Designing of Welding Defect Samples for Data Mining in Defect Detection and Classification using 3D Geometric Scanners

Using a deep learning technology for welding defect detection and classification could improve the quality and productivity of the welding industry. However, developing the deep learning model based on the supervised learning method requires a large amount of data for good and defective welding. The welding technicians nowadays have been well trained and rarely produce the imperfect welds. Thus, lack of defective welding samples poses a major challenge to design welds defect samples necessary for developing the deep learning model. In this paper, a model and a method are established based on the standard for initial quality assessment in external surface imperfection. The method is used to create imperfection on aluminum plate specimen and to design an experiment using a 3D laser scanner for detecting cracks on weld bead and to generate three-dimensional models for welding quality assessment by visual testing technic (VT) in accordance with the ISO 9712 and American Society of Mechanical Engineer (ASME). The model of the calibration specimen that has been designed is used as a tool to create imperfection (crack) according to ISO-6520-1. As per the design principles of creating cracks, these are at the root of weld, at the heat-affected zone (HAZ) and at the parent material. We determine the size of the cracks by dividing the size according to the location of the cracks in order to obtain a total of 47 cracks in 1 specimen. With this design principle, the specimens will be realistic and it is necessary for the detection software to be highly accurate to correctly detect and classify cracks on actual weld beads.