【作者】 吴杰；

【导师】 周建江；

【作者基本信息】 南京航空航天大学 ， 信号与信息处理， 2012， 博士

【摘要】 雷达自动目标识别技术在军事和民用上都有着巨大的应用价值。宽带雷达高分辨距离像（HRRP）具有目标结构信息丰富、易于获取的特点，是雷达目标识别重要的发展方向之一。本文以探寻稳健、实用的识别算法为目标，围绕HRRP目标识别的关键问题，系统地研究了姿态敏感性、特征提取、噪声背景下的稳健识别和序列识别问题，为HRRP目标识别的实用化进行了有益的探索。主要研究内容和创新点如下：1．基于散射点模型理论，研究了不同方位角下等角域划分对目标识别的影响，通过对HRRP方位敏感性较强的方位角区域细分，以相邻HRRP的互相关系数为基准自适应地划分角域，弥补了最大相关模板匹配法（MCC-TMM）均匀划分角域的缺陷，改善了识别性能。2．针对HRRP目标识别中简单套用经典概率分布模型存在的“模型失配”问题，研究了逆向云模型的建模问题，给出了云滴确定度和逆向云隶属度的计算方法，提出了基于逆向云模型的雷达目标识别方法。仿真表明，相比经典的Gaussian模型，该方法具有识别精度高、对训练数据量依赖性小、对目标姿态变化不敏感、抗噪性能好等优点。3．当目标在某一角域内的散射特性失配时，其对应角域HRRP能量呈现非线性分布特性。针对这一问题，提出了一种基于核主分量分析（KPCA）重构的雷达目标识别方法。该方法在等角域划分下利用核主分量分析提取每个角域内HRRP的特征子空间，再将测试样本投影到各角域特征子空间中进行重构，最后通过计算最小重构误差来判别测试样本的类别。相比主分量分析重构方法和最大相关系数模板匹配法，核主分量分析重构方法可以松弛角域划分范围，降低角域划分的精度要求，同时也具有较好的抗噪性能。4．针对传统的线性判别分析（LDA）算法应用于HRRP特征提取时存在的“四个缺陷”（要求假设待分类数据服从具有相同协方差矩阵的高斯分布；降维后的特征子空间维数受限；在计算散射矩阵时没有突出边界样本在目标识别中的作用；待分类样本的维数大于或接近于样本个数时，容易造成所谓的“小样本问题”），研究了HRRP非参数特征提取方法：（1）提出了基于非参数特征分析（NFA）和逆向云模型相结合的HRRP目标识别方法。NFA算法在计算散射矩阵时用局部KNN均值代替类均值；利用样本的局部信息来构建类间散射矩阵，增加了类间散射矩阵的秩；通过权函数增强了训练数据中类边界样本在分类中的作用。弥补了LDA算法的前三个应用缺陷。利用逆向云模型作为分类器改进了概率论和模糊数学在处理不确定性问题方面的不足，更加符合目标HRRP经过特征提取后特征子空间模糊分布的实际情况；（2）提出了基于非参数最大间隔准则的雷达目标识别方法。该方法结合最大间隔准则和非参数化方法的优点，采用“以差化商”的方法解决了LDA算法的“小样本”问题，用非参数的方法计算类内和类间散射矩阵弥补了LDA算法的其它三种应用缺陷。仿真实验表明，相比参数的特征提取方法，所提出的非参数特征提取方法可以增加HRRP样本的类内聚合性和类间可分性，从而提高目标识别率和抗噪性能。5．针对传统的HRRP识别方法对噪声环境适应性差的问题，提出了一种噪声背景下的HRRP目标识别方法。该方法通过分析不同信噪比下幂次变换(PT)参数的选取对识别效果的影响，利用线性回归的方法给出参数选取的经验公式；结合信噪比实时估算，研究了基于自适应幂次变换的数据预处理方法。根据自相关小波变换的时移不变性特性和较好的抗噪性能，构造自相关小波SVM分类器。仿真实验表明，该方法在目标识别率和噪声稳健性方面远优于高斯核SVM分类器。6．为获得更加稳健、可信的识别效果，设计了HRRP序列的雷达目标识别模型，并根据该模型详细论述了雷达目标的序列识别方法；针对模型中的主干环节，基于概率推理理论，引入灰色关联算子，构造了HRRP序列目标识别算法中MYCIN的不确定因子，提出了一种基于HRRP序列的雷达目标识别方法。与单样本的识别算法相比，所提出的算法具有识别精度高、稳定性好、抗干扰能力强等优点，具有较好的工程应用前景。更多还原

【Abstract】 The radar automatic target recognition is of great value in military and civilian use.High-Resolution Range Profile（HRRP） is characterized by rich information and accessibility，so itis one of the most important developments of radar target recognition. This thesis intends to explorerobust and practical recognition algorithms and the critical problems of HRRP recognition, includingtarget-aspect sensitivity, feature extraction, the robust recognition under noise background andsequence recognition, which is a useful exploration of the application of HRRP’ target recognition.The main contents and the new ideas are as follows:1. According to the scattering center model, the impacts of the equiangular division underdifferent target-aspect on the target recognition is studied in this paper. A new frame segmentationmethod based on cross correlation coefficient is proposed. Comparing with the Max CorrelationCoefficient-Template Matching Method (MCC-TMM), the presented method can efficiently improverecognition performance.2．In the radar target recognition based on the statistic model, the problem of model mismatch isexisted when using classical parametric probability density model to describe the statistical propertiesof HRRP. Due to this problem, this thesis studies the modeling of the backward cloud model, andfigures out the algorithms of the certainty degree of cloud drop and the backward cloud membership,and proposes the radar target recognition method based on backward cloud model. Simulation resultsdemonstrate that the proposed approach is effective for radar target identification. Compared withGaussian models, it has advantages of higher identification accuracy, better anti-noise performance,more relaxed requirements for azimuth division, and can also achieve good recognition results in thecase of fewer training samples.3. When scatter feature within one angular sector mismatches, the HRRP energy within itscorresponding angular sector shows the feature of non-linear distribution. In view of this, an approachbased on kernel principle component analysis reconstruction is proposed in this thesis. Kernelprinciple component analysis is used to extract eigen subspace in every equiangular sector for a start,and then test sample is reconstructed by projecting it onto the eigen subspace of each angular sector,finally, the type of test sample is determined by the minimum reconstruction error. Simulations resultsshow that the proposed approach relaxes the angular sector, requires lower angular division precision, and meanwhile it has a better anti-noise performance compared with principle component analysisreconstruction method and MCC-TMM.4. As an effective feature extraction method in radar target HRRP recognition community, LinearDiscriminant Analysis (LDA) faces four main shortcomings. First, it relies on the assumption that thesamples in each class satisfy Gaussian distribution with the same covariance matrix; Second, thedimension of the eigen subspace has an upper limit after dimension reduction; Third, the effect ofboundary samples are not highlighted when calculating scatter matrix; Fourth， when the dimensionof samples is more than or close to the number of samples, the so-called “Small Sample Sizeproblem” is likely to emerge. To tackle these problems, the methods of nonparametric featureextraction are studied in this paper:(1) a radar target recognition method based on nonparametricfeature analysis (NFA) and backward cloud mode is proposed. NFA algorithm uses the local KNNmean instead of class mean when calculating scatter matrix; the rank of between-class scatter matrixis increased when it’s calculated with samples’ local information; the effect of class boundary samplesare strengthened by weighting function. So the first three defects of LDA algorithm can be offset inNFA. The deficiency of probability theory and fuzzy mathematics in dealing with uncertain problemsis improved by using backward cloud model, it is more consistent with the fuzzy distribution of eigensubspace’s for the target HRRP after feature extraction;(2) a radar target recognition method based onNonparametric Maximum Margin Criterion (NMMC) is proposed. This method integrates theadvantages of maximum margin criterion and nonparametric feature extraction. The quotientoperation is replaced by difference operation in NMMC, in order to solve the problem of small samplesize in LDA algorithm, and the other three application defects of LDA algorithm are offset by usingnonparametric method to calculate within-class and between-class scatter matrix.Compared with parametric feature extraction, the simulation experiment demonstrates thatnonparametric feature extraction could increase within-class polymerization and between-classdivisibility, so it can improve target recognition rate and noise robustness.5. As traditional algorithms are not robust to noise, this thesis proposes HRRP’s targetrecognition against the noisy environment. In this method, the formula of parameter’s selection isgiven by analyzing the impacts of power transformation parameter’s selection with different SNRs onrecognition effect, and the data-preprocessing based on adaptive power transformation is studiedintegrating the estimation method for real-time SNR. And an auto-correlation wavelet support vectormachine is proposed as the classifier, the kernel of which is constructed with a compactly supportedwavelet satisfies the translation invariant property. The simulation results show that the proposed algorithm is of better noise-robustness and higher recognition rate compared with Gaussian kernelSVM.6. To achieve steadier and more reliable recognition result, a sequence model of HRRP for radartarget recognition was designed in the thesis. For the backbone of this model, a radar targetrecognition algorithm is proposed based on HRRP’s sequence. In this method, grey incidence operatoris introduced into the probability-reasoning theory, and the calculation method of MYCIN’s uncertainfactor is given based on the framework of Bayes theory. Compared with recognition algorithms with asingle sample, the proposed algorithm has good prospects for engineering applications with higherrecognition rate, better stability, and stronger anti-interference.更多还原