Deep learning-powered vision system for seam-tracking and gap width estimation in robotic laser welding

The increasing adoption of robotic laser welding highlights the need for advanced seam-tracking systems to ensure precision, adaptability, and high-quality welds. Conventional methods struggle with fixturing errors, part-to-part variations, thermal deformation, and challenges related to material reflectivity. This work presents a deep learning-based vision system designed for real-time seam tracking and gap width estimation, addressing these limitations by leveraging state-of-the-art computer vision techniques. The system employs the YOLO (You Only Look Once) architecture for seam detection and a MobileNet-based Convolutional Neural Network (CNN) for gap width estimation, ensuring robust performance under variable lighting conditions and reflective surfaces. By processing coaxial images acquired during welding processes, the system accurately identifies the seam position and discretizes the trajectory into a sequence of key points. This information can then be used for further processing or integration with robotic motion control strategies. Experimental validation on an industrial robotic laser welding setup demonstrates the system’s capability to enhance tracking precision, minimize positioning errors, and support high-speed welding operations. The results confirm that deep learning-based vision systems play a crucial role in advancing autonomous robotic welding, increasing flexibility and efficiency in smart manufacturing.