Source-free domain adaptation framework for rotating machinery fault diagnosis by Progressive Pseudo-Labeling and Prototype-Boundary Calibration

Abstract

Domain adaptation (DA) techniques have made significant advancements in the field of mechanical fault diagnosis. However, existing methods typically assume that source domain data is accessible during the DA phase. In real-world engineering scenarios, this assumption is often impractical due to limitations in data privacy, storage overheads, and transmission bandwidth. To address this issue, a novel source-free DA framework is proposed for rotating machinery fault diagnosis. First, the Progressive Pseudo-Labeling strategy is introduced, which gradually builds a reliable pseudo-label memory bank and dynamically updates it with historical information. This strategy effectively suppresses incorrect pseudo-labels. Then the Boundary Adversarial Calibration module is designed to incorporate low-confidence boundary samples into model training, enhancing feature discriminability. Furthermore, the Targeted Prototype Alignment constraint is introduced to promote intraclass compactness and interclass separation by pulling target samples toward their corresponding class prototypes while pushing them away from those of other classes. Extensive source-free cross-domain diagnostic experiments conducted on two rotating machinery datasets yielded average accuracies of 99.37 and 99.04%, respectively. The proposed framework achieves strong average performance and remains competitive across all evaluated transfer tasks, validating its feasibility and effectiveness in practical diagnostic scenarios.