Unsupervised Damage Localization Method Based on GAN-BiLSTM Response Modeling

The unsupervised structural damage identification approach has garnered significant focus recently. Currently, most unsupervised approaches lack consideration of spatial correlation among the sensors; thus, unsupervised damage localization performance that can be applied to actual engineering structures is rarely achieved. The generative adversarial network (GAN) excels at mining both low-frequency and high-frequency characteristics embedded in the data. In addition, bidirectional long short-term memory (BiLSTM) networks are superior at extracting the data’s spatial-temporal correlations. Therefore, by combining GAN and BiLSTM networks, GAN-BiLSTM is proposed in this paper to build the correlation among sensors by reconstructing the response. The change in sensor correlation caused by damage can be characterized by the reconstruction error. Finally, the reconstruction error is utilized for damage detection and localization in an unsupervised manner. The proposed method was verified by a numerical model and a practical civil engineering structure. The study indicates that the proposed approach can detect and localize minor structural damage despite only utilizing low-order modal information of acceleration response collected by the limited quantity of sensors. Moreover, in practical engineering applications under environmental and/or operational variability, the proposed approach can autonomously provide early warning of damage locations for the structure.