Naveen Loganathan, Joel Andersson, Vivek Patel
Weld quality in friction stir welding (FSW) is difficult to maintain because rapid changes in heat input and material flow can generate transient surface defects during welding. These defects cannot be detected in real time using conventional inspection approaches, resulting in increased inspection time and higher production cost. Real-time visual monitoring is therefore required to support stable and efficient production. This study investigates whether modern convolutional neural network (CNN) models can provide reliable, in-situ segmentation of FSW surface defects together with accurate geometric measurements during welding. A multi-class dataset of weld-surface video frames was created and annotated for flash, burrs, voids, galling, tool interaction, and weld-zone regions. Several CNN-based segmentation models were evaluated, and a lightweight architecture suitable for real-time deployment was selected and integrated with a high-dynamic-range industrial camera on the FSW setup. The system performs continuous segmentation and extracts weld width and defect area from live video at approximately 25 frames per second. Quantitative validation against optical-microscope measurements demonstrated near microscope-level accuracy, with sub-millimetre weld-width deviations and defect-area errors below 6 %. These results demonstrate that real-time visual segmentation can provide reliable weld-quality monitoring in FSW, support early defect detection, and establish a practical foundation for future automated process-control strategies in manufacturing environments. • Real-time weld defect segmentation achieved during active friction stir welding. • Pixel-level masks enable direct measurement of weld width and defect size. • System reaches microscope-level accuracy for weld geometry and defect areas. • High-speed vision pipeline operates at twenty five frames per second. • Study supports future closed-loop control and automated weld quality assurance.