Electrocardiograph (ECG) is one of the most critical physiological signals for arrhythmia diagnosis in clinical practice. In recent years, various algorithms based on deep learning have been proposed to solve the heartbeat classification problem and achieved saturated accuracy in intrapatient paradigm, but encountered performance degradation in inter-patient paradigm due to the drastic variation of ECG signals among different individuals. In this paper, we propose a novel unsupervised domain adaptation scheme to address this problem. Specifically, we first propose a robust baseline model called Multi-path Atrous Convolutional Network (MACN) to tackle ECG heartbeat classification. Further, we introduce Cluster-aligning loss and Cluster-separating loss to align the distributions of training and test data and increase the discriminability, respectively. The proposed method requires no expert annotations but a short period of unlabelled data in new records. Experimental results on the MIT-BIH database demonstrate that our scheme effectively intensifies the baseline model and achieves competitive performance with other state-of-the-arts.