Abstract: In view of bearing speed conditions and damage modes are complex and changeable, and the traditional fault diagnosis methods based on feature extraction are not satisfactory, a novel diagnosis method built on deep convolutional neural network is presented. The original vibration data is used as network input in this method, the convolutional layer of this model is applied to extract the features, then play down the dimensionality of features through the pooling layer, and the full connection layer and the classifier layer are used for fault identification. On the basis of setting reasonable network structure and parameters, the four-classification diagnosis model was established by using the fault data of under variable speed conditions and multi-mode. The diagnostic accuracy of the test sets reached 98.6%, higher than that of back propagation neural network (62.8%), support vector machine (72.9%) and shallow convolutional neural network (93.1%).