| Ultra high frequency (UHF) method has been widely applied in GIS partial discharge (PD) on-line monitoring because of its outstanding advantages, such as strong anti-jamming capability. Although the UHF method can avoid most of the low frequency interference, when the multiple random periodic narrow-band noise whose frequencies lie in the monitor band of the sensor intrude into the monitoring system, it will seriously affect the credibility and accuracy of monitoring, but it’s lack of good method to suppress this kind of noise. In addition, UHF PD signal contains abundant information, the feature extraction of UHF PD signal is important for the accurate identification of the GIS internal insulation defect types and guiding the repair work, more effective feature extraction method still need to be explored to lay a foundation for the subsequent fault diagnosis.This paper took advantage of the strict box-shaped spectral characteristic of harmonic wavelet and proposed a new de-noising method based on the optimal discrete harmonic wavelet packet transform (HWPT) to surpress the random narrow-band noise with mixed frequencies in PD signal. The energy of the narrow-band noise with different frequencies was concentrated in a single sub-band respectively through the optimal harmonic wavelet packet decomposition, and the sub-bands containing narrow-band noise were identified according to the value of the sub-band Shannon entropy ratio. After the corresponding sub-band was set to zero and reconstruction, the de-noised PD signal was obtained. The sub-band spectral leakage of discrete wavelet packet transform (WPT) was overcome and the adaptive optimization decomposition of PD monitoring signals was achieved. Through de-noising processing of the random narrow-band noise with mixed frequencies that lied in the frequency band of the simulated and measured PD signal, and compared with the result of WPT method, it shows that the energy loss and waveform distortion of the PD signal is smaller after the optimal HWPT de-noising. This is beneficial to the subsequent pattern recognition of the PD signal, and the de-noising problem when the noise frequencies lie in the monitoring band is solved.Four typical GIS internal defect models was designed in the laboratory, and a large number of UHF PD signal samples under different physical conditions were acquired by using the UHF detection system in the lab, then a UHF PD feature information extraction method based on harmonic wavelet packet transform was proposed.Multi-scale decomposition with HWPT was adopted to the UHF PD signal produced bythe four typical discharge models in the laboratory, and the false information containedthe sub-band signals by real wavelet packet decomposition was avoid. Using the energyand complexity difference of the UHF PD signal in different scales, the multi-scaleenergy and multi-scale sample entropy of the reconstructed sub-band signals wereextracted as the parameters for pattern recognition, so the time-frequency domaininformation of the UHF PD signal could be described more accurately. Then the Fisherlinear discriminant analysis was applied to the original feature set for key featureselection, the key features which conducive to the correct classification were preserved,the redundant features were eliminated, and the dimension of the original feature spacewas reduced effectively.The classification and recognition results with support vector machine shows thatthe multi-scale energy and multi-scale sample entropy feature parameters based onHWPT are both able to identify the four kinds of insulation defects effectively, theaverage recognition accuracy is higher than90%, which is much better than the featureextraction method based on WPT. With the feature selection operation, not only therecognition efficiency can be improved, but also the average recognition accuracy of theinsulation defects is improved effectively, which shows a good application prospect. |
PDF Full Text Request