INTERPRETABLE DATA-DRIVEN PREDICTION OF RESISTANCE SPOT WELD QUALITY

Abstract

Resistance Spot Welding (RSW) has been an important welding process in many industries, including automotive and aerospace, given its great automation when combined with robots and electrical control systems. But real-time RSW quality inspection and assurance have been a challenge on shop floors. This paper presents an interpretable data-driven modeling method, upon neural networks and game theory-based Shapley values, for efficient and also interpretable prediction of weld quality metrics (e.g., nugget diameter, thickness, HAZ diameter) in RSW. The interpretation of the model predictions can contribute to the physical understanding of: 1) how would the weld quality vary under process uncertainties, e.g., varying sheet fit-up conditions; 2) how would in-situ sensing (e.g., resistance, displacement) benefit characterization of process uncertainties; 3) what would be the contributions of individual process parameters (e.g., welding current, time) and different sensing measurements to the prediction of weld quality. Through analyzing experimental data, these questions can be quantitively answered by the interpretation, which hence leads to a better understanding of RSW process dynamics and defect occurrences towards improved process control and quality assurance.