Correcting ArcJet Contamination to GOES-16 MAG using Bi-Directional LSTM Networks with Luong Attention

Geostationary Operational Environmental Satellite - 16 (GOES-16) has two Magnetometers mounted on a 8.55 meter long boom: an Inboard Magnetometer (IB-MAG) which is around 6.35 meters away from the satellite bus, and an Outboard Magnetometer (OB-MAG) which is 2.2 meters away from the IB-MAG, placed at the end of the boom. The locations of the IB- and OB-MAGss were chosen to avoid spacecraft magnetic interference in the measurements of Earth's magnetic field. However, to keep the GOES-16 in a desired orbital configuration, the GOES-16 uses 4 hydrazine ArcJet thrusters, which are located away from the IB- and OB-Mags. The thruster nozzles point in the y-direction of the spacecraft frame, which is aligned with the orbit normal. Part of the hydrazine ArcJet gas is ionized, and therefore it interacts with the local space environment, causing contamination in the measurements of the Earth's magnetic field through the duration of the ArcJet firing maneuvers. In this study, we use Bidirectional Long Short-Term Memory Encoder-Decoder Networks with Luong Attention mechanisms for near-real time corrections of the ArcJet contaminations in the IB-and OB-MAGs. We use 842 days of minute resolution data from the voltages and currents from the ArcJets and magnetic field measurements from the IB- and OB-MAGs, 590 days (70%) of which is used for training, 168 days (20%) for validation, and remaining 84 days (10%) for testing. We determine the hyperparameters, such as sizes of the LSTM encoder and decoder layers, sizes of the fully connected layers, learning rate, dropout, and recurrent dropout, based on Bayesian Optimization. Using the best fitting hyperparameter space for the LSTM network, we show the efficiency of the network in correcting ArcJet contamination in GOES-16 MAG measurements.