A universal source-free domain adaptation transfer learning for cross-domain fault diagnosis of rotating machinery

Existing cross-domain fault diagnostic algorithms often presume prior knowledge of the target domain’s failure patterns and train models on labeled data from the source domain. However, in practical engineering, new domains may include new failure modes, making it difficult to acquire labeled fault samples. This paper proposes a multi-classifier residual universal domain adaptation network (MUDAN) for cross-domain rotor fault diagnosis without making explicit assumptions about fault modes. The proposed model effectively extracts feature information from source vibration data and migrates model parameters to the target domain. In target domain training, an open-set classifier is trained using open-set entropy minimization to identify unknown faults, while a closed-set classifier is trained by weighting classification loss with nearest neighbor labels. Simultaneously, pseudo-labels are refined using nearest-neighbor knowledge to reduce pseudo-label noise. To lessen reliance on experimental data, the dynamics of the Jeffcott rotor with crack faults are modeled, and the system’s vibration response serves as a simulation dataset for migration tests. Three migration experiments are conducted using a simulated dataset and two test bench datasets. Results show that the MUDAN approach achieves higher diagnostic accuracy in cross-domain tasks compared to other methods, providing new reference value for cross-device rotor fault transfer diagnosis when target domain labels are unknown.