Reliability Engineering and System Safety 216 (2021) 108020

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

journal homepage: www.elsevier.com/locate/ress

Reliability index function approximation based on adaptive double-loop

Kriging for reliability-based design optimization
Xiaobo Zhang , Zhenzhou Lu *, Kai Cheng
Northwestern Polytechnical University, School of Aeronautics Xi’an, Shaanxi 710072, China

A R T I C L E I N F O A B S T R A C T

Keywords: Reliability-based design optimization (RBDO) aims at minimizing general cost under the reliability constraints by
Reliability-based design optimization considering the inherent uncertainties in engineering. In this work, we develop a decoupled RBDO approach
Decoupled approach named RIFA-ADK which aims at reliability index function (RIF) approximation by adaptive double-loop Kriging.
Reliability index function
The proposed RIFA-ADK contains three main blocks, namely reliability analysis (Block 1), reliability index
Gradient-enhanced Kriging
Adaptive learning
function approximation (Block 2) and optimization (Block 3). In RIFA-ADK, RIF is approximated by the outer
loop adaptive gradient-enhanced Kriging (GEK) model which takes into account reliability sensitivity in addition
to reliability index. The required reliability analysis in GEK is based on the inner loop adaptive Kriging model
which focuses on approximating the performance function, and the required reliability sensitivity analysis in
GEK is a post-processing of reliability analysis. Then the optimization can be proceeded using the cheap GEK
model of RIF. In addition, an adaptive learning strategy which involves two stages of enrichment is also
developed to improve the surrogate precision in the region of interest. Finally, four mathematical and practical
engineering examples for RBDO are presented to illustrate the accuracy and the efficiency of the proposed RIFA-
ADK decoupled approach.

1. Introduction the reliability constraints are replaced by the reliability index con­
straints using the first-order reliability method (FORM). The perfor­
The inherent uncertainties in manufacturing tolerances, material mance measure approach (PMA) [8,9] transforms the reliability index
properties, loads or environmental and operational conditions, etc. constraints into equivalent target performance measure constraints
wildly exist in practical engineering problems [1,2]. Traditional deter­ through an inverse FORM analysis. Mono-level approaches aim at
ministic optimization design takes into account these uncertainties avoiding the reliability analysis in RBDO. In the earlier studies, some
through empirical safety factors. However, such empirical safety factors mono-level approaches based on Karush-Kuhn-Tucker (KKT) optimality
often lead to conservative and uneconomical design. In order to find a conditions have been developed in [10,11]. Then the single loop single
better compromise between cost and safety, the reliability-based design vector (SLSV) approach and single-loop approach (SLA) are proposed by
optimization (RBDO) which contains reliability constraints (probability Chen et al. [12] and Liang et al. [13] respectively, where the reliability
constraints) has been researched in the literature [3,4]. Various methods index constraints are transformed into approximate deterministic con­
have been developed for the solution of RBDO. We classify them into straints. The decoupled approaches aim at transforming the RBDO
two categories in this article: Classical approaches based on approxi­ problem into a sequence of deterministic optimization. Sequential
mation methods and surrogate model-based approaches [5]. optimization and reliability assessment (SORA), one of the most popular
Classical RBDO methods mainly include two-level approach, the decoupled approaches, is proposed by Du and Chen [14]. The deter­
mono-level approach and the decoupled approach. The most direct way ministic constraints related to the minimum performance target point
for RBDO is the two-level approach which considers the reliability (MPTP) are used to replace the reliability index constraints in SORA.
constraints inside the optimization loop. The outer loop is the optimi­ Cheng et al. [15] developed the sequential approximate programming
zation algorithm while the reliability analysis in the inner loop is per­ (SAP) to improve the efficiency of RBDO. A second A second order SAP
formed when it is required. In the reliability index approach (RIA) [6,7], algorithm is proposed by Torii et al. [16] to overcome the drawbacks of

* Corresponding author
E-mail address: [email protected] (Z. Lu).

https://doi.org/10.1016/j.ress.2021.108020
Received 4 June 2021; Received in revised form 20 August 2021; Accepted 26 August 2021
Available online 1 September 2021
0951-8320/© 2021 Elsevier Ltd. All rights reserved.
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

the first order SAP algorithm. Despite the Classical RBDO methods are Different from existing surrogate model-based approaches, this
very efficient in some cases, some limits cannot be ignored. When the paper develops a decoupled RBDO approach named RIFA-ADK which
initial design point of the optimization is far from the optimal design aims at the reliability index function (RIF) approximation by adaptive
point and the performance function is highly non-linear, classical RBDO double-loop Kriging. Inner loop Kriging is used to perform reliability
methods may fail to converge [5]. analysis. The gradient-enhanced Kriging (GEK) model [17], which al­
The surrogate model aims at constructing a cheap approximation lows us to take into account gradient information in addition to function
through the design of experiment (DOE) to replace the expensive-to- response, is used to approximate RIF. After the GEK model of RIF is
evaluate input-output relationship [17]. Various surrogate models constructed, the RBDO problem is transformed into an equivalent
have been applied to reliability and RBDO problems in recent years [5, deterministic optimization problem, and it can be solved by any suitable
18-20]. These RBDO approaches use different types of surrogate models optimization algorithm.
to approximate performance function or failure probability function This paper is formulated as follows. Formulation of RBDO and
with combination of different reliability analysis technologies and Kriging model are given in Section 2. The proposed decoupled RBDO
classical RBDO approaches. approach is given in Section 3. Section 4 presents reliability analysis
Most surrogate model-based RBDO methods focus on constructing an based on the adaptive Kriging model. RIF approximation by adaptive
approximation of the performance function [21-23]. A direct approach GEK model is proposed in Section 5. Four examples are employed to
is that the required reliability analysis in the inner loop of two-level demonstrate the rationality and efficiency of the proposed method in
approach is implemented by locally building surrogate models of the Section 6. Conclusions are summarized in section 7.
performance function. For instance, Agarwal and Renaud [24] built a
most probable point (MPP)-based second-order response surface model 2. The formulation of RBDO and Kriging model
to implement approximate reliability analysis in each inner loop.
Papadrakakis et al [25]. built a neural network to implement Monte 2.1. RBDO formulation
Carlo simulation in each inner loop. This kind of approach is inefficient
since multiple surrogate models need to be built in the process of opti­ RBDO aims to minimize general cost function (i.e. objective func­
mization. Recently, surrogate models combined with adaptive sampling tion) under probability constraints (reliability constraints). A typical
in global region are developed to solve RBDO problem. The purpose of RBDO formulation reads as follows:
adaptive sampling is to improve the accuracy of surrogate models in the
min : c(d)
region of interest. For instance, Lee and Jung [26] proposed the d
⎧
constraint boundary sampling (CBS) criterion for Kriging model by ⎪ P (d) = Pr[gi (X(d)) ≤ 0] ≤ PTfi i = 1, ⋯, m
⎪
⎪
adaptively sampling in the whole design space and implemented RBDO ⎨ fi (1)
by advanced mean-value (AMV) method. Chen et al. [27] and Wu et al. s.t. : hl (d) ≤ 0l = 1, ⋯, M
⎪
⎪
[28] proposed the importance boundary sampling (IBS) criterion for ⎪
⎩ dL ≤ d ≤ dU
Kriging model by adaptively sampling in the region of importance
boundary and the RBDO is solved by SORA method. Similarly, local where c is the cost function and hl is the deterministic soft constraint.
adaptive sampling (LAS) [29] and local most probable point (LMPP) The performance of a system is assessed by the performance function
[30] criterions for Kriging model are proposed by adaptively sampling in g(X). The vector X denotes the random variables with the joint proba­
the region of interests. The local ranging sampling and support vector bility density function (PDF) fX (x). X(d)={W(d), Z} contains two types
machine (LR-SVM) method [31] was developed to further improve the of random variables: random design variables W(d) and environmental
efficiency of Kriging-based RBDO with assistance of SVM. In recent random variables Z. Design variables d are considered as a portion of
years, the concept of augmented reliability space which links the design distribution parameters (e.g. the mean and standard deviation) of the
apace and the reliability space has attracted more attention [32]. Zhang random design variables W. The distribution parameters of environ­
et al. [33] built the Kriging model in the augmented input space and mental random variables Z are known. In addition to deterministic
adopted a sequential approximate optimization strategy for RBDO. constraints, probability constraints {Pf (d) = Pr[gi (X(d)) ≤ 0] ≤ PTfi ,
Dubourg et al. [34] solved the RBDO problem by combining augmented
i = 1, ⋯, m} are also considered in RBDO. PTfi is prescribed target failure
reliability space, adaptive Kriging and subset simulation. A
quantile-based RBDO problem is formulated by Moustapha et al [35], in probability, i.e. the maximum acceptable failure probability.
which Kriging model is constructed adaptatively in an augmented reli­
ability space. 2.2. Reliability index function (RIF)
In addition, several surrogate model-based RBDO methods aim to
approximate the failure probability function (resp. reliability index The failure probability for a specific design variable d can be
function) which gives the relationship between design variables and the expressed by the following multidimensional integrals:
corresponding failure probability (resp. reliability index). For instance, ∫
Foschi et al. [36] adopt response surface model to approximate the Pf (d)= IF (x)fX (x|d)dx (2)
relationship between different design choices and the corresponding Rn

reliability indexes estimated by importance sampling. Lehky et al. [37] where IF (x) is the indicator function of the failure domain F = {x : g(x) ≤
developed a double-loop RBDO approach where the failure probability 0} as shown in Eq.(3).
function is approximated by artificial neural networks and the required {
failure probability is estimated by FORM method. In these RBDO ap­ IF (x) =
1x ∈ F
(3)
proaches, multiple reliability analysis based on true performance func­ 0otherwise
tion are required to generate training set. It is time-consuming especially
Eq.(2) can be estimated by Monte Carlo Simulation (MCS). Various
for expensive-to-evaluate performance functions. And these RBDO ap­
efficient reliability analysis method also can be used to estimate Eq.(2),
proaches lack adaptive sampling in the process of constructing surrogate
e.g., FORM, importance sampling (IS) and Active Kriging-based MCS
model of failure probability function.
(AK-MCS) method [20]. For practical applications, the transformation

2
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

between Pf (d) and the reliability index function (RIF) can be expressed [ ]
∑
n
by identical equation in Eq.(4). ( )
R x(i) , x(j) , θ = exp − θk |x(i) − x (j) 2
| (8)
k k
( ) k=1
β(d) = − Φ− 1 Pf (d)
(4)
Pf (d) = Φ(− β(d))
where xk denotes the kth dimension component of the training point
(i)

where Φ(⋅) is the cumulative distribution function (CDF) of the standard x , θ = [θ1 , θ2 , ..., θn ]T is a vector of the so-called hyperparameters
(i)

normal distribution. Φ− 1 (⋅) is the inverse CDF. β(d) reflects the safety driving the auto-correlation function, which can be estimated by the
level of the structure. In fact, Eq.(4) is a mapping relation between β(d) Maximum Likelihood Estimation.
and Pf (d). For different reliability analysis methods of MCS, FORM and Gradient-enhanced Kriging (GEK) model [17] which can account for
AK-MCS, the corresponding failure probability is denoted as PMCS (d), function gradient information was introduced by Morris et al. [38], and
f
has been wildly applied to many fields.
PFORM (d) and PAK− MCS
(d) respectively, and the corresponding reliability
f f GEK is very similar to the classical Kriging approach. For a problem
index is denoted as βMCS (d), βFORM (d) and βAK− MCS (d) respectively. with n-dimensional variables, given a design of experiments (DOE) with:
Obviously, the accuracy of β(d) depends on Pf (d) estimated by the [
selected reliability analysis method. SDOE = x(1) , ⋯, x(i) , ⋯, x(m) ,
]T (9)
Based on the relationship between Pf and β in Eq.(4), Eq.(1) can also x , ⋯x , ⋯x(i) , ⋯x(i) , ⋯, x(m) , ⋯x(m)
(1) (1)

be rewritten equivalently as follows: [ ( ) ( )

yS = y x(1) , ⋯, y x(i) , ⋯, y(x(m) ),
min : c(d) ) ) ) )
d
⎧ ∂y(x(1) ∂y(x(1) ∂y(x(i) ∂y(x(i)
⎨ β (d) ≥ βT i = 1, ⋯, m
i i
(5) ∂x1
, ⋯,
∂xn
, ⋯,
∂x1
, ⋯,
∂xn
,
(10)
s.t. : ]
⎩ hj (d) ≤ 0j = 1, ⋯, M ∂y(x(m) ) ∂y(x(m) ) T
⋯, , ⋯,
∂x1 ∂xn
where βi (d) is the i-th reliability index function. βTi is the i-th target
reliability index corresponding to PTfi , i.e. βTi = − Φ− 1 (PTfi ). βTi represents where x(i) = [x1 ,⋯,xj ,⋯,xn ], i = 1,⋯,m. y(x(i) ) is the model response
(i) (i) (i)

the minimum safety requirement. at x(i) .∂y(x

∂xj is partial derivative of y with respect to the j-th component xj
(i)
)

It should be mentioned that the traditional RIA is a special case of Eq.

at x(i) . Therefore, the DOE contains m-size model responses and mn-size
(5). The RIA adopts a double-loop procedure and FORM is used as
model gradient information.
reliability analysis method. βFORM (d) is obtained by solving a con­
The best linear unbiased predictor (BLUP) of the sample response
strained optimization problem and then the corresponding failure
and gradient information makes the Kriging model. This predictor is a
probability is given by PFORM
f (d) = Φ( − βFORM (d)). For other reliability linear weighting of the sample response and gradient information at the
analysis methods (e.g., AK-MCS), the failure probability PAK−
f
MCS
(d) is sample points:
estimated first and then corresponding reliability index is given by ∑
m
( ) n ∑
∑ m
∂y(x(i) )
βAK− MCS (d) = − Φ− 1 (PAK−
f
MCS
(d)). y GEK (x) =
̂ ωi (x)y x(i) + λ(i)
j (x) (11)
∂xj
The lower and upper bounds of the failure probability function Pf (d)
i=1 j=1 i=1

are 0 and 1. In general, PTf is a small value (10− 2 ∼ 10− 5 ). According to In the case of gradient-enhanced Kriging, the covariances satisfy:
our experience, due to the magnitude difference in the failure proba­ [ ( ) ( )] (
Cov z x(i) , z x(j) = σ 2 R x(i) , x(j) , θ
)
bility corresponding to different design variables, it is difficult to
[ )] )
construct an accurate surrogate model of Pf (d). According to the rela­ ( ) ∂z(x(j) ∂R(x(i) , x(j) , θ
Cov z x(i) , = σ2
tionship between Pf (d) and β(d) in Eq.(4), the value range of β(d) is ∂xk ∂xk(j)
generally between 0 and 6 in the active feasible domain. Therefore, this [ ) ] (12)
paper aims at constructing explicit expression of RIF β(d) by double-loop = − Cov
∂z(x(i) ( (j) )
,z x
Kriging model. ∂xk
[ (i)
) )
(j) ]
)
∂z(x ∂z(x ∂R(x(i) , x(j) , θ
2.3. Kriging and gradient-enhanced Kriging (GEK) Cov , = − σ2
∂xk ∂xl (i)
∂xk ∂xl (j)

In general, Kriging model yK (x) is given as: Similar as classical Kriging model, the prediction ̂
y GEK (x) of y(x) in a
T
yK (x) = f (x)β + z(x) (6) given unknown point x for Gradient-Enhanced Kriging Model also can be
derived as:
where f T (x)β is the regression model, f (x) = [f1 (x), f2 (x), ..., fp (x)]T de­ ( )
y GEK (x) = ̂
̂ β + rTc (x)R−c 1 yS − ̂
βFc (13)
notes the vector of regression basis functions, β(x) =
T
[β1 (x), β2 (x), ..., βp (x)] denotes the vector of regression coefficients. In where
this paper, ordinary Kriging in which f T (x)β = β is a scalar is selected.
Fc = [1, ⋯, 1, 0, 0, ⋯, 0]T ∈ ℜm+mn (14)
z(x) is a stationary Gaussian process with a mean of zero and a standard
deviation of σ . Its covariance matrix is: ( )− 1
[ ( ) ( )] ( )
̂
β = FTc R−c 1 Fc FTc R−c 1 yS (15)
Cov z x(i) , z x(j) = σ 2 R x(i) , x(j) , θ (7)
Rc is the (mn + m)2 -size correlation matrix between each pair of
where R(x , x ) is the auto-correlation function. In this paper, the
(i) (j) points of the DOE.
anisotropic Gaussian model is selected to define the correlation function.
It can be expressed as follows:

3
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

⎡ ⎤ loop Kriging, RIFA-ADK) are presented in the following.

R0 RWY
Rc = ⎣ ⎦ First, the proposed decoupled approach contains three main blocks,
RTWY RWW namely reliability analysis (Block 1), reliability index function approx­
imation (Block 2) and optimization (Block 3).
⎡ R( x(1) , x(1) ) )
R(x(1) , x(m) ⎤
⋯ Second, the Kriging model used in the proposed framework is a
⎢
⎢
⎥
⎥ double-loop procedure, which is summarized below.
⎢ ⎥
R0 = ⎢ ⎥
⎢
⎢ ⋮ ⋱ ⋮ ⎥
⎥ l Outer loop: Build a GEK model βGEK (d) for RIF β(d) in block 2. The
⎣ ⎦ region of interest is the feasible region boundary with a low-cost
( )
R x(m) , x(1) ⋯ R(x(m) , x(m) ) value. The required gradient information is provided by reliability
) ) ) sensitivity analysis which is usually the by-product of reliability
⎡ ∂R(x(1) , x(1) ∂R(x(1) , x(1)
⋯
∂R(x(1) , x(m) ⎤ analysis.
⎢ ∂x(1)
1 ∂x2(1) ∂x(m)
n ⎥ l Inner loop: Build a Kriging model gK (x) for the performance function
⎢ ⎥
⎢ ⎥ g(x) in block 1 to perform reliability analysis. The region of interest is
⎢ ⎥
RWY =⎢ ⋮ ⋮ ⋮ ⋮ ⎥ the limit state surface.
⎢ ⎥
⎢ ⎥ (16)
⎢ ⎥
⎣
∂R(x(m) , x(1)
)
∂R(x(m) , x(1)
)
∂R(x(m) , x(m) )
⎦ Finally, an adaptive learning strategy for the double-loop Kriging
(1) (1)
⋯ model is developed to improve the surrogate precision on the region of
∂ x1 ∂x2 ∂x(m)
n
interest. In Block 1, the inner loop Kriging model gK (X) contains two
( ) ( ) ( )
2
⎡ ∂ R x ,x (1) (1) 2
∂ R x(1) , x(1) ∂2 R x(1) , x(m) ⎤ stages of enrichment. The global stage of enrichment is implemented in
(1) (1)
⋯ the augmented reliability space to construct a rough Kriging model
⎢ ∂x1 ∂x1 ∂x1(1) ∂x(1) ∂x1(1) ∂x(m) ⎥
⎢
2 n
⎥ gK (X). The local stage of enrichment is implemented to update the
⎢ ⎥
⎢
⎢ 2 ( (1) (1) ) ( ) ( )⎥⎥ Kriging model gK (X) in the specific reliability space defined by the
2 2
specific design point only when necessary. The detailed description is
(1) (1) (1) (m)
⎢ ∂ R x ,x ∂ R x ,x ∂ R x ,x ⎥
⎢ ⋯ ⎥
RWW
⎢ ∂x(1) ∂x(1)
=⎢ 2 1 ∂x2(1) ∂x(1)
2
(1) (m)
∂x2 ∂xn ⎥
⎥ presented in Section 4. In Block 2, the outer loop gradient-enhanced
⎢ ⎥
⎢ ⎥ Kriging model βGEK (d) also contains two stages of enrichment. The
⎢ ⎥
⎢ ⋮ ⋮ ⋱ ⋮ ⎥ first stage of enrichment aims at adding more training samples around
⎢ ⎥
⎢
⎢
⎥
⎥ the feasible region boundary with low cost value. The second stage of
⎣ 2 ( (1) (m) )
∂ R x ,x
(
∂2 R x(1) , x(m)
) 2
(
∂ R x ,x(1) (m)
) ⎦ enrichment focuses on the region around the optimal solution obtained
⋯ by Block 3. The detailed description is presented in Section 5. The
∂xn(m) ∂x1(1) ∂x(m) (1)
n ∂x2
∂x(m) (m)
n ∂xn
flowchart of the proposed decoupled approach is shown in Fig. 1.
rc is the correction vector between x and the points of the DOE.
[ ( ) 4. Reliability analysis based on adaptive Kriging model
rc = R x, x(1) , ⋯, R(x, x(m) ),
]T (17) In Block 1, the inner loop Kriging model gK (X) contains two stages of
∂R(x, x(1) ) ∂R(x, x(1) ) ∂R(x, x(m) )
(1)
, (1)
, ⋯, enrichment: the global stage of enrichment in the augmented reliability
∂x1 ∂x2 ∂xn(m) space and the local stage of enrichment in the specific reliability space.
The Kriging variance σ 2 (x) also can be obtained as: In general, the adaptive Kriging and MCS (AK-MCS [20]) method for
̂y GEK
reliability analysis mainly contains three parts: sample pool, adaptive
( ))
σ̂2y (x) = ̂
( )T (
σ 2 1 − rTc R−c 1 rc + FTc R−c 1 rc − 1 FTc R−c 1 rc
)− 1 ( T − 1
Fc Rc rc − 1 learning strategy for the design of experiment (DOE) and stopping cri­
GEK terion [20,40,41].
(18)

where 4.1. Global stage of enrichment-construction of a Kriging model in the

( )T ( ) augmented reliability space
1
σ2 =
̂ y − Fc ̂
β R−c 1 yS − Fc ̂
β (19)
m(n + 1) S The first stage of enrichment is implemented in the augmented
GRENAT MATLAB toolbox [39], which stands for gradient enhanced reliability space defined by Au SK [32]. Design variables d are artificially
approximation toolbox, is used in this paper. considered as random variables with PDF f(d). A uniform distribution
U(dL , dU ) is chosen for convenience in this paper. Augmented failure
3. The proposed decoupled RBDO approach probability Pf is defined by Eq.(20).
∫ ∫
RBDO considers uncertainty, making it more complex than deter­ Pf = IF (x)fX|d (x)f (d)dxdd (20)
ministic optimization. Many existing RBDO methods couple the opti­
mization process and the reliability analysis process together, making it It can be seen from Eq.(20) that Pf is the mean of the failure prob­
hard to understand especially for general engineering practitioners. ability function Pf (d) with respect to artificially random variables d with
A direct way to solve RBDO is to approximate the relationship be­ PDF f(d). Therefore, Pf contains “global ” information of Pf (d).
tween design variables and the corresponding reliability index, i.e. Sample pool: A set of samples S = {x(i) , i = 1, ⋯, N} are generated
reliability index function (RIF) β(d) by the cheap surrogate model. Then with PDFs f(d) and fX|d (x). Randomly select a set of samples from the
the optimization can be proceeded using the cheap surrogate model of sample pool as an initial DOE, a.k.a training set .
RIF. Reliability sensitivity (gradient information) can be easily obtained Adaptive learning strategy: In reliability analysis, the region of in­
as a post-processing of reliability analysis (This point is described in terest is the vicinity of the limit-state surface g(x) = 0. Several learning
detail in Section 4). Therefore, the cheap gradient information can be functions have been proposed to adaptive sampling, e.g., the Expected
used to improve the accuracy of surrogate model of RIF through GEK Feasibility Function (EFF) [42], U learning function [20] and H learning
model. The key points of the proposed decoupled RBDO approach function [43]. U learning function is selected in this paper since it is easy
(Reliability index function approximation based on adaptive double- to implement and possesses outstanding performance. U learning

4
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

( ( ))
xnext = arg min U x(i) (22)
x(i) ∈S

Stopping criterion: The enrichment is stopped when U(xnext ) > 2.

This stopping condition means that the probability of misjudging the
sign of performance function is less than Φ( − 2) = 0.023.

4.2. Local stage of enrichment

The Kriging model gK (x) constructed in the augmented reliability

space can get a roughly precision. The accurate failure probability Pf (d∗ )
at some specific design variables d∗ is needed when constructing the
outer loop GEK model βGEK (d). The local stage of enrichment with
respect to some specific design variables d∗ can be implemented based
on existing DOE.
Sample pool: The sample pool Sinner d∗ = {x(i) , i = 1, ⋯, Nd∗ } for d∗ is
generated according to fX|d∗ (x). Nd∗ is also a critical point affecting the
efficiency of AK-MCS. For a big failure probability, if a big Nd∗ is
selected, the efficiency of AK-MCS will reduce since too many samples
are generated around the limit-state surface. For a small failure proba­
bility, if a small Nd∗ is selected, the accuracy of AK-MCS will reduce.
Usually, the coefficient of variation (Cov) of failure probability is used to
check if the sample pool is large enough to obtain an accurate result. It is
estimated by Eq.(23).
√̅̅̅̅̅̅̅̅̅̅̅̅̅
[ ]̅
[ ] ̂f
V P √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1− P ̂f
̂f =
Cov P [ ] ≈ (23)
̂f ̂f
(N − 1) P
E P

A Cov[ P
̂ f ] below 5% is seen as acceptable. Benefit from Kriging model
constructed by existing DOE, a rough estimation P ̂ (0) (d∗ ) of failure
f
probability can be estimated by MCS using the current Kriging model.
The initial size Nd∗ of the sample pool can be set as Eq.(24) according to
Eq.(23).

1− P ̂ (0) ∗
f (d )
Nd ∗ = (0)
+1 (24)
̂ f (d∗ )
0.052 P

Adaptive learning strategy: U learning function is used to select the

best next sample.
Stopping criterion: The ESC stopping criterion introduced by Wang
et al. [44] is considered. The maximum error of the Kriging-based failure
̂ MCS with respect to MCS estimation PMCS can be
probability estimation P f f
calculated as:
⃒ ⃒ ⎛⃒ ⃒ ⃒ ⃒⎞
⃒ ⃒ ⃒ ⃒ ⃒ ⃒
⃒N̂f ⃒ ⃒ N ̂ ⃒ ⃒ N ̂ ⃒
⃒ ⃒ ⃒ ⃒⎟
ε = ⃒⃒ − 1⃒⃒ ≤ max⎜
⎝⃒
f
u − 1⃒, ⃒
f
ε max
u − 1⃒⎠ = ̂ (25)
N ⃒ ̂ ̂ ⃒ ⃒ ̂ ̂ ⃒
⃒ f ⃒ ⃒N f − S f ⃒ ⃒N f + S s ⃒

The adaptive learning process stops when the following condition is

reached:
[ ]
̂ f ≤ 5%
ε max ≤ εthe andCov P
̂ (26)
Fig. 1. The flowchart of the proposed decoupled approach.
where εthe is a prescribed error threshold. A detailed description of ESC
function reads: can be found in Ref. [44] . A brief introduction of ESC is presented in
⃒
⃒
⃒
⃒ Appendix A.
⃒̂
U(x) = ⃒⃒
g K (x) ⃒
⃒ (21) Then P̂ f (d∗ ) is estimated by MCS using the final Kriging model. The
⃒
⃒σ̂g K (x)⃒ β(d) = − Φ− 1 ( P
β(d∗ ) is calculated by ̂
corresponding ̂ ̂ f (d)).

where ̂g K (x) and σ̂g (x) are the mean and standard deviation of the 4.3. Reliability sensitivity
K

performance function predicted by the Kriging model, respectively. The

best next sample to add to the DOE is defined as: The reliability sensitivity ∂Pf (d)/∂d is the partial derivative of the
failure probability Pf for the distribution parameter d.

5
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

(∫ )
∂Pf (d) ∂ The proposed UC learning function is inspired by the U learning
= IF (x)fX|d (x)dx
∂d ∂d Rn function and the cost function is also considered simultaneously.
∫
∂fX|d (x) β(d) − βT ) is judged correctly
Φ(Uβ (d)) is the probability that the sign of (̂
= IF (x)
∂d
dx by the constructed GEK model βGEK (d)as shown in Fig. 2. The sample
Rn
∫ (27) which has a low probability Φ(Uβ (d)) of correct judgment and low-cost
∂fX|d (x) 1
= IF (x) f (x)dx function should be added to DOE {Souter
DOE , βDOE } to improve the accuracy
∂d fX|d (x) X|d
of βGEK (d) around the feasible region boundary with low-cost value.
Rn
[ ]
IF (x) ∂fX|d (x) Therefore, the best next sample is defined as:
=E ⋅
fX|d (x) ∂d
dnext = arg min
outer
(UC(d)) (32)
d∈S
According to the relationship between Pf and β (Eq.(4)). ∂β(d) /∂d
can be obtained according to chain derivation rule. where Souter = {d(i) , i = 1, ⋯, Nouter } is a large sample set generated in the
deterministic feasible region {d : gk (μX |d) ≥ 0}. For the applications in
∂β(d) ∂β(d) ∂Pf (d) − 1 ∂Pf (d)
= = (28) this paper, Nouter is set to 105.
∂d ∂Pf (d) ∂d ϕ(− β(d)) ∂d
̂
β(dnext ) and its sensitivity ∂ β (d∂dnext ) are estimated by sections 4.2 and
̂
{ }
where ϕ(⋅) is the PDF of the standard normal distribution. It can be seen
∂̂
from Eq.(27) and Eq.(28) that benefit from the definition of the failure 4.3. A new training sample dnext , β(dnext ), ∂d ̂ β (dnext )
is added to DOE
probability, they only require partial derivative of joint PDF fX|d (x) for d
which can be derived analytically. {Souter
DOE , βDOE } until the stopping criterion is achieved. In this work, the
following stopping criterion of the first stage of enrichment is used:
Therefore, the gradient information ∂∂dβ required in the construction of
the outer GEK model βGEK (d) can be easily obtained as a post-processing c(dnext ) ≥ cthr (33)
of reliability analysis.
where c(dnext ) is the cost function value of current dnext . cthr is a pre­
5. RIF approximation based on adaptive GEK model for RBDO scribed threshold which is different for the different problem. The
feasible region boundary with high-cost value is not the region of in­
Benefits from efficient reliability and reliability sensitivity analysis terest, the first stage of enrichment is stopped when c(dnext ) is larger than
technology, it is possible to efficiently construct a GEK model βGEK (d) of a prescribed threshold.
β(d). The adaptive learning strategy and stopping criterion to construct a
βGEK (d) are introduced in detail in this section. 5.3. Second stage of enrichment

The following deterministic optimization is implemented after the

5.1. Approximate RIF βGEK (d) based on GEK model
first stage of enrichment is completed.
Randomly select Nouter 1 samples as an initial DOE in the deterministic min : c(d)
feasible region {d : gk (μX |d) ≥ 0}. And their estimation of the β(d) and
d
⎧
gradient information ∂β(d)/∂d can be obtained by Section 4.2 and 4.3.
⎨β T
GEK,i (d) ≥ βi , i = 1, ⋯, m (34)
s.t. :
An initial GEK model βGEK (d) can be constructed according to the DOE ⎩ hj (d) ≤ 0, j = 1, ⋯, M
{Souter
DOE , βDOE }which can be expressed as:
[ And the optimal solution dopt
new is obtained by any suitable optimiza­
SDOE = d(1) , ⋯, d(N1 ) , ⋯,
outer outer

tion algorithm (Genetic algorithm is used in this paper). Then ̂β(dopt ) new
]T (29)
̂ opt
d(1) , ⋯d(1) , ⋯d(N1 ) , ⋯d(N1 ) and its sensitivity ∂ β(d∂dnew ) are estimated by sections 4.2 and 4.3. A new
outer outer

{ }
opt
[ ( ) ( outer ) training sample dopt ̂ opt ∂̂β(dnew ) is added to DOE {Souter , βDOE }to
new , β(dnew ), ∂d DOE
̂ β d(N1 ) ,
β d(1) , ⋯, ̂
β DOE = ̂
improve the local accuracy of βGEK (d) around the optimal solution. The
( outer ) ( outer )]T (30)
(
∂̂β d(1)
)
∂̂
(
β d(1)
)
β d(N1 )
∂̂ ∂̂β d(N1 ) second stage of enrichment is stopped until the difference of the cost
, ⋯, , ⋯, , ⋯, function between two consecutive iterations are small enough. The
∂d1 ∂dnd ∂d1 ∂dnd
stopping criterion is shown in Eq.(35).
The βGEK (d) also contains two-stage of enrichment. The first stage of ⃒ (
⃒ ) ( opt )⃒
⃒
enrichment aims at adding more training samples around the feasible η = ⃒c dopt ̃
new − c dnew ⃒ ≤ ηthr (35)
region boundary with low-cost value. The second stage of enrichment
focus on the region around the optimal solution obtained by Block 3. where c(dopt ̃opt
new ) and c(dnew ) are the cost value of the current and last
iteration respectively.
5.2. First stage of enrichment
6. Examples analysis
The learning function UC of the first stage is defined as:
( ) In this section, four mathematical and practical engineering exam­
UC(d) = Φ Uβ (d) × c(d) ples for RBDO are presented to illustrate the accuracy and the efficiency
⃒ ⃒
⃒ ⃒ of the proposed decoupled method (Approximate reliability index
⃒̂ T⃒ (31)
⃒ β GEK (d) − β ⃒
Uβ (d) = ⃒ ⃒ function based on adaptive double-loop Kriging, RIFA-ADK).
⃒ σ̂ (d) ⃒
⃒ β GEK ⃒ βiGEK denotes the outer GEK model of the RIF for i-th probability
constraint. gKi denotes the inner Kriging model of the performance

6
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

β(d) − βT ) judged correctly by βGEK (d).

Fig. 2. The probability of the sign of ( ̂

function for i-th probability constraint. Ncall denotes the number of the also fail to satisfy the reliability constrains since classical approaches
model evaluations. Reliability index βMCS
i at the optimal solution from (SORA and AMV) are used to solve RBDO after Kriging model is
different RBDO methods is assessed by MCS with 106 evaluations of the constructed.
true performance function. For comparison, the results of other methods The accuracy and characteristic of the proposed FIFA-ADK is
from existing literature are also provided in the results table. demonstrated by Fig. 3 ang Fig. 4. Fig. 3 shows the approximate feasible
region boundary (FRB) and its DOE from final outer-loop GEK βiGEK in
Example 1. The accuracy of outer-loop GEK βiGEK directly affects the final
6.1. Example 1: Standard test problem for RBDO
solution. The FRB computed by MCS is drawn with the solid line in Fig. 3
This standard text problem is studied in several existing literature, (a). The approximate FBR from the final outer-loop GEK βiGEK is drawn
which has three probabilistic constraints and two random variables. The with the dotted line. It can be seen from Fig. 3(a) that three accurate
RBDO problem reads: FRBs in the important region can be obtained by approximate RIF βiGEK .
Fig. 3(b) mainly shows the DOE of outer-loop GEK βiGEK , where the
min : c(d) = 10 − d1 + d2
d enrichment DOE is drawn with solid square. Fig. 3(b) demonstrates that
( ) the enrichment DOE is located in the feasible region boundary around
s.t. : Pr[gi (X(d)) ≤ 0] ≤ Φ − βTi i = 1, 2, 3
the optimal solution. It benefits from the proposed UC learning function
X12 X2 and the two-stage learning strategy. Fig. 4 shows the approximate limit-
g1 (X) = − 1
20 state surface (LSS) and its DOE from the final inner-loop Kriging gKi in
Example 1. The LSS computed by the original performance function is
(X1 + X2 − 5)2 (X1 − X2 − 12)2 (36)
g2 (X) = + − 1 drawn with the solid line in Fig. 4(a). The approximate LSS from the
30 120
approximate performance function gKi is drawn with the dotted line. It
g3 (X) = (
80 can be seen from Fig. 4(a) that an accurate limit-state surface can be
)2 − 1
X12 + 8X2 + 5 obtained by an approximate performance function gKi . Fig. 4(b) mainly
( ) shows the DOE of approximate performance function gKi , where the
0 ≤ dj ≤ 10, Xj ∼ N dj , 0.32 , j = 1, 2
enrichment DOE is drawn with solid square. Enrichment DOE selected
βT1 = βT2 = βT3 = 3 by U learning function is located around the limit-state surface. Overall,
it can be seen from Fig. 3 and Fig. 4 that the proposed double-loop
The results are summarized in Table 1. It can be seen that the pro­ Kriging can obtain the accurate FRB and LSS and the proposed strat­
posed decoupled approach RIFA-ADK has an accurate solution. Ncall of egy for enrichment DOE can add more training samples around the
the proposed decoupled approach is less than the Ncall of classical ap­ importance region of interest. Therefore, the proposed decoupled
proaches (SORA and PMA) and other Kriging-based methods (IBS and approach RIFA-ADK can obtain the satisfactory approximate model
CBS). The SORA and PMA show the same performance for this example. required in RBDO at a reasonable calculation cost.
However, βMCS3 at optimal design point is slightly less than βT3 and they
fail to satisfy the reliability constrains since approximation reliability 6.2. Example 2: Highly nonlinear constraint problem
analysis method is adopted. The IBS and CBS require less Ncall than SORA
and PMA since the RBDO is solved based on cheap Kriging model. They Example 2 has two probability constraints and two random design
variables with normal distributions. The main feature of this example is
Table 1 the strong nonlinearity of probability constrain 1. The RBDO problem
The results of RBDO for example 1. reads:
Methods Optimal design Ncall Obj. βMCS
1 βMCS
2 βMCS
3
point min : c(d) = (d1 − 3.7)2 + (d2 − 4)2
d
( )
Analytical (5.8633, 3.4159) // 7.5526 Inf. 3.0007 3.0004 s.t. : Pr[gi (X(d)) ≤ 0] ≤ Φ − βTj i = 1, 2
SORA (5.8542, 3.4238) 332 7.5695 Inf. 3.0183 2.9423
PMA (5.8542, 3.4238) 396 7.5695 Inf. 3.0183 2.9423 g1 (X) = − X1 sin(4X1 ) − 1.1X2 sin(2X2 )
CBS (5.8546, 3.4236) 53/53/ 7.5690 Inf. 3.0235 2.9952 (37)
g2 (X) = X1 + X2 − 3
53
IBS (5.8577, 3.4169) 48/48/ 7.5635 Inf. 3.0095 2.9971 0 ≤ d1 ≤ 3.7, 0 ≤ d2 ≤ 4
( )
48 Xj ∼ N dj , 0.12 , j = 1, 2
RIFA- (5.8575 3.4134) 15/12/ 7.5559 Inf. 3.0002 3.0064
ADK 21
βT1 = βT2 = 2

7
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

Fig. 3. The approximate feasible region boundary and its DOE from final outer-loop GEK βiGEK in Example 1.

Fig. 4. The approximate limit-state surface and its DOE from final inner-loop Kriging gKi in Example 1.

can be seen from Fig. 6(a) that an accurate limit-state surface can be
Table 2
obtained by an approximate performance function gKi . Enrichment DOE
The results of RBDO for example 2.
selected by U learning function is located around the limit-state surface.
Methods Optimal design point Ncall Obj. βMCS
1 βMCS
2 The results of Example 2 demonstrate that the proposed decoupled
Analytical (2.8421, 3.2320) // 1.3258 2.0031 Inf. approach RIFA-ADK is suitable for highly nonlinear constraint RBDO
SORA (2.8200, 3.2800) 138 1.3000 2.0096 Inf. problems.
PMA (2.8400, 3.2300) 296 1.3300 2.0088 Inf.
CBS (2.8153, 3.2782) 55/55 1.3038 1.8514 Inf. 6.3. Example 3: Roof truss RBDO problem
IBS (2.8162, 3.2769) 49/49 1.3032 1.8566 Inf.
RIFA-ADK (2.8414, 3.2312) 50/12 1.3282 2.0098 Inf.
Fig. 7 shows a roof truss. The top boom and compression members
are concrete, and the bottom boom is steel. The parameters of random
variables are given in Table 3. The roof is subjected to the vertical load of
The results are summarized in Table 2. As shown in Table 2, the q. The vertical deflection at the peak of the structure is limited to 3 cm,
proposed decoupled approach RIFA-ADK led to an accurate solution. and the objective is to minimize the material cost. The RBDO problem of
SORA and PMA require about 100-300 model evaluations. Kriging- roof truss is given as follows:
based methods (CBS and IBS) and the proposed method only require ( )
about 50 model evaluations for this highly nonlinear constraint RBDO min : c μAS , μAC = 20224μAS +364μAC
d
problem. [ ( 2 )( ) ]
( )
Fig. 5 shows the approximate FRB and its DOE from the final outer- s.t. : Pr g= 0.03−
ql 3.81
+
1.13
≤ 0 ≤ Φ − βT (38)
2 AC EC AS ES
loop GEK βiGEK in Example 2. As shown in Fig. 5(a), an accurate FRB in
the important region can be obtained by approximate RIF βiGEK . Benefit 6 × 10− 4
≤ μAS ≤ 1.2 × 10− 3 , 0.018 ≤ μAC ≤ 0.063βT1 = 3
from the proposed UC learning function and the two-stage learning
strategy, the enrichment DOE is located on the FRB around the optimal This RBDO problem involves four random variables and two design
solution as shown in Fig. 5(b). Fig. 6 shows the approximate limit-state variables. The two design variables follow two different distributions:
Normal and Lognormal. The results of the proposed decoupled approach
surface and its DOE from the final inner-loop Kriging gKi in Example 2. It

8
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

Fig. 5. The approximate feasible region boundary and its DOE from final outer-loop GEK βiGEK in Example 2.

Fig. 6. The approximate limit-state surface and its DOE from final inner-loop Kriging gKi in Example 2.

Table 3
Random variables for Example 3.
Variable Distribution Mean Standard
deviation

Uniform load q(N /m) Normal 20,000 1400

Length l(m) Normal 12 0.12
Sectional area Normal Design 5.892 × 10− 5

AS (m2 ) /Lognormal variable

Sectional area Normal Design 4.8 × 10− 3

AC (m2 ) /Lognormal variable

Elastic modulus Normal 1 × 1011 6 × 109
ES (Pa)
Elastic modulus Normal 2 × 1010 1.2 × 109
EC (Pa)

and other methods are given in Table 4. As shown in Table 4, the pro­
posed method requires fewer model evaluations than other methods and
can obtain a satisfactory solution in condition of normal and lognormal
design variables. The SORA and PMA in condition of normal design
variables require more model evaluations than lognormal design
Fig. 7. Roof truss [45].

9
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

Table 4
The results of RBDO for example 3.
Distribution Methods Optimal design point(AS , Ncall Obj. βMCS
AC )

Normal SORA (1.0887 × 10− 3 , 3.8355 × 1103 35.97 2.90

10− 2 )
PMA (1.0886 × 10− 3 , 3.8358 × 446 35.97 2.90
10− 2 )
CBS (1.0876 × 10− 3 , 3.8342 × 202 35.95 2.89
10− 2 )
IBS (1.0882 × 10− 3 , 3.8360 × 188 35.97 2.91
10− 2 )
RIFA- (1.1045 × 10− 3 , 3.8341 × 68 36.29 3.00
ADK 10− 2 )
Lognormal SORA (1.0991 × 10− 3 , 3.7362 × 468 35.82 2.94
10− 2 )
PMA (1.1003 × 10− 3 , 3.7293 × 296 35.82 2.95
10− 2 )
Fig. 9. A Twenty-five bar spatial truss [46].
CBS (1.1013 × 10− 3 , 3.7303 × 198 35.85 2.96
10− 2 )
IBS (1.1001 × 10− 3 , 3.7288 × 190 35.82 2.94
Table 5
10− 2 )
Element information for the twenty-five bar spatial truss.
RIFA- (1.0878 × 10− 3 , 3.8280 × 70 35.93 3.00
ADK 10− 2 ) Group 1 2 3 4 5 6

Member 1: 2: 6: 10: 14:(3,10) 18: 22:

(1,2) (1,4) (2,4) (6,3) (4,7) (10,6)
3: 7: 11: 15:(6,7) 19: 23:(3,7)
(2,3) (2,5) (5,4) (3,8) 24:(4,8)
4: 8: 12: 16:(4,9) 20: 25:(5,9)
(1,5) (1,3) (3,4) (5,10)
5: 9: 13: 17:(5,8) 21:
(2,6) (1,6) (6,5) (6,9)

Table 6
Loading conditions for the twenty-five bar spatial truss.
Node Case 1 Case 2
PX PY PZ PX PY PZ
kips(kN) kips(kN) kips(kN) kips(kN) kips(kN) kips(kN)
1 0.0 20.0 -5.0 1.0 10.0 -5.0
(89) (22.25) (4.45) (44.5) (22.25)
2 0.0 -20.0 -5.0 0.0 10.0 -5.0
(89) (22.25) (44.5) (22.25)
3 0.0 0.0 0.0 0.5 0.0 0.0
(2.22)
6 0.0 0.0 0.0 0.5 0.0 0.0
(2.22)

6.4. Example 4: Twenty-five bar spatial truss

Fig. 8. The approximate feasible region boundary (FRB) and its DOE from final
outer-loop GEK βiGEK in Example 3 Fig. 9 shows a Twenty-five bar spatial truss structure [46].
Twenty-five bar spatial truss problem is a common deterministic opti­
mization problem. This problem is extended to the RBDO problem by
variables. The proposed decoupled approach requires a similar level of
considering uncertainties in the loading conditions, material properties
model evaluations for two types of design variables. The object function
and section area.
value of SORA, PMA, CBS and IBS methods is smaller than the proposed
The structural members are arranged into eight groups, where all
method but they fail to satisfy the reliability constraint. Fig. 8 shows the
members in a group share the same material and cross-sectional prop­
approximate feasible region boundary and its DOE from the final outer-
erties. Table 5 defines each element group by member number. Table 6
loop GEK βiGEK in condition of normal design variables. As shown in
shows the two loading conditions. The material density is 0.1 lb/in3
Fig. 8, the feasible region boundary for this RBDO problem is not
(2767.990 kg/m3) and the modulus of elasticity is 10,000 ksi (68,950
complicated. Six enrichment DOE selected by the proposed two-stage
MPa). The parameters of random variables are given in Table 7. The two
learning strategy is located around the feasible region boundary. The
limit states are defined as the maximum displacement D1 and D2 should
approximate feasible region boundary from the final outer-loop GEK
be less than 4.5 in on the condition of loading case 1 and case 2,
βiGEK is very close to the true one.

10
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

Table 7 Table 8
Random variables for Example 4. The results of RBDO for example 3.
Variable Description Distribution Mean Standard Methods Optimal design point Ncall Obj. βAK−
1
MCS
βAK−
2
MCS

deviation (A1 ∼ A6 )

P1Y Loading Case 1: Y- Normal 20 1 SORA (0.50, 2.84, 3.37, 0.50, 2716 680.91 1.90 1.96
direction Load in node 1 1.52, 2.78)
and 2 PMA (0.50, 2.84, 3.37, 0.50, 1568 680.91 1.90 1.96
P1Z Loading Case 1: Z- Normal 5 0.25 1.52, 2.78)
direction Load in node 1 CBS (0.50, 2.84, 3.37, 0.50, 403 680.91 1.90 1.96
and 2 1.52, 2.78)
P2Y Loading Case 2: Y- Normal 10 0.5 IBS (0.50, 2.84, 3.37, 0.50, 387 680.91 1.90 1.96
direction Load in node 1 1.52, 2.78)
and 2 RIFA- (0.50, 2.83, 3.38, 0.50, 299 693.20 2.38 2.01
P2Z Loading Case 2: Z- Normal 5 0.25 ADK 1.64, 2.71)
direction Load in node 1
and 2
E Modulus of elasticity Normal 104 500
The core of the proposed RIFA-ADK decoupled approach is approxi­
A1 Cross-sectional area for Normal Design 0.1 mating RIF based on adaptive double-loop Kriging. In RIFA-ADK, inner
group 1 variable
A2 Cross-sectional area for Normal Design 0.1
loop Kriging is used to perform reliability analysis, and GEK model,
group 2 variable which takes into account gradient information, is used to approximate
A3 Cross-sectional area for Normal Design 0.1 RIF. Adaptive learning strategy which involves two stages of enrichment
group 3 variable is also developed to improve the surrogate precision on the region of
A4 Cross-sectional area for Normal Design 0.1
interest. The basic idea and example results show that the proposed UC
group 4 variable
A5 Cross-sectional area for Normal Design 0.1 learning function and two stages of enrichment strategy help improve
group 5 variable the surrogate precision on the important active feasible region of RBDO.
A6 Cross-sectional area for Normal Design 0.1 Four examples demonstrate that the RBDO problem can be solved by the
group 6 variable proposed RIFA-ADK decoupled approach with a reasonable computa­
tional cost.
respectively. The RBDO problem aims at minimizing the weight of the The limitation of the proposed method includes two aspects. One is
twenty-five bar spatial truss with design variables being the cross- that the proposed method is unsuitable for the problem with very small
sectional areas A = [A1 , A2 , A3 , A4 , A5 , A6 ]. The RBDO problem of variation coefficient of the random design variable, since reliability
twenty-five bar is given as follows: sensitivity is difficult to be obtained accurately in this case, which results
( ) the inaccuracy of the GEK and RBDO. The other is that the proposed
min : c μA 1 , μA2 , μA3 , μA4 , μA5 , μA6 method is unsuitable for the problem with high-dimensional input, due
to a large size of DOE including gradient information is required in this
d
[ ( ) ( ) ] ( T)
s.t. : Pr D1 A, P(1) , E = D1 A, P(1) (1)
Y , PZ , E ≥ 4.5in ≤ Φ − β (39) case, which makes training GEK unaffordable. In future, the more robust
[ ( ) ( ) ] ( )
Pr D2 A, P(2) , E = D2 A, P(2) (2)
Y , PZ , E ≥ 4.5in ≤ Φ − β
T
and accurate reliability sensitivity method and dimensionality reduction
0.5in2 ≤ μA i ≤ 4.5in2 , βT = 2 strategy should be developed to enlarge the applicability of the proposed
method. In addition, the combination of different types of gradient-
An initial DOE of 50 samples is selected to construct the initial inner based surrogate models and reliability analysis methods will be
Kriging gKD1 (A,P(1) ,E)andgKD2 (A,P(2) ,E), respectively. An initial DOE of 37 explored to verify their rationality and efficiency for different types of
samples is selected to construct the initial outer GEK βDGEK 1
andβDGEK
2
, RBDO problems.
respectively. The optimal solution by the proposed RIFA-ADK is Aopt =
[0.50, 2.83, 3.38, 0.50, 1.64, 2.71] with 156 model evaluations for prob­ CRediT authorship contribution statement
ability constraint 1 and 143 model evaluations for probability constraint
2. The optimal weight is 693.20 lb. Group 3 is assigned the largest cross- Xiaobo Zhang: Methodology, Software, Validation, Data curation,
sectional area with 3.38 in2. Group 1 and 3 have litter influence on the Writing – original draft, Project administration. Zhenzhou Lu:
maximum displacement and are hence set close to the lower bound of Conceptualization, Resources, Writing – review & editing, Supervision,
the value range. Funding acquisition. Kai Cheng: Formal analysis, Investigation,
The results of the proposed decoupled approach and other methods Visualization.
(SORA and PMA) are summarized in Table 8. The reliability index β at
optimal design point is estimated by AK-MCS in Table 8 since MCS re­
quires a large number of finite element runs and it is very time- Declaration of Competing Interest
consuming. The SORA and PMA requires 2716 and 1568 finite
element runs respectively. The CBS and IBS requires 403 and 387 finite The authors declare that they have no known competing financial
element runs respectively. And they fail to satisfy the reliability con­ interests or personal relationships that could have appeared to influence
strains. Overall, only 299 finite element runs are required to solve this the work reported in this paper.
RBDO problem with six design variable and multiple reliability con­
strains, which demonstrates that the proposed RIFA-ADK decoupled Acknowledgments
approach for RBDO can be applied to practical engineering problems
with an acceptable computational cost. This work was supported by the National Natural Science Foundation
of China (Grant no. NSFC 11902254) , National Science and Technology
7. Conclusion Major Project (2017-IV-0009-0046), and Innovation Foundation for
Doctor Dissertation of Northwestern Polytechnical University
This paper develops a decoupled approach to solve RBDO efficiently. (CX202018).

11
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

Appendix A. ESC stopping criterion

The ESC stopping criterion introduced by Wang et al. [44] is considered. The maximum error of the Kriging-based failure probability estimation
̂ MCS with respect to MCS estimation PMCS can be calculated as:
P f f
⃒ ⃒ ⎛⃒ ⃒ ⃒ ⃒⎞
⃒ ⃒ ⃒ ⃒ ⃒ ⃒
⃒N̂f ⃒ ⃒ N ̂ ⃒ ⃒ N ̂ ⃒
⎜⃒ ⃒ ⃒ ⃒⎟
(40)
⃒ ⃒ f f
ε = ⃒ − 1⃒ ≤ max⎝⃒ u − 1⃒, ⃒ ε max
u − 1⃒⎠ = ̂
⃒ Nf ⃒N̂ ̂ ⃒ ⃒N ̂ ̂ ⃒
⃒ ⃒ f − Sf ⃒ ⃒ f + Ss ⃒

where N
̂ f and N
̂ S denote the number of failure samples and safe samples from sample pool determined by the Kriging model gK (x), respectively. ̂ S f is
the number of failure samples judged by Kriging model that actually belongs to safe domain. ̂
S S is the number of safe samples judged by Kriging model
u u
that actually belongs to failure domain. ̂S f and ̂
S S are the upper bound of the confidence interval of ̂
S S and ̂
S f , respectively.
S f and S S follow Poisson distributions:
̂ ̂

( ) ̂
Nf
∑ ̂
Nf
∑ ( )
̂
S f ∼ PB μ̂ , σ̂2 , μ̂ = Pwse
i , σ̂2 = Pwse
i 1 − Pwse
i (41)
Sf Sf Sf Sf
i=1 i=1

( ) ̂
∑
NS ̂
∑
NS
( )
̂
S S ∼ PB μ̂ , σ̂2 , μ̂ = Pwse
i , σ̂2 = Pwse
i 1 − Pwse
i (42)
SS SS SS SS
i=1 i=1

⎛ ⎞
̂g
where Pwse = Φ⎝ − | σ (xi ) |⎠ is the probability of wrong sign estimation for xi . The confidence interval of ̂
S S and ̂
S f can be obtained by:
K (xi )
i
̂g K
[ ]
̂
S S ∈ μ̂ − γci σ̂ , μ̂ + γ ci σ̂ (43)
Sf Sf Sf Sf

[ ( ) ( ]
α α)
̂
S f ∈ Γ̂− 1 , Γ̂− 1 1 − (44)
Sf 2 Sf 2

Where γ ci = 1.96 for the confidence level α = 0.05. Γ− 1 (.) is the inverse CDF of the Poisson distribution with mean and variance equal to μ̂ .
̂S f Sf

References [15] Cheng G, Xu L, Jiang L. A sequential approximate programming strategy for

reliability-based structural optimization. Comput Struct 2006;84(21):1353–67.
[16] Torii AJ, Lopez RH, Miguel LFF. A second order SAP algorithm for risk and
[1] Aoues Y, Chateauneuf A. Benchmark study of numerical methods for reliability-
reliability based design optimization. Reliabil Eng Syst Saf 2019:190.
based design optimization. Struct Multidiscip Optim 2010;41(2):277–94.
[17] Laurent L, Le Riche R, Soulier B, Boucard P. An overview of gradient-enhanced
[2] Cheng K, Lu Z, Xiao S, Zhang X, Oladyshkin S, Nowak W. Resampling method for
metamodels with applications. Arch Comput Meth Eng 2019;26(1):61–106.
reliability-based design optimization based on thermodynamic integration and
[18] Kim J, Song J. Quantile surrogates and sensitivity by adaptive Gaussian process for
parallel tempering. Mech Syst Sig Process 2021:156.
efficient reliability-based design optimization. Mech Syst Signal Process 2021:161.
[3] Kroetz HM, Moustapha M, Beck AT, Sudret B. A two-level Kriging-based approach
[19] Echard B, Gayton N, Lemaire M. AK-MCS: An active learning reliability method
with active learning for solving time-variant risk optimization problems. Reliabil
combining Kriging and Monte Carlo Simulation. Struct Saf 2011;33(2):145–54.
Eng Syst Saf 2020:203.
[20] Cui D, Wang G, Lu Y, Sun K. Reliability design and optimization of the planetary
[4] Leimeister M, Kolios A. Reliability-based design optimization of a spar-type
gear by a GA based on the DEM and Kriging model. Reliabil Eng Syst Saf 2020:203.
floating offshore wind turbine support structure. Reliabil Eng Syst Saf 2021:213.
[21] Jiang C, Yan Y, Wang D, Qiu H, Gao L. Global and local Kriging limit state
[5] Moustapha M, Sudret B. Surrogate-assisted reliability-based design optimization: a
approximation for time-dependent reliability-based design optimization through
survey and a unified modular framework. Struct Multidiscip Optim 2019;60(5):
wrong-classification probability. Reliabil Eng Syst Saf 2021:208.
2157–76.
[22] Li M, Wang Z. Surrogate model uncertainty quantification for reliability-based
[6] Nikolaidis E, Burdisso R. Reliability based optimization: a safety index approach.
design optimization. Reliabil Eng Syst Saf 2019:192.
Comput Struct 1988;28(6):781–8.
[23] Yang M, Zhang D, Jiang C, Han X, Li Q. A hybrid adaptive Kriging-based single
[7] Lee JO, Yang YS, Ruy WS. A comparative study on reliability-index and target-
loop approach for complex reliability-based design optimization problems. Reliabil
performance-based probabilistic structural design optimization. Comput Struct
Eng Syst Saf 2021:215.
2002;80:257–69.
[24] Agarwal H, Renaud JE. Reliability based design optimization using response
[8] Tu J, Choi KK, Park YH. A new study on reliability-based design optimization.
surfaces in application to multidisciplinary systems. Eng Optim 2004;36(3):
J Mech Des 1999;121(4):557–64.
291–311.
[9] Youn BD, Choi KK, Du L. Enriched performance measure approach for reliability-
[25] Papadrakakis M, Lagaros ND, Plevris V. Design optimization of steel structures
based design optimization. AIAA J 2005;43(4):874–84.
considering uncertainties. Eng Struct 2005;27(9):1408–18.
[10] Kuschel N, Rackwitz R. Two basic problems in reliability-based structural
[26] Lee TH, Jung JJ. A sampling technique enhancing accuracy and efficiency of
optimization. Math Methods Oper Res 1997;46(3):309–33.
metamodel-based RBDO: Constraint boundary sampling. Comput Struct 2008;86
[11] Madsen HO, Friis Hansen P. A comparison of some algorithms for reliability-based
(13–14):1463–76.
structural optimization and sensitivity analysis. In: Rackwitz R, Thoft-
[27] Chen Z, Peng S, Li X, Qiu H, Xiong H, Gao L, et al. An important boundary sampling
Christensen P, editors. Proceedings of the 4th IFIP WG 7.5 conference. Springer-
method for reliability-based design optimization using kriging model. Struct
Verlag; 1992. p. 443–51. Berlin.
Multidiscip Optim 2015;52(1):55–70.
[12] Chen X, Hasselman T, Neill D. Reliability based structural design optimization for
[28] Wu Z, Chen Z, Chen G, Li X, et al. A probability feasible region enhanced important
practical applications. In: 38th Structures, structural dynamics, and materials
boundary sampling method for reliability-based design optimization. Struct
conference; 1997.
Multidiscip Optim 2020.
[13] Liang J, Mourelatos ZP, Nikolaidis E. A single-loop approach for system reliability-
[29] Chen Z, Qiu H, Liang G, Li X, Li P. A local adaptive sampling method for reliability-
based design optimization. J Mech Des 2007;129(12):1215–24.
based design optimization using Kriging model. Struct Multidiscip Optim 2014;49
[14] Du XP, Chen W. Sequential optimization and reliability assessment method for
(3):401–16.
efficient probabilistic design. J Mech Des 2004;126(2):225–33.

12
X. Zhang et al. Reliability Engineering and System Safety 216 (2021) 108020

[30] Li X, Qiu H, Chen Z, Gao L, Shao X. A local Kriging approximation method using [39] Laurent L. GRENAT (Matlab/Octave Toolbox). 2016. https://bitbucket.org/
MPP for reliability-based design optimization. Comput Struct 2016;162:102–15. luclaurent/grenat.
[31] Liu X, Wu Y, Wang B, Ding J, Jie H. An adaptive local range sampling method for [40] Echard B, Gayton N, Lemaire M, Relun N. A combined Importance Sampling and
reliability-based design optimization using support vector machine and Kriging Kriging reliability method for small failure probabilities with time-demanding
model. Struct Multidiscip Optim 2017;55(6):2285–304. numerical models. Reliab Eng Syst Saf 2013;111:232–40.
[32] Au SK. Reliability-based design sensitivity by efficient simulation. Comput Struct [41] Zhang X, Lu Z, Cheng K. AK-DS: An adaptive Kriging-based directional sampling
2005;83(14):1048–61. method for reliability analysis. Mech Syst Sig Process 2021;156:107610.
[33] Zhang J, Taflanidis AA, Medina JC. Sequential approximate optimization for [42] Bichon BJ, Eldred MS, Swiler LP, Mahadevan S, McFarland JM. Efficient Global
design under uncertainty problems utilizing Kriging metamodeling in augmented Reliability Analysis for Nonlinear Implicit Performance Functions. AIAA J 2008;46
input space. Comput Meth Appl Mech Eng 2017;315:369–95. (10):2459–68.
[34] Dubourg V, Sudret B, Bourinet J. Reliability-based design optimization using [43] Lv Z, Lu Z, Wang P. A new learning function for Kriging and its applications to
kriging surrogates and subset simulation. Struct Multidiscip Optim 2011;44(5): solve reliability problems in engineering. Comput Math Appl 2015;70(5):1182–97.
673–90. [44] Wang Z, Shafieezadeh A. ESC: an efficient error-based stopping criterion for
[35] Moustapha M, Sudret B, Bourinet J, Guillaume B. Quantile-based optimization kriging-based reliability analysis methods. Struct Multidiscip Optim 2019;59(5):
under uncertainties using adaptive Kriging surrogate models. Struct Multidiscip 1621–37.
Optim 2016;54(6SI):1403–21. [45] Rashki M, Miri M, Moghaddam MA. A simulation-based method for reliability
[36] Foschi RO, Li H, Zhang J. Reliability and performance-based design: a based design optimization problems with highly nonlinear constraints. Autom
computational approach and applications. Struct Saf 2002;24(2–4):205–18. Constr 2014;47:24–36.
[37] Lehky D, Slowik O, Novak D. Reliability-based design: Artificial neural networks [46] Kaveh A, Talatahari S. Particle swarm optimizer, ant colony strategy and harmony
and double-loop reliability-based optimization approaches. Adv Eng Soft 2017; search scheme hybridized for optimization of truss structures. Comput Struct 2009;
117:1–13. 87(5–6):267–83.
[38] Morris MD, Ylvisaker TJMA. Bayesian design and analysis of computer
experiments: use of derivatives in surface prediction. Technometrics 1993;35(3):
243–55.

13