Abstract: Aiming at the problems of high similarity between melt pool information and background, much noise, poor real-time prediction and low recognition accuracy, a GTAW welding fusion state recognition method based on improved CeiT network model is proposed. First, the Image-to-Tokens module is lightened and improved by MobileNetV3 to enhance the real-time performance of prediction; second, the DGCA module is designed to improve the LeFF module to enhance the remote dependencies among features and enrich the categorical information contained in the class labels; and lastly, the fusion of the underlying features and the high-level semantic features in the LeFF module improves the model's ability to represent the features of the melt pool. Simulation experiments show that the proposed model obtains higher accuracy and faster detection speed compared with MobileNetV3, ResNet50, ShuffleNetV2, DeiT, and CeiT models.