Research/ SPACE AI & AUTONOMY
A Hybrid Model and Data Driven Approach for Anomaly Detection in Space Power Systems
M A Carbone, M J Muscatello, A McCormick, J T Csank, C ATeubert, J D Frank
Autonomy, Task Planning & OperationsNTRS keyword: Fault Detection and Diagnosis · Learning-based MethodsNTRS keyword: Machine Learning
NTRSElectronics and Electrical Engineering
The future of crewed deep space exploration requires advanced system awareness and monitoring techniques to ensure the reliability of critical subsystems such as the electric power system (EPS). Currently, expert systems are used to detect and isolate large abrupt faults such as short circuits are sufficient to safe the power system. More nuanced and subtle fault behavior is analyzed by ground control, where corrective and troubleshooting actions can be made remotely or by the onboard crew. To ensure the safety of the crew, onboard computers are tasked with the detection and diagnosis of faults and disturbances onboard the vehicle. Spacecraft experience additional failure modes on a regular basis after launch that were unplanned, unexpected, or ambiguous, and NASA has identified improved software as a means of lowering the number of failures onboard spacecraft. This paper proposes a novel combination of model-based and data-driven techniques to monitor and identify known normal, faulty, and anomalous behavior. The method is demonstrated against a high-fidelity simulation of the spacecraft EPS and validated against real hardware.