基于机器学习的高频地波雷达复杂杂波识别技术研究

发布时间：2018-02-15 09:00

本文关键词： 高频地波雷达机器学习特征提取特征筛选卷积神经网络　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：高频地波雷达(High Frequency Surface Wave Radar,简称HFSWR),基于高频(3MHz-30MHz)垂直极化波与海面的相互作用,能探测超出地球曲率的远处目标,对于国家安全、航海安全以及环境检测具有至关重要的意义。然而返回的电磁波受到如海杂波、电离层杂波、大气噪声、流星余迹、电台干扰等多种因素影响,极大影响了HFSWR的目标检测。如今主流的杂波抑制方法,如空时、时频、图像、建模分析等,都有其针对性处理的杂波以及适应性应用的条件。若将高频地波雷达回波距离多普勒谱(Range-dopper,简称RD谱)自适应的区分成不同类型的杂波,不仅可以自适应的将不同类型的杂波送入不同的抑制方法模块里,后续的目标检测策略还可根据杂波的类型作出相应调整,实现更智能的雷达处理模式。本文的最终目的是能很好的区分高频地波雷达背景中的不同杂波。由于实测HFSWR回波RD谱中各类杂波的边界是未知的,本文首先利用海杂波产生的原理、雷达方程、地波衰减原理以及电离层杂波的统计特性对HFSWR模拟仿真。本文将介绍相关的原理,制作不同海态环境下关于杂波的仿真环境以及仿真数据的杂波类别标签。对杂波的认知和抑制一直是研究的重点。本文先针对现有对杂波认知的方法,建立杂波特征库,相继提取了杂波的功率谱幅值、无维分流参数、小波多尺度参数、Gabor方向性参数、统计拟合参数、Bragg峰理论位置信息、杂噪比以及联合参数作为杂波独特的特征向量。并基于特征统计、信息理论以及分类效果对特征分析和筛选,选取最优的特征组合进行基于支持向量机的杂波分类,得到了满意的分类效果。本文提出了一种基于卷积神经网络(Convolutional neural network,简称CNN)的杂波分类方法。介绍了对仿真RD谱的预处理以及送入CNN的子图像,经CNN训练学习后,将训练好的网络效果与特征筛选算法对比分析。这种方法同样具有比较满意的分类效果,为杂波的背景分类提供了一种新方法。最后,本文将上述方法应用在实测数据中,提出了一种自动分类算法来对实测数据定量评价,并提供了GUI(Graphical User Interface)界面便于自动提取算法的进一步完善。本文提出的方法在实测数据中也具有可行性。
[Abstract]:High Frequency Surface Wave radar (HFSWR), based on the interaction of high frequency 3MHz-30MHz waves with the sea surface, can detect distant targets beyond the curvature of the earth. Safety of navigation and environmental detection are of vital importance. However, returned electromagnetic waves are affected by many factors, such as sea clutter, ionospheric clutter, atmospheric noise, meteor trace, radio interference, etc. It has greatly affected the target detection of HFSWR. Nowadays, the mainstream clutter suppression methods, such as space-time, time-frequency, image, modeling and analysis, etc. If the echo range Doppler spectrum of HF ground wave radar is adaptively divided into different types of clutter, Not only can different types of clutter be adaptively sent into different suppression methods, but the subsequent target detection strategies can also be adjusted accordingly according to the type of clutter. The final purpose of this paper is to distinguish the different clutter in the background of high frequency ground wave radar. The boundary of all kinds of clutter in the measured HFSWR echo Rd spectrum is unknown. This paper first uses the principle of sea clutter generation, radar equation, ground wave attenuation principle and statistical characteristics of ionospheric clutter to simulate HFSWR. The simulation environment of clutter and the label of clutter category of simulation data in different sea environment are made. The recognition and suppression of clutter is always the focus of the research. Firstly, aiming at the existing methods of recognition of clutter, the clutter signature library is established in this paper. The power spectrum amplitude, dimensionless shunt parameters, wavelet multiscale parameters and Gabor directional parameters of clutter are extracted successively, and the theoretical position information of Bragg peak is obtained by statistical fitting parameters. Based on feature statistics, information theory and classification effect, the best feature combination is selected to classify clutter based on support vector machine. In this paper, a clutter classification method based on convolution neural network Convolutional neural Network (CNNs) is proposed. The preprocessing of simulated Rd spectrum and the sub-images sent to CNN are introduced. After training and learning by CNN, this paper proposes a new method of clutter classification based on Convolutional neural Network (CNNs). The trained network effect is compared with the feature selection algorithm. This method also has a satisfactory classification effect, which provides a new method for background classification of clutter. Finally, the above method is applied to the measured data. An automatic classification algorithm is proposed to quantitatively evaluate the measured data, and the GUI(Graphical User interface is provided to further improve the automatic extraction algorithm. The method proposed in this paper is also feasible in the field data.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN959;TP181

【参考文献】