An Improved Hybrid Lightweight Approach for Bearing Fault Detection and Classification in Three-Phase Squirrel Cage Induction Motors

Early and reliable detection of bearing faults is essential for ensuring the safe and efficient operation of rotating electrical machines, especially under varying loads and non-stationary operating conditions. However, traditional diagnostic approaches struggle to maintain accuracy when signals are noisy, high-dimensional, or affected by multiple fault patterns. To address these issues, this work presents RNN-XBoostNet, a lightweight hybrid framework that combines the temporal-feature extraction capability of Recurrent Neural Networks (RNNs) with the robust classification strength of XGBoost. A new feature-selection strategy, CoLaR-FS (integrating correlation analysis, Lasso regularization, and recursive feature elimination), is introduced to reduce redundancy and retain only the most discriminative fault features. The proposed framework is evaluated using the widely known CWRU dataset and a newly developed induction-machine dataset containing ten fault categories, including six newly introduced real-world conditions. Experimental results show significant performance improvements: accuracy increased from 87.01% to 99.35% on the CWRU dataset and from 79.98% to 99.57% on the laboratory dataset. The combination of high accuracy, reduced complexity, and strong generalization demonstrates that RNN-XBoostNet, supported by CoLaR-FS, is a practical and effective solution for modern condition-based monitoring systems.