【作者】 徐湛；

【导师】 万建伟；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2012， 博士

【摘要】 海杂波严重影响着海事雷达的工作性能，海杂波物理特性和统计特性的研究对雷达信号处理、海面目标检测、雷达模拟器设计等应用领域起着关键的作用。随着非线性科学的发展，海杂波被认识到具有混沌分形的许多特征，这使人们开始利用神经网络方法来处理海杂波时间序列和进行目标检测。本文针对电磁散射建模、海杂波统计建模与仿真、海杂波背景下的目标检测三个方面展开研究，主要内容包括：针对传统电磁散射系数计算的双尺度模型将粗糙海面视为高斯海面，没有考虑迎风和顺风方向上海面的倾斜和不对称性，本文通过引入高阶统计量双频谱刻画海面的这种不对称性，提出了一种基于未充分发展的完全海谱（Non-fullyDeveloped Full-range Sea Spectrum，NDFSS）分形海面的电磁散射系数计算的修正双尺度模型。针对处理多个脉冲雷达回波数据等情况需要同时考虑海杂波的时间相关特性和空间相关特性，提出了一种基于球不变随机过程（Spherically Invariant RandomProcess，SIRP）的时间空间相关K分布海杂波生成方法，该方法生成的海杂波序列满足指定的时间空间相关性要求，在某外协项目中该方法得到了具体应用。针对以K分布作为幅度统计模型存在对不均匀区域建模能力不足，及其概率密度函数（Probability Density Function，PDF）表达式中存在修正的贝塞尔函数，算法复杂度较高等问题，而G0分布则具有广泛均匀度变化下的杂波区域建模能力和较强的模型兼容性，同时具有较易的工程可实现性，且其表达式中仅含有指数函数等特点，因此利用G0分布作为幅度统计模型对实测海杂波数据进行拟合，并与K分布进行对比分析，发现G0分布与实测海杂波数据拟合效果更好，通过库尔贝克-莱伯勒（Kullback-Leibler，KL）值度量、均方误差（Mean Square Error，MSE）检验及柯尔莫诺夫-斯米尔诺夫（Kolmogorov-Smirnov，K-S）检验证明了此结论。根据海杂波具有混沌分形的许多特征，本文使用了回声状态网络（Echo StateNetwork，ESN）预测海杂波时间序列，并利用含有目标和不含目标海杂波时间序列预测值与实测数据之间显著的MSE差异进行目标检测。还利用了ESN的改进方法如解耦回声状态网络（Decoupled Echo State Network，DESN）、具有最大可用信息的解耦回声状态网络（DESN with Maximum Available Information，DESN+MaxInfo）、具有储备池预测的解耦回声状态网络（DESN with ReservoirPrediction，DESN+RP）、以及两种具有简化结构的前馈回声状态网络（Feed ForwardEcho State Network，FF-ESN）和输入抽头延迟线（Tapped Delay Line with Inputs，TDL-I）方法进行了类似地海杂波时间序列预测和目标检测。使用了ESN方法检测海尖峰，并检测目标所在距离门，为海尖峰的深入研究和分析提供了一种新的思路。更多还原

【Abstract】 Sea clutter impacts seriously on the performance of maritime radars. The researchon physical and statistical characteristic of sea clutter palys a key role in radar signalprocessing, target detection in sea surface and radar simulator design. Withdevelopment of nonlinear science, sea clutter is realized that it has much character inchaos and fractal. This makes people use neural network method to deal with sea cluttertime series and detect targets. We carry out research on electromagnetic scatteringmodel, statistical model and simulation of sea clutter, target detection in sea clutter. Thework mainly contains:The conventional two-scale model for electromagnetic scattering coefficentcomputation regards coarse sea surface as Gaussian sea surface, and does not considerslope and asymmetry of sea surface in upwind-downwind direction. In this paper theasymmetry of sea surface is depicted by introducing higher-order statistical bispectrum,and a modified two-scale model for electromagnetic scattering coefficent computationbased on a Non-fully Developed Full-range Sea Spectrum (NDFSS) fractal sea surfaceis presented.Temporal correlation characteristic and spatial correlation characteristic of seaclutter are considered simultaneously for multi-pulse radar echo data, a generationmethod of temporal-spatial correlated K-distributed sea clutter is proposed based onspherically invariant random processes (SIRP). The generated sea clutter satisfies thedemand of appointed temporal-spatial correlation, and has been applied in a cooperativeproject.K-distribution as amplitude statistical model for heterogeneous area is deficient.The expression of Probability Density Function (PDF) for K-distribution has themodified Bessel function, and the algorithm complexity is much higher.G0-distribution has modeling ability in clutter regional of widely homogeneous degreevariation and strong model compatibility. This distribution has easily engineeringrealizability, and it’s expression only contains index function. So we useG0-distribution as amplitude statistical model for the sea clutter data fitting. Comparedto K-distribution, we found the results ofG0-distribution and sea clutter real-life datahave better fitting effect. Kullback Leibler (KL) distance, Mean Squared Error (MSE) testand Kolmogorov-Smirnov (K-S) test are used for proving our conclusion.According to sea clutter with character in chaos and fractal, echo state network(ESN) is used in predicting sea clutter time series, and the great MSE differencesbetween prediction value for sea clutter time series with target and without target andreal-life data are available for detecting target. Improved ESN methods similarly predictsea clutter time setires and detect target, and these methods are decoupled echo state network (DESN), DESN with maximum available information (DESN+MaxInfo),DESN with reservoir prediction (DESN+RP), feed forward echo state network(FF-ESN), and tapped delay line with inputs (TDL-I).ESN method is used to predict sea spikes, and detect target that exists in one rangebin. This method provides a new route to further research and analyse the sea spikes.更多还原