Extraction Of Partial Discharge Signal Feature With Complex Wavelet Transform

The feature extraction of partial discharge signal of transformer is the front research project in online monitoring and diagnosing high voltage insulation failure. However, those present PD monitoring technologies are unsatisfactory for engineering application, it is because that pulse current is the common feature of all diagnosing methods, and noise restraining is the key trouble which haven't been resolved yet in online monitoring and diagnosing. Because of its excellent local time-frequency analysis characteristic, wavelet transform is widely used in signal analysis and feature extraction. Some useful and important research fruits of using WT in PD online monitoring and diagnosing are obtained although too much key technology is still vacant. This paper studies the principles and methods of complex wavelet in extracting PD signal features, and obtains some valuable results.(1) After analyzing principles and methods of real wavelet transform (RWT), I find that only real coefficients can be obtained with real wavelet technology, the signal feature information is extracted from exclusive amplitude value with frequency-domain analysis, namely, it only relies on amplitude-frequency characteristic (AFC) after real wavelet transform. Consider that AFC of PD signals is different from AFC of periodic narrow-bass disturbance and white noise, but similar with that of pulse noise, real wavelet technology can only be used to wipe off previous two noise, it is useless to separate PD signal from pulse noise. (2) CWT is composed of two RWTs, that is RWT of real part and imaginary part respectively. Real part and imaginary part of CWT are orthonormal, which means CWT of signal f(t,y) includes two RWTs in each orthonormal space, real part and imaginary part of coefficients include sectional information in these two orthonormal space. However, coefficients after RWT represent sectional information in only one space. Unusual relation information arctan(IW/RW) between two signal part in two orthonormal space is obtained after CWT, the one and only arctan(y/t), which shows orthonormal relation between t and y of original signal f(t,y), is transferred to arctan(IW/RW) the phase part of coefficient. We can extract phase information of original signal from imaginary part after CWT.(3) When I analyze the characteristic of PD signal and three kinds of noise that is white noise, periodic noise and pulse noise, I find their difference in phase-frequency characteristic or amplitude- frequency characteristic. CWT is more superior than RWT in extracting full- scale feature from both amplitude and phase, it is feasible to choose CWT to extract phase-frequency characteristic and amplitude- frequency characteristic of PD signal.(4) I find that phase-frequency characteristic of CWT directly reflect its analysis effect on PD signal. When phase-frequency characteristic of CWT is fairly similar with that of signal, the analysis result is the best. This paper considers CWT as the excellent method to extract PDS feature.(5) After analyzing the relationship between CWT coefficients and PDS feature value, I find amplitude part and phase part of CWT coefficients synthetically reflect similarity of PDS and wavelet radix function. If the eigenvalue includes synthesis information of both amplitude and phase, it will be feasible for CWT to open out tiny feature difference between PD signal and other noise.(6) In conclusion, if we select proper complex wavelet, and its eigenvalue includes synthesis information of both amplitude and phase, it will effectively restrain all kinds of disturbance such as white noise, periodic noise, and pulse noise, and extract tiny feature difference between PD signal and other noise.