A cross-condition benchmark and interpretability analysis for vision-based nugget size prediction in resistance spot welding

Resistance spot welding (RSW) demands reliable quality assessment, yet traditional inspection methods are often destructive, costly, or inefficient. While computer vision has been applied to RSW, existing studies are predominantly limited to the classification of surface defects. This study represents a pioneering attempt to achieve low-cost, non-destructive quality inspection by performing end-to-end regression prediction of internal nugget size directly from post-weld surface images. We constructed a comprehensive dataset spanning multiple sheet stack-ups and four typical fit-up conditions—standard, initial gap, electrode angle, and edge proximity. Five mainstream Convolutional Neural Network (CNN) architectures were trained and systematically evaluated to assess the potential of surface-to-internal quality mapping. The results highlight two core findings. First, regarding feasibility, the study demonstrates that advanced vision models combined with complete surface imagery can achieve practical prediction accuracy (R² >0.8) on in-distribution data, confirming that weld surface morphology contains physical cues highly predictive of nugget growth. Second, the study explicitly reveals and quantitatively analyzes a critical challenge: model generalization. All models exhibited severe performance degradation when the test distribution deviated from the training conditions, specifically under abnormal fit-up conditions and unseen material stack-ups. Grad-CAM interpretability analysis corroborates these trends, showing that while effective predictions rely on indentation borders and oxidation patterns, models struggling with generalization over-attend to background artifacts. This work clarifies both the practical applicability and the generalization boundaries of surface-vision-based regression, suggesting that future work must focus on domain adaptation techniques and data augmentation strategies to overcome distribution shifts in diverse industrial environments.