Application of unsupervised machine learning for resistance spot welding

The quality of resistance spot welding (RSW) is crucial in several sectors. For instance, one modern car contains up to 7000 spot welds. By examining electrode force and displacement signals, this work proposes an unsupervised machine learning method for the RSW process. The application of classical supervised algorithms requires labeled data which often are costly to gather. This solution solves the problem by potentially enabling automatic labeling. Therefore, the insights gained from the application of data clustering can be used then to improve the performance, robustness, efficiency, and interpretability of a subsequent supervised modeling strategy. An experimental campaign of several spot welds was done to collect force-displacement data that were recorded by high-frequency monitoring devices, allowing for the extraction of crucial characteristics. These features were subjected to clustering methods, specifically K-means++, Hierarchical Clustering, and DBSCAN with the aim of distinguishing between weld affected by expulsion and not.