The invention relates to an overheating fault diagnosis method of a transformer. The method comprises steps of calculating basic probability distribution functions of all kinds of faults on the evidence of codeless four ratios and characteristic gas by use of a posterior probability SVM; and then according to evidence theory definition, calculating fusion results of all kinds of the results under combined effects of multi-evidence bodies for belief functions of different evidence bodies. A probability estimation module is introduced in a 'one versus one' multiclass SVM by use of a pair-wise coupling algorithm, so defects of limitations on following data processing imposed by hard decision output when a traditional SVM is solving multiclass problems are overcome and an SVM probability output problem is effectively solved. According to the invention, by fully using redundant and complementation information of chromatographic data and electric test data and applying the D-S evidence theory, correction and reliability of comprehensive diagnosis results based on multi-default characteristics are greatly improved than the diagnosis based on single-default characteristics, which facilitates diagnosis of fault types of the transformer and has practical significance for formulation of maintenance strategies of the transformer.