Structural Damage Detection of Steel Corrugated Panels Using Computer Vision and Deep Learning

In recent years, steel corrugated panels have been introduced in Canada to construct long-span and low-rise frameless systems. However, such systems are prone to have critical structural damages such as buckling under extreme loading conditions such as earthquakes. Identification of these damages is crucial for owners to make informed decisions. In recent years, computer vision methods have been successfully developed and applied in structural visual damage detection for different types of structures including concrete, steel, masonry and timber structures, ranging from regional-scale post-disaster collapse identification, to localized applications such as metal surface defects detection, joint damage detection, concrete crack and spalling detection. However, there are almost no attempts in vision-based detection of buckling damage. In this paper, a hierarchical buckling damage detection framework has been proposed for steel structures, which consists of system-level buckling identification, component-level buckling localization. First, global buckling identification is performed on the image of the panels using the convolutional neural networks (CNN)-based classification algorithms. If the panel is identified as buckled, then the YOLOv3-tiny object detection algorithm is applied to localize the damaged area. Extensive monotonic and cyclic laboratory tests have been conducted on the steel corrugated panels, where image and video data are collected for training, validation, and testing of the CNN algorithms. Results indicate that the CNN-based vision methods can achieve high accuracy in detecting and localizing the buckling damage for the steel corrugated panels. Moreover, additional discussion about further investigations of these steel panels is also presented.