Image-Based Analysis and Estimation of Welding Features for Adaptive Control

Abstract

Real-time monitoring and control of Pulsed Gas Metal Arc Welding (GMAW-P) is important for maintaining process stability and weld quality. This thesis investigates whether image-derived welding features can be used to support data-driven estimation of the physical state of the welding process. High-speed imaging was used to quantify geometric process features, which were then synchronized with electrical measurements to create a dataset for training and evaluating soft-sensing models. Features related to arc behavior, droplet motion, weld pool position, and wire geometry were extracted from high-speed video. Two imaging configurations were evaluated: laser-lit and back-lit imaging. Laser illumination improved the visibility of fine structures but introduced strong reflections and artifacts that reduced feature extraction reliability. In contrast, back-lit imaging produced high-contrast silhouettes with cleaner object boundaries, making it more suitable for robust geometric feature extraction. A computer vision pipeline was used to extract geometric features, including wire edges, contact tip, wire tip, and weld pool position. Additionally, droplets and spatters were detected and tracked using blob detection and simple motion rules. These image-based features were then synchronized with electrical signals and process parameters to build a time-series dataset. Several statistical, machine learning, and deep learning models were evaluated for predicting image-derived features from electrical inputs, primarily voltage, current, and wire feed speed, along with additional engineered features. Overall, the results show that electrical signals contain information related to key geometric welding features, but the estimation accuracy remains limited. The models generally capture the cyclic patterns and overall trends in the signals, but struggle with amplitude variations between sequences. The results support the feasibility of soft sensing for GMAW-P monitoring, but not yet the accuracy and robustness necessary for reliable adaptive control.