Abstract: According to statistics, arrhythmia is the main cause of sudden cardiac death. An improved one-dimensional U-shaped network is proposed to classify and recognize the segmented ECG signals. The network is based on the MIT-BIHarrhythmia database, and adopts the classification standard established by the Association for the Advancement of Medical Instrumentation (AAMI). The improved 1D-Unet selects the appropriate convolution layer and convolution kernel.In the process of upsampling and splicing, try to preserve the characteristics of the original signal. In the case of ECG signal wavelet denoising, the recognition accuracy rate reaches 99.46%, and the F1 score is 97.57%. The average accuracy of network classification is 99.73%, and the precision is 98.23%, the sensitivity is 96.92%, the specificity is 99.17%. The accuracy of network recognition without denoising is 99.4%, and the F1 score is 97.12%. Due to the network is a full convolutional network, the output layer doesn't use the classical full connection layer, the parameters of the neural network are greatly reduced. The network is very helpful to diagnose arrhythmia by ECG.