In-process prediction of weld penetration depth using machine learning-based molten pool extraction technique in tungsten arc welding

Even though arc welding is widely utilized to join metallic parts with high reliability, the prediction and control of welding quality is challenging owing to difficulties in the prediction of weld penetration depth and the backside bead. In this study, an effective method for predicting weld penetration based on deep learning was proposed to control the welding quality in-process. The topside weld pool image was closely related to the welding quality and penetration depth and was also an accurate indicator of the state of welding over time. A prediction model for penetration depth using a topside weld pool image was constructed. Semantic segmentation based on a residual neural network was then performed on the acquired weld pool image. Consequently, an accurate weld pool shape was extracted. In addition, a penetration regression model was constructed based on a back-propagation neural network. Finally, the penetration depth (corresponding to the weld pool shape) was extracted via segmentation. The segmentation and regression models were combined to create a penetration prediction model. Considering a gas tungsten arc welding (GTAW) process, the predictions obtained from the proposed method were evaluated experimentally. In the validation process, the developed model quantitatively predicted the penetration depth in tungsten gas arc welding. The mean absolute error was 0.0596 mm with an R-2 value of 0.9974. The model developed in this study can be utilized to predict weld depth penetration and in-processing time using surface images of the weld pool.