Abstract： To address the challenges of noise interference and feature extraction difficulty for rotating machinery acoustic signals under complex working conditions, variable speeds, and multiple fault types, this paper proposes a bearing fault diagnosis method based on multi-domain feature fusion and the XGBoost classifier. Utilizing multi-condition bearing audio signals from the BJTU-RAO Bogie dataset and the University of Ottawa motor fault dataset, time-domain, frequency-domain, and Mel-frequency cepstral coefficients (MFCC) features are extracted from the raw signals. A core feature set characterized by high contribution and low redundancy is selected by integrating XGBoost feature importance ranking, Shapley additive explanations value analysis, and feature correlation heatmaps, thereby reducing model complexity and enhancing interpretability. During the classification phase, the strong generalization capability of XGBoost is leveraged for fault identification under multiple working conditions. Experimental results demonstrate that the proposed method achieves weighted average accuracy and recall rates exceeding 96% across all conditions in the BJTU-RAO dataset. In the University of Ottawa dataset, it achieves accuracy over 98% under constant-speed conditions and maintains above 83% under variable-speed conditions. These findings validate the high accuracy, robustness, and generalization performance of the method in multi-condition and multi-fault scenarios. The innovative integration of multi-domain acoustic feature fusion with feature importance analysis enables efficient diagnosis of rotating machinery acoustic signals, offering significant potential for online monitoring in noisy industrial environments.
Key words : fault diagnosis of rotating machinery；audio signal processing；Mel-frequency cepstral coefficients；multi-domain feature fusion
