Integrated Physics-Informed Machine Learning Framework for Structural Damage Detection, Localization, and Severity Classification

Abstract

Structural health monitoring (SHM) plays a critical role in the early identification and assessment of structural damage, thereby enhancing the safety and reliability of civil infrastructure. Structural damage identification generally encompasses three key tasks: damage detection, localization, and quantification. While extensive research has been conducted on each of these tasks individually, relatively few studies have integrated all three components into a unified framework for comprehensive structural condition assessment. Physics-based approaches require an accurate finite element model (FEM), which is often difficult to calibrate to accurately represent the behavior of the actual structure. In contrast, data-driven approaches rely on sufficient labeled data collected from the actual structure, which is likewise challenging to acquire in practice. To address these limitations, this work proposes an integrated hierarchical physics-informed domain adaptation (I-HierPhyDA) framework that performs damage detection, localization, and severity classification in a hierarchical manner. The proposed framework bridges the gap between the reduced-order FEM and the higher-fidelity FEM by generating vibration signatures that are consistent across both domains. Furthermore, the proposed framework enables structural damage localization without requiring damage-state data from the target domain during training, while damage severity classification is performed using transductive domain adaptation with unlabeled damaged-state data from the target domain. The proposed framework is systematically evaluated using the numerical ASCE benchmark models under structural uncertainties and measurement noise. The results demonstrate that the proposed approach achieves accurate structural damage detection and localization. For structural damage severity classification, it achieves the highest mean accuracy and Macro-F1 score while exhibiting the lowest standard deviations for both metrics among the baseline and ablation methods, demonstrating its effectiveness for comprehensive structural condition assessment. Future work will focus on experimentally validating the proposed approach using measured data from laboratory or field structures.