Previous Abstract Return to Session C5 Next Abstract

Session C5: Positioning Technologies and Machine Learning

Machine Learning-assisted GNSS Interference Monitoring through Crowdsourcing
Nisha Lakshmana Raichur, Tobias Brieger, Dorsaf Jdidi, Tobias Feigl, J. Rossouw van der Merwe, Birendra Ghimire, Felix Ott, Alexander Rügamer, and Wolfgang Felber, Fraunhofer Institute for Integrated Circuits IIS, Nuremberg, Germany
Date/Time: Friday, Sep. 23, 9:20 a.m.

The Global Navigation Satellite System (GNSS) community shows great interest in detecting and eliminating GNSS interference, i.e., jammers. State-of-the-art techniques employ either threshold-based mechanisms or supervised learning on raw data streams or features thereof. However, they require special expensive GNSS receiver hardware that needs to be placed in a fixed location. Instead, there is limited research on ubiquitous interference detection using mobile devices such as smartphones. But, advances in the smartphone ecosystem enable the support of GNSS measurements for real-time navigation and the Third-Generation Partnership Project (3GPP) enables the distribution of potential assistance information. However, there is limited research into crowdsourcing smartphone-based features to localize the source of any detected interference.

Hence, we employ supervised learning to map effective GNSS features of Android-based smartphones to corresponding reference labels to reliably detect and classify sources of interference. From there, we localize the identified sources of interference. We use the 5G platform for decentralized synchronization and collection of features and corresponding reference positions (either an unbiased GNSS- or a 5G-position). We evaluate both state-of-the-art and our methods on data from our large-scale real-world measurement campaign. Our dataset covers realistic effects such as multipath, motion dynamics, and variations in distance and power between jammer and sensors. We show that our selected features are optimal for detection, classification, and localization and are robust against multipath and dynamic environments. Our experiments also show that our novel deep learning pipeline (based on U-Net) outperforms state-of-the-art techniques, reliably detects (F2 > 93%) and classifies (F2 > 90.22%) six different interference classes (with 33 subclasses), and predicts uncertainty (< 16%). For localization, our supervised coarse-grained region classification and fine-grained position regression (MAE < 2.5 m) enable applications such as lane-detection when we combine the information from four smartphones over the 5G platform.



Previous Abstract Return to Session C5 Next Abstract