The application of compound networks in fault diagnosis of power transformer

Using the concepts of typical gas´s concentration and cumulative frequency in analysis of the reliability data for dealing with the pretreatment of data of DGA, two new normalized methods which named characteristic normalization and mix normalization are presented in this paper. The Fisher rule to evaluate the results of the two pretreatment methods is also introduced. The evaluation of the results indicates that both of the two data pretreatment methods can achieve the purpose of big difference in the value of mean between classes and small difference in dispersion of a class. The DGA data of the failure transformers are treated by different normalization methods as the training samples, and then the samples are trained in the compound neural networks which use the CP algorithm. The diagnosis results of the test samples indicate that the new methods may help to improve the precision of network diagnosis.