Developing a machine vision framework for real-time monitoring dynamic molten pool during aluminum alloy TIG welding

• Developing a modular machine vision framework that effectively balances accuracy, efficiency, and robustness for real-time molten pool monitoring in complex welding environments. • A large-scale, temporally annotated molten pool dataset is innovatively constructed using the Segment Anything Model 2, enabling efficient and low-cost image segmentation. • A lightweight CNN-based classification system combined with a U-Net segmentation network successfully enables high-precision state identification and feature extraction for adaptive weld monitoring. Aluminum alloy TIG welding serves as a critically important and widely adopted manufacturing process within the aerospace industry. Real-time visual monitoring in aluminum alloy TIG welding faces inherent challenges, including persistent visual interference, stringent robustness requirements, and competing accuracy-efficiency objectives. To overcome these limitations, this study proposes a machine vision framework based on a modular deep learning and image processing architecture. The framework integrates image preprocessing, pattern recognition, feature region segmentation, and feature calculation to enable accurate and robust extraction of molten pool characteristics under complex and unstable welding conditions. A high-quality segmentation dataset was efficiently constructed using an interactive annotation strategy, supporting the training of a precise and reliable convolutional network for feature measurement. To validate its performance, the framework was applied to double-pulsed variable polarity TIG welding of medium-thick aluminum alloys. Experimental results demonstrate a classification accuracy of 97.84%, segmentation accuracy of 95.21%, and dimensional measurement accuracy of 96.38%, with an overall processing speed of 31.11 FPS. The developed system exhibits high robustness, generalizability, and real-time responsiveness, making it well-suited for deployment in industrial aluminum alloy welding. Future work will focus on adapting the framework to support manual welding assistance and integration into intelligent welding systems across a broader range of materials and processes.