A novel source-free domain adaptation network for intelligent diagnosis of bearings under unknown faults

Abstract Domain adaptation has significantly advanced the field of cross-domain intelligent fault diagnosis of bearings. However, there are still some issues that hinder the progress of cross-domain fault diagnosis as follows. The dataset may not be shared due to privacy issues in some industrial scenarios, and unknown faults included in the dataset may lead to low diagnosis accuracy. To address these issues, this paper proposes a source-free domain adaptation network considering unknown faults (SDANU) for intelligent diagnosis of bearings. To address privacy issues, the source domain model is adapted to the target domain by optimizing an information maximization loss. This aligns feature representations while fixing the source classifier parameters, ensuring compatibility without exposing raw data. Then, a Bayesian inference-based module calculates the posterior distributions of both domains through Laplace approximation. By comparing the Kullback-Leibler (KL) divergence between source and target distributions, unknown faults in the target domain are identified and isolated, preventing misclassification. The proposed method is empirically validated through two fault diagnosis cases. The results show that the method effectively maintains data privacy while enabling model training without revealing sensitive information, and accurately diagnoses the health status of bearings by detecting the unknown faults.