The traditional fault diagnosis methods show low accuracy and poor generalization ability in identification， and the methods based on deep learning need huge amounts of training data. In light of these shortcomings， this paper proposes an intelligent rolling bearing fault diagnosis method based on the empirical mode decomposition （EMD） and convolution neural network （CNN）. The bearing vibration data are decomposed by EMD. The intrinsic mode function （IMF） component with the largest correlation coefficient are analyzed for spectral images， which are compressed into characteristic binarization images to feed in the CNN training. Then， CNN yields a model that is used to classify and identify various faults based on the compressed images both under normal and various fault conditions . The experimental results show that the presented method surpasses traditional fault diagnosis methods when the accuracy rate of the bearing fault diagnosis with less training data increased to 97.61%， which is much larger than that of the BP neural network and probabilistic neural network method. Besides， the method presents stronger generalization ability and better anti-noise performance with an accuracy rate of 96.19% when the original signals is superposed with a 6 dB white noise.