Adaptive subset searching-based deep neural network method for structural reliability analysis

Abstract

To deal with the problem of local optimal sampling in active learning methods for structural reliability analysis, an adaptive subset searching-based deep neural network (ASS-DNN) method is proposed. The proposed ASS-DNN method uses a global to local strategy to iteratively select important subsets close to the limit state surface from the Monte Carlo population to update a deep neural network (DNN) for calculating the structural failure probability. In each iteration, first a subset is selected from the previous subset of the Monte Carlo population to update the DNN of the limit state function. Second, the updated DNN is used to predict the failure probability and the conditional probability of the next subset in the current subset. Third, the convergence of the conditional probability is evaluated to evaluate the prediction accuracy of the DNN on the current subset. If the convergence condition is met, the sampling range of the next iteration is moved to a subset closer to the limit state surface, otherwise the sampling range remains in the current subset. Finally, the iteration is repeated until the required accuracy of the predicted failure probability is met. The efficiency and accuracy of the ASS-DNN method were demonstrated using three examples.