Physics-guided neural networks for bridge health monitoring: A state-of-the-art review

Appropriately utilizing the data recorded by the bridge health monitoring (BHM) system to identify structural damage, evaluate structural safety, and predict structural serviceability is the core work in the community of BHM. Neural network models are more flexible to describe multi-source, multi-dimension and non-linear relationships, comparing with traditional statistical and regression analysis, and have been widely used for data-driven evaluation of bridge performance. But it is easily influenced by noise and errors that are difficult to eliminate in the monitoring data. Physics-guided neural networks (PGNNs), which combine physical information with neural networks, have stronger accuracy, robustness, and reliability, and are becoming promising tools for bridge performance evaluation. In the past few years, numerous researchers all over the world paid intensive attention on this topic. This paper summarizes the latest developments of PGNN methods for BHM. The commonly used PGNNs are classified into three categories, including the physics-guided loss function, the physical data enhancement and the digital twin. Following that, the applications of the three types of PGNNs are presented through a summary of relevant literature. Finally, the challenges and prospects of PGNN methods in the field of BHM are discussed.