The Experimetal Study Of Neural Network Application In PD Pattern Recognition

The quantity of partial discharge (PD) is one of the most important technical parameters which are used to assess the insulation state of high voltage electrical apparatus on-line. The relation of insulation defects with PD sources is closely interrelated so that operating states of high voltage electrical apparatus can be clearly monitored,and the insulation failures can be simultaneously diagnosed, predicted, and the locations of insulation defects and the deteriorated degrees could be deduced logically from the PD data-base mined real time. Therefore, theoretical study and scientific experiment on pattern recognition method of PD signal waveform not only have important academic significance, but have great practical application values. Reviewed many researchers'working results and based on a great deal of research literatures published in domestic and aboard on pattern recognition of PD failures, the three kinds of PD models are structured according to the mechanism of generating PD for simulating and demonstrating the PD physical and chemistry phenomena while appearing the insulating failures of the operating apparatus in the Lab, which include point-point electrode system, point-plate electrode system and sphere-plate electrode system. And the devices for sampling PD signals are designed according to the electronic technique, furthermore, the computer software packages and the operating interface are made up for processing PD data acquired from the experiment models, and for the software being more maneuverable, the C++ Builder programming languages are combined with the MATLAB. The classification effects of PD pattern recognition depend on the features of the pattern, topology constitution of the networks and their training arithmetic. Based on the method of extracting PD feature, choosing the optimum feature vector space theory and method of pattern classification, a new kind of WNN is created, which possesses the PD signal features could be extracted by self-adaptive, and the training arithmetic is proved and demonstrated by the mathematics and the experiment results. The extracting feature self-adaptive wavelet neural network is a sort of feed-forward neural network 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. Meanwhile, the PD pattern recognition experiments are carried out individually by using of the extracting feature self adaptively wavelet neural network, combining the moment feature of PD picture gray with BP neural network and the PD statistical feature with orthogonal wavelet neural network, their results show that choosing the best efficiency feature of the PD pattern recognition and topology constitution of the networks and their training arithmetic are vital and great effect for correct classifying, and that extracting feature self-adaptive wavelet neural network can be determined easily by the theoretical base and the training methods are sampling the convergence rates are very quickly comparing extracting feature self-adaptive wavelet neural network to BP neural network and orthogonal wavelet neural network.