Burn-through prediction and weld depth estimation by deep learning model monitoring the molten pool in gas metal arc welding with gap fluctuation

Abstract In a single bevel GMAW (gas metal arc welding) with gap fluctuation, a deep learning model was constructed using the monitoring image during the welding to predict the welding quality. We utilized Python and the library Keras and created a CNN (Convolutional neural network) model using the top surface image including the molten pool as an input. The classification model was used to predict the burn-through, and the regression model was used to estimate the penetration depth. As a result, the excessive penetration and burn-through could be predicted in advance and more than 95 % of estimated results of penetration depth were less 1 mm error for stepped and tapered sample shapes.