Physics-Frequency-Guided Bidirectional Mamba Network for Cross-Domain Few-Shot Bearing Fault Diagnosis

Rolling bearing fault diagnosis is critical for the reliable operation of rotating machinery. However, vibration signals collected under varying sampling rates, devices, and operating conditions often exhibit domain shifts, and only limited labeled samples are available in the target domain. To address these challenges, this paper proposes a multi-scale physics-frequency bidirectional Mamba network (MS-PFMamba) for cross-sampling-rate, cross-device, and few-shot bearing fault diagnosis. The framework integrates a multi-scale convolutional front end to capture local impacts and periodic structures, a local context encoder and bidirectional Mamba module to model short-range correlations and long-range temporal dependencies, and a statistical feature branch to provide complementary global dynamic information. A physics-frequency selection strategy (PFSS) embeds bearing fault characteristic frequency priors into feature modulation, guiding the model toward fault-related frequency responses without relying on hard thresholding. Four diagnostic tasks are constructed using the CWRU and PU datasets to evaluate cross-sampling-rate transfer and cross-device few-shot adaptation. Experimental results demonstrate that MS-PFMamba achieves Accuracy/Macro-F1 of 0.9259/0.9221, 0.9608/0.9516, 0.9809/0.9809, and 0.8145/0.8095 on the four tasks, reported as the mean over three identical support-set draws. It ranks first among the compared methods on the three-class cross-device few-shot task and remains competitive on the remaining tasks, while requiring only 3.71M parameters. Comparative experiments, ablation studies, and feature visualizations further characterize the behavior of the proposed method under complex domain-shift conditions.