| Abstract Scope |
Real-time monitoring of welding processes remains difficult in practice due to the need for large amounts of fully labeled data, varying lighting and surface conditions, and the complexity of extracting useful geometric information from optical images. In this work, we develop a self-supervised learning based weld monitoring system that works with only 100 labeled images. The system uses a self-supervised model DINOv3 developed by Meta to segment the melt pool, wire, and torch from ordinary grayscale images. Then system can automatically extract key geometric features, such as melt pool contour, wire tip offset, and other. Based on these features together with welding parameters including current, speed, and voltage, the system predicts reinforcement width and classifies weld quality using two lightweight neural networks. The complete pipeline runs in real time and can be integrated with LabVIEW for live production use. Our results show that the system provides accurate geometric measurements and reliable quality predictions, even with limited labeled training data. By turning optical images into actionable information with minimal labeling effort, this work offers a feasible solution for low-cost, real-time weld monitoring that can provide live suggestions in adjusting welding parameters and support closed-loop control in automated manufacturing. |