Inline quality assurance of hairpin welding using a machine learning approach based on photodiode signals and images

This article discusses the challenges and solutions associated with quality assurance during laser welding of stators using hairpin technology. The high number of welds and the requirement for quality control throughout the process chain makes it essential to find suitable inspection methods. X-ray computed tomography (CT) is currently the most accurate method, but it is not feasible for inline quality control in series production. Alternative testing methods, such as image analysis and photodiode signals, are being investigated. In this study, a new end-to-end machine learning approach is developed that combines different sensor signals as input. The approach uses in-process radiation and pre- and post-process images to achieve a correlation on the cross-sectional area of the welding pin measured by CT. A coefficient of determination of 0.953 is achieved. Using the same sensor signals, the classification of the weld into six distinct categories is also investigated with a Support Vector Machine. The average accuracy of this classification is 98.63%. Both systems are suitable for integration into series production and show high potential for an inline quality control in hairpin production. Further improvements are expected with the use of new model architectures and high-resolution cameras.