A machine learning-based classification method for SynRM faults

Synchronous reluctance motors (SynRMs) are increasingly critical in industrial and traction applications due to their high efficiency, magnet-free construction, and thermal robustness. However, fault diagnosis frameworks specifically tailored to SynRMs remain scarce, with existing literature predominantly focusing on induction or permanent-magnet machines, isolated fault scenarios, or simulation-only validations. This paper presents a comprehensive multi-fault diagnosis framework that addresses critical gaps in SynRM condition monitoring through rigorous experimental validation and reproducible methodology. Inter-turn short-circuit faults (5% severity, 12/240 turns) and inner-race bearing defects were experimentally induced on a 2.2 kW laboratory SynRM under varying load conditions (no-load, 50%, and 100% rated load), while static/dynamic eccentricity faults were modelled via ANSYS Maxwell FEA and statistically aligned with experimental distributions through noise injection and domain adaptation. Discrete Wavelet Transform (Daubechies 4, 5-level decomposition) was employed to extract time-frequency features from stator currents, yielding a 12-dimensional feature space capturing harmonic signatures from 0 Hz to 5 kHz. Eight machine-learning classifiers were evaluated under standardized protocols: stratified 80/20 train-test splitting (group-based to prevent data leakage), 5-fold cross-validation, and systematic hyperparameter optimization via Grid Search. Results demonstrate that ensemble tree-based methods significantly outperform linear models (McNemar's test, p < 0.05). Random Forest achieved 99.975% accuracy with 100% recall for inter-turn fault detection and 99.975% accuracy with 100% recall for bearing faults, prioritizing zero false negatives essential for protection relaying. For eccentricity classification, AdaBoost and XGBoost attained 100% accuracy with O(NlogN) training complexity, avoiding the prohibitive O(N³) cost of equivalent-performance SVMs. In high-cardinality multi-fault scenarios (16 classes), CatBoost achieved 99.96% accuracy and 99.625% recall, significantly exceeding Random Forest (p = 0.0044 ) through effective handling of class imbalance via ordered boosting. All optimal classifiers satisfied real-time constraints (inference latency: 16-28µs; memory footprint: 2.1-18.4 MB), meeting IEC 61,850 protection standards. This work establishes the first statistically validated, multi-fault benchmark for SynRMs, demonstrating that recall-optimized ensemble learning enables reliable detection of incipient faults while providing deterministic latency bounds for embedded deployment. The framework bridges the gap between laboratory diagnostic accuracy and industrial condition monitoring requirements.