Novel physics-informed neural network approach for dynamic and static displacement reconstruction via strain and acceleration

Abstract

The real-time deformation reconstruction of the plate-beam composite structures is an immediate requirement to monitor and safeguard the structures, such as ships, aircraft, etc. We present a novel physics-guided machine learning method based on Convolutional Neural Network (CNN) to reconstruct the dynamic and static displacement of structures. The essence is to integrate the priori physics of the acceleration into deep learning. A network with two individual branches is designed to extract the comprehensive information from inputs. The multi-task weights are balanced by choosing appropriate scale and penalty factors to improve the physics-based training efficiency. The numerical examples and experiments for structural dynamics and statics are presented to demonstrate the superiority of the proposed methodology. It turns out the integration of the acceleration information can reduce measurement costs and improve the robustness of the trained model for more reliable displacement prediction.