Real-Time Information Extraction in Welding Image Acquisition System Based on Spatial Filtering

Real-time information extraction can provide real-time data support to improve welding quality and be used for the analysis and modeling of automatic welding. In gas tungsten arc welding, the weld pool and the arc play essential roles in weld shaping and penetration depth. In this study, built a welding image acquisition system based on spatial filtering, and the YOLOv5m algorithm was used to extract geometry information about the weld pool and the arc from the system. The dataset was collected by a homemade welding image acquisition system based on spatial filtering. The accuracy of YOLOv5m on mAP@0.5 is 99.5%, and the prediction speed is 52.63 frames per second (FPS). The algorithm was used to map weld stability and smoothness of weld reinforcement through information extracted in real-time such as weld pool width, weld pool length, arc width, and arc length. The reliability of the results is verified by comparing them with the weld seam images.