Unsupervised Spatiotemporal Framework for Structural Damage Localization and Severity Quantification

The research on structural damage identification based on unsupervised learning mainly focuses on the stage of whether damage has occurred, while the research on damage localization and damage quantification is relatively scarce. Based on this, this paper proposes an unsupervised structural damage localization and quantitative identification method based on the combination of Convolutional Autoencoder (CAE) and Long Short-Term Memory (LSTM), which fully combines the advantages of the model in spatiotemporal feature extraction. First, the data is reconstructed through the established CAE-LSTM network, and then the reconstruction error between the reconstructed data and the real data is used as the damage sensitive feature to locate the damage. Subsequently, based on the normal distribution model of the damage sensitive feature, a stable damage quantification result is obtained, thereby realizing the quantitative assessment of the damage degree. The effectiveness of the proposed method was verified through the data obtained from the Qatar grandstand model and the IASC-ASCE SHM benchmark model. The results show that the proposed method is applicable to both single-damage and multi-damage scenarios. A comparative analysis was also conducted with the identification results of the traditional CAE network. The results showed that the CAE-LSTM network exhibited outstanding identification capabilities for both minor damage caused by bolt loosening and severe damage resulting from the removal of intermediate supports. Its identification results were significantly superior to those of the traditional CAE network for structural damage localization and quantification under unsupervised conditions. CAE-LSTM also holds great application potential in the damage identification of complex structures.