Previous Abstract Return to Session B3 Next Abstract

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Artificial Intelligence for GNSS Interference Monitoring: Detection and Characterization
Antonia N. Ivan, Stefan S. Mihai, Alexandru C. Pandele, Ileana Mihu, Romanian InSpace Engineering; Julia Hauser, Roman Lesjak, JOANNEUM RESEARCH; Florin Stoican, Three Tensors
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 10:40 a.m.

Reliable GNSS interference detection and characterization are essential for operational monitoring at critical infrastructure sites (e.g., airports), where degraded GNSS performance can directly affect service quality and resilience. In this context, the ESA ECHO-NN project investigates the use of Artificial Intelligence (AI) for GNSS interference detection and characterization within the ECHO monitoring concept. This abstract presents the final-review results of the project, evaluating a neural-network-based approach alongside conventional rule-based indicators (e.g., AGC and C/N0 monitoring), with emphasis on AI design, validation methodology, and operational performance. The goal is to determine whether AI can provide more robust and informative interference monitoring than classical algorithms, while also identifying current deployment limitations across different hardware setups.
The work comprised two main activities: (1) dataset preparation and AI architecture design, and (2) neural network training, validation, and testing. Dataset preparation was treated as a critical step, since model quality depends strongly on data quality and labeling consistency. Data were recorded using the ECHO-NN monitoring station installed near Graz Airport (Austria) in a collaboration between JOANNEUM RESEARCH (JR) and Romanian InSpace Engineering S.R.L. (RISE). The monitoring station is based on a proprietary acquisition setup developed by RISE. The ECHO-NN station hardware includes a stack of five GNSS receivers (RISE HAPPi HAT) connected to a multi-antenna setup, with the directional receivers logging baseband samples in SBF format at 1 Hz. This architecture monitors both observables and time-frequency data suitable for AI feature generation. The AI dataset combines real and simulated data. In the operational-station context, dataset labels were generated using classical interference-detection/characterization algorithms. In addition, controlled jamming campaigns performed by JOANNEUM RESEARCH provided traceable event timing and scenario logs, enabling manual annotation and consistency checks. This mixed-label strategy was practical for building a sufficiently large dataset while retaining direct links to real operational behavior and known interference injections.
The AI model for interference detection and characterization was implemented as a multi-task Convolutional Neural Network (CNN) operating on a short analysis window. The network jointly predicts: (a) interference presence, (b) interference type, and (c) interference parameters, i.e., central frequency (CF) and bandwidth (BW). This multi-task formulation was selected to align with operational needs: a single inference pass can generate both an alarm and a first-order characterization of the interferer. In the training formulation, CF and BW regression were treated as conditional tasks, meaning they were optimized and evaluated only on windows where interference was present. In practice, this was implemented through an indicator-mask term in the loss function so that regression errors were not learned from interference-free windows. This improves training stability and avoids contaminating parameter regression with irrelevant samples. Model validation was performed with frozen weights in a pseudo-real-time, window-by-window evaluation loop, using task-specific metrics (precision/recall/F1 for detection, type accuracy, and regression errors for CF/BW) and runtime profiling of the end-to-end inference chain.
Model selection and validation in our training workflow were driven by a combined objective that prioritized strong detection while penalizing poor parameter estimation, using a single scalar score that merged detection performance (primarily F1) with regression error terms. This design choice was important because it discouraged models that achieved high detection rates at the cost of unusable CF/BW predictions.
The AI results highlight that neural-network performance depends strongly on data-domain consistency, especially hardware configuration.
When the trained interference detector was evaluated on data acquired with a different receiver/antenna setup than the one used for training, we noticed a substantial degradation in performance. The model entered a conservative operating regime: it was able to recognize many interference-free windows, but exhibited very low recall, missing most true interference events. This led to a low F1 Score and accuracy below a random baseline. This is an indicative of a pronounced domain shift, likely driven by differences in antenna patterns, polarization, and RF front-end characteristics, such as gain, filtering and noise floor. The current model is not directly transferable across hardware setups without retraining or adaptation.
A more favorable outcome was observed when the model was tested on data collected from the same ECHO-NN station configuration but on a different day, specifically during the later jamming campaign. In this case, the trained model outperformed simple baselines (e.g., random guessing and trivial always-positive, always-negative assumptions) for interference presence detection. The model achieved stronger and more balanced performance, with improved accuracy and precision relative to trivial rules, indicating that it had learned non-trivial interference patterns from the time-frequency inputs. Importantly, the model maintained this behavior days after training, supporting temporal generalization within the same measurement setup.
The characterization results were also encouraging, because the network produced physically meaningful CF and BW estimates, with per-hour mean absolute CF error in the few-MHz range and per-hour mean absolute BW error in the tens-of-kHz range. In addition to continuous parameter estimation, the interferer-type classification accuracy was moderate to high, indicating that the model can provide useful post-detection characterization rather than only a binary alarm. These results support the practical value of the multi-task formulation: once interference is detected, the same network can recover spectral descriptors that are relevant for incident triage and downstream decision-making.
The ECHO-NN results demonstrate that a multi-task AI approach can support GNSS interference monitoring by jointly performing detection and classification in a single model. Under consistent hardware conditions, the neural network generalizes across time and delivers meaningful characterization outputs with near-real-time-capable runtime behavior. Moreover, the study shows that cross-setup portability remains a major challenge. Performance degradation on different receiver/antenna configurations indicates that model deployment across heterogeneous stations will require retraining, calibration, or domain-adaptation techniques. This finding is particularly relevant for operational GNSS resilience networks, where sensor diversity is common.
This work contributes to a practical AI validation framework for GNSS interference monitoring that combines: (i) real datasets, (ii) multi-task CNN inference, (iii) conditional regression for interference characterization, and (iv) pseudo-real-time testing. The main significance is twofold: first, AI can provide richer and more actionable interference characterization than simple alarm indicators; second, dataset representativeness and hardware consistency are decisive factors for reliable field performance. These results provide a concrete step for future operational AI deployments in GNSS resilience systems and motivate further work on cross-hardware robustness and adaptive model transfer.

Previous Abstract Return to Session B3 Next Abstract