【作者】 李博；

【导师】 赵志欣； 邵启红；

【作者基本信息】 南昌大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 在数字调幅广播(Digital Radio Mondiale,DRM)外辐射源雷达中,由于采用了非合作的正交频分复用(Orthogonal Frequency Division Multiplex,OFDM)波形通信信号作为探测源,因此为实现对目标的检测、跟踪和定位,通常需设置参考通道以获取直达的DRM广播信号作为参考信号。为了从被多径干扰和噪声污染的参考通道信号中提取出纯净的DRM广播信号,传统方法利用数字广播信号本身优良的抗多径和噪声性能,通过“解调-再调制”的方法重构出纯净的参考信号。然而该方法涉及多种通信技术,其实现需全程依赖于具有专业知识的技术人员。随着深度学习方法在无线通信领域的迅速发展和广泛应用,特别是其高效的特征提取和灵活的参数调整能力,为通信领域中的智能恢复无线传输信号提供了多种解决方法。在此背景下,本文深度挖掘参考通道信号实测数据中的回波信息和导频信息等波形有用特征,探索基于深度学习的DRM外辐射源雷达智能参考信号提取方法。具体研究内容如下:(1)利用实测数据中导频信息所包含参考通道CSI特征,从优化参考信号提取部分模块的角度出发,借助卷积神经网络和双向长短期记忆网络实现传统信道估计模块的网络化,从而精确地估计参考通道进而提升参考信号提取的性能。同时针对DRM外辐射源雷达实测数据样本有限和无真实标签的情况,利用最小二乘(Least Square,LS)信道估计方法从实测数据中获得信道参数信息,通过仿真模拟信道生成训练数据集。最后,结合仿真数据以目标信噪比为指标验证所提方法的可行性。(2)利用实测数据的时域参考通道回波数据的时序相关性,从优化参考信号提取部分模块的角度出发,借助双向长短期记忆网络和全连接网络搭建去噪网络,从而减少参考通道中的噪声污染进而提高参考信号提取的性能。同时针对DRM外辐射源雷达实测数据样本有限和无真实标签的情况,基于Noise as Clean思想,通过对时域参考通道回波数据添加额外噪声生成自监督的训练标签,满足外辐射源雷达的实测应用需求。最后,结合仿真数据以目标信噪比为指标验证所提方法的可行性。(3)同时利用实测数据中导频所含的信道响应特征和参考通道回波数据中的时序相关性特征,从参考信号提取过程整体优化的角度出发,提出了基于编码-解码结构的导频条件参考信号提取网络(Pilot Conditional Reference Signal Extraction Network,PCRSENet)。PCRSENet以导频处CSI作为除接收回波外的条件输入,隐含地估计CSI并自适应的实现导频信息特征与参考通道接收信号特征融合,从而直接输出参考信号。同时,为了解决实测数据中网络训练标签无法获取的难题,提出了LS信道估计辅助的训练集生成方法。最后,结合实测数据以目标信噪比为指标验证所提方法的有效性。更多还原

【Abstract】 In the Digital Radio Mondiale(DRM)-based passive radar,non-cooperative communication signals with Orthogonal Frequency Division Multiplexing(OFDM)waveform are used as the detection source.Therefore,to achieve detection,tracking and positioning of the targets,the radar system typically needs to set a reference channel to obtain a direct DRM broadcast signal as the reference signal.In order to extract the pure DRM broadcasting signal from the reference channel signal polluted by multipath interference and noise,the traditional method utilizes the excellent antimultipath and anti-noise performance of the digital broadcasting signal itself to reconstruct the pure reference signal through the "demodulation-remodulation" method.However,this method involves multiple communication technologies and requires the full reliance on technical personnel with professional knowledge for its implementation.With the rapid development and wide application of the deep learning method in wireless communication field,especially its efficient feature extraction and flexible parameter adjustment capabilities,it has provided a variety of solutions for intelligent recovery of wireless transmission signals in the communication field.In this context,this paper deeply utilizes the useful waveform features such as temporal echo and pilot information in the reference channel signals of the measured data,and explores the intelligent reference signal extraction method based on deep learning for DRM-based passive radar.The specific research contents are as follows:(1)Using the CSI characteristics contained in the pilot information of the reference channel in the measured data,from the perspective of optimizing part module of reference signal extraction,the traditional channel estimation module is networked by Convolutional Neural Network(CNN)and Bidirectional Long and Short Term Memory(BiLSTM)network,so that the reference channel is accurately estimated and the performance of reference signal extraction is improved.At the same time,in view of the limited sample size and lack of real labels for the DRM-based passive radar measurement data,the least square(LS)channel estimation method was used to obtain channel parameter information from the measured data,and then the information is used to simulate the analog channel to generate a training dataset.Finally,the feasibility of the proposed method is verified by combining simulation data and using the target Signal-to-Noise Ratio(SNR)as the evaluation metric.(2)Using the time-domain correlation of the reference channel echo data in the measured data,from the perspective of optimizing part module of reference signal extraction,a denoising network is constructed by the BiLSTM network and the fully connected(FC)network,thereby reducing noise pollution in the reference channel and improving the performance of the reference signal extraction.For DRM-based passive radar with limited data samples and no real label,based on the idea of Noise as Clean,the self-supervised training label is generated by adding additional noise to the timedomain reference channel echo data,which satisfies the practical application of passive radar.Finally,the feasibility of the proposed method is verified by combining simulation data and using the target SNR as the evaluation metric.(3)Using the CSI characteristics contained in the pilot information of the reference channel and the timing correlation of the reference channel echo data in the measured data,from the perspective of overall optimization of the reference signal extraction process,a pilot conditional reference signal extraction network(PCRSENet)based on encoding decoding structure is proposed.PCRSENet takes the CSI at the pilot as the conditional input in addition to the received echo,implicitly estimating the CSI and adaptively implementing the fusion of pilot information features and reference channel received signal features,thereby directly outputting the reference signal.At the same time,in order to solve the problem that network training labels cannot be obtained from measured data,we propose a LS channel estimation aided training sets generation method.Finally,the effectiveness of the proposed method is verified by measured data and using the target SNR as the evaluation metric.更多还原