Enhancing the reliability of self-excited induction generators used in wind energy microgrids through a hybrid FEM–CNN-based fault diagnosis framework

Reliable condition monitoring of self-excited induction generators (SEIGs) is essential for ensuring the stability of isolated wind-powered microgrids (MGs), where progressive degradation of excitation capacitors remains one of the most frequent and least detectable faults. This research proposes a high-fidelity hybrid diagnostic framework that incorporates an advanced two-dimensional time stepping finite-element method (TSFEM) model of a 4.087 MW SEIG with machine-learning classifiers to detect healthy, disturbed, and progressively faulted states with high accuracy. The TSFEM model captures electromagnetic transients under realistic operating scenarios—including balanced/unbalanced loading, open-phase events, and staged internal capacitor-element losses—while generating synchronized voltage, current, and torque waveforms. Time-domain, spectral, and continuous-wavelet scalogram features are extracted to train multiple classifiers: Random forest (RF), k-nearest neighbors (KNN), support vector machine (SVM), Naïve Bayes (NB), and a lightweight convolutional neural network (CNN). Results confirm that the proposed CNN significantly outperforms the classical models with an accuracy of 99.91%, F1-score of 99.90%, and an average area under the curve of 0.98, thus representing improvements of +2.89% over RF, +6.03% over KNN, +9.66% over SVM, and +13.12% over NB. The CNN provides an enhancement to early-fault recognizability of subtle capacitor-degradation stages by up to 18.7% compared with traditional feature-based classifiers, underlining its far superior sensitivity to nonstationary electromagnetic signatures. These results confirm the effectiveness of combining physics-driven simulation with data-driven classification, which makes it a reproducible, computationally efficient framework for early detection of capacitor degradation in SEIG-based renewable MGs. This hybrid approach inherently supports predictive maintenance policies, enhances system reliability, and enables more resilient wind-energy integration in isolated power networks.