Probabilistic method for time-varying reliability analysis of structure via variational bayesian neural network

This study proposes a novel framework computing the dynamic reliability and associated uncertainty quantification of structures under time-varying excitation with significantly reduced time complexity. For this purpose, the deep neural network's power and the Bayesian theory's probabilistic ability are leveraged, forming a Bayesian neural network data-driven model (BNN). The BNN-based surrogate model can yield a probability distribution of outputs of interest, e.g., a limit state function and its derived statistics such as median value, confidence interval rather than only a deterministic quantity. The effectiveness and correctness of the proposed method are reaffirmed via three case studies involving examples from the literature and a 3D numeral model of a prestressed reinforced concrete bridge structure, showing a reduction in time complexity up to three orders of magnitude compared to the Monte Carlo method only using finite element models. As a result, an 11-year maintenance routine is recommended for a marine and chemically aggressive environment to ensure the high reliability of prestressed bridge structures when accounting for uncertainty estimation.

Automated summary

This study proposes a novel framework using a Bayesian neural network data-driven model to compute the dynamic reliability and uncertainty quantification of structures under time-varying excitation, significantly reducing time complexity. The effectiveness and correctness of the proposed method are validated through three case studies, recommending an 11-year maintenance routine for prestressed bridge structures in marine and chemically aggressive environments.