Welding robots have significantly advanced the state of automotive welding by improving operational efficiency and precision. However, robotic welding is not fully autonomous yet. Still relying heavily on preprogrammed welding parameters, current welding robots cannot adaptively adjust parameters to compensate for task variability, material heterogeneity, and process uncertainty as skilled human welders do. The transition from preprogrammed to adaptive robotic welding requires the seamless integration of in-situ perception and process characterization, process-quality relationship quantification, closed-loop control, and robotic execution.
This presentation presents an online adaptive control approach to robotic arc welding, where weld pool width is controlled in a closed-loop. From in-situ optical imaging, an efficient pixel-level image segmentation network is first developed to outline the pool area and estimate the pool width. Then based on a perceptron that describes the dependence of pool width on parameters (i.e., welding speed, current, torch angle), an advanced gradient descent algorithm is investigated to backpropagate the needed pool geometry changes to parameter adjustments based on individual parameters’ contributions to pool status change, while minimizing the settling time and steady-state error band. The main contribution of this work lies in the following two areas:
Efficiency: in-situ images can be processed within milliseconds, thus enabling sub-second periodical process adjustment. Also, the control algorithm greatly reduces the settling time of process adjustment.
Physical interpretation: no black-box mapping from pool state to action determination is involved, as compared to the RL approach. The outputs from the networks and controller can be readily interpreted and evaluated by humans.
Experimental results indicate that the process can be adjusted within 7 adjustment periods to a less than 11% error band. Since weld pool status is a critical and measurable metric in all types of welding processes, choosing it as the control target enables direct physical interpretability and generalizability of the developed system. Assuming the welding path is known and preprogramed into the robotic motion path, the system can be readily transferred to other welding processes (e.g., gas metal arc welding) and application scenarios (e.g., curved welding). The developed system will be expanded by including more capabilities (e.g., sophisticated torch weave, real-time welding path calculation), and be also compared to manual welding as part of the future work.