Interpretable Intelligent Fault Diagnosis of Active Magnetic Bearing–Rotor Systems Based on Multidimensional Feature Fusions with Robustness Assessment

To address the limited interpretability of feature extraction in terms of fault mechanisms and the inefficient fusion of multisource information in fault diagnosis of active magnetic bearing–rotor systems, an interpretable intelligent diagnostic method based on multidimensional features and a graph convolutional network is proposed. Using multichannel displacement sensor signals acquired by the integrated sensing system, the method extracts single-channel features and multichannel spatiotemporal relational features from shaft orbits. A trajectory complexity feature set is designed to characterize the relationship between orbit morphology and fault mechanisms, while a spatiotemporal relational feature set is developed to explicitly describe the coupling relationships among multiple channels. Considering the different characteristics of the two feature sets, a dual-stage feature fusion diagnostic model is constructed. The graph convolutional network is employed to learn interaction patterns among channel-specific features, which are subsequently fused with the spatiotemporal relational features, followed by a classifier for fault identification. Experimental results obtained from five fault categories collected on a physical magnetic-levitation rotor test rig demonstrate that the proposed method achieves a diagnostic accuracy of 97.8% under class-imbalanced conditions. The designed feature sets provide good interpretability for the diagnostic results. Specifically, the diagnostic accuracy reaches 89.0% at a signal-to-noise ratio of 5 dB and 83.2% when the sampling rate is reduced from 2500 to 1250 Hz, corresponding to decreases of 8.8 and 14.6 percentage points, respectively. These separately evaluated perturbations provide quantitative evidence of robustness within the investigated noise and sampling-rate ranges at a fixed rotor speed. Moreover, the proposed method exhibits strong robustness to variations in sampling rate and noise interference, providing an effective solution for intelligent fault diagnosis of active magnetic bearing–rotor systems.