Online prediction of resistance spot weld quality and model explanation under fluctuating conditions

Based on the features of multi-sensing signals in resistance spot welding process, online prediction models were established for the spot weld quality of different stack-ups in this paper. The influence of fluctuating welding conditions fluctuation on four machine learning regression models was studied, and the adaptability of different models and input variables on the database containing data of abnormal conditions was analyzed. Shapley value, and t-SNE methods were used to explain the model performance under fluctuating conditions. The results show that the Gaussian process regression model and resistance + force signal input had the best prediction performance of nugget diameter. Features of welding current, heat input and peak value of electrode displacement had good universality under fluctuating conditions. Besides, the difference of feature distribution caused by condition fluctuation could significantly influence the generalization performance of regression models. Thereby, the reduction of the difference between training set and test set could improve the prediction accuracy.