Considering the non-stationary of rolling machinery fault data, the wavelet threshold denoising method based on CEEMD and Permutation Entropy (PE)is proposed to cover the shortage of Complementary Ensemble Empirical Mode Decomposition (CEEMD) denoising method and wavelet threshold denoising method. CEEMD is used to decompose signals into a series of IMF components, the Permutation Entropy is presentedto evaluate noise contained in each IMF component, and wavelet threshold method is adopted for IMF components containingmuch noise for denoising, in order to retain the useful information of these components. The test results ofsimulation data and experimental data show that, the wavelet threshold denoising method based on CEEMD and PE is better than pure CEEMD denoising method and the wavelet threshold denoising method.