Deep learning-based image analysis to study arc characteristics and metal transfer in GMAW

Abstract

This paper presents a novel application of deep learning to quantify arc region characteristics and metal transfer behavior in gas metal arc welding using high-speed videography and synchronized electrical signals. The motivation for this work is the need for scalable and objective methods to analyze features such as arc length, droplet size, and droplet transfer frequency. A U-Net model was modified to achieve multi-class segmentation and trained to process approximately 15,000 frames per video across globular and spray transfer modes using ER4043 aluminum wire. The model segments four distinct classes within the arc region: internal arc, external arc, molten consumable, and droplet. During training, the model achieved an average intersection over union of 0.912 with a low loss of \(3.2\times 10^{-3}\). The model achieved errors below 10% across all arc-length definitions and 6.6% for droplet transfer frequency when compared with manual measurements. The results show that the projected droplet area increases with voltage and decreases with current, consistent with observed frequency trends. This method provides an automated alternative to manual image analysis for the conditions studied and establishes a foundation for future welding arc research, waveform development, and adaptive control. Extension to other materials, transfer modes, and broader welding conditions will require further validation.