<p>The basic purpose of multivariate time series anomaly detection is to identify anomalous data points from a large amount of normal data. However, building a model capable of detecting anomalies in real-time in the Internet era is a major challenge. This is due to the enormous amount of data available today and the complex patterns of dependencies between data. Traditional methods are insufficient to discern interdependencies between temporal sequences in large-scale datasets. Although the emergence of deep learning has driven the rapid development of anomaly detection tasks, there is still room for improvement in capturing the correlations between temporal sequences from multiple perspectives. To address these issues, this paper proposes a novel Time-Frequency Augmented Perception Method for Multivariate Time Series Anomaly Detection (TFAP). TFAP contains a temporal analyzer, a frequency analyzer, as well as a two-branch contrastive learning and adversarial training mechanism. Among them, the time analyzer uses a decoupling module to obtain stable input data. The multi-granularity information capture module acquires the features of the input data from local and global perspectives. Meanwhile, the data in the frequency analyzer is converted into frequency domain information by Fast Fourier transform. The Contrastive Learning mechanism exploits the consistent relationship between different perspectives of the same data, enabling the model to fully learn the features inherent in the data itself and reduce the interference of irrelevant information. The purpose of adversarial training is to prevent overfitting during the training procedure. Our method demonstrates superior performance compared to 16 baseline methods.</p>