Penetration state recognition based on stereo vision in GMAW process by deep learning

Abstract

100 % penetration state with consistent and reliable weld is typically required in Gas metal arc welding (GMAW). Three-dimensional (3D) information of the weld pool surface is more sensitive to the welding penetration state. In order to take advantage of this critical information, a biprism passive stereo visual system based on a single camera is designed and established in this paper. A novel high-performance deep learning strategy is invested and discussed to recognize penetration state. Based on 3D information of the weld pool surface, the proposed strategy directly correlating the depth of the weld pool to penetration states. Deep learning stereo matching network based on residual structure and multiscale spatial aggregation is built and the depth of the weld pool is estimated based on the input left and right image pairs. A reliable dataset for optimizing stereo matching network to estimate the depth of the weld pool surface is established by using the globally optimized variational stereo matching algorithm. Output weld pool depth information is input into a high-performance classification network, which has been weighted with efficiency and precision. In this study, deep learning intelligibility approach is adopted to visualize the class activation regions (CAM), and it is found that CAM shows high activity around the concave regions on the weld pool surface. The optimized model comes into effect on the test set, recognition accuracy reached 99.6 %, which demonstrates sufficient performance.