Comparison and explanation of data-driven modeling for weld quality prediction in resistance spot welding

Resistance spot welding (RSW) is an important manufacturing process across major industries due to its high production speed and ease of automation. Though conceptually straightforward, the process combines complex electrical, thermal, fluidic, and mechanical phenomena to permanently assemble sheet metal components. These complex process dynamics make RSW prone to inconsistencies, even with modern automation techniques. This motivates online process monitoring and quality evaluation systems for quality assurance. This study investigates in-situ process sensing and neural networks-based modeling to understand key aspects of RSW process monitoring and offers three contributions: (1) a comparison of two data-driven modeling approaches, a feature-based Multilayer Perceptron (MLP) and a raw sensing-based convolutional neural network (CNN), (2) a comparison of how electrical and mechanical sensing data affect the model's performance, and (3) an explanation of MLP behavior using Shapley Additive Explanation (SHAP) values to interpret the contribution of sensing features to weld quality metric predictions. Both the MLP and CNN can predict weld quality metrics (e.g., nugget geometry) and detect a process defect (i.e., expulsion) using in-situ current and resistance sensing signals. Including force and displacement measurements improved performance, and the SHAP values revealed salient features underlying the RSW process (e.g., displacement contributes significantly to predicting axial nugget growth). Future work will explore additional architectural developments, explore ways to translate lab-developed models to production plants, and leverage these models to optimize RSW processes and improve quality consistency.