In order to address common challenges in industrial environments, including limited labeled samples, variable operating conditions, and cross-machine fault diagnosis, this study proposes a novel vibration–acoustic multimodal contrastive learning framework. First, a multi-level constrained contrastive learning method is designed to jointly regularize vibration–acoustic features at the levels of sample distribution, feature statistics, and feature structure. This enhances multi modal consistency, feature discriminability, and information diversity, enabling the learning of robust joint representations. Second, a dynamic frequency domain collaborative enhancement module is introduced during the fine-tuning stage, which progressively integrates frequency domain features into vibration time series features while adaptively adjusting feature weights, thereby improving discriminative capability and representation stability. Finally, the proposed framework is validated on two motors with significant differences in structure and operating conditions, as well as on a series of variable condition and cross-machine fault diagnosis tasks. Experimental results demonstrate that, with only 10 labeled samples per class, the method achieves over 99.57% diagnostic accuracy in cross-machine tasks, highlighting its adaptability to distribution shifts and structural differences and confirming its robust generalization performance in cross-machine and cross condition fault diagnosis.
Yuan Zhuang, Deqiang He, Zhenzhen Jin et al.· Measurement science and tech...· 0 citations
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.