Bearing Fault Diagnosis Based on FR-ABGWO-XGBoost

Abstract

Weak fault modulation components in motor stator current signals are easily masked by dominant low-frequency components, and redundant multi-domain features can further degrade bearing fault classification. To address these problems, this paper proposes a feature selection method based on an adaptive binary grey wolf optimizer (ABGWO). Feature reconstruction (FR) is first used to suppress the dominant low-frequency component of the current signal, and time-domain, frequency-domain, spectral-peak and wavelet packet energy features are extracted from the residual signal. On this basis, a nonlinear convergence factor, an adaptive population update mechanism and a feature-number-penalized objective function are introduced into BGWO to improve the fixed search schedule, susceptibility to local convergence and redundant feature retention. The selected features are then fed into XGBoost for bearing state identification. Under the baseline operating condition, the complete ABGWO selects six features from the 36-dimensional candidate feature set, achieving a feature compression rate of 83.33%, an average test accuracy of 99.63% and a Macro-F1 of 99.62%. On the self-measured data set, the proposed method achieves an average accuracy above 99% under all four operating conditions C1-C4. Further validation on the Paderborn University public data set with real bearing damage shows an average accuracy above 98.7% under different damage severities and operating conditions. The results indicate that the proposed method effectively reduces feature redundancy while maintaining high diagnostic performance across different data sources and operating conditions.