Anomaly Detection for Spacecraft using Hierarchical Agglomerative Clustering based on Maximal Information Coefficient

The spacecraft's telemetry data is the only basis for the ground transportation management system to monitor its on-orbit operating status. Anomaly detection of spacecraft has become an important means to enhance the reliability of spacecraft on-orbit operation. There are many ways to detect anomalies in spacecraft. With the increasing amount of telemetry data and the improvement of modern computing capabilities, anomaly detection methods have gradually transitioned to data-driven methods. Because the data-driven approach does not require a large amount of expert experience, it also tolerates that operators do not have sufficient theoretical knowledge. However, telemetry data has the characteristics of large scale, high dimensions, complex relationships, and strong professionalism. These bring severe challenges to achieve high detection rates, low false detection rates, and strong interpretive goals for anomaly detection methods. Current spacecraft monitoring systems generally only perform anomaly detection for a single parameter, and few studies have provided clear and effective methods for multivariate anomaly detection. This paper proposes an anomaly detection method for multivariate telemetry data. The idea is to propose a hierarchical clustering method based on the maximum information coefficient, mining the correlation between telemetry parameters, grouping the telemetry parameters to form a subspace; using the LSTM method to perform single-parameter anomaly detection on the subspace; using weighting The averaging method integrates the anomaly detection results in the subspace to achieve multivariate anomaly detection. The experiments were performed on a real satellite historical data set of the Beijing Aerospace Flight Control Center. The expert evaluation of the agency proves that the method proposed in this paper is feasible and can preliminary excavate the correlation between telemetry parameters. Although the accuracy needs to be improved, there is still room for optimization.