GIS is currently widely used in the power system and is the Developing Trends of Electrical appliances manufacturing. At present, a lot of research on GIS PD, but the study of the recognition of the PD severity is less. By analyzing a large number of partial discharge sample data, a new GIS PD severity diagnostic method is proposed in this paper. Partial discharge is divided into three stages (initial stage, development stage and dangerous stage) in the method according to the severity of the partial discharge. These three stages are fitted with non-linear mapping and get the fitting parameters. GIS partial discharge fault reference database can be dynamically generated with these fitting parameters. The similarity between the detected data and the reference data is discriminated with the least square method and Supplier division method, and then the GIS PD severity can be diagnosed. A large number of simulation results initially confirmed the effectiveness and feasibility of the proposed method.