An efficient welding state monitoring model for robotic welding based on ensemble learning and generative adversarial knowledge distillation

•A multi-model ensemble decision framework is created using DS voting theory.•A generative adversarial knowledge distillation algorithm is proposed to compress multiple models into one through knowledge transfer.•The model trained with proposed algorithm maintains predictive accuracy while ensuring computational efficiency for real-time welding. Autonomous robotic welding is a crucial aspect of modern welding technology. However, due to the complexities inherent in the welding process, achieving real-time welding state monitoring remains a significant challenge in robotic welding research. Among various approaches, the utilization of deep neural networks for welding state prediction has emerged as a potent means to disentangle the complexities of the welding process. To meet the real-time and precision requirements for online welding quality monitoring, this study has devised a welding state prediction model grounded in ensemble learning and generative adversarial knowledge distillation. Primarily, a knowledge distillation approach based on generative adversarial has been proposed. It uses a discriminative network to distinguish features from teacher and student networks, facilitating effective knowledge transfer between them. Subsequently, leveraging the DS decision theory, a multi-model ensemble decision algorithm has been formulated, enhancing the predictive accuracy of the system without altering the complexity of the models. Through this mode, the knowledge from multiple teacher networks can be transposed into a single student network, mitigating the computational resource dependency issue in traditional multi-model ensemble systems. Experimental results demonstrate the model trained through proposed algorithm can uphold predictive accuracy while ensuring computational efficiency. The effectiveness of this approach is validated in practical welding scenarios, suggesting its potential to enable autonomous monitoring in robotic welding processes.