For the problems of complex network structure, long training time and other issues in classical neural networks, a method of RBF neural network combined with rough sets is proposed in this paper for transformers fault diagnosis. Firstly, the minimum reduction is calculated after data mining based on distinguishes matrix and information entropy in rough set theory. Then, using a Gaussian function as the radial basis function, the variance and weights of layers is calculated by the processed data collection as training samples. Finally, the transformer fault diagnosis model is built. By comparing the test results, although the algorithm in terms of diagnostic accuracy is slightly lower, but the simple network structure, training speed and strong generalization ability of neural networks to improve application performance in transformer fault diagnosis have better guidance.