A novel method to enhance partial discharge localization accuracy using sectional winding analysis and a neural network-based learning vector quantization model

Timely detection of insulation faults in power transformers is essential to prevent damage, outages, and financial losses. Locating partial discharge (PD) sources in transformers with complex windings poses a significant challenge, leading to insulation failure. This paper presents a novel approach combining a learning vector quantization (LVQ) network with frequency response analysis (FRA) to locate PDs along the windings, even in noisy environments, without using noise suppression techniques. The method obtains sectional winding transfer functions (SWTFs) using vector fitting (VF). A Gaussian function is injected into each SWTF, and the responses are captured as PD reference signals to train the LVQ network. In addition, a 5000 pC PD test pulse from a calibrator is injected into randomly selected laboratory winding sections, while a traveling rectangular wave is injected into the estimated SWTFs, obtained from FRA using vector fitting. The resulting responses are utilized as PD test signals to assess the proposed method across various PD pulse waveforms. The simulation and experimental results demonstrate that the proposed method achieves superior performance, particularly in high-level background noise, leading to significantly improved PD localization accuracy. Finally, the method's efficiency is compared with previous studies, particularly regarding performance in noisy conditions.