| 摘要: |
| 针对辐射源个体识别任务中训练数据与实际应用场景数据之间存在的特征分布差异问题，提出了一种新颖的基于元学习的辐射源个体识别方法。该方法能够通过学习不同任务间的共性特征，提高特征提取器的泛化能力，在目标域样本较少和不需要重新训练模型的情况下实现粗粒度的特征对齐。该方法旨在通过有效的数据增强策略，扩展目标域数据的多样性，在元学习的框架下训练并得到泛化性强的初始化参数。在此基础上引入微调策略以实现更为精准的特征对齐与分类效果，显著增强了其在新场景下的泛化性能。实验结果表明，该方法不需要重新训练模型，在LoRa和WiSig数据集上的迁移准确率可达到96%。此外，该方法凭借其轻量化设计，有效降低了计算与存储需求，可以部署于分布式节点。 |
| 关键词: 辐射源个体识别 闭集识别 迁移学习 元学习 |
| DOI：10.20079/j.issn.1001-893x.241019001 |
|
| 基金项目:国家自然科学基金资助项目(62371463) |
|
| Meta-learning-driven Individual Identification of Radiation Sources |
| MEN Haoxuan,WU Hao,QIAO Xiaoqiang,DU Yihang,ZHANG Tao,ZHANG Jiang |
| (The 63rd Research Institute，National University of Defense Technology，Nanjing 210007，China) |
| Abstract: |
| For the problem of feature distribution differences between the training data and the data of real application scenarios in the task of radiation source individual identification，a novel meta-learning-based method for radiation source individual identification is proposed.The method is able to improve the generalization ability of the feature extractor by learning common features across different tasks，and achieve coarse-grained feature alignment with fewer samples in the target domain and without the need to re-train the model.The method aims to extend the diversity of target domain data through an effective data enhancement strategy to train and obtain initialization parameters with strong generalization in the framework of meta-learning.On this basis，a fine-tuning strategy is introduced to achieve more accurate feature alignment and classification results，which significantly enhances its generalization performance in new scenarios.Experimental results show that the method does not need to retrain the model and can achieve 96% transfer accuracy on LoRa and WiSig datasets.In addition，the method effectively reduces the computation and storage requirements by virtue of its lightweight design and can be deployed in distributed nodes. |
| Key words: radiation source individual identification closed set identification transfer learning meta-learning |
《电讯技术》编辑部 