Deep-learning based analysis of metal-transfer images in GMAW process

Gas metal arc welding (GMAW) is a widely used metal-joining method in industrial manufacturing. In this method, the metal-transfer process plays an important role in determining welding quality. It is common in-dustrial practice to optimize GMAW waveforms and process parameters using high-speed videography. The assessment of videos is currently done by subjective human interpretation, which is time consuming and prone to errors. Computer vision techniques before deep learning have not been able to overcome the confounding aspects of the images. This work uses deep learning segmentation models to isolate droplets in the video footage of the GMAW metal-transfer process. Segmentation masks are used to compute the geometric and kinematic properties of the droplet to illustrate the dynamic characterization process. The proposed deep learning model is a fully convolutional network (FCN) approach. Several architectures are considered and compared here. The main result shows that the FCN-based approach can reliably segment droplets within an image with the benefit of processing thousands of images within minutes. The image features isolated in this work allow for the calculation of valuable process variables such as droplet trajectory, velocity, acceleration, and detachment frequency, which agree with those found in the literature.