Anomaly detection based on telemetry data is a key technology for the on-orbit operation and maintenance management of satellite. However, most of the existing methods only use normal samples to build models, while the anomaly detection results are sensitive to the detection threshold, resulting in a high false positive rate. To address this problem, this paper proposes an anomaly detection method based on contrastive time-series reconstruction satellite of telemetry data, which makes full use of the prior knowledge of limited abnormal telemetry samples to enhance the differences between normal and abnormal samples. First, variational autoencoders is used to extract the time-series evolutionary characteristics of telemetry data, specifically the contrastive learning method is introduced to establish an encoder with differentiated outputs of abnormal and normal telemetry data, which uses a large amount of normal telemetry data to further train the whole model to achieve precise time-series reconstruction of normal telemetry data and form a time-series data reconstruction model sensitive to abnormal data. Then the anomaly detection threshold of satellite telemetry data is deduced based on the kernel density estimation method to further improve the detection rate of abnormal samples. Experimental verification was conducted using real satellite telemetry data and the results show that the proposed method can effectively use historical abnormal samples to establish an anomaly detection model, effectively reduce the false positive rate ( all below 0. 002) of anomaly detection and maintain a high detection rate at the same time, keeping a good practical application level.