无线电信号识别作为智能频谱监管的关键技术之一，其主要目标是通过检测非法或恶意的频谱使用，以确保频谱资源的合法利用。然而，随着无线电设备数量的爆炸式增长，日益复杂的电磁环境使得无线电信号的准确识别变得更加困难。近年来，深度学习技术凭借其强大的自动特征提取能力，提高了无线电信号的识别精度。然而，基于深度学习的无线电信号识别方法仍存泛化能力较差及小样本场景下识别准确率低等问题。因此，本文针对无线电信号精确识别问题，围绕数据增强，知识与数据融合两个内容开展研究，对智能频谱管控和认知侦察具有重要意义。具体而言，本文的主要工作包括以下两个方面：

（1）为了提高数据驱动的无线电信号识别方法的特征提取能力，本文提出了基于混合数据增强的无线电信号识别方法。首先，在数据层面引入了变分模态分解算法，结合时间序列变换设计了混合数据增强方法。其次，在网络模型层面针对算力和存储受限的边缘设备，基于稀疏字典理论设计了轻量级的稀疏残差网络(SRNet)。在仿真数据和实测辐射源数据下开展了对比实验，验证了所提混合数据增强方法的有效性和无线电信号识别模型SRNet的识别性能，该模型在保证较少参数量和较快推理速度的同时，有效地提高了模型在数据量充足和小样本场景下识别精度。

（2）在充分利用数据驱动方法提取特征能力的基础上，本文提出了基于数据和知识混合驱动的无线电信号识别方法。进一步地，引入无线电领域的专家知识辅助模型来挖掘潜在的特征信息。为了最大限度地发挥数据驱动和专家知识各自的优势，设计了混合驱动神经网络(HDNet)，提高了小样本场景下的识别准确性和稳定性。HDNet通过将数据驱动模型提取的深度特征与基于专家知识提取的人工特征相融合，改善了小样本场景中特征空间的不完备性。为了验证所提方法在真实场景中性能表现，搭建了基于USRP的无线电信号采集实验平台。在仿真数据和实测接收机数据下的实验结果表明，相比于基于数据驱动的识别方法， HDNet在小样本场景下能够实现更精确地无线电信号识别。

As one of the key technologies in intelligent spectrum management, the radio signal recognition aims primarily to detect illegal or malicious spectrum usage to ensure the legitimate utilization of spectrum resources. However, with the explosive growth in the number of wireless devices, the increasingly complex electromagnetic environment has made the accurate recognition of wireless signals more challenging. In recent years, deep learning technology has improved the recognition accuracy of wireless signals by virtue of its powerful automated feature extraction capabilities. Nevertheless, deep learning-based wireless signal recognition methods still suffer from poor generalization ability in complex environments and low recognition accuracy in few-shot scenarios. Therefore, this paper addresses the precise identification of wireless signals by focusing on two aspects: data augmentation and the fusion of knowledge and data, which are crucial for intelligent spectrum regulation and cognitive reconnaissance. Specifically, the main contributions of this paper are as follows:

(1) To enhance the feature extraction capabilities of data-driven methods for radio signal recognition, this paper proposes hybrid data augmentation-based approach. Firstly, a variational mode decomposition algorithm is introduced at the data level, coupled with temporal sequence transformations to devise the hybrid data augmentation method. Furthermore, at the network model level, aiming at edge devices with limited computing power and storage, a lightweight sparse residual network (SRNet) is designed based on sparse dictionary theory. Comparative experiments conducted on simulated datasets and measured radiation source data validate the effectiveness of the proposed hybrid data augmentation method and the recognition performance of the wireless signal recognition model SRNet. This model effectively improves the recognition accuracy of the model in scenarios with sufficient data and few-shot samples while ensuring fewer parameters and faster inference speed.

(2) Building upon the robust feature extraction capabilities inherent in data-driven methods, this paper proposes a hybrid data augmentation-based approach for radio signal recognition. Furthermore, expertise from the domain of radio communication is incorporated to aid the model in uncovering latent feature information. To maximize the advantages of data-driven and expert knowledge, a Hybrid-Driven Neural Network (HDNet) is designed to improve the recognition accuracy and stability in few-shot scenarios. HDNet achieves this by integrating the high-dimensional features extracted by data-driven models with manually extracted features based on expert knowledge, thus addressing the incompleteness of feature space in few-shot scenarios. To validate the performance of the proposed methods on real wireless data, a wireless signal collection experimental platform based on USRP devices is constructed. Experimental results on simulated datasets and measured receiver data demonstrate that compared to data-driven wireless signal recognition methods, the designed HDNet achieves more precise wireless signal recognition in few-shot scenarios.