Semi-Supervised Hybrid Ensemble Learning for Fault Detection in 20kV XLPE Cables

Abstract

Cable fault detection is critical for ensuring the reliability and safety of high-voltage 20 kV XLPE cable systems, minimizing downtime and maintenance costs. This research introduces a semi-supervised hybrid ensemble model combining Random Forest, Gradient Boosting, and XGBoost within a Voting Classifier framework. Data preprocessing involves feature scaling and Gaussian noise injection (σ = 0.01) to enhance robustness, followed by training on 3943 labeled samples and iteratively incorporating highconfidence predictions (threshold > 0.9) from 11829 unlabeled samples. Evaluated on a dataset of 15772 samples with diverse features like cable age, partial discharge, corrosion, and loading conditions, the model achieves 98% accuracy, 97.5% recall, 97% precision, and 97% F1-score. Compared to SOTA supervised models such as SVM, CNN, and ANN, it demonstrates superior performance and scalability by leveraging unlabeled data. This approach offers an efficient, accurate solution for cable fault diagnosis in industrial applications.