Physics-informed neural network for first-passage reliability assessment of structural dynamic systems

Abstract

The first-passage failure of structural dynamic systems is a typical failure mode in engineering. Most of the existing methods use ordinary or partial differential equations (ODEs/PDEs) to describe the dynamic systems, and transform the first-passage reliability assessment into the solution of differential equations. Thus, the efficient solution of ODEs/PDEs is the main difficulty of the first-passage problems, especially for the structures with multi-degree-of-freedom and unknown parameters. For addressing this issue, a novel first-passage reliability assessment method based on physics-informed neural network (PINN) is proposed. In this paper, we firstly use ODEs/PDEs to describe the dynamic properties of the two classical first-passage models: Poisson process model and Markov diffusion process model. Then, the ODEs/PDEs are used to construct loss functions of the neural network, and the PINN evaluation model of the first-passage reliability is obtained finally. The effectiveness and accuracy of the proposed method are verified by several examples, and the potential value of the deep learning framework combined with physical properties and sample data in the field of dynamic reliability analysis is proved.