Prediction of Weld Penetration Using Dynamic Weld Pool Arc Images

Welding has been automated/robotized greatly. However, in typical automated/robotic welding applications, the welding parameters are preset and not adjusted adaptively to overcome the effect from unpredicted disturbances. This imperfection cannot meet the increasing requirements from the welding/manufacturing industry on quality, efficiency, and flexibility. Combining information sensing/processing with traditional welding manufacturing techniques has been a major directive to revolutionize the welding industry (Ref. 1). In practical welding, the weld penetration, as measured by the back-side weld bead width, is a critical factor determining the integrity of the weld produced. However, the back-side bead width is difficult to monitor directly during manufacturing because it occurs underneath the surface of the workpiece being processed. Therefore, predicting the back-side bead width using conveniently sensible information from the welding process becomes a fundamental issue in intelligent welding. Many studies have been done to predict the weld penetration using different characteristic information from the welding process. They typically 1) sense observable phenomena from the welding process using, or based on, different sensors/phenomena such as infrared, pool oscillation, laser ultrasonic, and active vision methods (Refs. 2–5); 2) define and extract characteristic features from sensed phenomena; and 3) build a model to correlate the extracted characteristic features to the penetration state (Refs. 6, 7). However, the characteristic features are proposed subjectively based on the individual’s understanding of the physics, thus lacking a systematic way to ensure success in leading to a good model. Iteration is often needed such that the development efficiency is low. To address this general challenge, researchers recently started to apply deep-learning-based methods to extract the information automatically. Therefore, the major remaining challenge is reduced to acquiring adequate information from the welding process. Skilled welders can judge the weld penetration per their observed welding phenomena during the process. The welding community believes that images from the observable welding scene, including the 3D weld pool surface, contain sufficient information to predict the weld penetration (Ref. 8). While earlier efforts followed the aforementioned procedure to first propose characteristic features, the deep learning method has recently been applied, with a concentration on using convolutional neural networks (CNNs), to directly map images to the penetration (Refs. 9–14). The training for the parameters, including the convolutional kernels and Prediction of Weld Penetration Using Dynamic Weld Pool Arc Images