A new online quality monitoring method of chain resistance upset butt welding based on Isolation Forest and Local Outlier Factor

Abstract Small chain resistance butt welding often has unstable welding quality due to high production frequency and short welding time. However, there is still a lack of research on online monitoring of resistance upset butt welding. This study proposes to collect the state information in the welding process and combine with the unsupervised learning method to predict the final welding quality. The collected state information includes the dynamic resistance curve, electrode displacement curve, and upset pressure curve. And state information of 308,274 welding joints was collected. Because of the high production frequency of chains, it is difficult to obtain the quality label by the destructive detection method. The commonly used supervised learning is no longer applicable. In view of this challenge, the unsupervised learning methods, Isolation Forest and Local Outlier Factor are proposed to predict the welding quality online for the first time. Another critical problem is that the unsupervised learning model lacks evaluation criteria. This paper presented the concept of separation degree between anomalous data set and normal data set to solve this problem. The classification performance of the model is judged by comparing the separation degree. According to the separation degree calculation results, the Isolation Forest's classification performance is better than the Local Outlier Factor. Finally, according to the classification results of Isolation Forest, the correlation between state information and welding quality is analyzed. It is found that the upset pressure and dynamic resistance can reflect the change of welding quality, but the correlation between electrode displacement and welding quality is not significant.