Evaluation of Machine Learning for Quality Monitoring of Laser Welding Using the Example of the Contacting of Hairpin Windings

In a world of growing electrification, the demand for high-quality, well-optimized electric motors continues to rise. The hairpin winding is one such optimization, improving the slot-fill ratio and handling during production. As this winding technology leads to a high amount of contact points, special attention is drawn to contacting processes, with laser welding being one promising choice. The challenge now is to make the process more stable by means of advanced methods for quality monitoring. Therefore, this paper proposes a novel, cost-efficient quality monitoring system for the laser welding process using a machine learning architecture. The investigated data sources are machine parameters as well as visual information acquired by a CCD camera. Firstly, the usage of machine parameters to predict weld defects and the overall quality of a weld seam before contacting is investigated. In the case of hairpin windings, not only the mechanical but also the electrical properties of each contact point contribute to the overall quality. Secondly, it is illustrated that convolutional neural networks are well suited to analyze image data. Thereby, different network architectures for directly assessing the weld quality as well as for classifying visible weld defects by their severity in a post-process manner are presented. Thirdly, these results are compared to a more explainable two-stage approach which detects weld defects in a first step and uses this information for weld quality prediction in a second step. Finally, these applications are combined into a quality monitoring system consisting of a pre-process plausibility test as well as a post-process quality assessment and defect classification. The proposed system architecture is not only applicable to the contacting of hairpin windings but also to other applications of laser welding.