For the problem of the lack of encoding, critical value difficult to determine and faults at the same time difficult to distinguish in the three-ratio method caused by the fuzzy relationship between transformer feature and fault types, a fault diagnosis method for power transformer based on optimizing characteristics and the fuzzy theory is proposed. Each fault data in the measurement-space is transformed to Hilbert space through the Gaussian kernel function firstly. Then, the method of PCA is used to extract fault features, which is regarded as the most effective features. Membership functions describing the fault character is built, and the type of fault can be identified according to the maximum membership degree. Feature subspace not only contains the information of fault characters from measurement space, but also generates the new characters as the ability of the kernel theory. Thus, the clear relation between fault characteristics and fault type is built. Example analysis shows that this method can make up for the inadequacy of three-ratio method. Similarities between fault data and its own fault and differences between fault data and not its own fault can be observed. When faults occur at the same time, the method can help the worker.