Autonomous Fault Detection, Isolation, and Recovery (FDIR) is essential for the safety and continuity of deep-space missions, particularly where bandwidth constraints and communication delays limit the effectiveness of ground-based monitoring. This paper presents a near production-ready Machine Learning (ML) system designed for on-board anomaly detection within the HERA mission, specifically targeting the Reaction Wheel Assembly subsystem. Moving beyond traditional Out-of-Limits (OOL) alarms and previous two-stage proof-of-concepts, we evaluate and implement single-stage deep learning architectures, determining that a Convolutional Soft-Boundary Deep Support Vector Data Description (Deep SVDD) model yields optimal performance (F1 score of 0.84) on flight telemetry. To bridge the gap between theoretical ML and space-grade deployment, we address specific hardware constraints of the radiation-hardened LEON 3 processor. We introduce a novel patch for TensorFlow Lite for Microcontrollers (TFLM) that enables robust Big-Endian support and strict static memory allocation, ensuring compliance with space software safety standards. Furthermore, we integrate Uncertainty Quantification (UQ) using Conformal Prediction to establish statistically valid confidence bounds, offering "Reckless" and "Cautious" decision strategies to manage the trade-off between sensitivity and false alarm rates. The final system demonstrates a minimal memory footprint (~2.1 MB) and efficient inference latency (~123 ms), providing a reusable and scalable framework for future ESA missions.
Voldřich et al. (Tue,) studied this question.