Aiming at the problem of accurate classification is difficult which is due to a certain correlation between fault feature attribute domain, a kind of rotor fault data classification method of considering the influence of the correlation coefficient is proposed. This method is the result of concept combining the Correlation Manifold Distance Marginal Fisher Analysis (Correlation Manifold Distance Marginal Fisher Analysis, CDMFA) and Correlation Manifold Distance K - Nearest Neighbor (Correlation Manifold Distance K-Nearest Neighbor, CDKNN) classifier together. First of all, the vibration signal are converted into high-dimensional data-set of multi-domain and multi-channel. Then, using correlation manifold distance of the fused correlation coefficient to measure neighbors and weights of fault samples by the CDMFA, which can better reflect the similarity relation between high-dimensional data and extract low-dimensional feature subset of making the bigger distance between the class. Finally, the low-dimensional feature subset is input into CDKNN classifier for fault pattern recognition. The proposed method is verified by using a double span rotor system data-set and simulation data-set. The results show that the method has better dimension reduction effect and higher fault classification accuracy. The study finds that the manifold distance fault data classification method can reveal more realistic high-dimensional feature geometry relation. This method provides the theory reference of data preprocessing for feature attribute reduction and classification of the high dimensional fault data-set.