Prediction of Weld Penetration Using Dynamic Weld Pool Arc Images

Abstract

This work aims to study an improved method to predict weld penetration that is not directly observable during manufacturing but is critical for the integrity of the weld produced. Previous methods used signals acquired at a time, typically a single image or multiple images/signals from the process, to derive the penetration at that given time. Al-though deep learning appears to extract data well, analyses of weld pool physics, previous studies, and skilled weld operation all suggest that the dynamic welding phenomena give a more solid mechanism to assure the adequacy of the needed information. Therefore, this paper proposes to fuse the present weld pool arc image with two previous images, acquired 1⁄6 and 2⁄6 s earlier. The fused single image thus reflects the dynamic welding phenomena. Due to the extraordinary complexity, the weld penetration is correlated to the fused image through a convolutional neural network (CNN). Welding experiments have been conducted in a variety of welding conditions to synchronously collect the needed data pairs to train the CNN. Results show that this method improved the prediction accuracy from 92.7 to 94.2%. Due to the critical role of weld penetration and the negligible impact on system/implementation, this method represents major progress in the important field of weld penetration monitoring and is expected to provide more significant improvements during welding using pulsed current, where the process becomes highly dynamic.