Unsupervised deep learning framework for damage identification under ambient excitations: trait of damage localization and demonstrative applications

Developing unsupervised deep learning-based damage identification (uDLdi) under ambient excitations is crucial for civil engineering but remains hindered by insufficient feature extraction from complex original acceleration responses and a lack of damage localization capabilities. To address these issues, this study proposes a novel framework of uDLdi under ambient excitation (F-uDLdi-ae) by adopting the following innovative techniques: (1) a cube-shaped sample set is constructed to theoretically protrude damage information while suppressing interference information; (2) a self-attention module dedicated to protruding damage feature so as to largely identify damage samples is constructed; and (3) an uncertainty-weakened damage probability strategy for localizing damage accurately and reliably is proposed. Validation using a numerical steel truss bridge and a laboratory three-story frame structure demonstrates the superior damage discrimination and localization performance of the proposed framework. Specifically, in the numerical case under a high-noise environment (SNR = 5 dB), F-uDLdi-ae achieves an F‑score of 91.98%, outperforming several advanced unsupervised methods by 6.27% to 20.65%. In the three-story frame structure, F-uDLdi-ae achieves an F‑score of 100%. Furthermore, the framework is applied to a long‑span cable‑stayed bridge, Yonghe Bridge, where it achieves an F‑score of 99.17% and successfully localizes the damage via the maximum damage probability. The proposed F-uDLdi-ae overcomes the limitations of existing uDLdi, holding the promise for viably and intelligently identifying damage in practical engineering structures.