Source-free domain adaptation framework for rotating machinery fault diagnosis by reliable self-learning and auxiliary contrastive learning

Domain adaptation techniques have been extensively studied and applied in rotating machinery fault diagnosis to improve diagnostic performance. However, most existing approaches require direct access to source domain samples, which are often unavailable in industrial applications due to the limitations of privacy protection, storage space, and transmission bandwidth. To address these challenges, this paper proposes a novel source-free domain adaptation framework for rotating machinery fault diagnosis, which can disentangle the domain adaptation from the need of source domain samples. First, a nearest neighbor knowledge aggregation strategy is designed to generate more reliable pseudo-labels. Then, the classification loss is re-weighted according to the reliability of pseudo-labels that are quantified through uncertainty estimation. Second, an auxiliary contrastive learning framework is applied in the target feature space to facilitate knowledge aggregation. In particular, a new negative pair exclusion scheme is introduced to recognize and exclude negative pairs composed of same-category samples, even in the existence of some noisy pseudo-labels. The cross-condition and cross-device experiments on three datasets are implemented to verify the feasibility and superiority of the proposed method.