A Foundation Model Approach for Anomaly Detection in DoD Spacecraft Telemetry
Name
Cagle_Final.pdf
Size
1.45 MB
Format
Adobe PDF
Checksum (MD5)
43d281ebf785552cf1a084fac509831a
Author(s)
Cagle, Kevin
Tierney, Michael
Date Issued
August 20, 2026
Abstract
Spacecraft operators monitor large volumes of multivariate telemetry to identify off-nominal behavior, but anomaly events are rare, mission-specific, and expensive to screen manually at scale. This paper presents ARGUS, a three-phase effort to adapt the UniTS time-series foundation model to spacecraft telemetry from the European Space Agency Anomaly Detection Benchmark, historical STPSat-4 data, and live STPSat-7 telemetry. UniTS is pretrained on a broad range of time-series data, making it a candidate backbone for missions where labeled anomaly data are scarce. The approach combines mission-specific preprocessing, threshold and training-configuration sweeps on MIT SuperCloud, and an operator-facing anomaly investigation application. The study evaluates how a pretrained time-series model can be adapted across spacecraft telemetry contexts with different levels of supervision, and it emphasizes both quantitative anomaly-detection behavior and operational reviewability. The result is a reproducible path from benchmark validation to DoD mission telemetry adaptation and human-in-the-loop anomaly investigation.
Subjects
spacecraft telemetry
anomaly detection
time series foundation models
UniTS
prompt tuning
operator decision support
MIT Department
Lincoln Laboratory
Terms of Use
All rights reserved
Persistent DSpace Link
