The invention discloses a power transformer fault diagnosis method and device. The method comprises the steps that: after dissolved gas analysis (DGA) data in power transformer oil is obtained, faulttype identification on a fault sample is carried out through a support vector regression-variable prediction model (SVR-VPMCD); firstly, ratios of all kinds of gas in the total gas content are selected as various characteristic values, so that corresponding characteristic vectors are formed; then, polynomial response surface regression in VPMCD is replaced by SVR; variable prediction models (VPMs)corresponding to various fault types are constructed by utilization of a training sample; finally, each test sample is predicted by sequential utilization of the constructed VPMs of all kinds of faults; and the fault type corresponding to the VPM having the minimum prediction error is the fault type of the test sample. The SVR-VPMCD method sufficiently considers mutual correlation among various characteristic values, and overcomes the disadvantage that the VPMCD method has poor processing capability to high-dimensional data; therefore, the identification precision of the mode identification method on the sample type is increased; and the method has good small sample data identification capability.