Anomaly detection plays a key role in the reliable and efficient operation of the satellite. However, most current algorithms tend to focus more on modeling the data in the time domain for anomaly detection. The anomaly detection model lacks sufficient learning of frequency features, leading to false alarms or missed detections. To address these challenges, this paper proposes an anomaly detection approach that integrates contrastive learning with time-frequency augmentation into a data-driven reconstruction framework. First, the encoder is pre-trained through time-frequency contrastive learning, where temporal and frequency views of the telemetry sequences are exploited to obtain discriminative feature representations. Then, a decoder is appended to the pre-trained encoder, and the entire autoencoder is retrained by minimizing the reconstruction loss to enhance its representation and reconstruction capability. Finally, anomaly detection is performed by computing the reconstruction error of test samples, where deviations beyond a predefined threshold are identified as anomalies. Experimental results demonstrate that the proposed anomaly detection method achieves superior detection performance on the real satellite telemetry data set.

anomaly detection,time-frequency contrastive learning,reconstruction model