Three-Dimensional Weld Pool Monitoring and Penetration State Recognition for Variable-Gap Keyhole Tungsten Inert Gas Welding Based on Stereo Vision

K-TIG welding offers the advantages of single-sided welding and double-sided formation, making it widely used for medium/thick-plate welding. The welding quality of K-TIG is closely linked to its penetration state. However, the assembly gap in medium/thick-plate workpieces can easily result in an unstable penetration state. In K-TIG welding, the geometric characteristics of the weld pool are closely related to the penetration state. Compared to arc voltage sensing and acoustic signal sensing, visual sensing is a method capable of obtaining the three-dimensional geometric features of the weld pool. To this end, a K-TIG weld pool three-dimensional monitoring algorithm based on a semantic segmentation network using a stereo vision system with a single High-Dynamic-Range (HDR) camera is proposed in this paper. In order to identify the assembly gap of medium/thick-plate workpieces, a gap width extraction algorithm based on the watershed method is proposed. Subsequently, a penetration state recognition model is constructed, taking the three-dimensional geometric features of the weld pool and the gap width as inputs, with the penetration state as the output. The relationship between the input features and the accuracy of penetration recognition is analyzed through feature ablation experiments. The findings reveal that gap width is the most critical feature influencing the accuracy of penetration recognition, while the area feature negatively affects this accuracy. After removing the area feature, the accuracy of the proposed penetration recognition model reaches 96.7%.