Modeling and Optimization of Adjustment of Human Welder on Weld Pool Dynamics for Intelligent Robot Welding

An improved machine–human cooperative control system was developed to obtain sufficient data pairs for modeling welder’s adjustment on weld pool dynamics by data-driven approaches. Spectral analysis shows that weld widths are apparently changed due to low-frequency variations of the welder’s hand movement, which can be filtered by a low-pass filter to remove the high-frequency components. The effect of welding torch orientation on weld pool dynamics was numerically studied to understand the mechanism of human welder’s adjustment and to provide the useful data for control system optimization. A gay multiple linear regression model (GMLRM) was employed to analyze the contribution and interactive compensation of each adjusted parameter on the weld widths as the welding current randomly changes in a given range. A nonlinear adaptive kernel radial basis function neural network (AK-RBFNN) was also proposed to improve the model accuracy. Results indicate that the redundant, coupled, and integrated hand adjustments are adopted to maintain the desired weld pool status, and the human welder’s adjustment reflect nonlinear, complex characteristics. Results also show that the proposed AK-RBFNN model can appraise the weld widths with a good accuracy.