Accurate structural displacement reconstruction from acceleration and computer vision measurements using physics-informed neural networks

Abstract

Displacement is a critical parameter in structural health monitoring (SHM). However, existing approaches, including direct measurement, indirect measurement, and data fusion techniques, face inherent limitations. To address these challenges and achieve high-precision displacement reconstruction by leveraging multiple structural responses, this study proposes a physics-informed neural networks (PINNs) method that integrates accelerations and computer vision-based displacements sampled at limited resolution and frame rate. The governing equation for displacement reconstruction is derived and incorporated as a physical constraint within the loss function to enhance the interpretability of model training. This method enables the estimation of high-precision displacements for various conditions and measurement locations not included in the training dataset. To validate the effectiveness of the proposed approach, experimental tests were performed on a scaled model bridge, as well as field tests on a 50-meter-span tied arch bridge. The results indicate that the integration of physical information can effectively guide and constrain the process of network training. In addition, the displacements reconstructed by the proposed PINNs method exhibit lower relative errors and higher correlation compared to those obtained by different vision-based methods, traditional neural networks, and conventional data fusion approaches. Furthermore, the spectral analysis results demonstrate that PINNs can accurately reconstruct the high-frequency components of structural responses that are not captured by the vision-based displacements.