节点文献
【作者】
李晓东；
【导师】
宿富林；
于长军；
【作者基本信息】
哈尔滨工业大学
，
信息与通信工程，
2020，
博士
【摘要】 高频地波雷达(High Frequency Surface Wave Radar,HFSWR)利用高频电磁波(High Frequency Electromagnetic Wave,HFEW)在海面传播损耗低的特点,可对专属经济区甚至更远海域实现大面积和全天候的实时监测,因此,HFSWR受到世界各国的高度重视。HFSWR可分为宽波束HFSWR和窄波束HFSWR,其中窄波束HFSWR不仅可以提供超视距目标的位置和运动信息,而且可以进行远距离海洋表面动力学要素(Ocean Surface Dynamics Elements,OSDEs)即海流流速、有效波高、风速和风向等的精细化提取,监测面积更大且具有更好的角度分辨力。因此,窄波束HFSWR有广泛的应用前景。目前HFSWR海流流速提取技术非常成熟且已经商业化,然而有效波高和风速提取技术还有待进一步研究。本文以单站窄波束HFSWR为研究平台,以有效波高和风速为研究对象,重点研究远程多海况下有效波高和风速提取方法,并在国内外首次开展三维海洋表面动力学要素(Three-dimensional Ocean Surface Dynamics Element,3D-OSDE)即海流流速、有效波高和风速的联合预测方法研究。本研究成果为开展目标和海洋兼容探测提供理论依据和方法,有助于在现有目标探测HFSWR系统基础上,建立起一套可同时进行OSDEs提取及预测的信号处理方法。本文的主要研究内容如下:1.基于一阶谱的有效波高提取方法研究。首先详细介绍目前单站窄波束HFSWR中常用的有效波高提取方法,包括Barrick算法和基于一阶Bragg峰的有效波高提取方法,并指出这些方法中存在的问题,然后提出基于一阶谱的有效波高提取新方法,可突破Barrick算法有效波高提取下限,解决基于一阶Bragg峰的有效波高提取方法中存在的问题。2.远程多海况下有效波高提取方法研究。当进行远距离OSDEs提取时需要采用HFSWR载频低端,此时Barrick算法、基于一阶谱的有效波高提取新方法以及基于双频融合模式的有效波高提取方法均无法进行远距离多海况下有效波高提取。为此,本研究首先利用人工智能技术充分融合Barrick算法和基于一阶谱的有效波高提取新方法,提出基于单频融合模式的有效波高提取方法,进而实现单站窄波束HFSWR远程多海况下有效波高提取,实测数据分析表明本方法可获得满意的有效波高提取结果。随后对基于单频融合模式的有效波高提取方法中的分类方法进行进一步研究,针对传统前馈神经网络(Feedforward Neural Network,FNN)对时间序列数据建模性能有限和分类精度低等问题提出一种基于长短时记忆神经网络(Long Short-term Memory Neural Network,LSTMNN)的分类算法,同时针对梯度下降(Gradient Descent,GD)算法收敛速度慢和容易陷入局部极小值等问题提出一种基于无迹卡尔曼滤波器(Unscented Kalman Filter,UKF)的训练方法。实测数据分析表明本方法可以更好的进行远程多海况下有效波高提取。3.远程多海况下风速提取方法研究。在获取远距离多海况下的有效波高提取结果后,通常使用各种参数或非参数模型进行风速提取,其中参数模型拟合能力和泛化性能较差,常用的非参数模型不能充分利用历史时刻的有效波高信息,不适用于具有不同长度时序依赖特性的时间序列,而且输入矢量长度难以确定。为此,本研究提出一种基于LSTMNN的远程多海况下风速提取方法,本方法可以把时序特性包含在其反馈连接中,从而充分利用历史时刻的有效波高信息,且不需要已知输入矢量长度,是解决上述问题的一种有效手段。4.HFSWR OSDEs预测方法研究。对于3D-OSDE预测目前未见相关文献报道,本研究是对这一领域的全新探索。本研究依据多维信号处理领域最新成果,在四元数域进行3D-OSDE联合预测,并提出一种基于四元数值神经网络(Quaternion-valued Neural Network,QNN)的3D-OSDE联合预测方法。考虑到多海况下3D-OSDE时间序列具有很强的非平稳性,提出一种基于二阶导数的在线训练方法对QNN中的参数进行实时调整。
【Abstract】 By employing the propertie of low attenuation of high frequency electromagnetic wave(HFEW)propagating on the sea surface,the high frequency surface wave radar(HFSWR)can achieve large-area,all-weather,and real-time monitoring of coastal area.Therefore HFSWR is highly valued by countries around the world.HFSWR can be divided into wide-beam HFSWR and narrow-beam HFSWR,where the narrow-beam HFSWR can not only provide the position and motion information of the over-the-horizon target,but also perform fine extraction of ocean surface dynamics elements(OSDEs)in remote sea areas which are ocean current velocity,significant wave height,wind speed,and wind direction etc.Narrow-beam HFSWR is able to monitor larger area and has better azimuth resolution,hence it has broad application prospect.At present,HFSWR ocean current velocity extraction technology is very mature and has been commercialized.However the significant wave height and wind speed extraction technologies need to be further studied.This paper takes single-station narrow-beam HFSWR as the study platform,takes the significant wave height and wind speed as the research objects,and mainly studies the algorithms for extracting significant wave height and wind speed under multiple sea states in remote sea areas.Furthermore,we carry out the research on the joint prediction of three-dimensional ocean surface dynamics element(3D-OSDE)which consists of ocean current velocity,significant wave height,and wind speed for the first time.The research results provide theoretical basis and methods for achieving target and marine compatible detection,and are helpful to develop a set of signal processing methods that can simultaneously extract and predict OSDEs based on our school’s existing target detection HFSWR system.The main research contents are as following:1.Research on significant wave height extraction method based on the first-order spectrum.Firstly,the significant wave height extraction methods used in single-site narrow-beam HFSWR are introduced in detail,including Barrick algorithm and the first-order Bragg peaks-based method,and the problems in these methods are pointed out.Then a new method is proposed which is based on the first-order spectrum.The method can break through the lower limit of Barrick algorithm and solve the problem of the first-order Bragg peaks-based method.2.Research on significant wave height extraction method under multiple sea states in remote sea areas.Generally,a low-frequency HFSWR is more suitable to extract the OSDEs at a distance.However,Barrick algorithm,the new method based on the first-order spectrum,and the significant wave height extraction method based on dual-frequency fusion mode are not able to extract significant wave height using a low-frequency HFSWR under multiple sea states in remote sea areas.Therefore we employ artificial intelligence technology to fully integrate Barrick algorithm and the new method based on the first-order spectrum,and propose the significant wave height extraction method based on single-frequency fusion mode which can extract significant wave height using a low-frequency HFSWR under multiple sea states in remote sea areas.Simulations on real-world data show the significant wave height can be successfully extracted by the method and coincide well with the buoy significant wave height.Then the classification method in the significant wave height extraction method based on single-frequency fusion mode is further studied.Since feedforward neural network(FNN)has limited modeling capability and low classification accuracy for time series data,a classification algorithm based on long-short-term memory neural network(LSTMNN)is proposed.Because the gradient descent(GD)algorithm often suffers from a slow convergence speed and poor local optima,an unscented Kalman filter(UKF)based training algorithm is developed.Simulations on real-world data show the algorithm can achieve better performance.3.Research on wind speed extraction method under multiple sea states in remote sea areas.After obtaining the significant wave height,the wind speed can be extracted by employing the parametric or non-parametric models.However the parametric models have poor fitting and generalization performance.Moreover the commonly used non-parametric models can not make full use of the significant wave height information at historical moments,are not suitable for time series with timing-dependent characteristics of different lengths,and the input vector length is difficult to determine.Therefore we propose a wind speed extraction method based on the LSTMNN and the method can extract wind speed under multiple sea states in remote sea areas.The method can absorb timing-dependent characteristics into its feedback connections,make full use of the significant wave height information at historical moments,and does not need to determine the input vector length.Therefore this method is an effective way to solve the above problems.4.Research on prediction methods of OSDEs extracted by HFSWR.There is no related literature on the prediction of 3D-OSDE and this research is a new exploration in this field.Based on the latest research results in the field of multi-dimensional signal processing,we perform joint prediction of 3D-OSDE in quaternion domain and propose a 3D-OSDE joint prediction method based on quaternion-valued neural network(QNN).Considering that the 3D-OSDE time series is highly nonstationary under multiple sea states,we propose an online training method based on second-order derivative to adjust the parameters of QNN in real time.
【关键词】 高频地波雷达；
海洋表面动力学要素提取；
长短时记忆神经网络；
无迹卡尔曼滤波器；
四元数值信号处理；
【Key words】 HFSWR；
OSDEs extraction；
LSTMNN；
UKF；
quaternion-valued signal processing；
【网络出版投稿人】
哈尔滨工业大学
【网络出版年期】2021年
01期
【分类号】P714;TN958
【下载频次】132
攻读期成果
本文链接的文献网络图示:本文的引文网络}