A Simplified Convolutional Autoencoder for Denoising of UHF Partial Discharge Signals

Abstract

Partial discharge (PD) is a nondestructive testing method to identify defects inside electrical insulation. Hence, accurate PD measurement plays a crucial role in the condition monitoring of high-voltage apparatus. Since PD signals have very low amplitude, the real-life PD signals are often perturbed with field noise and interference from the mains. In this study, a deep learning (DL) framework is proposed for accurate denoising of PD signals. The work implements a simplified convolutional denoising autoencoder (SCDAE) designed to denoise both simulated and real-life PD signals. Initially, two standard PD signals were mathematically modeled and subjected to white Gaussian noise across varying signal-to-noise ratios (SNRs). The pretrained SCDAE was then applied, with the performance evaluated using the SNR, peak SNR (PSNR), root-mean-square error (RMSE), and normalized correlation coefficient (NCC). For experimental validation, laboratory experiments were conducted to replicate corona, surface discharge, and internal cavity discharge sources, with PD signals captured via a wideband ultrahigh-frequency (UHF) sensor. A comparative analysis against existing DL methods demonstrated that the model achieved an output SNR of 4.4–4.94 dB at a noise level of −10 dB for simulated data and an SNR improvement of up to 7.61–16.03 dB for laboratory-generated signals, confirming its effectiveness even under severely noisy conditions and its suitability for real-world PD denoising applications.