Previous Abstract Return to Session C5 Next Abstract

Session C5: Positioning Technologies and Machine Learning

Machine Learning-assisted GNSS Interference Monitoring through Crowd-sourcing
Nisha Lakshmana Raichur, Tobias Brieger, Fraunhofer Institute for Integrated Circuits (IIS); Dorsaf Jdidi, Friedrich-Alexander-University (FAU), IIS; Carlo Schmitt, IIS; Tobias Feigl, IIS, FAU; J. Rossouw van der Merwe, Birendra Ghimire, Felix Ott, Alexander Ruegamer, Wolfgang Felber, IIS
Date/Time: Friday, Sep. 23, 9:20 a.m.

Objectives.
Interference signals harm Global Navigation Satellite System (GNSS) processing. Therefore, interference signals need to be mitigated or the interference sources need to be localized and removed. A monitoring network is needed to localize an interference source, and the more monitoring receivers are available the better. Luckily, smartphones are ubiquitous and typically include a GNSS receiver for positioning, making them ideal as a monitoring network through crowdsourcing. Further, with recent advances in the Android ecosystem, GNSS measurement support is mandatory to capture the raw data for many real-time tracking applications. This opens a door to more advanced GNSS processing and applications. Since Rel. 9, the Third-Generation Partnership Project (3GPP) supports the distribution of assistance information for mobile devices, e.g., GNSS signal to improve the quality of network services or even precise positioning. It is possible, that such crowdsourcing approaches could provide immense benefits to GNSS integrity if it is included in future 3GPP releases. In the extended abstract, Fig. 1 illustrates the benefit along with a typical scenario: a moving interference affects the signal reception of two UEs within a certain area. The signaling changes dynamically so that the situation requires monitoring by more than one entity, motivating the real-time and widespread monitoring to adapt to the dynamic environment. Therefore, we propose an adaptive framework to detect, (potentially) classify, and localize such sources of interference using artificial intelligence and crowdsourcing techniques along our processing chain.
In the extended abstract, Figure 2 shows the exemplary pipeline of our framework that bases machine learning (ML). The idea of our paper is that mobile devices such as smartphones may detect such local GNSS interference events using data-driven methods and their reliability and make them available to the global network, e.g., by using 5G. Additionally, our crowdsourcing framework allows for the collection of interference events from many devices to characterize their nature and impact. Our framework publishes the collected information to all other mobile devices in the network to significantly improve the accuracy of localization and signaling and reduce unnecessary processing of corrupted signals.

Methodology.
Our framework for machine learning-assisted GNSS interference monitoring through crowdsourcing consists of four main components: (1) feature importance analysis [8, 3, 2, 5, 9], that provides insights into the raw data and the model, and enables necessary dimensionality reduction that may improve the efficiency of a predictive model; (2) Segmentation of the time-series GNSS data using the U-Net-based architecture [7], which maps the inputs of arbitrary length to a sequence of class labels on a freely chosen time scale; (3) Sorting of the received signal into five classes (four with interference one class without interference); (4) Uncertainty estimation of our method [6] to ensure that our algorithm detects reliably.
We thus propose the combination of our deep feature importance analysis on real-time data with our U-Net-based detection algorithm in our novel crowdsource-based localization framework to localize the sources of interference and to (potentially) improve the localization of a mobile phone in static and dynamic scenarios that may even suffer from multipath. By crowdsourcing the characteristics, interference detection, and type of interference to the network, we support cellular-based localization to further improve the integrity of positioning solutions.
To analyze feature importance in-depth, we use raw GNSS measurements of the Android ecosystem. We analyze essential (i.e., DOPs, Elevation, Azimuth, and C/N0) and complex features (i.e., Automatic Gain Control (AGC), signal states, and indicators) that may be used to detect interference. The most optimal selection of features enables interference processing even when GNSS signal reception is degraded or signal tracking is lost entirely.
Our methods to detect and localize interference are based on the latest state of-the-art: Kraemer et al. [4] investigate a pedestrian dead reckoning method to study the effects of static interference on Android smartphones. Strizic et al. [10] exploit the high density of smartphones to localize interference with time or power-difference-of-arrival and static receivers. Borio et al. [1] propose two methods of synthetic array principle and crowdsourcing for static interference localization. However, these approaches are either not suitable for real-time use, employ unsuitable characteristics, or only evaluate methods on simulated data. And there is almost no research on dynamic detectors and interference. Thus, it remains unclear which features contribute positively to an ML-based processing chain for optimal interference treatment in a realistic application scenario. Furthermore, we are not aware of any published evaluation using features of systems with dual-band signal reception and multi-GNSS processing. However, our preliminary findings show that these features significantly improve detection. Hence, we determine the importance and theoretical limits of each feature, to find the representative features that generalize well across sensors and devices, and enable detection and classification tasks on mobile phones. We examine the basic coefficient-based [8], the decision tree-based [3], the permutation-based [2], and the advanced rank-based stability metrics of SHAP (SHapley Additive ex- Planations) [5] and LIME (Local Interpretable Model-agnostic Explanations) [9] algorithms to analyze the feature importance of raw GNSS data for detection and classification tasks. Our experiments show that the selected features are more robust (generally valid) against multipath and dynamic environments and significantly improve the prediction metrics over the state-of-the-art.
Learning the temporal data structure is crucial in detecting and classifying interference since interference affects GNSS signaling for different lengths of time. To detect even variable-duration interference, we use the renowned U-Net-architecture [7] was originally proposed for image segmentation. U-Net enables the entire time segment to be segmented in a single forward pass and outputs different classes with any desired temporal resolution. In experiments, we therefore adapt and optimize different variants of the U-Net-architecture for interference detection and classification. Based on our studies, we propose ageneralizable architecture with a robust training process that does not require architecture or hyperparameter adjustment for different scenarios or devices.
Mobile phones may be extremely valuable sensors for locating GNSS interference. They are ubiquitous and available in all areas subject to potential interference. If we collect information about interference from multiple sensors (crowdsourcing), a few of these devices are sufficient to make a valuable contribution to locating sources of interference, e.g., using the 5G platform. To enable crowdsource-based localization of interference, we propose a heatmap-based localization approach that leverages our supervised ML-based classification. Our experiments examine two scenarios: One scenario examines the accuracy and reliability of static detectors (e.g., parked mobile phone/car on a freeway) to detect passing noise signals. Another scenario examines the detection accuracy and reliability of dynamic detectors and moving interference. Since in both cases the affected area changes dynamically, the situation must be monitored by more than one detector (crowdsourcing). Our experiments show how mobile devices (smartphones) process the information, detect interference events and publish the findings to the network. The crowdsourced information improves the characterization of the nature and impact of the interference signal. We found that sharing the information improves accuracy, eliminates unnecessary redundant processing, and avoids using corrupted signals. Hence, potential interference is detected, the detector shares its history of recent and previous reliable location information, current interference type, time, and detection reliability with the network. Our crowdsourcing framework collects the synchronized information and formulates an optimization problem to estimate the location of the interference. The key factor here is the history and time of the location of all detecting participants and their level of confidence (uncertainty estimate).

Anticipated Results.
We evaluate our method using realistic data collected in our test center. The data contains real-world signals with and without interference, with different distances between transmitter and receiver, with different signal strengths, different dynamics, and different multipath propagations between transmitter (i.e., jammer) and receiver and the environment. In a deterministic indoor scenario, we attenuate the signals to limit interference with nearby real-time GNSS receivers. The GNSS signals are relayed via a repeater with a receiver mounted on the building roof and a ceiling transmitter. This is how we counteract the 20 to 50 dB attenuation caused by the reinforced concrete building.
In this paper, we present our comprehensive experimental study of different methods of feature selection algorithms for classification tasks. We show that
Shapley Additive Explanations (SHAP) and Permutation Importance methods are the most effective single methods for feature importance analysis. We evaluate
the classification performance of the different variants of U-Net-architectures using the class-wise F-?-scores, which are calculated globally across all datasets.
Our U-Net model achieves high performance even across different window sizes, sample rates, allegation protocols, and locations. We show that our method
delivers accurate results with low uncertainties even without adapting to task-specific changes, thus enabling a cost-effective, low-effort crowdsourcing solution.

Conclusion.
In this paper, we present our advanced GNSS processing approach for interference detection, classification, and localization by implementing a machine learning-based pipeline. We lay the practical foundation for using Android phones to crowdsource GNSS interference information using a global location and communication platform, such as the 5G network platform. Our pipeline implicitly analyzes the raw GNSS signals for important features w.r.t. their contribution to detection and classification. We use deep neural networks for robust and generalizable detection, classification, and localization of an interference. Also, we use crowdsourcing of the information about, e.g., the 5G network, to inform other Android phones about the event and significantly reduce the influence of an interference signal and prevent unnecessary processing of corrupted signals from other participants to save energy.

Significance.
Eliminating the GNSS interference by detecting, classifying, and locating them has become paramount to enable smooth GNSS-based localization and save computational and power costs. Given a huge variety of GNSS data with multiple features thereof, that may drastically differ depending on the sensor, we perform a feature importance analysis. In contrast to other state-of-the-art methods, we derive generous features that play a central role in building a good machine learning model for generalizable and robust interference analysis. On the other hand, we observe that our neural network-based U-Net-architecture for interference classification is much more robust than the other state-of-the-art data-driven methods. In addition, the training process of U-Net is significantly more robust than other ML-based methods and requires no architecture or hyperparameter adjustment to work across tasks, favoring generalization and dramatically reducing cost and effort. In addition, our approach localizes the sources of interference along the process chain for both static and dynamic detectors using artificial intelligence and crowdsourcing techniques.

References
[1] Daniele Borio, Ciro Gioia, Andrej ?Stern, Franc Dimc, and Gianmarco Baldini. Jammer localization: From crowdsourcing to synthetic-detection. In Intl. Technical Meeting of The Satellite Division of the Institute of Navigation (ION GNSS 2016), Portland, Oregon, 2016.
[2] Leo Breiman. Random forests. In Machine learning, volume 45, pages 5–32. Springer, 2001.
[3] Krzysztof Grabczewski and Norbert Jankowski. Feature selection with decision tree criterion. In Intl. Conf. on Hybrid Intelligent Systems (HIS’05), pages 6–pp. IEEE, 2005.
[4] I Kraemer, P Dykta, R Bauernfeind, and B Eissfeller. Android GPS jammer localizer application based on c/n0 measurements and pedestrian dead reckoning. In Intl. Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS 2012), Nashville, TN, 2012.
[5] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017.
[6] Aishwarya Manivannan. A comparative study of uncertainty estimation methods in deep learning-based classification models. Deutsche Nationalbibliothek, 2020.
[7] Mathias Perslev, Michael Jensen, Sune Darkner, Poul Jorgen Jennum, and Christian Igel. U-time: A fully convolutional network for time series segmentation
applied to sleep staging. In Advances in Neural Information Processing Systems (NIPS), 2019.
[8] Ping Qiu and Zhendong Niu. Tcic fs: Total correlation information coefficient-based feature selection method for high-dimensional data. In Knowledge-Based Systems, volume 231, page 107418. Elsevier, 2021.
[9] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ” why should I trust you?” explaining the predictions of any classifier. In ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data mining, pages 1135–1144, 2016.
[10] Luka Strizic, Dennis M Akos, and Sherman Lo. Crowdsourcing GNSS jammer detection and localization. In Intl. Technical Meeting of The Institute of Navigation, Reston, Virginia, 2018.



Previous Abstract Return to Session C5 Next Abstract