声源DOA估计中的TDOA-DOA映射方法研究
发布时间：2018-08-04 07:46
【摘要】：声源波达方向(Direction Of Arrival,DOA)估计作为麦克风阵列信号处理中的一项关键技术,在视频会议系统、故障检测、医疗诊断、军事等许多领域都有广泛应用。基于多通道到达时间差(Time Differences Of Arrival,TDOA)的方法是声源DOA估计中的一种重要方法。然而当前研究工作主要集中在TDOA获取,而对TDOA-DOA映射方法研究较少。基于最小二乘支持向量回归机(Least Squares Support Vector Regression,LS-SVR)的TDOA-DOA映射方法有较好的声源DOA估计效果,但其研究并不全面。本文针对基于LS-SVR的TDOA-DOA映射方法,从LS-SVR中的核函数选取、多核LS-SVR构造以及稀疏化分析等方面进行了深入研究。此外,本文提出一种基于稀疏表示理论的无需调节参数的TDOA-DOA映射方法。本文的主要工作有:1)由于不同核函数具有不同的映射性能,因而本文研究了径向基核、多项式核以及线性核这三种常见核函数构造的LS-SVR在混响和噪声环境中的声源DOA估计性能,并与最小二乘映射方式进行了比较。研究结果表明,采用径向基核函数具有更高的估计性能。2)针对估计时延在混响较为严重的环境中出现离群值的问题,本文根据TDOA-DOA的映射特点,提出一种基于中值滤波的TDOA处理方法以消除离群值。研究结果表明,采用该方法后,在混响较为严重的环境中声源DOA映射性能得到了有效提升。3)为了进一步提升声源DOA估计性能,本文结合多核学习理论以及K-means聚类算法,提出了基于聚类方法的多核LS-SVR映射方法。仿真结果表明,多核LS-SVR的性能要优于单核LS-SVR以及最小二乘法;一般情况下,核的个数越多,多核LS-SVR的性能越好,并且混响时间越大,多核的性能优势体现得越明显。4)针对LS-SVR映射方法中训练集存在冗余这一问题,本文将基于最小支持权重的剪枝稀疏方法运用到声源DOA估计中,分别对单核和多核LS-SVR映射方法进行了稀疏化分析。研究结果表明,与基本LS-SVR相比,稀疏LS-SVR方法不仅能保持良好的DOA估计性能,而且有效减小了测试时的运算量。5)提出了一种基于稀疏表示理论的无需调节参数的TDOA-DOA映射方法。在此基础上,为进一步降低运算量,本文应用一种双步网格搜索方法来匹配TDOA向量和数据字典。研究结果表明,与传统的无需调节参数的映射方法相比,该算法存在一定的性能优势。
[Abstract]:As a key technology in microphone array signal processing, acoustic source DOA estimation (DOA) estimation is widely used in many fields such as video conferencing system, fault detection, medical diagnosis, military and so on. The method based on multi-channel time-of-arrival (Time Differences Of ArrivalTDOA) is an important method in sound source DOA estimation. However, the current research focuses on TDOA acquisition, but less on TDOA-DOA mapping methods. The TDOA-DOA mapping method based on least squares support vector regression machine (Least Squares Support Vector) has a good effect on sound source DOA estimation, but its research is not comprehensive. In this paper, the method of TDOA-DOA mapping based on LS-SVR is studied in detail from the aspects of kernel function selection, multi-core LS-SVR construction and sparse analysis in LS-SVR. In addition, this paper presents a TDOA-DOA mapping method based on sparse representation theory without adjusting parameters. The main work of this paper is: (1) because different kernel functions have different mapping performance, this paper studies the DOA estimation performance of LS-SVR in reverberation and noise environments with three common kernel functions, radial basis kernel, polynomial kernel and linear kernel. And compared with the least square mapping method. The results show that the radial basis function has higher estimation performance (.2). In order to solve the problem of outliers in the reverberation environment, the mapping characteristics of TDOA-DOA are discussed in this paper. A TDOA processing method based on median filter is proposed to eliminate outliers. The results show that the performance of sound source DOA mapping is improved by using this method in a more serious reverberation environment. In order to further improve the performance of sound source DOA estimation, this paper combines multi-core learning theory and K-means clustering algorithm. A multi-core LS-SVR mapping method based on clustering method is proposed. The simulation results show that the performance of multi-core LS-SVR is better than that of single core LS-SVR and least square method. In general, the more the number of cores, the better the performance and reverberation time of multi-core LS-SVR. The more obvious the performance advantage of multi-kernel is, the more obvious is the redundancy of training set in LS-SVR mapping method. In this paper, the pruning sparse method based on minimum support weight is applied to the sound source DOA estimation. The single core and multi-core LS-SVR mapping methods are analyzed respectively. The results show that compared with the basic LS-SVR, the sparse LS-SVR method can not only maintain good DOA estimation performance, but also reduce the computational complexity of the test effectively.) A new TDOA-DOA mapping method based on sparse representation theory without adjusting parameters is proposed. On this basis, a two-step grid search method is applied to match TDOA vectors and data dictionaries in order to further reduce the computational complexity. The results show that the proposed algorithm has some performance advantages compared with the traditional mapping method without adjusting parameters.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.23
本文编号：2163138
[Abstract]:As a key technology in microphone array signal processing, acoustic source DOA estimation (DOA) estimation is widely used in many fields such as video conferencing system, fault detection, medical diagnosis, military and so on. The method based on multi-channel time-of-arrival (Time Differences Of ArrivalTDOA) is an important method in sound source DOA estimation. However, the current research focuses on TDOA acquisition, but less on TDOA-DOA mapping methods. The TDOA-DOA mapping method based on least squares support vector regression machine (Least Squares Support Vector) has a good effect on sound source DOA estimation, but its research is not comprehensive. In this paper, the method of TDOA-DOA mapping based on LS-SVR is studied in detail from the aspects of kernel function selection, multi-core LS-SVR construction and sparse analysis in LS-SVR. In addition, this paper presents a TDOA-DOA mapping method based on sparse representation theory without adjusting parameters. The main work of this paper is: (1) because different kernel functions have different mapping performance, this paper studies the DOA estimation performance of LS-SVR in reverberation and noise environments with three common kernel functions, radial basis kernel, polynomial kernel and linear kernel. And compared with the least square mapping method. The results show that the radial basis function has higher estimation performance (.2). In order to solve the problem of outliers in the reverberation environment, the mapping characteristics of TDOA-DOA are discussed in this paper. A TDOA processing method based on median filter is proposed to eliminate outliers. The results show that the performance of sound source DOA mapping is improved by using this method in a more serious reverberation environment. In order to further improve the performance of sound source DOA estimation, this paper combines multi-core learning theory and K-means clustering algorithm. A multi-core LS-SVR mapping method based on clustering method is proposed. The simulation results show that the performance of multi-core LS-SVR is better than that of single core LS-SVR and least square method. In general, the more the number of cores, the better the performance and reverberation time of multi-core LS-SVR. The more obvious the performance advantage of multi-kernel is, the more obvious is the redundancy of training set in LS-SVR mapping method. In this paper, the pruning sparse method based on minimum support weight is applied to the sound source DOA estimation. The single core and multi-core LS-SVR mapping methods are analyzed respectively. The results show that compared with the basic LS-SVR, the sparse LS-SVR method can not only maintain good DOA estimation performance, but also reduce the computational complexity of the test effectively.) A new TDOA-DOA mapping method based on sparse representation theory without adjusting parameters is proposed. On this basis, a two-step grid search method is applied to match TDOA vectors and data dictionaries in order to further reduce the computational complexity. The results show that the proposed algorithm has some performance advantages compared with the traditional mapping method without adjusting parameters.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.23
【参考文献】
相关期刊论文 前1条
1 谭颖;殷福亮;李细林;;改进的SRP-PHAT声源定位方法[J];电子与信息学报;2006年07期
，本文编号：2163138
本文链接：https://www.wllwen.com/kejilunwen/wltx/2163138.html
