Damage localization using a deep learning-based response modeling method

Existing multi-damage localization methods usually need to be trained using labeled data obtained from various damage cases, and such methods can identify multiple damages with high accuracy. However, it’s extremely challenging to obtain labeled data from engineered structures under various damage states, especially in multiple damages case. Thus, damage localization methods that need to be trained using only structural health data have received much attention as an alternative. In addition, existing multi-damage localization methods are mainly based on structural dynamic responses, such as acceleration, whereas structural quasi-static responses are also sensitive to damage location and perform well in damage localization, such as strain response. However, existing quasi-static response-based damage localization methods usually focus on the single-damage localization problem, ignoring the double- and multi-damage localization problems. Therefore, this paper develops a multi-damage localization method based on strain response modeling using the DL-AR-ATT model. The proposed method was compared with one of the latest methods, the BiLSTNet-A-based method, and validated using both simulation and experimental datasets. The results illustrate that the proposed method can accurately locate single and double damage and outperformed the BiLSTNet-A-based method, especially in high noise levels and minor damage cases.