A new method for transformer fault diagnosis based on cluster analysis and statistical theory is presented. First, the fault diagnosis results are obtained according to the distances between the state sorts of transformer. Then, the final fault diagnosis is accomplished according to the concentration distribution of typical fault gases in higher dimensional space. The proposed approach is constructing the most accuracy model from few training samples supporting. Moreover, by comparing with the other methods, it cost less time for diagnosing by the proposed model and the accuracy for transformer fault diagnosis is improved using our proposed model.