Aiming at extracting rotating machinery fault characteristics submerged in strong background noise, a new method based on refined composite multiscale fuzzy entropy and locality preserving projection (LPP) is proposed for fault diagnosis. Firstly, by introducing slip construction short-time series and refined composite multiscale fuzzy entropy, The feature information in different scale and the fault potential characteristics can be acquired, which accurately describe the complexity and uncertainty of the vibration signal. Secondly, LPP is applied to reduce dimension and retain local signal feature. Then the designed optimized bandpass filter successfully extract the fault feature of the rolling bearing, which is separately verified by the simulation signal and experimental data. It is verified by the simulation and experimental result that the proposed method shows better performance and advantage in restraining noise and recognizing weak shock features of rolling bearing.