Developing physics-informed neural networks for virtual sensing in beams with moving loads

Physics-informed neural networks (PINNs) show promise for structural health monitoring and related applications since they have the potential to represent physical systems and processes by conforming to governing differential equations as well as physical constraints such as displacement and force boundary conditions. This paper addresses the challenge of developing PINNs that can predict the response of bridge structures under a moving concentrated load such as from the axle load for a vehicle. It also demonstrates the potential of PINNs for virtual sensing, i.e., predicting response parameters at locations where sensors may not be available. The proposed PINNs assume that the load moves at constant speed and uses the time at which the load enters the bridge to evaluate its position with respect to time. The study initially proposes a one-dimensional PINN that takes only response location as input and gradually increases dimensionality (number of inputs) to three by including the vehicle passage time and load magnitude as additional inputs. The study conducts a thorough sensitivity analysis of the PINN model’s parameters to understand their influence on model training and performance. The performance of the final PINN model is investigated for a real-world bridge girder for which monitoring data is available. The capacity of the PINN to predict strains, deflections and internal forces in locations that are not physically equipped with sensors is demonstrated, and strain predictions are observed to closely follow actual measurements from field monitoring. • Proposing PINN models to predict the response of bridge girders under moving loads. • Sensitivity analysis validates and optimises parameters for PINN setup, training. • Case studies show PINN applications to bridge girders with performance analysis. • PINN predicts strain profiles for a real bridge, matching in-situ monitoring data.