Predicting the Quality of High-power Connector Joints with Different Machine Learning Methods

State-of-the-art manufacturing processes used in the electric drive production show a high degree of automation and provide a large amount of process data. Often these data remain unused even though they contain potentially valuable process information. The efficient processing and evaluation of these data bears enormous potential for improving the electric drives production, for example with regard to contacting processes. Innovative machine learning (ML) methods already proved to be a powerful tool for big data set evaluation and continuously enter the manufacturing domain. However, the comprehensive and feasible ML application in manufacturing is hindered by the large effort necessary for adequate data preparation. This work lays the foundation for ML application in ultrasonic metal welding and related contacting techniques, which play an important role in the electric drives production. Therefore, on the one hand side, a data pipeline is developed which covers necessary steps of data preparation. On the other hand side, two data sets generated with a validated US metal welding process model are processed with the data pipeline and quality prediction is performed with three different regression methods, which include classical linear regression as well as advanced ML methods as a neural network. For quality prediction, the mean absolute percentage error reaches values as low as 6.9 %.