Traditional intelligent diagnosis methods rely too much on the experience of signal processing and fault diagnosis to extract fault features, and generalization ability of models is poor. Based on the theory of deep learning, a convolutional neural network algorithm combined with the softmax classifier is proposed to introduce weighting to the solution of data set imbalance problem. Model optimization techniques such as weighted loss function, regularization, and batch normalization are applied to the construction of an improved deep convolutional neural network model for rolling bearing fault diagnosis. The model learns from the original measured bearing vibration signal by layer-by-layer learning to achieve feature extraction and target classification. Experimental results show that the optimized deep learning model can achieve accurate recognition of early weak faults and different levels of faults, and its recognition accuracy on unbalanced data sets can reach 95%. Furthermore the model has faster convergence speed and strong generalization ability.