【作者】 陈凤；

【导师】 保铮；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2009， 博士

【摘要】 随着现代战争的不断信息化、智能化,雷达自动目标识别技术受到广泛关注。现代雷达工艺技术的发展为雷达目标识别提供了多种可用的信号形式,近年来雷达目标识别领域不断有新的成果面世,使其逐渐从理论研究向工程实现迈进。本文主要围绕着预研项目“目标识别技术”以及“重点实验室基金项目”等的研究任务,从宽带雷达一维高分辨距离像(HRRP)自动目标识别和窄带雷达自动目标分类两方面进行了研究,主要内容概括如下:1.基于宽带雷达HRRP的目标识别研究包括四部分内容:第一部分:较为详细地分析了HRRP的姿态敏感性,主要讨论了飞机类目标对偏航、俯仰、侧摆三维姿态角变化的敏感性、飞机类目标在正侧视附近的特点以及测试样本的“相干峰”现象。进而,为了合理划分HRRP方位角域,提出了一种自适应递归划分角域的建模方法,利用有概率含义的统计模型分类器,从雷达数据中提取连续HRRP序列中包含的非线性结构信息,递归地对雷达数据自适应划分角域。第二部分:研究了HRRP强度和平移敏感性。现有HRRP雷达自动目标识别(RATR)方法一般对平移敏感性采用包络对齐及一些改进方法,对强度敏感性简单地能量归一化。显然,现有处理方式对强度敏感性与平移敏感性的优化是分离开来的,因此其匹配精度并不高。我们提出强度和平移联合优化匹配思想,并将其应用到独立高斯模型、PCA子空间统计模型和PPCA子空间统计模型,以提高平移和强度匹配精度。该思想可用于模板库建立和测试阶段。第三部分:提出了一种稳健噪声的自适应统计识别方法。雷达目标识别希望能够在较远距离实现,因此识别算法对噪声的稳健性是HRRP目标识别工程化需要研究的一个问题。我们分别基于PPCA和FA统计模型提出了强噪声污染的测试样本如何匹配无噪声或弱噪声样本训练的模板。第四部分:在线建库是HRRP目标识别工程化的一条途径。这一部分提出两种在线自适应HRRP识别方法:1).通过一种在线混合专家(OME)将HRRP数据在线地分割成若干个近似平稳的区域,在各个区域内使用平稳协方差函数的在线高斯过程分类器(OGPC)。针对迭代在线高斯过程分类器(IOGPC)的参数学习算法EP和EM,提出了一种双链高斯过程(Bi-OGP)来使OGPC的参数得以在单次数据扫掠的情况下实现在线更新。针对迭代在线混合专家门网络参数的学习算法EM,提出了基于初值选择的单次数据扫掠学习方法。2).在独立高斯模型假设下,推导了参数在线学习的公式,提出双门限法来剔除“坏值”样本并适当缓减假设模型与实时HRRP多模数据的失配。2.基于窄带雷达的目标分类研究内容由于窄带信号分类可以作为宽带HRRP识别的一种预处理,并且大量现役装备雷达是分辨率较低的窄带雷达,本论文研究了如何利用窄带信号喷气发动机调制(JEM)特征实现对喷气式飞机、螺旋桨飞机和直升机目标的分类。大多数雷达,特别是地面警戒雷达,其脉冲重复频率相对较低,会导致JEM回波多普勒模糊,此外,对目标的观测时间(扫掠时间)相对较短,多普勒分辨率较低,会导致目标分类性能下降,这正是利用JEM特征进行目标分类的需要克服的困难。我们从模式分类的角度提出了利用JEM特征谱散布程度特征进行目标分类的方法。JEM回波在多普勒域近似看作是一系列线谱,采用谐波和的数学模型提取特征谱作为分类特征,分别给出脉间、脉内的特征谱提取方法及特征的降维方法。该特征不补偿机身回波,对机身多谱勒变化不敏感。更多还原

【Abstract】 With the tendency of modern battle to become more and more information-based and intelligent, radar automatic target recognition(RATR) techniques have received intensive attentions. The development of modern radar technology supplies many usable signal forms for RATR, and many new productions in the RATR field have come forth in recent years. As a result, RATR is striding to practical realization from theoretical study. This dissertation provides our researches for RATR from two aspects, i.e. wideband radar High-resolution range profile (HRRP) automatic target recognition and narrow band radar automatic classification.The main contents of this dissertation are summarized as follows.1. The research of HRRP automatic recognition based on wideband radar includes four parts:In the first part, we discuss the sensitivities of airplane targets with the variation of course, pitch and roll angle, the property of the HRRPs near the airplane target’s broadside, and coherent apex phenomena in test data. Furthermore, due to the target-aspect sensitivity, we propose a recursive algorithm for adaptively angular sector segmenting which exploit the nonlinear structure characteristic embedded in HRRP data through classifiers that have probability meaning.The second part focuses on the sensitivities of amplitude-scale and time-shift in radar HRRP statistical recognition. Most approaches available just use the slide correlation processing or its modifications for time-shift sensitivity while simply normalizing the amplitude-scale. Obviously, approaches available solve the two sensitivity problems disjointedly, which leads to mismatch and limits recognition performance. In order to improve the matching precision, we propose two algorithms to jointly match the amplitude and time-shift, based on independent Gaussian model, PCA subspace and PPCA subspace statistical model. One of the algorithms is used in the training phase and the other one is used in the test phase.The third part presents a noise-robust adaptive statistical recognition method. We hope to recognize target at long distance in RATR, and therefore, the robustness study of recognition algorithm is necessary. In this part, based on PPCA model and FA model, a robust method for adaptive statistical recognition is presented when test SNR is lower than training SNR.The fourth part focuses on how to get data, learn and build the template database interactively and concurrently. We present two online methods. 1) An online mixture of experts (OME) is used to divide the HRRP data into several regions within which online Gaussian process classifier (OGPC) make predictions. Due to EM and EP, parameter learning methods for iterative online Gaussian process (IOGP), a Bi-online Gaussian process (Bi-OGP) is proposed to learn parameters by single pass of data. Due to EM algorithm for OME’s parameter learning, a single data pass method based on proper initial values is presented. 2) With the assumption that returned echoes in range cells is independent Gaussian distributed, online parameters is first given, and then a two thresholds method is proposed to pick the outliers and reduce the mismatch between the statistics model and the online real HRRP data.2. Narrow band radar automatic classificationBecause narrow bandwidth signal target classification is a pre-processing method of HRRP RATR and the bandwidth of a lot of radar equipments in use is narrow, we study how to use Jet Engine Modulation (JEM) characteristic of low-resolution signal to categorize aeroplanes into three kinds, i.e., turbojet aircraft, prop aircraft and helicopter.Most low-resolution radar systems, especially ground surveillance radar systems, work at relatively low pulse repeat frequency (PRF) and with short time-on-target (TOT) (duration in scanning). Low PRF leads to Doppler ambiguity and short TOT results in low Doppler resolution, which poses a problem to target classification with low-resolution radar based on the JEM characteristic of radar echo. From the pattern classification viewpoint, using dispersion situations of JEM eigenvalue spectra, we propose a method for categorizing aeroplanes into three kinds, i.e., turbojet aircraft, prop aircraft and helicopter. We analyze the mathematical model of JEM echoes consisting of a series of line spectra and regard them as a sum of several series of harmonious waves. Classification features can be extracted based on the harmonious wave sum model. Some schemes for extracting features from echoes within or between pulses are proposed. Low-dimensional features are extracted to reduce computation burden. Our methods do not compensate for the fuselage echoes and are insensitive to the variation of fuselage Doppler.更多还原