Simulation-Driven Diagnosis of Stator Phase-Current Imbalance in Wind Turbine Induction Generators Using Multi-Signal Features and Explainable Machine Learning

Abstract

Wind turbine reliability depends on timely identification of electromechanical faults, especially in generator-related subsystems under variable mechanical loads. This study presents a simulation-based, multi-signal, and physically interpretable diagnostic workflow for wind turbine electrical systems. It combines multiphysics simulation, FFT feature extraction, and explainable machine learning, emphasising the integration of existing methods rather than developing new AI models. A COMSOL Multiphysics (2D electromagnetic with 3D multibody dynamics) model of an induction machine simulated both healthy and imbalanced operating conditions with increasing stator phase-A current imbalance (parameter ε). The verified fault mechanism was incorporated into the model. From these simulations, a multisignal dataset was built using electromagnetic torque, rotor speed, electromagnetic force, and foundation force responses across 9 configurations, resulting in 54 samples (each with 100 features) classified into healthy, minor, and major imbalance groups. We tested Support Vector Machine, Multilayer Perceptron, and Random Forest algorithms. Repeated stratified cross-validation showed Random Forest performed best, with an average accuracy of 92.3% (±4.1%) and macro-F1 of 0.764 (±0.146). A leave-one-configuration-out test, where no data from the same configuration appear in both training and testing, produced more conservative results: 46.3% accuracy and 0.317 macro-F1, with no healthy-condition samples correctly classified, because only one independent healthy configuration was available. SHAP analysis identified foundation-force spectral energy in the 50-150 Hz range as the most important predictor, suggesting imbalance severity at the configuration level. Since foundation-force features are fixed within each configuration, this indicator should be considered a configuration-level marker rather than an individual sample marker. Overall, the sample-level results are promising, indicating that the multi-signal, physics-based feature set and interpretability are useful. However, the configuration-level results suggest that the current 9-configuration simulation setup is not yet sufficient for definitive diagnostic accuracy.