Abstract
This paper focuses on the challenge of online predicting resistance spot welding (RSW) quality, which is affected by the fit-up condition that inevitably varies in production due to assembly deviation but is rarely considered in previous research. To address this gap, this paper develops online prediction models for RSW quality based on multi-sensing signal features. Four widely-used machine learning regression models, namely multiple linear regression (MLR), Gaussian process regression (GPR), support vector regression (SVR), and multilayer perceptron regression (MLPR) were trained on datasets with varying fit-up conditions. The models’ adaptabilities to input signals and their generalizations to different fit-up conditions were assessed. Specifically, SHAP values were adopted to determine the contribution of each signal feature for model interpretation, and the t-SNE method was utilized to explain the influence of fit-up conditions on the model performance. The results demonstrated that the GPR model has the best prediction performance, with an accuracy of 92.2% and RMSE of 0.591 mm. Among the 26 features, welding current, heat input, and average electrode displacement exhibit the highest universality for varying conditions. When only limited signals are allowed as input, resistance together with displacement signals can achieve the best prediction. However, the model generalizations to varying conditions are low, particularly to the edge proximity condition. Therefore, it is important to minimize the discrepancy between the training and test datasets to improve the model generalization.
