Abstract: The vibration signals of rolling bearings in the actual operation of roller crusher are seriously coupled, and the fault pulse tends to be weak.Thus, the fault identification becomes difficult, especially, in the fault distinguish between the rolling body and the inner ring.A feature extraction method of bearing fault based on time-frequency image has been put forward by introducing the image texture feature extraction technology to the fault diagnosis.Firstly, combing the Ensemble Empirical Mode Decomposition (EEMD) with Wigner-Ville Dis-tribution (WVD) can obtain the vibration frequency representation with no cross-term interference.Secondly, the corresponding Local Binary Patterns spectrum can be generated by introducing the LBP, which help to enhance the texture features in time-frequency grayscale images.Thirdly, the LBP gray histogram can be taken as feature quantity, and reduced by means of PCA after compressing the feature dimension.Finally, the low-dimensional characteristic quantity can input into BP neural network for fault classification.In the experiment of bearing fault diagnosis, the method proposed above is proved to have a high fault recognition accuracy through comparing with other algorithms, and its accuracy of 99.5% fully demonstrates the effectiveness of the method.This method is reliable to accurately extract the bearing fault characteristics of motor in the gear roller crusher.