Quality Assessment of RSW Based on Transfer Learning and Imbalanced Multi-class Classification Algorithm

• Published in: IEEE access Vol. 10; p. 1
• Main Authors: Guo, Peijin, Zhu, Qinmiao, Kang, Jingran, Wang, Yuhui, Hu, Wenqiang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Automotive engineering; class-imbalanced classification; Classification; Data models; Electrodes; Machine learning; Model accuracy; Parameter identification; Prediction models; Predictive models; Process parameters; Product safety; Quality assessment; Regression models; Resistance; Resistance spot welding; Shear strength; Training data; Transfer learning; Welding parameters
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3212410

In automobile manufacturing, the quality assessment of resistance spot welding (RSW) plays a decisive role in the quality and safety of products. Recently, it has become very popular to use machine learning to evaluate the quality of welding nuggets. However, there are two obstacles: data imbalance caused by limited defective samples, and data shortage due to expensive time and labor costs. This paper proposes a novel method. On one hand, the self-paced ensemble (SPE) algorithm for binary classification is improved to handle imbalanced multi-class classification of quality levels. On the other hand, an instance-based ensemble transfer learning approach is proposed to predict the tensile-shear strength of RSW for precise control of the weld quality. In detail, a quality level identification model is formulated with the process and material parameters as the input at first. Secondly, an explainable algorithm SHapley Additive exPlanations (SHAP) was introduced to anatomize the impacts of welding parameters on the welding quality predictions. Finally, a hybrid dataset including actual historic production data and 454 spot-welding cases is constructed, and then the eXtreme Gradient Boosting (XGBoost) is introduced as the base learner of TrAdaBoost.R2 to train the prediction model. Compared with conventional methods, the SPE provides the greatest macro geometric-mean score of 0.923, and the proposed regression model yields superior accuracy R 2 of 0.952, which shows the potential of assisting welding process design.