<p>Anomaly detection techniques can effectively identify observations that deviate from normal behavior patterns in time series. The current reconstruction-based unsupervised anomaly detection methods still face some challenges, such as key feature extraction and data distribution drift. To address these issues, we propose an Efficient Channel Attention-based Time-frequency Masked AutoEncoder model (ECAT-MAE) for anomaly detection in time series. The proposed method innovatively integrates time-frequency analysis and channel attention mechanisms through comparative criteria within the dual-branch transformer autoencoder architecture. First, an efficient channel attention module is developed via lightweight convolutional operations to dynamically calibrate channel weights and suppress redundant features. Subsequently, a weighted time-frequency masking strategy guided by the recalibrated channel weights is employed to facilitate the learning of unbiased representations of normal patterns, thereby enhancing sensitivity to critical time-frequency features. Finally, we employ a time-frequency consistency contrastive loss function combined with an adversarial training strategy to prevent overfitting and thereby alleviate distribution shift issues. The experimental results indicate that the proposed anomaly detection method achieves higher detection accuracy across several benchmark datasets.</p>