Research on fault feature extraction and diagnosis method of rolling bearing based on vibration signal

Abstract

This paper proposes a rolling bearing fault diagnosis approach based on vibration signal analysis. The collected vibration signals are first processed through denoising, normalization, and segmentation to improve data quality and provide reliable inputs for subsequent fault feature extraction and diagnosis. A multidomain feature extraction strategy is then applied to obtain time-domain, frequency-domain, and time-frequency-domain features, including RMS, kurtosis, spectral entropy, and wavelet energy features. A structured feature dataset is constructed through feature fusion, and PCA is used for dimensionality reduction. Finally, SVM and Random Forest classifiers are adopted for fault identification. Experiments on the Case Western Reserve University (CWRU) bearing dataset, including 12kHz vibration data under 0–3 hp loads and multiple fault sizes, demonstrate that the proposed method achieves over 95% classification accuracy across normal, inner race, outer race, and ball fault conditions, outperforming traditional methods by 5%–10%.