Reproduce “Deep Learning-Based Anomaly Detection in Spacecraft Telemetry on Edge Devices” on OpenResearch

About this paper

Christopher Goetze, Tim Schlippe, Daniel Lakey

Researchers from IU International University of Applied Sciences optimized deep learning models to perform anomaly detection on spacecraft telemetry directly on edge devices. Their work demonstrated that an optimized forecasting & threshold model achieved an 88.8% Corrected Event-wise F0.5-score with a 99.4% reduction in MACs and a 97.1% reduction in RAM, successfully enabling high-performance anomaly detection within CubeSat resource constraints.

Key insights

• The forecasting & threshold deep learning approach, particularly using an XceptionTimePlus CNN, demonstrated the highest baseline performance for spacecraft telemetry anomaly detection.
• Neural architecture optimization can dramatically reduce the computational and memory footprint of deep learning models by over 95%, making them feasible for deployment on highly resource-constrained space-grade edge devices.
• The optimized forecasting & threshold model maintained high detection performance (88.8% CEF0.5) with near-perfect precision (99.9%) while fitting into minimal memory (59 KB RAM) and computational budgets.

Summary by alphaXiv.