For the exiting staging methods, the accuracy is limited by insufficient feature extraction and class imbalance. To solve the problem, the residual shrinkage network is applied to design a convolutional neural network to extract feature efficiently. Meanwhile, the idea of re-weighting is used to design the loss function to address the problem that N1 stage gets low accuracy due to less samples. Finally, experiments are designed based on data of the Fpz-Cz and Pz-Oz channel in the Sleep-EDF dataset. The accuracy rates are 85. 4% and 82. 2% , respectively. The MF1 values are 79. 6% and 75. 4% , respectively. Results show that the method achieves higher accuracy and MF1 than the benchmark algorithm and current advanced comparison algorithms. It proves the effectiveness and advancement of the proposed algorithm.