Structural reliability assessment based on low-discrepancy adaptive importance sampling and artificial neural network

In the field of structural reliability, the estimation of failure probability often requires large numbers of time-consuming performance function calls. It is a great challenge to keep the number of function calls to a minimum extent. The aim of this paper is to propose an approach to assess the structural reliability in an efficient way. The proposed method could be viewed as a hybrid reliability method which combines the advantages of adaptive importance sampling, low-discrepancy sampling and artificial neural network. In the proposed method, artificial neural network is introduced to alleviate the computational burden of deterministic and boring engineering analysis, and its introduction guarantees the computational efficiency of the proposed method. While the Markov chain process is adopted to generate the experimental samples which are used to construct the artificial neural network, the introduction of Markov chain process guarantees the adaptivity of the proposed method and makes the proposed method applicable for various reliability problems. The proposed method is shown to be very efficient as the estimated failure probability is very accurate and only a small number of calls to the actual performance function are needed. The effectiveness and engineering applicability of the proposed method are demonstrated by several test examples.