【作者】 刘强；

【导师】 曹建蜀；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2020， 硕士

【摘要】 随着现代电子战技术的快速发展,雷达所面临的新型有源干扰具有低功率、高相参和强欺骗特性,对雷达的生存和作战带来了巨大的威胁。为了更好的进行干扰抑制,雷达需要对干扰样式进行识别,以采取针对性的抗干扰措施。然而,传统的干扰识别方法需要人工分析和提取各类特征,通用性差,泛化能力弱,难以适应瞬息万变的复杂对抗环境,因此,迫切需要提出更为稳健和智能的干扰识别方法。本文针对雷达有源干扰的识别分类问题,分别采用卷积神经网络和半监督生成对抗网络实现干扰信号端到端的监督和半监督学习,相比传统方法取得了更好的效果。全文的主要工作围绕干扰的建模分析、数据集设计、干扰提取和识别分类方法展开,具体研究内容如下:1、分析九种典型的雷达有源干扰的信号的产生机理并建立模型,给出其时域、频域和时频域的波形,为干扰识别提供理论支持。2、干扰信号深度学习需要统一格式数据集,论文仿真各类干扰信号,对其进行时频变换、归一化、平滑滤波、自适应裁剪等处理以突出特征统一格式,并通过设置不同干噪比、不同参数分别得到大量训练和测试数据集。3、论文给出了基于信号特征提取的干扰识别过程,分析并选择干扰信号多域的特征应用于支持向量机算法中,实现对干扰的分类。仿真结果证明选取的特征和分类方法对九种干扰有较好的分类效果,可作为深度学习方法的对照。4、传统方法需要人工特征提取效率不高且对干扰参数变化敏感,针对这一问题,本文基于统一格式干扰数据集首先采用卷积神经网络(CNN)深度学习方法实现干扰的监督学习,仿真表明其总体识别率相比传统方法显著提高并具有更好的鲁棒性;其次采用GAP-CNN进行改进,在大幅度减少网络参数和训练时间情况下仍然保持很高的识别率,提高了雷达干扰深度学习识别的应用潜力。5、实际应用场景中干扰样本难以广泛标注,需针对少量有标注干扰及大量无标签干扰情况进行学习,本文采用半监督生成对抗网络SSGAN和改进模型GC-SSGAN实现了干扰样本的半监督学习,设计了合适的识别器和生成器网络,在较低的标注率(小于2%)样本下进行有效学习,达到了较高的识别率(大于80%),与同样条件下基于CNN的干扰识别方法相比,显著提高了识别率。更多还原

【Abstract】 With the rapid development of modern electronic warfare technology,the new type of active interference that radar faces is characterized by low power,high coherence and strong deception,which poses a huge threat to the survival and operation of radar.In order to suppress interference,the radar first needs to identify the interference so that can take targeted anti-interference measures.However,traditional interference recognition methods require manual analysis and extraction of various features,which are poor in generality and weak in generalization,and it is difficult to adapt to rapid changes of complicated adversarial environment,so it is urgent to propose a more robust and intelligent interference identification method.For the recognition of radar active interference,this paper respectively uses convolutional neural network algorithm and generative adversarial network algorithm to realize supervised and semi-supervised learning of interference,its results are better than traditional methods.The main work of this paper focuses on the modeling and analysis of interference,dataset design,interference feature extraction and recognition algorithm.The specific research content is as follows:1.Model and analyze the 9 types of typical radar active interference,analyze their characteristics of time domain,frequency domain and time-frequency domain,and lay a theoretical foundation for their identification.2.Interference signals deep learning requires a unified format dataset.The paper simulates various types of interference signals,and performs them through time-frequency conversion,normalization,smooth filtering,adaptive cropping and other processing to get the unified format feature,and obtain training and test datasets of different JNR and different parameters.3.The paper presents the interference recognition process based on signal feature extraction,analyzes and selects the multi-domain features of the interference signal and applies them to the support vector machine algorithm to achieve the classification of interference.The simulation results prove that the selected features and classification methods have a good classification effect on the 9 types of interference,which can be used as a control for deep learning methods.4.Traditional methods require artificial feature extraction and sensitivity to changes in interference parameters.To solve this problem,this paper uses convolutional neural network(CNN)deep learning methods to achieve supervisedlearning of interference based on a unified format dataset.Simulation shows that its overall recognition rate is significantly improved and model has better robustness compared with traditional methods.Secondly,the improved model GAP-CNN keeps a high recognition rate under the condition of greatly reducing network parameters and save much training time,and improves application potential of the radar interference deep learning.5.Interference samples are difficult to be widely labeled in actual application scene,So it is necessary to learn from a small amount of labeled interference and a large number of unlabeled interference.In this paper,semi-supervised generative adversarial nets(SSGAN)and GC-SSGAN is used to implement semi-supervised learning of interference samples,and a suitable discriminator and generator is designed.The model learns effectively under a lower labeling rate(less than 2%)samples,and achieves a high recognition rate(greater than 0.8).Compared with CNN model under the same conditions,the recognition rate is significantly improved.更多还原