Physics-guided spatiotemporal learning with invariant graph for cross-domain structural damage identification

Abstract

The scarcity of real-world damage data poses a significant challenge to data-driven structural damage identification. While transfer learning between numerical and experimental domains offers a promising solution, existing methods often overlook the crucial spatial relationships among sensors, which are vital for accurate damage localization. To address this, the work proposes a novel spatiotemporal feature learning network (SfLN) guided by structural physics. Its core innovation lies in an invariant graph structure, constructed deterministically from sensor layout and structural connectivity, which provides a consistent spatial prior across domains. The SfLN synergistically combines a one-dimensional convolutional neural network for extracting discriminative local temporal features from raw signals and graph convolutional network for propagating information through the physics-guided graph. This design explicitly models spatiotemporal dependencies in an inherently transferable manner, enabling effective cross-domain knowledge transfer through fine-tuning without complex domain alignment. Validation on a cross-domain case study of a tower steel structure demonstrates the efficacy of our approach, achieving a damage classification Accuracy and F1 Score of 0.9351 and 0.9346 on the target domain, significantly outperforming direct training. Furthermore, robustness tests under varying noise levels confirm the model’s stability and practical applicability.