Deep Learning-Based Anomaly Detection in Spacecraft Telemetry on Edge Devices
Spacecraft anomaly detection is a critical component for ensuring mission safety in the field of aerospace engineering. As missions grow more complex, the need for sophisticated models capable of real-time anomaly detection becomes paramount. However, deploying such models on-board spacecraft poses significant challenges due to hardware constraints inherent in edge devices.
This article discusses a recent study published under arXiv:2603.29375v1, which investigates three distinct approaches for spacecraft telemetry anomaly detection: forecasting & threshold, direct classification, and image classification. The research emphasizes the optimization of these methodologies for edge deployment by utilizing multi-objective neural architecture optimization, specifically focusing on the European Space Agency Anomaly Dataset.
Key Approaches to Anomaly Detection
- Forecasting & Threshold: This approach leverages predictive modeling to anticipate telemetry values, setting thresholds to identify deviations from expected behavior.
- Direct Classification: This method involves training models to classify telemetry data directly, identifying whether a given set of data points indicate an anomaly.
- Image Classification: Utilizing visual representations of telemetry data, this technique applies image recognition methods to detect anomalies.
Performance and Optimization
The baseline experiments conducted as part of this study indicated that the forecasting & threshold method achieved the highest detection performance, recording a Corrected Event-wise F0.5-score (CEF0.5) of 92.7%. This performance was notably superior when compared to the alternative methods explored in the research.
Through the application of Pareto-optimal architecture optimization, the researchers were able to significantly reduce computational requirements without sacrificing model capabilities. The optimized forecasting & threshold model maintained 88.8% of the original CEF0.5 score while dramatically decreasing RAM usage by 97.1%, bringing it down to just 59 KB. Additionally, operations were reduced by an impressive 99.4%.
Deployment Viability
Analysis of the deployment viability for these optimized models suggests that they require only a fraction of the RAM available in CubeSat systems, ranging from just 0.36% to 6.25%. This finding indicates that on-board anomaly detection can be practical, even in environments with highly constrained hardware resources.
Conclusion
This research demonstrates that it is indeed possible to deploy sophisticated anomaly detection capabilities within the constraints of spacecraft edge computing. By enabling near-instantaneous detection of anomalies without exceeding hardware limitations, this study contributes significantly to the safety and reliability of space missions. The findings pave the way for future advancements in spacecraft telemetry monitoring, ensuring that missions can be conducted with greater confidence in their safety measures.