In the field of fault diagnosis of power transformers， the main disadvantage of the back propagation algorithm of wavelet neural network commonly used lies in the optimization procedure getting easily stacked into the minimal value locally and strict requirement on the initial value， which would make fault diagnosis difficult to some extent. This paper presents a fault diagnostic method based on the genetic algorithm evolving wavelet neural network. The selection of network structure and parameters is carried out by use of the real value encoding genetic algorithm instead of artificial setting. Throughout the process， compromise is satisfactorily reached among the network complexity， the convergence and the generalization ability. The results of numerical examples show that the algorithm proposed has good classifying capability of both single-fault and multiple-fault samples as well as high accuracy of fault diagnosis.