Domain-Adversarial Disentanglement and Physics Constraints Enable Robust Structural Damage Identification Under Environmental Variability

Abstract

Structural damage identification under environmental and operational variations (EOVs) remains a major challenge in civil infrastructure health monitoring because environmental shifts can be statistically correlated with damage-sensitive vibration features. This study presents a physics-informed spatiotemporal graph neural network with domain-adversarial feature disentanglement (CP-STGNN). The framework models the sensor network as a physics-guided graph, extracts spatiotemporal features with Chebyshev graph convolution and temporal convolution, suppresses environmental-domain information in the damage representation through gradient-reversal-based adversarial training, and regularizes signal reconstruction using structural dynamics. The term “causal disentanglement” is used here only as a design motivation for reducing environmental confounding; the implemented objective is domain-adversarial representation learning and does not constitute formal causal identification. The framework is evaluated on the Z-24 Bridge, KW51 Railway Bridge, and LANL three-story structure benchmarks. In the Z-24 cross-domain experiment, CP-STGNN attains an F1-score of 0.97 under the reported protocol, while the additional benchmark studies indicate improved robustness to operational variability and nonlinear response patterns. The results support the value of combining structural topology, physics-based regularization, and domain-invariant representation learning, while the applicability of the approach remains dependent on sensor configuration, reference structural information, and the representativeness of the training domains.