Literature Review and Validation of Anomaly Detection in Spacecraft Telemetry Data

Abstract

Spacecraft anomaly detection is critical for mission safety, but deploying capable models on space-grade hardware remains a practical barrier. Ground-based monitoring introduces delays that grow with distance and disappear entirely during communication blackouts, while on-board systems are constrained by processors that are orders of magnitude less powerful than their ground counterparts. This paper evaluates whether an existing LSTM-based detection pipeline can be adapted and compressed to better meet those constraints without sacrificing detection capability. Using the ESA Anomaly Detection Benchmark Mission 1 lightweight subset, the Telemanom framework from Hundman et al. (2018) is replicated, tuned, and quantized across three phases. Phase 1 establishes a baseline CEF0.5 score of 43.57%. Phase 2 adjusts two hardcoded parameters originally calibrated for NASA telemetry, bringing CEF0.5 to 81.0% — a substantial improvement that demonstrates the sensitivity of threshold-based anomaly scoring to dataset-specific error distributions. Phase 3 applies post-training quantization, compressing the model from 310.7 KB to 91.9 KB while preserving prediction accuracy within 0.14% of the original output, establishing a path to onboard deployment without retraining or architectural redesign. Taken together, the three phases show that a well-understood LSTM architecture, with targeted adjustments and post-training compression, can approach competitive detection performance within the memory constraints of real flight hardware.