Abstract: Existing domain adaptation methods require access to raw data, increasing the risk of data leakage. Fuzzy domain adaptation generates fuzzy rules and relationships without accessing source data. Leveraging knowledge from the source, this research proposes a novel method for source-free domain rotary machinery fault diagnosis using fuzzy domain adaptation. Specifically, during source domain training, a joint training strategy simultaneously trains all source models by sharing model parameters. This approach significantly enhances the multi-domain performance of each private source model. Additionally, a model with a fused fuzzy strategy is constructed to derive fuzzy rules from source samples and anchor points corresponding to various categories. In the target domain adaptation phase, the proposed method retrains the source feature extractor using target data while keeping the fuzzy rules fixed. Subsequently, an anchor point alignment strategy is introduced to transfer diagnostic knowledge. Finally, a filtering strategy is designed to obtain high-confidence target pseudo-labels, combining a self-supervised learning strategy for the ultimate optimization of the target model. After verification by rolling bearing cross-domain diagnosis cases and gear cross-domain diagnosis cases, the framework is an efficient and secure solution for fault diagnosis.

Key words: rotary machinery fault diagnosis, data privacy leakage, source-free domain, fuzzy domain adaptation, self-supervised learning strategy