In order to improve the accuracy of bearing fault feature extraction and operation condition evaluation，a diagnosis method based on wavelet packet dispersion entropy and Meanshift probability density estimation is proposed. Firstly，wavelet packet transform is used to increase the dimension of bearing vibration signal data，and the dispersion entropy (DE) of each sub-band is calculated to construct the characteristic matrix. Then，PCA is used to reduce the dimension of multi-dimensional matrix visually. Meanshift nonparametric estimation is used to obtain the maximum probability density position of training samples as the clustering center. Finally，the Euclidean distance between the test sample distribution entropy coordinates and each cluster center is calculated to determine the test sample category. The experimental data of CWRU and QPZZ-II are used to verify the effectiveness of the proposed method for identifying different fault types and fault degrees. Due to the complete theoretical model of wavelet packet and the ability of signal band decomposition sparsity，combined with the good robustness of the DE index，the constructed feature matrix has good aggregation and large inter-class distance. At the same time，Meanshift aims at maximizing probability density，and can effectively classify different data samples by adaptive iterative clustering center and membership.