Physics-guided recurrent neural network trained with approximate Bayesian computation: A case study on structural response prognostics

Abstract

A new physics-guided Bayesian recurrent neural network is proposed in this manuscript. This hybrid algorithm benefits from the knowledge in physics-based models, the capability of recurrent neural networks to handle sequential data, and the flexibility of Bayesian methods to quantify the uncertainty. The introduction of physics in the forward pass of the neural network significantly improves the results in multistep-ahead forecasting, and the gradient-free nature of the Bayesian learning engine provides great flexibility to adapt to the observed data. The proposed algorithm has been applied to a data-driven problem about fatigue in composites, and a case study about accelerations in concrete buildings, where a comparison against the state-of-the-art algorithms is also provided. The results revealed: (1) the accuracy of the proposals, comparable to the state-of-the-art recurrent neural networks; (2) its stability during multiple runs of the algorithm, proving that it is a more reliable option; (3) its precise quantification of the uncertainty, which provides useful information for the subsequent decision-making process. As potential future applications to real world scenarios, the proposed Bayesian recurrent neural network could be used in on-board PHM systems in the aerospace industry, or as an on-site prediction tool in buildings for seismic events and/or aftershocks.