Automated One-Sided Learning Fault Detection System for Reaction Wheel Bearing Friction Anomalies

Monitoring a satellite’s health during a mission has become critical in space engineering. In-flight anomaly detection is difficult for ground operators, placing space missions at risk of failure. Machine learning algorithms are data-driven methods that could autonomously detect faults in situ. In this paper, a new application of machine learning algorithms in space engineering is introduced for detecting reaction wheel bearing anomalies that only relies on nominal data (no failure data) for training and with no prior knowledge of the system dynamics. Using a one-sided regression method, an automated fault detection system was designed to monitor the attitude dynamics control system for a small satellite. The proposed detection algorithm was first trained using a simulated attitude dynamics control system for the small satellite. Next, the detection system was trained with only nominal behavioral data of the control system for a designated period of time. Then, different types of bearing friction failures were added to the simulated system to test the trained fault detection system. The empirical rule (68-95-99.7 rule) was used as a failure detection criterion to differentiate failure data from nominal. Similar physical tests were conducted using a combination of a brushless motor and drone propellers. Both simulation and experimental results demonstrated the robustness of detection accuracy, were model-free, and verified the feasibility of an easy-to-use, accurate, and autonomous anomaly detection system for reaction wheels that could be extended to other space systems.