The accumulation of electrocardiogram (ECG) data and the absence of labeled samples make ECG anomaly detection a self-supervised deep learning task. Traditional ECG augmented samples, mainly generated by jittering, permutation, and flipping etc., are similar to abnormal ECG samples, causing abnormal samples to be incorrectly drawn closer to normal samples, triggering issues of ambiguous boundaries. Motivated by this, we design a feature augmentation generator to generate feature-augmented samples by sampling the latent variables within the feature space of examples, and constructing feature-level positive pairs. Specifically, the method contains time encoder, frequency encoder, intra-time-frequency module, and inter-time-frequency module, respectively. Time and frequency encoder take the original ECG and its frequency counterpart as inputs to learn time-frequency representations. Then in intra-time-frequency module, the time-frequency fused features are fed into feature augmentation generator to generate augmentation feature, constructing feature level positive sample pairs to learn intra-time-frequency data anomaly distribution, and inter-time-frequency module makes time-based and frequency-based representation adjacent to each other to learn the inter-time-frequency data anomaly distribution. We evaluate the new method on six datasets, including atrial fibrillation detection, ventricular arrhythmia detection and ECG quality assessment, demonstrating that TF-A outperforms against ten state-of-the-art methods.