Edge AI Optimization for Satellite Telemetry Anomaly Detection: INT8 Autoencoders, Energy Estimates and Downlink Savings on OPS-SAT and ESA Mission 2

Onboard anomaly detection is motivated by limited satellite power and costly downlink. This preprint evaluates compact LSTM (3,281 parameters) and dense (428 parameters) telemetry autoencoders in 32-bit floating-point and fully integer 8-bit (INT8) LiteRT form on OPS-SAT and on a hash-locked confirmation year of ESA Mission 2 telemetry. INT8 lowers estimated arithmetic and weight-fetch energy about 5.2-fold, while measured CPU latency improves only 1.11-fold. A policy that downlinks a short summary for every block and raw readings only for flagged blocks removes about 98% of the raw telemetry volume while still transmitting at least one block of about 95% of annotated rare events; random flagging at the same budget captures about 12%, and the policy remains about 96% below losslessly compressed telemetry. Simple statistical controls achieve the same saving. All energy and cost figures are model-based estimates, not hardware measurements. Code and results: https://github.com/sasmina96-sudo/-satellite-anomaly-detection