【作者】 余志斌；

【导师】 金炜东；

【作者基本信息】 西南交通大学 ， 信号与信息处理， 2010， 博士

【摘要】 雷达辐射源信号(RES)识别是雷达对抗信号处理中的关键技术,其识别水平已成为衡量电子对抗装备先进程度的重要标志。长期以来,人们主要依靠常规五参数实现RES的识别处理,这只有在常规雷达信号和信号密集度较低的情况下才能获得满意的识别效果。随着现代电子和雷达技术的发展以及它们在现代战争中的广泛应用,新型RES调制式样更加灵活,参数日益多变,电磁信号越来越密集,致使传统的识别技术已无法满足现代电子战的实际需要。因此,迫切需要探索新的RES识别方法,以提高我国电子对抗装备的技术水平。近年来,国内外学者对RES识别进行了大量的研究,提出了许多新方法,以提高RES识别水平。但是,这些方法一般要求较高的信噪比、较为明显的信号类型差异,且大都基于常规参数相对稳定、辐射源数量较少等情况,很少涉及调制类型相同、参数变化和低信噪比下调制类型不同的信号识别问题。一般都只关注特征提取和识别算法的研究,很少涉及特征识别能力和高密集电磁环境下异常脉冲的处理等RES识别问题的分析。而这些实际问题,已严重制约RES识别技术的进一步提高。因此,本文从特征的可分性、脉内特征提取、识别模型和算法设计等方面,针对上述复杂体制RES识别中需要解决的关键理论问题展开了系统、深入的研究,主要贡献有以下5个方面。1.基于概率论与统计理论,构建了一维特征参数的可分性模型,以分析提取特征的识别能力,为RES特征提取提供理论支持。在模型构建过程中,依据特征参数服从近似正态分布,研究并推导出了特征正确分类概率随相应特征统计参数变化的关系子模型。从而导出了特征参数要在理论上达到90%以上的正确分类概率,两类特征估计均值之差的绝对值与估计精度之比至少应大于3.3的结论。然后,定量地分析了常用特征参数的可分性度量。2.提出了脊线-频率特征(脊频特征)及其级联特征提取方法,为在低信噪比条件下,实现不同调制类型的RES的有效识别补充新的特征参数。首先从脊线定义出发,基于信号时-频尺度原理导出了小波脊频特征(WRF)提取的条件约束模型。然后,提出一种新小波原子和脊线提取策略对脊线提取算法进行改进,并提取了典型RES的WRF。在此基础上,对WRF进行特征降维再挖掘,提取了一组能较好描述RES脉内调变统计规律的级联特征。新的增量模糊支持向量机被用于检验提取特征的有效性。3.提出了小波包融合和融合熵特征提取方法,为实现类型相同、仅仅某些参数具有差异的近似雷达辐射源信号(ARES)识别构建有效的特征向量。首先使用小波包变换和主分量分析构建小波包融合算法。在此基础上,提取融合特征的融合香农熵、范数熵和概率熵,并对三种熵特征的抗噪性能进行详细分析。然后,进一步研究了LFM信号参数的识别问题。考虑到信号分解层数、特征维数和参数等多种影响因素,论文深入分析了不同识别算法与不同特征组合的识别性能。实验结果表明,使用融合熵特征向量不仅具有较好的ARES识别效果,而且算法复杂度远小于传统方法,基于实测和仿真数据给出了识别结果。4.对雷达功率放大器的非线性特性进行详细分析,从而导出了反映非线性特性的谐波功率约束模型,并提出了相应的谐波功率约束特征(HPRF)提取算法,为实现辐射源个体识别提供新的无意调制参数。在估计谐波功率时,为削弱噪声影响、提高估计精度,基于二项展开式,导出了谐波功率谱的自相关估计模型。比较放大器输入功率固定和变化两种情况下HPRF的稳健性结果,发现在输入功率变化时,HPRF的二维分布成近似线性关系。最后的测试结果表明HPRF具有良好的识别性能,并得出测试信号需要不小于400个脉冲的能量积累,才能达到论文中识别结果的结论。5.提出了一种增量模糊支持向量机识别算法用于雷达辐射源信号识别,以提高信号的识别率,并解决当前识别算法难以处理非库属目标、训练时间较长等问题,深入研究了算法设计过程中所涉及的相关理论问题和解决方案。其中,构建了训练样本的类隶属度模型,提出了确定最小超球体半径的支持向量模糊数据描述方法,并引入了平凡训练数据的概念。在此基础上,提出增量模糊学习算法。然后,基于属性理论构建了处理未知雷达信号的拒判规则,以控制虚警率。最后,考虑多种影响因素,通过若干实验,深入研究了不同参数和样本数量条件下算法的识别性能。本论文的研究工作得到国家自然科学基金(No.60572143, No.60702026)和国防科技重点实验室基金的共同资助。更多还原

【Abstract】 The recognition of radar emitter signals(RES) is a key technology in signal processing of radar countermeasures, and the recognition level of RES has become an important symbol of the technical merit of the radar countermeasures equipment. For a long time, traditional methods of recognizing radar signals are generally based on five conventional parameters, which these methods are effective and can obtain satisfactory results in the low dense environment. However, with the rapid development of electronic technology and radar technology, the modulation manner of RES became more and more complex and various, and the circumstance of RES became increasing denseness. As a result, the performances of these traditional methods descend rapidly. Therefore, only some new and valid approaches are explored to improve the technical merit of electronic countermeasure equipments.In recent years, though many scholars have helpfully explored a great of new methods to improve the recognition level of RES, these proposed approaches only analyze those signals in high SNR, obvious difference, fixed signal parameters, and fewer emitters. The existing methods are difficult to identify same modulation signals, parameter-changes signals and different modulation signals in low SNR. The feature-separability of radar signals and the singularity pulse processing in dense circumstance of RES are not also studied. In fact, the recognition technology of RES has been restricted by these problems. Thereout, from the view of four important facets, which are feature-separability, intra-pulse feature extraction, novel model and algorithm of signal recognition, recognition methods of advanced RES have been studied. The research fruits are as follows.1. In order to explore the problem of the radar signal feature-separability, the one-dimensional feature-separability model of the signal feature is built based on the probability and statistics theory. According to parameters obey approximate normal distribution, and the relationship sub-model of the correct classification probability and correspondence feature statistics parameters is proposed. As a result, the correct classification probability is more than 90% when the ratio value of the measure precision and the absolute value of the difference of two feature mean values is not less than 3.3. Afterwards, the separability measurement of convention parameters in signal recognition is gained.2. The feature extraction algorithms of ridge-frequency features and Cscade Cnnection features of the ridge-frequency feature are proposed. Via these algorithms, some new parameters can be extracted, so that different modulation signals can be recognized. According to the time-frequency principle and the definition of the ridge-line, the condition restriction model of the ridge-frequency feature is constructed, and then the improved ridge-line feature extraction algorithm is proposed based on a new wavelet atom and extraction strategy of ridge-line. After extracted the ridge-frequency of radar emitter signals, the Cscade-Cnnection features of the feature are extracted to describe the modulation characteristics of the signal. The results of classification experiments based on increment fuzzy support vector machine demonstrate that Cscade-Cnnection features of the ridge-frequency featrure can reflect the difference of different modulation signals, and have a good ability to resist noise.3. The wavelet packet fusion algorithm and the feature extraction algorithms of fusion entropy features are proposed to construct the effective recognition feature vector for approximately radar emitter signals(i.e., the modulation manner of the signal is identical, but some parameters of the signal are different). In this method, the choice rule of the wavelet is presented, and the fusion algorithm is reconstructed based on the wavelet packet decomposition and the principle component analysis. Similarly, the fusion Shannon entropy, fusion Norm entropy and fusion Probability entropy are extracted to describe the energy structure of the signal, and analyzing the resisting noise capability of three entropy features. Afterward, the parameters of LFM are estimated. Considering the number of wavelet decomposition layers, dimension numbers of the feature and various parameters, it is been researched detailedly that the recognition performances of the different features based on different recognition algorithms in this paper. The experiment results show that the proposed approach not only achieves good in recognition effect, but also suffers less computational burden than traditional methods.4. According to analyses of the nonlinearity of the radar power amplifier, the harmonic power restriction model is constructed to describe the nonlinear characteristics of the amplifier, and the correspondence feature extraction algorithm of harmonic power restriction (HPR) is proposed. Via this algorithm, some unintentional-modulation features are extracted to recognize the emitter. In this algorithm, the correlation estimation model of the harmonic power of the signal is proposed based on two-term formula. Comparing with the solidity of the HPR feature in different power conditions, the linear relationship of HPR features is obtained when input power of the amplifier is various. The experiment results have shown that these conclusions can be drawn in this paper, if energy accumulation of the pulse signal is enough.5. Aiming at unknown radar signal processing, lower signal recognition rate and longer training time of the existence signal recognition algorithm, an increment fuzzy support vector machine algorithm is proposed to improve the signal recognition rate, and the correspondence theory and solving-scheme are studied detailedly in the design of the algorithm. In this algorithm, the combine membership function of every training example and the support vector fuzzy data description method for confirming the radius of the hypersphere are proposed, the conception of the common training data set is introduced and the increment fuzzy training algorithm is also presented. Then, in order to control the loss alert ratio, the rejection strategy of unknown radar emitter signals is proposed based on the attribute theory. Lastly, considering all kinds of effect factors, it is been studied deeply that the recognition performances of the recognition algorithm based on different parameters and training example numbers.This work is supported by the National Natural Science Foundation of China (No.60572143, No.60702026) and the National Electronic Warfare Laboratory Foundation.更多还原