Datasetpublic

jgalego/orbit2vec-windows

Orbit2Vec windows Windows of daily orbital elements for 70,953 objects, the training data of jgalego/orbit2vec. Configs windows: 32-day windows of daily mean elements. a and b are the two windows of a pair (adjacent windows of one object for train; the earlier and later test windows for test), as flattened float32 arrays of 32 days by 7 elements. 302,056 train rows and 28,230 test rows. objects: one row per object with its catalogue number, name, constellation or… See the full description on the dataset page: https://huggingface.co/datasets/jgalego/orbit2vec-windows.

sourceHugging Facecc-by-4.0updated 2d agoView on Hugging Face

0likes24downloads

Dataset Card

Orbit2Vec windows

Windows of daily orbital elements for 70,953 objects, the training data of jgalego/orbit2vec.

Configs

—windows: 32-day windows of daily mean elements. a and b are the two windows of a pair (adjacent windows of one object for train; the earlier and later test windows for test), as flattened float32 arrays of 32 days by 7 elements. 302,056 train rows and 28,230 test rows.
—objects: one row per object with its catalogue number, name, constellation or payload group, owner, and seen, whether it was in training (the other 10% are held out).
—maneuvers: 823 labelled Starlink cases used only for evaluation. a and b are the 10 days before and after (10 days by 7 elements), and maneuver is true for a manoeuvre and false for a quiet day.

Sources

Elements come from juliensimon/space-track-tle-history (CC BY 4.0, originally Space-Track), with repeated epochs collapsed to the latest TLE per UTC day. Groups and owners come from juliensimon/space-track-satcat. Manoeuvre labels come from RhynoWu/starlink-maneuver-db. The script that builds the windows is the data command in the model repo.

Mean elements from public TLEs are accurate to a few km, and classified objects are missing.