Anomaly detection of satellite telemetry data based on extended dominant sets clustering

Abstract

To mine out anomalies in satellite telemetry data under unsupervised conditions, a cluster-based method is proposed in this paper. Firstly, an extended dominant sets clustering algorithm is proposed to cluster the telemetry data with arbitrary shapes. Secondly, objects that do not belong to any cluster or belong to small clusters are traditionally identified as anomalies. Thirdly, the anomalies in large clusters are detected according to the relative similarity. Finally, the information on anomaly windows in the telemetry sequence is obtained according to the local anomaly rate, which provides more characteristics of the anomalies. Experimental results show that: 1) The proposed extended dominant sets clustering algorithm can deal with the dataset containing multiple and arbitrarily shaped clusters; 2) The introduction of relative similarity increases the AUC values of anomaly detection by 3%~10%; 3) The proposed anomaly detection method can effectively detect the anomalies in magnetometer telemetry data of Tianping-2B satellite. Therefore, the proposed method can achieve anomaly detection of satellite telemetry data under unsupervised conditions, and provide support for improving satellite safety and reliability.