Cardiovascular disease is one of the main causes of human death. Based on the modified residual network, we identify ECG signals and combine the modified residual network with dilated convolution to extract global information as much as possible while keeping local information unchanged in feature extraction. The MIT-BIH arrhythmia data set is trained, validated and tested using K-fold cross validation. In the experiment, firstly, the convolution layer is used to collect the input images. Secondly, the modified network is used to extract the features. Finally, the Softmax classifier is used for classification. In the MIT-BIH arrhythmia database, the proposed model achieves 97.20% accuracy, 92.85% sensitivity, 98.29% specificity, 93.16% accuracy and 93.00% F₁ score without any additional artificial features and data augmentation. This research will provide technical support for the detection and recognition of ECG signals to reduce the workload of professional doctors in medical institutions.