Previous Abstract Return to Session D8 Next Abstract

Session D8: AI/Machine Learning: MagNav

Code for Limiting Electromagnetic Anomalies Nearby User Platforms (CLEANUP)
Jonathan M. Wheeler, Michael Bulatowicz, Adrian Hernandez, Michaela Villareal, Austin Parrish, Michael Larsen, Northrop Grumman Systems Corporation
Location: Ballroom E
Date/Time: Wednesday, Jun. 5, 9:15 a.m.

Code for Limiting Electromagnetic Anomalies Nearby User Platforms (CLEANUP) is an algorithm that enables high-accuracy magnetic navigation in a GPS-denied environment. Magnetic navigation using map-matching approaches requires knowledge of the earth magnetic field accurate to a few tens of nanotesla or less, corresponding to the typical amount of error in data products such as the Enhanced Magnetic Model 2017 (EMM2017). However, the magnetic material in the hulls of many platforms produce their own magnetic fields that are typically on the order of tens of thousands of nanotesla. Thus, magnetic navigation requires an algorithm that can discriminate the magnetic contamination of the platform from the background earth field with at least with 99.9% accuracy.
CLEANUP takes readings from strap-down magnetometers positioned around the platform and feeds them into a mathematical model of the structure of the ship. The CLEANUP mathematical model is trained by magnetometer data that is tagged by a priori estimates of the earth magnetic field (e.g., with EMM2017).
Compared to other approaches that use neural networks to remove platform magnetic contamination, our technique offers three unique advantages. First, CLEANUP utilizes a relatively small number of free parameters. Compared to other machine learning approaches (which can have millions of free parameters), our approach can be trained with a small dataset. Furthermore, once trained, the cleanup of the platform contamination requires only a few dozen floating-point operations and about 1 kilobyte of memory, which enhances the deployability of CLEANUP on embedded systems that would otherwise have insufficient CPU memory or FPGA fabric to accommodate a neural network.
Second, the mathematical model does not require a priori knowledge of the location and orientation of the sensors or the composition of the platform. This allows our approach to be used on government platforms without disclosing any information about classified components on board the platform to CLEANUP. However, in settings where the sensor arrangement and an approximate model of the platform's magnetic domains is known, this further accelerates the training of the model. Furthermore, if a sensor needs to be moved, or if a large bulky magnetic component on the platform has moved, CLEANUP is designed in such a way that only discrete subcomponents of the mathematical model need to be retrained.
Third, CLEANUP is resilient to degradation in the sensors. For example, if CLEANUP detects that a channel from one of the magnetometers is outputting bad data, it can safely ignore that individual channel and continue to remove the platform-dependent field with high accuracy using the remaining channels.
We have assessed the performance of CLEANUP in simulation to understand the performance limits of the algorithm itself (i.e., using ideal sensors with no noise and drift, perfect knowledge of the earth field in the training data set). The algorithm accurately captures the effects of both permanent and induced magnetization as well as hysteresis. In a simulation of a large watercraft that produces 50,000 nT of platform field contamination, the model trains on a data set of a few hundred samples, and CLEANUP is able to recover earth field in a validation data with very high accuracy. The residual error improves when the sensors are suspended using a boom.

Previous Abstract Return to Session D8 Next Abstract