A Clustering-Guided Source-Free Domain Transfer Learning Diagnostic Method for Rotating Machinery

When transfer learning is applied to fault diagnosis, firstly, it will bring serious negative transfer problem due to the distribution difference between source and target domain data, secondly, the excessive reliance on source data will bring certain privacy and security problems. A clustering guided source-free domain transfer diagnosis method is proposed in this paper. First, to address the problem that the samples in the source and target domains of mechanical devices do not conform to a normal distribution, a Gaussian mixture model (GMM) is introduced to divide the data and generate sample weight scores as a data-level judgment. Secondly, the predictive probability of the model is regarded as the model score as the model level judgment, and a hybrid weighting strategy is used to combine the sample weights with the model weights. The overall effective fusion of source domain classification prediction and sample Gaussian mixture model distribution prediction. In addition, classification target diversity loss was introduced to improve model accuracy. Finally, the effectiveness of the proposed method is verified using a cross-device migration diagnostic test. The experimental results show that the proposed method can fully exploit the fault feature information and improve the diagnostic accuracy under passive and unsupervised cross-domain conditions, which has higher diagnostic accuracy than the current popular methods.