Synthetic Leakage Current Data Generation for Machine Learning-Based Pollution Classification in High Voltage Insulators

• cGAN framework generates synthetic leakage current data for insulator monitoring. • Synthetic data improves ML classification accuracy from 43-59% to 89-99%. • Six ML models validated: XGBoost, LightGBM, CatBoost, DNN, 1D CNN, Transformer. • Domain shift between insulator configurations successfully addressed via cGANs. • Transformer model achieves highest improvement of 55 percentage points. Reliable condition monitoring of high-voltage insulators is essential to prevent flashovers caused by surface contamination. However, developing robust machine learning models for leakage current analysis is hindered by limited datasets, particularly for multi-string configurations under diverse pollution levels. This paper presents a framework for generating synthetic leakage current data using Conditional Generative Adversarial Networks (cGANs) to address this data scarcity. The approach was validated using experimental data from ceramic insulators tested under three contamination levels and three string configurations (1-disc, 2-disc, and 3-disc). Time-domain and frequency-domain features were extracted and used to train six classifiers: XGBoost, LightGBM, CatBoost, DNN, 1D CNN, and Transformer. Two cases were evaluated: models trained on real 1–2 disc data only, and models trained on combined real and synthetic 3-disc data. Results show that cGAN-generated data closely matched real measurements statistically. When synthetic data was included, test accuracy on unseen 3-disc configurations improved from 43–59% to 89–99%, with Transformer achieving 98.89%. These findings demonstrate that generative models effectively address domain shift, offering a practical solution for AI-based insulator monitoring.