In order to improve the fault diagnosis rate of rolling bearings, make full use of the difference in the recognition ability of the bearing operating state by the time domain, frequency domain and frequency domain features, and take into account that the features are prone to irrelevance, redundant interference and other issues, as well as the simple, fast and effective feature evaluation of the actual project method needs. Position optimized Fisher distance measure ( POFDM) method is proposed and applied to bearing fault characteristic select. The method is based on Fisher’s criterion, and the positional relationship between multi-class samples is used to correct the evaluation coefficient by the median method, which could reflect sensitivity of the state separation and aggregation. Thus, the features that can suppress the degree of state coincidence are selected. In addition, aiming at the problem that the intelligent diagnosis model is inefficient in seeking optimal feature set, feature set evaluation method based on multi-dimensional spatial measure-Fisher is proposed. The optimal feature set is selected based on the maximum value principle by calculating the distance measure index of different dimension candidate feature sets in multidimensional space. Finally, the proposed algorithm is verified by the bearing fault experiment. The experimental results show that the optimal low-dimensional feature set obtained by the proposed method achieves 99. 17% diagnostic accuracy of the SVM classifier when the number of feature combinations is 3, which can effectively diagnose bearing faults.