A Domain-Adversarial Graph-Based Framework for Scalable Damage Localization in Real Monitored Bridges

This paper presents a transfer-learning framework for vibration-based damage localization in real monitored bridges without requiring a calibrated model of the target structure. The method exploits an archetypal finite element (FE) model as a labelled source domain, from which synthetic damage scenarios are generated, and long-term monitoring data from the real bridge as an unlabeled target domain affected by environmental and operational variability. After mapping modal frequencies and mode shapes into a common reference space, the data are processed through a Domain-Adversarial Attention-Based Graph Neural Network (DA-AGNN), in which mode-shape variations are represented as graph signals and frequency shifts are used as complementary global descriptors. The network is trained to predict damage-sensitive parameters associated with different regions of the bridge while enforcing feature invariance between source and target domains through adversarial learning. The framework is validated on real monitored data of an operational bridge in Spain. Results show that the transferred model is capable of detecting meaningful parameter changes under damage conditions and of localizing the imposed structural modifications with good spatial consistency. These findings support the potential of the proposed approach as a scalable strategy for localization-oriented SHM of real monitored bridges.