Aiming at the problem that the rotor vibration signal is non-stationary and the weak fault features are difficult to extract, a fault feature extraction method based on ensemble empirical mode decomposition (EEMD) and singular value entropy and manifold learning algorithm is proposed. Firstly, the original vibration signal is decomposed by EEMD, and some intrinsic mode function (IMF) components are obtained. According to the kurtosis-European distance evaluation index, the sensitive components with rich fault information are selected to form the initial eigenvector, and the singularity is obtained. Value entropy. Then, using the near probability distance Laplacian eigenmap (NPDLE), the feature matrix composed of singular value entropy is reduced. Finally, the obtained low-dimensional feature subset is input into the K-nearest neighbor (KNN) for fault pattern recognition. The proposed method was validated by a two-span rotor test bench dataset and Iris simulation dataset. The experimental results show that the combination of IMF singular value entropy and NPDLE can effectively extract the rotor fault features and improve the accuracy of fault identification.