Previous Abstract Return to Session D8

Session D8: AI/Machine Learning: MagNav

Magnetic Navigation Flight Testing on a C-17A by the DAF-MIT AI Accelerator
Jonathan Taylor, Glenn Carl, Allan Wollaber, Albert Gnadt, MIT Lincoln Laboratory
Location: Ballroom E
Date/Time: Wednesday, Jun. 5, 9:35 a.m.

In conjunction with Air Mobility Command, the DAF-MIT AI Accelerator (AIA) conducted a demonstration of magnetic navigation (MagNav) on a C-17A in May 2023. This demonstration leveraged the sensing capabilities of an Air Force Institute of Technology (AFIT) Mag-in-a-Box, as well as an Air Force Research Laboratory (AFRL) Vampire module to interface with the C-17A inertial navigation system (INS). This hardware combination allowed for real-time magnetic measurements to be paired with an onboard INS, to enable a field test of magnetic navigation as an alternate-navigation position signal.
MagNav determines aircraft (or other vehicle) location by measuring the strength of the Earth's magnetic field and comparing that to a known map of the Earth's magnetic field. The majority of the useful variation in this map arises from the crustal anomaly field. MagNav must overcome the challenge of removing the corrupting magnetic signal and noise of the aircraft in order to measure a sufficiently clean and accurate magnetic signal to recover the crustal field, which is referred to as compensation. MagNav can then be used as an alternative to GPS as an update source to the INS. MagNav is especially of interest because it is almost impossible to jam or spoof, and it does not require clear weather, visibility of the ground, or other external dependencies (e.g., satellites). MagNav does require magnetic maps of operating areas, and performance is constrained by the quality of those maps. This work focuses primarily on addressing the calibration/compensation problem for a C-17A over Edwards Air Force Base.
As part of this testing, the AIA MagNav team demonstrated the ability to train and produce a real-time MagNav signal during a C-17A flight. Due to data degradation, analysis is primarily conducted on an offline replay of data. This analysis, focusing on compensation, yielded high-quality compensation performance, as well as acceptable navigation accuracy based off of the real-time data.
The compensation modeling leveraged both conventional Tolles-Lawson compensation, as well as a neural network augmentation in order to deal with additional noise that is not adequately modeled by the baseline. We will discuss our approaches to compensation, including training data limitations, before moving to a full evaluation of our navigation results.

Previous Abstract Return to Session D8