Previous Abstract Return to Session C5 Next Abstract

Session C5: Positioning Technologies and Machine Learning

Unsupervised Disentanglement for PostIdentification of GNSS Interference in the Wild
Dorsaf Jdidi, Tobias Brieger, Friedrich-Alexander-University (FAU) & Fraunhofer Institute for Integrated Circuits IIS; Tobias Feigl, David Contreras Franco, J. Rossouw van der Merwe, Alexander Rügamer, Jochen Seitz, and Wolfgang Felber, Fraunhofer Institute for Integrated Circuits IIS
Date/Time: Friday, Sep. 23, 9:43 a.m.

Interference signals (i.e., jammers) affect the processing chain of global navigation satellite system (GNSS) receivers, affecting or disrupting their accuracy. State-of-the-art supervised, learning-based methods detect and classify jammers at the highest accuracies. However, they require an elaborate labeling process that is scenario-specific and covers variations within each class. It is critical for practical applications to enable model training that covers environment and application-specific requirements.
We propose a quasi-unsupervised framework that does not rely on prior knowledge and actively adapts to different environment-specific factors such as multipath, dynamics between participants, dynamics in the environment, and variations in signal strength or distance. Our experiments show that our pipeline is robust to six families and 21 sub-classes of different (un)-known jammers, multipath propagation, different jammer-receiver distances, and dynamics. Our results indicate that in 95% of all cases, the classification of known jammers is accurate at low uncertainty (SD±10%). In addition, even for unknown types of jammers, unsupervised almost error-free detection is possible (accuracy?90%, SD±10%). In addition, our disentanglement method allows our monitoring system to reliably interpret and identify outliers, i.e., unknown classes (V-measure?0.9; F2?95%, SD±5%). Experts may actively provide reference labels to consistently improve the performance of our surveillance system.



Previous Abstract Return to Session C5 Next Abstract