Prediction of GTAW penetration state based on enhanced knowledge of dynamic morphology molten pool

To achieve online monitoring of welding penetration status, ensure weld quality, and promote the development of robotic intelligent technology, a knowledge-enhanced prediction method for Gas Tungsten Arc Welding (GTAW) penetration status based on molten pool dynamic deformation is proposed. High-speed, high-dynamic-range industrial cameras are employed to capture molten pool images, and the DeepLabv3+ semantic segmentation model is utilized for dynamic segmentation of the molten pool to obtain precise molten pool regions. On this basis, multi-frame molten pool contour image fusion is performed to describe the dynamic deformation of the molten pool during the welding process. The fused molten pool contour images and original molten pool images are combined and input into a CNN to learn pixel-level changes in the molten pool contours at the same location, enabling the CNN to predict penetration status. Experimental results demonstrate that the CNN enhanced with molten pool dynamic deformation knowledge can accurately identify three typical weld states: partial penetration, adequate penetration, and excessive penetration, achieving a classification accuracy of 97.1% with a single-frame prediction time of 0.86 ms. Compared to deep learning methods without integration of expert knowledge on molten pool dynamic deformation features, this method exhibits higher robustness and accuracy in scenarios with limited sample data.