Travis Yeager, Lawrence Livermore National Laboratory; Talin Guliani, Lawrence Livermore National Laboratory; Joel Diamond, Lawrence Livermore National Laboratory; Denvir Higgins, Lawrence Livermore National Laboratory
https://doi.org/10.64861/WJNI8891
Keywords: TLE, SGP4, orbit propagation, orbital feature engineering, time series, anomaly detection, LSTM autoencoder, Isolation Forest, satellite maneuver detection, SDA, SSA, SSAPy
Abstract:
Two-Line Elements (TLEs) remain one of the most accessible and widely used data products for monitoring Earth-orbiting objects. However, practical SDA workflows often require more than raw element sets: analysts need a scalable way to transform evolving TLE streams into interpretable time-series signals and to highlight time periods that warrant investigation (e.g., maneuvers, abrupt orbit changes, or anomalous propagation behavior). This work presents an end-to-end, TLE-driven anomaly detection pipeline that automates data collection, orbit propagation, orbital feature engineering, and machine-learning-based anomaly scoring to support rapid screening of satellite behavior over user-defined time spans. The approach has been demonstrated on real satellite data, with detected anomalies correlated to known maneuver events, validating the operational utility of the method for SDA applications. The pipeline begins by downloading historical daily TLE files over a specified date range using an incremental, idempotent process designed to avoid redundant downloads and to support repeated execution in operational environments. For a target object defined by NORAD catalog ID, TLE records are parsed and filtered into an ordered set of element updates. Each element set is then propagated to generate an ephemeris time series of position and velocity states at a configurable cadence. Propagation is performed using standard SGP4 and, when available, an SSAPy-based option, enabling comparison between propagation implementations and potential integration into broader SSA toolchains. From the resulting ephemeris, the system computes a compact set of physically meaningful orbital features intended to improve interpretability and reduce dependence on coordinate-frame-specific raw state vectors. Engineered features include radius, velocity magnitude, specific orbital energy, specific angular momentum, inclination, eccentricity, semi-major axis, and orbital period. These features form a multivariate time series that is normalized and segmented into sliding windows for machine learning. Windowing enables detection methods to capture both short-duration deviations and longer-term changes relative to learned nominal patterns. Two complementary anomaly detection approaches are implemented. First, a recurrent LSTM autoencoder is trained on the windowed feature sequences to learn a compressed representation of nominal orbital evolution. Anomalies are flagged using reconstruction error with a percentile-based threshold; training stability measures include early stopping, learning-rate reduction, and gradient clipping. Second, an Isolation Forest model provides a non-neural baseline detector with a contamination-based operating point, enabling rapid scoring and comparison against the autoencoder results. The pipeline produces analyst-facing artifacts including anomaly score plots and CSV reports that associate anomalous timestamps with feature-level z-scores to aid triage and root-cause assessment. Optional visualization capability supports 3D trajectory rendering for contextual review of anomalous periods. Preliminary demonstrations on real world operational satellite data have shown promising correlation between detected anomaly windows and documented maneuver events, confirming the pipeline’s ability to identify meaningful orbit changes from TLE histories alone. The expected contribution is a reproducible, configurable workflow that bridges raw TLE ingestion and ML-driven screening, supporting SDA applications where rapid identification of unusual orbit evolution is required. The approach is designed to be extensible: additional feature sets, alternative time-series models, or auxiliary data sources can be integrated while preserving the same TLE-to-report operational pattern. Additional validation runs on classified datasets are planned prior to the conference to further demonstrate detection performance across a broader range of satellite types and maneuver profiles.
Date of Conference: September 15-18, 2026
Track: Astrodynamics