A RBFNN based active learning surrogate model for evaluating low failure probability in reliability analysis

This paper presents a novel active learning surrogate model for estimating low failure probability in the reliability analysis of complex structures based on a radial basis function neural network (RBFNN). The RBFNN surrogate model, which possesses global approximation capability, is constructed by randomly selecting an initial design of experiments (DoEs) from a Monte Carlo Simulation (MCS) population using the minimax distance method. Furthermore, the RBFNN surrogate model is updated by sequentially adding highly representative samples to improve the local prediction accuracy. The highly representative samples are selected using minimax distance method from each level's failure region of subset simulation (SS), and are further modified by the calculation of actual limit state function (LSF). Consequently, the low failure probability is obtained through an iterative framework that combines SS and RBFNN surrogate model. The proposed approach significantly reduces the number of experiments required, resulting in lower costs and higher efficiency. Three examples are provided to demonstrate the effectiveness and accuracy of the proposed method.