An FEM-Informed Statistical Feature Extraction and Comparative Machine Learning Framework for Dynamic Eccentricity Fault Diagnosis in Interior Permanent Magnet Synchronous Motors

Abstract

Interior Permanent Magnet Synchronous Motors (IPMSMs) are widely used in traction and industrial drive systems because they combine high efficiency, high power density, and excellent performance over a wide speed range. Rotor eccentricity, however, remains one of the most significant faults affecting their reliable operation; diagnosing it early is crucial for avoiding unexpected breakdowns. To achieve this, the analysis utilizes a simulation-based fault diagnosis framework that combines the Finite Element Method (FEM) with machine learning. A 550 W, 220 V IPMSM was modeled in ANSYS Maxwell to simulate dynamic eccentricity faults at three severity levels: 10%, 20%, and 40%. A fixed-length, non-overlapping window segmentation approach was used to pull statistical features from the stator current and radial air-gap flux density signals. These features were then fed into several supervised machine learning algorithms, evaluated using a consistent 5-fold cross-validation protocol across all investigated classifiers, with the Ensemble Bagged Trees classifier achieving validation accuracies of 93.12% for radial air-gap flux density and 87.86% for stator current. By integrating finite-element analysis, statistical feature extraction, and comparative machine learning, the proposed framework demonstrates the feasibility of simulation-based dynamic eccentricity severity classification in IPMSMs. The results indicate that Ensemble Bagged Trees provide the best classification performance among the evaluated classifiers, establishing a foundation for future experimental validation and real-time condition-monitoring applications.