Enhancing Partial Discharge Classification Through Augmented Fault Data Balancing

Partial discharge (PD) is a prevalent phenomenon in high-voltage (HV) equipment, and its accurate classification is crucial for ensuring the reliability of power systems. For in situ systems, different types of faults, such as corona, floating electrode, surface, and void discharge, exhibit varying occurrences, posing challenges to accurate classification. This research addresses the inherent issues of classification accuracy caused by unbalanced fault data. Employing z-score normalization and combined synthetic data generation using a random undersampling and synthetic minority oversampling technique (SMOTE) ensures a fair representation of different fault types, leading to more accurate classification results. Further, after applying grid-search to optimize the hyperparameters, k-nearest neighbor (KNN), random forest (RF), and gradient boosting (GB) have achieved accuracies of 98.43%, 95.29%, and 88.54% for balanced denoised, unbalanced denoised, and unbalanced noisy datasets, respectively. The presented results also demonstrate a significant statistical difference in classifier accuracies between the three datasets, as confirmed by the analysis of variance (ANOVA) test. This emphasizes the efficacy of balancing the denoised signal features for improved classification performance. The findings of this work contribute valuable insights into the optimization of PD classification models, paving the way for more reliable fault detection and classification in HV equipment.