This paper presents a rolling bearing fault diagnosis method based on convolution neural network for the complicated diagnosis progress, and bad effect of traditional fault diagnosis method. First of all, different fault vibration signals which were normalized. Next, converted the one-dimensional vibration signals into two-dimensional grey image, to take advantage of the relationship between each element and its neighbors, use the method of the overlap sampling to strengthen data sets. In the convolution neural network, the tensorflow was used to build a network framework to establish network structures. 4 different convolution neural network structures were used to train the samples. In order to avoid the randomness of the experiments, train many times for each scheme, the average of the results were selected to as the optimal model. According to the accuracy of the test sets, the best model for bearing fault diagnosis was selected. At the same time, the structural parameters of the network were optimized to improve the recognition rate and operation efficiency of the model. The experimental results showed that the method can identify and classify the faults of rolling bearings well.