Deep-learning-based real-time monitoring of full-penetration laser keyhole welding by using the synchronized coaxial observation method

Abstract

In laser keyhole welding, the keyhole exhibits inherently unstable behavior, and the laser beam absorptance inside a keyhole varies rapidly. In this study, a real-time full-penetration laser keyhole welding monitoring system was established using a synchronized high-speed coaxial observation method in combination with deep learning models. Considering the images pertaining to the simultaneous observation of the top and bottom surfaces of the welding process, an object detection model (YOLOv4) was used to automatically measure the keyhole top and bottom apertures. The optimized model exhibited mean intersection over union accuracies of 98.23% (top) and 95.6% (bottom) and had a prediction speed of 156 fps. For 2-D images involving the measured keyhole top and bottom apertures, ResNet-34, which is a representative image classification AI model, was employed with an image regressor to predict the laser beam absorptance inside a keyhole; the model achieved an R-2 accuracy of 99.76% with a prediction time of 1.66 s for 740 keyhole geometries. The keyhole variations and absorptance fluctuated considerably during the welding progress, and the absorptance increased as the number of opened keyhole bottom apertures decreased, area of the keyhole bottom aperture reduced, and tilting angle increased. When a defect was generated, the laser absorptance declined rapidly as the keyhole bottom aperture size increased. Moreover, the width of the melt pool dramatically reduced.