Weld Surface Defect Detection Based on a Laser Scanning System and YOLOv5

Abstract

Most welding defects are critical and being stress concentrators, their presence always has a bad influence on the tensile strength and other mechanical properties of steel structures. The purpose of this research was to develop a system for high-precision inspection of welded joints using a welding robot scanning computer system. This paper describes a machine-learning approach and the development of an artificial neural network model based on the YOLO structure [1]–[3] for the detection of surface defects in welded joints. The dataset includes images of welds with a laser line provided by an iRVision computer vision system. An artificial neural network model was developed based on the data obtained to analyze this dependence and predict the presence of defects in the images. The possibility of the welding defects inspection using the iRVision 3DL laser scanning system on a Fanuc robot was established. The key metrics of the developed model reached an accuracy of 0.95, a precision of 0.86, and a recall of 0.92 with a confidence parameter of 0.4 on the test data set. The model could become a tool for the implementation of robotic weld inspection to improve the iRVision 3DL system or other systems with the ability to acquire laser line profile images on the scanning object. Further research could be focused on increasing the value of the recall metric without a decrease in the precision metric value.