•ECG episodes are repeated and resampled for augmenting small datasets.•The data augmentation balances samples among different categories.•The augmentation enhances deep neural networks in atrial fibrillation detection. Automated detection of atrial fibrillation (AF) from electrocardiogram (ECG) recordings remains challenging in real clinical settings. Deep neural networks (DNN) emerge as a promising tool for the task of AF detection. However, the success of DNN for AF detection is hampered by limited size and imbalanced number of samples in datasets. We propose a novel data augmentation strategy based on duplication, concatenation and resampling of ECG episodes to balance the number of samples among different categories as well as to increase the diversity of samples. The performance of the data augmentation method was examined on an AF database from Computing in Cardiology (CinC) challenge 2017. A 2-layer long short-term memory (LSTM) network was trained with the augmented dataset. Its ability of AF detection was evaluated using a 10-fold cross validation approach. And F1 score was adopted as the metrics. The AF detection results show that the proposed method was superior to two conventional data augmentation methods: window slicing and permutation. The network was also submitted to the evaluation system of the CinC challenge 2017. The F1 score obtained by the network using the proposed data augmentation method was close to the winner (0.82 vs. 0.83). In summary, the proposed data augmentation method provides an effective solution to enhance the dataset for improving the performance of DNN in ECG analysis. Such a method promotes the application of deep learning in the analysis of ECG, particularly when the dataset is small and imbalanced.