Detecting and Classifying GNSS Interference with Machine Learning on Real-Field Jammertest Data: A SWaP-Constrained Comparison of Tree-Based Ensembles and Sequence-Based Deep Learning
Rissa, Aleksi Juho Rudolf (2026)
2026
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-08-27
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202608279311
https://urn.fi/URN:NBN:fi:tuni-202608279311
Tiivistelmä
Global Navigation Satellite Systems (GNSS) provide the Positioning, Navigation and Timing (PNT) backbone of critical infrastructure. The inherently low received signal power and unencrypted civilian signal structure make GNSS susceptible to intentional interference. Mitigation logic in commercial receivers remains proprietary, while most published machine learning (ML) detectors rely on simulated data that cannot replicate real multipath, ionospheric fades, or receiver-dependent noise.
An ML framework is developed and evaluated for detecting and classifying GNSS interference using post-correlation measurements from the 2025 Jammertest field trials at Andøya, Norway. A set of 55 candidate features, comprising ten statistical measures computed over a 30-epoch sliding window across five base tracking parameters, is reduced through Pearson correlation filtering and Shannon mutual information ranking. Three tree-based ensemble classifiers (XGBoost, LightGBM, Random Forest) and three sequence-based deep learning models (Transformer, CRNN, TCN) are trained and evaluated against a data-calibrated threshold baseline using AGC and pseudorange standard deviation.
Contrary to initial expectations, the tree-based ensembles outperformed the Transformer. XGBoost classified 86% of the held-out evaluation windows correctly, compared with 84% for the Transformer and 60% for the threshold baseline. Explicit upstream feature engineering captured the physics of the interference more accurately than allowing a sequence model to learn temporal structure from windowed feature vectors.
Each model was also profiled for SWaP (Size, Weight and Power)-constrained edge deployment, where LightGBM offered the lowest resource footprint (213 MB peak RAM, 0.23 ms P99 latency). For embedded GNSS interference detection, engineered features paired with lightweight gradient-boosted ensembles provide a better accuracy-to-cost trade-off than sequence-based deep learning models.
An ML framework is developed and evaluated for detecting and classifying GNSS interference using post-correlation measurements from the 2025 Jammertest field trials at Andøya, Norway. A set of 55 candidate features, comprising ten statistical measures computed over a 30-epoch sliding window across five base tracking parameters, is reduced through Pearson correlation filtering and Shannon mutual information ranking. Three tree-based ensemble classifiers (XGBoost, LightGBM, Random Forest) and three sequence-based deep learning models (Transformer, CRNN, TCN) are trained and evaluated against a data-calibrated threshold baseline using AGC and pseudorange standard deviation.
Contrary to initial expectations, the tree-based ensembles outperformed the Transformer. XGBoost classified 86% of the held-out evaluation windows correctly, compared with 84% for the Transformer and 60% for the threshold baseline. Explicit upstream feature engineering captured the physics of the interference more accurately than allowing a sequence model to learn temporal structure from windowed feature vectors.
Each model was also profiled for SWaP (Size, Weight and Power)-constrained edge deployment, where LightGBM offered the lowest resource footprint (213 MB peak RAM, 0.23 ms P99 latency). For embedded GNSS interference detection, engineered features paired with lightweight gradient-boosted ensembles provide a better accuracy-to-cost trade-off than sequence-based deep learning models.