Automated anomaly detection ground station for small satellite attitude dynamics control system
MetadataShow full item record
This thesis presents an automated ground-based anomaly detection system for an attitude dynamics system of a small satellite. In a space mission, a ground station works as the operation center to communicate and monitor the satellite. Human operators are required to stay at the ground station at all times to collect, track, and analyze the performance of the satellite. With the growing popularity of the Artificial Intelligence (AI) application in recent years, processes relating to data collection, transmission, and analysis have been replaced by machine learning which reduces human labour, expense and time, whereas shortcomings in the traditional human expert systems are improving, and applications to space systems fault detection are now possible. Therefore, in this research, machine learning algorithms were applied to accomplish the proposed anomaly detection system. The detection system aimed to detect subtle failures in the spacecraft's attitude dynamics system, mainly in the reaction wheel subsystem, by only learning from nominal behavioural data from the spacecraft. The system was developed from a small satellite attitude dynamics control system using reaction wheels that may exhibit bearing failures. There were two types of anomaly detection systems introduced, including a two-sided learning anomaly detection system and a one-sided learning anomaly detection system. For this study, I first developed the two-sided learning anomaly detection system using the Logistic Regression (LR) method, which provided a background of how the training process would be undertaken using a machine learning method. By learning from both nominal and failure behaviours from the satellite, the system was expected to detect small reaction wheel friction failures. Then, the one-sided learning anomaly detection system was built by only learning from nominal behaviours from the satellite and was expected to detect the same reaction wheel failures. The methods used to develope the one-sided learning system were: One-Class Support Vector Machine (OC-SVM) and One-Class Linear Regression (OC-LR). Two simulation tests with different friction failures were given to the two-sided and one-sided learning systems. Detection performance for each system was discussed. After demonstrating the proposed system in simulation, the one-sided learning system was verified by a real motor test. Similar to the simulations, the detection system was only trained by nominal behaviours from the motor and was expected to detect friction failures to distinguish between normal and abnormal motor motion. The simulation test and the motor anomaly test illustrated the feasibility and generality of the proposed one-sided learning fault detection system for space systems.
- FGS - Electronic Theses and Practica 
- Manitoba Heritage Theses