With the continuous development of deep learning, convolutional neural network (CNN) has been widely applied to the field of fault diagnosis. However, most of the previous methods rely on complex signal processing knowledge or feature extraction methods, and thus cannot achieve end-to-end fault diagnosis. In this paper, we put forward a multiscale dense convolutional network (MSDCN) for the fault identification of rolling bearing. First of all, a modified coarse-graining procedure is introduced to incorporate multiscale learning ability into CNN model. Then a novel dense convolutional neural network architecture (DCNN) is designed for the feature extraction of the mechanical vibration signals. Finally, an end-to-end fault diagnosis framework which is based on the improved coarse-grained process and the designed DCNN, is presented. The bearing data collected from the fault simulator is used to verify the effectiveness of the proposed method. Experiments show that the proposed approach outperforms some competitive methods in terms of diagnostic accuracy.