Penetration State Recognition for GMAW Process Using Stereo Vision Based on Deep Learning

Abstract

As a major arc welding process, pulsed gas metal arc welding (GMAW-P) is extensively applied in the production of various critical components due to its benefits such as high weld strength and ease of automation. The geometry of the weld pool contains a lot of welding quality characteristics in the process of welding, which is an important basis for judging the penetration states of the weld pool. For such key features, we develop a single camera stereo vision system based on a biprism and propose a deep learning-based penetration states recognition strategy. Using a stereo matching network based on attention concatenation volume, filter redundant information in concatenation volumes by generating attention weights to predict the accurate disparity map, and compare experimental results to evaluate its validity in weld pool disparity estimation. Secondly, to train the penetration states recognition model, a depth-wise separable convolutions neural network is built. This model directly maps the weld pool penetration relationship to disparity without any preprocessing and extracts deeper features such as pool contours and depth of fusion, then accurately identify the penetration states. According to the results of the experiments, based on disparity characteristics, penetration states classification accuracy is 99.83%.