To solve the problem of sensitive feature extraction from the non-stationary and nonlinear vibration signals of rolling bearing, local mean decomposition (LMD) was carried out. and the time/frequency domain features were extracted from the sensitive fault components quantified by the correlation coefficient method. Then, the feature sets of different faults states were established and used to train the state classifier. In order to achieve the higher accuracy of bearing fault states identification, an improved K-nearest neighbor classifier (KNNC) algorithm based on dichotomy K-means clustering was proposed, in which the big training samples were simplified, and the bad samples and interference points were effectively removed. Finally, the effectiveness of the method was verified through diagnostic analysis of experimental data of bearings.