PD Pattern Recognition Based On BP Network

Partial discharge (PD) level is one of important technical indexes which are used to assessing the insulation state of high voltage electrical apparatus. It is not only reflection of the insulation conditions, but also predicts the service expectancy-life of the insulation used in the high voltage electrical apparatus. However, the randomicity of the partial discharge affects the partial discharge sampling. Meanwhile, it is also make some troubles for pattern recognition of partial discharge. The projects caaried out in this dissertation have discussed the possibility in details for PD pattern reconginition constructed new kinds of classifiers, extracted new features and invented new method to de-noise, aimed at to improve the classifying effects and been avoided aimless constructing the network topology. The PD pattern recognition methods are compared and new ways are brought forward in this dissertation, including a method to use statistical features of partial discharge grey intensity images for PD pattern recognition, the statistical features are consisted of moments, and statistical features of the PD signals based on the orthogonal WNN and another is self-adaptive extracting feature wavelet neural network constructed by author. The moments based on the grey intensity images, which is short for MGII and a special characteristics, are first used in the features extracting of the PD signals. The PD signals are taken as a picture which has some distribution of the grey intensity images, so the method dealing with images for recognition can be used for the recognizing PD. The MGII features of PD signals sampling in the Lab are input the BPNN for learning and training, the correct recognizing rate can be up to 75%. The statistical features are maken use of taking as the input matrix of the orthogonal WNN for classifying the PD paatren extracted from the PD signals by the experiment models, because the exciting functions of orthogonal WNN possess orthogonality, and the rates of pattern recognition for PD failures are much higher and maybe up to 82%. After analyzing the defects of the BP neural network and orthogonal WNN, a kind of new classifier, called self-adaptive extracting feature wavelet neural network that both optimum feature are extracted and patterns are recognized simultaneously, is first presented and constructed in the research project and its classification effects of PD pattern recognition are demonstrated by the theoretical and experiment, and its learning and training arithmetic are proved by the mathematics and the experiment results. The self-adaptive extracting feature wavelet neural network is a sort of feed-forward neural network that has both function of choosing the best efficiency PD time-frequency features and pattern sorts, in order to increase the network convergence rate, the adaptive learning step size arithmetic is employed to adjust the network learning rates. In the project, sampling three kinds PD signals (needle-needle, needle-plane, sphere-plane electrode systems) and their features are extracted to train the self-adaptive extracting feature wavelet neural network; its recognition rates are 90%. To meet the demand for PD pattern recognition , the singularity of PD pulse signals and typical noising signals (function) are theoretically analyzed and demonstrated, respectively, furthermore, their different characteristics are presented during the wavelet transforms in different scales, these distinct feathers presented to the wavelet transforms can help to extracting the PD pulse signals from the any noise disturbances and make sure the PD pulse signals are fidelity and could be really expressive the characteristic of the PD source. A sort of de-noise method is presented and the de-noise procedure and valid arithmetic are explained in details and carry out successfully simulation with the computer. These results achieved from the research project could supply for technique assurance and theoretical references for PD detective on-line, and the researchapproaches and ideas on PD pattern recognition are radical basises for realizing theory breakthrough in the future.