Development of Portable Crack Evaluation System for Welding Bend Test

This paper describes a system to evaluate the crack severity for the welding bend test fragment. The examination in the welding qualification test in Japan, is conducted by human visual inspection and its burden is concern. The authors constructed an equipment to photograph the fragment specimens under the stable optical condition. The proposed system is also designed for portability to assist the evaluator in the field. We employed Resnet18 to evaluate the given image. The image input layer of original Resnet18 was remodeled from 224-by-224 to 500-by-500 to capture the crack feature in detail. The output layer is replaced with three classification nodes such as “Bad,” “Good,” and “Neutral” expressing the crack severity levels. Experiments showed that 83% accuracy were obtained, confirming that CNN adequately captured the surface crack conditions. The experimental details, remarks, and future works are discussed.