Atrial fibrillation (AF) is the most common persistent arrhythmia, and it is crucial to develop generalizable automatic AF detection methods. However, supervised AF detection is often limited in performance due to the difficulty in obtaining labeled data. To address the gap between limited labeled data and the requirements for model robustness and generalization in single-lead ECG AF detection, we proposed a semi-supervised contrastive learning method named MLMCL for AF detection. The MLMCL method utilizes the multi-level feature representations of the encoder to perform multiple contrastive learning to fully exploit temporal consistency, channel consistency, and label consistency. Meanwhile, it combines labeled and unlabeled data for pre-training to obtain robust features for downstream tasks. In addition, it uses the domain knowledge in the field of AF diagnosis for domain knowledge augmentation to generate hard samples and improve the distinguishability of ECG representations. In the cross-dataset testing mode, MLMCL had better performance and good stability on different test sets, demonstrating its effectiveness and robustness in the AF detection task. The comparison results with existing studies show that MLMCL outperformed existing methods in external tests. The MLMCL method can be extended and applied to multi-lead scenarios and has reference significance for the development of contrastive learning methods for other arrhythmia.