Rotating machinery is a key component in modern industry, and its operating condition directly affects equipment safety and production reliability. However, discrepancies among different machines cause source–target distribution shifts, while fault annotation for target machines is costly, limiting the performance of deep learning-based diagnosis under cross-machine scenarios with limited labels. To address these issues, this paper proposes a multi-scale time–frequency semantic consistency model based on self-supervised transferable representation learning, termed MTFSC. First, augmented waveform views and multi-scale frequency-domain views are constructed from unlabeled source-domain vibration signals for self-supervised pre-training without source labels. Then, a time-domain impulse-aware feature extractor and a time–frequency decoupled spectral feature extractor are designed to enhance local impulsive responses and emphasize fault-sensitive time–frequency patterns. Furthermore, a semantic-aware soft contrastive loss is developed to mine potential semantic neighbors from multi-scale frequency-domain structural similarity, reducing false-negative effects in conventional hard-label contrastive learning. Finally, the pre-trained time-domain extractor is transferred to the target machine and fine-tuned with limited labeled samples. Experimental results show that MTFSC outperforms comparison methods under different labeled sample ratios and achieves an average accuracy of 97.5% across four cross-machine diagnostic tasks.

Rotating machinery plays a critical role in transmission systems, while the scarcity of fault samples and labeled data limits the performance of existing diagnostic methods under few-shot conditions. This paper proposes a few-shot fault diagnosis method for rotating machinery based on a time-frequency dual-stream Mamba architecture and joint metric learning. Raw vibration signals and multi-scale short-time Fourier transform time-frequency views are constructed to characterize transient impacts and frequency-band energy distributions from complementary perspectives. A dual-stream feature extraction network is developed, where the time-domain branch captures impulsive fault patterns, and the time-frequency spatial Mamba and channel attention branches extract spatial dependencies and fault-sensitive responses. Moreover, a hierarchical selective fusion mechanism is introduced to adaptively integrate complementary features across different domains. Finally, a covariance-based joint metric learning module is designed to model class distributions using second-order statistics of support samples and classify query samples through local similarity aggregation. Experimental results show that the proposed method consistently outperforms comparison methods under different rotational speeds, training sample scales, and 8-way 1-shot/5-shot tasks, validating its effectiveness and robustness for few-shot rotating machinery fault diagnosis.