This paper presents an analysis of partial discharge (PD) signals in high voltage current transformers (HVCTs). Our approach centers on using a convolutional neural network with an autoencoder architecture to filter out electrical noise, enabling assessment of the characteristic waveform shapes of PDs. The denoising convolutional autoencoder (DCAE) was trained using an extensive database composed of synthetic PD signals, with several characteristic waveforms, noise levels, quantities, and positions of occurrences. Furthermore, the DCAE was tested using real data collected from HVCTs during commissioning in electrical substations from various regions of Brazil. A comparative analysis between the DCAE and wavelet denoising methods is conducted. The results indicate that the proposed DCAE neural-network effectively denoises both synthetic and real signals, preserving the impulsive characteristics of PDs.
