A locality margin discriminant projection （LMDP） algorithm is proposed to reduce the dimension of the fault feature set．The algorithm defines the local intra-class similarity and the local inter-class similarity to separate the neighboring samples of different classes and join those in the same class. Then， the time-domain and frequency-domain statistical characteristics of rotor vibration signals are extracted to form the original fault feature set. The LMDP algorithm fuse the feature set to select a low-dimensional sensitive feature subset that contains the most intrinsic information. Finally， the K-nearest neighbor （KNN） classifier trains the subset and classify the faults. The vibration signal seta of two different double-span rotor systems verify the effectiveness of the proposed method.