Fault Feature Extraction for Centrifugal Pump Impellers via EMD and Cyclic Bispectral Slicing

As a widely used device in the industrial field, the operational efficiency and stability of centrifugal pumps are critical to industrial production. Therefore, to prevent equipment damage and potential casualties during production, accurate monitoring and diagnosis of centrifugal pump impeller failures are essential. This study aims to develop a novel feature extraction method for analyzing fault vibration signals of centrifugal pump impellers. The extracted features are expected to provide higher accuracy and efficiency in fault detection. Accordingly, the cyclic bispectrum secondary slicing technique is employed to extract fault information from nonlinear periodic vibration signals. Based on this approach, the optimal slicing method is further investigated. Since the frequency characteristics of the original signal contain substantial clutter, empirical mode decomposition (EMD) is first applied to separate intrinsic mode functions (IMFs). The results demonstrate that the test set accuracy of the EMD–cyclic bispectrum secondary slicing fusion method in the SVM and XGBoost classification models reaches 85.71% and 92.14%, respectively, which is markedly higher than that of the single cyclic bispectrum secondary slicing method (76.43%/78.57%) and the EMD method alone (80.71%/85%). The proposed method exhibits a notable effect on extracting modulation frequency characteristics: after EMD preprocessing, clutter frequency amplitudes in the spectrogram are reduced by more than 60%, while the signal-to-noise ratio of the modulation frequency component increases by 2.3 times. In the anti-corner slice mode, the amplitude resolution of the modulation frequency component increases by 40%, and the carrier frequency along with its harmonic interference is effectively suppressed.