Cross-domain damage identification based on conditional adversarial domain adaptation

To address the limitations of labelled data and generalisation in machine learning-based structural damage identification, a conditional adversarial domain adaptation network is proposed for damage identification of the structure without labelled data. First, a one-dimensional residual network is used to extract features from raw vibration data, and then a classifier is trained with labelled data from the source domain. A novel domain adaptation strategy is applied to align both features and classes across domains, which is achieved by leveraging the classifier's output as a conditional variable for domain adaptation. Additionally, all training samples are reweighted by entropy-aware to mitigate the influence of these difficult-to-transfer samples. This approach guarantees that the network acquires damage-sensitive but domain-invariant features, allowing for effective damage identification in the unlabelled target domain by the classifier trained on the source domain. Cross-domain damage identification experiments on numerical simply supported beams and the YuXi River Bridge demonstrate the effectiveness of the proposed network. The results indicate that the proposed method effectively overcomes the training difficulty associated with limited labelled data and has superior damage identification performance. Additionally, the method significantly compensates for the modelling discrepancies, allowing the network trained on numerical models to accurately identify damage in physical structures.