Based on one-dimensional ECG signals, an improved convolution bidirectional long short-term memory network has been proposed for a realization of an automatic classification of arrhythmias. On the basis of the convolutional neural network (CNN) as well as its attention mechanism, with key features extracted, the bi-directional long short memory network (BiLSTM) is built to mine the temporal correlation of ECG signals, thus finally realizing the automatic classification of ECG signals. The experimental results on MIT-BIH arrhythmia data-set show that the proposed method achieves an improvement of rare category classification based on the overall accuracy of 99.32%, with the accuracy of S and F classification improved by 1.02% and 10.07% respectively; meanwhile, with the recall rate increased by 12.52% and 4.25% respectively as well, thus meeting the requirements of automatic classification of arrhythmia.