【作者】 赵振兴；

【导师】 杨凌；

【作者基本信息】 兰州大学 ， 信息与通信工程·通信与信息系统， 2016， 硕士

【摘要】 短波通信系统中,信号在信道中传输时受到时延、多径衰落和噪声等的干扰,会产生码间干扰(InterSymbol Interference,ISI),从而使发送信号在接收端无法正确识别。消除或减小ISI的主要技术是信道均衡。盲均衡是不需要训练序列,就能消除ISI而恢复输入信号的信道均衡技术。短波突发通信由于抗干扰和抗截获能力强、保密性能高而被广泛应用在军事通信领域。短波突发信号与传统短波信号相比,符号数通常是几百,有时甚至仅是几十,较少的数据符号对盲均衡技术提出了更高的要求。盲均衡技术中,最传统的方法是高阶统计量法,但由于其利用数据的高阶统计特性使得算法运算量特别大,而且需要大量的数据才能收敛并且收敛的速度很慢。这些缺点限制了其在小数据信号处理领域的应用。近些年,支持向量机(Support Vector Machnie,SVM)和相关向量机(Relevance Vector Machine,RVM)等机器学习方法被广泛应用于盲均衡。与基于高阶统计量的盲均衡方法相比,基于SVM和RVM的盲均衡器需要较小数据样本即可达到所要求的均衡水平。相比于SVM,由于RVM基于贝叶斯架构,所以RVM均衡器比SVM均衡器具有更好的收敛性及更稀疏的检测模型。传统的SVM或RVM通常使用单一核函数,插值能力和外推能力不能兼而得之,而核函数的选择对均衡器的性能有很大的影响。因此,本文提出采用混合核函数(Hybrid-kernel,Hk)产生设计矩阵,使得基于混合核函数的RVM(Hybrid-kernel Relevance Vector Machine,Hk-RVM)比传统的RVM有更好的插值能力和外推能力。本文研究了盲均衡的基本原理,分析了传统的盲均衡方法及各自的优缺点,对RVM的原理及基于RVM的盲均衡方法进行了详细阐述并做了实验仿真。提出了基于Hk-RVM的盲均衡理论和算法,分别对比了Hk-RVM与RVM以及Hk-RVM与SVM的盲均衡算法性能。实验结果表明,Hk-RVM盲均衡器比RVM及SVM盲均衡器具有更好的稀疏性且误码率更低。此外,Hk-RVM和RVM盲均衡器基于贝叶斯架构,使得其均衡稳定性高于SVM盲均衡器。更多还原

【Abstract】 In the shortwave communication system, the Inter-Symbol Interference(ISI) is often encountered so that sending signals cannot be recognized correctly by receiving terminal, which because sending signals are interfered by time delay, multipath fading and noise, etc.. Channel equalization is the main technology to eliminate or reduce ISI. Blind equalization is a signal processing technique without resorting to the training sequence, which could suppress ISI and recover the input signals. With better security, anti-jamming and anti-interception capability, shortwave burst communication is widely used in the military communicate field. Comparing with the traditional shortwave signal, shortwave burst signal has less symbolic number, usually a few hundred, sometimes only a few decades. However, less data symbols put forward higher requirement for the blind equalization.Higher-order statistics(HOS) is the most classical method in the blind equalization, but it has a a large amount of calculation by using the characteristics of HOS and needs plenty of data to be convergent in slow speed, which restrains blind equalization to be used in small data signal processing field. In recent years, machine learning technologies are widely applied in the blind equalization, such as Support Vector Machine(SVM), Relevance Vector Machine(RVM), etc.. The blind equalization based on SVM and RVM can meet the required equilibrium levels with less sample data than that based on HOS. Basing on the Bayesian framework, RVM equalizer has better convergence and sparser detection model compared with SVM.Usaually, SVM or RVM uses single kernel function, and can’t get interpolation and extrapolation ability simultaneously, moreover, the choice of kernel function influences the performance of equalizer greatly. Therefore, a Hybrid-kernel Relevance Vector Machine(Hk-RVM) which uses Hybrid-kernel to generate design matrix is proposed in this paper in order to make the Hk-RVM get superior interpolation and extrapolation ability than traditional RVM.We study the basic principle of blind equalization, analyse traditional method of blind equalization and it’s advantages and disadvantages. Detailing RVM’s principle and the blind equalization method based on RVM, then make plenty of simulation experiments. Proposing a blind equalization theory and algorithm based on Hk-RVM, compare performances of the blind equalization algorithm’s basde on Hk-RVM with RVM and SVM, respectively. The experiment results show that Hk-RVM equalizer is sparser and has lower bit-error rate than RVM and SVM equalizer. Furthermore, because of the Bayesian framework, the blind equalization based on Hk-RVM and RVM have higher evolutionary stability than that on SVM.更多还原