Needle-plate, suspended metal particles and metal particles fixed on insulator surface were placed separately in GIS entity model. The discharge waveforms detected by using ultrasonic sensor under the same voltage acquisition to partial discharge waveform good. Aiming at the waveform selection of 7 parameters, the discharge type recognition with LVQ neural network, the recognition rate of 96.7%, higher than that of S_Kohonen neural network, has the very good practical value.