In order to diagnose fault effectively by using vibration signal, a bearing fault diagnosis model based on neighborhood adaptive locality preserving projections was proposed. Firstly, the bearing vibration signal was decomposed into several smooth intrinsic mode functions (IMFs) by EMD and the auto-regressive (AR) model of IMF was established to construct original characteristic subset. Then, the algorithm of neighborhood adaptive locality preserving projections was used to reduce the dimension of original characteristic subset to gain low-dimension eigenvectors and projection matrix. The best reduced dimension and the best corresponding projection matrix could be determined by studying the relationship between fault recognition rate and dimension of the low-dimension eigenspace with the low-dimension eigenvectors as inputs and least square support vector machine(LS-SVM)as classifier. At last Low-dimension eigenvectors converted from original characteristic subset based on the best reduced dimension were put into LS-SVM for recognizing the conditions and fault states of bearing. Results of the test indicated that the proposed model diagnosed bearing fault with high accuracy.