Development of Automatic Crack Detection Technology in Welded Surface using Laser Active Thermography and CNN Deep Learning

In this study, automatic crack detection for welded surfaces was studied through the development of a laser active thermography system and a crack detection algorithm. The laser active thermography system observes thermal wave concentrations in the crack while exciting the surface of the weld. The crack detection algorithm (1) visualizes the cracks by merging the infrared (IR) images using the temperature distribution characteristics; (2) employs input image generation with a specific method to prevent overfitting; (3) analyzes and classifies the characteristics of the cracks using a deep learning convolutional neural network (CNN); and (4) marks the location of the cracks in the original IR image. The system and algorithm were verified using two SUS specimens (#1 and #2) and compared with actual crack data obtained by microscopy and penetration test. The CNN was trained with 618 images of cracks and 1834 images of intact specimen #1. For performance verification, a total of nine areas of specimens #1 and #2 were divided into 300 test images; 13 out of 14 cracks were detected while four intact images were overdetected. Thus, the developed algorithm can detect cracks in welded surfaces by distinuishing them from complex patterns of welding.