Aiming at the problem of faults’ being difficult to identify due to their characteristics intersecting, a new measure function called the neighbor probability distance was proposed considering the probability for adjacent points to become neighbor ones. The neighbor probability distance was applied into the locality preserving projection (LPP) algorithm and K-nearest neighbor (KNN) classifier, then the neighbor probability distance-based locality preserving projection (NPDLPP) algorithm and the neighbor probability distance-based K-nearest neighbor (NPDKNN) classifier were proposed. Firstly, vibration signals were converted into high-dimensional data sets using feature-extracting methods in time domain and frequency domain. Then high-dimensional data sets were projected into a lower-dimensional space with NPDLPP. Finally, lower-dimensional sensitive feature sets obtained with dimension reduction were input into NPDKNN for pattern recognition. The vibration signal sets of a double-span rotor system were used to verify the proposed method. It was shown that the proposed method has an obvious effect to better separate various fault types; compared with the traditional Euclidean distance measure, the neighbor probability distance can minimize the divergence within a class and maximize the separation between classes.