Prediction of the Joint Cross-Section of Laser-Welded Hairpin Windings Based on 2D Image Data Using Convolutional Neural Networks

Due to the increasing electrification, the demand for high-quality, cost-effective electric motors continues to rise. The hairpin winding is one such optimization, improving the slot fill factor of the motor while simultaneously facilitating handling operations in production. Since this winding technology is accompanied by a high number of joints, special attention is paid to the contacting process, with laser welding being a promising choice. Due to the many influencing and disturbing factors, the current challenge is to better monitor the welding process and to stabilize the resulting quality. Up to now, quality characteristics such as the joint cross-section can only be measured with a costly computer tomograph or destructive random testing using microsections. Therefore, this paper proposes a novel, cost-effective approach to predicting the cross-section of laser-welded joints based on 2D image data in combination with deep learning techniques. In order to investigate what kind of images are most expressive, image data are acquired at different times from different perspectives. On the one hand, pre-process images are taken to predict whether the actual position and surface condition of the pins allow a sufficient weld seam. On the other hand, post-process images are analyzed to estimate the achieved cross-section and thus the joint's functionality. For the classification of the images, a pre-trained convolutional neural network is adapted using transfer learning. To reduce the error slip, pre-and post-process images are finally evaluated in combination, resulting in a recall of 100% in the present experiment. Of course, the proposed quality prediction is not only applicable to the contacting of hairpin windings but also to other laser welding applications.