Cross-domain fault diagnosis of RV reducer rolling bearings based on relative shaft synchronous resampling and adaptive gradient freezing transfer learning

As the core transmission component of industrial robots, the rotate vector (RV) reducer is required by the industry to be more and more reliable. Therefore, a set of reliable, accurate, and all-condition applicable condition monitoring systems is needed for RV reducer. Since a modern robot can perform complex maneuvers, its RV reducer operates under different operational conditions, such as different speeds and loads, etc. These variations lead to significant differences in the vibration signals of RV reducer during operations. For example, with the same number of sampling points, the interval between fault shocks can vary depending on the rotational speed. This brings a great challenge to realize cross-domain fault diagnosis. Although traditional deep learning methods have achieved good results in fault diagnosis, there are still challenges, such as high computational complexity and limited cross-domain adaptation capability, when facing cross-domain diagnosis tasks. To address this problem, this paper proposed a novel method based on relative shaft synchronous resampling, lightweight transformer module, and adaptive gradient freezing transfer learning (RS-LWformer). Firstly, the relative shaft synchronous resampling method is adopted to process the vibration signal, effectively eliminating the problem of inconsistency in the signal features caused by speed variations. Then, a lightweight transformer module is designed to reduce the computational complexity, while enhancing the capability of the model to extract local features and long-range dependency relationships from the signal. Finally, an adaptive gradient freezing transfer learning strategy is proposed to improve the generalization capability and robustness of the model under cross working conditions, by dynamically freezing the parameters that contribute less to the target task and fine-tuning the key layer parameters. Experimental results demonstrate that, in multiple cross working condition fault diagnosis experiments, the proposed method exhibits superior classification performance compared with other existing advanced methods.