A VAE-Based Framework for GNSS Jammer Classification Using Time-Frequency Image Representations

Global Navigation Satellite System (GNSS) signals are highly susceptible to intentional radio-frequency jamming, threatening safety- and time-critical services. We propose a data-efficient monitoring pipeline that converts raw GNSS snapshots into time–frequency spectrograms and employs a Variational Autoencoder (VAE) for unsupervised feature learning. Trained exclusively on nominal data, the VAE captures a latent distribution that enables faithful reconstructions; jamming is flagged whenever the reconstruction error exceeds a learned threshold. On a synthetic dataset covering six representative jammer classes—AM, chirp, FM, pulse, narrowband and Distance-Measuring-Equipment (DME) — the approach reached an overall anomaly-detection accuracy of 90.0 %, correctly identifying 97 % of all jammed examples. It consistently surpassed CNN and SVM baselines, yielding a precision of 93.25 % and an F1-score of 91.39 %. Detection was perfect for DME jammers (100 %) and exceeded 92 % for both AM and FM jammers, underscoring the model’s ability to isolate structured interference. Latent-space visualization further reveals clear separability between normal and jammed signals. The proposed framework therefore offers an interpretable, real-time solution for GNSS interference surveillance and provides a foundation for recognizing emerging jamming patterns without expensive annotation effort.