A physics-guided residual learning framework for acceleration response-based structural damage identification

Abstract

Structural damage identification through acceleration response measurements is a noteworthy but challenging task because damage-induced changes are often localized, sensor information is limited, and measured responses are affected by noise. This study proposes a physics-guided residual learning framework (PGRLF) for acceleration-response-based structural damage identification. The framework combines a physics-guided branch, which encodes sensor–element relationships through structural topology and finite element model (FEM)-based sensitivity information, with a data-driven residual learning branch based on a multi-scale one-dimensional convolutional neural network (CNN) through channel-wise attention. A residual fusion strategy is then used to integrate the physics-guided bias with learned signal features in order to generate element-wise damage indices. The proposed method is validated using experimental fixed–fixed aluminum beam specimens tested under hammer-impact excitation and further evaluated on a numerical ninety-eight-element steel frame subjected to seismic excitation. For the steel frame, both noise-free and noise-contaminated response data are considered, and the proposed model is compared with an ablated version without the physics-guided bias and a support vector regression baseline. The results show accurate localization of damaged elements in the experimental beam and proper damage identification for the frame, while maintaining stable performance under high measurement noise levels. The ablation and baseline comparisons confirm that the physics-guided bias improves localization accuracy and robustness. These findings demonstrate that structural topology, mechanical sensitivity, and residual deep learning can provide an effective acceleration-response-based damage identification method.