辐射源个体识别中的深度特征表示与多核融合方法研究

Specific emitter identification (SEI) system discerns radio emitters based on the extracted individual features. Specific radar emitter identification, i.e. distinguishing each individual radar emitter, is one of the important research spots in radar countermeasures field, which is related to national security and thus we must pay great attention. However, for the current complex electronic warfare environment, the mechanism analysis of individual features has become more and more difficult and the problem of small sample size in non-cooperative mode becomes prominent, which constrain the further development of specific radar emitter identification. Therefore, considering the existing research results and the pre-research experiments of our team, this project will resort to the popular deep learning and multiple kernel learning methods to deal with the above mentioned problems. The contents of the project are: (1) in regular cooperative mode, studying the deep learning based identification methods by means of mining the discriminant information from the envelope or ambiguity function features, which utilizes the existing mechanism analysis results and can reduce the difficulty of the mechanism analysis; (2) in non-cooperative mode, studying three strategies to cope with the small sample size issue, i.e., the identification method of deep feature learning combined with kernel classifiers, the identification method of data augmentation and the identification method of deep metric learning, respectively; (3) in the framework of information fusion, studying two strategies to fuse a variety of feature representations via multiple kernel learning based kernel fusion methods, i.e., fusing the patches of ambiguity function image and fusing the primary features with the output features extracted from each hidden layer of deep networks. By improving the discriminant of primary individual features and fusing different feature representations, it is hoped that the technology of specific radar emitter identification will enter a new stage of development.

辐射源个体识别通过提取个体特征来辨识无线电辐射源个体。雷达辐射源个体识别，即识别雷达个体，是雷达对抗领域的热点研究课题，关乎国家安全，必须高度重视。然而，在当前复杂的电子战环境下，雷达辐射源个体识别面临机理分析困难、非合作工作模式下小样本问题突出的挑战，制约着个体识别技术的向前发展。基于此，本项目借助深度学习和多核学习，主要研究：（1）合作模式下基于深度学习的识别方法，从包络和模糊函数等源自既有机理分析的个体特征中深挖判别信息，降低机理研究的难度；（2）非合作模式即小样本条件下，基于深度特征学习与核分类器相结合、基于数据扩增以及基于深度度量学习的识别方法；（3）在多核学习理论和算法框架下，基于模糊函数各分块特征相融合以及各初级特征与深度网络各隐层输出特征相融合的识别方法。本项目从增强雷达辐射源信号初级特征的判别能力、融合多元信息表达等思路出发，希冀推动雷达辐射源个体识别技术进入新发展阶段。

雷达辐射源个体识别是电子对抗领域的研究热点，它对所截获的雷达信号进行特征分析与分类识别，从而区分不同的雷达个体，为威胁分析和告警提供支撑。围绕该任务，本项目主要开展了以下工作：（1）对2020年之前（包括2020年）雷达辐射源个体识别方向的研究进展进行了文献综述，包括雷达辐射源个体特征机理分析、基于手工特征的识别方法、基于深度学习的识别方法以及数据集构建四个方面，并对当前现状和未来方向进行了总结与展望。该综述是迄今为止雷达辐射源个体识别领域最新、最全的综述，旨在推动雷达辐射源个体识别理论和方法研究的新发展，对整个领域的发展具有积极作用。（2）基于多核融合的思想，提出了一种基于模糊函数分块特征和两阶段多核极限学习机的雷达辐射源个体识别方法。该方法有效利用了信号模糊函数平面上的有用信息，并且采用了两阶段多核学习算法，识别精度和效率均得到了提升。特别地，该方法在小样本条件下具有较大优势，能够缓解目标处于非合作模式时训练样本有限的问题。此外，该方法所采用的核判别比值准则能够自动地识别出时频平面上信息量丰富的区域，非常适用于复杂多变的雷达信号。（3）基于多核学习和判别分析，提出了一种基于多核判别分析框架的雷达辐射源融合识别方法。该方法能够同时满足识别率和计算效率两方面的要求，很适合雷达辐射源个体识别系统。（4）基于深度学习的思想并考虑模型训练效率，提出一种基于多核多层极限学习机的雷达辐射源个体识别方法，取得了良好的性能。较之其他深度学习算法，多层极限学习机的计算效率较高，更适合用于实时系统。