Joint quality assessment by machine learning using characterized surface thermal radiation images of laser welding process

In recent years, laser welding has been widely used as an alternative to arc welding because of its high power and faster welding speed with local heating. In the welding process, particularly for e-mobility applications, the demand for quality control via all-point inspection is increasing. The laser process enables real-time observation of the welding area during processing, making all-point inspection possible. In this study, we investigated the possibility of predicting weld bead width from a set of images acquired using a CMOS camera with a band-pass filter. Machine learning was used for the prediction, and the prediction accuracy was determined using the Root Mean Squared Error (RMSE). The laser parameters, such as irradiation power and scan speed, and 13 feature values, such as the area, centroid, and rotation angle of the light emission acquired from the images and were used as training data. The RMSE of 0.16 mm was achieved for a bead width of 0.5-1.5 mm, confirming that the prediction was sufficiently accurate. Furthermore, we conducted an analysis with and without spectroscopic images to verify whether spectroscopic images are effective for the evaluation of laser welding using machine learning.