Anomaly Detection for Agile Satellite Attitude Control System Using Hybrid Deep-Learning Technique

Abstract

Agile low-Earth-orbit (LEO) observation satellites need a robust attitude control and determination system. It is a critical satellite subsystem, which stabilizes the satellite to different desired orientations during its mission using different actuators. The detection of satellite misorientation is a highly challenging problem because it requires continuous monitoring of data from hundreds of satellite sensors to guarantee healthy operability. In this paper, the authors propose a data-driven deep-learning framework for detecting satellite misorientation by analyzing attitude control subsystem telemetry data. The proposed approach combines a hybrid predictive deep-learning model that consists of long short-term memory and convolutional neural networks in two parallel paths to predict telemetry data and a robust isolation forest classifier for anomaly detection purposes that can classify output residuals as normal or anomalous. The hybrid model was optimized by the particle swarm optimization algorithm to ensure faster fitness function convergence with optimal model hyperparameters. The suggested data-driven model was validated using real telemetry datasets, including real anomalous case studies. The experimental results proved the suggested approach’s superiority for identifying satellite misorientation as well as helping satellite operators monitor the system’s health and deduce the causes of anomalies to aid in decision-making.