Deep learning-based bridge damage identification approach inspired by internal force redistribution effects

Damage identification has always been one of the core functions of bridge structural health monitoring (SHM) systems. Damage identification techniques based on deep learning (DL) approaches have shown great promise recently. However, DL methods still need to be improved owing to their poor interpretability and generalization performance. The fundamental reason lies in the separation between physics-based mechanical principles and data-driven DL methods. To address this issue, this paper proposes a physics-inspired approach combining the data-driven method and the internal force redistribution effects to perform efficient damage identification. Firstly, the mechanical derivation of internal force redistribution is given based on a simplified three-span continuous bridge. Then, two types of typical damage scenarios including segment stiffness decrease and prestress loss are simulated to formulate the damage dataset with monitored field data noise added. Next, a modified Transformer model with multi-dimensional output is trained to obtain the complex dynamic spatiotemporal mapping among multiple measurement points from the intact structure as a benchmark model. Finally, the relationship between multiple damage patterns and the corresponding output regression residual distribution is studied, based on which the flexible combinations of the sensors are proposed as the test set to characterize the internal force redistribution due to damage. Validation on the extended dataset showed that this approach is effective to realize preliminary identification of damage patterns and resist interference from noise at the monitoring site.