A novel few-shot learning based feature relation model for robotic welding states monitoring

Amidst the evolution of contemporary welding technologies, real-time monitoring of the welding process has emerged as an indispensable element within intelligent welding systems. Prior research has demonstrated that welding process modeling methods based on deep neural networks exhibit high accuracy and robustness in predicting welding quality. Nevertheless, data dependency-related challenges, including the onerous task of data annotation and the paucity of model translatability, have constrained their utility in practical applications. To address these challenges, this paper proposes a feature relation model based on few-shot learning for welding state monitoring. First, we design a hybrid supervised training strategy suitable for welding monitoring models, leveraging both unlabeled data and commonly labeled data to enhance the representation ability and transferability of deep molten pool features. Thereafter, we developed a feature relational architecture leveraging attention mechanisms and Brownian distance covariance, enabling the recalibration of network feature distributions to align with specific tasks. This feature re-embedding improves the discriminative capability of the model, facilitating accurate identification of various welding states in few-shot scenarios. Experimental results indicate that our algorithm achieves a prediction accuracy of 96.5 % using only 15 samples per class, significantly reducing the data requirements for model training. Compared to traditional algorithms, this model's low dependency on sample size enhances its transferability and generalizes, thereby promoting the practical application of intelligent monitoring technologies in real-world welding environments.