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 pioneers end-to-end regression for predicting internal nugget size from post-weld surface images, enabling low‑cost, non‑destructive inspection. A comprehensive dataset covering multiple sheet stack‑ups and four fit‑up conditions (standard, initial gap, electrode angle, edge proximity) was constructed. Five mainstream convolutional neural network (CNN) architectures were trained and evaluated. Results show two key findings: (1) feasibility - advanced vision models achieve practical accuracy (R² > 0.8) on in‑distribution data, confirming that surface morphology contains strong predictive cues for nugget growth; (2) generalization - all models suffer severe performance drops under abnormal fit‑ups and unseen stack‑ups. Grad‑CAM analysis indicates that while effective predictions rely on indentation borders and oxidation patterns, poorly generalizing models 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.