| Abstract Scope |
Welding performance is influenced by current and voltage parameters. Understanding their impact on metal transfer is essential for optimizing power source design and welding output prediction. Recent advances in computer vision, driven by transformer-based foundation models, have moved beyond conventional frame-by-frame segmentation by enabling coherent video interpretation through object temporal evolution. EfficientTAM, a promptable segmentation architecture that integrates a lightweight vision transformer with efficient temporal memory, was adapted to analyze high-speed videos of GMAW using ER4043 aluminum wire. The model was fine-tuned in two phases to segment process regions, including detached droplets, molten consumable, and arc, and to improve droplet temporal continuity during metal transfer. Based on these predictions, a droplet-tracking system quantified transfer frequency, projected area, volume, and post-detachment kinematics in sequences of approximately 15,000 frames. The refined segmentation achieved an Intersection over Union of 0.862 for the droplet class, while transfer-frequency estimation reached an absolute percentage error of 1.25% +/- 1.39% relative to manual reference. Droplet volume was estimated using a weighted least-squares formulation based on temporal trajectory and an approximately constant-volume assumption during flight. Total transferred volume was obtained by summing reconstructed droplets and comparing it with a reference calculated from wire diameter and wire feed speed. Using the WLS estimator with 1-pixel erosion, the absolute percentage error in volume was 23.61% +/- 19.61%. These results show that temporally coherent segmentation combined with droplet-trajectory analysis can automate quantitative characterization of metal transfer in GMAW. The proposed pipeline provides a basis for relating electrical parameters to droplet frequency, size, and dynamics, supporting future analysis of waveform-design strategies. The main limitation remains volume estimation, although improved segmentation accuracy could strengthen volumetric reconstruction and enable more robust estimation of plasma drag force on droplets. |