Toward Trainable AI Agents for Satellite Operations

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Low Earth orbit is increasingly crowded, and a growing share of it is flown by universities and small operators. These teams must detect and respond to on-orbit anomalies with far less staff, expertise, and tooling than large operators, and a single missed anomaly can end a mission. We present COMET, a ground-based operator agent that helps a small team both detect anomalies and diagnose them. COMET pairs a learned detector with a retrieval-augmented diagnosis stage. An LSTM autoencoder flags anomalous telemetry windows, and a language model explains each one against the mission’s own documentation, procedures, and prior operational records. Because the detector is trained on a mission’s own telemetry and the diagnosis is grounded in its own documents, an organization can build an agent tailored to its spacecraft, before launch and during operations. The agent is advisory rather than autonomous, and every output is traceable back to the telemetry and documents that produced it. We describe the design and a proof-of-concept implementation, and evaluate COMET on a public benchmark of simulated multivariate spacecraft telemetry with injected anomalies. The detector recovers most of the labeled anomalies, favoring recall so that few are missed, and for a flagged event the diagnosis stage produces a traceable, documentation-grounded explanation. These results show the pipeline is feasible and its outputs auditable, a step toward reducing operator burden as deployments grow.

Links and sources

Comments