Research On The Application Of Fractional Fourier Transform In Extraction Of Partial Discharge

On-line Partial Discharge (PD) monitoring is meaningful in precise interpretation of insulating condition and accident prevention.One of the most important problems in on-line PD monitoring is how to detect the weak stochastic PD pulses from the strong background noises. PD signal is a kind of non-stationary time-variable signal. Frequency Spectrum analysis technology is a traditional method for signal detecting and analyzing. This method is effective when signal is stationary and spectrum is different from noise. But in fact, what we face usually is nonstationary signal. It is necessary that every frequency component in every moment needs to be analyzed. So the traditional method has some disadvantages and it is difficult to detect the PD signal from the interference effectively. A new tool for time-frequency analysis, Fractional Fourier Transform (FRFT), is introduced in this paper. The Fractional Fourier Transform which is a version of generalized common Fourier Transform (FT) can overcome the incomplete of FT in feature extraction of fault signal. FRFT also reveals the features of PD signal from time and frequency simultaneously.At the weak noise condition, the paper uses Wavelet Transform and Fractional Fourier Transform to extract the PD signal respectively. The auto-threshold function was used to deal with the noise signal in the Wavelet Transform, while in the Fractional Fourier domain optimal linear filter method was used. Both of them take good effect, but under the strong noise, the result is not very perfect.This paper proposes a novel method using Fractional Fourier Transform and Wavelet Transform for the analysis of partial discharge signals buried in strongly excessive noise. In Fractional Fourier domain, FRFT spectrums of PD signals show different time-frequency congregation. The algorithm for scanning the optimal order of FRFT according to the minimum mean square error (MMSE) criterion is derived and 3σrule based on FRFT is introduced to the pretreatment of PD data. To extract the PD features more exactly, the pretreated signals need to be de-noised based on Wavelet Transform. Db8 wavelet and the soft threshold based on Stein's Unbiased Estimate of Risk are used to smoothly denoise. The simulation and experimental results demonstrate effectiveness and feasibility of the proposed method.