Open Collections

Mobaraki, Mobina

University of British Columbia

Gas Metal Arc Welding (GMAW) is widely used in manufacturing, but unstable molten metal can lead to defective welds, increasing costs due to rework. Defects often arise from improper torch positioning or process inconsistencies. Real-time monitoring of molten metal behavior can help prevent defects, but this is challenging due to fluctuations in the electric arc caused by assembly errors, thermal deformations, and varying welding parameters. Fully automated welding systems can eliminate human error and labor shortages but are expensive and impractical for small-batch or highly variable applications like pipe welding. Semi-automated monitoring, where human operators collaborate with sensor-equipped robots and Artificial Intelligence (AI) algorithms, offers a more adaptable solution. However, challenges remain, particularly for fillet welds due to non-perpendicular camera angles and light reflections. This thesis proposes AI-driven methods to enhance semi-automated real-time monitoring in fillet welds. We use an industrial collaborative welding robot with a Complementary Metal-Oxide Semiconductor (CMOS) camera and a microphone to collect real-time image and sound data. Our approach includes: • Vision-Based Seam Tracking: We develop a deep learning classification model to detect tacks, where the seam is not visible, and a keypoint detection model to locate the seam in non-tack images. Optimizing image input size and leveraging temporal information improved seam tracking accuracy to over 80 % for errors below 0.3 mm, aligning with professional welder standards. • Defect Detection: We introduce uni-modal and multi-modal deep learning models to identify common welding defects, including Lack of Penetration, Lack of Fusion, Porosity, Undercut, and Cold Lap. Sound-based models were particularly effective for detecting Porosity, while image-based models performed better for other defects. Combining image and sound data improved defect detection by up to 55.17 % over uni-modal models, and attention-based temporal techniques further enhanced accuracy by 12.5 %. • explainable AI (XAI): We applied XAI techniques to highlight the most relevant areas in welding images for defect detection, increasing model transparency and trustworthiness. Our findings demonstrate that AI-enhanced semi-automated monitoring can improve welding precision, reduce defects, and lower manufacturing costs, making real-time defect detection more feasible for industrial applications. Future work could integrate our AI models into robotic controllers for autonomous corrective actions.

Text

2025-04-07

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/90598

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/