Deep Learning-Based Welding Process Optimization and Adaptive Control Method for Welding Robots

GMAW (gas metal arc welding) robots are frequently used in contemporary manufacturing as they offer high levels of productivity, repeatable process stability and less need for manual labour. However, even today practical robotic welding is often based upon manual design, teaching and offline programming. Welding parameters and trajectories must be redesigned whenever the workpiece, joint geometries or assembly conditions change. Furthermore, during welding, changes to the weld-seam state caused by machining error, assembly deviation and cyclic heat input result in weld parameters planned for ideal conditions becoming inadequate and the risk of incomplete fusion, porosity, cracking, slag inclusions and unstable backside formations greatly increased. In this paper a method based on deep learning is presented to provide optimization of the welding process and adaptive control for GMAW robots. This method consists of integrated quality grade oriented process planning; site-specifically improved multi-objective parameter optimization; semi-supervised semantic segmentation for the perception of weld seams and molten pools and their associated spatter suppression through neural networks predicting weld formation; and fuzzy adaptive control logic. Butt joints of Q235 carbon structural steel were tested experimentally and a response surface model was developed linking welding current, welding speed, torch angle and nozzle height to geometric properties of the welds.