In this paper, high voltage conductor metal protrusions, suspended particles and immobilized metal particles on gas insulated switchgear(GIS) insulators were simulated in the GIS model. The high voltage conductor metal protrusions defect was simulated by a needle-plate model. The GIS model was filled with 0.4MPa SF6 gas. When the voltage was added to 60kV, the three models all had stable discharge. Ultrasonic sensor was used to measure the discharge waveform for 100 groups. The absolute value of difference between the amplitude of adjacent half wave as Udif and the absolute sum of a cycle of the signal as Utal were chosen as the characteristic parameters. The defect types were recognized with BP neural network and the recognition rate is about 80%.