An evaluation of total variation signal denoising methods for partial discharge signals

Abstract

Partial discharge (PD) monitoring in high voltage power plants is a useful tool to detect the presence of insulation degradation in the electrical systems [1]. Specifically, condition assessment of power systems has the advantage of reducing maintenance and repair costs, downtime and reducing risks. This requires measurement of field PD signals which are often observed in an environment contaminated by Additive White Gaussian Noise (AWGN) which may mask relevant PD information [2]. The use of Electromagnetic Interference (EMI) testing has also been suggested for condition assessment of insulation systems in turbine-driven generators. These techniques allow online testing which avoids generator shutdown during long inspections [3]. Power assets such as transformers, generators and cables surround the EMI signals making them very susceptible to noise. This affects the relevant information contained in the signal which hinders the analysis for PD identification and may affect the classification accuracy of PD. The proposed solution is to employ signal denoising techniques for noise mitigation in the captured field signals. Several denoising methods have been used on PD signals. The most popular technique is wavelet-based which has successfully denoised simulated and real PD signals combined with hard and soft thresholding of the wavelet coefficients [4]. The Wavelet Transform (WT) was exploited in [5] using high spatial correlation and Support Vector Machine (SVM) methods to distinguish the relevant PD coefficients from noise. A similar decomposition method using the eigen approach was implemented in [6] to reduce noise collected in field PD data. Different techniques for wavelet coefficient selection were introduced for enhanced PD denoising in [7,8]. The Complex Wavelet Transform (CWT) was used instead of WT in [9,10], in combination with simple and combined information, for signal denoising because of CWT adaptivity to the non-stationary PD signal. However, PD signal denoising is still an ongoing area of research which aims to achieve improved results for non-stationariy nature signals.