Unsupervised Disentanglement for Post-Identification of GNSS Interference in the Wild

Jdidi D, Brieger T, Feigl T, Contreras Franco D, Van Der Merwe JR, Ruegamer A, Seitz J, Felber W (2022)


Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2022

Pages Range: 1176-1208

Conference Proceedings Title: Proc. Intl. Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+)

Event location: Denver, CO US

DOI: 10.33012/2022.18493

Abstract

Interference influences Global Navigation Satellite System (GNSS) signals, degrading localization performance.
Therefore, the interference signals must be handled or the transmitter (i.e., jammer or interference source) detected, localized and removed.
However, interference must first be identified, possibly decorrelated, and then located to remove interference to enable robust localization over multipath.In addition, a successful classification of the waveform of an interference helps to deduce the purpose of the interference and separate it from environmental multipath effects, making it easier to find and handle the interference. For instance, a chirp waveform could indicate a Privacy Protection Device (PPD) most likely mounted in a vehicle; or a pair of pulses could indicate that it is a Distance Measurement Equipment (DME) located at an airport. Therefore, both detection and signal classification are critical to dealing with sources of interference.
Classic approaches use the spectrogram [13] to classify sources of interference, as it offers a time-frequency representation of the received signal, or they extract features (i.e., entropy) from raw data (I/Q samples) of the received signal [15]. Recent research employs Deep Learning (DL) (e.g., Convolutional Neural Networks(CNN)) or Machine Learning (ML) methods [13], e.g., Support Vector Machine (SVM) or Random Forest (RF) for classification. However, the current state-of-the-art shows that ML and DL methods may only detect and classify interference signals accurately and reliably when working in laboratory environments or using simulated interference signals. Several interference monitoring systems can detect and classify interference [2]. However, these typically require expensive and sophisticated hardware and software to enable accurate detection and classification, thus limiting the number of participants. However, to enable an efficient network that increases the probability of interception (POI) of an interference source, the efficient exchange of information about known interference and environmental effects between the participants is essential. Accordingly, an application-oriented processing chain that reliably detects, classifies, and localizes various sources of interference in a real environment with multipath propagation and environmental dynamics is not yet publicly available.
Additionally, interference handling depends on the quality and meaningfulness of the data it processes and the efficient exchange of data between devices that sense interference. However, the data volume and the complexity of lossless information are enormous even for a short time, which leads to an enormous functional space, also known as the curse of dimensionality. It requires significant memory resources, resulting in rapidly depleting memory, compute, and power resources [7].
In order to circumvent these limitations, GNSS data compression has been proposed for several applications [7, 1, 14]. Typically, they determine the size and dimensions of the GNSS data packets a priori. Thus, this forces them to transmit fixed data packets that may not adapt to environmental changes, changes in the communication link, and may not preserve all significant and characteristic information of the raw data. Practically, this means that the worse (lossier) the compression is, the worse the downstream tasks (detection, classification, and so localization) perform. It is why data compression, which we describe as an unsupervised non-linear dimensionality reduction function, is a crucial part of our ML framework.
Another limitation is the creation of multipath-resistant classification algorithms. It is challenging but crucial for real-world applications. A traditional approach to interference classification works well with deterministically well-defined signals in a laboratory setup. In the real world, however, multipath effects cause frequency-selective fading, which, for example, disturbs the spectrogram and consequently degrades classification accuracy. Consequently, this would affect the localization performance of an interference source and thus disable efficient mitigation. Conventional localization approaches apply detection, attenuation, and either Time Difference of Arrival (TDoA) or Power Difference of Arrival (PDoA), or a mixture thereof, to the time and/or frequency domain of GNSS signals to locate interference. Although these techniques prove to be optimal interference locators in terms of high Interference to Noise Ratios (INRs) and Line-of-Sight (LoS) conditions, we claim that unsupervised channel charting [12] yields competitive accuracy and robustness even in situations with Non-Line-of-Sight (NLoS) and low INRs and achieves efficient compression.
Since our compression component inherently disentangles information in its latent space, both spatial information of the interfering signal and multipath effects are preserved. Reconstruction errors are also kept low, allowing an accurate reconstruction of raw data from the compressed latent version with virtually no loss.
Moreover, our compression method is based on the theory that similar information is also stored in close neighborhoods/ clusters. In this context, similarity metrics (e.g., Euclidean or Pearson distance) quantify the distance between different clusters in the latent space allowing reliable outlier detection [8]. It enables a coarse-grained detection and localization of outliers in the latent space for the post-processing phase. Hence, it is sufficient if, instead of the high dimensional received GNSS signal channels, we only exchange the most representative latent components between participants in our surveillance system, and as a result, the communication load is reduced significantly.
In this paper, we propose our novel interference handling framework (along with detection and classification) and highlight the critical effects and benefits of our compression technique. We evaluate our surveillance system in a realistic test environment, where we create low-power tests with controlled multipath at different distances and motion dynamics. We also evaluate our pipeline for compressing data with an external real-world measurement campaign to verify its performance in real-world scenarios.

Authors with CRIS profile

Involved external institutions

How to cite

APA:

Jdidi, D., Brieger, T., Feigl, T., Contreras Franco, D., Van Der Merwe, J.R., Ruegamer, A.,... Felber, W. (2022). Unsupervised Disentanglement for Post-Identification of GNSS Interference in the Wild. In Proc. Intl. Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+) (pp. 1176-1208). Denver, CO, US.

MLA:

Jdidi, Dorsaf, et al. "Unsupervised Disentanglement for Post-Identification of GNSS Interference in the Wild." Proceedings of the Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+), Denver, CO 2022. 1176-1208.

BibTeX: Download