【作者】 徐彬；

【导师】 刘宏伟； 金林；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2019， 博士

【摘要】 雷达高分辨距离像数据(High Resolution Range Profile,HRRP)是宽带雷达信号在发射到目标后得到的散射点子回波在雷达射线方向上的向量和,它反映了目标的物理结构特性,具有十分重要的研究价值。相比于二维SAR图像和ISAR图像,HRRP数据不需要目标相对于雷达有一定的转角,更易于获取和处理。随着雷达技术的逐渐发展,军事战争对于HRRP目标识别技术的需求越发迫切。因此,本文主要是围绕国家自然科学基金,青年千人基金以及国防预研项目,针对雷达高分辨距离像识别,分别从HRRP距离单元的时序相关性、HRRP的平移敏感性以及HRRP的方位敏感性等问题作为切入点,提出了不同的机器学习算法来有针对性地解决具体的问题。本论文的内容主要包括如下五个部分:1、简要地回顾了机器学习的一些基础理论知识,对分类任务进行了定义。除此之外,我们还详细地介绍了两种经典的神经网络模型,即深度置信网络和堆栈自编码器。2、传统的神经网络方法在进行识别的过程中仅仅考虑了HRRP样本的包络信息,而忽略了样本距离单元之间的相关性。本章针对这个问题,提出了一种循环神经网络投票模型(Recurrent Neural Network with voting Strategy,RNNvot),该模型将原始的HRRP数据通过滑窗转化为序列形式,并且采用循环神经网络模型对其进行提取特征并输出类别。循环神经网络模型输出的是类别序列,而HRRP序列特征对应于单个类别,因此,我们又进一步地采用投票机制将所有时刻的信息进行融合,输出最终的类别。3、传统的解决HRRP平移敏感性的方法有两种,一种是采用相对对齐方法,将所有的样本都同模板进行平移对齐,但是这种方法的计算量较大。另一种方法是提取平移不变特征,将原始时域数据转化为其他平移稳健的特征空间,该方法确实可以有效地解决平移敏感性问题,但是它改变了数据的时序相关特性,丢失了可分性的信息。本章提出了一种双向截断长短时记忆网络模型(Bidirectional Truncated Long-short Term Memory,BTLSTM),该方法首先采用一种截断机制提取目标区域,并依据目标区域提取平移稳健的输入特征,为了考虑双向的时序相关性,模型采用双向LSTM模型对稳健的输入提取可分性特征,并采用投票机制输出样本的类别。基于实测数据的实验结果表明当前的方法不仅可以有效地进行识别,而且对于平移敏感性非常稳健。4、针对HRRP数据不同距离单元对于最终识别的贡献不同的问题,提出了一种注意循环神经网络模型(Target-Aware Recurrent Attentional Network,TARAN)。具体地,我们利用循环神经网络模型来探索HRRP样本中距离单元之间的时序相关性,并设计了一种注意机制,该机制通过自动地学习不同时刻输入数据的权值来决定不同区域在识别中的贡献,然后对不同时刻的隐层特征进行加权求和作为最终的识别特征,并采用分类器输出类别。由于RNN的记忆功能和注意机制,与传统方法相比,TARAN不仅可以有效地进行识别,而且对于平移敏感性是鲁棒的。此外,我们还分析了谱图特征和时域序列特征对于识别性能的影响。5、HRRP数据存在方位敏感性,这使得不同方位角的HRRP样本的距离单元之间的时序相关性存在一定的差异。为了解决这个问题,本章提出了一个高斯混合-张量循环神经网络模型(Gaussian Mixture Model-Tensor Recurrent Neural Network,GMMTRNN)。具体地,该模型首先采用三维张量参数代替传统循环神经网络中的二维参数,然后采用高斯混合模型对不同时刻的输入进行聚类,并由聚类的结果选择三维张量参数中的矩阵作为当前时刻的参数提取特征,输出类别。相比于传统的循环神经网络模型,GMM-TRNN模型在不同时刻的参数是非共享的,且不同时刻的参数由输入数据来决定,可以更有效地表达HRRP数据。更多还原

【Abstract】 A High-Resolution Range Profile(HRRP)is composed of the amplitude of the coherent summations of the complex returns from target scatterers in each range cell,which represents the projection of the complex returned echoes from the target scattering centers onto the radar Line-Of-Sight(LOS).Compared with the two-dimensional image(SAR image,ISAR image),HRRP data does not need a certain angle of the target relative to the radar,which is easier to obtain and process.With the gradual development of radar technology,the military requirement for HRRP target recognition technology is becoming more and more urgent.Therefore,this dissertation provide our researches for HRRP target recognition,which are supported by the Thousand Young Talent Program of China,NSFC(61771361),111Project(B18039),and the National Science Fund for Distinguished Young Scholars of China(61525105).This dissertation aims at proposing different machine learning algorithms to deal with the aspect sensitivity,time-shift sensitivity and temporal correlation of HRRP.The content of this paper mainly includes the following five parts:1.This chapter briefly reviews some basic theories of machine learning and defines the classification tasks.In addition,two classical neural network models,deep belief network(DBN)and stacked autoencoder(SAE),are introduced in detail.2.The traditional neural network methods only consider the envelope information of HRRP samples,ignoring the temporal correlation between the range cells of HRRP sample.In order to deal with this issue,a Recurrent Neural Network with voting Strategy(RNNvot)is proposed in this chapter,which converts the original HRRP data into a sequence and extracts discriminative features with the Recurrent Neural Network model.The output of the recurrent neural network is a sequence,while the HRRP sample corresponds to a single category.Therefore,we further combine the information at all timesteps with the voting strategy to predict the final label.3.There are two traditional methods to solve the time-shift sensitivity of HRRP.One is the alignment method,which aligns all samples with the template sample.However,this method requires a large amount of computation burden.Another is to extract translation invariant features,such as frequency domain features,which can effectively solve the time-shift sensitivity.However,this method changes the characteristics of data and loses the discriminative information.This chapter provides a Bidirectional Truncated Long-short Term Memory(BTLSTM).Specifically,the method firstly uses a truncation mechanism to extract the target area and extracts time-shift robust representation based on the target area.In order to consider the bidirectional temporal correlation,the model adopts the bidirectional LSTM to deal with the input representation and uses the voting strategy to predict the sample category.The experimental results based on measured data show that the current method is not only effective for recognition,but also very robust for time-shift sensitivity.4.In this chapter,we develop a Target-Aware Recurrent Attentional Network(TARAN)for Radar Automatic Target Recognition(RATR)based on High-Resolution Range Profile(HRRP)to make use of the temporal dependence and find the informative areas in HRRP,since it reflects the distribution of scatterers in target along the range dimension.Specifically,we utilize RNN to explore the sequential relationship between the range cells within a HRRP sample and employ the attention mechanism to weight up each timestep in the hidden state so as to discover the target area,which is more discriminative and informative.Effectiveness and efficiency are evaluated on the measured data.Compared with traditional methods,besides the competitive recognition performance,TARAN is also more robust to time-shift sensitivity thanks to the memory function of RNN and attention mechanism.Furthermore,detailed analysis of TARAN model are provided based on time domain and spectrogram features.5.In this chapter,a Gaussian Mixture Model Tensor Recurrent Neural Network(GMMTRNN)is proposed for Radar Automatic Target Recognition(RATR)based on High-Resolution Range Profile(HRRP),since it reflects the distribution of scatterers in target along the range dimension.Specifically,in order to consider the temporal correlation between the range cells within a HRRP sample,we transform the sample into sequential feature and utilize the RNN model to process it.Furthermore,based on the hypothesis that the correlation between range cells being different at different timesteps in the sequential feature,a Tensor Recurrent Neural Network(TRNN)model is designed to deal with the sequential feature,whose parameters are not shared at different timesteps,but depends on the input representation at each timestep.In order to solve the problem that the continuous HRRP sample being not suitable for determining the parameters,we employ the Gassian Mixture Model to cluster the input representations in sequential features and utilize the clustering category to decide the parameters at each timestep in TRNN.Experimental results based on measured data show that the GMM-TRNN achieves competitive recognition performance compared with traditional methods.更多还原