Robotic grinding systems are pivotal in precision manufacturing for their flexibility and cost-effectiveness. However, high-precision online monitoring of robotic grinding state remains challenging due to complex machining mechanisms and unstable environments. Existing deep learning methods, reliant on the assumption of identically distributed data, suffer from poor generalization under domain shifts, such as feature shifts caused by varying processing parameters (cross-condition) or differences between machines (cross-device). To address this issue, this paper proposes a Multi-Scale Attention-Fused Domain Adaptation Network (MADAN) for robust state monitoring in robotic grinding, where an integrated robust transfer framework is constructed by an enhanced signal augmentation module, a multi-scale attention-aware fusion module and an adversarial domain adaptation strategy. The robustness of the monitoring model is greatly improved in processing scenarios involving noise and disturbances with the enhanced signal augmentation module. Based on a Cross-Scale Self-Attention Fusion Network mechanism, the multi-scale attention-aware fusion module adaptively aligns and integrates deep features of vibration signals across different receptive fields. Furthermore, the adversarial domain adaptation strategy is implemented to reduce significant distribution discrepancies between different domains. Two sets of experiments on a robotic grinding platform are carried out including cross-condition and cross-device state monitoring, respectively. Experimental results demonstrate that MADAN achieves superior performance in these two state monitoring tasks. Specifically, the model attains an average accuracy of 96.83% in cross-condition scenarios and 95.68% in cross-device scenarios, significantly outperforming State-of-the-Art transfer learning methods in both detection precision and generalization stability.
Xiong et al. (Sun,) studied this question.