Using UHF method to detect GIS partial discharging, it tend to be different types of noise. The UHF signals which are collected by the antenna are easy to be submerged, which cannot cause partial discharge detection accurately or system working normally. For above problems, putting forward a sort of GIS partial discharge UHF signal noise reduction method based on improved EMD. This method uses dual-tree complex wavelet transform and empirical mode decomposition noise reduction method to improve and reduce the GIS partial discharge UHF signal noise. Using EMD method to decompose the noise signals into a series of IMF components,and then using joint distribution model for each IMF component of DT-CWT noise wavelet coefficient to estimate. Finally, reduced IMF signals are rebuilt to get reduction noise signals. GIS partial discharge UHF signals reduction test results show that this method achieves a good effect to separate noise and non-noise signals. Besides it, the method also has the characteristics which contain high signal-to-noise ratio and can keep early partial discharge UHF signals.