A GNSS Interference Signal Identification Scheme Based on Meta-Learning for Few-Shot Conditions

With the widespread adoption of Global Navigation Satellite System (GNSS), the radio magnetic environment surrounding GNSS receivers has become increasingly complicated. The satellite navigation signal may be subject to intentional or unintentional interference, contributing to the reduction of the positioning accuracy of GNSS. Therefore, it is indispensable for the system to identify the types of interferences. Recognizing the types of interferences enables the prompt implementation of specific anti-interference measures, thereby enhancing the performance of the satellite navigation system. The current research on GNSS interference recognition is primarily based on sufficient data-driven approaches. However, in real-world scenarios, the intelligent recognition of GNSS interference signals encounters challenges, including low-quality datasets and a scarcity of labeled samples, resulting in a few-shot problem. The mainstream approach to solve the few-shot problem involves expanding and enhancing the number of samples or features at the data level, rather than focusing on the model algorithm level. Addressing the above set of issues, this paper proposes a GNSS interference signal identification scheme based on meta-learning for fewshot conditions. In the scheme, we utilize the Model-Agnostic Meta-Learning (MAML) as an algorithm to optimize the model parameters and improve the traditional training process. Replacing the traditional training process with MAML is to develop a model with strong generalization capabilities that can quickly adapt to new tasks with minimal interference signal samples. This paper combines MAML with an enhanced Residual Neural Network (ResNet) architecture incorporating Squeeze-and-Excitation (SE) channel attention mechanisms, introducing the MAML+SE-ResNet model. SE-ResNet is used for feature extraction and recognition monitoring of GNSS interference signal time-frequency maps in a small number of samples scenarios. MAML, on the other hand, randomly extracts multiple sample data from the dataset to form a task so that the network could be trained at the task level. Actual experiments show that under different Jamming-to-Noise Ratios (JNR), the proposed MAML+SE-ResNet model exhibits significantly higher recognition accuracy for few-shot GNSS interference signals compared to state-of-the-art traditional machine learning models. The results indicate that MAML+SE-ResNet outperforms the pre-trained models in generalization performance across various JNR. The meta-learning-based GNSS interference signal recognition solution for few-shot conditions proposed in this paper holds great practical significance and application value.