The invention relates to a transformer fault diagnosis method based on support vector description and K periapse neighbor. The method comprises the following steps: acquiring content data of characteristic gas H2, CH4, C2H6, C2H4 and C2H2 dissolved in transformer oil by virtue of a technology for analyzing dissolved gas in the transformer oil, preprocessing and normalizing the data, forming a transformer fault diagnosis sample set, establishing a multi-classification suprasphere model based on a single-classification support vector description method, diagnosing a fault of a transformer, further classifying an aliasing domain sample in a support vector description multi-classification process by using a K periapse neighbor classification algorithm, so that the fault diagnosis accuracy of the transformer is increased. By combining the multi-classification support vector description method and the K periapse neighbor classification method, small samples and imbalance sample sets can be accurately classified, the fault diagnosis precision of the transformer can be remarkably improved, and powerful support can be provided for making a state maintenance decision of the transformer.