Deep Learning-based Initial Structural Damage Detection Approach via Sub-structuring Class Activation Map

Initial defects or damages upon a structure will be propagated throughout the entire structure. Therefore, it is important to detect damage at an early stage to prevent such influence of the damage to the entire structure. Recently, digital image correlation (DIC) has been utilized to measure the deformation or monitor the robustness of structures. Since the damages upon the structure affect the displacement/strain, if the degree of damage is large enough, the location of the damage can be predicted with the naked eye. However, there may be a limit to visual analysis of initial damage which may be the case when considering DIC measurements. In this paper, class activation map (CAM), an explainable artificial intelligence, is used to predict the presence and location of damages. Herein, the DIC measurements are assumed. Thus, the relevant displacements and strains are obtained via the finite element method. The resulting CAM model, trained on the relationship between strain and damage, predicted the presence and location of damages, and shows good accuracy as higher than 99%.