Aiming at the problem of classification difficulty caused by the strong nonlinearity and the high dimensionality of fault dataset of rotating machinery，a fault dataset dimension reduction algorithm local centroid mean minimum-distance discriminant projection（LCMMDP）is proposed. The algorithm can maintain the local geometric structure information of the sample while considering the cohesion and separation of the sample，reflecting the close relationship between the sample and the local centroid mean.The hybrid characteristics of rotor vibration signals are extracted from multiple angles，the original high-dimensional feature sets are constructed，and low-dimensional sensitive feature subsets are extracted by LCMMDP. The improved k-nearest centroid neighbor classification based on local mean and class mean is used（KNCNCM）for fault pattern recognition. The proposed method integrates the advantages of LCMMDP in dimension reduction and KNCNCM in pattern recognition and provides higher fault identification accuracy. The validity of the proposed method is verified by the instance of the fault diagnosis of a double-span rotor system dataset and simulation dataset.