Abstract: Transformer is the key equipment in a power system. Partial discharge (Partial Discharge, PD) is the main cause of transformer insulation deterioration, and also an important manifestation of insulation deterioration. Aiming at the complex noise information such as the narrowband interference or the white noise in the PD signals, an improved adaptive parameterless empirical wavelet transform noise reduction based on the mutual information and the spline interpolation fitting to optimize the spectrum segmentation is proposed. The method combines and optimizes the mutual information values of each time domain component decomposed by the parameterless empirical wavelet transform to obtain new spectral division points. Then, by extracting the peak points in each frequency band, the cubic spline interpolation fitting is performed to obtain a fitting curve. The minimum value point of the fitted curve is used as the latest spectrum division point to obtain the component in each time domain. Finally, the kurtosis value of each component is calculated. By removing components with a kurtosis value less than 3, the remaining components are subjected to the general threshold noise reduction and reconstruction, which achieves the PD signal noise reduction. Simulation, laboratory and field tests show that, compared with the existing wavelet transform noise reduction and the empirical mode decomposition noise reduction methods, the method in this paper more effectively suppresses the noise information in the PD signals.