and r signifies the given threshold. at 130.72 s. Compared to the AK-MCS-U method, the
The parameters of each random variable are presented in AK-MCS-WU method reduces the number of calls to the
table 2. The random variables c1 , c2 , r, m, t and F (t) are all performance function to some extent while achieving the
independent and follow a normal distribution. smallest relative error in estimating the failure probability.

10
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 7. Modeling results of various methods when k = 7.

Figure 8. The convergence curve and relative error curve of the failure probability when k = 6.

11
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 9. The convergence curve and relative error curve of the failure probability when k = 7.

Table 1. The reliability calculation result of Case 1.

k Method Ncall Pf /10−3 CovPf /% εerror /% T/s

6 MCS 1 × 106 4.473 1.49 — 4.32

AK-MCS-U 12 + 59.3 4.437 1.50 0.80 150.13
AK-MCS-WU 12 + 51.1 4.500 1.49 0.60 143.83
AK-MCS-EFF 12 + 14.2 4.262 1.53 4.7 638.89
AK-MCS-H 12 + 32.8 4.673 1.46 4.5 688.21

7 MCS 1 × 106 2.223 2.12 — 4.13

AK-MCS-U 12 + 44.1 2.192 2.13 1.39 101.30
AK-MCS-WU 12 + 37.2 2.229 2.12 0.27 95.55
AK-MCS-EFF 12 + 14.6 2.271 2.10 2.16 655.91
AK-MCS-H 12 + 30.9 2.116 2.17 4.81 672.66

Figure 10. Box plot of the number of times various methods are called to a performance function.

12
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 11. Schematic diagram of a nonlinear oscillator system.

Table 2. Distribution of random variables.

Variable Distribution Mean value Standard deviation

c1 Normal 1 0.1
c2 Normal 0.1 0.01
r Normal 0.5 0.05
F1 Normal 1 0.2
m Normal 1 0.05
t1 Normal 1 0.2

Table 3. The reliability calculation results of Case 2.

Method Ncall Pf /10−2 CovPf /% εerror /% T/s

MCS 1 × 105 2.859 1.84 — 0.77

AK-MCS-U 12 + 59.4 2.872 1.84 0.45 32.03
AK-MCS-WU 12 + 53.9 2.849 1.85 0.35 27.30
AK-MCS-EFF 12 + 14.2 2.713 1.89 5.1 72.18
AK-MCS-H 12 + 52.9 2.837 1.85 0.77 130.72

Additionally, the box plots of the number of calls to the alumina and silicon carbide to ensure that they have high-
actual performance function by various methods are shown in temperature resistance, wear resistance, and other properties,
figure 12. enabling them to work stably in a high-temperature environ-
In a certain test, the convergence curves of the failure prob- ment of up to 1000 ◦ C. The physical model of the ceramic
ability and the relative-error curves of various methods are roller drive system is shown in figure 14, which mainly con-
shown in figure 13. sists of ceramic rollers, iron heads, drive shafts, bearing hous-
Figures 12 and 13 further prove that although the AK-MCS- ings, and sprockets.
EFF method has the fewest calls to the performance func- Due to the presence of numerous uncertain factors, ceramic
tion, its relative error in predicting the failure probability is rollers may experience failures during operation, which sig-
the largest. Compared with the AK-MCS-U method, the AK- nificantly impacts the safety and reliability of hot-forming
MCS-WU method has fewer calls to the performance func- equipment. This section thoroughly considers the inherent
tion, and the relative error of the obtained failure probabil- uncertainties in material dimensions, performance, and other
ity can converge well near 0, and the convergence speed is parameters. Using finite element simulation technology, the
faster. mechanical response of the ceramic roller during the heating
process is modeled to obtain response data. Finally, based on
the proposed AK-MCS-WU method, the failure probability
4.3. Engineering example: structural reliability of ceramic
roller of the ceramic roller is calculated. Since the solution process
involves finite element simulation, which cannot be directly
Ceramic rollers play a crucial role in the hot-forming pro- compared with the MCS method, the AK-MCS-U method is
cess of high-strength steel by transporting the billets. They used as the control group.
are key components of hot-forming equipment, and their per- In terms of dimensional parameters, the outer diameter and
formance has a significant impact on the entire hot-forming wall thickness of the ceramic roller directly determine its load-
process and product quality. Ceramic rollers are usually made bearing capacity and stress distribution. Regarding material
of high-purity high-temperature ceramic materials such as performance parameters, the coefficient of thermal expansion

13
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 12. Box plot of the number of times various methods are called to a performance function.

Figure 13. The convergence curve and relative error curve of the failure probability in example 2.

Figure 14. Physical model of a ceramic roller drive system.

is a key indicator for measuring the degree of linear expan- distribution of heat transfer within the ceramic roller, thereby
sion of materials during temperature changes. Its value dir- influencing the uniformity of its internal temperature field and
ectly affects the deformation of the roller during variations in ultimately having an indirect impact on its mechanical proper-
the temperature field and plays a crucial role in the generation ties. Furthermore, parameters such as the elastic modulus and
of thermal stress. Thermal conductivity governs the rate and Poisson’s ratio are also vital in the static reliability analysis

14
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Table 4. Distribution information of random variables in ceramic rollers.

Variable Distribution Mean value Standard deviation

R (mm) Normal 80 4
D (mm) Normal 10 1
α (10−6 /◦ C) Normal 5 0.5
K W (m∗ K)−1 Normal 2.5 0.2
E (GPa) Normal 85 5
ν Normal 0.15 0.05
ρ g cm3 Normal 3.5 0.4

Table 5. Random parameters and response values.

Random parameter
−6 ◦
Test number R (mm) D (mm) α (10 / C) K W (m∗ K)−1 E (GPa) ν ρ (g/cm3 ) Maximum stress (MPa)

1 77.61 9.62 5.39 2.38 84.64 0.17 3.23 560.85

2 82.11 7.51 5.43 2.19 88.91 0.19 3.84 554.31
3 82.02 10.44 5.87 2.56 81.85 0.13 3.64 602.40
4 76.71 10.64 3.97 2.51 94.55 0.18 3.49 468.26
5 76.41 11.79 4.94 2.61 89.72 0.13 3.04 524.68
.. .. .. .. .. .. .. .. ..
. . . . . . . . .
16 78.65 9.39 4.25 2.32 84.17 0.11 2.70 468.26
17 88.75 10.13 5.20 2.80 75.68 0.20 3.29 514.66
18 83.51 9.92 5.57 2.27 79.19 0.17 3.39 589.15
19 81.22 9.22 4.65 2.45 86.66 0.21 3.09 480.79
20 73.64 10.38 5.07 2.72 81.31 0.08 3.73 496.86

of ceramic rollers, as they define the roller’s response charac- response values of the initial sample points of the ceramic
teristics under stress. Based on the above analysis, the outer roller. The initial sample points and their response val-
diameter, wall thickness, coefficient of thermal expansion, ues are shown in table 5. An initial Kriging model is
thermal conductivity, Young’s modulus, and Poisson’s ratio established based on the random variables and the max-
are set as random variables, denoted by R, D, α, K, E, and ν, imum stress response values to form an initial sample
respectively. The distributions and specific parameters of each set.
random variable are shown in table 4. The simulation stress contour diagrams of the data in
Since the ceramic roller mainly bears axial compressive Group 1 and Group 16 in table 5 are shown in figure 15.
stress during the heating process, the axial stress of the ceramic The maximum stress values are 560.85 MPa and 468.26 MPa
roller is used as the reliability evaluation index, and the per- respectively.
formance function of the ceramic roller is established as After constructing the initial Kriging model, the WU active
follows: learning function is used for the iterative calculation of can-
didate sample points, and the best sample points are added to
G(X) = σ max (X) − σ c (35) the initial sample set. The finite-element simulation is called
to obtain the maximum stress response values. The Kriging
where X is a random variable, including random variables model is updated by continuously selecting points through the
such as the outer diameter, wall thickness, thermal expan- WU learning function until the stop criterion is met. Finally,
sion coefficient, and thermal conductivity of the ceramic roller, a Kriging model that meets the accuracy requirements is
X = (R, D, α, K, E, υ, ρ)T ; σmax (·) is the actual maximum axial obtained, and the failure probability of the ceramic roller is
stress of the ceramic roller, obtained through ANSYS simu- predicted based on this. The calculation results are shown in
lation; σc is the critical axial stress of the ceramic roller. By table 6, and the convergence curve of the failure probability of
referring to relevant materials, σc = 2990 MPa. a certain test is shown in figure 16.
A sample space of NMCS = 106 is given. The LHS sampling Under the relevant parameters of the ceramic roller set in
method is used to sample the random input variables of this paper, the failure probability of the ceramic roller obtained
the ceramic roller, and 20 groups of initial sample points by the AK-MCS-WU method is 4.874 × 10−4 , which is sim-
are extracted. The ANSYS parametric design language is ilar to the calculation result of the AK-MCS-U method. In
used to parametrically design the maximum axial stress terms of the number of calls to the finite-element simulation
of the ceramic roller to obtain the maximum axial stress during the calculation process, the AK-MCS-WU method has

15
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 15. Initial sample point simulation stress contours.

Table 6. Calculation results of static failure probability of ceramic rollers.

Methon Ncall Pf /10−4 Cov (Pf ) /%

AK-MCS-U 20 + 108.2 4.575 4.54

AK-MCS-WU 20 + 96.4 4.874 4.67

certain advantages and can ensure the stable convergence of developing the WU-learning function, selecting an appropri-
the failure probability. ate adaptive directional function is crucial. This paper com-
pares the effects of different adaptive directional functions
on the selection of sample points by the WU learning func-
5. Conclusion tion. The final chosen adaptive directional function is ρ2 (x) =
1 + exp(− µĜ (x) ), and based on this, the AK-MCS-WU reli-
Currently, the Kriging model is an important surrogate model ability analysis method is established.
for structural reliability analysis. Based on the U learning In two numerical examples, although the AK-MCS-EFF
function, the WU learning function considers the influence of and AK-MCS-H methods can effectively reduce the number
both the predicted values and the prediction variances of the of calls to the performance function, their relative error in pre-
sample points on the accuracy of failure probability estima- dicting the failure probability is relatively large. Moreover,
tion. It prioritizes selecting sample points closer to the LSS in these two methods require longer average solution times per
the sample space and assigns a larger weight to the prediction run, resulting in higher time costs. Compared to the AK-MCS-
variance of the sample points near the LSS, thereby improv- U method, the AK-MCS-WU method can further reduce the
ing the efficiency of sample point selection. In the process of number of calls to the performance function while ensuring

16
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Figure 16. Convergence curve of failure probability of ceramic roller.

the accuracy of failure probability prediction and maintain- References

ing lower time costs. Engineering case studies further demon-
strate that the AK-MCS-WU method, when applied to reliab- [1] Zhao W, Chen Y and Liu J 2020 An effective first order
reliability method based on Barzilai–Borwein step Appl.
ility problems with complex performance functions, ensures
Math. Modelling 77 1545–63
stable convergence of failure probability, reduces the num- [2] Keshtegar B and Chakraborty S 2018 A hybrid self-adaptive
ber of calls to the performance function, and offers significant conjugate first order reliability method for robust structural
practical value. reliability analysis Appl. Math. Modelling 53 319–32
The AK-MCS-WU structural reliability analysis method [3] Wang Y, Hao P, Yang H, Wang B and Gao Q 2020 A
proposed in this paper provides high-accuracy failure prob- confidence-based reliability optimization with single loop
strategy and second-order reliability method Comput.
ability predictions, but it also has certain limitations. First, Methods Appl. Mech. Eng. 372 113436
compared to the AK-MCS-EFF and AK-MCS-H methods, this [4] Meng Z, Zhou H, Hu H and Keshtegar B 2018 Enhanced
approach still requires a relatively large number of calls to sequential approximate programming using second order
the performance function. When dealing with complex finite reliability method for accurate and efficient structural
element simulation models, frequent calls to the finite ele- reliability-based design optimization Appl. Math. Modelling
ment simulations result in higher computational costs. Second, 62 562–79
[5] Zhu Z and Du X 2016 Reliability analysis with Monte Carlo
the WU learning function proposed in this paper has weights simulation and dependent Kriging predictions J. Mech. Des.
that are only related to the predicted values throughout the 138 121403
entire training phase. In future research, we plan to intro- [6] Papaioannou I, Breitung K and Straub D 2018 Reliability
duce a dynamic weight adjustment mechanism based on train- sensitivity estimation with sequential importance sampling
ing progress. In the early stages of training, the focus will Struct. Saf. 75 24–34
be on enhancing the global search capability of the learn- [7] Xiao S, Oladyshkin S and Nowak W 2020 Reliability analysis
with stratified importance sampling based on adaptive
ing function, encouraging the algorithm to more effectively Kriging Reliab. Eng. Syst. Saf. 197 106852
explore potential multiple failure regions. In the later stages of [8] Chan J, Papaioannou I and Straub D 2022 An adaptive subset
training, the emphasis will shift to improving its local search simulation algorithm for system reliability analysis with
capability, allowing for more efficient extraction of local discontinuous limit states Reliab. Eng. Syst. Saf. 225 108607
information and accelerating convergence. This approach will [9] Lima J P S, Evangelista F Jr and Soares C G 2023
comprehensively improve the overall performance of the Hyperparameter-optimized multi-fidelity deep neural
network model associated with subset simulation for
method. structural reliability analysis Reliab. Eng. Syst. Saf.
239 109492
[10] Valdebenito M A, Wei P, Song J, Beer M and Broggi M 2021
Failure probability estimation of a class of series systems by
Acknowledgments multidomain line sampling Reliab. Eng. Syst. Saf.
213 107673
This research was supported by the 2022 industrial technology [11] Ameryan A, Ghalehnovi M and Rashki M 2022 AK-SESC: a
basic public service platform (No. 2022-232-223-02). novel reliability procedure based on the integration of active

17
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

learning kriging and sequential space conversion method [31] Wang Z and Shafieezadeh A 2019 ESC: an efficient
Reliab. Eng. Syst. Saf. 217 108036 error-based stopping criterion for kriging-based reliability
[12] Liu Z and Choe Y 2021 Data-driven sensitivity indices for analysis methods Struct. Multidiscip. Optim. 59 1621–37
models with dependent inputs using polynomial chaos [32] Zhang Y, Dong Y and Xu J 2023 An accelerated active
expansions Struct. Saf. 88 101984 learning Kriging model with the distance-based subdomain
[13] Zheng S, Jiang A N and Feng K S 2021 A reliability and a new stopping criterion for reliability analysis Reliab.
evaluation method for intermittent jointed rock slope based Eng. Syst. Saf. 231 109034
on evolutionary support vector machine Comput. Model. [33] Peijuan Z, Ming W C, Zhouhong Z and Liqi W 2017 A new
Eng. Sci. 129 149 active learning method based on the learning function U of
[14] Hong L, Li H and Peng K 2021 A combined radial basis the AK-MCS reliability analysis method Eng. Struct.
function and adaptive sequential sampling method for 148 185–94
structural reliability analysis Appl. Math. Modelling [34] Meng Y, Zhang D, Shi B, Wang D and Wang F 2024 An active
90 375–93 learning Kriging model with approximating parallel
[15] Ren C, Aoues Y, Lemosse D and De Cursi E S 2022 Ensemble strategy for structural reliability analysis Reliab. Eng. Syst.
of surrogates combining Kriging and artificial neural Saf. 247 110098
networks for reliability analysis with local goodness [35] Cao R N, Sun Z L, Zhang Y B and Wang J 2020 Reliability
measurement Struct. Saf. 96 102186 analysis based on active learning for complex mechanical
[16] El Haj A-K and Soubra A-H 2021 Improved active learning structure J. Northeastern Univ. 41 223
probabilistic approach for the computation of failure
probability Struct. Saf. 88 102011
[17] Ma Y-Z, Jin X-X, Wu X-L, Xu C, Li H-S and Zhao Z-Z 2023 Hongyan Chu received the Ph.D. degree in mech-
Reliability-based design optimization using adaptive anical design and theory from Beijing University
Kriging-A single-loop strategy and a double-loop one of Technology, Beijing, China, in 2003. She is
Reliab. Eng. Syst. Saf. 237 109386 currently a professor of the Beijing Key Laboratory
[18] Krige D G 1951 A statistical approach to some basic mine of Design and Intelligent Machining Technology for
valuation problems on the Witwatersrand J. South. Afr. Inst. High Precision Machine Tools, Beijing University
Min. Metall. 52 119–39 of Technology. She has published more than 40
[19] Matheron G 1963 Principles of geostatistics Econ. Geol. SCI/EI papers. Her research interests include
58 1246–66 reliability and accuracy retention of manufacturing
[20] Lophaven S N, Nielsen H B and Søndergaard J 2002 DACE-A equipment, intelligent manufacturing technology,
matlab Kriging toolbox version 2.0 and optimization algorithm.
[21] Kaymaz I 2005 Application of kriging method to structural
reliability problems Struct. Saf. 27 133–51 Wenzhong Li received a bachelor’s degree from
[22] Echard B, Gayton N and Lemaire M 2011 AK-MCS: an active Beijing University of Technology (Beijing, China) in
learning reliability method combining Kriging and Monte 2018. He is currently pursuing a master’s degree at
the College of Mechanical and Energy Engineering,
Carlo simulation Struct. Saf. 33 145–54
Beijing University of Technology (Beijing, China).
[23] Huang X, Chen J and Zhu H 2016 Assessing small failure
His main research interests are the innovation and
probabilities by AK–SS: an active learning method application of reliability analysis methods.
combining Kriging and subset simulation Struct. Saf.
59 86–95
[24] Echard B, Gayton N, Lemaire M and Relun N 2013 A
combined importance sampling and kriging reliability
method for small failure probabilities with time-demanding Dongyang Sun received his master’s degree from
numerical models Reliab. Eng. Syst. Saf. 111 232–40 Beijing University of Technology (Beijing, China) in
[25] Bichon B J, Eldred M S, Swiler L P, Mahadevan S and 2019. He is currently pursuing his doctor’s degree at
McFarland J M 2008 Efficient global reliability analysis for the College of Mechanical and Energy Engineering,
nonlinear implicit performance functions AIAA J. Beijing University of Technology (Beijing, China).
46 2459–68 His research interests include vacuum furnace
[26] Lv Z, Lu Z and Wang P 2015 A new learning function for reliability design and evaluation.
Kriging and its applications to solve reliability problems in
engineering Comput. Math. Appl. 70 1182–97
[27] Zhang X, Wang L and Sørensen J D 2019 REIF: a novel
active-learning function toward adaptive Kriging surrogate
models for structural reliability analysis Reliab. Eng. Syst.
Zhifeng Liu received the Ph.D. degree in mechanical
Saf. 185 440–54
design from the Mechanical Engineering and
[28] Shi Y, Lu Z, He R, Zhou Y and Chen S 2020 A novel learning Automation College, Northeastern University, in
function based on Kriging for reliability analysis Reliab. 2001. He is currently a professor of Key Laboratory
Eng. Syst. Saf. 198 106857 of CNC Equipment Reliability, Ministry of
[29] Yang X, Liu Y, Gao Y, Zhang Y and Gao Z 2015 An active Education, School of Mechanical and Aerospace
learning kriging model for hybrid reliability analysis with Engineering, Jilin University; and a distinguished
both random and interval variables Struct. Multidiscip. Professor of Institute of Advanced Manufacturing
Optim. 51 1003–16 and Intelligent Technology, Beijing University
[30] Sun Z, Wang J, Li R and Tong C 2017 LIF: a new Kriging of Technology. His research interests include
based learning function and its application to structural robotic motion control, machine vision technology,
reliability analysis Reliab. Eng. Syst. Saf. 157 152–65 networked manufacturing, and manufacturing system information.

18
J. Reliab. Sci. Eng. 1 (2025) 025003 H Chu et al

Weixu Zheng is a postgraduate student at Beijing Qiang Cheng received the Ph.D. degree from
University of Technology. His research interests Huazhong University of Science and Technology
include time-variant reliability and reliability in 2009. He currently serves as a professor and Ph.D.
prediction. supervisor at Beijing University of Technology. He
has published more than 100 SCI papers and been
authorized with over 80 invention patents in the UK,
Japan, and China. His main research interests include
high-end manufacturing equipment, intelligent
manufacturing, and medical robots.

Jingjing Xu received the Ph.D. degree in mech-

anical engineering from the Beijing University of
Technology, in 2020. She is currently a Lecturer
with the Institute of Advanced Manufacturing and
Intelligent Technology, Faculty of Materials and
Manufacturing, Beijing University of Technology.
Her research interests include the robotic control
and arms coordination. She has made some progress
in the studying on kinematic analysis and motion
control of redundant manipulators.

19