The electrocardiogram signal (ECG) intuitively reflects the physiologically electrical activities of heart, and has important reference value in diagnosing heart diseases. In this paper, we proposed a kind of twoclass classification method for ECG signals using convolutional neural networks. The network convolution layer used different convolution kernels to maximize the use of local features for classification and detection of abnormal heart beats. The method has utilized the MITBIH Arrhythmia Database proposed by Massachusetts Institute of Technology. Calculating performance metrics through confusion matrix and applying crossvalidation against three traditional machine learning methods, experiments show that the model accuracy rate can even reach 9686%, which increases 339%, compared with the support vector machine dichotomy method with the highest accuracy performance. This method simplified the feature extraction process and fully improved the accuracy of abnormal heartbeat detection.