Satellite orbit correction and maneuver detection

Orbit correction

SGP4 is the standard model for satellite orbit propagation from two-line element (TLE) data. Its error grows over a 7-day horizon. I built a deep residual-learning model that predicts and corrects this error.

• Architecture. A WaveNet-style dilated gated temporal convolutional network (TCN) plus a bidirectional RNN, with FiLM conditioning.
• Calibrated uncertainty. A heteroscedastic loss gives a calibrated uncertainty for each axis.

Maneuver detection

When a satellite fires its thrusters, its orbit changes. Operators want to know that this happened, and when. I built an end-to-end detection pipeline on TLE data:

1. Compute 200+ physics-informed features.
2. Train with focal and cost-sensitive losses, because maneuvers are rare.
3. Choose the operating point with a decision-threshold optimizer.
4. Find the maneuver epoch with a new probability-weighted gradient localization algorithm.

The pipeline runs live-data inference, and an alerting state machine reports each detected maneuver.