Partial Discharge Identification in High-Voltage Systems Using Frequency Extraction and Machine Learning Techniques

This study focuses on the analysis and identification of four types of partial discharge (PD) signals: internal discharge, surface discharge, positive needle corona discharge (Air Corona; High Voltage Needle), and negative needle corona discharge (Air Corona; Needle to Ground). The PD signals were measured in a high-voltage laboratory environment and recorded using an oscilloscope connected to a notebook equipped with IO Libraries suite, capturing timedomain data. Signal characteristics were extracted through Fast Fourier Transform (FFT), noise filtering, and histogram distribution analysis. Gaussian Curve Fitting was employed to identify primary and secondary features, which were subsequently used as inputs for a Feedforward Neural Network (FNN) model with one hidden layer. The model's performance was optimized by varying key parameters, including the number of input features, hidden layer nodes, validation set partitioning (k-fold), and learning rate. The optimal configuration, utilizing two input features (mean frequency of primary and secondary characteristics), five hidden nodes, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k=4$</tex> validation set, and a learning rate of 0.1, achieved superior identification performance with an accuracy of 90 % and an F1-score of 90.7 %. The results demonstrate the efficacy of the proposed methodology in accurately identifying partial discharge patterns, highlighting its potential for application in highvoltage electrical system diagnostics.