Class activation map을 활용한 변위/변형률 정보 기반 구조물 손상 탐지 기법1

Initial structural damage to a structure propagates and leads to failure. Detecting damage at an early stage is important to prevent failure. Recently, studies on measuring the displacement of a structure and predicting the strain field using digital image correlation (DIC) have been actively conducted. However, when the degree of damage is small or occurs in an area where stress is concentrated, it is difficult to identify it by visual inspection. Yet, it is impossible to process the real-time strain field data of large structures such as wind turbine blades, bridges, etc. In this paper, a novel real-time damage detection method using a class activation map (CAM) is developed. The CAM learns the relationship between strain field and structural damage location. After training is finished, the CAM is not only able to detect damage location from an untrained structural strain field, but also to print images of damage locations. The developed model predicts damage with 99.2 % accuracy.