Monitoring of distance between laser beam and arc in laser-arc hybrid welding based on deep learning

Abstract

The laser beam to arc distance (DLA) in laser-arc hybrid welding will significantly affect the heat source’s thermal flow distribution and the weld joint qualities. Therefore, the development of real-time monitoring methods for DLA is highly urgent. This study proposes classification and regression methods for DLA monitoring in laser-arc hybrid welding based on top-surface molten pool images and the convolution neural network (CNN) model. Molten pool images obtained under six fixed DLA values are used to train and test the classification CNN model. The classification CNN model can discriminate six kinds of DLA conditions with an accuracy of 99.83 %. Further, workpieces with different thicknesses are welded with different change strategies of DLA to capture molten pool images. Four training and testing cases are designed to verify the accuracy and generalization ability of the regression CNN model. The mean absolute errors of the estimation results in the four training and testing cases are less than 0.3 mm. The regression CNN model can predict the DLA value within 2 ms, proving that the proposed model exhibits tremendous applicability for monitoring DLA in real-time.