This paper simulates four typical defects in GIS for PD detection, and uses the pulse, amplitude, phase and number of PD to form the three-dimensional PQN matrix. Based on the PQN, three two-dimensional distributions of Hqmax similar to Phi, Hqmean similar to Phi and Hn similar to Phi can be achieved. Then the new G-K clustering method is introduced to separate the four different defects according to the parameters of Sk, Ku, Pe, Q and CC. At last the comparison with different clustering methods is achieved based on the evaluation criterions which are PC, CE, PI and SI.