Deep Learning to Characterize the Morphology of the Arc and Metal Transfer in GMAW

The heat input is a key component of a welding procedure, which is dependent on fall voltages and amperage settings that directly impact the mode of metal transfer in GMAW. The mode of metal transfer achieved significantly influences the characteristics and quality of the weld bead deposited.

Traditionally, high-speed videography for analyzing metal transfer characteristics (e.g. droplet formation and arc length measurements) has depended on human interpretation for both qualitative and quantitative assessments. Although effective, human interpretations are time consuming and prone to error. Thus, making it desirable to interpret these high-speed videos through automated means, which allow for a larger number of frames to be interpreted in a fraction of the time. This presentation focuses on the quantification of key features in the GMAW arc region using a deep learning architecture called U-Net. Here, a series of welds using ER4043 were analyzed with synchronized data (voltage, amperage, and high-speed videography) to obtain results for key features including droplet size, droplet frequency, arc length, and respective relationships to voltage and amperage. Performance results of the deep learning algorithm will be shared along with a sample comparison to human interpretations of the same sample set.

J. Eduardo Alvarez Rocha, MSc Student, University of Alberta
Presented at: CanWeld Conference 2024