Damage identification of hinge joints is crucial for ensuring operational safety of prefabricated bridges. However, the engineering application of existing data-driven methods is restricted by scarce labeled data in actual bridges and significant domain shifts between simulation (source domain) and monitoring data (target domain). Furthermore, actual bridges typically exhibit only limited bridge states, which can induce negative transfer in conventional domain adaptation. This paper proposes an attention-guided Partial Domain Adaptation (PDA) method for damage identification of hinge joints in prefabricated bridge subjected to random traffic flow. Firstly, a transformer feature extractor is employed to learn discriminative representations from acceleration responses at mid-span measured by limited sensors. Then, a transferability evaluation module and linear cosine cross-domain attention mechanism are introduced to construct Weighted Maximum Mean Discrepancy (WMMD), which facilitates fine-grained alignment of shared classes between the source and target domains at the instance level, and effectively suppresses interference from private classes in the source domain. Besides, the cross-domain attention weight distribution provides visual explanations for transfer effectiveness and decision mechanisms of network model. Numerical examples demonstrate that even under the dual challenges of random traffic excitation and limited sensors, the proposed method effectively mitigates the adverse effects from label space differences and environmental noise. The identification accuracy is significantly improved compared with traditional supervised method and global domain adaptation method, offering a low-cost and effective solution for hinge joint damage identification with sparse observations.
Li et al. (Thu,) studied this question.