Generalized Damage Diagnostics of Two Miter Gates Through Domain Adaptation

Scarce monitoring data spanning unique civil infrastructure classes limits damage diagnostic generalization using data-driven approaches. While physics-based methods can address some limitations, they struggle to accurately simulate complex damage mechanisms, such as corrosion and fatigue cracks. To alleviate the data scarcity challenge, this study proposes generalizing diagnostic algorithms from structures with abundant data on condition and performance (source) to similar structures with little or not data (target). By leveraging diagnostic mappings through domain adaptation, it becomes possible to diagnose analogous structures within a class effectively. This research focuses on multidamage diagnostics between two miter gates using an adversarial domain adaptation approach. Two types of damage are simulated using validated finite element models. Diagnostic knowledge from the source gate is transferred to the target gate, enabling accurate diagnostics despite the lack of labeled data for the latter. The results demonstrate that the proposed approach achieves reasonable accuracy in diagnosing unlabeled target structures, highlighting the feasibility of transferring diagnostic information across domains with similar label ranges.