Development of Machine Learning Based Models for Detecting GNSS Signal Jamming in Real-World Scenarios Using AGC Data

Although Global Navigation Satellite Systems (GNSS) provide high-precision positioning under optimal situations, they remain vulnerable to intentional Radio Frequency Interference (RFI) due to the low signal power received at ground level and the increasing availability of jamming devices. Previous studies have demonstrated that traditional jamming detection methods, such as fixed-threshold-based detection, are limited in handling dynamic scenarios and therefore remain less efficient under real-world conditions. This study addresses these limitations by applying machine learning (ML) techniques to detect GNSS jamming using AGC data. The primary objective of the study is to analyse supervised and unsupervised machine learning (ML) models for detecting jamming using multidimensional features derived from AGC signals. The theoretical foundation is based on anomaly detection, statistical learning, and time-series signal analysis. Two machine learning models were applied: a supervised classifier (XGBoost) and an unsupervised anomaly detector (Isolation Forest). Both were trained and validated on AGC data measured during a controlled jamming test in Norway. The experiment is conducted under various jamming conditions, which enable realistic and reproducible data collection in a dynamic vehicular environment. Feature extraction was performed using the sliding-window approach across various GNSS frequency bands, retaining the temporal and spectral dynamics of AGC. The supervised model was trained on labelled samples to distinguish normal and jammed windows, while the unsupervised model was trained on normal data only to identify anomalies. The test consisted of typical classification metrics and interpretability methods, such as feature importance analysis, to understand the model's decisions. Both models demonstrated the ability to detect jamming effectively. The supervised model showed strong performance, revealing that AGC features from specific frequency bands, particularly the G1 band, were most indicative of jamming. The unsupervised model performed reliably on normal data, although it generated some false predictions, which may be attributed to sudden variations in positioning data. Visual analysis offers a deeper understanding of the relationship between prediction results and the behavior of positioning parameters. The proposed ML-based approaches provide better adaptability, eliminate the need for manual threshold tuning, and scale effectively across various conditions. They are well-suited for real-time deployment in GNSS receivers. Future work may enhance these models by combining them to provide a total localization solution.