Due to the non-stationary property of partial discharge（PD） signals

the joint time-frequency domain analysis of discharge pulses can reflect the discharge characteristics more perfectly compared with time domain or frequency domain analysis. Hence

in this study a needle-plate defect model

a floating defect model and an internal cavity defect model were designed according to possible defects in actual GIS. Gabor distribution characteristics of PD pulses of different defects under power frequency voltage were obtained

and the characteristic parameters were input to a back propagation artificial neural network for training and test. The validity of Gabor characteristic analysis on discharge pulses was analyzed through field PD test of 110 kV GIS defects. The results showed that the time-frequency spectrograms of the discharge pulses were extracted effectively for discharge assessment

and the training and test identification rates of the characteristic parameters were high enough for identification of discharge types.