Causal discovery of signal morphology drivers in steel resistance spot welding: a high-dimensional data analysis

Abstract Resistance spot welding is a primary joining technique in the automotive and general metal manufacturing industries due to its cost-efficiency and potential for automation. As manufacturing transitions to Industry 4.0, an unprecedented volume of high-frequency process data is recorded. Recent advances rely on machine learning to correlate process data with weld quality. However, existing correlational models are typically trained under restricted laboratory conditions, lacking both physical interpretability and generalization across heterogeneous stack-up configurations with varying materials, coatings, thicknesses, and electrode guns. Here we show that by applying a two-phase causal discovery framework and time series clustering to 15,096 representative profiles derived from over 1.8 million welds, one can extract the structural dependencies of signal morphology from multicollinear stack-up configurations. Increased alloying content is associated with accelerated localized heating, advancing dynamic resistance peaks and suppressing secondary peaks in press-hardened steel. Sheet thickness proportionally scales dynamic resistance, displacement, and force amplitudes, while serving as a primary indicator for post-expansion stability. Reducing tip face diameter from 8 to 6 mm increases the initial resistance peak by 9 % and alters post-expansion mechanics. Finally, at a constant thickness, internal faying surface quantity remains statistically insignificant to macroscopic resistance.