Multi-task Weld Seam Recognition Network for Welding Robot based on Structured-light Vision

Robot welding system with structured light vision has been widely used in chemical industry construction, aerospace, and other fields for the welding tasks that require high precision standard. The accuracy of weld recognition is the key to improving weld quality and the robot servo performance. The challenge of weld seam detection is how to deal with the influence of the arc and splash which are existing commonly during the welding process. In this paper, a novel contour-based approach has been developed to detect the weld seam in terms of key feature points which are useful for welding robots to make the welding planning. The proposed network can process multitask simultaneously, and it can not only estimate the location of the weld key-point but also can identify different weld joints without any preprocessing or postprocessing. The experimental result shows that the actual weld seam detection error is within 0.3 mm in the noisy welding process and that illustrates the effectiveness of the proposed work.