Ensemble empirical mode decomposition method (EEMD) is important improvement of the empirical mode decomposition (EMD), but it still leaves some annoying difficulties unresolved, such as large computational cost and reconstruction error. To solve above problems of the EEMD, a new method named EEMD of adaptive noise (EEMDAN) is proposed, which can obtain accurate time-frequency distribution of the original signal. In addition, aiming at the non-linear, non-stationary and non-Gaussian characteristics of underwater acoustic signal, a noise reduction method based on the EEMDAN is also proposed. Each order intrinsic mode function (IMF) can be obtained by the EEMDAN for noisy Lorenz signal and three different types of real ship radiated noises. The components of the signal and noise can be adaptively determined by IMF correlation coefficient based on energy density and average cycle. The results show that the method is effective and a very clean signal can be obtained.