It is well known that the second generation wavelet is the best de-noising means, but the result of de-noising depends on how to set up the basis function, decomposition layers and threshold parameters. Without parameter setting empirical mode decomposition (EMD) decomposes the signal into intrinsic mode functions (IMF), then structuring IMF filter and the de-noising process is adaptive. It is worth noting that the signal and the noise are mixed together in very high frequency, that is to say that there has been mode overlap, and what happened will affect the de-noising effect. It was found that ensemble empirical mode decomposition (EEMD) decomposes Raman spectrum into the signal and the noise effectively avoiding from mode overlap in high frequency in the experiments, and it is similar with wavelet in de-nosing effect fortunately. At first, a period of non-linear and non-smooth bean greases Raman spectrum was decomposed by EMD in the paper, there was mode overlap, but the authors have got clear characteristic components by EEMD. Secondly noisy spectrum was processed by fast Fourier transform (FFT), wavelet, EMD and EEMD independently, and signal to noise ratio, root mean square error and correlation coefficient indicate that FFT is the worse means in high frequency de-noising than EMD, and the appropriate wavelet is similar with EEMD in de-noising result, but the de-noising process of EEMD is adaptive. In the last section, a brief research direction of the spectrum study method in time frequency field and noise properties criterion on IMF are given for the future.