支持向量机在频率估计算法中的应用研究

发布时间：2018-01-14 05:00

本文关键词：支持向量机在频率估计算法中的应用研究　出处：《解放军信息工程大学》2014年博士论文　论文类型：学位论文

【摘要】：动态频谱聚合、机会使用其他业务空闲频谱资源的重要前提是不能干扰授权用户的正常通信。这首先需要对某固定频段进行实时频谱感知,即通过对此频段不间断的搜索、判别与分析,实时检测授权用户是否正在使用该频段,如果正在使用,进一步提取其信号的调制方式、频率以及带宽等特征参数,从而全面评估此频段频谱的特性及频率占用情况,找出适合通信的“频谱空洞”,在不影响已有通信的前提下伺机工作。其中授权用户信号的频率估计是整个频谱感知过程的重要环节,它的正确与否直接反映了授权用户对频谱的使用情况,从而客观地描述了此频段频谱的利用率。然而由于受到信号参数时变性、地理位置、传输距离等因素的影响,数据的小样本、低信噪比(Signal-to-Noise Ratio, SNR)特性很大程度上制约了频谱感知的性能。统计学习理论(Statistical Learning Theory, SLT)是一门专门研究小样本情况下机器学习规律的理论,它的出现为解决数据的小样本、低SNR问题带来一定的契机。支持向量机(Support Vector Machine, SVM)作为其具体实现方式,充分体现了结构风险最小化(Structural Risk Minimization, SRM)思想的精髓。它具有较好的泛化能力、高维处理能力和非线性处理能力,并且能够克服人工神经网络等机器学习方法存在的模型过拟合以及算法存在局部极小值等一些难以解决的具体问题。本文将SVM应用于频率估计相关问题的解决,重点针对基于离散Fourier变换(Discrete Fourier Transform, DFT)和基于相位频率估计算法,以及线性调频信号和Gaussian噪声条件下的频率估计问题等方面展开研究,并最终给出了基于SVM的频谱感知系统设计方案。主要工作和创新点如下：1.针对基于DFT的频率估计算法存在计算量与估计性能之间的矛盾,本文将最小二乘支持向量机回归(Least Squares Support Vector Regression, LS-SVR)看成一个线性内插器,同时将内插范围缩小到已知离散幅度谱最大值左右相邻两根谱线之间,提出了一种基于LS-SVR的DFT内插频率估计算法。2.针对影响基于相位频率估计算法的两个主要因素：相位噪声模型的近似性和相位解绕过程的不正确性,本文从接收信号实际相位与时间序列之间存在的线性关系出发,充分发挥SLT对小样本数据良好的学习能力与泛化能力,提出一种基于支持向量机回归(Support Vector Regression, SVR)的相位解绕与频率估计算法。3.针对噪声分布未知条件下的频率估计问题,本文从SLT出发,利用调制信息的有限字符集特性构造关于频率的SRM函数,将参数估计问题转化为求分类问题的极值,从而提出一种基于最小二乘支持向量机分类(Least Squares Support Vector Classification, LS-SVC)的星座图频率估计算法。4.针对线性调频信号实际相位与时间序列满足的二次关系，本文选用二次多项式核函数完成相位解绕过程。并利用SVR较强的非线性处理能力,准确估计线性调频信号的瞬时频率(Instantaneous Frequency, IF)、瞬时频率变化率(Instantaneous Frequency Rate, IFR)以及初始相位。5.针对国家科技重大专项“新一代无线宽带移动网”的子课题"IMT-A频谱聚合技术研发”中频谱感知的需求,本文设计了基于SVM的频谱感知系统设计方案。
[Abstract]:Dynamic spectrum aggregation, an important prerequisite for other business opportunities to use the idle spectrum resources is not normal communication interference to licensed users. This first needs to carry on the real-time spectrum sensing of a fixed frequency, namely the frequency of uninterrupted search, identification and analysis, real-time detection of authorized users is the use of the band, if you are using, further extraction the signal modulation, frequency and bandwidth parameters, thus the comprehensive assessment of the characteristics and frequency spectrum occupancy, to find a suitable communication "spectrum hole", without affecting the existing communication under the work. The authorized user to signal frequency estimation is an important part of the spectrum sensing process, it correct or not directly reflect the use of authorized users of the spectrum, so as to objectively describe the use of the frequency spectrum but due rate. By signal time-varying parameters, geographical location, transmission distance and other factors, the small sample data, low signal-to-noise ratio (Signal-to-Noise Ratio SNR) features largely restricted the performance of spectrum sensing. The statistical learning theory (Statistical Learning Theory, SLT) is a specialized research on machine learning method under the condition of small samples in theory, it appears to solve the small sample data, bring certain opportunity low SNR problem. Support vector machine (Support Vector Machine, SVM) as its implementation mode, fully embodies the structural risk minimization (Structural Risk, Minimization, SRM). The essence of it has good generalization ability, high the dimension of ability and nonlinear processing ability, and can overcome the artificial neural network in machine learning methods such as the model in the presence of local minima and some difficult to solve the over fitting and algorithm Specific issues. The application of SVM in frequency estimation to solve related problems, focusing on based on the discrete Fourier transform (Discrete Fourier Transform, DFT) and phase estimation algorithm based on frequency and linear frequency modulation signal and Gaussian noise under the condition of frequency estimation problem is studied, and finally gives the design scheme of spectrum sensing system SVM based on the main work and innovation are as follows: 1. aiming at the contradiction between algorithm and estimation performance estimation based on the frequency of DFT, the least squares support vector machine regression (Least Squares Support Vector Regression, LS-SVR) as a linear interpolator, the interpolation range between the maximum known discrete amplitude spectrum about two adjacent spectral lines, we propose a LS-SVR based DFT interpolation algorithm for frequency estimation based on phase frequency estimation for.2. Two main factors: the phase noise model of the algorithm and the approximation of the phase unwrapping process is not correct, the linear relationship between the received signal and the actual phase of time series, give full play to SLT on the learning ability and generalization ability of small sample data, proposed a regression based on support vector machine (Support Vector Regression, SVR) of the phase unwrapping and frequency estimation algorithm for.3. noise distribution under the condition of unknown frequency estimation problem, this paper starting from SLT, SRM function characters using modulation information set about frequency characteristics, the parameter estimation problem into extremum classification problems, and put forward a least squares support vector machine (Least Squares Support Vector classification based on Classification, LS-SVC) constellation.4. frequency estimation algorithm for LFM signal phase and the actual time The two time sequence relationship, the complete polynomial kernel function two phase unwrapping process. And the nonlinear processing ability of SVR strong, accurate instantaneous frequency estimation of LFM signals (Instantaneous, Frequency, IF), the instantaneous frequency rate (Instantaneous Frequency, Rate, IFR) and.5. in the initial phase of major national science and technology projects "a new generation of broadband wireless mobile network project of IMT-A spectrum aggregation technology research" spectrum sensing requirements, this paper designs the design of spectrum sensing system of SVM based on the case.

【学位授予单位】：解放军信息工程大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN911.23;TP181

【参考文献】