Dynamic reliability analysis via data aggregation and selective state-space deep learning architecture

Performing dynamic reliability of structures is an resource-intensive task because it involves a threenested loop computation: the first loop over different structural members, the second loop over discretized time steps, and the third loop over random variables sampled from pre-described probability distributions. To address this challenging problem, this study first designed a feature aggregation process using various techniques such as zeropadding, broadcasting, and position encoding to combine different time-invariant random variables, historical responses, and ground motion into a unified three-dimensional feature tensor. Next, we developed a computationally efficient yet highly accurate metamodel based on a modern selective state space deep learning architecture. Third, a mixed-multiple layer perceptron projection layer is employed to generate outputs with desirable prediction lengths and number of channels. To demonstrate the viability of the proposed method, two examples are presented: one involving a two-dimensional planar frame structure and the other a three-dimensional spatial frame structure. The obtained results demonstrate that the proposed method is consistently more accurate than numerous counterparts, accelerates the seismic reliability analysis of large-size Monte Carlo populations by approximately 5.5 times compared with FEM, and achieves a mean deviation of the reliability index of about 6%. Supplemental studies, including hyperparameter optimization and comparison studies, have been performed, providing further insights into the performance of the proposed method.