SF6/N2 mixed gas is beneficial for reducing the amount of SF6 gas used, but its partial discharge characteristics differ from pure SF6. This paper sets up four mixed gas ratios of SF6 and N2, and conducts partial discharge experiments with three types of defects and different gas pressures. The experiments research discharge ultra-high frequency (UHF) signals, and the PRPD spectrum of UHF signals is established, and an improved EfficientNetV2 neural network is used for partial discharge pattern recognition. Research shows that the UHF method has good detection sensitivity for tip corona discharge, surface discharge, and metal particle deficiency discharge. The PRPD spectrum shows that the signal phase of tip discharge has three different distribution characteristics. The signal of surface discharge is similar to that of tip discharge, but there are signals with the same discharge amplitude and phase difference of 180°. The particle discharge signal exhibits a characteristic similar to “W”. The recognition accuracy of improved EfficientNetV2 can achieve 98%.