In view of the difficulty of identifying different fault states of bearings, the feature selection method was applied to rolling bearing fault diagnosis. A feature selection method based on nonparametric mutual information(NPMI) was presented on the basis of mutual information method. First, the statistical characteristics in time and frequency domains that could represent the change of the bearing state were extracted from the original signal and the multidomain feature set was set up. Then, the NPMI feature selection method was used to remove the unrelated features and redundant features in the feature set, the sensitive feature set was established, and multidimensional scaling was used to visualize the sensitive feature sets, then the classification and clustering ability of the features were compared. Finally, the feature vectors after dimension reduction were input into the support vector machine to get the recognition results of different faults. Based on the accuracy of classifier, the validity and superiority of feature selection method based on nonparametric mutual information was verified.