A global strategy based on deep learning for time-dependent optimal reliability design

Time-dependent reliability-based design optimization (RBDO) is a computationally tough problem that needs to be addressed urgently. The difficulty of solving the time-dependent RBDO mainly comes from the time-dependent reliability analysis involved in probabilistic constraints, which itself is one of the thorny problems in the reliability community and makes the computational cost become much more onerous. In this paper, a deep-learning-assisted approach is proposed to solve the time-dependent RBDO. The proposed approach leverages the classification capability of the deep learning, and constructs the alternative model for the actual probabilistic constraint function in the so-called augmented reliability space, so as to make the trained alternative model accurate wherever it will be invoked. Moreover, a sequential sampling technique utilizing the classification probability provided by the deep learning is proposed to further reduce the computational cost. Then, the time-dependent reliability analysis involved in the time-dependent RBDO is conducted by the cheaper alternative model instead of the original computing-intensive probabilistic constraint function, which evidently reduces the computational burden. The presented examples showcase the performance of the proposed approach. Especially, for the complicated engineering application, the proposed approach saves about 10% of the computational cost compared with the existing methods.