Efficient Bayesian model updating of large-scale bridges based on physics-informed neural networks surrogate model

An accurate baseline finite element model (FEM) is essential for the subsequent structural condition assessment and damage identification. Bayesian framework is widely adopted to update the initial FEM for a more accurate baseline, owing to its noise robustness and uncertainty quantification. However, the solution of the posterior distribution of structural updating factors usually relies on Markov chain Monte Carlo (MCMC) method. The redundant and time-consuming matrix assembly in repeated numerical modelling causes a serious computational burden, preventing the application to large-scale structures. To address this drawback, a novel physics-informed neural network (PINN) surrogate model is proposed based on natural frequency order correctness and mode shape mass orthogonality. The PINN surrogate model, combined with Hamiltonian Monte Carlo sampling, forms an efficient Bayesian FEM updating method applicable to large-scale continuous rigid-frame bridges. Comparative analyses with other surrogate models, including neural network, Gaussian process regression, and Polynomial chaos expansion, show that the proposed PINN surrogate model can ensure physical consistency on the limited and unseen datasets, resulting in better fitting performance and extrapolation ability, and significantly enhances the updating efficiency through fast inference. Results also demonstrate that the proposed approach achieves good updating performance on the large-scale bridge.