【作者】 于鑫；

【导师】 国强；

【作者基本信息】 哈尔滨工程大学 ， 信息与通信工程， 2020， 硕士

【摘要】 在现代信息化战争中,有效准确的识别截获的雷达波形对电子侦察具有重要意义。随着低截获概率(Low Probability of Intercept,LPI)雷达等复杂体制雷达的出现,雷达在功能和性能上得到了很大的提升,传统的识别算法难以对其进行分类,这给电子侦察带来了巨大的挑战。本文主要针对复杂电磁环境下LPI雷达波形特征提取困难和识别率低等问题,提出了基于模糊函数(Ambiguity Function,AF)主脊切面包络特征和时频图像结构特征以及两种特征相融合的三种LPI雷达波形识别算法。论文主要研究内容如下:1.在模糊函数主脊切面包络特征提取与识别方面,提出了基于AF主脊切面和TPOT的雷达波形识别算法。该算法首先提取雷达波形AF主脊切面包络和最大旋转角;接着通过奇异值分解算法对主脊切面包络去噪;然后提取AF主脊切面包络特征,包括维数特征和相像系数特征,并联合最大旋转角特征构成特征集;最后,选择并优化分类器,使用基于树结构的流程优化工具(Tree-based Pipeline Optimization Tool,TPOT)实现雷达波形分类。仿真结果表明,本文所提算法在低信噪比下对特征提取以及分类都具有良好的效果。2.在自主学习时频图像结构特征提取与识别方面,首先,针对低信噪比条件下LPI雷达波形特征提取困难以及识别率低等问题,提出了基于卷积神经网络和支持向量机(Lenet-5-SVM)相结合的LPI雷达波形识别算法。该算法将雷达波形通过时频分析转换为二维时频图像,利用图像处理算法提取时频图像中的主成分信息,并采用卷积神经网络联合支持向量机实现雷达波形的特征提取与识别;然后,针对LPI雷达波形训练样本数目少,深层CNN的参数训练困难的问题,将迁移学习的思想引入到雷达波形识别算法中,分析支持向量机与深层网络模型分类器的关系,提出了两种基于深度迁移学习的雷达波形识别算法(Inception-v3-SVM、Res Net-V2-152-SVM)。仿真实验表明,所提算法提高了低信噪比下雷达波形识别率,验证了所提网络模型在LPI雷达波形识别领域应用的可行性与有效性。3.在联合特征识别方面,针对低信噪比下迁移学习模型识别率低的问题,设计了基于AF主脊切面特征联合图像特征迁移学习的网络。通过将AF主脊切面包络特征与迁移学习算法提取的时频图像特征进行融合,实现了LPI雷达波形的分类识别。仿真结果表明,所提联合特征识别算法提高了低信噪比下LPI雷达波形的识别准确率。更多还原

【Abstract】 In the modern information war,it is of great significance for electronic reconnaissance to identify the intercepted radar waveform effectively and accurately.With the emergence of low probability of intercept(LPI)radar and other complex radar systems,the radar has been greatly improved in function and performance.The traditional recognition algorithm is difficult to classify it,which brings great challenges to electronic reconnaissance.In this paper,three LPI radar waveform recognition algorithms are proposed to solve the problems of the difficulty in extracting LPI radar waveform features and the low recognition rate in the complex electromagnetic environment,which include of the envelope feature of the main ridge section of the ambiguity function(AF)and the structure feature of the time-frequency image,besides that the fusion of the two kind features.The main work of the thesis are as follows:1.In the aspect of envelope feature extraction and recognition of main ridge section of ambiguity function,a radar waveform recognition algorithm based on AF main ridge section and TPOT is proposed.The algorithm first extracts the envelope and the maximum rotation angle of the AF main ridge section of the radar waveform;then extracts the envelope features of the AF main ridge section,including the dimension features and resemblance coefficient features,and combines the maximum rotation angle features to form the feature set;finally,selects and optimizes the classifier,to realize the radar waveform classification using the treebased pipeline optimization tool(TPOT).The simulation results show that the algorithm proposed in this paper has good effect on feature extraction and classification at low signal-tonoise ratio(SNR).2.In the aspect of self-learning time-frequency image structure features extraction and recognition,firstly,aiming at the problems of low SNR LPI radar waveform feature extraction and low recognition rate,an LPI radar waveform recognition algorithm based on convolution neural network(CNN)and support vector machine(SVM)is proposed.The algorithm transforms radar waveform into two-dimensional time-frequency image by time-frequency analysis,extracts the principal component information of time-frequency image by image processing algorithm,and realizes the feature extraction and recognition of radar waveform by convolutional neural network and support vector machine.Then,aiming at the problem that LPI radar waveform training samples small and deep CNN parameters training difficult,the idea of transfer learning is put forward in the radar waveform recognition algorithm.The relationship between SVM and deep network model classifier is analyzed,and two radar waveform recognition algorithms based on depth transfer learning(perception-v3-svm,resnetv2-152-svm)are proposed.The simulation results show that the proposed algorithm improves the radar waveform recognition rate under low SNR,and verifies the feasibility and effectiveness of the proposed network model in the field of LPI radar waveform recognition.3.In the aspect of joint feature recognition,aiming at the problem of low recognition rate of transfer learning model under low SNR,a network based on AF main ridge section feature and image feature depth transfer learning is designed.By fusing the envelope feature of AF main ridge section with the time-frequency image feature extracted by depth transfer learning algorithm,the LPI radar waveform classification is realized.The simulation results show that the proposed joint feature recognition algorithm improves the recognition accuracy of LPI radar waveform at low SNR.更多还原