Abstract
Motor eccentricity faults, stemming from the misalignment of the rotor's center and pivot point, lead to significant vibrations and noise, compromising motor reliability. This study emphasizes the need for an efficient diagnostic system to enable early detection and correction of these faults. Our research proposes a novel Eccentricity Fault Diagnosis Network (E-FDNet), designed for integration into a Motor Eccentricity Fault Diagnosis System (MEFDS), utilizing neural networks for detection. Evaluation tests reveal that a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture is ideal as the internal neural network within the E-FDNet. Key contributions of this research include (1) E-FDNet that stabilizes transition predictions among SEF/DEF/MEF; (2) a steady-state characteristic normalization (SSCN) improving feature consistency under dynamic responses; (3) an integrated physics-FEM-experiment pipeline for controlled analysis and validation; (4) approximately 98.86% accuracy/F1 outperforming classical and deep baselines; and (5) a non-invasive, current-only sensing design suited for deployment.