Rolling bearing prognostic and health management (PHM) method can extract a large number of fault characterization data. Those data are of great potential value, because of their characteristics of high-dimensionality and high-redundancy. However, direct analysis and utilization of them are impossible. Therefore, aiming at reducing the redundancy data and screening sensitive features, a two-stage feature selection algorithm is proposed. In the first stage of the method, the Laplacian score (LS) is used to sort the original features based on their locality preserving power, and the mutual information-based clustering algorithm is utilized to remove the redundant features of the original feature set. In the second stage, the Mahalanobis-Taguchi system (MTS), as a useful multivariate pattern recognition method, is employed to comprehensively evaluate the remaining features, unearthing features which are prone to fault classification. The verification results of the bearing degradation simulation test data show that the proposed two-stage feature selection algorithm can effectively remove redundancy and improve the accuracy of fault monitoring. This method can be effectively applied to the initial fault detection of rolling bearings.