An Improved Rolling Bearing Fault Diagnosis Method based on Enhanced Feature Extraction and Regularized SVM

Abstract

Rolling bearing fault diagnosis is critical for ensuring the safety and reliability of rotating machinery. This paper proposes an improved fault diagnosis method that addresses the overfitting problem commonly observed in traditional approaches. The method employs a 9-dimensional enhanced feature vector combining time-domain statistical features (RMS, kurtosis, crest factor, skewness) with envelope spectrum features (BPFO, BPFI, BSF amplitudes, band energy ratio, spectral entropy). A data augmentation strategy using Gaussian noise perturbation is introduced to increase training sample diversity. The RBF-SVM classifier is regularized with optimized parameters (C=1, γ=0.1) to prevent overfitting. Validated on the CWRU bearing dataset with multiple damage severities, the proposed method achieves 98.5% accuracy under strict temporal-split validation with a train-test gap of only 1.5%, compared to 87.0% for the baseline model. The results demonstrate that enhanced features combined with proper regularization effectively resolve overfitting while maintaining high diagnostic performance.