A structural reliability analysis method based on extreme learning machine

ZHENG Jianxiao; CUI Mengmeng; WANG Wenbo; ZHANG Mao

doi:10.16579/j.issn.1001.9669.2026.08.011

Abstract：[Objective]Excessive computational cost is encountered in traditional Monte Carlo simulation(MCS)method for structural reliability analysis.Poor high-dimensional fitting performance and large memory occupation under low failure probability conditions are presented in Kriging surrogate model.An active learning structural reliability analysis method combining adaptive extreme learning machine(AELM)surrogate model with MCS method was proposed.[Methods]Firstly,an AELM surrogate model was constructed,and the number of hidden layer nodes and activation function were optimized to improve the generalization ability of the model.Secondly,a Bayesian framework was introduced to model the uncertainty of hidden layer weights,and the posterior distribution was calculated to estimate the prediction variance.The problem that extreme learning machine cannot directly quantify prediction uncertainty was solved.Then,a active learning function was constructed to select the sample points with the maximum uncertainty for iterative model updating.Finally,the failure probability stability criterion was adopted as the convergence condition,and the structural failure probability was calculated in combination with the MCS method.[Results]The results show that compared with the Kriging surrogate model algorithms,the number of limit state function calls of the proposed method is reduced by up to 71.64％,and the calculation error is as low as 0.10％.The method has both higher computational efficiency and accuracy in engineering problems with multiple failure domains,strong nonlinearity and low failure probability,and provides a reference for engineering structural reliability analysis.