Defects detection of GMAW process based on convolutional neural network algorithm

Abstract

AbstractIt is significant to predict welding quality during gas metal arc welding process. The welding defect detection algorithm has been developed based on convolutional neural network (CNN). The sensing system and image processing algorithm for molten pools has been developed. It overcomes the interference caused by the arc light to obtain clear images of the molten pool's boundaries. The molten pools images are used to build up training set and test set for training and testing the CNN model. The model is designed to extract the visual features of molten pool images to predict the penetration state, the welding crater, and slags. Through optimizing the network parameters such as kernel-size, batch-size and learning rate, the prediction accuracy is higher than 95%. Moreover, the model enhances additional focus on the welding crater based on the welder experience. The mechanisms between molten pool characteristics and welding defects were analyzed based on the welder experience and the visual features of the model. It is found that the model judges the occurrence of burn-through with the black hole in the middle zone of the molten pool. When the surface pores are generated, the model exhibits a strong response to circular voids in the semi-solid region at the trailing end of the molten pool. The size and shape of fusion holes exhibit a strong correlation with the molten state. When the shape of the crater does not appear concave, it often signifies excessive penetration. It contributes to enhancing the algorithm's robustness during various welding scenarios.