In order to solve the problem of weak fault feature extraction in one-dimensional vibration signal of rolling bearing; In order to solve the problem that the deepening of deep learning model layer is easy to lead to the disappearance of gradient or the deterioration of gradient explosion, which leads to the low accuracy and poor robustness of fault diagnosis, this paper proposes a rolling bearing fault diagnosis method based on EMD-GAF and improved SERE-DenseNet. One-dimensional vibration signals of rolling bearings were decomposed and reconstructed by EMD after rolling sampling, and the reconstructed one-dimensional signals were converted into two-dimensional images by GAF as model input. DenseNet121 was selected as the main task in terms of model, and SERE module was introduced. The Dense Layer with 2 convolution layers is improved into 3 sparse modules with base number of 8. Feature extraction and fault classification are carried out by using 2D image as input. The bearing data set of Case Western Reserve University was used for simulation experiments. The experimental results show that the proposed method can accurately diagnose rolling bearings, with the maximum accuracy of 100% and the average accuracy of 99.91% in 10 experiments. Compared with the common deep learning model, the proposed method has great advantages. The fault diagnosis accuracy is 96.48% when the signal to noise ratio is 10 dB, and the proposed method has strong robustness.