The key of rolling bearing fault diagnosis is the extraction of sensitive fault features. Multiscale fuzzy entropy (MFE) is an effective analysis method for complexity measurement of time series and has been used for fault features extraction from rolling bearing vibration signals. Considering the defects existed in the MFE coarse graining , its process is replaced by sliding average and then an improved MFE algorithm is proposed in this paper. It is also compared with MFE by using simulation signal analysis. In this case, a new fault diagnosis method for rolling bearing is proposed based on the improved multiscale fuzzy entropy and support vector machine. Finally, the proposed fault diagnosis method is applied to data analysis of rolling bearing experiment by comparing with the traditional MFE method, and the analysis results verify the effectiveness and superiority of the proposed method.