In the detection of power system transformer stability, it is usually found that the transformer discharge signal has a variety of interference, especially noise. Aiming at the noise interference of transformer discharge signal, this paper presents a noise reduction method based on wavelet trace method. In this method, the traditional wavelet denoising algorithm has been improved. Based on the traditional wavelet transform, the wavelet trace dictionary is redefined, and the discharge signal to be processed is circularly translated. This method can effectively eliminate the translational dependence of wavelet basis and improve the particularity of wavelet transform in the vicinity of singularity so as to eliminate the random oscillation. By comparing with traditional wavelet packet de-noising and traditional wavelet denoising, threshold, wavelet base and decomposition layer number are selected, and the best value of signal denoising is obtained. Then a stationary wavelet trace threshold de-noising method suitable for the study of partial discharge signal is constructed. Combined with MATLAB simulation experiment signal and partial transformer partial discharge actual signal verification, this method can effectively restrain the noise of partial discharge and extract the effective signal. The noise elimination effect is better than the traditional wavelet and wavelet packet method, which can meet the needs of identification.