QGSL-GNSS: Q-Learning-Based Adaptive Graph Structure Learning for Semi-Supervised GNSS Interference Classification

Abstract

Jamming devices often pose a significant threat to Global Navigation Satellite System (GNSS) signals by disrupting position navigation and timing (PNT) services. Consequently, reliable detection and classification of interference types are essential for developing effective anti-jamming strategies. However, most existing approaches rely on supervised learning and require large volumes of labeled datasets for training. In practice, obtaining such fully labeled GNSS interference data in time sensitive scenarios is challenging, which limits the effectiveness of conventional supervised approaches. Additionally, these approaches lack the ability to capture global relationships between samples. To address this limitation, this work proposes a reinforcement-learning-guided graph-based framework for GNSS interference classification. Specifically, a Q-learning agent is employed to learn an optimal graph structure from both labeled and unlabeled data, where Q-values encode similarity relationships between samples. Additionally, these learned Q-values are further utilized for quadruplet selection to improve metric learning using informative positive and negative interference samples. Subsequently, a Graph Convolutional Neural Network (GCN) is trained using a joint objective consisting of cross-entropy loss on labeled data and quadruplet loss on unlabeled data to simultaneously enable jammer classification while enforcing structured embedding separation. By integrating RL-guided graph construction with metric learning, the proposed approach effectively improves inter-class separability and intra-class compactness, enabling efficient and robust interference classification under limited labeled data conditions. Particularly, only 5% of the training samples per class are labeled, and the remaining samples are treated as unlabeled. Finally, the proposed approach is evaluated on both a real-world collected dataset and a publicly available benchmark dataset. The real-world dataset consists of 7 jammer classes with 2000 samples per class, whereas the public benchmark dataset contains 6 jammer classes with 1000 samples per class. For both datasets, identical preprocessing and train-test split protocols are applied, and the same proportion of 5% labeled-samples per class is maintained to ensure a fair semi-supervised evaluation. Experimental results demonstrate jammer classification accuracies of 91.31% on the real-world dataset and 94.74% on the public benchmark dataset, respectively.