Structural Seismic Response Reconstruction Using Physics-Guided Neural Networks

Abstract

Reconstruction of data loss in structural seismic responses is important for structural health monitoring to evaluate the safety of structures. A physics-guided neural network that leverages the prior knowledge was proposed for reconstructing structural seismic responses that were inaccessible to measure or missing during earthquakes. The presented methodology consisted of convolutional neural networks with dilated kernel and fully connected neural networks, which were developed to achieve a multitask learning that involved the regression task with measured labeled displacement data and the reconstruction task of seismic response without any labels. To better balance the loss gradient across different tasks, a probabilistic model was introduced to optimize the weight coefficient for each task by quantifying the task-dependent uncertainty based on Bayesian statistics. The weight coefficient for each task can be dynamically updated during the training process, thereby improving the learning efficacy and performance accuracy of the neural networks. The probabilistic model with task-dependent uncertainty was validated to outperform the equal-weighted model (i.e. equal weight for each task) in reconstructing the structural seismic responses based on numerical data, even when the relevant physical information (i.e. Bouc–Wen model) was not complete.