Research progress on deep learning-based arc welding molten pool detection methods

Molten pool acts as direct carrier of welding energy input and solidification shaping, and its geometric morphology and dynamic behavior play a decisive role in weld formation quality. In practical arc welding, intense arc radiation, spatter occlusion, and fluctuations in process parameters often result in image overexposure, contour loss, and abrupt morphological variations, which limits effectiveness of traditional methods based on threshold segmentation and manually engineered features under complex conditions. In recent years, deep learning has provided a data-driven approach for molten pool detection, enabling automatic learning of morphological features and evolution patterns from large-scale image data. Compared with conventional methods, deep learning models demonstrate superior robustness, generalization capability, and adaptability across varying working conditions, offering important technical support for stable molten pool information acquisition. To meet the demands of arc welding process sensing and quality control, this paper systematically reviews the relevant research progress and summarizes the application status of deep learning in key tasks, including molten pool geometric parameter extraction, welding state recognition, dynamic behavior prediction, defect detection, anomaly warning, and multimodal fusion. In addition, advantages and applicability of different models are analyzed, and training and optimization strategies oriented toward engineering applications are further summarized from four aspects of data, learning paradigm, objective function, and network architecture. The analysis indicates that the current deep learning-based molten pool detection is still constrained by scarcity of labeled data under diverse conditions, insufficient generalization capability, and limitations in real-time performance and computational resources. Looking forward, it is necessary to establish unified evaluation benchmarks, develop transferable model frameworks, and achieve deeper integration with welding process control, thereby promoting the advancement of intelligent sensing and control technologies in arc welding.