Static and Dynamic Eccentricity Fault Detection and Quantification in an Inverter-Fed Reluctance Synchronous Machine Using Machine Learning

Inverter-driven Reluctance Synchronous Machines (RSMs) pose unique diagnostic challenges due to pronounced harmonic distortions and noisy operating conditions, complicating accurate eccentricity fault classification and severity estimation. To address these issues, this paper proposes an ensemble Hierarchical Convolutional Neural Network (HCNN) designed to directly analyze raw stator current waveforms combined with time-delay embeddings, eliminating the need for manual feature extraction. Experimental data from an inverter-driven RSM, subjected to static and dynamic eccentricity faults under various severities and loading conditions, were collected in a laboratory environment to train and validate the model. Load information was initially included during training to provide essential contextual details about the machine's operating state, as preliminary results showed significant improvements in fault classification and severity estimation with explicit load data. However, the proposed ensemble HCNN achieved consistently excellent classification accuracy and severity estimation performance even without explicit load information, suggesting reduced sensitivity to load variations. This highlights the practical value of the ensemble method for reliable fault monitoring in realistic inverter- fed operating environments.