Розробка гібридної бортової діагностичної архітектури з легким машинним навчанням для систем CubeSat з обмеженими ресурсами

The object of the study is the onboard diagnostic process in CubeSat nanosatellite systems, with particular emphasis on fault detection based on real-time analysis of multivariate telemetry data under real-time embedded operating conditions. The problem statement is the gap between the relatively low computational expenses and predictability of classical FDIR approaches and their limited ability to detect complex anomalies, and, conversely, the greater anomaly sensitivity and lower embedded performance of machine-learning models. A hybrid onboard diagnostic architecture that integrates telemetry acquisition, telemetry preprocessing, statistical feature extraction, a deterministic FDIR branch, a lightweight machine-learning branch, and decision fusion into one onboard diagnostics session has been elaborated. The proposed hybrid onboard diagnostic architecture has been evaluated based on hardware-in-the-loop testing using representative telemetry and fault cases for CubeSat systems. The evaluation metrics were the fault-detection accuracy, the detection latency, the CPU consumption, the memory consumption, and the resistance to telemetry noise. The hybrid onboard diagnostic architecture was compared to the conventional threshold-based FDIR system and standalone machine learning algorithms under the same HIL conditions. In the reported experiment setup, the proposed hybrid onboard diagnostic architecture demonstrated the fault-detection accuracy of 94%, while the classical FDIR method provided 71%, and the standalone machine-learning approach provided 88%. The average detection latency was decreased to 83 ms, in contrast to 120 ms and 95 ms, respectively. The embedded solution requires only 17.6% CPU and 58 KB of memory. Under the highest telemetry-noise level, the fault-detection accuracy of the proposed hybrid onboard diagnostic architecture decreases to 80%, whereas the standalone machine-learning and the classical FDIR baselines provide only 65% and 43% fault-detection accuracy, respectively.