Comparative Investigation and Determination of Partial Discharge Source Using Gaussian Naive Bayes (GNB)And K- Nearest Neighbour Method

Abstract

Partial discharge (PD) detection and classification play a vital role in ensuring the reliability and longevity of high-voltage insulation systems. This study presents a comparative investigation into the identification and determination of PD sources using two supervised machine learning techniques—Gaussian Naive Bayes (GNB) and K-Nearest Neighbor (KNN). Experimental PD data were acquired under controlled conditions from different defect models such as surface discharge, internal discharge, and corona discharge. Key statistical and time–frequency features were extracted from the acquired signals to form an efficient feature set for classification. The GNB algorithm, based on probabilistic reasoning and the assumption of feature independence, provides a fast and computationally efficient classification mechanism. In contrast, the KNN method, a non-parametric approach, classifies data based on similarity measures in multidimensional space. The comparative analysis evaluates both models in terms of accuracy, precision, recall, and computational complexity. Experimental results demonstrate that while KNN yields superior accuracy for nonlinear PD patterns, GNB performs efficiently for real-time and low-complexity scenarios. The study highlights the trade-offs between model performance and computational demand, suggesting a hybrid or ensemble approach for future PD source identification systems in high-voltage applications. This work contributes to improving predictive maintenance and insulation health assessment methodologies.