基于改进ResNet18的GNSS压制式干扰分类识别方法

GNSS Jamming Classification and Identification Method Based on the Improved ResNet18

摘要:

作为现代社会关键的时空基础设施，卫星导航系统（GNSS）的定位、导航与授时服务广泛应用于交通、航空航天等核心领域，但其信号传输过程中功率衰减显著的固有脆弱性，使得干扰信号识别面临小样本、特征区分度低的突出难题。针对这一问题，提出一种改进ResNet18算法。该算法采用预训练权重迁移与选择性微调策略，重构含BatchNorm1d和Dropout的轻量化分类头，结合标签平滑交叉熵与StepLR调度优化训练流程，构建了涵盖北斗、GPS多频段的单音、脉冲、线性调频三类干扰信号数据集，完成了图像预处理与数据增强后训练。实验结果表明，模型测试准确率达99.64%，三类干扰单独识别准确率均≥98%，较原始ResNet18测试损失降低66.7%，训练-测试准确率差值仅3.1%，有效抑制过拟合，收敛效率提升50%，为GNSS压制式干扰信号精准识别提供技术支撑。

Abstract:

As a pivotal spatiotemporal infrastructure in modern society, the Positioning, Navigation and Timing (PNT) services of the Global Navigation Satellite System (GNSS) are widely applied in core fields such as transportation and aerospace. However, its inherent vulnerability of significant power attenuation during signal transmission brings about prominent challenges of small samples and low feature discriminability for jamming signal recognition. To solve this problem, an improved ResNet18 algorithm was proposed. This algorithm adopted the strategies of pre-trained weight transfer and selective fine-tuning, reconstructed a lightweight classification head with BatchNorm1d and Dropout, and optimized the training process by combining label-smoothed cross-entropy with the StepLR scheduler. A dataset of three types of jamming signals including single-tone, pulse and linear frequency modulation, covering the multi-frequency bands of BDS and GPS, was constructed, and the model training was completed after image preprocessing and data augmentation. Experimental results show that the model achieves a test accuracy of 99.64%, with the individual recognition accuracy of each of the three jamming types being no less than 98%. Compared with the original ResNet18, the model’s test loss is reduced by 66.7%, the difference between training and test accuracy is only 3.1%, which effectively suppresses overfitting and improves the convergence efficiency by 50%. It provides technical support for the accurate recognition of GNSS barrage jamming signals.