Machine learning-based enhanced Monte Carlo simulation for low failure probability structural reliability analysis

Abstract

Low failure probability problems with high computational costs are difficult to solve. Regarding this, a structural reliability analysis method (called AK-EMCS-SVR) combining active Kriging model, support vector regression and enhanced Monte Carlo simulation is proposed. To achieve global modeling, the uniform sampling strategy and expected feasibility function are also adopted. Besides, this paper proposes an adaptive training interval combining with the support vector regression algorithm to achieve more accurate and robust prediction. Five numerical cases and a finite element engineering case are used to illustrate the effectiveness of the proposed method. The results comparison showed that the AK-EMCS-SVR has an advantage in the number of calls to the limit state function and to the surrogate model. The method shows higher accuracy and robustness solving low failure probability problems with high computational cost.