In order to deal with the shortcoming of the ensemble empirical mode decomposition (EEMD) method, where the amplitude of Gaussian white noise is required to be determined artificially, an improved EEMD method is proposed based on the characteristic of empirical mode decomposition (EMD) filter to determine the optimal amplitude of Gaussian white noise adaptively. The principle of determining the amplitude of the noise in EEMD is established by discussing the reasons of EMD mode confusion. The distribution of signal extreme points is taken as an evaluation index to determine the optimal amplitude of Gaussian white noise through an ergodic process. The validity of this improved EEMD method is verified by a simulation. Finally, the method is applied into the rotor early fault diagnosis. The result shows that the method is able to determine the optimal amplitude of Gaussian white noise adaptively, thus overcoming the mode mixing and extracting the fault information effectively.