In order to classify the ultrahighfrequency (UHF) partial discharge (PD) signals resulting from four types of insulation defects in gas insulated switchgear (GIS), the complex wavelet transform is applied to extract features of UHF PD signals. Five statistical parameters including mean, variance, kurtosis, skewness and energy are used to quantize the scaling coefficients of the complex wavelet transform and describe the feature subsets of UHF PD signals.A critical index J is defined to select features according to their classification performance. Using the J criterion, five optimal features are selected from sixty UHF PD features and taken as the input of radial basis function neural network. The classification results show that the information of real part and image part of complex wavelet coefficients indicates the characteristics of UHF PD singles and the recognition effect is pretty good. To use db4 complex wavelet can get the best classification performance.