maneuver-detect

Open dataset, models, and benchmark for detecting orbital maneuvers from public TLE history.

maneuver-detect takes a satellite's public TLE history and returns a DataFrame of detected maneuvers — each with a detection epoch, a calibrated confidence, a maneuver type (in-track / cross-track / radial), and a Δv estimate. It ships a curated, reconstructable, labelled dataset built from public catalog data and operator maneuver announcements; a classical reference detector, learned (BiLSTM and transformer) baselines, and a foundation-model (Chronos forecast-residual) baseline distributed through the Hugging Face Hub, all sharing a vis-viva / Gauss-variational Δv inversion; and a frozen, leak-free benchmark protocol — with a public leaderboard — so a new detection method can be measured against prior work on the same splits. See the changelog for released functionality.

What this is

Detecting maneuvers from public TLEs is a long-running space-situational-awareness problem, but the open ecosystem has no shared answer to it: every paper rebuilds its own dataset, cleaning pipeline, detector, and evaluation, so results are not comparable and the data is rarely published. maneuver-detect provides the missing shared piece — an open, citable dataset and a reproducible benchmark. The load-bearing engineering is the dataset and the benchmark protocol (leak-free splits, the detection-matching rule, the metric, and the physics of the Δv inversion), not the model code, which is deliberately small and standard. The classical reference detector is the baseline every learned model must beat.

Quick start

from maneuver_detect import detect, datasets

history = datasets.tle_history(norad_id=25544, start="2024-01-01")
maneuvers = detect(history, model="transformer-base")   # learned baseline, pulled from the Hub on first use
# DataFrame columns: epoch, confidence, type, delta_v_estimate, plus provenance

"classical" is the no-download default; the learned "bilstm-base" and "transformer-base" models pull their checkpoints from the Hugging Face Hub on first use (CPU-only, cached on disk).

From the command line, on a NORAD id (fetched live) or a local TLE file:

maneuver-detect detect 25544

Runnable examples live in examples/: detect_norad.py detects maneuvers for a NORAD id, and reproduce_baseline.py runs the classical detector through the benchmark scorer on a labelled series.

Dataset and benchmark

• Dataset — a curated, labelled set spanning LEO altimetry, MEO navigation (GPS and Galileo), GEO, and IGSO (QZSS) satellites, built from public TLE history and operator maneuver announcements. The raw multi-year history comes from Space-Track, whose terms do not permit redistributing the data or analysis derived from it, so the dataset ships as a pinned, byte-deterministic reconstruction recipe plus a content-hash manifest: you rebuild it locally from your own Space-Track account with maneuver-detect dataset build, and the manifest verifies the result byte-for-byte. The authored artifacts (labels, recipe, splits, manifest) are CC-BY-4.0.
• Classical baseline — a rule-based reference detector (time-aware Holt smoothing, a multi-element jump rule, and the vis-viva / Gauss Δv inversion) that ships in the package. It is the baseline every learned model must beat.
• Learned baselines — a BiLSTM and a ~10M-parameter transformer detector, distributed through the Hugging Face Hub and selected by name (detect(history, model="transformer-base")). Each localises maneuvers in the element series and reuses the same Δv inversion; inference is CPU-only and pulled on first use.
• Foundation baseline — an optional forecast-residual detector built on a pretrained time-series model (Chronos), selected as model="chronos-residual" and installed via the [foundation] extra. It replaces the classical detector's hand-built quiet-dynamics prior with a learned forecast and reuses the same matching, Δv inversion, and scorer.
• Benchmark — frozen, leak-free train / val / test splits (by satellite and time), the detection-matching rule, the per-class metric (precision/recall at a fixed false-alarm rate), and a deterministic scorer that reproduces the published numbers from a committed predictions file.

The dataset and the learned-model checkpoints are distributed through the Hugging Face Hub: detect(history, model="bilstm-base") (or "transformer-base") pulls the checkpoint on first use — CPU-only, cached on disk, nothing fetched at install time — and the dataset is downloadable with the datasets accessor (see Models and the Hub). A public leaderboard scores any method on the frozen test split with the shipped deterministic scorer and ranks it against the baselines (see the how-to-submit guide). The dataset and benchmark references document the source terms and the full protocol.

Installation

pip install maneuver-detect

The base install carries only permissive dependencies. It includes the PyTorch / Lightning modelling stack and the Hugging Face Hub / datasets libraries — the learned baselines and the Hub-distributed dataset and checkpoints build on them — so a GPU is needed only to train new baselines, never to install the package or run the classical detector. The optional time-series foundation-model baseline lives behind the [foundation] extra:

pip install "maneuver-detect[foundation]"

maneuver-detect supports Python 3.10, 3.11, and 3.12.

What this is not

• Not a maneuver predictor. It detects maneuvers that have already happened; forecasting future maneuvers is a different problem, deliberately out of scope.
• Not real-time or streaming. It is batch — a TLE history in, a maneuver DataFrame out.
• Not a new propagator or orbit-determination engine. It consumes SGP4 mean elements and the small inversions the Δv estimate requires; it does not do precise propagation.
• Not a general time-series-anomaly framework. The detectors are maneuver detectors on orbital element series, not a reusable anomaly library.
• Not a cross-catalog correlation or object-association tool. It works one catalogued object's history at a time; it does not correlate observations across catalogs or sensors.
• No closed or commercial data. Only publicly available TLEs and publicly released maneuver labels are used; redistribution-restricted commercial SSA products are excluded.

Documentation

Full documentation is at astro-tools.github.io/maneuver-detect — getting started, the dataset and label-source reference, the models-and-Hub reference, the benchmark protocol, the leaderboard how-to-submit guide, the output schema and Δv-inversion reference, the frozen design decisions, and the API reference.

Development

git clone https://github.com/astro-tools/maneuver-detect.git
cd maneuver-detect
uv sync --all-groups

See CONTRIBUTING.md for the workflow and local checks. The frozen design decisions live in docs/design/.

License

MIT — see LICENSE.