Introduction

Satellites in Low Earth Orbit (LEO) operate in harsh, unpredictable environments where component degradation or cyber-attacks can lead to catastrophic failure. Traditional monitoring relies on ground-based analysis of telemetry logs, introducing critical delays.

The ArkSpace project explores using Spiking Neural Networks (SNNs) for onboard anomaly detection. By processing telemetry data directly on the satellite’s neuromorphic payload, we can identify issues—from battery degradation (“Satellite-Battery” anomalies) to spectrum interference—in milliseconds rather than minutes.

Why SNNs for Anomaly Detection?

SNNs are fundamentally different from traditional Deep Learning models. They process information as sparse, asynchronous events over time. This makes them uniquely suited for monitoring satellite subsystems:

1. Temporal Precision: Satellite failures often manifest as subtle timing glitches or transient spikes in voltage/current. SNNs inherently encode time, making them hypersensitive to these temporal patterns.
2. Unsupervised Learning: Through Spike-Timing-Dependent Plasticity (STDP), SNNs can learn the “normal” baseline of satellite operation without needing labeled training datasets. Anything deviating from this baseline triggers an alert.
3. Low Power: Monitoring runs 24/7. SNNs on chips like Intel Loihi 2 consume milliwatts during nominal operation, spiking in power only when an anomaly (event) occurs.

Key Use Cases

1. Battery Health Monitoring

Li-ion batteries are the lifeblood of a CubeSat. Anomalies like dendritic growth or capacity fade often show precursor signals in the discharge voltage curve. An SNN can monitor these curves in real-time, detecting deviations that indicate imminent failure (“Satellite-Battery” faults). Research suggests SNNs can predict remaining useful life (RUL) with higher accuracy than Kalman filters in dynamic load scenarios.

2. Interference and Attack Detection

As highlighted in recent research (Lagunas et al., 2024), SNNs excel at detecting RF interference. This capability extends to cyber-security. “Satellite-Based” attacks, such as jamming or protocol fuzzing, create anomalous traffic patterns. An onboard SNN can fingerprint these patterns and trigger defensive countermeasures (e.g., frequency hopping) instantly.

3. Attitude Control Stability

Reaction wheels and magnetorquers degrade mechanically. SNNs can ingest data from star trackers and gyroscopes to detect micro-vibrations or friction increases that precede mechanical failure.

Implementation in ArkSpace

The Exocortex Constellation dedicates a portion of its 100-million-neuron payload to self-monitoring:

• Input Layer: Direct feed from the satellite bus (voltage, temperature, current, spectral density).
• Processing: A “Guardian” SNN network trained on nominal telemetry.
• Output: Prediction error signal. If the predicted state deviates from the actual state, an interrupt is sent to the main flight computer.

Conclusion

Integrating SNN-based anomaly detection transforms a satellite from a passive machine into a self-aware system. By detecting “Satellite-Based” faults and attacks onboard, we increase the resilience of the entire constellation. This capability is essential for LEO networks where ground contact is intermittent and light-speed latency prevents real-time human intervention.

---

Official Sources

1. General SNN Anomaly Detection: M. Davies et al., “Advancing Neuromorphic Computing with Loihi 2,” Intel Labs, 2021.
2. Battery Monitoring: Wu, J., et al. (2020). “Deep Neural Network for Lithium-Ion Battery State of Health Estimation.” (Contextual reference for AI in battery monitoring).
3. RF Anomaly Detection: Lagunas, E., et al. (2024). “Performance Evaluation of Neuromorphic Hardware for Onboard Satellite Communication Applications.” arXiv:2401.06911.
4. ArkSpace Architecture: arkspace-core/docs/architecture/satellite-node.md