A physics-guided hybrid framework for online pre-expulsion prediction in resistance spot welding

Expulsion in resistance spot welding (RSW) causes weld quality fluctuations and increases quality-control effort in high-volume manufacturing. Existing data-driven studies have mainly addressed post-occurrence expulsion detection, process-end classification, or the identification of influencing factors, whereas online monitoring requires short-term risk estimation before the event occurs. In this study, expulsion prediction is formulated as a sliding-window-based pre-expulsion risk estimation task for the currently welded spot. A physics-guided hybrid GRU-XGBoost ensemble is developed to combine temporal learning from dynamic resistance and electrode-force signals with process-physics-related scalar features describing heat input, resistance state, and force response. The framework was evaluated on 2730 valid welds, including 588 expulsion and 2142 non-expulsion welds, using weld-grouped five-fold cross-validation with fold-level working-point selection. The ensemble achieved an area under the ROC curve of 0.945 ± 0.004 and a weld-level recall of 90.6 ± 3.7% at an average false alarm rate of 9.8 ± 0.2%, outperforming both individual branches. For the 533 correctly warned expulsion welds, the median early-warning lead time was 56 ms. These results indicate that online, physically interpretable pre-expulsion risk prediction is feasible under low-false-alarm constraints within the investigated RSW configuration and provide a basis for future adaptive monitoring and control studies.