TY - JOUR
T1 - An improved AK-MCS for reliability analysis by an efficient and simple reduction strategy of candidate sample pool
AU - Liu, Ziyi
AU - Lu, Zhenzhou
AU - Ling, Chunyan
AU - Feng, Kaixuan
AU - Hu, Yingshi
N1 - Publisher Copyright:
© 2021 Institution of Structural Engineers
PY - 2022/1
Y1 - 2022/1
N2 - Adaptive Kriging combined with Monte Carlo simulation (AK-MCS) is popularly used to estimate the failure probability, but its candidate sample pool (CSP) size should be large enough to guarantee the accuracy of failure probability estimation, especially for small failure probability. To select training point for updating the Kriging model, AK-MCS needs to traverse the CSP to calculate the U-learning function of every point, which is time-demanding. Thus, an enhanced AK-MCS (shorten as EAK-MCS Ⅰ) is proposed by use of an efficient CSP reduction strategy. In EAK-MCS Ⅰ, the candidate samples are classified into two types, i.e., the samples with accurately judged states and the other samples. Obviously, removing the first type samples from CSP does not need extra computational cost, which has no effect on the precision of estimating failure probability and can improve the efficiency of training Kriging model. Hence, EAK-MCS Ⅰ is more efficient than AK-MCS in estimating failure probability. Additionally, for the risk of misjudging the states of the samples by the Kriging model trained in the reduced CSP for complicated problem, another version of EAK-MCS Ⅰ (shorten as EAK-MCS Ⅱ) is proposed by nesting an adaptive strategy to calibrate the Kriging model. The proposed methods improve the efficiency of AK-MCS and inherit all advantages of AK-MCS.
AB - Adaptive Kriging combined with Monte Carlo simulation (AK-MCS) is popularly used to estimate the failure probability, but its candidate sample pool (CSP) size should be large enough to guarantee the accuracy of failure probability estimation, especially for small failure probability. To select training point for updating the Kriging model, AK-MCS needs to traverse the CSP to calculate the U-learning function of every point, which is time-demanding. Thus, an enhanced AK-MCS (shorten as EAK-MCS Ⅰ) is proposed by use of an efficient CSP reduction strategy. In EAK-MCS Ⅰ, the candidate samples are classified into two types, i.e., the samples with accurately judged states and the other samples. Obviously, removing the first type samples from CSP does not need extra computational cost, which has no effect on the precision of estimating failure probability and can improve the efficiency of training Kriging model. Hence, EAK-MCS Ⅰ is more efficient than AK-MCS in estimating failure probability. Additionally, for the risk of misjudging the states of the samples by the Kriging model trained in the reduced CSP for complicated problem, another version of EAK-MCS Ⅰ (shorten as EAK-MCS Ⅱ) is proposed by nesting an adaptive strategy to calibrate the Kriging model. The proposed methods improve the efficiency of AK-MCS and inherit all advantages of AK-MCS.
KW - Candidate sample pool reduction
KW - Failure probability
KW - Kriging
KW - MCS
KW - U-learning function
UR - http://www.scopus.com/inward/record.url?scp=85119327579&partnerID=8YFLogxK
U2 - 10.1016/j.istruc.2021.10.090
DO - 10.1016/j.istruc.2021.10.090
M3 - 文章
AN - SCOPUS:85119327579
SN - 2352-0124
VL - 35
SP - 373
EP - 387
JO - Structures
JF - Structures
ER -