Abstract: Due to limitations such as the size of the convolution filter, U-net cannot learn the long timing correlation of electrocardiographic (ECG) signals and the correlation between tags. Therefore, this paper proposes a multi-target detection method for arrhythmia based on U-net-BiLSTM-CRF, which can simultaneously output the type and location of the target heartbeat. U-net was used to learn the fusion features. The fusion features were input into the bi-directional long short-term memory (BiLSTM) to learn long time-dependent features. Conditional random field (CRF) was used to model the relationship between tags to optimize the classification results. According to the heartbeat classification standard of ANSI/AAMI EC57:2012, a dataset was built in this paper based on a total of 85 609 heartbeat records in the MIT-BIH arrhythmia database, which was used to verify the proposed method. The results show that the accuracy of this method for heartbeat classification is 99.11%, the specificity is 99.76%, and the sensitivity is 97.21%, all of which are better than the classification performance of traditional U-net on the MIT-BIH arrhythmia database.