A Data-Driven Multiscale Convolutional Adaptive Network for Welding Robot Operating State Recognition

The operating states of welding robots are a critical component in the automotive body-in-white assembly process, directly affecting the product quality and production efficiency of the manufacturing line. Therefore, accurate recognition of the operating state patterns is of great importance. Traditional methods relying on sensor signal threshold changes and operator observation are subjective, dependent on human experience, and difficult to implement in intelligent and automated production processes. This study proposes a novel approach to recognize the operating states of welding robots without additional sensors, using a multiscale convolutional adaptive network (MSCAN). First, motion data collection was achieved by leveraging the welding robot’s installed sensors, providing estimates of angular acceleration, angular velocity, and angle of the XYZ-axes. To address the issue of class imbalance in the collected data, the synthetic minority over-sampling technique (SMOTE) algorithm was adopted to generate synthetic samples of the minority class. Then, a MSCAN was constructed, where an attention mechanism was embedded into the convolutional architecture, and a domain adaptation measure was further constructed to mitigate the data distribution discrepancy induced by different operation speeds. Finally, the proposed approach was evaluated and validated on a real welding robot dataset in the body-in-white assembly process. The results showed that the proposed method achieved an accuracy, precision, recall, and ${F}1$ -score of 99.25%, 99.25%, 99.25%, and 99.25%, respectively, outperforming other comparative models. This demonstrates that the proposed model can effectively recognize the operating states of welding robots, possessing significant theoretical and engineering application value in automotive body-in-white assembly.