Arrhythmia disease can be extremely damaging to the heart, and in severe cases can even lead to death. The ECG smart monitoring device is an effective way for detecting arrhythmia disease, and as wearable devices spreads, it also places certain requirement on lightweight arrhythmia detection algorithms. It is of great importance to implement an efficient arrhythmia detection algorithm with strong generalization performance. This work trains an arrhythmia detection model on the Georgia 12-lead ECG Challenge (G12EC) database and the China Physiological Signaling Challenge 2018 (CPSC2018) database using xResNet18 as the backbone network and momentum contrast learning as the framework, which allows contrast learning of positive samples and a large number of negative samples by introducing queue and momentum update encoder parameters to obtain a more comprehensive information representation. The model was pre-trained using the Georgia 12-lead ECG Challenge (G12EC) Database to obtain better characterization of initialization information and fine-tuned using the China Physiological Signal Challenge 2018 (CPSC2018) database to perform arrhythmia classification test. Among them, the CPSC2018 database contains arrhythmia ECG data in nine sample proportions in a balanced manner. During the pre-trained data enhancement, we added Gaussian noise of different strengths to the signal and compared the performance of the model in the results section. The experimental results showed that the model was effective with an AUC of 0.861, an Acc of 77.04% on the CPSC2018 database.