Developing physics-informed neural networks for structural parameters identification of beam with moving loads

Physics-Informed Neural Networks (PINNs) seamlessly integrate the predictive capabilities of neural networks with established physical principles. By integrating constraints such as displacement and force boundary conditions alongside governing equations, PINNs can generate digital twins of physical systems and processes. This fusion allows for more accurate modelling and simulation of complex physical phenomena, bridging the gap between data-driven approaches and traditional physics-based methods. Nevertheless, the practical implementation of PINNs remains challenging, primarily due to numerous influential hyperparameters and the complex nature of modelling the governing physics through partial differential equations (PDEs). This challenge becomes especially critical in the context of dynamic loads, where higher-order PDEs encompassing both spatial and temporal domains, alongside relevant structural parameters and generalised (distributed) load’s function, must be carefully optimised during the PINNs training process. This study presents a novel application of PINNs model, developed, trained, and validated using real-world bridge monitoring data, for the inverse problem of predicting structural parameters of a girder subjected to moving loads. Two case studies are considered. In the first, PINNs model is utilised to estimate the structural parameters of a bridge girder under varying levels of noise in the data. In the second, the model is trained with actual field monitoring measurements to estimate structural parameters while predicting girder deflection and other internal forces. The findings advance the existing body of knowledge in structural health monitoring (SHM) by demonstrating a practical PINNs-based solution for bridge girders under moving loads.