Convolutional Neural Network Model for the Prediction of Back-Bead Occurrence in GMA Root Pass Welding of V-groove Butt Joint

Abstract

Gas metal arc (GMA) welding is widely used in the machinery industry. The quality of a welded joint is affected by the penetration of root pass welding in the V-groove joint. Automation using GMA welding is continuously required, and root pass welding automation is required to automate the entire welding process. In particular, the development of a prediction model that can ensure full penetration back-bead is required for the automation of root pass welding. In this study, a convolutional neural network (CNN) model was applied to predict the occurrence of back-bead in V-groove butt joint GMA root pass welding. The bead profile was measured using a laser vision sensor system and it was used as the input data for the prediction model, and the bead occurrence was used as the output data for the model. A total of 12,873 bead profiles were extracted and pre-processed through cutting, resizing, and thresholding. The CNN model consists of nine layers, and performs three convolution and two pooling operations. The accuracy of the prediction model was 99.5%, and through this study, it was demonstrated that the quality of root-pass welding can be controlled by using convolutional neural network and it can contribute to automation.