Data-driven process characterization and adaptive control in robotic arc welding

Robotic arc welding (RAW) has been an essential process in various assembly systems, such as automotive manufacturing. However, its implementations lack adaptivity to compensate for process variations. This paper presents a data-driven process characterization and online adaptive control framework for RAW to automatically and efficiently achieve desired weld pool condition, given any initial conditions. Based on optical imaging, pool width is characterized through a pixel-level image segmentation network and then used for determining the parameter adjustment for robotic execution through a gradient-based controller. Experiments demonstrate quick process convergence within 7 adjustment periods and an error band within 10.9%.