A New Comprehensive Neural Network and Its Application in Partial Discharge Pattern Recognition of XLPE

Abstract

In this paper, a SOM-BP combined neural network model is designed by connecting two networks in series. The self-organized mapping (SOM) neural network is used as the primary network and the back propagation (BP) neural network as the secondary network. This network can avoid the shortcomings of SOM neural network that can't express pattern recognition results in vector form and that BP neural network needs a large number of training samples. In order to evaluate the performance of the model, the simulation experiments of partial discharge in cross-linked cables were carried out. The third and fourth order statistical characteristics of the ultra-wideband single discharge pulse in time domain were extracted as discharge fingerprints. Three kinds of neural networks, SOM, BP and SOM-BP, are used as classifiers to complete partial discharge pattern recognition. By comparing the recognition results of these three kinds of neural networks, it is found that when SOM-BP combined neural network is used as classifier, the recognition rate of each pattern is 90%. The network has the best recognition effect in both all kinds of recognition rates and the overall recognition rates. It is proved that the SOM-BP combined model is effective and reasonable.