Feature extraction is particularly important for diagnosing the faults of aero-engine rotor. Based on the nonlinearity and nonstationarity of vibration signal of aero-engine rotor， first of all， Fourier decomposition method （FDM） is applied to signal of aero-engine rotor. The marginal spectrum centroid and the power spectral centroid of maximum energy layer are extracted. Then the refined composite multiscale dispersion entropy （RCMDE） of vibration signal is calculated. Finally， the method of two-stage adaptive wavecluster is applied to fault classification and recognition of eigenvector space. Through the sample verification of the aero-engine rotor test rig， it is shown that the extracted characteristic vectors are accurate and fluctuated little， the value of the same fault type is concentrated， and the difference between different fault types is large， which are helpful to improve the diagnosis accuracy of multiple-faults.