Title: Simultaneous Localization of Multiple GNSS Interference Sources via Neural Networks
Author(s): David Besson
Published in: Proceedings of the 30th International Technical Meeting of The Satellite Division of the Institute of Navigation (ION GNSS+ 2017)
September 25 - 29, 2017
Oregon Convention Center
Portland, Oregon
Pages: 2812 - 2829
Cite this article: Besson, David, "Simultaneous Localization of Multiple GNSS Interference Sources via Neural Networks," Proceedings of the 30th International Technical Meeting of The Satellite Division of the Institute of Navigation (ION GNSS+ 2017), Portland, Oregon, September 2017, pp. 2812-2829.
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In
Abstract: The accuracy and availability of Global Navigation Satellite System (GNSS) applications hinge on the rapid localization and mitigation of interference sources. This research proposes to localize multiple interference sources simultaneously using artificial neural networks; one of the primary tools in the rapidly expanding field of machine learning. In particular, we leverage the techniques of handwritten character recognition to reconstruct the localization problem in a tractable multi-label, multi-class classification framework. We pose the multiple GNSS interference source localization problem as determining which of 400 cells in a 100 km2 grid contain transmitters interfering with the Global Positioning System (GPS) L1 signal based on in-band power measurements taken by 25 sensors from an array of known, fixed positions. In this preliminary work we only consider stationary and isotropic interference sources and train the neural network to simultaneously classify the cell positions of up to ten 100 W interference sources. We develop the training sets for the neural network by simulating the aggregate interference of tens of thousands of interference source combinations. The aggregate interference is recorded by simulated drones that are hovering at a fixed altitude of 121.92 m (400 ft) above the search grid. These measurements roughly correspond to the array of pixel intensities that are fed into neural networks for handwritten character recognition. The neural network in our problem is comprised of five layers: the input layer of 25 aggregate interference power measurements, three hidden layers with 100 perceptrons each, and the output layer of 400 values corresponding to the probabilities that each of the 400 cells contain interference sources. The output layer identifies only the probability that the cells contain interference sources, not the location of the sources within the cells or the transmitted interference power. The precision score of the trained neural network is approximately 80%. We simulate this approach using Python 2.7 and its scikit-learn 0.18 machine learning library, both of which are free software packages.