Satellite Maneuver Detection

Model by Booz Allen

This model detects if a satellite performed a maneuver since it was last observed. As inputs, it uses a series of Two Line Elements (TLEs), also referred to as Keplerian Elements. As output, it provides a determination of whether the satellite recently made a maneuver. This model can be used for space domain awareness (SDA). Detecting if a satellite performed a maneuver provides the ability to update satellite trajectory predictions and reveals potential collisions with other space objects, allowing operators to take measures to avoid them. This model calculates orbital energy and the semi major axis from the Keplerian Elements in a Two Line Element (TLE). Based on changes in Orbital Elements combined with our calculated values, the model uses an Extra Trees Classifier to determine if a maneuver recently took place by splitting the inputs into normal and anomalous data. The model was trained on TLEs with given times that a satellite made a maneuver.

  • Description

    Product Description

    PERFORMANCE METRICS:

    96.3% Precision

    99.9% Recall

    The Satellite Maneuver Detection model is based on an Extra Trees Classifier with 100 estimators; the total training dataset was composed of less than 50,000 records from 9 satellites. Based upon this training data, precision and recall metrics were calculated to characterize performance. The model achieved a precision of 96.3% and a recall of 99.9%.

    A higher precision score indicates that the majority of labels predicted by the model for different classes are accurate. Further information here.

    A higher recall score indicates that the model finds and predicts correct labels for the majority of the classes it is supposed to find. Further information here.

    OVERVIEW:

    This model calculates orbital energy and the semi major axis from Two Line Element’s (TLEs) Keplerian elements. Based on changes in these values along with all Keplerian elements from the TLEs, the model can determine if a maneuver took place. This model uses an Extra Trees Classifier to split data into normal and anomalous data. The model was trained on TLEs with given times that a satellite made a maneuver.

    TRAINING:

    This model was trained using an Extra Trees Classifier. The nine training sets, one for each satellite, varied in size from 720 to 5,320 records. Training time on a 2.2 GHz Intel Core i7 took less than one second for all models.

    VALIDATION:

    There was a validation set for each of the nine models; these varied from 240 to 1,770 records. Test time was less than 0.1 seconds for all models.