Unsupervised Quantitative Structural Damage Identification Method Based on BiLSTM Networks and Probability Distribution Model

Current unsupervised structural damage identification methods mainly focus on unsupervised damage detection, i.e., detecting whether damage occurs, but rarely involve damage localization and quantification in practical civil engineering structures. Thus, based on bidirectional long short-term memory (BiLSTM) networks and a probability distribution model, i.e., generalized extreme value distribution (GEVD), this paper proposes an unsupervised structural damage detection, localization, and quantification method. The response correlations model among sensors under healthy conditions was built by employing the BiLSTM networks for reconstructing response data. Then, reconstruction error between reconstructed and true data is used as damage-sensitive features and thus is applied to detect and localized damage. In addition, a stable damage quantification result can be obtained based on the probability distribution model of damage-sensitive features. The proposed method is validated by using the data obtained from a numerical steel beam model and a practical long-span cable-stayed bridge. The research results demonstrate that the proposed method can effectively detect, localize, and quantify damage in an unsupervised manner, even with a limited number of sensors and low-order modal information in acceleration data. The proposed method possesses the capability to autonomously furnish alerts regarding structural damage location and severity under Environmental and Operational Conditions.