On-board Telemetry Monitoring in Autonomous Satellites: Challenges and Opportunities
A new 'peephole' framework makes AI fault-detection on autonomous satellites explainable — a requirement for real spacecraft deployment that black-box models fail to meet.
The Thesis
Autonomous satellites need systems that can detect, isolate, and recover from hardware faults without waiting for a signal to reach Earth. The problem is that modern neural networks — the best anomaly detectors available — are opaque: they flag a problem but cannot say which component failed or why. This paper proposes a framework called 'peepholes' that extracts low-dimensional, human-readable summaries from the internal activations of a neural network, turning a black box into something an engineer can audit. The target application is the Attitude and Orbit Control Subsystem (AOCS) — the set of hardware, including reaction wheels (spinning gyroscopes that orient the spacecraft), that keeps a satellite pointed correctly. The catch is that this has only been demonstrated in simulation on a single subsystem, and the gap between a promising laboratory result and a flight-certified system is very large.
Catalyst
The commercial satellite industry is deploying increasingly autonomous spacecraft — including large constellations — where ground-in-the-loop fault response is too slow and too expensive. At the same time, space agencies such as ESA and NASA are actively writing requirements for explainable AI in flight software, creating a direct procurement pull. Edge-deployable neural inference is now practical on modern space-grade processors, making onboard ML feasible where it was not five years ago.
What's New
Prior fault-detection approaches for spacecraft either used rule-based threshold monitors — fast and explainable but brittle — or neural anomaly detectors such as autoencoders that are accurate but uninterpretable. Explainable AI (XAI) methods developed for terrestrial applications, such as SHAP and LIME, were not designed for the compute constraints or the real-time telemetry structure of embedded space systems. This paper adapts the idea of probing intermediate neural representations (a technique from natural-language-processing research) to a convolutional autoencoder trained on reaction-wheel telemetry, producing lightweight semantic indicators the authors call peepholes, and claims this adds only marginal computational overhead suitable for onboard deployment.
The Counter
This paper presents a promising concept demonstrated entirely in simulation on a single subsystem with no flight heritage. The 'peepholes' idea is essentially a probe layer — a well-known technique from NLP research — applied to a new domain, which limits the novelty claim. The paper does not provide head-to-head comparison against existing space-qualified FDIR methods on shared, publicly available datasets, making it hard to assess whether the accuracy-interpretability trade-off is actually favorable. Space software certification is a multi-year, multi-million-dollar process governed by standards like ECSS and DO-178C; a research prototype that runs at low computational cost is very far from a certifiable flight product. Perhaps most importantly, the reaction-wheel telemetry used here is relatively structured and low-dimensional — the hardest cases in real spacecraft involve subtle, multi-subsystem fault interactions that may not yield to this approach so cleanly.
Longs
- RKLB (Rocket Lab) — vertically integrated small-sat bus and software stack
- BWXT — space-grade power and hardware for autonomous deep-space missions
- KTOS (Kratos Defense) — satellite ground and autonomy software
- ARKX (ARK Space Exploration ETF) — broad exposure to autonomous spacecraft
- Lockheed Martin (LMT) — prime contractor on military autonomous satellite programs
Shorts
- Traditional spacecraft FDIR (Fault Detection, Isolation and Recovery) software vendors relying on rule-based threshold systems — their moat erodes if neural methods become both accurate and certifiable
- Ground-station-heavy satellite operations companies — autonomous onboard fault recovery reduces dependence on expensive 24/7 ground monitoring infrastructure
Enablers (Picks & Shovels)
- ESA's OPS-SAT mission and open telemetry datasets — direct source of flight-representative data for this type of research
- GreenWaves Technologies and Ubotica — providers of low-power AI inference chips qualified or being qualified for space
- PyTorch and TensorFlow Lite — open-source frameworks used to build and compress the autoencoder models described
- NASA's cFS (core Flight System) — open-source flight software architecture that could host this type of plugin
Private Watchlist
- Slingshot Aerospace — space situational awareness and satellite analytics
- Pixxel — autonomous Earth-observation satellites with onboard processing
- Orbital Sidekick — hyperspectral satellites with edge AI inference
- Cognitive Space — autonomous mission planning software for satellite operators
Resources
The Paper
The increasing autonomy of spacecraft demands fault-detection systems that are both reliable and explainable. This work addresses eXplainable Artificial Intelligence for onboard Fault Detection, Isolation and Recovery within the Attitude and Orbit Control Subsystem by introducing a framework that enhances interpretability in neural anomaly detectors. We propose a method to derive low-dimensional, semantically annotated encodings from intermediate neural activations, called peepholes. Applied to a convolutional autoencoder, the framework produces interpretable indicators that enable the identification and localization of anomalies in reaction-wheel telemetry. Peepholes analysis further reveals bias detection and supports fault localization. The proposed framework enables the semantic characterization of detected anomalies while requiring only a marginal increase in computational resources, thus supporting its feasibility for on-board deployment.