SPECTRAL FEATURE-BASED NEURAL CLASSIFICATION FOR EFFICIENT BEARING AGING ASSESSMENT IN ELECTRIC MOTORS

This study proposes a novel methodology for classifying bearing aging stages in induction motors by leveraging a compact and effective set of spectral features. Two advanced neural network classifiers-a Pattern Recognition Neural Network (PRNN) trained with the Levenberg-Marquardt algorithm and a Feedforward Neural Network (FFNN) optimized with the Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm-were compared. Experimental results demonstrate the FFNN's superior accuracy and robustness in classifying eight distinct aging grades. The primary innovation of this study lies in the use of five key spectral features extracted from the critical 2-4 kHz frequency band. This feature set significantly reduces dimensionality while preserving the descriptive features needed to characterize the aging process, enabling efficient and precise diagnostics. By employing this approach, the methodology not only enhances computational efficiency but also facilitates seamless integration into real-world fault detection and maintenance systems. Beyond fault detection, this work establishes a foundation for accurately determining bearing aging stages, creating opportunities to estimate bearing lifespan more precisely. By providing actionable insights into the aging process, it enables proactive maintenance strategies that reduce downtime and operational costs while enhancing machinery reliability. Future applications may extend this methodology to broader predictive maintenance frameworks and condition assessment tasks across various industrial domains.