A robust multi-class bearing fault diagnosis framework using envelope analysis, cepstrum prewhitening and machine learning
This paper presents a systematic framework for real-time fault-type identification in rolling element bearings using vibration signal analysis. The Case Western Reserve University bearing dataset is employed as the primary data source, comprising vibration signals recorded under different load conditions (0–3 HP). Initially, Envelope Analysis (EA) is applied to extract fault-related characteristic frequencies. While computationally efficient, EA successfully identifies fault features in 66.67% of the signals but shows limitations under noisy and spectrally smeared conditions. To address this, Cepstrum Prewhitening Analysis (CPA) is selectively applied to unresolved signals, achieving a 75% detection success in these cases and improving the overall detection rate to 91.67%. Thereafter, 36 time-domain, frequency-domain, and EA-CPA based features were extracted from segmented vibration signals. A sequential feature optimization strategy comprising variance threshold filtering, correlation analysis, Z-score normalization, ANOVA F-test feature ranking, and Recursive Feature Elimination reduced the feature set to the 10 most discriminative features. To eliminate sample- and group-level information leakage, a leakage-free Nested GroupKFold framework was developed, in which preprocessing, feature selection, and hyperparameter optimization was performed exclusively within the training folds using GridSearchCV. Six machine learning classifiers, namely Random Forest, XGBoost, LightGBM, Support Vector Machine, K-Nearest Neighbors, and Logistic Regression, were comparatively evaluated. XGBoost achieved the highest mean classification accuracy of 97.62%, while RF attained a comparable accuracy of 97.51% with lower fold-to-fold variation, indicating superior robustness and stability. Consequently, RF was selected for independent cross-condition validation, in which it was trained solely on the 0 HP operating condition and evaluated on the unseen 1 HP, 2 HP, and 3 HP datasets, demonstrating strong generalization across varying load conditions. Feature importance analysis further confirmed that the characteristic bearing defect frequencies (BPFO, BPFI, and BSF) are the dominant contributors to classification performance. The proposed framework provides an effective balance between diagnostic accuracy, robustness, interpretability, and computational efficiency, making it a promising solution for intelligent condition monitoring and predictive maintenance of rotating machinery.