Characteristic defect detection in laser micro-welding using distinct optical signal signatures and machine learning

The term ‘defect’ is widely used in the fields of laser welding and process monitoring for a broad range of deviations from what is considered a standard or good weld. However, whether a given anomalous weld is to be considered critically defective often depends on the intended application and on the specific kind of its abnormality. In this work we present a detailed investigation on characteristic defect recognition and classification during high-speed laser micro-welding of thin stainless-steel sheets with varying laser incidence angle. The study focuses on the correlation between optical emission signals, defect type, and physical origin under dynamic welding conditions. A pure optical sensing system was assembled and deployed to simultaneously monitor the laser reflection, near-infrared thermal emission and visible emission during the process. Controlled variations in process parameters such as welding power, focus variation, and inter-sheet gap were used to intentionally generate typical defects, including lack of weld penetration, hole, thickening formation and transitions between keyhole and conduction welding regimes. The relation between signal patterns, welding defects, and their physical origin was studied. The results reveal distinct and reproducible signal signatures associated with each of these process states; the keyhole-to-conduction transition is characterized qualitatively. Further machine learning experiments using a simple algorithm to classify chunks of the optical signal demonstrate that our easily deployable setup can exploit optical signals to recognize specific defects. The classifier is trained on four classes: good weld, lack of weld, hole and thickening. For the off-axis welds of the present dataset, recall values exceed 90% for all but one of the four classes. Finally, a single-run, exploratory case study suggests that knowledge can be partially transferable from intentionally generated defects to real, spontaneous defects. The proposed method therefore provides a practical basis for application-specific online defect assessment in laser micro-welding.