Jaedong Seong, Korea Aerospace Research Institute; Saehan Song, Korea Aerospace Research Institute; Seo-Eun Lee, Korea Aerospace Research Institute; Okchul Jung, Korea Aerospace Research Institute
https://doi.org/10.64861/DWVZ6452
Keywords: Space Domain Awareness, Ontology, Autonomous Analysis, Maneuver Detection, Large Language Model, TLE, SGP4, Rule-Based Reasoning
Abstract:
Maneuver detection is a mature discipline: comparing consecutive Two-Line Element (TLE) sets to find discontinuities in the orbital elements is a widely used technique. This paper addresses what comes after detection: making a machine explain what a detected maneuver is for, and on what grounds that judgment rests.
We present SENTINEL, an operational system built on a strict layering principle: lower layers establish reproducible facts, and interpretation happens only above them. A two-stage detection pipeline, consisting of orbital-element screening followed by SGP4 V-pattern epoch estimation, characterizes each maneuver, and delta-V is decomposed into radial, in-track, and cross-track components through the Gauss variational equations, with a measured noise floor of 0.4 m/s on the radial channel.
A domain ontology written in OWL 2, comprising 25 classes including an 11-category maneuver-purpose taxonomy, provides the vocabulary over which eight weighted rules assign maneuver purposes. Each judgment carries an explicit evidence chain and a three-tier confidence grade. A large language model (LLM) layer then generates contextual interpretation, threat assessment, and recommendations exclusively from the structured facts established by the ontology. Raw orbital data are never provided directly to the LLM.
Applied to 5,958,884 TLEs covering 17,859 objects over approximately eight months, the system detected 702,161 maneuvers and assigned purposes to 486,542 of them, achieving 69.3% coverage. Validation against two satellites with published weekly maneuver schedules (26 events) recovered the maneuver day of week in 24 of 26 cases, with median epoch errors of 11% to 19% of one orbital period and a radial component correctly reported as zero in all 26 events. Across 340 TLE pairs from non-maneuverable rocket bodies, no radial false positive exceeded the noise floor.
We also report the limitations candidly: TLE data misses most of one test satellites micro-burns, purpose coverage is not equivalent to purpose accuracy, and the quality of LLM-generated narrative is not quantitatively evaluated in this work.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications