Objective:To improve the performance of ECG arrhythmia classification algorithm and provide auxiliary basis for clinical ECG diagnosis.Methods:The one-dimensional ECG data was segmented according to the R point, and the segmented data was generated into a 2D image. The samples were expanded by data augmentation technology, and the image features were extracted by the 2D convolutional layer, 2D maximum pooling layer, Flatten layer and fully connected layer in 2D-CNN. Then, the samples were classified with Softmax classifier. The loss function with weight coefficients was used to enhance the model's learning of class S and class V. The MIT-BIH data set was used for model training and algorithm performance evaluation.Results:Sample expansion and the use of loss functions with weight coefficients can improve the recall rate and specificity index of the model, while maintaining the model's accuracy index of the classificatio on VEB and SVEB.Conclusions:The accuracy of the proposed model is 99.02%, and the recall rate of SVEB is 96.4%, indicating that this classification method can assist medical staff in diagnosing heart diseases.