Research On Recognition Of Image Of Partial Discharge Based On The Technology Of Wavelet Multi-Scale Transform

The technology of monitor on line and pattern recognition of partial discharge (PD) is an important means to discover internal local defects and hidden danger of failures of transformer. At present, the research content of pattern recognition of PD are mainly discharging characteristic amounts and pattern classifiers etc.. Based on existing work at home and abroad, by experiment datas of laboratory partial discharge, this article studies systematically multi-scale wavelet transform of image of PD and wavelet neural network (WNN) based on genetic algorithm optimizing, and based on features extractions from sub-images after multi wavelet, recognizes by pattern classifier constructed by this article. The main contents are as follows:①This article constructs 3-D spectrums and gray images of PD by sample datas of PD obtained from the test of models using five models of transformer discharge. Based on the basic theory and specific algorithm of wavelet multi-scale transform, Sinsin picture and spectrums of PD are analyzed using multi-resolution wavelet transform, the energy distribution of the scales of the image are obtained.②This article extracts feature set made of fractal and moments characteristic parameters from the original image and three-scale transformed sub-images of image of PD recognizes feature sets of images by using BP neural network (BPNN) classifiers as pattern classifier. The recognition results shows that, extracting this feature set, the recognition result of high-frequency on vertical LH3 sub-image is better than the original image and other sub-images.③This article constructs WNN to recognize original image and all of the sub-images, the speed of convergence and results of recognition in WNN are better than in BPNN. Genetic algorithm (GA) combined with, a method, which parameters and network hidden nodes of WNN are optimized first, which are trained in neural network then, is proposed. The result shows that, the optimal neural network has faster convergence speed.