A novel source-free domain adaptation with high-confidence sample selection and feature disentanglement for machinery fault diagnosis

Abstract

Motivated by privacy concerns and the high cost of measured data transmission, source-free domain adaptation (SFDA) has attracted increasing attention for intelligent fault diagnosis. Instead of accessing raw measured source-domain data, SFDA transfers knowledge from pre-trained source models to target domains. However, existing methods often rely on pseudo-labels during the stage of adaptation, which may be noisy and consequently lead to negative optimization. Moreover, the domain-invariant features learned by the source model may still contain target-irrelevant information, making the adapted model susceptible to local optima under varying working conditions. To address these issues, this paper proposes a novel SFDA with high-confidence sample selection and feature disentanglement for machinery fault diagnosis. First, a source model is pre-trained using multi-source domain data to learn transferable domain-invariant representations, while an anchor generator is introduced to preserve class-level prior knowledge. During target-domain adaptation, a high-confidence sample selection strategy with the entropy measure is designed to identify reliable target samples for model updating. By dynamically selecting trustworthy samples, the proposed strategy effectively alleviates the adverse influence of noisy pseudo-labels. Furthermore, a feature disentanglement framework is developed to decompose mixed representations into fault features (FFs) and working-condition features, thereby further purifying domain-invariant fault representations. Finally, an auxiliary classifier consisting of C binary classifiers is introduced to characterize the distribution of FFs. Experimental results on the PU bearing dataset and the self-made SU wheelset bearing dataset demonstrate the effectiveness of the proposed method, achieving average diagnostic accuracies of 98.6% and 99.9%, respectively, which are superior to those of the best-performing compared method.