The NARX neural network time series prediction technique and highly efficient reliability analysis method

Abstract

The reliability analysis theory represented by the stochastic simulation method requires repeated deterministic structural response analysis, which undoubtedly consumes a large amount of computational cost; meanwhile, the reliability obtained based on the stochastic simulation method has the characteristic of random convergence, so it is not worth advocating in practical engineering structures, which limits the application of the stochastic simulation method. In view of the above dilemma, this thesis proposes a strategy to train a nonlinear autoregressive with exogenous input model based on a neural network, and then measure the response characteristics of random samples in probability space, which greatly reduces the number of deterministic analyses and saves computational costs. Meanwhile, in order to circumvent the characteristics of random convergence of stochastic simulation methods, this study invokes the physical synthesis method based on probability density evolution theory to solve the system reliability. Finally, the accuracy and efficiency of the proposed method are systematically verified by two numerical academic examples. The analysis results show that the sample response characteristics predicted by the nonlinear autoregressive with exogenous input model are in high agreement with the real values. The proposed reliability analysis strategy has high computational efficiency without sacrificing high computational accuracy compared with the stochastic simulation method and the direct PDEM method.