In the light of eliminating the interference of noise, the limitation of traditional wavelet packet threshold denoising method is analyzed and an improved wavelet packet threshold denoising algorithm based on sample entropy is proposed. At the aspect of threshold function, the method takes sample entropy as a characteristic parameter to represent noise distribution of noisy signal wavelet packet coefficients, and improves the threshold function according to the parameter value, so that it can adjust adaptively based on the noise distribution of signal wavelet packet coefficients. In terms of threshold selection, define the difference between denoised signal and original signal as the estimation of the noise signal, and use sample entropy as the determination criterion to choose the optimal threshold which makes the noise estimation sample entropy maximum. Finally, the method is used to denoise rolling bearing vibration. Comparison with the traditional method shows that the method removed noise effectively and restored signal frequency characteristic better, it′s a greater denoising algorithm.