The rotor is the most important part of the whole rotating machinery. Whether the rotor is normal directly determines the normal operation of the whole rotating machinery. Aiming at the problem of classification difficulty caused by multi-class and high-dimensional complex characteristics of rotor fault data, a fault data set reduction method based on Local-Global Balanced Orthogonal Discriminant Projection (LGBODP) is proposed. The algorithm comprehensively considers the intra-class local information, intra-class non-local information, inter-class local information and inter-class non-local information of the data, so as to avoid the loss of structure information in the dimension reduction process. By maximizing the inter-class distance and minimizing the intra-class distance, the intrinsic manifold structure information of the fault feature data set is effectively extracted while maintaining the global feature information. First of all, the mixed feature of the rotor vibration signal was extracted from multiple angles in time domain, frequency domain and time-frequency domain, and the high-dimensional feature set was constructed. The low-dimensional fault sensitive feature subsets are extracted by the proposed LGBODP algorithm. Then, the K-nearest neighbor (KNN) method is used as a fault feature classifier to recognize different fault types of rotors. The effectiveness of the proposed algorithm is verified by the vibration signal sets of two different types of double-span rotor systems. Application examples show that this method can be used to comprehensively extract the global and local discriminant information of vibration signals of rotors and effectively diagnosis the fault of rotors. © 2020 Elsevier Ltd