Magnetic navigation project by Nathan Laoué
GPS navigation is very popular nowadays, but the problem is that the signal can be jammed. It is therefore necessary to implement methods with the same advantages, i.e:
- Available in any weather
- Available at any location
- Available 24 hours a day and 7 days a week
but which cannot be jammed. This is where magnetic navigation comes in. In addition to having the advantages of GPS, magnetic navigation is much more difficult to jam or impossible in the air.
However, there is a current blocking element, the magnetic disturbance of the carrier. In order to navigate by means of the magnetic field of the earth, measurements of the magnetic field are made from the earth and a map of magnetic anomalies is referred to. However, the carrier from which the measurements are made emits magnetic disturbances. In order to obtain good measurements, it is necessary to be able to remove the carrier disturbance from the magnetometer measurements.
There are currently some techniques such as placing the magnetometer on a pole at about 2-3 meters from the aircraft and then perform a Tolles-Lawson calibration. This is very impractical and imposes constraints. This project aims to explore new solutions to use the sensors placed in the aircraft by exploring new techniques such as deep learning.
The dataset is provided by the United States Air Force pursuant to Cooperative Agreement Number FA8750-19-2-1000.
Albert R Gnadt, Joseph Belarge, Aaron Canciani, Lauren Conger, Joseph
Curro, Alan Edelman, Peter Morales, Michael F O’Keeffe, Jonathan Taylor,
and Christopher Rackauckas. Signal enhancement for magnetic navigation challenge problem. arXiv e-prints, pages arXiv–2007, 2020.
The data used for this project comes from an MIT challenge. The data can be downloaded here. The goal of the challenge is the same as the one of our project but we do not take into account their restrictions on the dataset. A datasheet of the data is available here
This dataset was created by SGL. They have made several flights by placing magnetometers in several places of the plane and especially a magnetometer on a pole at the end of the plane which will serve as truth. They also took measurements of various elements of the aircraft such as roll angle and battery voltage.
In this project we use 3 different magnetic anomaly maps. Maps of Renfrew and Eastern area are provided with the dataset. For the map of Canada, it is available from this link. Also for the world map of magnetic anomalies, it is available from this link.
You can also download the IGRF model here.
To be able to run the notebooks, you have to paste the downloaded flight data (1002, 1003, 1004, 1005, 1006, 1007 flights) in the raw folder in data. All that remains is to run the notebooks in order and the additional data will be created automatically (all steps are in the notebooks). There is also a requirements.txt file, just open a terminal and go to the project folder and do pip install requirements.txt. This will download the packages needed to run the project.
├── LICENSE
├── README.md <- The top-level README of this project.
├── data
│ ├── external <- Data from third party sources.
│ ├── processed <- The final, canonical data sets for modeling.
│ └── raw <- The original, immutable data dump.
├── models <- Trained models.
├── notebooks <- Jupyter notebooks.
├── references <- Data dictionaries, manuals, and all other explanatory materials.
├── reports <- Report in french and english of this work.
├── requirements.txt <- The requirements file for reproducing the analysis environment.
├── src <- Source code to train models.
│ ├── models <- Code of models.
Three main types of models were tested: MLP, CNN and LSTM. During the different tests, the CNN seems to be the most inetressing model allowing to obtain an RMSE of 23 nT on section 1007.06 of flight 1007. We can see the prediction of the CNN model in the figure below for flight 1007 :
For more information on performance comparisons, a full report is available in the reports section in french (original) and in english (translated with DeepL).