The 2-D hidden Markov model (2D-HMM) has the ability of representing the signals, which are non-stationary and worse reappearance in the rotor run-up process. Analyzing the relationship of different level wavelet coefficients and the same levels, a new filtering method based on wavelet 2D-HMM was proposed and its detail steps were presented. Finally, this method was validated by the data collected from Bently-Nevada rotor experimental system (Model 24755). The results demonstrate that it is effective to remove the noise of non-stationary signals of rotor system.